xref: /titanic_44/usr/src/uts/common/fs/zfs/dsl_dataset.c (revision 528b7d8ba791f2da280ff1ddd45c61eb47a2744e)
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) 2011 by Delphix. All rights reserved.
24  */
25 
26 #include <sys/dmu_objset.h>
27 #include <sys/dsl_dataset.h>
28 #include <sys/dsl_dir.h>
29 #include <sys/dsl_prop.h>
30 #include <sys/dsl_synctask.h>
31 #include <sys/dmu_traverse.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/arc.h>
34 #include <sys/zio.h>
35 #include <sys/zap.h>
36 #include <sys/unique.h>
37 #include <sys/zfs_context.h>
38 #include <sys/zfs_ioctl.h>
39 #include <sys/spa.h>
40 #include <sys/zfs_znode.h>
41 #include <sys/zfs_onexit.h>
42 #include <sys/zvol.h>
43 #include <sys/dsl_scan.h>
44 #include <sys/dsl_deadlist.h>
45 
46 static char *dsl_reaper = "the grim reaper";
47 
48 static dsl_checkfunc_t dsl_dataset_destroy_begin_check;
49 static dsl_syncfunc_t dsl_dataset_destroy_begin_sync;
50 static dsl_syncfunc_t dsl_dataset_set_reservation_sync;
51 
52 #define	SWITCH64(x, y) \
53 	{ \
54 		uint64_t __tmp = (x); \
55 		(x) = (y); \
56 		(y) = __tmp; \
57 	}
58 
59 #define	DS_REF_MAX	(1ULL << 62)
60 
61 #define	DSL_DEADLIST_BLOCKSIZE	SPA_MAXBLOCKSIZE
62 
63 #define	DSL_DATASET_IS_DESTROYED(ds)	((ds)->ds_owner == dsl_reaper)
64 
65 
66 /*
67  * Figure out how much of this delta should be propogated to the dsl_dir
68  * layer.  If there's a refreservation, that space has already been
69  * partially accounted for in our ancestors.
70  */
71 static int64_t
72 parent_delta(dsl_dataset_t *ds, int64_t delta)
73 {
74 	uint64_t old_bytes, new_bytes;
75 
76 	if (ds->ds_reserved == 0)
77 		return (delta);
78 
79 	old_bytes = MAX(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
80 	new_bytes = MAX(ds->ds_phys->ds_unique_bytes + delta, ds->ds_reserved);
81 
82 	ASSERT3U(ABS((int64_t)(new_bytes - old_bytes)), <=, ABS(delta));
83 	return (new_bytes - old_bytes);
84 }
85 
86 void
87 dsl_dataset_block_born(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx)
88 {
89 	int used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
90 	int compressed = BP_GET_PSIZE(bp);
91 	int uncompressed = BP_GET_UCSIZE(bp);
92 	int64_t delta;
93 
94 	dprintf_bp(bp, "ds=%p", ds);
95 
96 	ASSERT(dmu_tx_is_syncing(tx));
97 	/* It could have been compressed away to nothing */
98 	if (BP_IS_HOLE(bp))
99 		return;
100 	ASSERT(BP_GET_TYPE(bp) != DMU_OT_NONE);
101 	ASSERT3U(BP_GET_TYPE(bp), <, DMU_OT_NUMTYPES);
102 	if (ds == NULL) {
103 		/*
104 		 * Account for the meta-objset space in its placeholder
105 		 * dsl_dir.
106 		 */
107 		ASSERT3U(compressed, ==, uncompressed); /* it's all metadata */
108 		dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
109 		    used, compressed, uncompressed, tx);
110 		dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
111 		return;
112 	}
113 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
114 
115 	mutex_enter(&ds->ds_dir->dd_lock);
116 	mutex_enter(&ds->ds_lock);
117 	delta = parent_delta(ds, used);
118 	ds->ds_phys->ds_used_bytes += used;
119 	ds->ds_phys->ds_compressed_bytes += compressed;
120 	ds->ds_phys->ds_uncompressed_bytes += uncompressed;
121 	ds->ds_phys->ds_unique_bytes += used;
122 	mutex_exit(&ds->ds_lock);
123 	dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD, delta,
124 	    compressed, uncompressed, tx);
125 	dsl_dir_transfer_space(ds->ds_dir, used - delta,
126 	    DD_USED_REFRSRV, DD_USED_HEAD, tx);
127 	mutex_exit(&ds->ds_dir->dd_lock);
128 }
129 
130 int
131 dsl_dataset_block_kill(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx,
132     boolean_t async)
133 {
134 	if (BP_IS_HOLE(bp))
135 		return (0);
136 
137 	ASSERT(dmu_tx_is_syncing(tx));
138 	ASSERT(bp->blk_birth <= tx->tx_txg);
139 
140 	int used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
141 	int compressed = BP_GET_PSIZE(bp);
142 	int uncompressed = BP_GET_UCSIZE(bp);
143 
144 	ASSERT(used > 0);
145 	if (ds == NULL) {
146 		/*
147 		 * Account for the meta-objset space in its placeholder
148 		 * dataset.
149 		 */
150 		dsl_free(tx->tx_pool, tx->tx_txg, bp);
151 
152 		dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
153 		    -used, -compressed, -uncompressed, tx);
154 		dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
155 		return (used);
156 	}
157 	ASSERT3P(tx->tx_pool, ==, ds->ds_dir->dd_pool);
158 
159 	ASSERT(!dsl_dataset_is_snapshot(ds));
160 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
161 
162 	if (bp->blk_birth > ds->ds_phys->ds_prev_snap_txg) {
163 		int64_t delta;
164 
165 		dprintf_bp(bp, "freeing ds=%llu", ds->ds_object);
166 		dsl_free(tx->tx_pool, tx->tx_txg, bp);
167 
168 		mutex_enter(&ds->ds_dir->dd_lock);
169 		mutex_enter(&ds->ds_lock);
170 		ASSERT(ds->ds_phys->ds_unique_bytes >= used ||
171 		    !DS_UNIQUE_IS_ACCURATE(ds));
172 		delta = parent_delta(ds, -used);
173 		ds->ds_phys->ds_unique_bytes -= used;
174 		mutex_exit(&ds->ds_lock);
175 		dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
176 		    delta, -compressed, -uncompressed, tx);
177 		dsl_dir_transfer_space(ds->ds_dir, -used - delta,
178 		    DD_USED_REFRSRV, DD_USED_HEAD, tx);
179 		mutex_exit(&ds->ds_dir->dd_lock);
180 	} else {
181 		dprintf_bp(bp, "putting on dead list: %s", "");
182 		if (async) {
183 			/*
184 			 * We are here as part of zio's write done callback,
185 			 * which means we're a zio interrupt thread.  We can't
186 			 * call dsl_deadlist_insert() now because it may block
187 			 * waiting for I/O.  Instead, put bp on the deferred
188 			 * queue and let dsl_pool_sync() finish the job.
189 			 */
190 			bplist_append(&ds->ds_pending_deadlist, bp);
191 		} else {
192 			dsl_deadlist_insert(&ds->ds_deadlist, bp, tx);
193 		}
194 		ASSERT3U(ds->ds_prev->ds_object, ==,
195 		    ds->ds_phys->ds_prev_snap_obj);
196 		ASSERT(ds->ds_prev->ds_phys->ds_num_children > 0);
197 		/* if (bp->blk_birth > prev prev snap txg) prev unique += bs */
198 		if (ds->ds_prev->ds_phys->ds_next_snap_obj ==
199 		    ds->ds_object && bp->blk_birth >
200 		    ds->ds_prev->ds_phys->ds_prev_snap_txg) {
201 			dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
202 			mutex_enter(&ds->ds_prev->ds_lock);
203 			ds->ds_prev->ds_phys->ds_unique_bytes += used;
204 			mutex_exit(&ds->ds_prev->ds_lock);
205 		}
206 		if (bp->blk_birth > ds->ds_dir->dd_origin_txg) {
207 			dsl_dir_transfer_space(ds->ds_dir, used,
208 			    DD_USED_HEAD, DD_USED_SNAP, tx);
209 		}
210 	}
211 	mutex_enter(&ds->ds_lock);
212 	ASSERT3U(ds->ds_phys->ds_used_bytes, >=, used);
213 	ds->ds_phys->ds_used_bytes -= used;
214 	ASSERT3U(ds->ds_phys->ds_compressed_bytes, >=, compressed);
215 	ds->ds_phys->ds_compressed_bytes -= compressed;
216 	ASSERT3U(ds->ds_phys->ds_uncompressed_bytes, >=, uncompressed);
217 	ds->ds_phys->ds_uncompressed_bytes -= uncompressed;
218 	mutex_exit(&ds->ds_lock);
219 
220 	return (used);
221 }
222 
223 uint64_t
224 dsl_dataset_prev_snap_txg(dsl_dataset_t *ds)
225 {
226 	uint64_t trysnap = 0;
227 
228 	if (ds == NULL)
229 		return (0);
230 	/*
231 	 * The snapshot creation could fail, but that would cause an
232 	 * incorrect FALSE return, which would only result in an
233 	 * overestimation of the amount of space that an operation would
234 	 * consume, which is OK.
235 	 *
236 	 * There's also a small window where we could miss a pending
237 	 * snapshot, because we could set the sync task in the quiescing
238 	 * phase.  So this should only be used as a guess.
239 	 */
240 	if (ds->ds_trysnap_txg >
241 	    spa_last_synced_txg(ds->ds_dir->dd_pool->dp_spa))
242 		trysnap = ds->ds_trysnap_txg;
243 	return (MAX(ds->ds_phys->ds_prev_snap_txg, trysnap));
244 }
245 
246 boolean_t
247 dsl_dataset_block_freeable(dsl_dataset_t *ds, const blkptr_t *bp,
248     uint64_t blk_birth)
249 {
250 	if (blk_birth <= dsl_dataset_prev_snap_txg(ds))
251 		return (B_FALSE);
252 
253 	ddt_prefetch(dsl_dataset_get_spa(ds), bp);
254 
255 	return (B_TRUE);
256 }
257 
258 /* ARGSUSED */
259 static void
260 dsl_dataset_evict(dmu_buf_t *db, void *dsv)
261 {
262 	dsl_dataset_t *ds = dsv;
263 
264 	ASSERT(ds->ds_owner == NULL || DSL_DATASET_IS_DESTROYED(ds));
265 
266 	unique_remove(ds->ds_fsid_guid);
267 
268 	if (ds->ds_objset != NULL)
269 		dmu_objset_evict(ds->ds_objset);
270 
271 	if (ds->ds_prev) {
272 		dsl_dataset_drop_ref(ds->ds_prev, ds);
273 		ds->ds_prev = NULL;
274 	}
275 
276 	bplist_destroy(&ds->ds_pending_deadlist);
277 	if (db != NULL) {
278 		dsl_deadlist_close(&ds->ds_deadlist);
279 	} else {
280 		ASSERT(ds->ds_deadlist.dl_dbuf == NULL);
281 		ASSERT(!ds->ds_deadlist.dl_oldfmt);
282 	}
283 	if (ds->ds_dir)
284 		dsl_dir_close(ds->ds_dir, ds);
285 
286 	ASSERT(!list_link_active(&ds->ds_synced_link));
287 
288 	mutex_destroy(&ds->ds_lock);
289 	mutex_destroy(&ds->ds_recvlock);
290 	mutex_destroy(&ds->ds_opening_lock);
291 	rw_destroy(&ds->ds_rwlock);
292 	cv_destroy(&ds->ds_exclusive_cv);
293 
294 	kmem_free(ds, sizeof (dsl_dataset_t));
295 }
296 
297 static int
298 dsl_dataset_get_snapname(dsl_dataset_t *ds)
299 {
300 	dsl_dataset_phys_t *headphys;
301 	int err;
302 	dmu_buf_t *headdbuf;
303 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
304 	objset_t *mos = dp->dp_meta_objset;
305 
306 	if (ds->ds_snapname[0])
307 		return (0);
308 	if (ds->ds_phys->ds_next_snap_obj == 0)
309 		return (0);
310 
311 	err = dmu_bonus_hold(mos, ds->ds_dir->dd_phys->dd_head_dataset_obj,
312 	    FTAG, &headdbuf);
313 	if (err)
314 		return (err);
315 	headphys = headdbuf->db_data;
316 	err = zap_value_search(dp->dp_meta_objset,
317 	    headphys->ds_snapnames_zapobj, ds->ds_object, 0, ds->ds_snapname);
318 	dmu_buf_rele(headdbuf, FTAG);
319 	return (err);
320 }
321 
322 static int
323 dsl_dataset_snap_lookup(dsl_dataset_t *ds, const char *name, uint64_t *value)
324 {
325 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
326 	uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
327 	matchtype_t mt;
328 	int err;
329 
330 	if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
331 		mt = MT_FIRST;
332 	else
333 		mt = MT_EXACT;
334 
335 	err = zap_lookup_norm(mos, snapobj, name, 8, 1,
336 	    value, mt, NULL, 0, NULL);
337 	if (err == ENOTSUP && mt == MT_FIRST)
338 		err = zap_lookup(mos, snapobj, name, 8, 1, value);
339 	return (err);
340 }
341 
342 static int
343 dsl_dataset_snap_remove(dsl_dataset_t *ds, char *name, dmu_tx_t *tx)
344 {
345 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
346 	uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
347 	matchtype_t mt;
348 	int err;
349 
350 	dsl_dir_snap_cmtime_update(ds->ds_dir);
351 
352 	if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
353 		mt = MT_FIRST;
354 	else
355 		mt = MT_EXACT;
356 
357 	err = zap_remove_norm(mos, snapobj, name, mt, tx);
358 	if (err == ENOTSUP && mt == MT_FIRST)
359 		err = zap_remove(mos, snapobj, name, tx);
360 	return (err);
361 }
362 
363 static int
364 dsl_dataset_get_ref(dsl_pool_t *dp, uint64_t dsobj, void *tag,
365     dsl_dataset_t **dsp)
366 {
367 	objset_t *mos = dp->dp_meta_objset;
368 	dmu_buf_t *dbuf;
369 	dsl_dataset_t *ds;
370 	int err;
371 	dmu_object_info_t doi;
372 
373 	ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
374 	    dsl_pool_sync_context(dp));
375 
376 	err = dmu_bonus_hold(mos, dsobj, tag, &dbuf);
377 	if (err)
378 		return (err);
379 
380 	/* Make sure dsobj has the correct object type. */
381 	dmu_object_info_from_db(dbuf, &doi);
382 	if (doi.doi_type != DMU_OT_DSL_DATASET)
383 		return (EINVAL);
384 
385 	ds = dmu_buf_get_user(dbuf);
386 	if (ds == NULL) {
387 		dsl_dataset_t *winner;
388 
389 		ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_SLEEP);
390 		ds->ds_dbuf = dbuf;
391 		ds->ds_object = dsobj;
392 		ds->ds_phys = dbuf->db_data;
393 
394 		mutex_init(&ds->ds_lock, NULL, MUTEX_DEFAULT, NULL);
395 		mutex_init(&ds->ds_recvlock, NULL, MUTEX_DEFAULT, NULL);
396 		mutex_init(&ds->ds_opening_lock, NULL, MUTEX_DEFAULT, NULL);
397 		rw_init(&ds->ds_rwlock, 0, 0, 0);
398 		cv_init(&ds->ds_exclusive_cv, NULL, CV_DEFAULT, NULL);
399 
400 		bplist_create(&ds->ds_pending_deadlist);
401 		dsl_deadlist_open(&ds->ds_deadlist,
402 		    mos, ds->ds_phys->ds_deadlist_obj);
403 
404 		if (err == 0) {
405 			err = dsl_dir_open_obj(dp,
406 			    ds->ds_phys->ds_dir_obj, NULL, ds, &ds->ds_dir);
407 		}
408 		if (err) {
409 			mutex_destroy(&ds->ds_lock);
410 			mutex_destroy(&ds->ds_recvlock);
411 			mutex_destroy(&ds->ds_opening_lock);
412 			rw_destroy(&ds->ds_rwlock);
413 			cv_destroy(&ds->ds_exclusive_cv);
414 			bplist_destroy(&ds->ds_pending_deadlist);
415 			dsl_deadlist_close(&ds->ds_deadlist);
416 			kmem_free(ds, sizeof (dsl_dataset_t));
417 			dmu_buf_rele(dbuf, tag);
418 			return (err);
419 		}
420 
421 		if (!dsl_dataset_is_snapshot(ds)) {
422 			ds->ds_snapname[0] = '\0';
423 			if (ds->ds_phys->ds_prev_snap_obj) {
424 				err = dsl_dataset_get_ref(dp,
425 				    ds->ds_phys->ds_prev_snap_obj,
426 				    ds, &ds->ds_prev);
427 			}
428 		} else {
429 			if (zfs_flags & ZFS_DEBUG_SNAPNAMES)
430 				err = dsl_dataset_get_snapname(ds);
431 			if (err == 0 && ds->ds_phys->ds_userrefs_obj != 0) {
432 				err = zap_count(
433 				    ds->ds_dir->dd_pool->dp_meta_objset,
434 				    ds->ds_phys->ds_userrefs_obj,
435 				    &ds->ds_userrefs);
436 			}
437 		}
438 
439 		if (err == 0 && !dsl_dataset_is_snapshot(ds)) {
440 			/*
441 			 * In sync context, we're called with either no lock
442 			 * or with the write lock.  If we're not syncing,
443 			 * we're always called with the read lock held.
444 			 */
445 			boolean_t need_lock =
446 			    !RW_WRITE_HELD(&dp->dp_config_rwlock) &&
447 			    dsl_pool_sync_context(dp);
448 
449 			if (need_lock)
450 				rw_enter(&dp->dp_config_rwlock, RW_READER);
451 
452 			err = dsl_prop_get_ds(ds,
453 			    "refreservation", sizeof (uint64_t), 1,
454 			    &ds->ds_reserved, NULL);
455 			if (err == 0) {
456 				err = dsl_prop_get_ds(ds,
457 				    "refquota", sizeof (uint64_t), 1,
458 				    &ds->ds_quota, NULL);
459 			}
460 
461 			if (need_lock)
462 				rw_exit(&dp->dp_config_rwlock);
463 		} else {
464 			ds->ds_reserved = ds->ds_quota = 0;
465 		}
466 
467 		if (err == 0) {
468 			winner = dmu_buf_set_user_ie(dbuf, ds, &ds->ds_phys,
469 			    dsl_dataset_evict);
470 		}
471 		if (err || winner) {
472 			bplist_destroy(&ds->ds_pending_deadlist);
473 			dsl_deadlist_close(&ds->ds_deadlist);
474 			if (ds->ds_prev)
475 				dsl_dataset_drop_ref(ds->ds_prev, ds);
476 			dsl_dir_close(ds->ds_dir, ds);
477 			mutex_destroy(&ds->ds_lock);
478 			mutex_destroy(&ds->ds_recvlock);
479 			mutex_destroy(&ds->ds_opening_lock);
480 			rw_destroy(&ds->ds_rwlock);
481 			cv_destroy(&ds->ds_exclusive_cv);
482 			kmem_free(ds, sizeof (dsl_dataset_t));
483 			if (err) {
484 				dmu_buf_rele(dbuf, tag);
485 				return (err);
486 			}
487 			ds = winner;
488 		} else {
489 			ds->ds_fsid_guid =
490 			    unique_insert(ds->ds_phys->ds_fsid_guid);
491 		}
492 	}
493 	ASSERT3P(ds->ds_dbuf, ==, dbuf);
494 	ASSERT3P(ds->ds_phys, ==, dbuf->db_data);
495 	ASSERT(ds->ds_phys->ds_prev_snap_obj != 0 ||
496 	    spa_version(dp->dp_spa) < SPA_VERSION_ORIGIN ||
497 	    dp->dp_origin_snap == NULL || ds == dp->dp_origin_snap);
498 	mutex_enter(&ds->ds_lock);
499 	if (!dsl_pool_sync_context(dp) && DSL_DATASET_IS_DESTROYED(ds)) {
500 		mutex_exit(&ds->ds_lock);
501 		dmu_buf_rele(ds->ds_dbuf, tag);
502 		return (ENOENT);
503 	}
504 	mutex_exit(&ds->ds_lock);
505 	*dsp = ds;
506 	return (0);
507 }
508 
509 static int
510 dsl_dataset_hold_ref(dsl_dataset_t *ds, void *tag)
511 {
512 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
513 
514 	/*
515 	 * In syncing context we don't want the rwlock lock: there
516 	 * may be an existing writer waiting for sync phase to
517 	 * finish.  We don't need to worry about such writers, since
518 	 * sync phase is single-threaded, so the writer can't be
519 	 * doing anything while we are active.
520 	 */
521 	if (dsl_pool_sync_context(dp)) {
522 		ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
523 		return (0);
524 	}
525 
526 	/*
527 	 * Normal users will hold the ds_rwlock as a READER until they
528 	 * are finished (i.e., call dsl_dataset_rele()).  "Owners" will
529 	 * drop their READER lock after they set the ds_owner field.
530 	 *
531 	 * If the dataset is being destroyed, the destroy thread will
532 	 * obtain a WRITER lock for exclusive access after it's done its
533 	 * open-context work and then change the ds_owner to
534 	 * dsl_reaper once destruction is assured.  So threads
535 	 * may block here temporarily, until the "destructability" of
536 	 * the dataset is determined.
537 	 */
538 	ASSERT(!RW_WRITE_HELD(&dp->dp_config_rwlock));
539 	mutex_enter(&ds->ds_lock);
540 	while (!rw_tryenter(&ds->ds_rwlock, RW_READER)) {
541 		rw_exit(&dp->dp_config_rwlock);
542 		cv_wait(&ds->ds_exclusive_cv, &ds->ds_lock);
543 		if (DSL_DATASET_IS_DESTROYED(ds)) {
544 			mutex_exit(&ds->ds_lock);
545 			dsl_dataset_drop_ref(ds, tag);
546 			rw_enter(&dp->dp_config_rwlock, RW_READER);
547 			return (ENOENT);
548 		}
549 		/*
550 		 * The dp_config_rwlock lives above the ds_lock. And
551 		 * we need to check DSL_DATASET_IS_DESTROYED() while
552 		 * holding the ds_lock, so we have to drop and reacquire
553 		 * the ds_lock here.
554 		 */
555 		mutex_exit(&ds->ds_lock);
556 		rw_enter(&dp->dp_config_rwlock, RW_READER);
557 		mutex_enter(&ds->ds_lock);
558 	}
559 	mutex_exit(&ds->ds_lock);
560 	return (0);
561 }
562 
563 int
564 dsl_dataset_hold_obj(dsl_pool_t *dp, uint64_t dsobj, void *tag,
565     dsl_dataset_t **dsp)
566 {
567 	int err = dsl_dataset_get_ref(dp, dsobj, tag, dsp);
568 
569 	if (err)
570 		return (err);
571 	return (dsl_dataset_hold_ref(*dsp, tag));
572 }
573 
574 int
575 dsl_dataset_own_obj(dsl_pool_t *dp, uint64_t dsobj, boolean_t inconsistentok,
576     void *tag, dsl_dataset_t **dsp)
577 {
578 	int err = dsl_dataset_hold_obj(dp, dsobj, tag, dsp);
579 	if (err)
580 		return (err);
581 	if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
582 		dsl_dataset_rele(*dsp, tag);
583 		*dsp = NULL;
584 		return (EBUSY);
585 	}
586 	return (0);
587 }
588 
589 int
590 dsl_dataset_hold(const char *name, void *tag, dsl_dataset_t **dsp)
591 {
592 	dsl_dir_t *dd;
593 	dsl_pool_t *dp;
594 	const char *snapname;
595 	uint64_t obj;
596 	int err = 0;
597 
598 	err = dsl_dir_open_spa(NULL, name, FTAG, &dd, &snapname);
599 	if (err)
600 		return (err);
601 
602 	dp = dd->dd_pool;
603 	obj = dd->dd_phys->dd_head_dataset_obj;
604 	rw_enter(&dp->dp_config_rwlock, RW_READER);
605 	if (obj)
606 		err = dsl_dataset_get_ref(dp, obj, tag, dsp);
607 	else
608 		err = ENOENT;
609 	if (err)
610 		goto out;
611 
612 	err = dsl_dataset_hold_ref(*dsp, tag);
613 
614 	/* we may be looking for a snapshot */
615 	if (err == 0 && snapname != NULL) {
616 		dsl_dataset_t *ds = NULL;
617 
618 		if (*snapname++ != '@') {
619 			dsl_dataset_rele(*dsp, tag);
620 			err = ENOENT;
621 			goto out;
622 		}
623 
624 		dprintf("looking for snapshot '%s'\n", snapname);
625 		err = dsl_dataset_snap_lookup(*dsp, snapname, &obj);
626 		if (err == 0)
627 			err = dsl_dataset_get_ref(dp, obj, tag, &ds);
628 		dsl_dataset_rele(*dsp, tag);
629 
630 		ASSERT3U((err == 0), ==, (ds != NULL));
631 
632 		if (ds) {
633 			mutex_enter(&ds->ds_lock);
634 			if (ds->ds_snapname[0] == 0)
635 				(void) strlcpy(ds->ds_snapname, snapname,
636 				    sizeof (ds->ds_snapname));
637 			mutex_exit(&ds->ds_lock);
638 			err = dsl_dataset_hold_ref(ds, tag);
639 			*dsp = err ? NULL : ds;
640 		}
641 	}
642 out:
643 	rw_exit(&dp->dp_config_rwlock);
644 	dsl_dir_close(dd, FTAG);
645 	return (err);
646 }
647 
648 int
649 dsl_dataset_own(const char *name, boolean_t inconsistentok,
650     void *tag, dsl_dataset_t **dsp)
651 {
652 	int err = dsl_dataset_hold(name, tag, dsp);
653 	if (err)
654 		return (err);
655 	if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
656 		dsl_dataset_rele(*dsp, tag);
657 		return (EBUSY);
658 	}
659 	return (0);
660 }
661 
662 void
663 dsl_dataset_name(dsl_dataset_t *ds, char *name)
664 {
665 	if (ds == NULL) {
666 		(void) strcpy(name, "mos");
667 	} else {
668 		dsl_dir_name(ds->ds_dir, name);
669 		VERIFY(0 == dsl_dataset_get_snapname(ds));
670 		if (ds->ds_snapname[0]) {
671 			(void) strcat(name, "@");
672 			/*
673 			 * We use a "recursive" mutex so that we
674 			 * can call dprintf_ds() with ds_lock held.
675 			 */
676 			if (!MUTEX_HELD(&ds->ds_lock)) {
677 				mutex_enter(&ds->ds_lock);
678 				(void) strcat(name, ds->ds_snapname);
679 				mutex_exit(&ds->ds_lock);
680 			} else {
681 				(void) strcat(name, ds->ds_snapname);
682 			}
683 		}
684 	}
685 }
686 
687 static int
688 dsl_dataset_namelen(dsl_dataset_t *ds)
689 {
690 	int result;
691 
692 	if (ds == NULL) {
693 		result = 3;	/* "mos" */
694 	} else {
695 		result = dsl_dir_namelen(ds->ds_dir);
696 		VERIFY(0 == dsl_dataset_get_snapname(ds));
697 		if (ds->ds_snapname[0]) {
698 			++result;	/* adding one for the @-sign */
699 			if (!MUTEX_HELD(&ds->ds_lock)) {
700 				mutex_enter(&ds->ds_lock);
701 				result += strlen(ds->ds_snapname);
702 				mutex_exit(&ds->ds_lock);
703 			} else {
704 				result += strlen(ds->ds_snapname);
705 			}
706 		}
707 	}
708 
709 	return (result);
710 }
711 
712 void
713 dsl_dataset_drop_ref(dsl_dataset_t *ds, void *tag)
714 {
715 	dmu_buf_rele(ds->ds_dbuf, tag);
716 }
717 
718 void
719 dsl_dataset_rele(dsl_dataset_t *ds, void *tag)
720 {
721 	if (!dsl_pool_sync_context(ds->ds_dir->dd_pool)) {
722 		rw_exit(&ds->ds_rwlock);
723 	}
724 	dsl_dataset_drop_ref(ds, tag);
725 }
726 
727 void
728 dsl_dataset_disown(dsl_dataset_t *ds, void *tag)
729 {
730 	ASSERT((ds->ds_owner == tag && ds->ds_dbuf) ||
731 	    (DSL_DATASET_IS_DESTROYED(ds) && ds->ds_dbuf == NULL));
732 
733 	mutex_enter(&ds->ds_lock);
734 	ds->ds_owner = NULL;
735 	if (RW_WRITE_HELD(&ds->ds_rwlock)) {
736 		rw_exit(&ds->ds_rwlock);
737 		cv_broadcast(&ds->ds_exclusive_cv);
738 	}
739 	mutex_exit(&ds->ds_lock);
740 	if (ds->ds_dbuf)
741 		dsl_dataset_drop_ref(ds, tag);
742 	else
743 		dsl_dataset_evict(NULL, ds);
744 }
745 
746 boolean_t
747 dsl_dataset_tryown(dsl_dataset_t *ds, boolean_t inconsistentok, void *tag)
748 {
749 	boolean_t gotit = FALSE;
750 
751 	mutex_enter(&ds->ds_lock);
752 	if (ds->ds_owner == NULL &&
753 	    (!DS_IS_INCONSISTENT(ds) || inconsistentok)) {
754 		ds->ds_owner = tag;
755 		if (!dsl_pool_sync_context(ds->ds_dir->dd_pool))
756 			rw_exit(&ds->ds_rwlock);
757 		gotit = TRUE;
758 	}
759 	mutex_exit(&ds->ds_lock);
760 	return (gotit);
761 }
762 
763 void
764 dsl_dataset_make_exclusive(dsl_dataset_t *ds, void *owner)
765 {
766 	ASSERT3P(owner, ==, ds->ds_owner);
767 	if (!RW_WRITE_HELD(&ds->ds_rwlock))
768 		rw_enter(&ds->ds_rwlock, RW_WRITER);
769 }
770 
771 uint64_t
772 dsl_dataset_create_sync_dd(dsl_dir_t *dd, dsl_dataset_t *origin,
773     uint64_t flags, dmu_tx_t *tx)
774 {
775 	dsl_pool_t *dp = dd->dd_pool;
776 	dmu_buf_t *dbuf;
777 	dsl_dataset_phys_t *dsphys;
778 	uint64_t dsobj;
779 	objset_t *mos = dp->dp_meta_objset;
780 
781 	if (origin == NULL)
782 		origin = dp->dp_origin_snap;
783 
784 	ASSERT(origin == NULL || origin->ds_dir->dd_pool == dp);
785 	ASSERT(origin == NULL || origin->ds_phys->ds_num_children > 0);
786 	ASSERT(dmu_tx_is_syncing(tx));
787 	ASSERT(dd->dd_phys->dd_head_dataset_obj == 0);
788 
789 	dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
790 	    DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
791 	VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
792 	dmu_buf_will_dirty(dbuf, tx);
793 	dsphys = dbuf->db_data;
794 	bzero(dsphys, sizeof (dsl_dataset_phys_t));
795 	dsphys->ds_dir_obj = dd->dd_object;
796 	dsphys->ds_flags = flags;
797 	dsphys->ds_fsid_guid = unique_create();
798 	(void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
799 	    sizeof (dsphys->ds_guid));
800 	dsphys->ds_snapnames_zapobj =
801 	    zap_create_norm(mos, U8_TEXTPREP_TOUPPER, DMU_OT_DSL_DS_SNAP_MAP,
802 	    DMU_OT_NONE, 0, tx);
803 	dsphys->ds_creation_time = gethrestime_sec();
804 	dsphys->ds_creation_txg = tx->tx_txg == TXG_INITIAL ? 1 : tx->tx_txg;
805 
806 	if (origin == NULL) {
807 		dsphys->ds_deadlist_obj = dsl_deadlist_alloc(mos, tx);
808 	} else {
809 		dsl_dataset_t *ohds;
810 
811 		dsphys->ds_prev_snap_obj = origin->ds_object;
812 		dsphys->ds_prev_snap_txg =
813 		    origin->ds_phys->ds_creation_txg;
814 		dsphys->ds_used_bytes =
815 		    origin->ds_phys->ds_used_bytes;
816 		dsphys->ds_compressed_bytes =
817 		    origin->ds_phys->ds_compressed_bytes;
818 		dsphys->ds_uncompressed_bytes =
819 		    origin->ds_phys->ds_uncompressed_bytes;
820 		dsphys->ds_bp = origin->ds_phys->ds_bp;
821 		dsphys->ds_flags |= origin->ds_phys->ds_flags;
822 
823 		dmu_buf_will_dirty(origin->ds_dbuf, tx);
824 		origin->ds_phys->ds_num_children++;
825 
826 		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
827 		    origin->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ohds));
828 		dsphys->ds_deadlist_obj = dsl_deadlist_clone(&ohds->ds_deadlist,
829 		    dsphys->ds_prev_snap_txg, dsphys->ds_prev_snap_obj, tx);
830 		dsl_dataset_rele(ohds, FTAG);
831 
832 		if (spa_version(dp->dp_spa) >= SPA_VERSION_NEXT_CLONES) {
833 			if (origin->ds_phys->ds_next_clones_obj == 0) {
834 				origin->ds_phys->ds_next_clones_obj =
835 				    zap_create(mos,
836 				    DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
837 			}
838 			VERIFY(0 == zap_add_int(mos,
839 			    origin->ds_phys->ds_next_clones_obj,
840 			    dsobj, tx));
841 		}
842 
843 		dmu_buf_will_dirty(dd->dd_dbuf, tx);
844 		dd->dd_phys->dd_origin_obj = origin->ds_object;
845 		if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
846 			if (origin->ds_dir->dd_phys->dd_clones == 0) {
847 				dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
848 				origin->ds_dir->dd_phys->dd_clones =
849 				    zap_create(mos,
850 				    DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
851 			}
852 			VERIFY3U(0, ==, zap_add_int(mos,
853 			    origin->ds_dir->dd_phys->dd_clones, dsobj, tx));
854 		}
855 	}
856 
857 	if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
858 		dsphys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
859 
860 	dmu_buf_rele(dbuf, FTAG);
861 
862 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
863 	dd->dd_phys->dd_head_dataset_obj = dsobj;
864 
865 	return (dsobj);
866 }
867 
868 uint64_t
869 dsl_dataset_create_sync(dsl_dir_t *pdd, const char *lastname,
870     dsl_dataset_t *origin, uint64_t flags, cred_t *cr, dmu_tx_t *tx)
871 {
872 	dsl_pool_t *dp = pdd->dd_pool;
873 	uint64_t dsobj, ddobj;
874 	dsl_dir_t *dd;
875 
876 	ASSERT(lastname[0] != '@');
877 
878 	ddobj = dsl_dir_create_sync(dp, pdd, lastname, tx);
879 	VERIFY(0 == dsl_dir_open_obj(dp, ddobj, lastname, FTAG, &dd));
880 
881 	dsobj = dsl_dataset_create_sync_dd(dd, origin, flags, tx);
882 
883 	dsl_deleg_set_create_perms(dd, tx, cr);
884 
885 	dsl_dir_close(dd, FTAG);
886 
887 	/*
888 	 * If we are creating a clone, make sure we zero out any stale
889 	 * data from the origin snapshots zil header.
890 	 */
891 	if (origin != NULL) {
892 		dsl_dataset_t *ds;
893 		objset_t *os;
894 
895 		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
896 		VERIFY3U(0, ==, dmu_objset_from_ds(ds, &os));
897 		bzero(&os->os_zil_header, sizeof (os->os_zil_header));
898 		dsl_dataset_dirty(ds, tx);
899 		dsl_dataset_rele(ds, FTAG);
900 	}
901 
902 	return (dsobj);
903 }
904 
905 /*
906  * The snapshots must all be in the same pool.
907  */
908 int
909 dmu_snapshots_destroy_nvl(nvlist_t *snaps, boolean_t defer, char *failed)
910 {
911 	int err;
912 	dsl_sync_task_t *dst;
913 	spa_t *spa;
914 	nvpair_t *pair;
915 	dsl_sync_task_group_t *dstg;
916 
917 	pair = nvlist_next_nvpair(snaps, NULL);
918 	if (pair == NULL)
919 		return (0);
920 
921 	err = spa_open(nvpair_name(pair), &spa, FTAG);
922 	if (err)
923 		return (err);
924 	dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
925 
926 	for (pair = nvlist_next_nvpair(snaps, NULL); pair != NULL;
927 	    pair = nvlist_next_nvpair(snaps, pair)) {
928 		dsl_dataset_t *ds;
929 		int err;
930 
931 		err = dsl_dataset_own(nvpair_name(pair), B_TRUE, dstg, &ds);
932 		if (err == 0) {
933 			struct dsl_ds_destroyarg *dsda;
934 
935 			dsl_dataset_make_exclusive(ds, dstg);
936 			dsda = kmem_zalloc(sizeof (struct dsl_ds_destroyarg),
937 			    KM_SLEEP);
938 			dsda->ds = ds;
939 			dsda->defer = defer;
940 			dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
941 			    dsl_dataset_destroy_sync, dsda, dstg, 0);
942 		} else if (err == ENOENT) {
943 			err = 0;
944 		} else {
945 			(void) strcpy(failed, nvpair_name(pair));
946 			break;
947 		}
948 	}
949 
950 	if (err == 0)
951 		err = dsl_sync_task_group_wait(dstg);
952 
953 	for (dst = list_head(&dstg->dstg_tasks); dst;
954 	    dst = list_next(&dstg->dstg_tasks, dst)) {
955 		struct dsl_ds_destroyarg *dsda = dst->dst_arg1;
956 		dsl_dataset_t *ds = dsda->ds;
957 
958 		/*
959 		 * Return the file system name that triggered the error
960 		 */
961 		if (dst->dst_err) {
962 			dsl_dataset_name(ds, failed);
963 		}
964 		ASSERT3P(dsda->rm_origin, ==, NULL);
965 		dsl_dataset_disown(ds, dstg);
966 		kmem_free(dsda, sizeof (struct dsl_ds_destroyarg));
967 	}
968 
969 	dsl_sync_task_group_destroy(dstg);
970 	spa_close(spa, FTAG);
971 	return (err);
972 
973 }
974 
975 static boolean_t
976 dsl_dataset_might_destroy_origin(dsl_dataset_t *ds)
977 {
978 	boolean_t might_destroy = B_FALSE;
979 
980 	mutex_enter(&ds->ds_lock);
981 	if (ds->ds_phys->ds_num_children == 2 && ds->ds_userrefs == 0 &&
982 	    DS_IS_DEFER_DESTROY(ds))
983 		might_destroy = B_TRUE;
984 	mutex_exit(&ds->ds_lock);
985 
986 	return (might_destroy);
987 }
988 
989 /*
990  * If we're removing a clone, and these three conditions are true:
991  *	1) the clone's origin has no other children
992  *	2) the clone's origin has no user references
993  *	3) the clone's origin has been marked for deferred destruction
994  * Then, prepare to remove the origin as part of this sync task group.
995  */
996 static int
997 dsl_dataset_origin_rm_prep(struct dsl_ds_destroyarg *dsda, void *tag)
998 {
999 	dsl_dataset_t *ds = dsda->ds;
1000 	dsl_dataset_t *origin = ds->ds_prev;
1001 
1002 	if (dsl_dataset_might_destroy_origin(origin)) {
1003 		char *name;
1004 		int namelen;
1005 		int error;
1006 
1007 		namelen = dsl_dataset_namelen(origin) + 1;
1008 		name = kmem_alloc(namelen, KM_SLEEP);
1009 		dsl_dataset_name(origin, name);
1010 #ifdef _KERNEL
1011 		error = zfs_unmount_snap(name, NULL);
1012 		if (error) {
1013 			kmem_free(name, namelen);
1014 			return (error);
1015 		}
1016 #endif
1017 		error = dsl_dataset_own(name, B_TRUE, tag, &origin);
1018 		kmem_free(name, namelen);
1019 		if (error)
1020 			return (error);
1021 		dsda->rm_origin = origin;
1022 		dsl_dataset_make_exclusive(origin, tag);
1023 	}
1024 
1025 	return (0);
1026 }
1027 
1028 /*
1029  * ds must be opened as OWNER.  On return (whether successful or not),
1030  * ds will be closed and caller can no longer dereference it.
1031  */
1032 int
1033 dsl_dataset_destroy(dsl_dataset_t *ds, void *tag, boolean_t defer)
1034 {
1035 	int err;
1036 	dsl_sync_task_group_t *dstg;
1037 	objset_t *os;
1038 	dsl_dir_t *dd;
1039 	uint64_t obj;
1040 	struct dsl_ds_destroyarg dsda = { 0 };
1041 	dsl_dataset_t dummy_ds = { 0 };
1042 
1043 	dsda.ds = ds;
1044 
1045 	if (dsl_dataset_is_snapshot(ds)) {
1046 		/* Destroying a snapshot is simpler */
1047 		dsl_dataset_make_exclusive(ds, tag);
1048 
1049 		dsda.defer = defer;
1050 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1051 		    dsl_dataset_destroy_check, dsl_dataset_destroy_sync,
1052 		    &dsda, tag, 0);
1053 		ASSERT3P(dsda.rm_origin, ==, NULL);
1054 		goto out;
1055 	} else if (defer) {
1056 		err = EINVAL;
1057 		goto out;
1058 	}
1059 
1060 	dd = ds->ds_dir;
1061 	dummy_ds.ds_dir = dd;
1062 	dummy_ds.ds_object = ds->ds_object;
1063 
1064 	/*
1065 	 * Check for errors and mark this ds as inconsistent, in
1066 	 * case we crash while freeing the objects.
1067 	 */
1068 	err = dsl_sync_task_do(dd->dd_pool, dsl_dataset_destroy_begin_check,
1069 	    dsl_dataset_destroy_begin_sync, ds, NULL, 0);
1070 	if (err)
1071 		goto out;
1072 
1073 	err = dmu_objset_from_ds(ds, &os);
1074 	if (err)
1075 		goto out;
1076 
1077 	/*
1078 	 * remove the objects in open context, so that we won't
1079 	 * have too much to do in syncing context.
1080 	 */
1081 	for (obj = 0; err == 0; err = dmu_object_next(os, &obj, FALSE,
1082 	    ds->ds_phys->ds_prev_snap_txg)) {
1083 		/*
1084 		 * Ignore errors, if there is not enough disk space
1085 		 * we will deal with it in dsl_dataset_destroy_sync().
1086 		 */
1087 		(void) dmu_free_object(os, obj);
1088 	}
1089 	if (err != ESRCH)
1090 		goto out;
1091 
1092 	/*
1093 	 * Only the ZIL knows how to free log blocks.
1094 	 */
1095 	zil_destroy(dmu_objset_zil(os), B_FALSE);
1096 
1097 	/*
1098 	 * Sync out all in-flight IO.
1099 	 */
1100 	txg_wait_synced(dd->dd_pool, 0);
1101 
1102 	/*
1103 	 * If we managed to free all the objects in open
1104 	 * context, the user space accounting should be zero.
1105 	 */
1106 	if (ds->ds_phys->ds_bp.blk_fill == 0 &&
1107 	    dmu_objset_userused_enabled(os)) {
1108 		uint64_t count;
1109 
1110 		ASSERT(zap_count(os, DMU_USERUSED_OBJECT, &count) != 0 ||
1111 		    count == 0);
1112 		ASSERT(zap_count(os, DMU_GROUPUSED_OBJECT, &count) != 0 ||
1113 		    count == 0);
1114 	}
1115 
1116 	rw_enter(&dd->dd_pool->dp_config_rwlock, RW_READER);
1117 	err = dsl_dir_open_obj(dd->dd_pool, dd->dd_object, NULL, FTAG, &dd);
1118 	rw_exit(&dd->dd_pool->dp_config_rwlock);
1119 
1120 	if (err)
1121 		goto out;
1122 
1123 	/*
1124 	 * Blow away the dsl_dir + head dataset.
1125 	 */
1126 	dsl_dataset_make_exclusive(ds, tag);
1127 	/*
1128 	 * If we're removing a clone, we might also need to remove its
1129 	 * origin.
1130 	 */
1131 	do {
1132 		dsda.need_prep = B_FALSE;
1133 		if (dsl_dir_is_clone(dd)) {
1134 			err = dsl_dataset_origin_rm_prep(&dsda, tag);
1135 			if (err) {
1136 				dsl_dir_close(dd, FTAG);
1137 				goto out;
1138 			}
1139 		}
1140 
1141 		dstg = dsl_sync_task_group_create(ds->ds_dir->dd_pool);
1142 		dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
1143 		    dsl_dataset_destroy_sync, &dsda, tag, 0);
1144 		dsl_sync_task_create(dstg, dsl_dir_destroy_check,
1145 		    dsl_dir_destroy_sync, &dummy_ds, FTAG, 0);
1146 		err = dsl_sync_task_group_wait(dstg);
1147 		dsl_sync_task_group_destroy(dstg);
1148 
1149 		/*
1150 		 * We could be racing against 'zfs release' or 'zfs destroy -d'
1151 		 * on the origin snap, in which case we can get EBUSY if we
1152 		 * needed to destroy the origin snap but were not ready to
1153 		 * do so.
1154 		 */
1155 		if (dsda.need_prep) {
1156 			ASSERT(err == EBUSY);
1157 			ASSERT(dsl_dir_is_clone(dd));
1158 			ASSERT(dsda.rm_origin == NULL);
1159 		}
1160 	} while (dsda.need_prep);
1161 
1162 	if (dsda.rm_origin != NULL)
1163 		dsl_dataset_disown(dsda.rm_origin, tag);
1164 
1165 	/* if it is successful, dsl_dir_destroy_sync will close the dd */
1166 	if (err)
1167 		dsl_dir_close(dd, FTAG);
1168 out:
1169 	dsl_dataset_disown(ds, tag);
1170 	return (err);
1171 }
1172 
1173 blkptr_t *
1174 dsl_dataset_get_blkptr(dsl_dataset_t *ds)
1175 {
1176 	return (&ds->ds_phys->ds_bp);
1177 }
1178 
1179 void
1180 dsl_dataset_set_blkptr(dsl_dataset_t *ds, blkptr_t *bp, dmu_tx_t *tx)
1181 {
1182 	ASSERT(dmu_tx_is_syncing(tx));
1183 	/* If it's the meta-objset, set dp_meta_rootbp */
1184 	if (ds == NULL) {
1185 		tx->tx_pool->dp_meta_rootbp = *bp;
1186 	} else {
1187 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
1188 		ds->ds_phys->ds_bp = *bp;
1189 	}
1190 }
1191 
1192 spa_t *
1193 dsl_dataset_get_spa(dsl_dataset_t *ds)
1194 {
1195 	return (ds->ds_dir->dd_pool->dp_spa);
1196 }
1197 
1198 void
1199 dsl_dataset_dirty(dsl_dataset_t *ds, dmu_tx_t *tx)
1200 {
1201 	dsl_pool_t *dp;
1202 
1203 	if (ds == NULL) /* this is the meta-objset */
1204 		return;
1205 
1206 	ASSERT(ds->ds_objset != NULL);
1207 
1208 	if (ds->ds_phys->ds_next_snap_obj != 0)
1209 		panic("dirtying snapshot!");
1210 
1211 	dp = ds->ds_dir->dd_pool;
1212 
1213 	if (txg_list_add(&dp->dp_dirty_datasets, ds, tx->tx_txg) == 0) {
1214 		/* up the hold count until we can be written out */
1215 		dmu_buf_add_ref(ds->ds_dbuf, ds);
1216 	}
1217 }
1218 
1219 /*
1220  * The unique space in the head dataset can be calculated by subtracting
1221  * the space used in the most recent snapshot, that is still being used
1222  * in this file system, from the space currently in use.  To figure out
1223  * the space in the most recent snapshot still in use, we need to take
1224  * the total space used in the snapshot and subtract out the space that
1225  * has been freed up since the snapshot was taken.
1226  */
1227 static void
1228 dsl_dataset_recalc_head_uniq(dsl_dataset_t *ds)
1229 {
1230 	uint64_t mrs_used;
1231 	uint64_t dlused, dlcomp, dluncomp;
1232 
1233 	ASSERT(!dsl_dataset_is_snapshot(ds));
1234 
1235 	if (ds->ds_phys->ds_prev_snap_obj != 0)
1236 		mrs_used = ds->ds_prev->ds_phys->ds_used_bytes;
1237 	else
1238 		mrs_used = 0;
1239 
1240 	dsl_deadlist_space(&ds->ds_deadlist, &dlused, &dlcomp, &dluncomp);
1241 
1242 	ASSERT3U(dlused, <=, mrs_used);
1243 	ds->ds_phys->ds_unique_bytes =
1244 	    ds->ds_phys->ds_used_bytes - (mrs_used - dlused);
1245 
1246 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) >=
1247 	    SPA_VERSION_UNIQUE_ACCURATE)
1248 		ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
1249 }
1250 
1251 struct killarg {
1252 	dsl_dataset_t *ds;
1253 	dmu_tx_t *tx;
1254 };
1255 
1256 /* ARGSUSED */
1257 static int
1258 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, arc_buf_t *pbuf,
1259     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
1260 {
1261 	struct killarg *ka = arg;
1262 	dmu_tx_t *tx = ka->tx;
1263 
1264 	if (bp == NULL)
1265 		return (0);
1266 
1267 	if (zb->zb_level == ZB_ZIL_LEVEL) {
1268 		ASSERT(zilog != NULL);
1269 		/*
1270 		 * It's a block in the intent log.  It has no
1271 		 * accounting, so just free it.
1272 		 */
1273 		dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
1274 	} else {
1275 		ASSERT(zilog == NULL);
1276 		ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
1277 		(void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
1278 	}
1279 
1280 	return (0);
1281 }
1282 
1283 /* ARGSUSED */
1284 static int
1285 dsl_dataset_destroy_begin_check(void *arg1, void *arg2, dmu_tx_t *tx)
1286 {
1287 	dsl_dataset_t *ds = arg1;
1288 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1289 	uint64_t count;
1290 	int err;
1291 
1292 	/*
1293 	 * Can't delete a head dataset if there are snapshots of it.
1294 	 * (Except if the only snapshots are from the branch we cloned
1295 	 * from.)
1296 	 */
1297 	if (ds->ds_prev != NULL &&
1298 	    ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1299 		return (EBUSY);
1300 
1301 	/*
1302 	 * This is really a dsl_dir thing, but check it here so that
1303 	 * we'll be less likely to leave this dataset inconsistent &
1304 	 * nearly destroyed.
1305 	 */
1306 	err = zap_count(mos, ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
1307 	if (err)
1308 		return (err);
1309 	if (count != 0)
1310 		return (EEXIST);
1311 
1312 	return (0);
1313 }
1314 
1315 /* ARGSUSED */
1316 static void
1317 dsl_dataset_destroy_begin_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1318 {
1319 	dsl_dataset_t *ds = arg1;
1320 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1321 
1322 	/* Mark it as inconsistent on-disk, in case we crash */
1323 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1324 	ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
1325 
1326 	spa_history_log_internal(LOG_DS_DESTROY_BEGIN, dp->dp_spa, tx,
1327 	    "dataset = %llu", ds->ds_object);
1328 }
1329 
1330 static int
1331 dsl_dataset_origin_check(struct dsl_ds_destroyarg *dsda, void *tag,
1332     dmu_tx_t *tx)
1333 {
1334 	dsl_dataset_t *ds = dsda->ds;
1335 	dsl_dataset_t *ds_prev = ds->ds_prev;
1336 
1337 	if (dsl_dataset_might_destroy_origin(ds_prev)) {
1338 		struct dsl_ds_destroyarg ndsda = {0};
1339 
1340 		/*
1341 		 * If we're not prepared to remove the origin, don't remove
1342 		 * the clone either.
1343 		 */
1344 		if (dsda->rm_origin == NULL) {
1345 			dsda->need_prep = B_TRUE;
1346 			return (EBUSY);
1347 		}
1348 
1349 		ndsda.ds = ds_prev;
1350 		ndsda.is_origin_rm = B_TRUE;
1351 		return (dsl_dataset_destroy_check(&ndsda, tag, tx));
1352 	}
1353 
1354 	/*
1355 	 * If we're not going to remove the origin after all,
1356 	 * undo the open context setup.
1357 	 */
1358 	if (dsda->rm_origin != NULL) {
1359 		dsl_dataset_disown(dsda->rm_origin, tag);
1360 		dsda->rm_origin = NULL;
1361 	}
1362 
1363 	return (0);
1364 }
1365 
1366 /*
1367  * If you add new checks here, you may need to add
1368  * additional checks to the "temporary" case in
1369  * snapshot_check() in dmu_objset.c.
1370  */
1371 /* ARGSUSED */
1372 int
1373 dsl_dataset_destroy_check(void *arg1, void *arg2, dmu_tx_t *tx)
1374 {
1375 	struct dsl_ds_destroyarg *dsda = arg1;
1376 	dsl_dataset_t *ds = dsda->ds;
1377 
1378 	/* we have an owner hold, so noone else can destroy us */
1379 	ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
1380 
1381 	/*
1382 	 * Only allow deferred destroy on pools that support it.
1383 	 * NOTE: deferred destroy is only supported on snapshots.
1384 	 */
1385 	if (dsda->defer) {
1386 		if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
1387 		    SPA_VERSION_USERREFS)
1388 			return (ENOTSUP);
1389 		ASSERT(dsl_dataset_is_snapshot(ds));
1390 		return (0);
1391 	}
1392 
1393 	/*
1394 	 * Can't delete a head dataset if there are snapshots of it.
1395 	 * (Except if the only snapshots are from the branch we cloned
1396 	 * from.)
1397 	 */
1398 	if (ds->ds_prev != NULL &&
1399 	    ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1400 		return (EBUSY);
1401 
1402 	/*
1403 	 * If we made changes this txg, traverse_dsl_dataset won't find
1404 	 * them.  Try again.
1405 	 */
1406 	if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
1407 		return (EAGAIN);
1408 
1409 	if (dsl_dataset_is_snapshot(ds)) {
1410 		/*
1411 		 * If this snapshot has an elevated user reference count,
1412 		 * we can't destroy it yet.
1413 		 */
1414 		if (ds->ds_userrefs > 0 && !dsda->releasing)
1415 			return (EBUSY);
1416 
1417 		mutex_enter(&ds->ds_lock);
1418 		/*
1419 		 * Can't delete a branch point. However, if we're destroying
1420 		 * a clone and removing its origin due to it having a user
1421 		 * hold count of 0 and having been marked for deferred destroy,
1422 		 * it's OK for the origin to have a single clone.
1423 		 */
1424 		if (ds->ds_phys->ds_num_children >
1425 		    (dsda->is_origin_rm ? 2 : 1)) {
1426 			mutex_exit(&ds->ds_lock);
1427 			return (EEXIST);
1428 		}
1429 		mutex_exit(&ds->ds_lock);
1430 	} else if (dsl_dir_is_clone(ds->ds_dir)) {
1431 		return (dsl_dataset_origin_check(dsda, arg2, tx));
1432 	}
1433 
1434 	/* XXX we should do some i/o error checking... */
1435 	return (0);
1436 }
1437 
1438 struct refsarg {
1439 	kmutex_t lock;
1440 	boolean_t gone;
1441 	kcondvar_t cv;
1442 };
1443 
1444 /* ARGSUSED */
1445 static void
1446 dsl_dataset_refs_gone(dmu_buf_t *db, void *argv)
1447 {
1448 	struct refsarg *arg = argv;
1449 
1450 	mutex_enter(&arg->lock);
1451 	arg->gone = TRUE;
1452 	cv_signal(&arg->cv);
1453 	mutex_exit(&arg->lock);
1454 }
1455 
1456 static void
1457 dsl_dataset_drain_refs(dsl_dataset_t *ds, void *tag)
1458 {
1459 	struct refsarg arg;
1460 
1461 	mutex_init(&arg.lock, NULL, MUTEX_DEFAULT, NULL);
1462 	cv_init(&arg.cv, NULL, CV_DEFAULT, NULL);
1463 	arg.gone = FALSE;
1464 	(void) dmu_buf_update_user(ds->ds_dbuf, ds, &arg, &ds->ds_phys,
1465 	    dsl_dataset_refs_gone);
1466 	dmu_buf_rele(ds->ds_dbuf, tag);
1467 	mutex_enter(&arg.lock);
1468 	while (!arg.gone)
1469 		cv_wait(&arg.cv, &arg.lock);
1470 	ASSERT(arg.gone);
1471 	mutex_exit(&arg.lock);
1472 	ds->ds_dbuf = NULL;
1473 	ds->ds_phys = NULL;
1474 	mutex_destroy(&arg.lock);
1475 	cv_destroy(&arg.cv);
1476 }
1477 
1478 static void
1479 remove_from_next_clones(dsl_dataset_t *ds, uint64_t obj, dmu_tx_t *tx)
1480 {
1481 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1482 	uint64_t count;
1483 	int err;
1484 
1485 	ASSERT(ds->ds_phys->ds_num_children >= 2);
1486 	err = zap_remove_int(mos, ds->ds_phys->ds_next_clones_obj, obj, tx);
1487 	/*
1488 	 * The err should not be ENOENT, but a bug in a previous version
1489 	 * of the code could cause upgrade_clones_cb() to not set
1490 	 * ds_next_snap_obj when it should, leading to a missing entry.
1491 	 * If we knew that the pool was created after
1492 	 * SPA_VERSION_NEXT_CLONES, we could assert that it isn't
1493 	 * ENOENT.  However, at least we can check that we don't have
1494 	 * too many entries in the next_clones_obj even after failing to
1495 	 * remove this one.
1496 	 */
1497 	if (err != ENOENT) {
1498 		VERIFY3U(err, ==, 0);
1499 	}
1500 	ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
1501 	    &count));
1502 	ASSERT3U(count, <=, ds->ds_phys->ds_num_children - 2);
1503 }
1504 
1505 static void
1506 dsl_dataset_remove_clones_key(dsl_dataset_t *ds, uint64_t mintxg, dmu_tx_t *tx)
1507 {
1508 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1509 	zap_cursor_t zc;
1510 	zap_attribute_t za;
1511 
1512 	/*
1513 	 * If it is the old version, dd_clones doesn't exist so we can't
1514 	 * find the clones, but deadlist_remove_key() is a no-op so it
1515 	 * doesn't matter.
1516 	 */
1517 	if (ds->ds_dir->dd_phys->dd_clones == 0)
1518 		return;
1519 
1520 	for (zap_cursor_init(&zc, mos, ds->ds_dir->dd_phys->dd_clones);
1521 	    zap_cursor_retrieve(&zc, &za) == 0;
1522 	    zap_cursor_advance(&zc)) {
1523 		dsl_dataset_t *clone;
1524 
1525 		VERIFY3U(0, ==, dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
1526 		    za.za_first_integer, FTAG, &clone));
1527 		if (clone->ds_dir->dd_origin_txg > mintxg) {
1528 			dsl_deadlist_remove_key(&clone->ds_deadlist,
1529 			    mintxg, tx);
1530 			dsl_dataset_remove_clones_key(clone, mintxg, tx);
1531 		}
1532 		dsl_dataset_rele(clone, FTAG);
1533 	}
1534 	zap_cursor_fini(&zc);
1535 }
1536 
1537 struct process_old_arg {
1538 	dsl_dataset_t *ds;
1539 	dsl_dataset_t *ds_prev;
1540 	boolean_t after_branch_point;
1541 	zio_t *pio;
1542 	uint64_t used, comp, uncomp;
1543 };
1544 
1545 static int
1546 process_old_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1547 {
1548 	struct process_old_arg *poa = arg;
1549 	dsl_pool_t *dp = poa->ds->ds_dir->dd_pool;
1550 
1551 	if (bp->blk_birth <= poa->ds->ds_phys->ds_prev_snap_txg) {
1552 		dsl_deadlist_insert(&poa->ds->ds_deadlist, bp, tx);
1553 		if (poa->ds_prev && !poa->after_branch_point &&
1554 		    bp->blk_birth >
1555 		    poa->ds_prev->ds_phys->ds_prev_snap_txg) {
1556 			poa->ds_prev->ds_phys->ds_unique_bytes +=
1557 			    bp_get_dsize_sync(dp->dp_spa, bp);
1558 		}
1559 	} else {
1560 		poa->used += bp_get_dsize_sync(dp->dp_spa, bp);
1561 		poa->comp += BP_GET_PSIZE(bp);
1562 		poa->uncomp += BP_GET_UCSIZE(bp);
1563 		dsl_free_sync(poa->pio, dp, tx->tx_txg, bp);
1564 	}
1565 	return (0);
1566 }
1567 
1568 static void
1569 process_old_deadlist(dsl_dataset_t *ds, dsl_dataset_t *ds_prev,
1570     dsl_dataset_t *ds_next, boolean_t after_branch_point, dmu_tx_t *tx)
1571 {
1572 	struct process_old_arg poa = { 0 };
1573 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1574 	objset_t *mos = dp->dp_meta_objset;
1575 
1576 	ASSERT(ds->ds_deadlist.dl_oldfmt);
1577 	ASSERT(ds_next->ds_deadlist.dl_oldfmt);
1578 
1579 	poa.ds = ds;
1580 	poa.ds_prev = ds_prev;
1581 	poa.after_branch_point = after_branch_point;
1582 	poa.pio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
1583 	VERIFY3U(0, ==, bpobj_iterate(&ds_next->ds_deadlist.dl_bpobj,
1584 	    process_old_cb, &poa, tx));
1585 	VERIFY3U(zio_wait(poa.pio), ==, 0);
1586 	ASSERT3U(poa.used, ==, ds->ds_phys->ds_unique_bytes);
1587 
1588 	/* change snapused */
1589 	dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1590 	    -poa.used, -poa.comp, -poa.uncomp, tx);
1591 
1592 	/* swap next's deadlist to our deadlist */
1593 	dsl_deadlist_close(&ds->ds_deadlist);
1594 	dsl_deadlist_close(&ds_next->ds_deadlist);
1595 	SWITCH64(ds_next->ds_phys->ds_deadlist_obj,
1596 	    ds->ds_phys->ds_deadlist_obj);
1597 	dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
1598 	dsl_deadlist_open(&ds_next->ds_deadlist, mos,
1599 	    ds_next->ds_phys->ds_deadlist_obj);
1600 }
1601 
1602 void
1603 dsl_dataset_destroy_sync(void *arg1, void *tag, dmu_tx_t *tx)
1604 {
1605 	struct dsl_ds_destroyarg *dsda = arg1;
1606 	dsl_dataset_t *ds = dsda->ds;
1607 	int err;
1608 	int after_branch_point = FALSE;
1609 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1610 	objset_t *mos = dp->dp_meta_objset;
1611 	dsl_dataset_t *ds_prev = NULL;
1612 	boolean_t wont_destroy;
1613 	uint64_t obj;
1614 
1615 	wont_destroy = (dsda->defer &&
1616 	    (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1));
1617 
1618 	ASSERT(ds->ds_owner || wont_destroy);
1619 	ASSERT(dsda->defer || ds->ds_phys->ds_num_children <= 1);
1620 	ASSERT(ds->ds_prev == NULL ||
1621 	    ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
1622 	ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
1623 
1624 	if (wont_destroy) {
1625 		ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1626 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
1627 		ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
1628 		return;
1629 	}
1630 
1631 	/* signal any waiters that this dataset is going away */
1632 	mutex_enter(&ds->ds_lock);
1633 	ds->ds_owner = dsl_reaper;
1634 	cv_broadcast(&ds->ds_exclusive_cv);
1635 	mutex_exit(&ds->ds_lock);
1636 
1637 	/* Remove our reservation */
1638 	if (ds->ds_reserved != 0) {
1639 		dsl_prop_setarg_t psa;
1640 		uint64_t value = 0;
1641 
1642 		dsl_prop_setarg_init_uint64(&psa, "refreservation",
1643 		    (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
1644 		    &value);
1645 		psa.psa_effective_value = 0;	/* predict default value */
1646 
1647 		dsl_dataset_set_reservation_sync(ds, &psa, tx);
1648 		ASSERT3U(ds->ds_reserved, ==, 0);
1649 	}
1650 
1651 	ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1652 
1653 	dsl_scan_ds_destroyed(ds, tx);
1654 
1655 	obj = ds->ds_object;
1656 
1657 	if (ds->ds_phys->ds_prev_snap_obj != 0) {
1658 		if (ds->ds_prev) {
1659 			ds_prev = ds->ds_prev;
1660 		} else {
1661 			VERIFY(0 == dsl_dataset_hold_obj(dp,
1662 			    ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
1663 		}
1664 		after_branch_point =
1665 		    (ds_prev->ds_phys->ds_next_snap_obj != obj);
1666 
1667 		dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
1668 		if (after_branch_point &&
1669 		    ds_prev->ds_phys->ds_next_clones_obj != 0) {
1670 			remove_from_next_clones(ds_prev, obj, tx);
1671 			if (ds->ds_phys->ds_next_snap_obj != 0) {
1672 				VERIFY(0 == zap_add_int(mos,
1673 				    ds_prev->ds_phys->ds_next_clones_obj,
1674 				    ds->ds_phys->ds_next_snap_obj, tx));
1675 			}
1676 		}
1677 		if (after_branch_point &&
1678 		    ds->ds_phys->ds_next_snap_obj == 0) {
1679 			/* This clone is toast. */
1680 			ASSERT(ds_prev->ds_phys->ds_num_children > 1);
1681 			ds_prev->ds_phys->ds_num_children--;
1682 
1683 			/*
1684 			 * If the clone's origin has no other clones, no
1685 			 * user holds, and has been marked for deferred
1686 			 * deletion, then we should have done the necessary
1687 			 * destroy setup for it.
1688 			 */
1689 			if (ds_prev->ds_phys->ds_num_children == 1 &&
1690 			    ds_prev->ds_userrefs == 0 &&
1691 			    DS_IS_DEFER_DESTROY(ds_prev)) {
1692 				ASSERT3P(dsda->rm_origin, !=, NULL);
1693 			} else {
1694 				ASSERT3P(dsda->rm_origin, ==, NULL);
1695 			}
1696 		} else if (!after_branch_point) {
1697 			ds_prev->ds_phys->ds_next_snap_obj =
1698 			    ds->ds_phys->ds_next_snap_obj;
1699 		}
1700 	}
1701 
1702 	if (dsl_dataset_is_snapshot(ds)) {
1703 		dsl_dataset_t *ds_next;
1704 		uint64_t old_unique;
1705 		uint64_t used = 0, comp = 0, uncomp = 0;
1706 
1707 		VERIFY(0 == dsl_dataset_hold_obj(dp,
1708 		    ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
1709 		ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
1710 
1711 		old_unique = ds_next->ds_phys->ds_unique_bytes;
1712 
1713 		dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
1714 		ds_next->ds_phys->ds_prev_snap_obj =
1715 		    ds->ds_phys->ds_prev_snap_obj;
1716 		ds_next->ds_phys->ds_prev_snap_txg =
1717 		    ds->ds_phys->ds_prev_snap_txg;
1718 		ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1719 		    ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
1720 
1721 
1722 		if (ds_next->ds_deadlist.dl_oldfmt) {
1723 			process_old_deadlist(ds, ds_prev, ds_next,
1724 			    after_branch_point, tx);
1725 		} else {
1726 			/* Adjust prev's unique space. */
1727 			if (ds_prev && !after_branch_point) {
1728 				dsl_deadlist_space_range(&ds_next->ds_deadlist,
1729 				    ds_prev->ds_phys->ds_prev_snap_txg,
1730 				    ds->ds_phys->ds_prev_snap_txg,
1731 				    &used, &comp, &uncomp);
1732 				ds_prev->ds_phys->ds_unique_bytes += used;
1733 			}
1734 
1735 			/* Adjust snapused. */
1736 			dsl_deadlist_space_range(&ds_next->ds_deadlist,
1737 			    ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
1738 			    &used, &comp, &uncomp);
1739 			dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1740 			    -used, -comp, -uncomp, tx);
1741 
1742 			/* Move blocks to be freed to pool's free list. */
1743 			dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
1744 			    &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
1745 			    tx);
1746 			dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
1747 			    DD_USED_HEAD, used, comp, uncomp, tx);
1748 			dsl_dir_dirty(tx->tx_pool->dp_free_dir, tx);
1749 
1750 			/* Merge our deadlist into next's and free it. */
1751 			dsl_deadlist_merge(&ds_next->ds_deadlist,
1752 			    ds->ds_phys->ds_deadlist_obj, tx);
1753 		}
1754 		dsl_deadlist_close(&ds->ds_deadlist);
1755 		dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1756 
1757 		/* Collapse range in clone heads */
1758 		dsl_dataset_remove_clones_key(ds,
1759 		    ds->ds_phys->ds_creation_txg, tx);
1760 
1761 		if (dsl_dataset_is_snapshot(ds_next)) {
1762 			dsl_dataset_t *ds_nextnext;
1763 
1764 			/*
1765 			 * Update next's unique to include blocks which
1766 			 * were previously shared by only this snapshot
1767 			 * and it.  Those blocks will be born after the
1768 			 * prev snap and before this snap, and will have
1769 			 * died after the next snap and before the one
1770 			 * after that (ie. be on the snap after next's
1771 			 * deadlist).
1772 			 */
1773 			VERIFY(0 == dsl_dataset_hold_obj(dp,
1774 			    ds_next->ds_phys->ds_next_snap_obj,
1775 			    FTAG, &ds_nextnext));
1776 			dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
1777 			    ds->ds_phys->ds_prev_snap_txg,
1778 			    ds->ds_phys->ds_creation_txg,
1779 			    &used, &comp, &uncomp);
1780 			ds_next->ds_phys->ds_unique_bytes += used;
1781 			dsl_dataset_rele(ds_nextnext, FTAG);
1782 			ASSERT3P(ds_next->ds_prev, ==, NULL);
1783 
1784 			/* Collapse range in this head. */
1785 			dsl_dataset_t *hds;
1786 			VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
1787 			    ds->ds_dir->dd_phys->dd_head_dataset_obj,
1788 			    FTAG, &hds));
1789 			dsl_deadlist_remove_key(&hds->ds_deadlist,
1790 			    ds->ds_phys->ds_creation_txg, tx);
1791 			dsl_dataset_rele(hds, FTAG);
1792 
1793 		} else {
1794 			ASSERT3P(ds_next->ds_prev, ==, ds);
1795 			dsl_dataset_drop_ref(ds_next->ds_prev, ds_next);
1796 			ds_next->ds_prev = NULL;
1797 			if (ds_prev) {
1798 				VERIFY(0 == dsl_dataset_get_ref(dp,
1799 				    ds->ds_phys->ds_prev_snap_obj,
1800 				    ds_next, &ds_next->ds_prev));
1801 			}
1802 
1803 			dsl_dataset_recalc_head_uniq(ds_next);
1804 
1805 			/*
1806 			 * Reduce the amount of our unconsmed refreservation
1807 			 * being charged to our parent by the amount of
1808 			 * new unique data we have gained.
1809 			 */
1810 			if (old_unique < ds_next->ds_reserved) {
1811 				int64_t mrsdelta;
1812 				uint64_t new_unique =
1813 				    ds_next->ds_phys->ds_unique_bytes;
1814 
1815 				ASSERT(old_unique <= new_unique);
1816 				mrsdelta = MIN(new_unique - old_unique,
1817 				    ds_next->ds_reserved - old_unique);
1818 				dsl_dir_diduse_space(ds->ds_dir,
1819 				    DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
1820 			}
1821 		}
1822 		dsl_dataset_rele(ds_next, FTAG);
1823 	} else {
1824 		/*
1825 		 * There's no next snapshot, so this is a head dataset.
1826 		 * Destroy the deadlist.  Unless it's a clone, the
1827 		 * deadlist should be empty.  (If it's a clone, it's
1828 		 * safe to ignore the deadlist contents.)
1829 		 */
1830 		struct killarg ka;
1831 
1832 		dsl_deadlist_close(&ds->ds_deadlist);
1833 		dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1834 		ds->ds_phys->ds_deadlist_obj = 0;
1835 
1836 		/*
1837 		 * Free everything that we point to (that's born after
1838 		 * the previous snapshot, if we are a clone)
1839 		 *
1840 		 * NB: this should be very quick, because we already
1841 		 * freed all the objects in open context.
1842 		 */
1843 		ka.ds = ds;
1844 		ka.tx = tx;
1845 		err = traverse_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
1846 		    TRAVERSE_POST, kill_blkptr, &ka);
1847 		ASSERT3U(err, ==, 0);
1848 		ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
1849 		    ds->ds_phys->ds_unique_bytes == 0);
1850 
1851 		if (ds->ds_prev != NULL) {
1852 			if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
1853 				VERIFY3U(0, ==, zap_remove_int(mos,
1854 				    ds->ds_prev->ds_dir->dd_phys->dd_clones,
1855 				    ds->ds_object, tx));
1856 			}
1857 			dsl_dataset_rele(ds->ds_prev, ds);
1858 			ds->ds_prev = ds_prev = NULL;
1859 		}
1860 	}
1861 
1862 	/*
1863 	 * This must be done after the dsl_traverse(), because it will
1864 	 * re-open the objset.
1865 	 */
1866 	if (ds->ds_objset) {
1867 		dmu_objset_evict(ds->ds_objset);
1868 		ds->ds_objset = NULL;
1869 	}
1870 
1871 	if (ds->ds_dir->dd_phys->dd_head_dataset_obj == ds->ds_object) {
1872 		/* Erase the link in the dir */
1873 		dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
1874 		ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
1875 		ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
1876 		err = zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx);
1877 		ASSERT(err == 0);
1878 	} else {
1879 		/* remove from snapshot namespace */
1880 		dsl_dataset_t *ds_head;
1881 		ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
1882 		VERIFY(0 == dsl_dataset_hold_obj(dp,
1883 		    ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
1884 		VERIFY(0 == dsl_dataset_get_snapname(ds));
1885 #ifdef ZFS_DEBUG
1886 		{
1887 			uint64_t val;
1888 
1889 			err = dsl_dataset_snap_lookup(ds_head,
1890 			    ds->ds_snapname, &val);
1891 			ASSERT3U(err, ==, 0);
1892 			ASSERT3U(val, ==, obj);
1893 		}
1894 #endif
1895 		err = dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx);
1896 		ASSERT(err == 0);
1897 		dsl_dataset_rele(ds_head, FTAG);
1898 	}
1899 
1900 	if (ds_prev && ds->ds_prev != ds_prev)
1901 		dsl_dataset_rele(ds_prev, FTAG);
1902 
1903 	spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
1904 	spa_history_log_internal(LOG_DS_DESTROY, dp->dp_spa, tx,
1905 	    "dataset = %llu", ds->ds_object);
1906 
1907 	if (ds->ds_phys->ds_next_clones_obj != 0) {
1908 		uint64_t count;
1909 		ASSERT(0 == zap_count(mos,
1910 		    ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
1911 		VERIFY(0 == dmu_object_free(mos,
1912 		    ds->ds_phys->ds_next_clones_obj, tx));
1913 	}
1914 	if (ds->ds_phys->ds_props_obj != 0)
1915 		VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
1916 	if (ds->ds_phys->ds_userrefs_obj != 0)
1917 		VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
1918 	dsl_dir_close(ds->ds_dir, ds);
1919 	ds->ds_dir = NULL;
1920 	dsl_dataset_drain_refs(ds, tag);
1921 	VERIFY(0 == dmu_object_free(mos, obj, tx));
1922 
1923 	if (dsda->rm_origin) {
1924 		/*
1925 		 * Remove the origin of the clone we just destroyed.
1926 		 */
1927 		struct dsl_ds_destroyarg ndsda = {0};
1928 
1929 		ndsda.ds = dsda->rm_origin;
1930 		dsl_dataset_destroy_sync(&ndsda, tag, tx);
1931 	}
1932 }
1933 
1934 static int
1935 dsl_dataset_snapshot_reserve_space(dsl_dataset_t *ds, dmu_tx_t *tx)
1936 {
1937 	uint64_t asize;
1938 
1939 	if (!dmu_tx_is_syncing(tx))
1940 		return (0);
1941 
1942 	/*
1943 	 * If there's an fs-only reservation, any blocks that might become
1944 	 * owned by the snapshot dataset must be accommodated by space
1945 	 * outside of the reservation.
1946 	 */
1947 	ASSERT(ds->ds_reserved == 0 || DS_UNIQUE_IS_ACCURATE(ds));
1948 	asize = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
1949 	if (asize > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
1950 		return (ENOSPC);
1951 
1952 	/*
1953 	 * Propogate any reserved space for this snapshot to other
1954 	 * snapshot checks in this sync group.
1955 	 */
1956 	if (asize > 0)
1957 		dsl_dir_willuse_space(ds->ds_dir, asize, tx);
1958 
1959 	return (0);
1960 }
1961 
1962 int
1963 dsl_dataset_snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
1964 {
1965 	dsl_dataset_t *ds = arg1;
1966 	const char *snapname = arg2;
1967 	int err;
1968 	uint64_t value;
1969 
1970 	/*
1971 	 * We don't allow multiple snapshots of the same txg.  If there
1972 	 * is already one, try again.
1973 	 */
1974 	if (ds->ds_phys->ds_prev_snap_txg >= tx->tx_txg)
1975 		return (EAGAIN);
1976 
1977 	/*
1978 	 * Check for conflicting name snapshot name.
1979 	 */
1980 	err = dsl_dataset_snap_lookup(ds, snapname, &value);
1981 	if (err == 0)
1982 		return (EEXIST);
1983 	if (err != ENOENT)
1984 		return (err);
1985 
1986 	/*
1987 	 * Check that the dataset's name is not too long.  Name consists
1988 	 * of the dataset's length + 1 for the @-sign + snapshot name's length
1989 	 */
1990 	if (dsl_dataset_namelen(ds) + 1 + strlen(snapname) >= MAXNAMELEN)
1991 		return (ENAMETOOLONG);
1992 
1993 	err = dsl_dataset_snapshot_reserve_space(ds, tx);
1994 	if (err)
1995 		return (err);
1996 
1997 	ds->ds_trysnap_txg = tx->tx_txg;
1998 	return (0);
1999 }
2000 
2001 void
2002 dsl_dataset_snapshot_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2003 {
2004 	dsl_dataset_t *ds = arg1;
2005 	const char *snapname = arg2;
2006 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
2007 	dmu_buf_t *dbuf;
2008 	dsl_dataset_phys_t *dsphys;
2009 	uint64_t dsobj, crtxg;
2010 	objset_t *mos = dp->dp_meta_objset;
2011 	int err;
2012 
2013 	ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
2014 
2015 	/*
2016 	 * The origin's ds_creation_txg has to be < TXG_INITIAL
2017 	 */
2018 	if (strcmp(snapname, ORIGIN_DIR_NAME) == 0)
2019 		crtxg = 1;
2020 	else
2021 		crtxg = tx->tx_txg;
2022 
2023 	dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
2024 	    DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
2025 	VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
2026 	dmu_buf_will_dirty(dbuf, tx);
2027 	dsphys = dbuf->db_data;
2028 	bzero(dsphys, sizeof (dsl_dataset_phys_t));
2029 	dsphys->ds_dir_obj = ds->ds_dir->dd_object;
2030 	dsphys->ds_fsid_guid = unique_create();
2031 	(void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
2032 	    sizeof (dsphys->ds_guid));
2033 	dsphys->ds_prev_snap_obj = ds->ds_phys->ds_prev_snap_obj;
2034 	dsphys->ds_prev_snap_txg = ds->ds_phys->ds_prev_snap_txg;
2035 	dsphys->ds_next_snap_obj = ds->ds_object;
2036 	dsphys->ds_num_children = 1;
2037 	dsphys->ds_creation_time = gethrestime_sec();
2038 	dsphys->ds_creation_txg = crtxg;
2039 	dsphys->ds_deadlist_obj = ds->ds_phys->ds_deadlist_obj;
2040 	dsphys->ds_used_bytes = ds->ds_phys->ds_used_bytes;
2041 	dsphys->ds_compressed_bytes = ds->ds_phys->ds_compressed_bytes;
2042 	dsphys->ds_uncompressed_bytes = ds->ds_phys->ds_uncompressed_bytes;
2043 	dsphys->ds_flags = ds->ds_phys->ds_flags;
2044 	dsphys->ds_bp = ds->ds_phys->ds_bp;
2045 	dmu_buf_rele(dbuf, FTAG);
2046 
2047 	ASSERT3U(ds->ds_prev != 0, ==, ds->ds_phys->ds_prev_snap_obj != 0);
2048 	if (ds->ds_prev) {
2049 		uint64_t next_clones_obj =
2050 		    ds->ds_prev->ds_phys->ds_next_clones_obj;
2051 		ASSERT(ds->ds_prev->ds_phys->ds_next_snap_obj ==
2052 		    ds->ds_object ||
2053 		    ds->ds_prev->ds_phys->ds_num_children > 1);
2054 		if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
2055 			dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2056 			ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
2057 			    ds->ds_prev->ds_phys->ds_creation_txg);
2058 			ds->ds_prev->ds_phys->ds_next_snap_obj = dsobj;
2059 		} else if (next_clones_obj != 0) {
2060 			remove_from_next_clones(ds->ds_prev,
2061 			    dsphys->ds_next_snap_obj, tx);
2062 			VERIFY3U(0, ==, zap_add_int(mos,
2063 			    next_clones_obj, dsobj, tx));
2064 		}
2065 	}
2066 
2067 	/*
2068 	 * If we have a reference-reservation on this dataset, we will
2069 	 * need to increase the amount of refreservation being charged
2070 	 * since our unique space is going to zero.
2071 	 */
2072 	if (ds->ds_reserved) {
2073 		int64_t delta;
2074 		ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
2075 		delta = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
2076 		dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV,
2077 		    delta, 0, 0, tx);
2078 	}
2079 
2080 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
2081 	zfs_dbgmsg("taking snapshot %s@%s/%llu; newkey=%llu",
2082 	    ds->ds_dir->dd_myname, snapname, dsobj,
2083 	    ds->ds_phys->ds_prev_snap_txg);
2084 	ds->ds_phys->ds_deadlist_obj = dsl_deadlist_clone(&ds->ds_deadlist,
2085 	    UINT64_MAX, ds->ds_phys->ds_prev_snap_obj, tx);
2086 	dsl_deadlist_close(&ds->ds_deadlist);
2087 	dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
2088 	dsl_deadlist_add_key(&ds->ds_deadlist,
2089 	    ds->ds_phys->ds_prev_snap_txg, tx);
2090 
2091 	ASSERT3U(ds->ds_phys->ds_prev_snap_txg, <, tx->tx_txg);
2092 	ds->ds_phys->ds_prev_snap_obj = dsobj;
2093 	ds->ds_phys->ds_prev_snap_txg = crtxg;
2094 	ds->ds_phys->ds_unique_bytes = 0;
2095 	if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
2096 		ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
2097 
2098 	err = zap_add(mos, ds->ds_phys->ds_snapnames_zapobj,
2099 	    snapname, 8, 1, &dsobj, tx);
2100 	ASSERT(err == 0);
2101 
2102 	if (ds->ds_prev)
2103 		dsl_dataset_drop_ref(ds->ds_prev, ds);
2104 	VERIFY(0 == dsl_dataset_get_ref(dp,
2105 	    ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
2106 
2107 	dsl_scan_ds_snapshotted(ds, tx);
2108 
2109 	dsl_dir_snap_cmtime_update(ds->ds_dir);
2110 
2111 	spa_history_log_internal(LOG_DS_SNAPSHOT, dp->dp_spa, tx,
2112 	    "dataset = %llu", dsobj);
2113 }
2114 
2115 void
2116 dsl_dataset_sync(dsl_dataset_t *ds, zio_t *zio, dmu_tx_t *tx)
2117 {
2118 	ASSERT(dmu_tx_is_syncing(tx));
2119 	ASSERT(ds->ds_objset != NULL);
2120 	ASSERT(ds->ds_phys->ds_next_snap_obj == 0);
2121 
2122 	/*
2123 	 * in case we had to change ds_fsid_guid when we opened it,
2124 	 * sync it out now.
2125 	 */
2126 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
2127 	ds->ds_phys->ds_fsid_guid = ds->ds_fsid_guid;
2128 
2129 	dsl_dir_dirty(ds->ds_dir, tx);
2130 	dmu_objset_sync(ds->ds_objset, zio, tx);
2131 }
2132 
2133 static void
2134 get_clones_stat(dsl_dataset_t *ds, nvlist_t *nv)
2135 {
2136 	uint64_t count = 0;
2137 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
2138 	zap_cursor_t zc;
2139 	zap_attribute_t za;
2140 	nvlist_t *propval;
2141 	nvlist_t *val;
2142 
2143 	rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2144 	VERIFY(nvlist_alloc(&propval, NV_UNIQUE_NAME, KM_SLEEP) == 0);
2145 	VERIFY(nvlist_alloc(&val, NV_UNIQUE_NAME, KM_SLEEP) == 0);
2146 
2147 	/*
2148 	 * There may me missing entries in ds_next_clones_obj
2149 	 * due to a bug in a previous version of the code.
2150 	 * Only trust it if it has the right number of entries.
2151 	 */
2152 	if (ds->ds_phys->ds_next_clones_obj != 0) {
2153 		ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
2154 		    &count));
2155 	}
2156 	if (count != ds->ds_phys->ds_num_children - 1) {
2157 		goto fail;
2158 	}
2159 	for (zap_cursor_init(&zc, mos, ds->ds_phys->ds_next_clones_obj);
2160 	    zap_cursor_retrieve(&zc, &za) == 0;
2161 	    zap_cursor_advance(&zc)) {
2162 		dsl_dataset_t *clone;
2163 		char buf[ZFS_MAXNAMELEN];
2164 		if (dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
2165 		    za.za_first_integer, FTAG, &clone) != 0) {
2166 			goto fail;
2167 		}
2168 		dsl_dir_name(clone->ds_dir, buf);
2169 		VERIFY(nvlist_add_boolean(val, buf) == 0);
2170 		dsl_dataset_rele(clone, FTAG);
2171 	}
2172 	zap_cursor_fini(&zc);
2173 	VERIFY(nvlist_add_nvlist(propval, ZPROP_VALUE, val) == 0);
2174 	VERIFY(nvlist_add_nvlist(nv, zfs_prop_to_name(ZFS_PROP_CLONES),
2175 	    propval) == 0);
2176 fail:
2177 	nvlist_free(val);
2178 	nvlist_free(propval);
2179 	rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2180 }
2181 
2182 void
2183 dsl_dataset_stats(dsl_dataset_t *ds, nvlist_t *nv)
2184 {
2185 	uint64_t refd, avail, uobjs, aobjs, ratio;
2186 
2187 	dsl_dir_stats(ds->ds_dir, nv);
2188 
2189 	dsl_dataset_space(ds, &refd, &avail, &uobjs, &aobjs);
2190 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_AVAILABLE, avail);
2191 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFERENCED, refd);
2192 
2193 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATION,
2194 	    ds->ds_phys->ds_creation_time);
2195 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATETXG,
2196 	    ds->ds_phys->ds_creation_txg);
2197 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFQUOTA,
2198 	    ds->ds_quota);
2199 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRESERVATION,
2200 	    ds->ds_reserved);
2201 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_GUID,
2202 	    ds->ds_phys->ds_guid);
2203 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_UNIQUE,
2204 	    ds->ds_phys->ds_unique_bytes);
2205 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_OBJSETID,
2206 	    ds->ds_object);
2207 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USERREFS,
2208 	    ds->ds_userrefs);
2209 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_DEFER_DESTROY,
2210 	    DS_IS_DEFER_DESTROY(ds) ? 1 : 0);
2211 
2212 	if (ds->ds_phys->ds_prev_snap_obj != 0) {
2213 		uint64_t written, comp, uncomp;
2214 		dsl_pool_t *dp = ds->ds_dir->dd_pool;
2215 		dsl_dataset_t *prev;
2216 
2217 		rw_enter(&dp->dp_config_rwlock, RW_READER);
2218 		int err = dsl_dataset_hold_obj(dp,
2219 		    ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
2220 		rw_exit(&dp->dp_config_rwlock);
2221 		if (err == 0) {
2222 			err = dsl_dataset_space_written(prev, ds, &written,
2223 			    &comp, &uncomp);
2224 			dsl_dataset_rele(prev, FTAG);
2225 			if (err == 0) {
2226 				dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_WRITTEN,
2227 				    written);
2228 			}
2229 		}
2230 	}
2231 
2232 	ratio = ds->ds_phys->ds_compressed_bytes == 0 ? 100 :
2233 	    (ds->ds_phys->ds_uncompressed_bytes * 100 /
2234 	    ds->ds_phys->ds_compressed_bytes);
2235 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRATIO, ratio);
2236 
2237 	if (ds->ds_phys->ds_next_snap_obj) {
2238 		/*
2239 		 * This is a snapshot; override the dd's space used with
2240 		 * our unique space and compression ratio.
2241 		 */
2242 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USED,
2243 		    ds->ds_phys->ds_unique_bytes);
2244 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_COMPRESSRATIO, ratio);
2245 
2246 		get_clones_stat(ds, nv);
2247 	}
2248 }
2249 
2250 void
2251 dsl_dataset_fast_stat(dsl_dataset_t *ds, dmu_objset_stats_t *stat)
2252 {
2253 	stat->dds_creation_txg = ds->ds_phys->ds_creation_txg;
2254 	stat->dds_inconsistent = ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT;
2255 	stat->dds_guid = ds->ds_phys->ds_guid;
2256 	if (ds->ds_phys->ds_next_snap_obj) {
2257 		stat->dds_is_snapshot = B_TRUE;
2258 		stat->dds_num_clones = ds->ds_phys->ds_num_children - 1;
2259 	} else {
2260 		stat->dds_is_snapshot = B_FALSE;
2261 		stat->dds_num_clones = 0;
2262 	}
2263 
2264 	/* clone origin is really a dsl_dir thing... */
2265 	rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2266 	if (dsl_dir_is_clone(ds->ds_dir)) {
2267 		dsl_dataset_t *ods;
2268 
2269 		VERIFY(0 == dsl_dataset_get_ref(ds->ds_dir->dd_pool,
2270 		    ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &ods));
2271 		dsl_dataset_name(ods, stat->dds_origin);
2272 		dsl_dataset_drop_ref(ods, FTAG);
2273 	} else {
2274 		stat->dds_origin[0] = '\0';
2275 	}
2276 	rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2277 }
2278 
2279 uint64_t
2280 dsl_dataset_fsid_guid(dsl_dataset_t *ds)
2281 {
2282 	return (ds->ds_fsid_guid);
2283 }
2284 
2285 void
2286 dsl_dataset_space(dsl_dataset_t *ds,
2287     uint64_t *refdbytesp, uint64_t *availbytesp,
2288     uint64_t *usedobjsp, uint64_t *availobjsp)
2289 {
2290 	*refdbytesp = ds->ds_phys->ds_used_bytes;
2291 	*availbytesp = dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE);
2292 	if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes)
2293 		*availbytesp += ds->ds_reserved - ds->ds_phys->ds_unique_bytes;
2294 	if (ds->ds_quota != 0) {
2295 		/*
2296 		 * Adjust available bytes according to refquota
2297 		 */
2298 		if (*refdbytesp < ds->ds_quota)
2299 			*availbytesp = MIN(*availbytesp,
2300 			    ds->ds_quota - *refdbytesp);
2301 		else
2302 			*availbytesp = 0;
2303 	}
2304 	*usedobjsp = ds->ds_phys->ds_bp.blk_fill;
2305 	*availobjsp = DN_MAX_OBJECT - *usedobjsp;
2306 }
2307 
2308 boolean_t
2309 dsl_dataset_modified_since_lastsnap(dsl_dataset_t *ds)
2310 {
2311 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
2312 
2313 	ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
2314 	    dsl_pool_sync_context(dp));
2315 	if (ds->ds_prev == NULL)
2316 		return (B_FALSE);
2317 	if (ds->ds_phys->ds_bp.blk_birth >
2318 	    ds->ds_prev->ds_phys->ds_creation_txg) {
2319 		objset_t *os, *os_prev;
2320 		/*
2321 		 * It may be that only the ZIL differs, because it was
2322 		 * reset in the head.  Don't count that as being
2323 		 * modified.
2324 		 */
2325 		if (dmu_objset_from_ds(ds, &os) != 0)
2326 			return (B_TRUE);
2327 		if (dmu_objset_from_ds(ds->ds_prev, &os_prev) != 0)
2328 			return (B_TRUE);
2329 		return (bcmp(&os->os_phys->os_meta_dnode,
2330 		    &os_prev->os_phys->os_meta_dnode,
2331 		    sizeof (os->os_phys->os_meta_dnode)) != 0);
2332 	}
2333 	return (B_FALSE);
2334 }
2335 
2336 /* ARGSUSED */
2337 static int
2338 dsl_dataset_snapshot_rename_check(void *arg1, void *arg2, dmu_tx_t *tx)
2339 {
2340 	dsl_dataset_t *ds = arg1;
2341 	char *newsnapname = arg2;
2342 	dsl_dir_t *dd = ds->ds_dir;
2343 	dsl_dataset_t *hds;
2344 	uint64_t val;
2345 	int err;
2346 
2347 	err = dsl_dataset_hold_obj(dd->dd_pool,
2348 	    dd->dd_phys->dd_head_dataset_obj, FTAG, &hds);
2349 	if (err)
2350 		return (err);
2351 
2352 	/* new name better not be in use */
2353 	err = dsl_dataset_snap_lookup(hds, newsnapname, &val);
2354 	dsl_dataset_rele(hds, FTAG);
2355 
2356 	if (err == 0)
2357 		err = EEXIST;
2358 	else if (err == ENOENT)
2359 		err = 0;
2360 
2361 	/* dataset name + 1 for the "@" + the new snapshot name must fit */
2362 	if (dsl_dir_namelen(ds->ds_dir) + 1 + strlen(newsnapname) >= MAXNAMELEN)
2363 		err = ENAMETOOLONG;
2364 
2365 	return (err);
2366 }
2367 
2368 static void
2369 dsl_dataset_snapshot_rename_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2370 {
2371 	dsl_dataset_t *ds = arg1;
2372 	const char *newsnapname = arg2;
2373 	dsl_dir_t *dd = ds->ds_dir;
2374 	objset_t *mos = dd->dd_pool->dp_meta_objset;
2375 	dsl_dataset_t *hds;
2376 	int err;
2377 
2378 	ASSERT(ds->ds_phys->ds_next_snap_obj != 0);
2379 
2380 	VERIFY(0 == dsl_dataset_hold_obj(dd->dd_pool,
2381 	    dd->dd_phys->dd_head_dataset_obj, FTAG, &hds));
2382 
2383 	VERIFY(0 == dsl_dataset_get_snapname(ds));
2384 	err = dsl_dataset_snap_remove(hds, ds->ds_snapname, tx);
2385 	ASSERT3U(err, ==, 0);
2386 	mutex_enter(&ds->ds_lock);
2387 	(void) strcpy(ds->ds_snapname, newsnapname);
2388 	mutex_exit(&ds->ds_lock);
2389 	err = zap_add(mos, hds->ds_phys->ds_snapnames_zapobj,
2390 	    ds->ds_snapname, 8, 1, &ds->ds_object, tx);
2391 	ASSERT3U(err, ==, 0);
2392 
2393 	spa_history_log_internal(LOG_DS_RENAME, dd->dd_pool->dp_spa, tx,
2394 	    "dataset = %llu", ds->ds_object);
2395 	dsl_dataset_rele(hds, FTAG);
2396 }
2397 
2398 struct renamesnaparg {
2399 	dsl_sync_task_group_t *dstg;
2400 	char failed[MAXPATHLEN];
2401 	char *oldsnap;
2402 	char *newsnap;
2403 };
2404 
2405 static int
2406 dsl_snapshot_rename_one(const char *name, void *arg)
2407 {
2408 	struct renamesnaparg *ra = arg;
2409 	dsl_dataset_t *ds = NULL;
2410 	char *snapname;
2411 	int err;
2412 
2413 	snapname = kmem_asprintf("%s@%s", name, ra->oldsnap);
2414 	(void) strlcpy(ra->failed, snapname, sizeof (ra->failed));
2415 
2416 	/*
2417 	 * For recursive snapshot renames the parent won't be changing
2418 	 * so we just pass name for both the to/from argument.
2419 	 */
2420 	err = zfs_secpolicy_rename_perms(snapname, snapname, CRED());
2421 	if (err != 0) {
2422 		strfree(snapname);
2423 		return (err == ENOENT ? 0 : err);
2424 	}
2425 
2426 #ifdef _KERNEL
2427 	/*
2428 	 * For all filesystems undergoing rename, we'll need to unmount it.
2429 	 */
2430 	(void) zfs_unmount_snap(snapname, NULL);
2431 #endif
2432 	err = dsl_dataset_hold(snapname, ra->dstg, &ds);
2433 	strfree(snapname);
2434 	if (err != 0)
2435 		return (err == ENOENT ? 0 : err);
2436 
2437 	dsl_sync_task_create(ra->dstg, dsl_dataset_snapshot_rename_check,
2438 	    dsl_dataset_snapshot_rename_sync, ds, ra->newsnap, 0);
2439 
2440 	return (0);
2441 }
2442 
2443 static int
2444 dsl_recursive_rename(char *oldname, const char *newname)
2445 {
2446 	int err;
2447 	struct renamesnaparg *ra;
2448 	dsl_sync_task_t *dst;
2449 	spa_t *spa;
2450 	char *cp, *fsname = spa_strdup(oldname);
2451 	int len = strlen(oldname) + 1;
2452 
2453 	/* truncate the snapshot name to get the fsname */
2454 	cp = strchr(fsname, '@');
2455 	*cp = '\0';
2456 
2457 	err = spa_open(fsname, &spa, FTAG);
2458 	if (err) {
2459 		kmem_free(fsname, len);
2460 		return (err);
2461 	}
2462 	ra = kmem_alloc(sizeof (struct renamesnaparg), KM_SLEEP);
2463 	ra->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
2464 
2465 	ra->oldsnap = strchr(oldname, '@') + 1;
2466 	ra->newsnap = strchr(newname, '@') + 1;
2467 	*ra->failed = '\0';
2468 
2469 	err = dmu_objset_find(fsname, dsl_snapshot_rename_one, ra,
2470 	    DS_FIND_CHILDREN);
2471 	kmem_free(fsname, len);
2472 
2473 	if (err == 0) {
2474 		err = dsl_sync_task_group_wait(ra->dstg);
2475 	}
2476 
2477 	for (dst = list_head(&ra->dstg->dstg_tasks); dst;
2478 	    dst = list_next(&ra->dstg->dstg_tasks, dst)) {
2479 		dsl_dataset_t *ds = dst->dst_arg1;
2480 		if (dst->dst_err) {
2481 			dsl_dir_name(ds->ds_dir, ra->failed);
2482 			(void) strlcat(ra->failed, "@", sizeof (ra->failed));
2483 			(void) strlcat(ra->failed, ra->newsnap,
2484 			    sizeof (ra->failed));
2485 		}
2486 		dsl_dataset_rele(ds, ra->dstg);
2487 	}
2488 
2489 	if (err)
2490 		(void) strlcpy(oldname, ra->failed, sizeof (ra->failed));
2491 
2492 	dsl_sync_task_group_destroy(ra->dstg);
2493 	kmem_free(ra, sizeof (struct renamesnaparg));
2494 	spa_close(spa, FTAG);
2495 	return (err);
2496 }
2497 
2498 static int
2499 dsl_valid_rename(const char *oldname, void *arg)
2500 {
2501 	int delta = *(int *)arg;
2502 
2503 	if (strlen(oldname) + delta >= MAXNAMELEN)
2504 		return (ENAMETOOLONG);
2505 
2506 	return (0);
2507 }
2508 
2509 #pragma weak dmu_objset_rename = dsl_dataset_rename
2510 int
2511 dsl_dataset_rename(char *oldname, const char *newname, boolean_t recursive)
2512 {
2513 	dsl_dir_t *dd;
2514 	dsl_dataset_t *ds;
2515 	const char *tail;
2516 	int err;
2517 
2518 	err = dsl_dir_open(oldname, FTAG, &dd, &tail);
2519 	if (err)
2520 		return (err);
2521 
2522 	if (tail == NULL) {
2523 		int delta = strlen(newname) - strlen(oldname);
2524 
2525 		/* if we're growing, validate child name lengths */
2526 		if (delta > 0)
2527 			err = dmu_objset_find(oldname, dsl_valid_rename,
2528 			    &delta, DS_FIND_CHILDREN | DS_FIND_SNAPSHOTS);
2529 
2530 		if (err == 0)
2531 			err = dsl_dir_rename(dd, newname);
2532 		dsl_dir_close(dd, FTAG);
2533 		return (err);
2534 	}
2535 
2536 	if (tail[0] != '@') {
2537 		/* the name ended in a nonexistent component */
2538 		dsl_dir_close(dd, FTAG);
2539 		return (ENOENT);
2540 	}
2541 
2542 	dsl_dir_close(dd, FTAG);
2543 
2544 	/* new name must be snapshot in same filesystem */
2545 	tail = strchr(newname, '@');
2546 	if (tail == NULL)
2547 		return (EINVAL);
2548 	tail++;
2549 	if (strncmp(oldname, newname, tail - newname) != 0)
2550 		return (EXDEV);
2551 
2552 	if (recursive) {
2553 		err = dsl_recursive_rename(oldname, newname);
2554 	} else {
2555 		err = dsl_dataset_hold(oldname, FTAG, &ds);
2556 		if (err)
2557 			return (err);
2558 
2559 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
2560 		    dsl_dataset_snapshot_rename_check,
2561 		    dsl_dataset_snapshot_rename_sync, ds, (char *)tail, 1);
2562 
2563 		dsl_dataset_rele(ds, FTAG);
2564 	}
2565 
2566 	return (err);
2567 }
2568 
2569 struct promotenode {
2570 	list_node_t link;
2571 	dsl_dataset_t *ds;
2572 };
2573 
2574 struct promotearg {
2575 	list_t shared_snaps, origin_snaps, clone_snaps;
2576 	dsl_dataset_t *origin_origin;
2577 	uint64_t used, comp, uncomp, unique, cloneusedsnap, originusedsnap;
2578 	char *err_ds;
2579 };
2580 
2581 static int snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep);
2582 static boolean_t snaplist_unstable(list_t *l);
2583 
2584 static int
2585 dsl_dataset_promote_check(void *arg1, void *arg2, dmu_tx_t *tx)
2586 {
2587 	dsl_dataset_t *hds = arg1;
2588 	struct promotearg *pa = arg2;
2589 	struct promotenode *snap = list_head(&pa->shared_snaps);
2590 	dsl_dataset_t *origin_ds = snap->ds;
2591 	int err;
2592 	uint64_t unused;
2593 
2594 	/* Check that it is a real clone */
2595 	if (!dsl_dir_is_clone(hds->ds_dir))
2596 		return (EINVAL);
2597 
2598 	/* Since this is so expensive, don't do the preliminary check */
2599 	if (!dmu_tx_is_syncing(tx))
2600 		return (0);
2601 
2602 	if (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE)
2603 		return (EXDEV);
2604 
2605 	/* compute origin's new unique space */
2606 	snap = list_tail(&pa->clone_snaps);
2607 	ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2608 	dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2609 	    origin_ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
2610 	    &pa->unique, &unused, &unused);
2611 
2612 	/*
2613 	 * Walk the snapshots that we are moving
2614 	 *
2615 	 * Compute space to transfer.  Consider the incremental changes
2616 	 * to used for each snapshot:
2617 	 * (my used) = (prev's used) + (blocks born) - (blocks killed)
2618 	 * So each snapshot gave birth to:
2619 	 * (blocks born) = (my used) - (prev's used) + (blocks killed)
2620 	 * So a sequence would look like:
2621 	 * (uN - u(N-1) + kN) + ... + (u1 - u0 + k1) + (u0 - 0 + k0)
2622 	 * Which simplifies to:
2623 	 * uN + kN + kN-1 + ... + k1 + k0
2624 	 * Note however, if we stop before we reach the ORIGIN we get:
2625 	 * uN + kN + kN-1 + ... + kM - uM-1
2626 	 */
2627 	pa->used = origin_ds->ds_phys->ds_used_bytes;
2628 	pa->comp = origin_ds->ds_phys->ds_compressed_bytes;
2629 	pa->uncomp = origin_ds->ds_phys->ds_uncompressed_bytes;
2630 	for (snap = list_head(&pa->shared_snaps); snap;
2631 	    snap = list_next(&pa->shared_snaps, snap)) {
2632 		uint64_t val, dlused, dlcomp, dluncomp;
2633 		dsl_dataset_t *ds = snap->ds;
2634 
2635 		/* Check that the snapshot name does not conflict */
2636 		VERIFY(0 == dsl_dataset_get_snapname(ds));
2637 		err = dsl_dataset_snap_lookup(hds, ds->ds_snapname, &val);
2638 		if (err == 0) {
2639 			err = EEXIST;
2640 			goto out;
2641 		}
2642 		if (err != ENOENT)
2643 			goto out;
2644 
2645 		/* The very first snapshot does not have a deadlist */
2646 		if (ds->ds_phys->ds_prev_snap_obj == 0)
2647 			continue;
2648 
2649 		dsl_deadlist_space(&ds->ds_deadlist,
2650 		    &dlused, &dlcomp, &dluncomp);
2651 		pa->used += dlused;
2652 		pa->comp += dlcomp;
2653 		pa->uncomp += dluncomp;
2654 	}
2655 
2656 	/*
2657 	 * If we are a clone of a clone then we never reached ORIGIN,
2658 	 * so we need to subtract out the clone origin's used space.
2659 	 */
2660 	if (pa->origin_origin) {
2661 		pa->used -= pa->origin_origin->ds_phys->ds_used_bytes;
2662 		pa->comp -= pa->origin_origin->ds_phys->ds_compressed_bytes;
2663 		pa->uncomp -= pa->origin_origin->ds_phys->ds_uncompressed_bytes;
2664 	}
2665 
2666 	/* Check that there is enough space here */
2667 	err = dsl_dir_transfer_possible(origin_ds->ds_dir, hds->ds_dir,
2668 	    pa->used);
2669 	if (err)
2670 		return (err);
2671 
2672 	/*
2673 	 * Compute the amounts of space that will be used by snapshots
2674 	 * after the promotion (for both origin and clone).  For each,
2675 	 * it is the amount of space that will be on all of their
2676 	 * deadlists (that was not born before their new origin).
2677 	 */
2678 	if (hds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2679 		uint64_t space;
2680 
2681 		/*
2682 		 * Note, typically this will not be a clone of a clone,
2683 		 * so dd_origin_txg will be < TXG_INITIAL, so
2684 		 * these snaplist_space() -> dsl_deadlist_space_range()
2685 		 * calls will be fast because they do not have to
2686 		 * iterate over all bps.
2687 		 */
2688 		snap = list_head(&pa->origin_snaps);
2689 		err = snaplist_space(&pa->shared_snaps,
2690 		    snap->ds->ds_dir->dd_origin_txg, &pa->cloneusedsnap);
2691 		if (err)
2692 			return (err);
2693 
2694 		err = snaplist_space(&pa->clone_snaps,
2695 		    snap->ds->ds_dir->dd_origin_txg, &space);
2696 		if (err)
2697 			return (err);
2698 		pa->cloneusedsnap += space;
2699 	}
2700 	if (origin_ds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2701 		err = snaplist_space(&pa->origin_snaps,
2702 		    origin_ds->ds_phys->ds_creation_txg, &pa->originusedsnap);
2703 		if (err)
2704 			return (err);
2705 	}
2706 
2707 	return (0);
2708 out:
2709 	pa->err_ds =  snap->ds->ds_snapname;
2710 	return (err);
2711 }
2712 
2713 static void
2714 dsl_dataset_promote_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2715 {
2716 	dsl_dataset_t *hds = arg1;
2717 	struct promotearg *pa = arg2;
2718 	struct promotenode *snap = list_head(&pa->shared_snaps);
2719 	dsl_dataset_t *origin_ds = snap->ds;
2720 	dsl_dataset_t *origin_head;
2721 	dsl_dir_t *dd = hds->ds_dir;
2722 	dsl_pool_t *dp = hds->ds_dir->dd_pool;
2723 	dsl_dir_t *odd = NULL;
2724 	uint64_t oldnext_obj;
2725 	int64_t delta;
2726 
2727 	ASSERT(0 == (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE));
2728 
2729 	snap = list_head(&pa->origin_snaps);
2730 	origin_head = snap->ds;
2731 
2732 	/*
2733 	 * We need to explicitly open odd, since origin_ds's dd will be
2734 	 * changing.
2735 	 */
2736 	VERIFY(0 == dsl_dir_open_obj(dp, origin_ds->ds_dir->dd_object,
2737 	    NULL, FTAG, &odd));
2738 
2739 	/* change origin's next snap */
2740 	dmu_buf_will_dirty(origin_ds->ds_dbuf, tx);
2741 	oldnext_obj = origin_ds->ds_phys->ds_next_snap_obj;
2742 	snap = list_tail(&pa->clone_snaps);
2743 	ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2744 	origin_ds->ds_phys->ds_next_snap_obj = snap->ds->ds_object;
2745 
2746 	/* change the origin's next clone */
2747 	if (origin_ds->ds_phys->ds_next_clones_obj) {
2748 		remove_from_next_clones(origin_ds, snap->ds->ds_object, tx);
2749 		VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2750 		    origin_ds->ds_phys->ds_next_clones_obj,
2751 		    oldnext_obj, tx));
2752 	}
2753 
2754 	/* change origin */
2755 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
2756 	ASSERT3U(dd->dd_phys->dd_origin_obj, ==, origin_ds->ds_object);
2757 	dd->dd_phys->dd_origin_obj = odd->dd_phys->dd_origin_obj;
2758 	dd->dd_origin_txg = origin_head->ds_dir->dd_origin_txg;
2759 	dmu_buf_will_dirty(odd->dd_dbuf, tx);
2760 	odd->dd_phys->dd_origin_obj = origin_ds->ds_object;
2761 	origin_head->ds_dir->dd_origin_txg =
2762 	    origin_ds->ds_phys->ds_creation_txg;
2763 
2764 	/* change dd_clone entries */
2765 	if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2766 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2767 		    odd->dd_phys->dd_clones, hds->ds_object, tx));
2768 		VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2769 		    pa->origin_origin->ds_dir->dd_phys->dd_clones,
2770 		    hds->ds_object, tx));
2771 
2772 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2773 		    pa->origin_origin->ds_dir->dd_phys->dd_clones,
2774 		    origin_head->ds_object, tx));
2775 		if (dd->dd_phys->dd_clones == 0) {
2776 			dd->dd_phys->dd_clones = zap_create(dp->dp_meta_objset,
2777 			    DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
2778 		}
2779 		VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2780 		    dd->dd_phys->dd_clones, origin_head->ds_object, tx));
2781 
2782 	}
2783 
2784 	/* move snapshots to this dir */
2785 	for (snap = list_head(&pa->shared_snaps); snap;
2786 	    snap = list_next(&pa->shared_snaps, snap)) {
2787 		dsl_dataset_t *ds = snap->ds;
2788 
2789 		/* unregister props as dsl_dir is changing */
2790 		if (ds->ds_objset) {
2791 			dmu_objset_evict(ds->ds_objset);
2792 			ds->ds_objset = NULL;
2793 		}
2794 		/* move snap name entry */
2795 		VERIFY(0 == dsl_dataset_get_snapname(ds));
2796 		VERIFY(0 == dsl_dataset_snap_remove(origin_head,
2797 		    ds->ds_snapname, tx));
2798 		VERIFY(0 == zap_add(dp->dp_meta_objset,
2799 		    hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
2800 		    8, 1, &ds->ds_object, tx));
2801 
2802 		/* change containing dsl_dir */
2803 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
2804 		ASSERT3U(ds->ds_phys->ds_dir_obj, ==, odd->dd_object);
2805 		ds->ds_phys->ds_dir_obj = dd->dd_object;
2806 		ASSERT3P(ds->ds_dir, ==, odd);
2807 		dsl_dir_close(ds->ds_dir, ds);
2808 		VERIFY(0 == dsl_dir_open_obj(dp, dd->dd_object,
2809 		    NULL, ds, &ds->ds_dir));
2810 
2811 		/* move any clone references */
2812 		if (ds->ds_phys->ds_next_clones_obj &&
2813 		    spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2814 			zap_cursor_t zc;
2815 			zap_attribute_t za;
2816 
2817 			for (zap_cursor_init(&zc, dp->dp_meta_objset,
2818 			    ds->ds_phys->ds_next_clones_obj);
2819 			    zap_cursor_retrieve(&zc, &za) == 0;
2820 			    zap_cursor_advance(&zc)) {
2821 				dsl_dataset_t *cnds;
2822 				uint64_t o;
2823 
2824 				if (za.za_first_integer == oldnext_obj) {
2825 					/*
2826 					 * We've already moved the
2827 					 * origin's reference.
2828 					 */
2829 					continue;
2830 				}
2831 
2832 				VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
2833 				    za.za_first_integer, FTAG, &cnds));
2834 				o = cnds->ds_dir->dd_phys->dd_head_dataset_obj;
2835 
2836 				VERIFY3U(zap_remove_int(dp->dp_meta_objset,
2837 				    odd->dd_phys->dd_clones, o, tx), ==, 0);
2838 				VERIFY3U(zap_add_int(dp->dp_meta_objset,
2839 				    dd->dd_phys->dd_clones, o, tx), ==, 0);
2840 				dsl_dataset_rele(cnds, FTAG);
2841 			}
2842 			zap_cursor_fini(&zc);
2843 		}
2844 
2845 		ASSERT3U(dsl_prop_numcb(ds), ==, 0);
2846 	}
2847 
2848 	/*
2849 	 * Change space accounting.
2850 	 * Note, pa->*usedsnap and dd_used_breakdown[SNAP] will either
2851 	 * both be valid, or both be 0 (resulting in delta == 0).  This
2852 	 * is true for each of {clone,origin} independently.
2853 	 */
2854 
2855 	delta = pa->cloneusedsnap -
2856 	    dd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2857 	ASSERT3S(delta, >=, 0);
2858 	ASSERT3U(pa->used, >=, delta);
2859 	dsl_dir_diduse_space(dd, DD_USED_SNAP, delta, 0, 0, tx);
2860 	dsl_dir_diduse_space(dd, DD_USED_HEAD,
2861 	    pa->used - delta, pa->comp, pa->uncomp, tx);
2862 
2863 	delta = pa->originusedsnap -
2864 	    odd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2865 	ASSERT3S(delta, <=, 0);
2866 	ASSERT3U(pa->used, >=, -delta);
2867 	dsl_dir_diduse_space(odd, DD_USED_SNAP, delta, 0, 0, tx);
2868 	dsl_dir_diduse_space(odd, DD_USED_HEAD,
2869 	    -pa->used - delta, -pa->comp, -pa->uncomp, tx);
2870 
2871 	origin_ds->ds_phys->ds_unique_bytes = pa->unique;
2872 
2873 	/* log history record */
2874 	spa_history_log_internal(LOG_DS_PROMOTE, dd->dd_pool->dp_spa, tx,
2875 	    "dataset = %llu", hds->ds_object);
2876 
2877 	dsl_dir_close(odd, FTAG);
2878 }
2879 
2880 static char *snaplist_tag = "snaplist";
2881 /*
2882  * Make a list of dsl_dataset_t's for the snapshots between first_obj
2883  * (exclusive) and last_obj (inclusive).  The list will be in reverse
2884  * order (last_obj will be the list_head()).  If first_obj == 0, do all
2885  * snapshots back to this dataset's origin.
2886  */
2887 static int
2888 snaplist_make(dsl_pool_t *dp, boolean_t own,
2889     uint64_t first_obj, uint64_t last_obj, list_t *l)
2890 {
2891 	uint64_t obj = last_obj;
2892 
2893 	ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock));
2894 
2895 	list_create(l, sizeof (struct promotenode),
2896 	    offsetof(struct promotenode, link));
2897 
2898 	while (obj != first_obj) {
2899 		dsl_dataset_t *ds;
2900 		struct promotenode *snap;
2901 		int err;
2902 
2903 		if (own) {
2904 			err = dsl_dataset_own_obj(dp, obj,
2905 			    0, snaplist_tag, &ds);
2906 			if (err == 0)
2907 				dsl_dataset_make_exclusive(ds, snaplist_tag);
2908 		} else {
2909 			err = dsl_dataset_hold_obj(dp, obj, snaplist_tag, &ds);
2910 		}
2911 		if (err == ENOENT) {
2912 			/* lost race with snapshot destroy */
2913 			struct promotenode *last = list_tail(l);
2914 			ASSERT(obj != last->ds->ds_phys->ds_prev_snap_obj);
2915 			obj = last->ds->ds_phys->ds_prev_snap_obj;
2916 			continue;
2917 		} else if (err) {
2918 			return (err);
2919 		}
2920 
2921 		if (first_obj == 0)
2922 			first_obj = ds->ds_dir->dd_phys->dd_origin_obj;
2923 
2924 		snap = kmem_alloc(sizeof (struct promotenode), KM_SLEEP);
2925 		snap->ds = ds;
2926 		list_insert_tail(l, snap);
2927 		obj = ds->ds_phys->ds_prev_snap_obj;
2928 	}
2929 
2930 	return (0);
2931 }
2932 
2933 static int
2934 snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep)
2935 {
2936 	struct promotenode *snap;
2937 
2938 	*spacep = 0;
2939 	for (snap = list_head(l); snap; snap = list_next(l, snap)) {
2940 		uint64_t used, comp, uncomp;
2941 		dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2942 		    mintxg, UINT64_MAX, &used, &comp, &uncomp);
2943 		*spacep += used;
2944 	}
2945 	return (0);
2946 }
2947 
2948 static void
2949 snaplist_destroy(list_t *l, boolean_t own)
2950 {
2951 	struct promotenode *snap;
2952 
2953 	if (!l || !list_link_active(&l->list_head))
2954 		return;
2955 
2956 	while ((snap = list_tail(l)) != NULL) {
2957 		list_remove(l, snap);
2958 		if (own)
2959 			dsl_dataset_disown(snap->ds, snaplist_tag);
2960 		else
2961 			dsl_dataset_rele(snap->ds, snaplist_tag);
2962 		kmem_free(snap, sizeof (struct promotenode));
2963 	}
2964 	list_destroy(l);
2965 }
2966 
2967 /*
2968  * Promote a clone.  Nomenclature note:
2969  * "clone" or "cds": the original clone which is being promoted
2970  * "origin" or "ods": the snapshot which is originally clone's origin
2971  * "origin head" or "ohds": the dataset which is the head
2972  * (filesystem/volume) for the origin
2973  * "origin origin": the origin of the origin's filesystem (typically
2974  * NULL, indicating that the clone is not a clone of a clone).
2975  */
2976 int
2977 dsl_dataset_promote(const char *name, char *conflsnap)
2978 {
2979 	dsl_dataset_t *ds;
2980 	dsl_dir_t *dd;
2981 	dsl_pool_t *dp;
2982 	dmu_object_info_t doi;
2983 	struct promotearg pa = { 0 };
2984 	struct promotenode *snap;
2985 	int err;
2986 
2987 	err = dsl_dataset_hold(name, FTAG, &ds);
2988 	if (err)
2989 		return (err);
2990 	dd = ds->ds_dir;
2991 	dp = dd->dd_pool;
2992 
2993 	err = dmu_object_info(dp->dp_meta_objset,
2994 	    ds->ds_phys->ds_snapnames_zapobj, &doi);
2995 	if (err) {
2996 		dsl_dataset_rele(ds, FTAG);
2997 		return (err);
2998 	}
2999 
3000 	if (dsl_dataset_is_snapshot(ds) || dd->dd_phys->dd_origin_obj == 0) {
3001 		dsl_dataset_rele(ds, FTAG);
3002 		return (EINVAL);
3003 	}
3004 
3005 	/*
3006 	 * We are going to inherit all the snapshots taken before our
3007 	 * origin (i.e., our new origin will be our parent's origin).
3008 	 * Take ownership of them so that we can rename them into our
3009 	 * namespace.
3010 	 */
3011 	rw_enter(&dp->dp_config_rwlock, RW_READER);
3012 
3013 	err = snaplist_make(dp, B_TRUE, 0, dd->dd_phys->dd_origin_obj,
3014 	    &pa.shared_snaps);
3015 	if (err != 0)
3016 		goto out;
3017 
3018 	err = snaplist_make(dp, B_FALSE, 0, ds->ds_object, &pa.clone_snaps);
3019 	if (err != 0)
3020 		goto out;
3021 
3022 	snap = list_head(&pa.shared_snaps);
3023 	ASSERT3U(snap->ds->ds_object, ==, dd->dd_phys->dd_origin_obj);
3024 	err = snaplist_make(dp, B_FALSE, dd->dd_phys->dd_origin_obj,
3025 	    snap->ds->ds_dir->dd_phys->dd_head_dataset_obj, &pa.origin_snaps);
3026 	if (err != 0)
3027 		goto out;
3028 
3029 	if (snap->ds->ds_dir->dd_phys->dd_origin_obj != 0) {
3030 		err = dsl_dataset_hold_obj(dp,
3031 		    snap->ds->ds_dir->dd_phys->dd_origin_obj,
3032 		    FTAG, &pa.origin_origin);
3033 		if (err != 0)
3034 			goto out;
3035 	}
3036 
3037 out:
3038 	rw_exit(&dp->dp_config_rwlock);
3039 
3040 	/*
3041 	 * Add in 128x the snapnames zapobj size, since we will be moving
3042 	 * a bunch of snapnames to the promoted ds, and dirtying their
3043 	 * bonus buffers.
3044 	 */
3045 	if (err == 0) {
3046 		err = dsl_sync_task_do(dp, dsl_dataset_promote_check,
3047 		    dsl_dataset_promote_sync, ds, &pa,
3048 		    2 + 2 * doi.doi_physical_blocks_512);
3049 		if (err && pa.err_ds && conflsnap)
3050 			(void) strncpy(conflsnap, pa.err_ds, MAXNAMELEN);
3051 	}
3052 
3053 	snaplist_destroy(&pa.shared_snaps, B_TRUE);
3054 	snaplist_destroy(&pa.clone_snaps, B_FALSE);
3055 	snaplist_destroy(&pa.origin_snaps, B_FALSE);
3056 	if (pa.origin_origin)
3057 		dsl_dataset_rele(pa.origin_origin, FTAG);
3058 	dsl_dataset_rele(ds, FTAG);
3059 	return (err);
3060 }
3061 
3062 struct cloneswaparg {
3063 	dsl_dataset_t *cds; /* clone dataset */
3064 	dsl_dataset_t *ohds; /* origin's head dataset */
3065 	boolean_t force;
3066 	int64_t unused_refres_delta; /* change in unconsumed refreservation */
3067 };
3068 
3069 /* ARGSUSED */
3070 static int
3071 dsl_dataset_clone_swap_check(void *arg1, void *arg2, dmu_tx_t *tx)
3072 {
3073 	struct cloneswaparg *csa = arg1;
3074 
3075 	/* they should both be heads */
3076 	if (dsl_dataset_is_snapshot(csa->cds) ||
3077 	    dsl_dataset_is_snapshot(csa->ohds))
3078 		return (EINVAL);
3079 
3080 	/* the branch point should be just before them */
3081 	if (csa->cds->ds_prev != csa->ohds->ds_prev)
3082 		return (EINVAL);
3083 
3084 	/* cds should be the clone (unless they are unrelated) */
3085 	if (csa->cds->ds_prev != NULL &&
3086 	    csa->cds->ds_prev != csa->cds->ds_dir->dd_pool->dp_origin_snap &&
3087 	    csa->ohds->ds_object !=
3088 	    csa->cds->ds_prev->ds_phys->ds_next_snap_obj)
3089 		return (EINVAL);
3090 
3091 	/* the clone should be a child of the origin */
3092 	if (csa->cds->ds_dir->dd_parent != csa->ohds->ds_dir)
3093 		return (EINVAL);
3094 
3095 	/* ohds shouldn't be modified unless 'force' */
3096 	if (!csa->force && dsl_dataset_modified_since_lastsnap(csa->ohds))
3097 		return (ETXTBSY);
3098 
3099 	/* adjust amount of any unconsumed refreservation */
3100 	csa->unused_refres_delta =
3101 	    (int64_t)MIN(csa->ohds->ds_reserved,
3102 	    csa->ohds->ds_phys->ds_unique_bytes) -
3103 	    (int64_t)MIN(csa->ohds->ds_reserved,
3104 	    csa->cds->ds_phys->ds_unique_bytes);
3105 
3106 	if (csa->unused_refres_delta > 0 &&
3107 	    csa->unused_refres_delta >
3108 	    dsl_dir_space_available(csa->ohds->ds_dir, NULL, 0, TRUE))
3109 		return (ENOSPC);
3110 
3111 	if (csa->ohds->ds_quota != 0 &&
3112 	    csa->cds->ds_phys->ds_unique_bytes > csa->ohds->ds_quota)
3113 		return (EDQUOT);
3114 
3115 	return (0);
3116 }
3117 
3118 /* ARGSUSED */
3119 static void
3120 dsl_dataset_clone_swap_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3121 {
3122 	struct cloneswaparg *csa = arg1;
3123 	dsl_pool_t *dp = csa->cds->ds_dir->dd_pool;
3124 
3125 	ASSERT(csa->cds->ds_reserved == 0);
3126 	ASSERT(csa->ohds->ds_quota == 0 ||
3127 	    csa->cds->ds_phys->ds_unique_bytes <= csa->ohds->ds_quota);
3128 
3129 	dmu_buf_will_dirty(csa->cds->ds_dbuf, tx);
3130 	dmu_buf_will_dirty(csa->ohds->ds_dbuf, tx);
3131 
3132 	if (csa->cds->ds_objset != NULL) {
3133 		dmu_objset_evict(csa->cds->ds_objset);
3134 		csa->cds->ds_objset = NULL;
3135 	}
3136 
3137 	if (csa->ohds->ds_objset != NULL) {
3138 		dmu_objset_evict(csa->ohds->ds_objset);
3139 		csa->ohds->ds_objset = NULL;
3140 	}
3141 
3142 	/*
3143 	 * Reset origin's unique bytes, if it exists.
3144 	 */
3145 	if (csa->cds->ds_prev) {
3146 		dsl_dataset_t *origin = csa->cds->ds_prev;
3147 		uint64_t comp, uncomp;
3148 
3149 		dmu_buf_will_dirty(origin->ds_dbuf, tx);
3150 		dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3151 		    origin->ds_phys->ds_prev_snap_txg, UINT64_MAX,
3152 		    &origin->ds_phys->ds_unique_bytes, &comp, &uncomp);
3153 	}
3154 
3155 	/* swap blkptrs */
3156 	{
3157 		blkptr_t tmp;
3158 		tmp = csa->ohds->ds_phys->ds_bp;
3159 		csa->ohds->ds_phys->ds_bp = csa->cds->ds_phys->ds_bp;
3160 		csa->cds->ds_phys->ds_bp = tmp;
3161 	}
3162 
3163 	/* set dd_*_bytes */
3164 	{
3165 		int64_t dused, dcomp, duncomp;
3166 		uint64_t cdl_used, cdl_comp, cdl_uncomp;
3167 		uint64_t odl_used, odl_comp, odl_uncomp;
3168 
3169 		ASSERT3U(csa->cds->ds_dir->dd_phys->
3170 		    dd_used_breakdown[DD_USED_SNAP], ==, 0);
3171 
3172 		dsl_deadlist_space(&csa->cds->ds_deadlist,
3173 		    &cdl_used, &cdl_comp, &cdl_uncomp);
3174 		dsl_deadlist_space(&csa->ohds->ds_deadlist,
3175 		    &odl_used, &odl_comp, &odl_uncomp);
3176 
3177 		dused = csa->cds->ds_phys->ds_used_bytes + cdl_used -
3178 		    (csa->ohds->ds_phys->ds_used_bytes + odl_used);
3179 		dcomp = csa->cds->ds_phys->ds_compressed_bytes + cdl_comp -
3180 		    (csa->ohds->ds_phys->ds_compressed_bytes + odl_comp);
3181 		duncomp = csa->cds->ds_phys->ds_uncompressed_bytes +
3182 		    cdl_uncomp -
3183 		    (csa->ohds->ds_phys->ds_uncompressed_bytes + odl_uncomp);
3184 
3185 		dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_HEAD,
3186 		    dused, dcomp, duncomp, tx);
3187 		dsl_dir_diduse_space(csa->cds->ds_dir, DD_USED_HEAD,
3188 		    -dused, -dcomp, -duncomp, tx);
3189 
3190 		/*
3191 		 * The difference in the space used by snapshots is the
3192 		 * difference in snapshot space due to the head's
3193 		 * deadlist (since that's the only thing that's
3194 		 * changing that affects the snapused).
3195 		 */
3196 		dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3197 		    csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3198 		    &cdl_used, &cdl_comp, &cdl_uncomp);
3199 		dsl_deadlist_space_range(&csa->ohds->ds_deadlist,
3200 		    csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3201 		    &odl_used, &odl_comp, &odl_uncomp);
3202 		dsl_dir_transfer_space(csa->ohds->ds_dir, cdl_used - odl_used,
3203 		    DD_USED_HEAD, DD_USED_SNAP, tx);
3204 	}
3205 
3206 	/* swap ds_*_bytes */
3207 	SWITCH64(csa->ohds->ds_phys->ds_used_bytes,
3208 	    csa->cds->ds_phys->ds_used_bytes);
3209 	SWITCH64(csa->ohds->ds_phys->ds_compressed_bytes,
3210 	    csa->cds->ds_phys->ds_compressed_bytes);
3211 	SWITCH64(csa->ohds->ds_phys->ds_uncompressed_bytes,
3212 	    csa->cds->ds_phys->ds_uncompressed_bytes);
3213 	SWITCH64(csa->ohds->ds_phys->ds_unique_bytes,
3214 	    csa->cds->ds_phys->ds_unique_bytes);
3215 
3216 	/* apply any parent delta for change in unconsumed refreservation */
3217 	dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_REFRSRV,
3218 	    csa->unused_refres_delta, 0, 0, tx);
3219 
3220 	/*
3221 	 * Swap deadlists.
3222 	 */
3223 	dsl_deadlist_close(&csa->cds->ds_deadlist);
3224 	dsl_deadlist_close(&csa->ohds->ds_deadlist);
3225 	SWITCH64(csa->ohds->ds_phys->ds_deadlist_obj,
3226 	    csa->cds->ds_phys->ds_deadlist_obj);
3227 	dsl_deadlist_open(&csa->cds->ds_deadlist, dp->dp_meta_objset,
3228 	    csa->cds->ds_phys->ds_deadlist_obj);
3229 	dsl_deadlist_open(&csa->ohds->ds_deadlist, dp->dp_meta_objset,
3230 	    csa->ohds->ds_phys->ds_deadlist_obj);
3231 
3232 	dsl_scan_ds_clone_swapped(csa->ohds, csa->cds, tx);
3233 }
3234 
3235 /*
3236  * Swap 'clone' with its origin head datasets.  Used at the end of "zfs
3237  * recv" into an existing fs to swizzle the file system to the new
3238  * version, and by "zfs rollback".  Can also be used to swap two
3239  * independent head datasets if neither has any snapshots.
3240  */
3241 int
3242 dsl_dataset_clone_swap(dsl_dataset_t *clone, dsl_dataset_t *origin_head,
3243     boolean_t force)
3244 {
3245 	struct cloneswaparg csa;
3246 	int error;
3247 
3248 	ASSERT(clone->ds_owner);
3249 	ASSERT(origin_head->ds_owner);
3250 retry:
3251 	/*
3252 	 * Need exclusive access for the swap. If we're swapping these
3253 	 * datasets back after an error, we already hold the locks.
3254 	 */
3255 	if (!RW_WRITE_HELD(&clone->ds_rwlock))
3256 		rw_enter(&clone->ds_rwlock, RW_WRITER);
3257 	if (!RW_WRITE_HELD(&origin_head->ds_rwlock) &&
3258 	    !rw_tryenter(&origin_head->ds_rwlock, RW_WRITER)) {
3259 		rw_exit(&clone->ds_rwlock);
3260 		rw_enter(&origin_head->ds_rwlock, RW_WRITER);
3261 		if (!rw_tryenter(&clone->ds_rwlock, RW_WRITER)) {
3262 			rw_exit(&origin_head->ds_rwlock);
3263 			goto retry;
3264 		}
3265 	}
3266 	csa.cds = clone;
3267 	csa.ohds = origin_head;
3268 	csa.force = force;
3269 	error = dsl_sync_task_do(clone->ds_dir->dd_pool,
3270 	    dsl_dataset_clone_swap_check,
3271 	    dsl_dataset_clone_swap_sync, &csa, NULL, 9);
3272 	return (error);
3273 }
3274 
3275 /*
3276  * Given a pool name and a dataset object number in that pool,
3277  * return the name of that dataset.
3278  */
3279 int
3280 dsl_dsobj_to_dsname(char *pname, uint64_t obj, char *buf)
3281 {
3282 	spa_t *spa;
3283 	dsl_pool_t *dp;
3284 	dsl_dataset_t *ds;
3285 	int error;
3286 
3287 	if ((error = spa_open(pname, &spa, FTAG)) != 0)
3288 		return (error);
3289 	dp = spa_get_dsl(spa);
3290 	rw_enter(&dp->dp_config_rwlock, RW_READER);
3291 	if ((error = dsl_dataset_hold_obj(dp, obj, FTAG, &ds)) == 0) {
3292 		dsl_dataset_name(ds, buf);
3293 		dsl_dataset_rele(ds, FTAG);
3294 	}
3295 	rw_exit(&dp->dp_config_rwlock);
3296 	spa_close(spa, FTAG);
3297 
3298 	return (error);
3299 }
3300 
3301 int
3302 dsl_dataset_check_quota(dsl_dataset_t *ds, boolean_t check_quota,
3303     uint64_t asize, uint64_t inflight, uint64_t *used, uint64_t *ref_rsrv)
3304 {
3305 	int error = 0;
3306 
3307 	ASSERT3S(asize, >, 0);
3308 
3309 	/*
3310 	 * *ref_rsrv is the portion of asize that will come from any
3311 	 * unconsumed refreservation space.
3312 	 */
3313 	*ref_rsrv = 0;
3314 
3315 	mutex_enter(&ds->ds_lock);
3316 	/*
3317 	 * Make a space adjustment for reserved bytes.
3318 	 */
3319 	if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes) {
3320 		ASSERT3U(*used, >=,
3321 		    ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3322 		*used -= (ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3323 		*ref_rsrv =
3324 		    asize - MIN(asize, parent_delta(ds, asize + inflight));
3325 	}
3326 
3327 	if (!check_quota || ds->ds_quota == 0) {
3328 		mutex_exit(&ds->ds_lock);
3329 		return (0);
3330 	}
3331 	/*
3332 	 * If they are requesting more space, and our current estimate
3333 	 * is over quota, they get to try again unless the actual
3334 	 * on-disk is over quota and there are no pending changes (which
3335 	 * may free up space for us).
3336 	 */
3337 	if (ds->ds_phys->ds_used_bytes + inflight >= ds->ds_quota) {
3338 		if (inflight > 0 || ds->ds_phys->ds_used_bytes < ds->ds_quota)
3339 			error = ERESTART;
3340 		else
3341 			error = EDQUOT;
3342 	}
3343 	mutex_exit(&ds->ds_lock);
3344 
3345 	return (error);
3346 }
3347 
3348 /* ARGSUSED */
3349 static int
3350 dsl_dataset_set_quota_check(void *arg1, void *arg2, dmu_tx_t *tx)
3351 {
3352 	dsl_dataset_t *ds = arg1;
3353 	dsl_prop_setarg_t *psa = arg2;
3354 	int err;
3355 
3356 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_REFQUOTA)
3357 		return (ENOTSUP);
3358 
3359 	if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3360 		return (err);
3361 
3362 	if (psa->psa_effective_value == 0)
3363 		return (0);
3364 
3365 	if (psa->psa_effective_value < ds->ds_phys->ds_used_bytes ||
3366 	    psa->psa_effective_value < ds->ds_reserved)
3367 		return (ENOSPC);
3368 
3369 	return (0);
3370 }
3371 
3372 extern void dsl_prop_set_sync(void *, void *, dmu_tx_t *);
3373 
3374 void
3375 dsl_dataset_set_quota_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3376 {
3377 	dsl_dataset_t *ds = arg1;
3378 	dsl_prop_setarg_t *psa = arg2;
3379 	uint64_t effective_value = psa->psa_effective_value;
3380 
3381 	dsl_prop_set_sync(ds, psa, tx);
3382 	DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3383 
3384 	if (ds->ds_quota != effective_value) {
3385 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
3386 		ds->ds_quota = effective_value;
3387 
3388 		spa_history_log_internal(LOG_DS_REFQUOTA,
3389 		    ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu ",
3390 		    (longlong_t)ds->ds_quota, ds->ds_object);
3391 	}
3392 }
3393 
3394 int
3395 dsl_dataset_set_quota(const char *dsname, zprop_source_t source, uint64_t quota)
3396 {
3397 	dsl_dataset_t *ds;
3398 	dsl_prop_setarg_t psa;
3399 	int err;
3400 
3401 	dsl_prop_setarg_init_uint64(&psa, "refquota", source, &quota);
3402 
3403 	err = dsl_dataset_hold(dsname, FTAG, &ds);
3404 	if (err)
3405 		return (err);
3406 
3407 	/*
3408 	 * If someone removes a file, then tries to set the quota, we
3409 	 * want to make sure the file freeing takes effect.
3410 	 */
3411 	txg_wait_open(ds->ds_dir->dd_pool, 0);
3412 
3413 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3414 	    dsl_dataset_set_quota_check, dsl_dataset_set_quota_sync,
3415 	    ds, &psa, 0);
3416 
3417 	dsl_dataset_rele(ds, FTAG);
3418 	return (err);
3419 }
3420 
3421 static int
3422 dsl_dataset_set_reservation_check(void *arg1, void *arg2, dmu_tx_t *tx)
3423 {
3424 	dsl_dataset_t *ds = arg1;
3425 	dsl_prop_setarg_t *psa = arg2;
3426 	uint64_t effective_value;
3427 	uint64_t unique;
3428 	int err;
3429 
3430 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
3431 	    SPA_VERSION_REFRESERVATION)
3432 		return (ENOTSUP);
3433 
3434 	if (dsl_dataset_is_snapshot(ds))
3435 		return (EINVAL);
3436 
3437 	if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3438 		return (err);
3439 
3440 	effective_value = psa->psa_effective_value;
3441 
3442 	/*
3443 	 * If we are doing the preliminary check in open context, the
3444 	 * space estimates may be inaccurate.
3445 	 */
3446 	if (!dmu_tx_is_syncing(tx))
3447 		return (0);
3448 
3449 	mutex_enter(&ds->ds_lock);
3450 	if (!DS_UNIQUE_IS_ACCURATE(ds))
3451 		dsl_dataset_recalc_head_uniq(ds);
3452 	unique = ds->ds_phys->ds_unique_bytes;
3453 	mutex_exit(&ds->ds_lock);
3454 
3455 	if (MAX(unique, effective_value) > MAX(unique, ds->ds_reserved)) {
3456 		uint64_t delta = MAX(unique, effective_value) -
3457 		    MAX(unique, ds->ds_reserved);
3458 
3459 		if (delta > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
3460 			return (ENOSPC);
3461 		if (ds->ds_quota > 0 &&
3462 		    effective_value > ds->ds_quota)
3463 			return (ENOSPC);
3464 	}
3465 
3466 	return (0);
3467 }
3468 
3469 static void
3470 dsl_dataset_set_reservation_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3471 {
3472 	dsl_dataset_t *ds = arg1;
3473 	dsl_prop_setarg_t *psa = arg2;
3474 	uint64_t effective_value = psa->psa_effective_value;
3475 	uint64_t unique;
3476 	int64_t delta;
3477 
3478 	dsl_prop_set_sync(ds, psa, tx);
3479 	DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3480 
3481 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
3482 
3483 	mutex_enter(&ds->ds_dir->dd_lock);
3484 	mutex_enter(&ds->ds_lock);
3485 	ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
3486 	unique = ds->ds_phys->ds_unique_bytes;
3487 	delta = MAX(0, (int64_t)(effective_value - unique)) -
3488 	    MAX(0, (int64_t)(ds->ds_reserved - unique));
3489 	ds->ds_reserved = effective_value;
3490 	mutex_exit(&ds->ds_lock);
3491 
3492 	dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV, delta, 0, 0, tx);
3493 	mutex_exit(&ds->ds_dir->dd_lock);
3494 
3495 	spa_history_log_internal(LOG_DS_REFRESERV,
3496 	    ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu",
3497 	    (longlong_t)effective_value, ds->ds_object);
3498 }
3499 
3500 int
3501 dsl_dataset_set_reservation(const char *dsname, zprop_source_t source,
3502     uint64_t reservation)
3503 {
3504 	dsl_dataset_t *ds;
3505 	dsl_prop_setarg_t psa;
3506 	int err;
3507 
3508 	dsl_prop_setarg_init_uint64(&psa, "refreservation", source,
3509 	    &reservation);
3510 
3511 	err = dsl_dataset_hold(dsname, FTAG, &ds);
3512 	if (err)
3513 		return (err);
3514 
3515 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3516 	    dsl_dataset_set_reservation_check,
3517 	    dsl_dataset_set_reservation_sync, ds, &psa, 0);
3518 
3519 	dsl_dataset_rele(ds, FTAG);
3520 	return (err);
3521 }
3522 
3523 typedef struct zfs_hold_cleanup_arg {
3524 	dsl_pool_t *dp;
3525 	uint64_t dsobj;
3526 	char htag[MAXNAMELEN];
3527 } zfs_hold_cleanup_arg_t;
3528 
3529 static void
3530 dsl_dataset_user_release_onexit(void *arg)
3531 {
3532 	zfs_hold_cleanup_arg_t *ca = arg;
3533 
3534 	(void) dsl_dataset_user_release_tmp(ca->dp, ca->dsobj, ca->htag,
3535 	    B_TRUE);
3536 	kmem_free(ca, sizeof (zfs_hold_cleanup_arg_t));
3537 }
3538 
3539 void
3540 dsl_register_onexit_hold_cleanup(dsl_dataset_t *ds, const char *htag,
3541     minor_t minor)
3542 {
3543 	zfs_hold_cleanup_arg_t *ca;
3544 
3545 	ca = kmem_alloc(sizeof (zfs_hold_cleanup_arg_t), KM_SLEEP);
3546 	ca->dp = ds->ds_dir->dd_pool;
3547 	ca->dsobj = ds->ds_object;
3548 	(void) strlcpy(ca->htag, htag, sizeof (ca->htag));
3549 	VERIFY3U(0, ==, zfs_onexit_add_cb(minor,
3550 	    dsl_dataset_user_release_onexit, ca, NULL));
3551 }
3552 
3553 /*
3554  * If you add new checks here, you may need to add
3555  * additional checks to the "temporary" case in
3556  * snapshot_check() in dmu_objset.c.
3557  */
3558 static int
3559 dsl_dataset_user_hold_check(void *arg1, void *arg2, dmu_tx_t *tx)
3560 {
3561 	dsl_dataset_t *ds = arg1;
3562 	struct dsl_ds_holdarg *ha = arg2;
3563 	char *htag = ha->htag;
3564 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3565 	int error = 0;
3566 
3567 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3568 		return (ENOTSUP);
3569 
3570 	if (!dsl_dataset_is_snapshot(ds))
3571 		return (EINVAL);
3572 
3573 	/* tags must be unique */
3574 	mutex_enter(&ds->ds_lock);
3575 	if (ds->ds_phys->ds_userrefs_obj) {
3576 		error = zap_lookup(mos, ds->ds_phys->ds_userrefs_obj, htag,
3577 		    8, 1, tx);
3578 		if (error == 0)
3579 			error = EEXIST;
3580 		else if (error == ENOENT)
3581 			error = 0;
3582 	}
3583 	mutex_exit(&ds->ds_lock);
3584 
3585 	if (error == 0 && ha->temphold &&
3586 	    strlen(htag) + MAX_TAG_PREFIX_LEN >= MAXNAMELEN)
3587 		error = E2BIG;
3588 
3589 	return (error);
3590 }
3591 
3592 void
3593 dsl_dataset_user_hold_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3594 {
3595 	dsl_dataset_t *ds = arg1;
3596 	struct dsl_ds_holdarg *ha = arg2;
3597 	char *htag = ha->htag;
3598 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
3599 	objset_t *mos = dp->dp_meta_objset;
3600 	uint64_t now = gethrestime_sec();
3601 	uint64_t zapobj;
3602 
3603 	mutex_enter(&ds->ds_lock);
3604 	if (ds->ds_phys->ds_userrefs_obj == 0) {
3605 		/*
3606 		 * This is the first user hold for this dataset.  Create
3607 		 * the userrefs zap object.
3608 		 */
3609 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
3610 		zapobj = ds->ds_phys->ds_userrefs_obj =
3611 		    zap_create(mos, DMU_OT_USERREFS, DMU_OT_NONE, 0, tx);
3612 	} else {
3613 		zapobj = ds->ds_phys->ds_userrefs_obj;
3614 	}
3615 	ds->ds_userrefs++;
3616 	mutex_exit(&ds->ds_lock);
3617 
3618 	VERIFY(0 == zap_add(mos, zapobj, htag, 8, 1, &now, tx));
3619 
3620 	if (ha->temphold) {
3621 		VERIFY(0 == dsl_pool_user_hold(dp, ds->ds_object,
3622 		    htag, &now, tx));
3623 	}
3624 
3625 	spa_history_log_internal(LOG_DS_USER_HOLD,
3626 	    dp->dp_spa, tx, "<%s> temp = %d dataset = %llu", htag,
3627 	    (int)ha->temphold, ds->ds_object);
3628 }
3629 
3630 static int
3631 dsl_dataset_user_hold_one(const char *dsname, void *arg)
3632 {
3633 	struct dsl_ds_holdarg *ha = arg;
3634 	dsl_dataset_t *ds;
3635 	int error;
3636 	char *name;
3637 
3638 	/* alloc a buffer to hold dsname@snapname plus terminating NULL */
3639 	name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3640 	error = dsl_dataset_hold(name, ha->dstg, &ds);
3641 	strfree(name);
3642 	if (error == 0) {
3643 		ha->gotone = B_TRUE;
3644 		dsl_sync_task_create(ha->dstg, dsl_dataset_user_hold_check,
3645 		    dsl_dataset_user_hold_sync, ds, ha, 0);
3646 	} else if (error == ENOENT && ha->recursive) {
3647 		error = 0;
3648 	} else {
3649 		(void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3650 	}
3651 	return (error);
3652 }
3653 
3654 int
3655 dsl_dataset_user_hold_for_send(dsl_dataset_t *ds, char *htag,
3656     boolean_t temphold)
3657 {
3658 	struct dsl_ds_holdarg *ha;
3659 	int error;
3660 
3661 	ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3662 	ha->htag = htag;
3663 	ha->temphold = temphold;
3664 	error = dsl_sync_task_do(ds->ds_dir->dd_pool,
3665 	    dsl_dataset_user_hold_check, dsl_dataset_user_hold_sync,
3666 	    ds, ha, 0);
3667 	kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3668 
3669 	return (error);
3670 }
3671 
3672 int
3673 dsl_dataset_user_hold(char *dsname, char *snapname, char *htag,
3674     boolean_t recursive, boolean_t temphold, int cleanup_fd)
3675 {
3676 	struct dsl_ds_holdarg *ha;
3677 	dsl_sync_task_t *dst;
3678 	spa_t *spa;
3679 	int error;
3680 	minor_t minor = 0;
3681 
3682 	if (cleanup_fd != -1) {
3683 		/* Currently we only support cleanup-on-exit of tempholds. */
3684 		if (!temphold)
3685 			return (EINVAL);
3686 		error = zfs_onexit_fd_hold(cleanup_fd, &minor);
3687 		if (error)
3688 			return (error);
3689 	}
3690 
3691 	ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3692 
3693 	(void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3694 
3695 	error = spa_open(dsname, &spa, FTAG);
3696 	if (error) {
3697 		kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3698 		if (cleanup_fd != -1)
3699 			zfs_onexit_fd_rele(cleanup_fd);
3700 		return (error);
3701 	}
3702 
3703 	ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3704 	ha->htag = htag;
3705 	ha->snapname = snapname;
3706 	ha->recursive = recursive;
3707 	ha->temphold = temphold;
3708 
3709 	if (recursive) {
3710 		error = dmu_objset_find(dsname, dsl_dataset_user_hold_one,
3711 		    ha, DS_FIND_CHILDREN);
3712 	} else {
3713 		error = dsl_dataset_user_hold_one(dsname, ha);
3714 	}
3715 	if (error == 0)
3716 		error = dsl_sync_task_group_wait(ha->dstg);
3717 
3718 	for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3719 	    dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3720 		dsl_dataset_t *ds = dst->dst_arg1;
3721 
3722 		if (dst->dst_err) {
3723 			dsl_dataset_name(ds, ha->failed);
3724 			*strchr(ha->failed, '@') = '\0';
3725 		} else if (error == 0 && minor != 0 && temphold) {
3726 			/*
3727 			 * If this hold is to be released upon process exit,
3728 			 * register that action now.
3729 			 */
3730 			dsl_register_onexit_hold_cleanup(ds, htag, minor);
3731 		}
3732 		dsl_dataset_rele(ds, ha->dstg);
3733 	}
3734 
3735 	if (error == 0 && recursive && !ha->gotone)
3736 		error = ENOENT;
3737 
3738 	if (error)
3739 		(void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3740 
3741 	dsl_sync_task_group_destroy(ha->dstg);
3742 
3743 	kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3744 	spa_close(spa, FTAG);
3745 	if (cleanup_fd != -1)
3746 		zfs_onexit_fd_rele(cleanup_fd);
3747 	return (error);
3748 }
3749 
3750 struct dsl_ds_releasearg {
3751 	dsl_dataset_t *ds;
3752 	const char *htag;
3753 	boolean_t own;		/* do we own or just hold ds? */
3754 };
3755 
3756 static int
3757 dsl_dataset_release_might_destroy(dsl_dataset_t *ds, const char *htag,
3758     boolean_t *might_destroy)
3759 {
3760 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3761 	uint64_t zapobj;
3762 	uint64_t tmp;
3763 	int error;
3764 
3765 	*might_destroy = B_FALSE;
3766 
3767 	mutex_enter(&ds->ds_lock);
3768 	zapobj = ds->ds_phys->ds_userrefs_obj;
3769 	if (zapobj == 0) {
3770 		/* The tag can't possibly exist */
3771 		mutex_exit(&ds->ds_lock);
3772 		return (ESRCH);
3773 	}
3774 
3775 	/* Make sure the tag exists */
3776 	error = zap_lookup(mos, zapobj, htag, 8, 1, &tmp);
3777 	if (error) {
3778 		mutex_exit(&ds->ds_lock);
3779 		if (error == ENOENT)
3780 			error = ESRCH;
3781 		return (error);
3782 	}
3783 
3784 	if (ds->ds_userrefs == 1 && ds->ds_phys->ds_num_children == 1 &&
3785 	    DS_IS_DEFER_DESTROY(ds))
3786 		*might_destroy = B_TRUE;
3787 
3788 	mutex_exit(&ds->ds_lock);
3789 	return (0);
3790 }
3791 
3792 static int
3793 dsl_dataset_user_release_check(void *arg1, void *tag, dmu_tx_t *tx)
3794 {
3795 	struct dsl_ds_releasearg *ra = arg1;
3796 	dsl_dataset_t *ds = ra->ds;
3797 	boolean_t might_destroy;
3798 	int error;
3799 
3800 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3801 		return (ENOTSUP);
3802 
3803 	error = dsl_dataset_release_might_destroy(ds, ra->htag, &might_destroy);
3804 	if (error)
3805 		return (error);
3806 
3807 	if (might_destroy) {
3808 		struct dsl_ds_destroyarg dsda = {0};
3809 
3810 		if (dmu_tx_is_syncing(tx)) {
3811 			/*
3812 			 * If we're not prepared to remove the snapshot,
3813 			 * we can't allow the release to happen right now.
3814 			 */
3815 			if (!ra->own)
3816 				return (EBUSY);
3817 		}
3818 		dsda.ds = ds;
3819 		dsda.releasing = B_TRUE;
3820 		return (dsl_dataset_destroy_check(&dsda, tag, tx));
3821 	}
3822 
3823 	return (0);
3824 }
3825 
3826 static void
3827 dsl_dataset_user_release_sync(void *arg1, void *tag, dmu_tx_t *tx)
3828 {
3829 	struct dsl_ds_releasearg *ra = arg1;
3830 	dsl_dataset_t *ds = ra->ds;
3831 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
3832 	objset_t *mos = dp->dp_meta_objset;
3833 	uint64_t zapobj;
3834 	uint64_t dsobj = ds->ds_object;
3835 	uint64_t refs;
3836 	int error;
3837 
3838 	mutex_enter(&ds->ds_lock);
3839 	ds->ds_userrefs--;
3840 	refs = ds->ds_userrefs;
3841 	mutex_exit(&ds->ds_lock);
3842 	error = dsl_pool_user_release(dp, ds->ds_object, ra->htag, tx);
3843 	VERIFY(error == 0 || error == ENOENT);
3844 	zapobj = ds->ds_phys->ds_userrefs_obj;
3845 	VERIFY(0 == zap_remove(mos, zapobj, ra->htag, tx));
3846 	if (ds->ds_userrefs == 0 && ds->ds_phys->ds_num_children == 1 &&
3847 	    DS_IS_DEFER_DESTROY(ds)) {
3848 		struct dsl_ds_destroyarg dsda = {0};
3849 
3850 		ASSERT(ra->own);
3851 		dsda.ds = ds;
3852 		dsda.releasing = B_TRUE;
3853 		/* We already did the destroy_check */
3854 		dsl_dataset_destroy_sync(&dsda, tag, tx);
3855 	}
3856 
3857 	spa_history_log_internal(LOG_DS_USER_RELEASE,
3858 	    dp->dp_spa, tx, "<%s> %lld dataset = %llu",
3859 	    ra->htag, (longlong_t)refs, dsobj);
3860 }
3861 
3862 static int
3863 dsl_dataset_user_release_one(const char *dsname, void *arg)
3864 {
3865 	struct dsl_ds_holdarg *ha = arg;
3866 	struct dsl_ds_releasearg *ra;
3867 	dsl_dataset_t *ds;
3868 	int error;
3869 	void *dtag = ha->dstg;
3870 	char *name;
3871 	boolean_t own = B_FALSE;
3872 	boolean_t might_destroy;
3873 
3874 	/* alloc a buffer to hold dsname@snapname, plus the terminating NULL */
3875 	name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3876 	error = dsl_dataset_hold(name, dtag, &ds);
3877 	strfree(name);
3878 	if (error == ENOENT && ha->recursive)
3879 		return (0);
3880 	(void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3881 	if (error)
3882 		return (error);
3883 
3884 	ha->gotone = B_TRUE;
3885 
3886 	ASSERT(dsl_dataset_is_snapshot(ds));
3887 
3888 	error = dsl_dataset_release_might_destroy(ds, ha->htag, &might_destroy);
3889 	if (error) {
3890 		dsl_dataset_rele(ds, dtag);
3891 		return (error);
3892 	}
3893 
3894 	if (might_destroy) {
3895 #ifdef _KERNEL
3896 		name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3897 		error = zfs_unmount_snap(name, NULL);
3898 		strfree(name);
3899 		if (error) {
3900 			dsl_dataset_rele(ds, dtag);
3901 			return (error);
3902 		}
3903 #endif
3904 		if (!dsl_dataset_tryown(ds, B_TRUE, dtag)) {
3905 			dsl_dataset_rele(ds, dtag);
3906 			return (EBUSY);
3907 		} else {
3908 			own = B_TRUE;
3909 			dsl_dataset_make_exclusive(ds, dtag);
3910 		}
3911 	}
3912 
3913 	ra = kmem_alloc(sizeof (struct dsl_ds_releasearg), KM_SLEEP);
3914 	ra->ds = ds;
3915 	ra->htag = ha->htag;
3916 	ra->own = own;
3917 	dsl_sync_task_create(ha->dstg, dsl_dataset_user_release_check,
3918 	    dsl_dataset_user_release_sync, ra, dtag, 0);
3919 
3920 	return (0);
3921 }
3922 
3923 int
3924 dsl_dataset_user_release(char *dsname, char *snapname, char *htag,
3925     boolean_t recursive)
3926 {
3927 	struct dsl_ds_holdarg *ha;
3928 	dsl_sync_task_t *dst;
3929 	spa_t *spa;
3930 	int error;
3931 
3932 top:
3933 	ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3934 
3935 	(void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3936 
3937 	error = spa_open(dsname, &spa, FTAG);
3938 	if (error) {
3939 		kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3940 		return (error);
3941 	}
3942 
3943 	ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3944 	ha->htag = htag;
3945 	ha->snapname = snapname;
3946 	ha->recursive = recursive;
3947 	if (recursive) {
3948 		error = dmu_objset_find(dsname, dsl_dataset_user_release_one,
3949 		    ha, DS_FIND_CHILDREN);
3950 	} else {
3951 		error = dsl_dataset_user_release_one(dsname, ha);
3952 	}
3953 	if (error == 0)
3954 		error = dsl_sync_task_group_wait(ha->dstg);
3955 
3956 	for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3957 	    dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3958 		struct dsl_ds_releasearg *ra = dst->dst_arg1;
3959 		dsl_dataset_t *ds = ra->ds;
3960 
3961 		if (dst->dst_err)
3962 			dsl_dataset_name(ds, ha->failed);
3963 
3964 		if (ra->own)
3965 			dsl_dataset_disown(ds, ha->dstg);
3966 		else
3967 			dsl_dataset_rele(ds, ha->dstg);
3968 
3969 		kmem_free(ra, sizeof (struct dsl_ds_releasearg));
3970 	}
3971 
3972 	if (error == 0 && recursive && !ha->gotone)
3973 		error = ENOENT;
3974 
3975 	if (error && error != EBUSY)
3976 		(void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3977 
3978 	dsl_sync_task_group_destroy(ha->dstg);
3979 	kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3980 	spa_close(spa, FTAG);
3981 
3982 	/*
3983 	 * We can get EBUSY if we were racing with deferred destroy and
3984 	 * dsl_dataset_user_release_check() hadn't done the necessary
3985 	 * open context setup.  We can also get EBUSY if we're racing
3986 	 * with destroy and that thread is the ds_owner.  Either way
3987 	 * the busy condition should be transient, and we should retry
3988 	 * the release operation.
3989 	 */
3990 	if (error == EBUSY)
3991 		goto top;
3992 
3993 	return (error);
3994 }
3995 
3996 /*
3997  * Called at spa_load time (with retry == B_FALSE) to release a stale
3998  * temporary user hold. Also called by the onexit code (with retry == B_TRUE).
3999  */
4000 int
4001 dsl_dataset_user_release_tmp(dsl_pool_t *dp, uint64_t dsobj, char *htag,
4002     boolean_t retry)
4003 {
4004 	dsl_dataset_t *ds;
4005 	char *snap;
4006 	char *name;
4007 	int namelen;
4008 	int error;
4009 
4010 	do {
4011 		rw_enter(&dp->dp_config_rwlock, RW_READER);
4012 		error = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
4013 		rw_exit(&dp->dp_config_rwlock);
4014 		if (error)
4015 			return (error);
4016 		namelen = dsl_dataset_namelen(ds)+1;
4017 		name = kmem_alloc(namelen, KM_SLEEP);
4018 		dsl_dataset_name(ds, name);
4019 		dsl_dataset_rele(ds, FTAG);
4020 
4021 		snap = strchr(name, '@');
4022 		*snap = '\0';
4023 		++snap;
4024 		error = dsl_dataset_user_release(name, snap, htag, B_FALSE);
4025 		kmem_free(name, namelen);
4026 
4027 		/*
4028 		 * The object can't have been destroyed because we have a hold,
4029 		 * but it might have been renamed, resulting in ENOENT.  Retry
4030 		 * if we've been requested to do so.
4031 		 *
4032 		 * It would be nice if we could use the dsobj all the way
4033 		 * through and avoid ENOENT entirely.  But we might need to
4034 		 * unmount the snapshot, and there's currently no way to lookup
4035 		 * a vfsp using a ZFS object id.
4036 		 */
4037 	} while ((error == ENOENT) && retry);
4038 
4039 	return (error);
4040 }
4041 
4042 int
4043 dsl_dataset_get_holds(const char *dsname, nvlist_t **nvp)
4044 {
4045 	dsl_dataset_t *ds;
4046 	int err;
4047 
4048 	err = dsl_dataset_hold(dsname, FTAG, &ds);
4049 	if (err)
4050 		return (err);
4051 
4052 	VERIFY(0 == nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP));
4053 	if (ds->ds_phys->ds_userrefs_obj != 0) {
4054 		zap_attribute_t *za;
4055 		zap_cursor_t zc;
4056 
4057 		za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
4058 		for (zap_cursor_init(&zc, ds->ds_dir->dd_pool->dp_meta_objset,
4059 		    ds->ds_phys->ds_userrefs_obj);
4060 		    zap_cursor_retrieve(&zc, za) == 0;
4061 		    zap_cursor_advance(&zc)) {
4062 			VERIFY(0 == nvlist_add_uint64(*nvp, za->za_name,
4063 			    za->za_first_integer));
4064 		}
4065 		zap_cursor_fini(&zc);
4066 		kmem_free(za, sizeof (zap_attribute_t));
4067 	}
4068 	dsl_dataset_rele(ds, FTAG);
4069 	return (0);
4070 }
4071 
4072 /*
4073  * Note, this function is used as the callback for dmu_objset_find().  We
4074  * always return 0 so that we will continue to find and process
4075  * inconsistent datasets, even if we encounter an error trying to
4076  * process one of them.
4077  */
4078 /* ARGSUSED */
4079 int
4080 dsl_destroy_inconsistent(const char *dsname, void *arg)
4081 {
4082 	dsl_dataset_t *ds;
4083 
4084 	if (dsl_dataset_own(dsname, B_TRUE, FTAG, &ds) == 0) {
4085 		if (DS_IS_INCONSISTENT(ds))
4086 			(void) dsl_dataset_destroy(ds, FTAG, B_FALSE);
4087 		else
4088 			dsl_dataset_disown(ds, FTAG);
4089 	}
4090 	return (0);
4091 }
4092 
4093 /*
4094  * Return (in *usedp) the amount of space written in new that is not
4095  * present in oldsnap.  New may be a snapshot or the head.  Old must be
4096  * a snapshot before new, in new's filesystem (or its origin).  If not then
4097  * fail and return EINVAL.
4098  *
4099  * The written space is calculated by considering two components:  First, we
4100  * ignore any freed space, and calculate the written as new's used space
4101  * minus old's used space.  Next, we add in the amount of space that was freed
4102  * between the two snapshots, thus reducing new's used space relative to old's.
4103  * Specifically, this is the space that was born before old->ds_creation_txg,
4104  * and freed before new (ie. on new's deadlist or a previous deadlist).
4105  *
4106  * space freed                         [---------------------]
4107  * snapshots                       ---O-------O--------O-------O------
4108  *                                         oldsnap            new
4109  */
4110 int
4111 dsl_dataset_space_written(dsl_dataset_t *oldsnap, dsl_dataset_t *new,
4112     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
4113 {
4114 	int err = 0;
4115 	uint64_t snapobj;
4116 	dsl_pool_t *dp = new->ds_dir->dd_pool;
4117 
4118 	*usedp = 0;
4119 	*usedp += new->ds_phys->ds_used_bytes;
4120 	*usedp -= oldsnap->ds_phys->ds_used_bytes;
4121 
4122 	*compp = 0;
4123 	*compp += new->ds_phys->ds_compressed_bytes;
4124 	*compp -= oldsnap->ds_phys->ds_compressed_bytes;
4125 
4126 	*uncompp = 0;
4127 	*uncompp += new->ds_phys->ds_uncompressed_bytes;
4128 	*uncompp -= oldsnap->ds_phys->ds_uncompressed_bytes;
4129 
4130 	rw_enter(&dp->dp_config_rwlock, RW_READER);
4131 	snapobj = new->ds_object;
4132 	while (snapobj != oldsnap->ds_object) {
4133 		dsl_dataset_t *snap;
4134 		uint64_t used, comp, uncomp;
4135 
4136 		err = dsl_dataset_hold_obj(dp, snapobj, FTAG, &snap);
4137 		if (err != 0)
4138 			break;
4139 
4140 		if (snap->ds_phys->ds_prev_snap_txg ==
4141 		    oldsnap->ds_phys->ds_creation_txg) {
4142 			/*
4143 			 * The blocks in the deadlist can not be born after
4144 			 * ds_prev_snap_txg, so get the whole deadlist space,
4145 			 * which is more efficient (especially for old-format
4146 			 * deadlists).  Unfortunately the deadlist code
4147 			 * doesn't have enough information to make this
4148 			 * optimization itself.
4149 			 */
4150 			dsl_deadlist_space(&snap->ds_deadlist,
4151 			    &used, &comp, &uncomp);
4152 		} else {
4153 			dsl_deadlist_space_range(&snap->ds_deadlist,
4154 			    0, oldsnap->ds_phys->ds_creation_txg,
4155 			    &used, &comp, &uncomp);
4156 		}
4157 		*usedp += used;
4158 		*compp += comp;
4159 		*uncompp += uncomp;
4160 
4161 		/*
4162 		 * If we get to the beginning of the chain of snapshots
4163 		 * (ds_prev_snap_obj == 0) before oldsnap, then oldsnap
4164 		 * was not a snapshot of/before new.
4165 		 */
4166 		snapobj = snap->ds_phys->ds_prev_snap_obj;
4167 		dsl_dataset_rele(snap, FTAG);
4168 		if (snapobj == 0) {
4169 			err = EINVAL;
4170 			break;
4171 		}
4172 
4173 	}
4174 	rw_exit(&dp->dp_config_rwlock);
4175 	return (err);
4176 }
4177 
4178 /*
4179  * Return (in *usedp) the amount of space that will be reclaimed if firstsnap,
4180  * lastsnap, and all snapshots in between are deleted.
4181  *
4182  * blocks that would be freed            [---------------------------]
4183  * snapshots                       ---O-------O--------O-------O--------O
4184  *                                        firstsnap        lastsnap
4185  *
4186  * This is the set of blocks that were born after the snap before firstsnap,
4187  * (birth > firstsnap->prev_snap_txg) and died before the snap after the
4188  * last snap (ie, is on lastsnap->ds_next->ds_deadlist or an earlier deadlist).
4189  * We calculate this by iterating over the relevant deadlists (from the snap
4190  * after lastsnap, backward to the snap after firstsnap), summing up the
4191  * space on the deadlist that was born after the snap before firstsnap.
4192  */
4193 int
4194 dsl_dataset_space_wouldfree(dsl_dataset_t *firstsnap,
4195     dsl_dataset_t *lastsnap,
4196     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp)
4197 {
4198 	int err = 0;
4199 	uint64_t snapobj;
4200 	dsl_pool_t *dp = firstsnap->ds_dir->dd_pool;
4201 
4202 	ASSERT(dsl_dataset_is_snapshot(firstsnap));
4203 	ASSERT(dsl_dataset_is_snapshot(lastsnap));
4204 
4205 	/*
4206 	 * Check that the snapshots are in the same dsl_dir, and firstsnap
4207 	 * is before lastsnap.
4208 	 */
4209 	if (firstsnap->ds_dir != lastsnap->ds_dir ||
4210 	    firstsnap->ds_phys->ds_creation_txg >
4211 	    lastsnap->ds_phys->ds_creation_txg)
4212 		return (EINVAL);
4213 
4214 	*usedp = *compp = *uncompp = 0;
4215 
4216 	rw_enter(&dp->dp_config_rwlock, RW_READER);
4217 	snapobj = lastsnap->ds_phys->ds_next_snap_obj;
4218 	while (snapobj != firstsnap->ds_object) {
4219 		dsl_dataset_t *ds;
4220 		uint64_t used, comp, uncomp;
4221 
4222 		err = dsl_dataset_hold_obj(dp, snapobj, FTAG, &ds);
4223 		if (err != 0)
4224 			break;
4225 
4226 		dsl_deadlist_space_range(&ds->ds_deadlist,
4227 		    firstsnap->ds_phys->ds_prev_snap_txg, UINT64_MAX,
4228 		    &used, &comp, &uncomp);
4229 		*usedp += used;
4230 		*compp += comp;
4231 		*uncompp += uncomp;
4232 
4233 		snapobj = ds->ds_phys->ds_prev_snap_obj;
4234 		ASSERT3U(snapobj, !=, 0);
4235 		dsl_dataset_rele(ds, FTAG);
4236 	}
4237 	rw_exit(&dp->dp_config_rwlock);
4238 	return (err);
4239 }
4240