xref: /titanic_41/usr/src/uts/common/fs/zfs/dsl_dataset.c (revision 6a634c9dca3093f3922e4b7ab826d7bdf17bf78e)
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  */
24 
25 #include <sys/dmu_objset.h>
26 #include <sys/dsl_dataset.h>
27 #include <sys/dsl_dir.h>
28 #include <sys/dsl_prop.h>
29 #include <sys/dsl_synctask.h>
30 #include <sys/dmu_traverse.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/arc.h>
33 #include <sys/zio.h>
34 #include <sys/zap.h>
35 #include <sys/unique.h>
36 #include <sys/zfs_context.h>
37 #include <sys/zfs_ioctl.h>
38 #include <sys/spa.h>
39 #include <sys/zfs_znode.h>
40 #include <sys/zfs_onexit.h>
41 #include <sys/zvol.h>
42 #include <sys/dsl_scan.h>
43 #include <sys/dsl_deadlist.h>
44 
45 static char *dsl_reaper = "the grim reaper";
46 
47 static dsl_checkfunc_t dsl_dataset_destroy_begin_check;
48 static dsl_syncfunc_t dsl_dataset_destroy_begin_sync;
49 static dsl_syncfunc_t dsl_dataset_set_reservation_sync;
50 
51 #define	SWITCH64(x, y) \
52 	{ \
53 		uint64_t __tmp = (x); \
54 		(x) = (y); \
55 		(y) = __tmp; \
56 	}
57 
58 #define	DS_REF_MAX	(1ULL << 62)
59 
60 #define	DSL_DEADLIST_BLOCKSIZE	SPA_MAXBLOCKSIZE
61 
62 #define	DSL_DATASET_IS_DESTROYED(ds)	((ds)->ds_owner == dsl_reaper)
63 
64 
65 /*
66  * Figure out how much of this delta should be propogated to the dsl_dir
67  * layer.  If there's a refreservation, that space has already been
68  * partially accounted for in our ancestors.
69  */
70 static int64_t
71 parent_delta(dsl_dataset_t *ds, int64_t delta)
72 {
73 	uint64_t old_bytes, new_bytes;
74 
75 	if (ds->ds_reserved == 0)
76 		return (delta);
77 
78 	old_bytes = MAX(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
79 	new_bytes = MAX(ds->ds_phys->ds_unique_bytes + delta, ds->ds_reserved);
80 
81 	ASSERT3U(ABS((int64_t)(new_bytes - old_bytes)), <=, ABS(delta));
82 	return (new_bytes - old_bytes);
83 }
84 
85 void
86 dsl_dataset_block_born(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx)
87 {
88 	int used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
89 	int compressed = BP_GET_PSIZE(bp);
90 	int uncompressed = BP_GET_UCSIZE(bp);
91 	int64_t delta;
92 
93 	dprintf_bp(bp, "ds=%p", ds);
94 
95 	ASSERT(dmu_tx_is_syncing(tx));
96 	/* It could have been compressed away to nothing */
97 	if (BP_IS_HOLE(bp))
98 		return;
99 	ASSERT(BP_GET_TYPE(bp) != DMU_OT_NONE);
100 	ASSERT3U(BP_GET_TYPE(bp), <, DMU_OT_NUMTYPES);
101 	if (ds == NULL) {
102 		/*
103 		 * Account for the meta-objset space in its placeholder
104 		 * dsl_dir.
105 		 */
106 		ASSERT3U(compressed, ==, uncompressed); /* it's all metadata */
107 		dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
108 		    used, compressed, uncompressed, tx);
109 		dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
110 		return;
111 	}
112 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
113 
114 	mutex_enter(&ds->ds_dir->dd_lock);
115 	mutex_enter(&ds->ds_lock);
116 	delta = parent_delta(ds, used);
117 	ds->ds_phys->ds_used_bytes += used;
118 	ds->ds_phys->ds_compressed_bytes += compressed;
119 	ds->ds_phys->ds_uncompressed_bytes += uncompressed;
120 	ds->ds_phys->ds_unique_bytes += used;
121 	mutex_exit(&ds->ds_lock);
122 	dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD, delta,
123 	    compressed, uncompressed, tx);
124 	dsl_dir_transfer_space(ds->ds_dir, used - delta,
125 	    DD_USED_REFRSRV, DD_USED_HEAD, tx);
126 	mutex_exit(&ds->ds_dir->dd_lock);
127 }
128 
129 int
130 dsl_dataset_block_kill(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx,
131     boolean_t async)
132 {
133 	if (BP_IS_HOLE(bp))
134 		return (0);
135 
136 	ASSERT(dmu_tx_is_syncing(tx));
137 	ASSERT(bp->blk_birth <= tx->tx_txg);
138 
139 	int used = bp_get_dsize_sync(tx->tx_pool->dp_spa, bp);
140 	int compressed = BP_GET_PSIZE(bp);
141 	int uncompressed = BP_GET_UCSIZE(bp);
142 
143 	ASSERT(used > 0);
144 	if (ds == NULL) {
145 		/*
146 		 * Account for the meta-objset space in its placeholder
147 		 * dataset.
148 		 */
149 		dsl_free(tx->tx_pool, tx->tx_txg, bp);
150 
151 		dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
152 		    -used, -compressed, -uncompressed, tx);
153 		dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
154 		return (used);
155 	}
156 	ASSERT3P(tx->tx_pool, ==, ds->ds_dir->dd_pool);
157 
158 	ASSERT(!dsl_dataset_is_snapshot(ds));
159 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
160 
161 	if (bp->blk_birth > ds->ds_phys->ds_prev_snap_txg) {
162 		int64_t delta;
163 
164 		dprintf_bp(bp, "freeing ds=%llu", ds->ds_object);
165 		dsl_free(tx->tx_pool, tx->tx_txg, bp);
166 
167 		mutex_enter(&ds->ds_dir->dd_lock);
168 		mutex_enter(&ds->ds_lock);
169 		ASSERT(ds->ds_phys->ds_unique_bytes >= used ||
170 		    !DS_UNIQUE_IS_ACCURATE(ds));
171 		delta = parent_delta(ds, -used);
172 		ds->ds_phys->ds_unique_bytes -= used;
173 		mutex_exit(&ds->ds_lock);
174 		dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
175 		    delta, -compressed, -uncompressed, tx);
176 		dsl_dir_transfer_space(ds->ds_dir, -used - delta,
177 		    DD_USED_REFRSRV, DD_USED_HEAD, tx);
178 		mutex_exit(&ds->ds_dir->dd_lock);
179 	} else {
180 		dprintf_bp(bp, "putting on dead list: %s", "");
181 		if (async) {
182 			/*
183 			 * We are here as part of zio's write done callback,
184 			 * which means we're a zio interrupt thread.  We can't
185 			 * call dsl_deadlist_insert() now because it may block
186 			 * waiting for I/O.  Instead, put bp on the deferred
187 			 * queue and let dsl_pool_sync() finish the job.
188 			 */
189 			bplist_append(&ds->ds_pending_deadlist, bp);
190 		} else {
191 			dsl_deadlist_insert(&ds->ds_deadlist, bp, tx);
192 		}
193 		ASSERT3U(ds->ds_prev->ds_object, ==,
194 		    ds->ds_phys->ds_prev_snap_obj);
195 		ASSERT(ds->ds_prev->ds_phys->ds_num_children > 0);
196 		/* if (bp->blk_birth > prev prev snap txg) prev unique += bs */
197 		if (ds->ds_prev->ds_phys->ds_next_snap_obj ==
198 		    ds->ds_object && bp->blk_birth >
199 		    ds->ds_prev->ds_phys->ds_prev_snap_txg) {
200 			dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
201 			mutex_enter(&ds->ds_prev->ds_lock);
202 			ds->ds_prev->ds_phys->ds_unique_bytes += used;
203 			mutex_exit(&ds->ds_prev->ds_lock);
204 		}
205 		if (bp->blk_birth > ds->ds_dir->dd_origin_txg) {
206 			dsl_dir_transfer_space(ds->ds_dir, used,
207 			    DD_USED_HEAD, DD_USED_SNAP, tx);
208 		}
209 	}
210 	mutex_enter(&ds->ds_lock);
211 	ASSERT3U(ds->ds_phys->ds_used_bytes, >=, used);
212 	ds->ds_phys->ds_used_bytes -= used;
213 	ASSERT3U(ds->ds_phys->ds_compressed_bytes, >=, compressed);
214 	ds->ds_phys->ds_compressed_bytes -= compressed;
215 	ASSERT3U(ds->ds_phys->ds_uncompressed_bytes, >=, uncompressed);
216 	ds->ds_phys->ds_uncompressed_bytes -= uncompressed;
217 	mutex_exit(&ds->ds_lock);
218 
219 	return (used);
220 }
221 
222 uint64_t
223 dsl_dataset_prev_snap_txg(dsl_dataset_t *ds)
224 {
225 	uint64_t trysnap = 0;
226 
227 	if (ds == NULL)
228 		return (0);
229 	/*
230 	 * The snapshot creation could fail, but that would cause an
231 	 * incorrect FALSE return, which would only result in an
232 	 * overestimation of the amount of space that an operation would
233 	 * consume, which is OK.
234 	 *
235 	 * There's also a small window where we could miss a pending
236 	 * snapshot, because we could set the sync task in the quiescing
237 	 * phase.  So this should only be used as a guess.
238 	 */
239 	if (ds->ds_trysnap_txg >
240 	    spa_last_synced_txg(ds->ds_dir->dd_pool->dp_spa))
241 		trysnap = ds->ds_trysnap_txg;
242 	return (MAX(ds->ds_phys->ds_prev_snap_txg, trysnap));
243 }
244 
245 boolean_t
246 dsl_dataset_block_freeable(dsl_dataset_t *ds, const blkptr_t *bp,
247     uint64_t blk_birth)
248 {
249 	if (blk_birth <= dsl_dataset_prev_snap_txg(ds))
250 		return (B_FALSE);
251 
252 	ddt_prefetch(dsl_dataset_get_spa(ds), bp);
253 
254 	return (B_TRUE);
255 }
256 
257 /* ARGSUSED */
258 static void
259 dsl_dataset_evict(dmu_buf_t *db, void *dsv)
260 {
261 	dsl_dataset_t *ds = dsv;
262 
263 	ASSERT(ds->ds_owner == NULL || DSL_DATASET_IS_DESTROYED(ds));
264 
265 	unique_remove(ds->ds_fsid_guid);
266 
267 	if (ds->ds_objset != NULL)
268 		dmu_objset_evict(ds->ds_objset);
269 
270 	if (ds->ds_prev) {
271 		dsl_dataset_drop_ref(ds->ds_prev, ds);
272 		ds->ds_prev = NULL;
273 	}
274 
275 	bplist_destroy(&ds->ds_pending_deadlist);
276 	if (db != NULL) {
277 		dsl_deadlist_close(&ds->ds_deadlist);
278 	} else {
279 		ASSERT(ds->ds_deadlist.dl_dbuf == NULL);
280 		ASSERT(!ds->ds_deadlist.dl_oldfmt);
281 	}
282 	if (ds->ds_dir)
283 		dsl_dir_close(ds->ds_dir, ds);
284 
285 	ASSERT(!list_link_active(&ds->ds_synced_link));
286 
287 	mutex_destroy(&ds->ds_lock);
288 	mutex_destroy(&ds->ds_recvlock);
289 	mutex_destroy(&ds->ds_opening_lock);
290 	rw_destroy(&ds->ds_rwlock);
291 	cv_destroy(&ds->ds_exclusive_cv);
292 
293 	kmem_free(ds, sizeof (dsl_dataset_t));
294 }
295 
296 static int
297 dsl_dataset_get_snapname(dsl_dataset_t *ds)
298 {
299 	dsl_dataset_phys_t *headphys;
300 	int err;
301 	dmu_buf_t *headdbuf;
302 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
303 	objset_t *mos = dp->dp_meta_objset;
304 
305 	if (ds->ds_snapname[0])
306 		return (0);
307 	if (ds->ds_phys->ds_next_snap_obj == 0)
308 		return (0);
309 
310 	err = dmu_bonus_hold(mos, ds->ds_dir->dd_phys->dd_head_dataset_obj,
311 	    FTAG, &headdbuf);
312 	if (err)
313 		return (err);
314 	headphys = headdbuf->db_data;
315 	err = zap_value_search(dp->dp_meta_objset,
316 	    headphys->ds_snapnames_zapobj, ds->ds_object, 0, ds->ds_snapname);
317 	dmu_buf_rele(headdbuf, FTAG);
318 	return (err);
319 }
320 
321 static int
322 dsl_dataset_snap_lookup(dsl_dataset_t *ds, const char *name, uint64_t *value)
323 {
324 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
325 	uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
326 	matchtype_t mt;
327 	int err;
328 
329 	if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
330 		mt = MT_FIRST;
331 	else
332 		mt = MT_EXACT;
333 
334 	err = zap_lookup_norm(mos, snapobj, name, 8, 1,
335 	    value, mt, NULL, 0, NULL);
336 	if (err == ENOTSUP && mt == MT_FIRST)
337 		err = zap_lookup(mos, snapobj, name, 8, 1, value);
338 	return (err);
339 }
340 
341 static int
342 dsl_dataset_snap_remove(dsl_dataset_t *ds, char *name, dmu_tx_t *tx)
343 {
344 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
345 	uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
346 	matchtype_t mt;
347 	int err;
348 
349 	dsl_dir_snap_cmtime_update(ds->ds_dir);
350 
351 	if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
352 		mt = MT_FIRST;
353 	else
354 		mt = MT_EXACT;
355 
356 	err = zap_remove_norm(mos, snapobj, name, mt, tx);
357 	if (err == ENOTSUP && mt == MT_FIRST)
358 		err = zap_remove(mos, snapobj, name, tx);
359 	return (err);
360 }
361 
362 static int
363 dsl_dataset_get_ref(dsl_pool_t *dp, uint64_t dsobj, void *tag,
364     dsl_dataset_t **dsp)
365 {
366 	objset_t *mos = dp->dp_meta_objset;
367 	dmu_buf_t *dbuf;
368 	dsl_dataset_t *ds;
369 	int err;
370 	dmu_object_info_t doi;
371 
372 	ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
373 	    dsl_pool_sync_context(dp));
374 
375 	err = dmu_bonus_hold(mos, dsobj, tag, &dbuf);
376 	if (err)
377 		return (err);
378 
379 	/* Make sure dsobj has the correct object type. */
380 	dmu_object_info_from_db(dbuf, &doi);
381 	if (doi.doi_type != DMU_OT_DSL_DATASET)
382 		return (EINVAL);
383 
384 	ds = dmu_buf_get_user(dbuf);
385 	if (ds == NULL) {
386 		dsl_dataset_t *winner;
387 
388 		ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_SLEEP);
389 		ds->ds_dbuf = dbuf;
390 		ds->ds_object = dsobj;
391 		ds->ds_phys = dbuf->db_data;
392 
393 		mutex_init(&ds->ds_lock, NULL, MUTEX_DEFAULT, NULL);
394 		mutex_init(&ds->ds_recvlock, NULL, MUTEX_DEFAULT, NULL);
395 		mutex_init(&ds->ds_opening_lock, NULL, MUTEX_DEFAULT, NULL);
396 		rw_init(&ds->ds_rwlock, 0, 0, 0);
397 		cv_init(&ds->ds_exclusive_cv, NULL, CV_DEFAULT, NULL);
398 
399 		bplist_create(&ds->ds_pending_deadlist);
400 		dsl_deadlist_open(&ds->ds_deadlist,
401 		    mos, ds->ds_phys->ds_deadlist_obj);
402 
403 		if (err == 0) {
404 			err = dsl_dir_open_obj(dp,
405 			    ds->ds_phys->ds_dir_obj, NULL, ds, &ds->ds_dir);
406 		}
407 		if (err) {
408 			mutex_destroy(&ds->ds_lock);
409 			mutex_destroy(&ds->ds_recvlock);
410 			mutex_destroy(&ds->ds_opening_lock);
411 			rw_destroy(&ds->ds_rwlock);
412 			cv_destroy(&ds->ds_exclusive_cv);
413 			bplist_destroy(&ds->ds_pending_deadlist);
414 			dsl_deadlist_close(&ds->ds_deadlist);
415 			kmem_free(ds, sizeof (dsl_dataset_t));
416 			dmu_buf_rele(dbuf, tag);
417 			return (err);
418 		}
419 
420 		if (!dsl_dataset_is_snapshot(ds)) {
421 			ds->ds_snapname[0] = '\0';
422 			if (ds->ds_phys->ds_prev_snap_obj) {
423 				err = dsl_dataset_get_ref(dp,
424 				    ds->ds_phys->ds_prev_snap_obj,
425 				    ds, &ds->ds_prev);
426 			}
427 		} else {
428 			if (zfs_flags & ZFS_DEBUG_SNAPNAMES)
429 				err = dsl_dataset_get_snapname(ds);
430 			if (err == 0 && ds->ds_phys->ds_userrefs_obj != 0) {
431 				err = zap_count(
432 				    ds->ds_dir->dd_pool->dp_meta_objset,
433 				    ds->ds_phys->ds_userrefs_obj,
434 				    &ds->ds_userrefs);
435 			}
436 		}
437 
438 		if (err == 0 && !dsl_dataset_is_snapshot(ds)) {
439 			/*
440 			 * In sync context, we're called with either no lock
441 			 * or with the write lock.  If we're not syncing,
442 			 * we're always called with the read lock held.
443 			 */
444 			boolean_t need_lock =
445 			    !RW_WRITE_HELD(&dp->dp_config_rwlock) &&
446 			    dsl_pool_sync_context(dp);
447 
448 			if (need_lock)
449 				rw_enter(&dp->dp_config_rwlock, RW_READER);
450 
451 			err = dsl_prop_get_ds(ds,
452 			    "refreservation", sizeof (uint64_t), 1,
453 			    &ds->ds_reserved, NULL);
454 			if (err == 0) {
455 				err = dsl_prop_get_ds(ds,
456 				    "refquota", sizeof (uint64_t), 1,
457 				    &ds->ds_quota, NULL);
458 			}
459 
460 			if (need_lock)
461 				rw_exit(&dp->dp_config_rwlock);
462 		} else {
463 			ds->ds_reserved = ds->ds_quota = 0;
464 		}
465 
466 		if (err == 0) {
467 			winner = dmu_buf_set_user_ie(dbuf, ds, &ds->ds_phys,
468 			    dsl_dataset_evict);
469 		}
470 		if (err || winner) {
471 			bplist_destroy(&ds->ds_pending_deadlist);
472 			dsl_deadlist_close(&ds->ds_deadlist);
473 			if (ds->ds_prev)
474 				dsl_dataset_drop_ref(ds->ds_prev, ds);
475 			dsl_dir_close(ds->ds_dir, ds);
476 			mutex_destroy(&ds->ds_lock);
477 			mutex_destroy(&ds->ds_recvlock);
478 			mutex_destroy(&ds->ds_opening_lock);
479 			rw_destroy(&ds->ds_rwlock);
480 			cv_destroy(&ds->ds_exclusive_cv);
481 			kmem_free(ds, sizeof (dsl_dataset_t));
482 			if (err) {
483 				dmu_buf_rele(dbuf, tag);
484 				return (err);
485 			}
486 			ds = winner;
487 		} else {
488 			ds->ds_fsid_guid =
489 			    unique_insert(ds->ds_phys->ds_fsid_guid);
490 		}
491 	}
492 	ASSERT3P(ds->ds_dbuf, ==, dbuf);
493 	ASSERT3P(ds->ds_phys, ==, dbuf->db_data);
494 	ASSERT(ds->ds_phys->ds_prev_snap_obj != 0 ||
495 	    spa_version(dp->dp_spa) < SPA_VERSION_ORIGIN ||
496 	    dp->dp_origin_snap == NULL || ds == dp->dp_origin_snap);
497 	mutex_enter(&ds->ds_lock);
498 	if (!dsl_pool_sync_context(dp) && DSL_DATASET_IS_DESTROYED(ds)) {
499 		mutex_exit(&ds->ds_lock);
500 		dmu_buf_rele(ds->ds_dbuf, tag);
501 		return (ENOENT);
502 	}
503 	mutex_exit(&ds->ds_lock);
504 	*dsp = ds;
505 	return (0);
506 }
507 
508 static int
509 dsl_dataset_hold_ref(dsl_dataset_t *ds, void *tag)
510 {
511 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
512 
513 	/*
514 	 * In syncing context we don't want the rwlock lock: there
515 	 * may be an existing writer waiting for sync phase to
516 	 * finish.  We don't need to worry about such writers, since
517 	 * sync phase is single-threaded, so the writer can't be
518 	 * doing anything while we are active.
519 	 */
520 	if (dsl_pool_sync_context(dp)) {
521 		ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
522 		return (0);
523 	}
524 
525 	/*
526 	 * Normal users will hold the ds_rwlock as a READER until they
527 	 * are finished (i.e., call dsl_dataset_rele()).  "Owners" will
528 	 * drop their READER lock after they set the ds_owner field.
529 	 *
530 	 * If the dataset is being destroyed, the destroy thread will
531 	 * obtain a WRITER lock for exclusive access after it's done its
532 	 * open-context work and then change the ds_owner to
533 	 * dsl_reaper once destruction is assured.  So threads
534 	 * may block here temporarily, until the "destructability" of
535 	 * the dataset is determined.
536 	 */
537 	ASSERT(!RW_WRITE_HELD(&dp->dp_config_rwlock));
538 	mutex_enter(&ds->ds_lock);
539 	while (!rw_tryenter(&ds->ds_rwlock, RW_READER)) {
540 		rw_exit(&dp->dp_config_rwlock);
541 		cv_wait(&ds->ds_exclusive_cv, &ds->ds_lock);
542 		if (DSL_DATASET_IS_DESTROYED(ds)) {
543 			mutex_exit(&ds->ds_lock);
544 			dsl_dataset_drop_ref(ds, tag);
545 			rw_enter(&dp->dp_config_rwlock, RW_READER);
546 			return (ENOENT);
547 		}
548 		/*
549 		 * The dp_config_rwlock lives above the ds_lock. And
550 		 * we need to check DSL_DATASET_IS_DESTROYED() while
551 		 * holding the ds_lock, so we have to drop and reacquire
552 		 * the ds_lock here.
553 		 */
554 		mutex_exit(&ds->ds_lock);
555 		rw_enter(&dp->dp_config_rwlock, RW_READER);
556 		mutex_enter(&ds->ds_lock);
557 	}
558 	mutex_exit(&ds->ds_lock);
559 	return (0);
560 }
561 
562 int
563 dsl_dataset_hold_obj(dsl_pool_t *dp, uint64_t dsobj, void *tag,
564     dsl_dataset_t **dsp)
565 {
566 	int err = dsl_dataset_get_ref(dp, dsobj, tag, dsp);
567 
568 	if (err)
569 		return (err);
570 	return (dsl_dataset_hold_ref(*dsp, tag));
571 }
572 
573 int
574 dsl_dataset_own_obj(dsl_pool_t *dp, uint64_t dsobj, boolean_t inconsistentok,
575     void *tag, dsl_dataset_t **dsp)
576 {
577 	int err = dsl_dataset_hold_obj(dp, dsobj, tag, dsp);
578 	if (err)
579 		return (err);
580 	if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
581 		dsl_dataset_rele(*dsp, tag);
582 		*dsp = NULL;
583 		return (EBUSY);
584 	}
585 	return (0);
586 }
587 
588 int
589 dsl_dataset_hold(const char *name, void *tag, dsl_dataset_t **dsp)
590 {
591 	dsl_dir_t *dd;
592 	dsl_pool_t *dp;
593 	const char *snapname;
594 	uint64_t obj;
595 	int err = 0;
596 
597 	err = dsl_dir_open_spa(NULL, name, FTAG, &dd, &snapname);
598 	if (err)
599 		return (err);
600 
601 	dp = dd->dd_pool;
602 	obj = dd->dd_phys->dd_head_dataset_obj;
603 	rw_enter(&dp->dp_config_rwlock, RW_READER);
604 	if (obj)
605 		err = dsl_dataset_get_ref(dp, obj, tag, dsp);
606 	else
607 		err = ENOENT;
608 	if (err)
609 		goto out;
610 
611 	err = dsl_dataset_hold_ref(*dsp, tag);
612 
613 	/* we may be looking for a snapshot */
614 	if (err == 0 && snapname != NULL) {
615 		dsl_dataset_t *ds = NULL;
616 
617 		if (*snapname++ != '@') {
618 			dsl_dataset_rele(*dsp, tag);
619 			err = ENOENT;
620 			goto out;
621 		}
622 
623 		dprintf("looking for snapshot '%s'\n", snapname);
624 		err = dsl_dataset_snap_lookup(*dsp, snapname, &obj);
625 		if (err == 0)
626 			err = dsl_dataset_get_ref(dp, obj, tag, &ds);
627 		dsl_dataset_rele(*dsp, tag);
628 
629 		ASSERT3U((err == 0), ==, (ds != NULL));
630 
631 		if (ds) {
632 			mutex_enter(&ds->ds_lock);
633 			if (ds->ds_snapname[0] == 0)
634 				(void) strlcpy(ds->ds_snapname, snapname,
635 				    sizeof (ds->ds_snapname));
636 			mutex_exit(&ds->ds_lock);
637 			err = dsl_dataset_hold_ref(ds, tag);
638 			*dsp = err ? NULL : ds;
639 		}
640 	}
641 out:
642 	rw_exit(&dp->dp_config_rwlock);
643 	dsl_dir_close(dd, FTAG);
644 	return (err);
645 }
646 
647 int
648 dsl_dataset_own(const char *name, boolean_t inconsistentok,
649     void *tag, dsl_dataset_t **dsp)
650 {
651 	int err = dsl_dataset_hold(name, tag, dsp);
652 	if (err)
653 		return (err);
654 	if (!dsl_dataset_tryown(*dsp, inconsistentok, tag)) {
655 		dsl_dataset_rele(*dsp, tag);
656 		return (EBUSY);
657 	}
658 	return (0);
659 }
660 
661 void
662 dsl_dataset_name(dsl_dataset_t *ds, char *name)
663 {
664 	if (ds == NULL) {
665 		(void) strcpy(name, "mos");
666 	} else {
667 		dsl_dir_name(ds->ds_dir, name);
668 		VERIFY(0 == dsl_dataset_get_snapname(ds));
669 		if (ds->ds_snapname[0]) {
670 			(void) strcat(name, "@");
671 			/*
672 			 * We use a "recursive" mutex so that we
673 			 * can call dprintf_ds() with ds_lock held.
674 			 */
675 			if (!MUTEX_HELD(&ds->ds_lock)) {
676 				mutex_enter(&ds->ds_lock);
677 				(void) strcat(name, ds->ds_snapname);
678 				mutex_exit(&ds->ds_lock);
679 			} else {
680 				(void) strcat(name, ds->ds_snapname);
681 			}
682 		}
683 	}
684 }
685 
686 static int
687 dsl_dataset_namelen(dsl_dataset_t *ds)
688 {
689 	int result;
690 
691 	if (ds == NULL) {
692 		result = 3;	/* "mos" */
693 	} else {
694 		result = dsl_dir_namelen(ds->ds_dir);
695 		VERIFY(0 == dsl_dataset_get_snapname(ds));
696 		if (ds->ds_snapname[0]) {
697 			++result;	/* adding one for the @-sign */
698 			if (!MUTEX_HELD(&ds->ds_lock)) {
699 				mutex_enter(&ds->ds_lock);
700 				result += strlen(ds->ds_snapname);
701 				mutex_exit(&ds->ds_lock);
702 			} else {
703 				result += strlen(ds->ds_snapname);
704 			}
705 		}
706 	}
707 
708 	return (result);
709 }
710 
711 void
712 dsl_dataset_drop_ref(dsl_dataset_t *ds, void *tag)
713 {
714 	dmu_buf_rele(ds->ds_dbuf, tag);
715 }
716 
717 void
718 dsl_dataset_rele(dsl_dataset_t *ds, void *tag)
719 {
720 	if (!dsl_pool_sync_context(ds->ds_dir->dd_pool)) {
721 		rw_exit(&ds->ds_rwlock);
722 	}
723 	dsl_dataset_drop_ref(ds, tag);
724 }
725 
726 void
727 dsl_dataset_disown(dsl_dataset_t *ds, void *tag)
728 {
729 	ASSERT((ds->ds_owner == tag && ds->ds_dbuf) ||
730 	    (DSL_DATASET_IS_DESTROYED(ds) && ds->ds_dbuf == NULL));
731 
732 	mutex_enter(&ds->ds_lock);
733 	ds->ds_owner = NULL;
734 	if (RW_WRITE_HELD(&ds->ds_rwlock)) {
735 		rw_exit(&ds->ds_rwlock);
736 		cv_broadcast(&ds->ds_exclusive_cv);
737 	}
738 	mutex_exit(&ds->ds_lock);
739 	if (ds->ds_dbuf)
740 		dsl_dataset_drop_ref(ds, tag);
741 	else
742 		dsl_dataset_evict(NULL, ds);
743 }
744 
745 boolean_t
746 dsl_dataset_tryown(dsl_dataset_t *ds, boolean_t inconsistentok, void *tag)
747 {
748 	boolean_t gotit = FALSE;
749 
750 	mutex_enter(&ds->ds_lock);
751 	if (ds->ds_owner == NULL &&
752 	    (!DS_IS_INCONSISTENT(ds) || inconsistentok)) {
753 		ds->ds_owner = tag;
754 		if (!dsl_pool_sync_context(ds->ds_dir->dd_pool))
755 			rw_exit(&ds->ds_rwlock);
756 		gotit = TRUE;
757 	}
758 	mutex_exit(&ds->ds_lock);
759 	return (gotit);
760 }
761 
762 void
763 dsl_dataset_make_exclusive(dsl_dataset_t *ds, void *owner)
764 {
765 	ASSERT3P(owner, ==, ds->ds_owner);
766 	if (!RW_WRITE_HELD(&ds->ds_rwlock))
767 		rw_enter(&ds->ds_rwlock, RW_WRITER);
768 }
769 
770 uint64_t
771 dsl_dataset_create_sync_dd(dsl_dir_t *dd, dsl_dataset_t *origin,
772     uint64_t flags, dmu_tx_t *tx)
773 {
774 	dsl_pool_t *dp = dd->dd_pool;
775 	dmu_buf_t *dbuf;
776 	dsl_dataset_phys_t *dsphys;
777 	uint64_t dsobj;
778 	objset_t *mos = dp->dp_meta_objset;
779 
780 	if (origin == NULL)
781 		origin = dp->dp_origin_snap;
782 
783 	ASSERT(origin == NULL || origin->ds_dir->dd_pool == dp);
784 	ASSERT(origin == NULL || origin->ds_phys->ds_num_children > 0);
785 	ASSERT(dmu_tx_is_syncing(tx));
786 	ASSERT(dd->dd_phys->dd_head_dataset_obj == 0);
787 
788 	dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
789 	    DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
790 	VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
791 	dmu_buf_will_dirty(dbuf, tx);
792 	dsphys = dbuf->db_data;
793 	bzero(dsphys, sizeof (dsl_dataset_phys_t));
794 	dsphys->ds_dir_obj = dd->dd_object;
795 	dsphys->ds_flags = flags;
796 	dsphys->ds_fsid_guid = unique_create();
797 	(void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
798 	    sizeof (dsphys->ds_guid));
799 	dsphys->ds_snapnames_zapobj =
800 	    zap_create_norm(mos, U8_TEXTPREP_TOUPPER, DMU_OT_DSL_DS_SNAP_MAP,
801 	    DMU_OT_NONE, 0, tx);
802 	dsphys->ds_creation_time = gethrestime_sec();
803 	dsphys->ds_creation_txg = tx->tx_txg == TXG_INITIAL ? 1 : tx->tx_txg;
804 
805 	if (origin == NULL) {
806 		dsphys->ds_deadlist_obj = dsl_deadlist_alloc(mos, tx);
807 	} else {
808 		dsl_dataset_t *ohds;
809 
810 		dsphys->ds_prev_snap_obj = origin->ds_object;
811 		dsphys->ds_prev_snap_txg =
812 		    origin->ds_phys->ds_creation_txg;
813 		dsphys->ds_used_bytes =
814 		    origin->ds_phys->ds_used_bytes;
815 		dsphys->ds_compressed_bytes =
816 		    origin->ds_phys->ds_compressed_bytes;
817 		dsphys->ds_uncompressed_bytes =
818 		    origin->ds_phys->ds_uncompressed_bytes;
819 		dsphys->ds_bp = origin->ds_phys->ds_bp;
820 		dsphys->ds_flags |= origin->ds_phys->ds_flags;
821 
822 		dmu_buf_will_dirty(origin->ds_dbuf, tx);
823 		origin->ds_phys->ds_num_children++;
824 
825 		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
826 		    origin->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ohds));
827 		dsphys->ds_deadlist_obj = dsl_deadlist_clone(&ohds->ds_deadlist,
828 		    dsphys->ds_prev_snap_txg, dsphys->ds_prev_snap_obj, tx);
829 		dsl_dataset_rele(ohds, FTAG);
830 
831 		if (spa_version(dp->dp_spa) >= SPA_VERSION_NEXT_CLONES) {
832 			if (origin->ds_phys->ds_next_clones_obj == 0) {
833 				origin->ds_phys->ds_next_clones_obj =
834 				    zap_create(mos,
835 				    DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
836 			}
837 			VERIFY(0 == zap_add_int(mos,
838 			    origin->ds_phys->ds_next_clones_obj,
839 			    dsobj, tx));
840 		}
841 
842 		dmu_buf_will_dirty(dd->dd_dbuf, tx);
843 		dd->dd_phys->dd_origin_obj = origin->ds_object;
844 		if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
845 			if (origin->ds_dir->dd_phys->dd_clones == 0) {
846 				dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
847 				origin->ds_dir->dd_phys->dd_clones =
848 				    zap_create(mos,
849 				    DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
850 			}
851 			VERIFY3U(0, ==, zap_add_int(mos,
852 			    origin->ds_dir->dd_phys->dd_clones, dsobj, tx));
853 		}
854 	}
855 
856 	if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
857 		dsphys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
858 
859 	dmu_buf_rele(dbuf, FTAG);
860 
861 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
862 	dd->dd_phys->dd_head_dataset_obj = dsobj;
863 
864 	return (dsobj);
865 }
866 
867 uint64_t
868 dsl_dataset_create_sync(dsl_dir_t *pdd, const char *lastname,
869     dsl_dataset_t *origin, uint64_t flags, cred_t *cr, dmu_tx_t *tx)
870 {
871 	dsl_pool_t *dp = pdd->dd_pool;
872 	uint64_t dsobj, ddobj;
873 	dsl_dir_t *dd;
874 
875 	ASSERT(lastname[0] != '@');
876 
877 	ddobj = dsl_dir_create_sync(dp, pdd, lastname, tx);
878 	VERIFY(0 == dsl_dir_open_obj(dp, ddobj, lastname, FTAG, &dd));
879 
880 	dsobj = dsl_dataset_create_sync_dd(dd, origin, flags, tx);
881 
882 	dsl_deleg_set_create_perms(dd, tx, cr);
883 
884 	dsl_dir_close(dd, FTAG);
885 
886 	/*
887 	 * If we are creating a clone, make sure we zero out any stale
888 	 * data from the origin snapshots zil header.
889 	 */
890 	if (origin != NULL) {
891 		dsl_dataset_t *ds;
892 		objset_t *os;
893 
894 		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
895 		VERIFY3U(0, ==, dmu_objset_from_ds(ds, &os));
896 		bzero(&os->os_zil_header, sizeof (os->os_zil_header));
897 		dsl_dataset_dirty(ds, tx);
898 		dsl_dataset_rele(ds, FTAG);
899 	}
900 
901 	return (dsobj);
902 }
903 
904 struct destroyarg {
905 	dsl_sync_task_group_t *dstg;
906 	char *snapname;
907 	char *failed;
908 	boolean_t defer;
909 };
910 
911 static int
912 dsl_snapshot_destroy_one(const char *name, void *arg)
913 {
914 	struct destroyarg *da = arg;
915 	dsl_dataset_t *ds;
916 	int err;
917 	char *dsname;
918 
919 	dsname = kmem_asprintf("%s@%s", name, da->snapname);
920 	err = dsl_dataset_own(dsname, B_TRUE, da->dstg, &ds);
921 	strfree(dsname);
922 	if (err == 0) {
923 		struct dsl_ds_destroyarg *dsda;
924 
925 		dsl_dataset_make_exclusive(ds, da->dstg);
926 		dsda = kmem_zalloc(sizeof (struct dsl_ds_destroyarg), KM_SLEEP);
927 		dsda->ds = ds;
928 		dsda->defer = da->defer;
929 		dsl_sync_task_create(da->dstg, dsl_dataset_destroy_check,
930 		    dsl_dataset_destroy_sync, dsda, da->dstg, 0);
931 	} else if (err == ENOENT) {
932 		err = 0;
933 	} else {
934 		(void) strcpy(da->failed, name);
935 	}
936 	return (err);
937 }
938 
939 /*
940  * Destroy 'snapname' in all descendants of 'fsname'.
941  */
942 #pragma weak dmu_snapshots_destroy = dsl_snapshots_destroy
943 int
944 dsl_snapshots_destroy(char *fsname, char *snapname, boolean_t defer)
945 {
946 	int err;
947 	struct destroyarg da;
948 	dsl_sync_task_t *dst;
949 	spa_t *spa;
950 
951 	err = spa_open(fsname, &spa, FTAG);
952 	if (err)
953 		return (err);
954 	da.dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
955 	da.snapname = snapname;
956 	da.failed = fsname;
957 	da.defer = defer;
958 
959 	err = dmu_objset_find(fsname,
960 	    dsl_snapshot_destroy_one, &da, DS_FIND_CHILDREN);
961 
962 	if (err == 0)
963 		err = dsl_sync_task_group_wait(da.dstg);
964 
965 	for (dst = list_head(&da.dstg->dstg_tasks); dst;
966 	    dst = list_next(&da.dstg->dstg_tasks, dst)) {
967 		struct dsl_ds_destroyarg *dsda = dst->dst_arg1;
968 		dsl_dataset_t *ds = dsda->ds;
969 
970 		/*
971 		 * Return the file system name that triggered the error
972 		 */
973 		if (dst->dst_err) {
974 			dsl_dataset_name(ds, fsname);
975 			*strchr(fsname, '@') = '\0';
976 		}
977 		ASSERT3P(dsda->rm_origin, ==, NULL);
978 		dsl_dataset_disown(ds, da.dstg);
979 		kmem_free(dsda, sizeof (struct dsl_ds_destroyarg));
980 	}
981 
982 	dsl_sync_task_group_destroy(da.dstg);
983 	spa_close(spa, FTAG);
984 	return (err);
985 }
986 
987 static boolean_t
988 dsl_dataset_might_destroy_origin(dsl_dataset_t *ds)
989 {
990 	boolean_t might_destroy = B_FALSE;
991 
992 	mutex_enter(&ds->ds_lock);
993 	if (ds->ds_phys->ds_num_children == 2 && ds->ds_userrefs == 0 &&
994 	    DS_IS_DEFER_DESTROY(ds))
995 		might_destroy = B_TRUE;
996 	mutex_exit(&ds->ds_lock);
997 
998 	return (might_destroy);
999 }
1000 
1001 /*
1002  * If we're removing a clone, and these three conditions are true:
1003  *	1) the clone's origin has no other children
1004  *	2) the clone's origin has no user references
1005  *	3) the clone's origin has been marked for deferred destruction
1006  * Then, prepare to remove the origin as part of this sync task group.
1007  */
1008 static int
1009 dsl_dataset_origin_rm_prep(struct dsl_ds_destroyarg *dsda, void *tag)
1010 {
1011 	dsl_dataset_t *ds = dsda->ds;
1012 	dsl_dataset_t *origin = ds->ds_prev;
1013 
1014 	if (dsl_dataset_might_destroy_origin(origin)) {
1015 		char *name;
1016 		int namelen;
1017 		int error;
1018 
1019 		namelen = dsl_dataset_namelen(origin) + 1;
1020 		name = kmem_alloc(namelen, KM_SLEEP);
1021 		dsl_dataset_name(origin, name);
1022 #ifdef _KERNEL
1023 		error = zfs_unmount_snap(name, NULL);
1024 		if (error) {
1025 			kmem_free(name, namelen);
1026 			return (error);
1027 		}
1028 #endif
1029 		error = dsl_dataset_own(name, B_TRUE, tag, &origin);
1030 		kmem_free(name, namelen);
1031 		if (error)
1032 			return (error);
1033 		dsda->rm_origin = origin;
1034 		dsl_dataset_make_exclusive(origin, tag);
1035 	}
1036 
1037 	return (0);
1038 }
1039 
1040 /*
1041  * ds must be opened as OWNER.  On return (whether successful or not),
1042  * ds will be closed and caller can no longer dereference it.
1043  */
1044 int
1045 dsl_dataset_destroy(dsl_dataset_t *ds, void *tag, boolean_t defer)
1046 {
1047 	int err;
1048 	dsl_sync_task_group_t *dstg;
1049 	objset_t *os;
1050 	dsl_dir_t *dd;
1051 	uint64_t obj;
1052 	struct dsl_ds_destroyarg dsda = { 0 };
1053 	dsl_dataset_t dummy_ds = { 0 };
1054 
1055 	dsda.ds = ds;
1056 
1057 	if (dsl_dataset_is_snapshot(ds)) {
1058 		/* Destroying a snapshot is simpler */
1059 		dsl_dataset_make_exclusive(ds, tag);
1060 
1061 		dsda.defer = defer;
1062 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1063 		    dsl_dataset_destroy_check, dsl_dataset_destroy_sync,
1064 		    &dsda, tag, 0);
1065 		ASSERT3P(dsda.rm_origin, ==, NULL);
1066 		goto out;
1067 	} else if (defer) {
1068 		err = EINVAL;
1069 		goto out;
1070 	}
1071 
1072 	dd = ds->ds_dir;
1073 	dummy_ds.ds_dir = dd;
1074 	dummy_ds.ds_object = ds->ds_object;
1075 
1076 	/*
1077 	 * Check for errors and mark this ds as inconsistent, in
1078 	 * case we crash while freeing the objects.
1079 	 */
1080 	err = dsl_sync_task_do(dd->dd_pool, dsl_dataset_destroy_begin_check,
1081 	    dsl_dataset_destroy_begin_sync, ds, NULL, 0);
1082 	if (err)
1083 		goto out;
1084 
1085 	err = dmu_objset_from_ds(ds, &os);
1086 	if (err)
1087 		goto out;
1088 
1089 	/*
1090 	 * remove the objects in open context, so that we won't
1091 	 * have too much to do in syncing context.
1092 	 */
1093 	for (obj = 0; err == 0; err = dmu_object_next(os, &obj, FALSE,
1094 	    ds->ds_phys->ds_prev_snap_txg)) {
1095 		/*
1096 		 * Ignore errors, if there is not enough disk space
1097 		 * we will deal with it in dsl_dataset_destroy_sync().
1098 		 */
1099 		(void) dmu_free_object(os, obj);
1100 	}
1101 	if (err != ESRCH)
1102 		goto out;
1103 
1104 	/*
1105 	 * Only the ZIL knows how to free log blocks.
1106 	 */
1107 	zil_destroy(dmu_objset_zil(os), B_FALSE);
1108 
1109 	/*
1110 	 * Sync out all in-flight IO.
1111 	 */
1112 	txg_wait_synced(dd->dd_pool, 0);
1113 
1114 	/*
1115 	 * If we managed to free all the objects in open
1116 	 * context, the user space accounting should be zero.
1117 	 */
1118 	if (ds->ds_phys->ds_bp.blk_fill == 0 &&
1119 	    dmu_objset_userused_enabled(os)) {
1120 		uint64_t count;
1121 
1122 		ASSERT(zap_count(os, DMU_USERUSED_OBJECT, &count) != 0 ||
1123 		    count == 0);
1124 		ASSERT(zap_count(os, DMU_GROUPUSED_OBJECT, &count) != 0 ||
1125 		    count == 0);
1126 	}
1127 
1128 	rw_enter(&dd->dd_pool->dp_config_rwlock, RW_READER);
1129 	err = dsl_dir_open_obj(dd->dd_pool, dd->dd_object, NULL, FTAG, &dd);
1130 	rw_exit(&dd->dd_pool->dp_config_rwlock);
1131 
1132 	if (err)
1133 		goto out;
1134 
1135 	/*
1136 	 * Blow away the dsl_dir + head dataset.
1137 	 */
1138 	dsl_dataset_make_exclusive(ds, tag);
1139 	/*
1140 	 * If we're removing a clone, we might also need to remove its
1141 	 * origin.
1142 	 */
1143 	do {
1144 		dsda.need_prep = B_FALSE;
1145 		if (dsl_dir_is_clone(dd)) {
1146 			err = dsl_dataset_origin_rm_prep(&dsda, tag);
1147 			if (err) {
1148 				dsl_dir_close(dd, FTAG);
1149 				goto out;
1150 			}
1151 		}
1152 
1153 		dstg = dsl_sync_task_group_create(ds->ds_dir->dd_pool);
1154 		dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
1155 		    dsl_dataset_destroy_sync, &dsda, tag, 0);
1156 		dsl_sync_task_create(dstg, dsl_dir_destroy_check,
1157 		    dsl_dir_destroy_sync, &dummy_ds, FTAG, 0);
1158 		err = dsl_sync_task_group_wait(dstg);
1159 		dsl_sync_task_group_destroy(dstg);
1160 
1161 		/*
1162 		 * We could be racing against 'zfs release' or 'zfs destroy -d'
1163 		 * on the origin snap, in which case we can get EBUSY if we
1164 		 * needed to destroy the origin snap but were not ready to
1165 		 * do so.
1166 		 */
1167 		if (dsda.need_prep) {
1168 			ASSERT(err == EBUSY);
1169 			ASSERT(dsl_dir_is_clone(dd));
1170 			ASSERT(dsda.rm_origin == NULL);
1171 		}
1172 	} while (dsda.need_prep);
1173 
1174 	if (dsda.rm_origin != NULL)
1175 		dsl_dataset_disown(dsda.rm_origin, tag);
1176 
1177 	/* if it is successful, dsl_dir_destroy_sync will close the dd */
1178 	if (err)
1179 		dsl_dir_close(dd, FTAG);
1180 out:
1181 	dsl_dataset_disown(ds, tag);
1182 	return (err);
1183 }
1184 
1185 blkptr_t *
1186 dsl_dataset_get_blkptr(dsl_dataset_t *ds)
1187 {
1188 	return (&ds->ds_phys->ds_bp);
1189 }
1190 
1191 void
1192 dsl_dataset_set_blkptr(dsl_dataset_t *ds, blkptr_t *bp, dmu_tx_t *tx)
1193 {
1194 	ASSERT(dmu_tx_is_syncing(tx));
1195 	/* If it's the meta-objset, set dp_meta_rootbp */
1196 	if (ds == NULL) {
1197 		tx->tx_pool->dp_meta_rootbp = *bp;
1198 	} else {
1199 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
1200 		ds->ds_phys->ds_bp = *bp;
1201 	}
1202 }
1203 
1204 spa_t *
1205 dsl_dataset_get_spa(dsl_dataset_t *ds)
1206 {
1207 	return (ds->ds_dir->dd_pool->dp_spa);
1208 }
1209 
1210 void
1211 dsl_dataset_dirty(dsl_dataset_t *ds, dmu_tx_t *tx)
1212 {
1213 	dsl_pool_t *dp;
1214 
1215 	if (ds == NULL) /* this is the meta-objset */
1216 		return;
1217 
1218 	ASSERT(ds->ds_objset != NULL);
1219 
1220 	if (ds->ds_phys->ds_next_snap_obj != 0)
1221 		panic("dirtying snapshot!");
1222 
1223 	dp = ds->ds_dir->dd_pool;
1224 
1225 	if (txg_list_add(&dp->dp_dirty_datasets, ds, tx->tx_txg) == 0) {
1226 		/* up the hold count until we can be written out */
1227 		dmu_buf_add_ref(ds->ds_dbuf, ds);
1228 	}
1229 }
1230 
1231 /*
1232  * The unique space in the head dataset can be calculated by subtracting
1233  * the space used in the most recent snapshot, that is still being used
1234  * in this file system, from the space currently in use.  To figure out
1235  * the space in the most recent snapshot still in use, we need to take
1236  * the total space used in the snapshot and subtract out the space that
1237  * has been freed up since the snapshot was taken.
1238  */
1239 static void
1240 dsl_dataset_recalc_head_uniq(dsl_dataset_t *ds)
1241 {
1242 	uint64_t mrs_used;
1243 	uint64_t dlused, dlcomp, dluncomp;
1244 
1245 	ASSERT(!dsl_dataset_is_snapshot(ds));
1246 
1247 	if (ds->ds_phys->ds_prev_snap_obj != 0)
1248 		mrs_used = ds->ds_prev->ds_phys->ds_used_bytes;
1249 	else
1250 		mrs_used = 0;
1251 
1252 	dsl_deadlist_space(&ds->ds_deadlist, &dlused, &dlcomp, &dluncomp);
1253 
1254 	ASSERT3U(dlused, <=, mrs_used);
1255 	ds->ds_phys->ds_unique_bytes =
1256 	    ds->ds_phys->ds_used_bytes - (mrs_used - dlused);
1257 
1258 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) >=
1259 	    SPA_VERSION_UNIQUE_ACCURATE)
1260 		ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
1261 }
1262 
1263 struct killarg {
1264 	dsl_dataset_t *ds;
1265 	dmu_tx_t *tx;
1266 };
1267 
1268 /* ARGSUSED */
1269 static int
1270 kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, arc_buf_t *pbuf,
1271     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
1272 {
1273 	struct killarg *ka = arg;
1274 	dmu_tx_t *tx = ka->tx;
1275 
1276 	if (bp == NULL)
1277 		return (0);
1278 
1279 	if (zb->zb_level == ZB_ZIL_LEVEL) {
1280 		ASSERT(zilog != NULL);
1281 		/*
1282 		 * It's a block in the intent log.  It has no
1283 		 * accounting, so just free it.
1284 		 */
1285 		dsl_free(ka->tx->tx_pool, ka->tx->tx_txg, bp);
1286 	} else {
1287 		ASSERT(zilog == NULL);
1288 		ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
1289 		(void) dsl_dataset_block_kill(ka->ds, bp, tx, B_FALSE);
1290 	}
1291 
1292 	return (0);
1293 }
1294 
1295 /* ARGSUSED */
1296 static int
1297 dsl_dataset_destroy_begin_check(void *arg1, void *arg2, dmu_tx_t *tx)
1298 {
1299 	dsl_dataset_t *ds = arg1;
1300 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1301 	uint64_t count;
1302 	int err;
1303 
1304 	/*
1305 	 * Can't delete a head dataset if there are snapshots of it.
1306 	 * (Except if the only snapshots are from the branch we cloned
1307 	 * from.)
1308 	 */
1309 	if (ds->ds_prev != NULL &&
1310 	    ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1311 		return (EBUSY);
1312 
1313 	/*
1314 	 * This is really a dsl_dir thing, but check it here so that
1315 	 * we'll be less likely to leave this dataset inconsistent &
1316 	 * nearly destroyed.
1317 	 */
1318 	err = zap_count(mos, ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
1319 	if (err)
1320 		return (err);
1321 	if (count != 0)
1322 		return (EEXIST);
1323 
1324 	return (0);
1325 }
1326 
1327 /* ARGSUSED */
1328 static void
1329 dsl_dataset_destroy_begin_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1330 {
1331 	dsl_dataset_t *ds = arg1;
1332 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1333 
1334 	/* Mark it as inconsistent on-disk, in case we crash */
1335 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1336 	ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
1337 
1338 	spa_history_log_internal(LOG_DS_DESTROY_BEGIN, dp->dp_spa, tx,
1339 	    "dataset = %llu", ds->ds_object);
1340 }
1341 
1342 static int
1343 dsl_dataset_origin_check(struct dsl_ds_destroyarg *dsda, void *tag,
1344     dmu_tx_t *tx)
1345 {
1346 	dsl_dataset_t *ds = dsda->ds;
1347 	dsl_dataset_t *ds_prev = ds->ds_prev;
1348 
1349 	if (dsl_dataset_might_destroy_origin(ds_prev)) {
1350 		struct dsl_ds_destroyarg ndsda = {0};
1351 
1352 		/*
1353 		 * If we're not prepared to remove the origin, don't remove
1354 		 * the clone either.
1355 		 */
1356 		if (dsda->rm_origin == NULL) {
1357 			dsda->need_prep = B_TRUE;
1358 			return (EBUSY);
1359 		}
1360 
1361 		ndsda.ds = ds_prev;
1362 		ndsda.is_origin_rm = B_TRUE;
1363 		return (dsl_dataset_destroy_check(&ndsda, tag, tx));
1364 	}
1365 
1366 	/*
1367 	 * If we're not going to remove the origin after all,
1368 	 * undo the open context setup.
1369 	 */
1370 	if (dsda->rm_origin != NULL) {
1371 		dsl_dataset_disown(dsda->rm_origin, tag);
1372 		dsda->rm_origin = NULL;
1373 	}
1374 
1375 	return (0);
1376 }
1377 
1378 /*
1379  * If you add new checks here, you may need to add
1380  * additional checks to the "temporary" case in
1381  * snapshot_check() in dmu_objset.c.
1382  */
1383 /* ARGSUSED */
1384 int
1385 dsl_dataset_destroy_check(void *arg1, void *arg2, dmu_tx_t *tx)
1386 {
1387 	struct dsl_ds_destroyarg *dsda = arg1;
1388 	dsl_dataset_t *ds = dsda->ds;
1389 
1390 	/* we have an owner hold, so noone else can destroy us */
1391 	ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
1392 
1393 	/*
1394 	 * Only allow deferred destroy on pools that support it.
1395 	 * NOTE: deferred destroy is only supported on snapshots.
1396 	 */
1397 	if (dsda->defer) {
1398 		if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
1399 		    SPA_VERSION_USERREFS)
1400 			return (ENOTSUP);
1401 		ASSERT(dsl_dataset_is_snapshot(ds));
1402 		return (0);
1403 	}
1404 
1405 	/*
1406 	 * Can't delete a head dataset if there are snapshots of it.
1407 	 * (Except if the only snapshots are from the branch we cloned
1408 	 * from.)
1409 	 */
1410 	if (ds->ds_prev != NULL &&
1411 	    ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1412 		return (EBUSY);
1413 
1414 	/*
1415 	 * If we made changes this txg, traverse_dsl_dataset won't find
1416 	 * them.  Try again.
1417 	 */
1418 	if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
1419 		return (EAGAIN);
1420 
1421 	if (dsl_dataset_is_snapshot(ds)) {
1422 		/*
1423 		 * If this snapshot has an elevated user reference count,
1424 		 * we can't destroy it yet.
1425 		 */
1426 		if (ds->ds_userrefs > 0 && !dsda->releasing)
1427 			return (EBUSY);
1428 
1429 		mutex_enter(&ds->ds_lock);
1430 		/*
1431 		 * Can't delete a branch point. However, if we're destroying
1432 		 * a clone and removing its origin due to it having a user
1433 		 * hold count of 0 and having been marked for deferred destroy,
1434 		 * it's OK for the origin to have a single clone.
1435 		 */
1436 		if (ds->ds_phys->ds_num_children >
1437 		    (dsda->is_origin_rm ? 2 : 1)) {
1438 			mutex_exit(&ds->ds_lock);
1439 			return (EEXIST);
1440 		}
1441 		mutex_exit(&ds->ds_lock);
1442 	} else if (dsl_dir_is_clone(ds->ds_dir)) {
1443 		return (dsl_dataset_origin_check(dsda, arg2, tx));
1444 	}
1445 
1446 	/* XXX we should do some i/o error checking... */
1447 	return (0);
1448 }
1449 
1450 struct refsarg {
1451 	kmutex_t lock;
1452 	boolean_t gone;
1453 	kcondvar_t cv;
1454 };
1455 
1456 /* ARGSUSED */
1457 static void
1458 dsl_dataset_refs_gone(dmu_buf_t *db, void *argv)
1459 {
1460 	struct refsarg *arg = argv;
1461 
1462 	mutex_enter(&arg->lock);
1463 	arg->gone = TRUE;
1464 	cv_signal(&arg->cv);
1465 	mutex_exit(&arg->lock);
1466 }
1467 
1468 static void
1469 dsl_dataset_drain_refs(dsl_dataset_t *ds, void *tag)
1470 {
1471 	struct refsarg arg;
1472 
1473 	mutex_init(&arg.lock, NULL, MUTEX_DEFAULT, NULL);
1474 	cv_init(&arg.cv, NULL, CV_DEFAULT, NULL);
1475 	arg.gone = FALSE;
1476 	(void) dmu_buf_update_user(ds->ds_dbuf, ds, &arg, &ds->ds_phys,
1477 	    dsl_dataset_refs_gone);
1478 	dmu_buf_rele(ds->ds_dbuf, tag);
1479 	mutex_enter(&arg.lock);
1480 	while (!arg.gone)
1481 		cv_wait(&arg.cv, &arg.lock);
1482 	ASSERT(arg.gone);
1483 	mutex_exit(&arg.lock);
1484 	ds->ds_dbuf = NULL;
1485 	ds->ds_phys = NULL;
1486 	mutex_destroy(&arg.lock);
1487 	cv_destroy(&arg.cv);
1488 }
1489 
1490 static void
1491 remove_from_next_clones(dsl_dataset_t *ds, uint64_t obj, dmu_tx_t *tx)
1492 {
1493 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1494 	uint64_t count;
1495 	int err;
1496 
1497 	ASSERT(ds->ds_phys->ds_num_children >= 2);
1498 	err = zap_remove_int(mos, ds->ds_phys->ds_next_clones_obj, obj, tx);
1499 	/*
1500 	 * The err should not be ENOENT, but a bug in a previous version
1501 	 * of the code could cause upgrade_clones_cb() to not set
1502 	 * ds_next_snap_obj when it should, leading to a missing entry.
1503 	 * If we knew that the pool was created after
1504 	 * SPA_VERSION_NEXT_CLONES, we could assert that it isn't
1505 	 * ENOENT.  However, at least we can check that we don't have
1506 	 * too many entries in the next_clones_obj even after failing to
1507 	 * remove this one.
1508 	 */
1509 	if (err != ENOENT) {
1510 		VERIFY3U(err, ==, 0);
1511 	}
1512 	ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
1513 	    &count));
1514 	ASSERT3U(count, <=, ds->ds_phys->ds_num_children - 2);
1515 }
1516 
1517 static void
1518 dsl_dataset_remove_clones_key(dsl_dataset_t *ds, uint64_t mintxg, dmu_tx_t *tx)
1519 {
1520 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1521 	zap_cursor_t zc;
1522 	zap_attribute_t za;
1523 
1524 	/*
1525 	 * If it is the old version, dd_clones doesn't exist so we can't
1526 	 * find the clones, but deadlist_remove_key() is a no-op so it
1527 	 * doesn't matter.
1528 	 */
1529 	if (ds->ds_dir->dd_phys->dd_clones == 0)
1530 		return;
1531 
1532 	for (zap_cursor_init(&zc, mos, ds->ds_dir->dd_phys->dd_clones);
1533 	    zap_cursor_retrieve(&zc, &za) == 0;
1534 	    zap_cursor_advance(&zc)) {
1535 		dsl_dataset_t *clone;
1536 
1537 		VERIFY3U(0, ==, dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
1538 		    za.za_first_integer, FTAG, &clone));
1539 		if (clone->ds_dir->dd_origin_txg > mintxg) {
1540 			dsl_deadlist_remove_key(&clone->ds_deadlist,
1541 			    mintxg, tx);
1542 			dsl_dataset_remove_clones_key(clone, mintxg, tx);
1543 		}
1544 		dsl_dataset_rele(clone, FTAG);
1545 	}
1546 	zap_cursor_fini(&zc);
1547 }
1548 
1549 struct process_old_arg {
1550 	dsl_dataset_t *ds;
1551 	dsl_dataset_t *ds_prev;
1552 	boolean_t after_branch_point;
1553 	zio_t *pio;
1554 	uint64_t used, comp, uncomp;
1555 };
1556 
1557 static int
1558 process_old_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1559 {
1560 	struct process_old_arg *poa = arg;
1561 	dsl_pool_t *dp = poa->ds->ds_dir->dd_pool;
1562 
1563 	if (bp->blk_birth <= poa->ds->ds_phys->ds_prev_snap_txg) {
1564 		dsl_deadlist_insert(&poa->ds->ds_deadlist, bp, tx);
1565 		if (poa->ds_prev && !poa->after_branch_point &&
1566 		    bp->blk_birth >
1567 		    poa->ds_prev->ds_phys->ds_prev_snap_txg) {
1568 			poa->ds_prev->ds_phys->ds_unique_bytes +=
1569 			    bp_get_dsize_sync(dp->dp_spa, bp);
1570 		}
1571 	} else {
1572 		poa->used += bp_get_dsize_sync(dp->dp_spa, bp);
1573 		poa->comp += BP_GET_PSIZE(bp);
1574 		poa->uncomp += BP_GET_UCSIZE(bp);
1575 		dsl_free_sync(poa->pio, dp, tx->tx_txg, bp);
1576 	}
1577 	return (0);
1578 }
1579 
1580 static void
1581 process_old_deadlist(dsl_dataset_t *ds, dsl_dataset_t *ds_prev,
1582     dsl_dataset_t *ds_next, boolean_t after_branch_point, dmu_tx_t *tx)
1583 {
1584 	struct process_old_arg poa = { 0 };
1585 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1586 	objset_t *mos = dp->dp_meta_objset;
1587 
1588 	ASSERT(ds->ds_deadlist.dl_oldfmt);
1589 	ASSERT(ds_next->ds_deadlist.dl_oldfmt);
1590 
1591 	poa.ds = ds;
1592 	poa.ds_prev = ds_prev;
1593 	poa.after_branch_point = after_branch_point;
1594 	poa.pio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
1595 	VERIFY3U(0, ==, bpobj_iterate(&ds_next->ds_deadlist.dl_bpobj,
1596 	    process_old_cb, &poa, tx));
1597 	VERIFY3U(zio_wait(poa.pio), ==, 0);
1598 	ASSERT3U(poa.used, ==, ds->ds_phys->ds_unique_bytes);
1599 
1600 	/* change snapused */
1601 	dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1602 	    -poa.used, -poa.comp, -poa.uncomp, tx);
1603 
1604 	/* swap next's deadlist to our deadlist */
1605 	dsl_deadlist_close(&ds->ds_deadlist);
1606 	dsl_deadlist_close(&ds_next->ds_deadlist);
1607 	SWITCH64(ds_next->ds_phys->ds_deadlist_obj,
1608 	    ds->ds_phys->ds_deadlist_obj);
1609 	dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
1610 	dsl_deadlist_open(&ds_next->ds_deadlist, mos,
1611 	    ds_next->ds_phys->ds_deadlist_obj);
1612 }
1613 
1614 void
1615 dsl_dataset_destroy_sync(void *arg1, void *tag, dmu_tx_t *tx)
1616 {
1617 	struct dsl_ds_destroyarg *dsda = arg1;
1618 	dsl_dataset_t *ds = dsda->ds;
1619 	int err;
1620 	int after_branch_point = FALSE;
1621 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1622 	objset_t *mos = dp->dp_meta_objset;
1623 	dsl_dataset_t *ds_prev = NULL;
1624 	boolean_t wont_destroy;
1625 	uint64_t obj;
1626 
1627 	wont_destroy = (dsda->defer &&
1628 	    (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1));
1629 
1630 	ASSERT(ds->ds_owner || wont_destroy);
1631 	ASSERT(dsda->defer || ds->ds_phys->ds_num_children <= 1);
1632 	ASSERT(ds->ds_prev == NULL ||
1633 	    ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
1634 	ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
1635 
1636 	if (wont_destroy) {
1637 		ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1638 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
1639 		ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
1640 		return;
1641 	}
1642 
1643 	/* signal any waiters that this dataset is going away */
1644 	mutex_enter(&ds->ds_lock);
1645 	ds->ds_owner = dsl_reaper;
1646 	cv_broadcast(&ds->ds_exclusive_cv);
1647 	mutex_exit(&ds->ds_lock);
1648 
1649 	/* Remove our reservation */
1650 	if (ds->ds_reserved != 0) {
1651 		dsl_prop_setarg_t psa;
1652 		uint64_t value = 0;
1653 
1654 		dsl_prop_setarg_init_uint64(&psa, "refreservation",
1655 		    (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
1656 		    &value);
1657 		psa.psa_effective_value = 0;	/* predict default value */
1658 
1659 		dsl_dataset_set_reservation_sync(ds, &psa, tx);
1660 		ASSERT3U(ds->ds_reserved, ==, 0);
1661 	}
1662 
1663 	ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1664 
1665 	dsl_scan_ds_destroyed(ds, tx);
1666 
1667 	obj = ds->ds_object;
1668 
1669 	if (ds->ds_phys->ds_prev_snap_obj != 0) {
1670 		if (ds->ds_prev) {
1671 			ds_prev = ds->ds_prev;
1672 		} else {
1673 			VERIFY(0 == dsl_dataset_hold_obj(dp,
1674 			    ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
1675 		}
1676 		after_branch_point =
1677 		    (ds_prev->ds_phys->ds_next_snap_obj != obj);
1678 
1679 		dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
1680 		if (after_branch_point &&
1681 		    ds_prev->ds_phys->ds_next_clones_obj != 0) {
1682 			remove_from_next_clones(ds_prev, obj, tx);
1683 			if (ds->ds_phys->ds_next_snap_obj != 0) {
1684 				VERIFY(0 == zap_add_int(mos,
1685 				    ds_prev->ds_phys->ds_next_clones_obj,
1686 				    ds->ds_phys->ds_next_snap_obj, tx));
1687 			}
1688 		}
1689 		if (after_branch_point &&
1690 		    ds->ds_phys->ds_next_snap_obj == 0) {
1691 			/* This clone is toast. */
1692 			ASSERT(ds_prev->ds_phys->ds_num_children > 1);
1693 			ds_prev->ds_phys->ds_num_children--;
1694 
1695 			/*
1696 			 * If the clone's origin has no other clones, no
1697 			 * user holds, and has been marked for deferred
1698 			 * deletion, then we should have done the necessary
1699 			 * destroy setup for it.
1700 			 */
1701 			if (ds_prev->ds_phys->ds_num_children == 1 &&
1702 			    ds_prev->ds_userrefs == 0 &&
1703 			    DS_IS_DEFER_DESTROY(ds_prev)) {
1704 				ASSERT3P(dsda->rm_origin, !=, NULL);
1705 			} else {
1706 				ASSERT3P(dsda->rm_origin, ==, NULL);
1707 			}
1708 		} else if (!after_branch_point) {
1709 			ds_prev->ds_phys->ds_next_snap_obj =
1710 			    ds->ds_phys->ds_next_snap_obj;
1711 		}
1712 	}
1713 
1714 	if (dsl_dataset_is_snapshot(ds)) {
1715 		dsl_dataset_t *ds_next;
1716 		uint64_t old_unique;
1717 		uint64_t used = 0, comp = 0, uncomp = 0;
1718 
1719 		VERIFY(0 == dsl_dataset_hold_obj(dp,
1720 		    ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
1721 		ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
1722 
1723 		old_unique = ds_next->ds_phys->ds_unique_bytes;
1724 
1725 		dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
1726 		ds_next->ds_phys->ds_prev_snap_obj =
1727 		    ds->ds_phys->ds_prev_snap_obj;
1728 		ds_next->ds_phys->ds_prev_snap_txg =
1729 		    ds->ds_phys->ds_prev_snap_txg;
1730 		ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1731 		    ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
1732 
1733 
1734 		if (ds_next->ds_deadlist.dl_oldfmt) {
1735 			process_old_deadlist(ds, ds_prev, ds_next,
1736 			    after_branch_point, tx);
1737 		} else {
1738 			/* Adjust prev's unique space. */
1739 			if (ds_prev && !after_branch_point) {
1740 				dsl_deadlist_space_range(&ds_next->ds_deadlist,
1741 				    ds_prev->ds_phys->ds_prev_snap_txg,
1742 				    ds->ds_phys->ds_prev_snap_txg,
1743 				    &used, &comp, &uncomp);
1744 				ds_prev->ds_phys->ds_unique_bytes += used;
1745 			}
1746 
1747 			/* Adjust snapused. */
1748 			dsl_deadlist_space_range(&ds_next->ds_deadlist,
1749 			    ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
1750 			    &used, &comp, &uncomp);
1751 			dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1752 			    -used, -comp, -uncomp, tx);
1753 
1754 			/* Move blocks to be freed to pool's free list. */
1755 			dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
1756 			    &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
1757 			    tx);
1758 			dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
1759 			    DD_USED_HEAD, used, comp, uncomp, tx);
1760 			dsl_dir_dirty(tx->tx_pool->dp_free_dir, tx);
1761 
1762 			/* Merge our deadlist into next's and free it. */
1763 			dsl_deadlist_merge(&ds_next->ds_deadlist,
1764 			    ds->ds_phys->ds_deadlist_obj, tx);
1765 		}
1766 		dsl_deadlist_close(&ds->ds_deadlist);
1767 		dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1768 
1769 		/* Collapse range in clone heads */
1770 		dsl_dataset_remove_clones_key(ds,
1771 		    ds->ds_phys->ds_creation_txg, tx);
1772 
1773 		if (dsl_dataset_is_snapshot(ds_next)) {
1774 			dsl_dataset_t *ds_nextnext;
1775 
1776 			/*
1777 			 * Update next's unique to include blocks which
1778 			 * were previously shared by only this snapshot
1779 			 * and it.  Those blocks will be born after the
1780 			 * prev snap and before this snap, and will have
1781 			 * died after the next snap and before the one
1782 			 * after that (ie. be on the snap after next's
1783 			 * deadlist).
1784 			 */
1785 			VERIFY(0 == dsl_dataset_hold_obj(dp,
1786 			    ds_next->ds_phys->ds_next_snap_obj,
1787 			    FTAG, &ds_nextnext));
1788 			dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
1789 			    ds->ds_phys->ds_prev_snap_txg,
1790 			    ds->ds_phys->ds_creation_txg,
1791 			    &used, &comp, &uncomp);
1792 			ds_next->ds_phys->ds_unique_bytes += used;
1793 			dsl_dataset_rele(ds_nextnext, FTAG);
1794 			ASSERT3P(ds_next->ds_prev, ==, NULL);
1795 
1796 			/* Collapse range in this head. */
1797 			dsl_dataset_t *hds;
1798 			VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
1799 			    ds->ds_dir->dd_phys->dd_head_dataset_obj,
1800 			    FTAG, &hds));
1801 			dsl_deadlist_remove_key(&hds->ds_deadlist,
1802 			    ds->ds_phys->ds_creation_txg, tx);
1803 			dsl_dataset_rele(hds, FTAG);
1804 
1805 		} else {
1806 			ASSERT3P(ds_next->ds_prev, ==, ds);
1807 			dsl_dataset_drop_ref(ds_next->ds_prev, ds_next);
1808 			ds_next->ds_prev = NULL;
1809 			if (ds_prev) {
1810 				VERIFY(0 == dsl_dataset_get_ref(dp,
1811 				    ds->ds_phys->ds_prev_snap_obj,
1812 				    ds_next, &ds_next->ds_prev));
1813 			}
1814 
1815 			dsl_dataset_recalc_head_uniq(ds_next);
1816 
1817 			/*
1818 			 * Reduce the amount of our unconsmed refreservation
1819 			 * being charged to our parent by the amount of
1820 			 * new unique data we have gained.
1821 			 */
1822 			if (old_unique < ds_next->ds_reserved) {
1823 				int64_t mrsdelta;
1824 				uint64_t new_unique =
1825 				    ds_next->ds_phys->ds_unique_bytes;
1826 
1827 				ASSERT(old_unique <= new_unique);
1828 				mrsdelta = MIN(new_unique - old_unique,
1829 				    ds_next->ds_reserved - old_unique);
1830 				dsl_dir_diduse_space(ds->ds_dir,
1831 				    DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
1832 			}
1833 		}
1834 		dsl_dataset_rele(ds_next, FTAG);
1835 	} else {
1836 		/*
1837 		 * There's no next snapshot, so this is a head dataset.
1838 		 * Destroy the deadlist.  Unless it's a clone, the
1839 		 * deadlist should be empty.  (If it's a clone, it's
1840 		 * safe to ignore the deadlist contents.)
1841 		 */
1842 		struct killarg ka;
1843 
1844 		dsl_deadlist_close(&ds->ds_deadlist);
1845 		dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1846 		ds->ds_phys->ds_deadlist_obj = 0;
1847 
1848 		/*
1849 		 * Free everything that we point to (that's born after
1850 		 * the previous snapshot, if we are a clone)
1851 		 *
1852 		 * NB: this should be very quick, because we already
1853 		 * freed all the objects in open context.
1854 		 */
1855 		ka.ds = ds;
1856 		ka.tx = tx;
1857 		err = traverse_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
1858 		    TRAVERSE_POST, kill_blkptr, &ka);
1859 		ASSERT3U(err, ==, 0);
1860 		ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
1861 		    ds->ds_phys->ds_unique_bytes == 0);
1862 
1863 		if (ds->ds_prev != NULL) {
1864 			if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
1865 				VERIFY3U(0, ==, zap_remove_int(mos,
1866 				    ds->ds_prev->ds_dir->dd_phys->dd_clones,
1867 				    ds->ds_object, tx));
1868 			}
1869 			dsl_dataset_rele(ds->ds_prev, ds);
1870 			ds->ds_prev = ds_prev = NULL;
1871 		}
1872 	}
1873 
1874 	/*
1875 	 * This must be done after the dsl_traverse(), because it will
1876 	 * re-open the objset.
1877 	 */
1878 	if (ds->ds_objset) {
1879 		dmu_objset_evict(ds->ds_objset);
1880 		ds->ds_objset = NULL;
1881 	}
1882 
1883 	if (ds->ds_dir->dd_phys->dd_head_dataset_obj == ds->ds_object) {
1884 		/* Erase the link in the dir */
1885 		dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
1886 		ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
1887 		ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
1888 		err = zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx);
1889 		ASSERT(err == 0);
1890 	} else {
1891 		/* remove from snapshot namespace */
1892 		dsl_dataset_t *ds_head;
1893 		ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
1894 		VERIFY(0 == dsl_dataset_hold_obj(dp,
1895 		    ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
1896 		VERIFY(0 == dsl_dataset_get_snapname(ds));
1897 #ifdef ZFS_DEBUG
1898 		{
1899 			uint64_t val;
1900 
1901 			err = dsl_dataset_snap_lookup(ds_head,
1902 			    ds->ds_snapname, &val);
1903 			ASSERT3U(err, ==, 0);
1904 			ASSERT3U(val, ==, obj);
1905 		}
1906 #endif
1907 		err = dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx);
1908 		ASSERT(err == 0);
1909 		dsl_dataset_rele(ds_head, FTAG);
1910 	}
1911 
1912 	if (ds_prev && ds->ds_prev != ds_prev)
1913 		dsl_dataset_rele(ds_prev, FTAG);
1914 
1915 	spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
1916 	spa_history_log_internal(LOG_DS_DESTROY, dp->dp_spa, tx,
1917 	    "dataset = %llu", ds->ds_object);
1918 
1919 	if (ds->ds_phys->ds_next_clones_obj != 0) {
1920 		uint64_t count;
1921 		ASSERT(0 == zap_count(mos,
1922 		    ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
1923 		VERIFY(0 == dmu_object_free(mos,
1924 		    ds->ds_phys->ds_next_clones_obj, tx));
1925 	}
1926 	if (ds->ds_phys->ds_props_obj != 0)
1927 		VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
1928 	if (ds->ds_phys->ds_userrefs_obj != 0)
1929 		VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
1930 	dsl_dir_close(ds->ds_dir, ds);
1931 	ds->ds_dir = NULL;
1932 	dsl_dataset_drain_refs(ds, tag);
1933 	VERIFY(0 == dmu_object_free(mos, obj, tx));
1934 
1935 	if (dsda->rm_origin) {
1936 		/*
1937 		 * Remove the origin of the clone we just destroyed.
1938 		 */
1939 		struct dsl_ds_destroyarg ndsda = {0};
1940 
1941 		ndsda.ds = dsda->rm_origin;
1942 		dsl_dataset_destroy_sync(&ndsda, tag, tx);
1943 	}
1944 }
1945 
1946 static int
1947 dsl_dataset_snapshot_reserve_space(dsl_dataset_t *ds, dmu_tx_t *tx)
1948 {
1949 	uint64_t asize;
1950 
1951 	if (!dmu_tx_is_syncing(tx))
1952 		return (0);
1953 
1954 	/*
1955 	 * If there's an fs-only reservation, any blocks that might become
1956 	 * owned by the snapshot dataset must be accommodated by space
1957 	 * outside of the reservation.
1958 	 */
1959 	ASSERT(ds->ds_reserved == 0 || DS_UNIQUE_IS_ACCURATE(ds));
1960 	asize = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
1961 	if (asize > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
1962 		return (ENOSPC);
1963 
1964 	/*
1965 	 * Propogate any reserved space for this snapshot to other
1966 	 * snapshot checks in this sync group.
1967 	 */
1968 	if (asize > 0)
1969 		dsl_dir_willuse_space(ds->ds_dir, asize, tx);
1970 
1971 	return (0);
1972 }
1973 
1974 int
1975 dsl_dataset_snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
1976 {
1977 	dsl_dataset_t *ds = arg1;
1978 	const char *snapname = arg2;
1979 	int err;
1980 	uint64_t value;
1981 
1982 	/*
1983 	 * We don't allow multiple snapshots of the same txg.  If there
1984 	 * is already one, try again.
1985 	 */
1986 	if (ds->ds_phys->ds_prev_snap_txg >= tx->tx_txg)
1987 		return (EAGAIN);
1988 
1989 	/*
1990 	 * Check for conflicting name snapshot name.
1991 	 */
1992 	err = dsl_dataset_snap_lookup(ds, snapname, &value);
1993 	if (err == 0)
1994 		return (EEXIST);
1995 	if (err != ENOENT)
1996 		return (err);
1997 
1998 	/*
1999 	 * Check that the dataset's name is not too long.  Name consists
2000 	 * of the dataset's length + 1 for the @-sign + snapshot name's length
2001 	 */
2002 	if (dsl_dataset_namelen(ds) + 1 + strlen(snapname) >= MAXNAMELEN)
2003 		return (ENAMETOOLONG);
2004 
2005 	err = dsl_dataset_snapshot_reserve_space(ds, tx);
2006 	if (err)
2007 		return (err);
2008 
2009 	ds->ds_trysnap_txg = tx->tx_txg;
2010 	return (0);
2011 }
2012 
2013 void
2014 dsl_dataset_snapshot_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2015 {
2016 	dsl_dataset_t *ds = arg1;
2017 	const char *snapname = arg2;
2018 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
2019 	dmu_buf_t *dbuf;
2020 	dsl_dataset_phys_t *dsphys;
2021 	uint64_t dsobj, crtxg;
2022 	objset_t *mos = dp->dp_meta_objset;
2023 	int err;
2024 
2025 	ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
2026 
2027 	/*
2028 	 * The origin's ds_creation_txg has to be < TXG_INITIAL
2029 	 */
2030 	if (strcmp(snapname, ORIGIN_DIR_NAME) == 0)
2031 		crtxg = 1;
2032 	else
2033 		crtxg = tx->tx_txg;
2034 
2035 	dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
2036 	    DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
2037 	VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
2038 	dmu_buf_will_dirty(dbuf, tx);
2039 	dsphys = dbuf->db_data;
2040 	bzero(dsphys, sizeof (dsl_dataset_phys_t));
2041 	dsphys->ds_dir_obj = ds->ds_dir->dd_object;
2042 	dsphys->ds_fsid_guid = unique_create();
2043 	(void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
2044 	    sizeof (dsphys->ds_guid));
2045 	dsphys->ds_prev_snap_obj = ds->ds_phys->ds_prev_snap_obj;
2046 	dsphys->ds_prev_snap_txg = ds->ds_phys->ds_prev_snap_txg;
2047 	dsphys->ds_next_snap_obj = ds->ds_object;
2048 	dsphys->ds_num_children = 1;
2049 	dsphys->ds_creation_time = gethrestime_sec();
2050 	dsphys->ds_creation_txg = crtxg;
2051 	dsphys->ds_deadlist_obj = ds->ds_phys->ds_deadlist_obj;
2052 	dsphys->ds_used_bytes = ds->ds_phys->ds_used_bytes;
2053 	dsphys->ds_compressed_bytes = ds->ds_phys->ds_compressed_bytes;
2054 	dsphys->ds_uncompressed_bytes = ds->ds_phys->ds_uncompressed_bytes;
2055 	dsphys->ds_flags = ds->ds_phys->ds_flags;
2056 	dsphys->ds_bp = ds->ds_phys->ds_bp;
2057 	dmu_buf_rele(dbuf, FTAG);
2058 
2059 	ASSERT3U(ds->ds_prev != 0, ==, ds->ds_phys->ds_prev_snap_obj != 0);
2060 	if (ds->ds_prev) {
2061 		uint64_t next_clones_obj =
2062 		    ds->ds_prev->ds_phys->ds_next_clones_obj;
2063 		ASSERT(ds->ds_prev->ds_phys->ds_next_snap_obj ==
2064 		    ds->ds_object ||
2065 		    ds->ds_prev->ds_phys->ds_num_children > 1);
2066 		if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
2067 			dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2068 			ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
2069 			    ds->ds_prev->ds_phys->ds_creation_txg);
2070 			ds->ds_prev->ds_phys->ds_next_snap_obj = dsobj;
2071 		} else if (next_clones_obj != 0) {
2072 			remove_from_next_clones(ds->ds_prev,
2073 			    dsphys->ds_next_snap_obj, tx);
2074 			VERIFY3U(0, ==, zap_add_int(mos,
2075 			    next_clones_obj, dsobj, tx));
2076 		}
2077 	}
2078 
2079 	/*
2080 	 * If we have a reference-reservation on this dataset, we will
2081 	 * need to increase the amount of refreservation being charged
2082 	 * since our unique space is going to zero.
2083 	 */
2084 	if (ds->ds_reserved) {
2085 		int64_t delta;
2086 		ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
2087 		delta = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
2088 		dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV,
2089 		    delta, 0, 0, tx);
2090 	}
2091 
2092 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
2093 	zfs_dbgmsg("taking snapshot %s@%s/%llu; newkey=%llu",
2094 	    ds->ds_dir->dd_myname, snapname, dsobj,
2095 	    ds->ds_phys->ds_prev_snap_txg);
2096 	ds->ds_phys->ds_deadlist_obj = dsl_deadlist_clone(&ds->ds_deadlist,
2097 	    UINT64_MAX, ds->ds_phys->ds_prev_snap_obj, tx);
2098 	dsl_deadlist_close(&ds->ds_deadlist);
2099 	dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
2100 	dsl_deadlist_add_key(&ds->ds_deadlist,
2101 	    ds->ds_phys->ds_prev_snap_txg, tx);
2102 
2103 	ASSERT3U(ds->ds_phys->ds_prev_snap_txg, <, tx->tx_txg);
2104 	ds->ds_phys->ds_prev_snap_obj = dsobj;
2105 	ds->ds_phys->ds_prev_snap_txg = crtxg;
2106 	ds->ds_phys->ds_unique_bytes = 0;
2107 	if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
2108 		ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
2109 
2110 	err = zap_add(mos, ds->ds_phys->ds_snapnames_zapobj,
2111 	    snapname, 8, 1, &dsobj, tx);
2112 	ASSERT(err == 0);
2113 
2114 	if (ds->ds_prev)
2115 		dsl_dataset_drop_ref(ds->ds_prev, ds);
2116 	VERIFY(0 == dsl_dataset_get_ref(dp,
2117 	    ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
2118 
2119 	dsl_scan_ds_snapshotted(ds, tx);
2120 
2121 	dsl_dir_snap_cmtime_update(ds->ds_dir);
2122 
2123 	spa_history_log_internal(LOG_DS_SNAPSHOT, dp->dp_spa, tx,
2124 	    "dataset = %llu", dsobj);
2125 }
2126 
2127 void
2128 dsl_dataset_sync(dsl_dataset_t *ds, zio_t *zio, dmu_tx_t *tx)
2129 {
2130 	ASSERT(dmu_tx_is_syncing(tx));
2131 	ASSERT(ds->ds_objset != NULL);
2132 	ASSERT(ds->ds_phys->ds_next_snap_obj == 0);
2133 
2134 	/*
2135 	 * in case we had to change ds_fsid_guid when we opened it,
2136 	 * sync it out now.
2137 	 */
2138 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
2139 	ds->ds_phys->ds_fsid_guid = ds->ds_fsid_guid;
2140 
2141 	dsl_dir_dirty(ds->ds_dir, tx);
2142 	dmu_objset_sync(ds->ds_objset, zio, tx);
2143 }
2144 
2145 void
2146 dsl_dataset_stats(dsl_dataset_t *ds, nvlist_t *nv)
2147 {
2148 	uint64_t refd, avail, uobjs, aobjs;
2149 
2150 	dsl_dir_stats(ds->ds_dir, nv);
2151 
2152 	dsl_dataset_space(ds, &refd, &avail, &uobjs, &aobjs);
2153 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_AVAILABLE, avail);
2154 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFERENCED, refd);
2155 
2156 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATION,
2157 	    ds->ds_phys->ds_creation_time);
2158 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATETXG,
2159 	    ds->ds_phys->ds_creation_txg);
2160 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFQUOTA,
2161 	    ds->ds_quota);
2162 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRESERVATION,
2163 	    ds->ds_reserved);
2164 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_GUID,
2165 	    ds->ds_phys->ds_guid);
2166 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_UNIQUE,
2167 	    ds->ds_phys->ds_unique_bytes);
2168 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_OBJSETID,
2169 	    ds->ds_object);
2170 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USERREFS,
2171 	    ds->ds_userrefs);
2172 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_DEFER_DESTROY,
2173 	    DS_IS_DEFER_DESTROY(ds) ? 1 : 0);
2174 
2175 	if (ds->ds_phys->ds_next_snap_obj) {
2176 		/*
2177 		 * This is a snapshot; override the dd's space used with
2178 		 * our unique space and compression ratio.
2179 		 */
2180 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USED,
2181 		    ds->ds_phys->ds_unique_bytes);
2182 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_COMPRESSRATIO,
2183 		    ds->ds_phys->ds_compressed_bytes == 0 ? 100 :
2184 		    (ds->ds_phys->ds_uncompressed_bytes * 100 /
2185 		    ds->ds_phys->ds_compressed_bytes));
2186 	}
2187 }
2188 
2189 void
2190 dsl_dataset_fast_stat(dsl_dataset_t *ds, dmu_objset_stats_t *stat)
2191 {
2192 	stat->dds_creation_txg = ds->ds_phys->ds_creation_txg;
2193 	stat->dds_inconsistent = ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT;
2194 	stat->dds_guid = ds->ds_phys->ds_guid;
2195 	if (ds->ds_phys->ds_next_snap_obj) {
2196 		stat->dds_is_snapshot = B_TRUE;
2197 		stat->dds_num_clones = ds->ds_phys->ds_num_children - 1;
2198 	} else {
2199 		stat->dds_is_snapshot = B_FALSE;
2200 		stat->dds_num_clones = 0;
2201 	}
2202 
2203 	/* clone origin is really a dsl_dir thing... */
2204 	rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2205 	if (dsl_dir_is_clone(ds->ds_dir)) {
2206 		dsl_dataset_t *ods;
2207 
2208 		VERIFY(0 == dsl_dataset_get_ref(ds->ds_dir->dd_pool,
2209 		    ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &ods));
2210 		dsl_dataset_name(ods, stat->dds_origin);
2211 		dsl_dataset_drop_ref(ods, FTAG);
2212 	} else {
2213 		stat->dds_origin[0] = '\0';
2214 	}
2215 	rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2216 }
2217 
2218 uint64_t
2219 dsl_dataset_fsid_guid(dsl_dataset_t *ds)
2220 {
2221 	return (ds->ds_fsid_guid);
2222 }
2223 
2224 void
2225 dsl_dataset_space(dsl_dataset_t *ds,
2226     uint64_t *refdbytesp, uint64_t *availbytesp,
2227     uint64_t *usedobjsp, uint64_t *availobjsp)
2228 {
2229 	*refdbytesp = ds->ds_phys->ds_used_bytes;
2230 	*availbytesp = dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE);
2231 	if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes)
2232 		*availbytesp += ds->ds_reserved - ds->ds_phys->ds_unique_bytes;
2233 	if (ds->ds_quota != 0) {
2234 		/*
2235 		 * Adjust available bytes according to refquota
2236 		 */
2237 		if (*refdbytesp < ds->ds_quota)
2238 			*availbytesp = MIN(*availbytesp,
2239 			    ds->ds_quota - *refdbytesp);
2240 		else
2241 			*availbytesp = 0;
2242 	}
2243 	*usedobjsp = ds->ds_phys->ds_bp.blk_fill;
2244 	*availobjsp = DN_MAX_OBJECT - *usedobjsp;
2245 }
2246 
2247 boolean_t
2248 dsl_dataset_modified_since_lastsnap(dsl_dataset_t *ds)
2249 {
2250 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
2251 
2252 	ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
2253 	    dsl_pool_sync_context(dp));
2254 	if (ds->ds_prev == NULL)
2255 		return (B_FALSE);
2256 	if (ds->ds_phys->ds_bp.blk_birth >
2257 	    ds->ds_prev->ds_phys->ds_creation_txg) {
2258 		objset_t *os, *os_prev;
2259 		/*
2260 		 * It may be that only the ZIL differs, because it was
2261 		 * reset in the head.  Don't count that as being
2262 		 * modified.
2263 		 */
2264 		if (dmu_objset_from_ds(ds, &os) != 0)
2265 			return (B_TRUE);
2266 		if (dmu_objset_from_ds(ds->ds_prev, &os_prev) != 0)
2267 			return (B_TRUE);
2268 		return (bcmp(&os->os_phys->os_meta_dnode,
2269 		    &os_prev->os_phys->os_meta_dnode,
2270 		    sizeof (os->os_phys->os_meta_dnode)) != 0);
2271 	}
2272 	return (B_FALSE);
2273 }
2274 
2275 /* ARGSUSED */
2276 static int
2277 dsl_dataset_snapshot_rename_check(void *arg1, void *arg2, dmu_tx_t *tx)
2278 {
2279 	dsl_dataset_t *ds = arg1;
2280 	char *newsnapname = arg2;
2281 	dsl_dir_t *dd = ds->ds_dir;
2282 	dsl_dataset_t *hds;
2283 	uint64_t val;
2284 	int err;
2285 
2286 	err = dsl_dataset_hold_obj(dd->dd_pool,
2287 	    dd->dd_phys->dd_head_dataset_obj, FTAG, &hds);
2288 	if (err)
2289 		return (err);
2290 
2291 	/* new name better not be in use */
2292 	err = dsl_dataset_snap_lookup(hds, newsnapname, &val);
2293 	dsl_dataset_rele(hds, FTAG);
2294 
2295 	if (err == 0)
2296 		err = EEXIST;
2297 	else if (err == ENOENT)
2298 		err = 0;
2299 
2300 	/* dataset name + 1 for the "@" + the new snapshot name must fit */
2301 	if (dsl_dir_namelen(ds->ds_dir) + 1 + strlen(newsnapname) >= MAXNAMELEN)
2302 		err = ENAMETOOLONG;
2303 
2304 	return (err);
2305 }
2306 
2307 static void
2308 dsl_dataset_snapshot_rename_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2309 {
2310 	dsl_dataset_t *ds = arg1;
2311 	const char *newsnapname = arg2;
2312 	dsl_dir_t *dd = ds->ds_dir;
2313 	objset_t *mos = dd->dd_pool->dp_meta_objset;
2314 	dsl_dataset_t *hds;
2315 	int err;
2316 
2317 	ASSERT(ds->ds_phys->ds_next_snap_obj != 0);
2318 
2319 	VERIFY(0 == dsl_dataset_hold_obj(dd->dd_pool,
2320 	    dd->dd_phys->dd_head_dataset_obj, FTAG, &hds));
2321 
2322 	VERIFY(0 == dsl_dataset_get_snapname(ds));
2323 	err = dsl_dataset_snap_remove(hds, ds->ds_snapname, tx);
2324 	ASSERT3U(err, ==, 0);
2325 	mutex_enter(&ds->ds_lock);
2326 	(void) strcpy(ds->ds_snapname, newsnapname);
2327 	mutex_exit(&ds->ds_lock);
2328 	err = zap_add(mos, hds->ds_phys->ds_snapnames_zapobj,
2329 	    ds->ds_snapname, 8, 1, &ds->ds_object, tx);
2330 	ASSERT3U(err, ==, 0);
2331 
2332 	spa_history_log_internal(LOG_DS_RENAME, dd->dd_pool->dp_spa, tx,
2333 	    "dataset = %llu", ds->ds_object);
2334 	dsl_dataset_rele(hds, FTAG);
2335 }
2336 
2337 struct renamesnaparg {
2338 	dsl_sync_task_group_t *dstg;
2339 	char failed[MAXPATHLEN];
2340 	char *oldsnap;
2341 	char *newsnap;
2342 };
2343 
2344 static int
2345 dsl_snapshot_rename_one(const char *name, void *arg)
2346 {
2347 	struct renamesnaparg *ra = arg;
2348 	dsl_dataset_t *ds = NULL;
2349 	char *snapname;
2350 	int err;
2351 
2352 	snapname = kmem_asprintf("%s@%s", name, ra->oldsnap);
2353 	(void) strlcpy(ra->failed, snapname, sizeof (ra->failed));
2354 
2355 	/*
2356 	 * For recursive snapshot renames the parent won't be changing
2357 	 * so we just pass name for both the to/from argument.
2358 	 */
2359 	err = zfs_secpolicy_rename_perms(snapname, snapname, CRED());
2360 	if (err != 0) {
2361 		strfree(snapname);
2362 		return (err == ENOENT ? 0 : err);
2363 	}
2364 
2365 #ifdef _KERNEL
2366 	/*
2367 	 * For all filesystems undergoing rename, we'll need to unmount it.
2368 	 */
2369 	(void) zfs_unmount_snap(snapname, NULL);
2370 #endif
2371 	err = dsl_dataset_hold(snapname, ra->dstg, &ds);
2372 	strfree(snapname);
2373 	if (err != 0)
2374 		return (err == ENOENT ? 0 : err);
2375 
2376 	dsl_sync_task_create(ra->dstg, dsl_dataset_snapshot_rename_check,
2377 	    dsl_dataset_snapshot_rename_sync, ds, ra->newsnap, 0);
2378 
2379 	return (0);
2380 }
2381 
2382 static int
2383 dsl_recursive_rename(char *oldname, const char *newname)
2384 {
2385 	int err;
2386 	struct renamesnaparg *ra;
2387 	dsl_sync_task_t *dst;
2388 	spa_t *spa;
2389 	char *cp, *fsname = spa_strdup(oldname);
2390 	int len = strlen(oldname) + 1;
2391 
2392 	/* truncate the snapshot name to get the fsname */
2393 	cp = strchr(fsname, '@');
2394 	*cp = '\0';
2395 
2396 	err = spa_open(fsname, &spa, FTAG);
2397 	if (err) {
2398 		kmem_free(fsname, len);
2399 		return (err);
2400 	}
2401 	ra = kmem_alloc(sizeof (struct renamesnaparg), KM_SLEEP);
2402 	ra->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
2403 
2404 	ra->oldsnap = strchr(oldname, '@') + 1;
2405 	ra->newsnap = strchr(newname, '@') + 1;
2406 	*ra->failed = '\0';
2407 
2408 	err = dmu_objset_find(fsname, dsl_snapshot_rename_one, ra,
2409 	    DS_FIND_CHILDREN);
2410 	kmem_free(fsname, len);
2411 
2412 	if (err == 0) {
2413 		err = dsl_sync_task_group_wait(ra->dstg);
2414 	}
2415 
2416 	for (dst = list_head(&ra->dstg->dstg_tasks); dst;
2417 	    dst = list_next(&ra->dstg->dstg_tasks, dst)) {
2418 		dsl_dataset_t *ds = dst->dst_arg1;
2419 		if (dst->dst_err) {
2420 			dsl_dir_name(ds->ds_dir, ra->failed);
2421 			(void) strlcat(ra->failed, "@", sizeof (ra->failed));
2422 			(void) strlcat(ra->failed, ra->newsnap,
2423 			    sizeof (ra->failed));
2424 		}
2425 		dsl_dataset_rele(ds, ra->dstg);
2426 	}
2427 
2428 	if (err)
2429 		(void) strlcpy(oldname, ra->failed, sizeof (ra->failed));
2430 
2431 	dsl_sync_task_group_destroy(ra->dstg);
2432 	kmem_free(ra, sizeof (struct renamesnaparg));
2433 	spa_close(spa, FTAG);
2434 	return (err);
2435 }
2436 
2437 static int
2438 dsl_valid_rename(const char *oldname, void *arg)
2439 {
2440 	int delta = *(int *)arg;
2441 
2442 	if (strlen(oldname) + delta >= MAXNAMELEN)
2443 		return (ENAMETOOLONG);
2444 
2445 	return (0);
2446 }
2447 
2448 #pragma weak dmu_objset_rename = dsl_dataset_rename
2449 int
2450 dsl_dataset_rename(char *oldname, const char *newname, boolean_t recursive)
2451 {
2452 	dsl_dir_t *dd;
2453 	dsl_dataset_t *ds;
2454 	const char *tail;
2455 	int err;
2456 
2457 	err = dsl_dir_open(oldname, FTAG, &dd, &tail);
2458 	if (err)
2459 		return (err);
2460 
2461 	if (tail == NULL) {
2462 		int delta = strlen(newname) - strlen(oldname);
2463 
2464 		/* if we're growing, validate child name lengths */
2465 		if (delta > 0)
2466 			err = dmu_objset_find(oldname, dsl_valid_rename,
2467 			    &delta, DS_FIND_CHILDREN | DS_FIND_SNAPSHOTS);
2468 
2469 		if (err == 0)
2470 			err = dsl_dir_rename(dd, newname);
2471 		dsl_dir_close(dd, FTAG);
2472 		return (err);
2473 	}
2474 
2475 	if (tail[0] != '@') {
2476 		/* the name ended in a nonexistent component */
2477 		dsl_dir_close(dd, FTAG);
2478 		return (ENOENT);
2479 	}
2480 
2481 	dsl_dir_close(dd, FTAG);
2482 
2483 	/* new name must be snapshot in same filesystem */
2484 	tail = strchr(newname, '@');
2485 	if (tail == NULL)
2486 		return (EINVAL);
2487 	tail++;
2488 	if (strncmp(oldname, newname, tail - newname) != 0)
2489 		return (EXDEV);
2490 
2491 	if (recursive) {
2492 		err = dsl_recursive_rename(oldname, newname);
2493 	} else {
2494 		err = dsl_dataset_hold(oldname, FTAG, &ds);
2495 		if (err)
2496 			return (err);
2497 
2498 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
2499 		    dsl_dataset_snapshot_rename_check,
2500 		    dsl_dataset_snapshot_rename_sync, ds, (char *)tail, 1);
2501 
2502 		dsl_dataset_rele(ds, FTAG);
2503 	}
2504 
2505 	return (err);
2506 }
2507 
2508 struct promotenode {
2509 	list_node_t link;
2510 	dsl_dataset_t *ds;
2511 };
2512 
2513 struct promotearg {
2514 	list_t shared_snaps, origin_snaps, clone_snaps;
2515 	dsl_dataset_t *origin_origin;
2516 	uint64_t used, comp, uncomp, unique, cloneusedsnap, originusedsnap;
2517 	char *err_ds;
2518 };
2519 
2520 static int snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep);
2521 static boolean_t snaplist_unstable(list_t *l);
2522 
2523 static int
2524 dsl_dataset_promote_check(void *arg1, void *arg2, dmu_tx_t *tx)
2525 {
2526 	dsl_dataset_t *hds = arg1;
2527 	struct promotearg *pa = arg2;
2528 	struct promotenode *snap = list_head(&pa->shared_snaps);
2529 	dsl_dataset_t *origin_ds = snap->ds;
2530 	int err;
2531 	uint64_t unused;
2532 
2533 	/* Check that it is a real clone */
2534 	if (!dsl_dir_is_clone(hds->ds_dir))
2535 		return (EINVAL);
2536 
2537 	/* Since this is so expensive, don't do the preliminary check */
2538 	if (!dmu_tx_is_syncing(tx))
2539 		return (0);
2540 
2541 	if (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE)
2542 		return (EXDEV);
2543 
2544 	/* compute origin's new unique space */
2545 	snap = list_tail(&pa->clone_snaps);
2546 	ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2547 	dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2548 	    origin_ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
2549 	    &pa->unique, &unused, &unused);
2550 
2551 	/*
2552 	 * Walk the snapshots that we are moving
2553 	 *
2554 	 * Compute space to transfer.  Consider the incremental changes
2555 	 * to used for each snapshot:
2556 	 * (my used) = (prev's used) + (blocks born) - (blocks killed)
2557 	 * So each snapshot gave birth to:
2558 	 * (blocks born) = (my used) - (prev's used) + (blocks killed)
2559 	 * So a sequence would look like:
2560 	 * (uN - u(N-1) + kN) + ... + (u1 - u0 + k1) + (u0 - 0 + k0)
2561 	 * Which simplifies to:
2562 	 * uN + kN + kN-1 + ... + k1 + k0
2563 	 * Note however, if we stop before we reach the ORIGIN we get:
2564 	 * uN + kN + kN-1 + ... + kM - uM-1
2565 	 */
2566 	pa->used = origin_ds->ds_phys->ds_used_bytes;
2567 	pa->comp = origin_ds->ds_phys->ds_compressed_bytes;
2568 	pa->uncomp = origin_ds->ds_phys->ds_uncompressed_bytes;
2569 	for (snap = list_head(&pa->shared_snaps); snap;
2570 	    snap = list_next(&pa->shared_snaps, snap)) {
2571 		uint64_t val, dlused, dlcomp, dluncomp;
2572 		dsl_dataset_t *ds = snap->ds;
2573 
2574 		/* Check that the snapshot name does not conflict */
2575 		VERIFY(0 == dsl_dataset_get_snapname(ds));
2576 		err = dsl_dataset_snap_lookup(hds, ds->ds_snapname, &val);
2577 		if (err == 0) {
2578 			err = EEXIST;
2579 			goto out;
2580 		}
2581 		if (err != ENOENT)
2582 			goto out;
2583 
2584 		/* The very first snapshot does not have a deadlist */
2585 		if (ds->ds_phys->ds_prev_snap_obj == 0)
2586 			continue;
2587 
2588 		dsl_deadlist_space(&ds->ds_deadlist,
2589 		    &dlused, &dlcomp, &dluncomp);
2590 		pa->used += dlused;
2591 		pa->comp += dlcomp;
2592 		pa->uncomp += dluncomp;
2593 	}
2594 
2595 	/*
2596 	 * If we are a clone of a clone then we never reached ORIGIN,
2597 	 * so we need to subtract out the clone origin's used space.
2598 	 */
2599 	if (pa->origin_origin) {
2600 		pa->used -= pa->origin_origin->ds_phys->ds_used_bytes;
2601 		pa->comp -= pa->origin_origin->ds_phys->ds_compressed_bytes;
2602 		pa->uncomp -= pa->origin_origin->ds_phys->ds_uncompressed_bytes;
2603 	}
2604 
2605 	/* Check that there is enough space here */
2606 	err = dsl_dir_transfer_possible(origin_ds->ds_dir, hds->ds_dir,
2607 	    pa->used);
2608 	if (err)
2609 		return (err);
2610 
2611 	/*
2612 	 * Compute the amounts of space that will be used by snapshots
2613 	 * after the promotion (for both origin and clone).  For each,
2614 	 * it is the amount of space that will be on all of their
2615 	 * deadlists (that was not born before their new origin).
2616 	 */
2617 	if (hds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2618 		uint64_t space;
2619 
2620 		/*
2621 		 * Note, typically this will not be a clone of a clone,
2622 		 * so dd_origin_txg will be < TXG_INITIAL, so
2623 		 * these snaplist_space() -> dsl_deadlist_space_range()
2624 		 * calls will be fast because they do not have to
2625 		 * iterate over all bps.
2626 		 */
2627 		snap = list_head(&pa->origin_snaps);
2628 		err = snaplist_space(&pa->shared_snaps,
2629 		    snap->ds->ds_dir->dd_origin_txg, &pa->cloneusedsnap);
2630 		if (err)
2631 			return (err);
2632 
2633 		err = snaplist_space(&pa->clone_snaps,
2634 		    snap->ds->ds_dir->dd_origin_txg, &space);
2635 		if (err)
2636 			return (err);
2637 		pa->cloneusedsnap += space;
2638 	}
2639 	if (origin_ds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2640 		err = snaplist_space(&pa->origin_snaps,
2641 		    origin_ds->ds_phys->ds_creation_txg, &pa->originusedsnap);
2642 		if (err)
2643 			return (err);
2644 	}
2645 
2646 	return (0);
2647 out:
2648 	pa->err_ds =  snap->ds->ds_snapname;
2649 	return (err);
2650 }
2651 
2652 static void
2653 dsl_dataset_promote_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2654 {
2655 	dsl_dataset_t *hds = arg1;
2656 	struct promotearg *pa = arg2;
2657 	struct promotenode *snap = list_head(&pa->shared_snaps);
2658 	dsl_dataset_t *origin_ds = snap->ds;
2659 	dsl_dataset_t *origin_head;
2660 	dsl_dir_t *dd = hds->ds_dir;
2661 	dsl_pool_t *dp = hds->ds_dir->dd_pool;
2662 	dsl_dir_t *odd = NULL;
2663 	uint64_t oldnext_obj;
2664 	int64_t delta;
2665 
2666 	ASSERT(0 == (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE));
2667 
2668 	snap = list_head(&pa->origin_snaps);
2669 	origin_head = snap->ds;
2670 
2671 	/*
2672 	 * We need to explicitly open odd, since origin_ds's dd will be
2673 	 * changing.
2674 	 */
2675 	VERIFY(0 == dsl_dir_open_obj(dp, origin_ds->ds_dir->dd_object,
2676 	    NULL, FTAG, &odd));
2677 
2678 	/* change origin's next snap */
2679 	dmu_buf_will_dirty(origin_ds->ds_dbuf, tx);
2680 	oldnext_obj = origin_ds->ds_phys->ds_next_snap_obj;
2681 	snap = list_tail(&pa->clone_snaps);
2682 	ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2683 	origin_ds->ds_phys->ds_next_snap_obj = snap->ds->ds_object;
2684 
2685 	/* change the origin's next clone */
2686 	if (origin_ds->ds_phys->ds_next_clones_obj) {
2687 		remove_from_next_clones(origin_ds, snap->ds->ds_object, tx);
2688 		VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2689 		    origin_ds->ds_phys->ds_next_clones_obj,
2690 		    oldnext_obj, tx));
2691 	}
2692 
2693 	/* change origin */
2694 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
2695 	ASSERT3U(dd->dd_phys->dd_origin_obj, ==, origin_ds->ds_object);
2696 	dd->dd_phys->dd_origin_obj = odd->dd_phys->dd_origin_obj;
2697 	dd->dd_origin_txg = origin_head->ds_dir->dd_origin_txg;
2698 	dmu_buf_will_dirty(odd->dd_dbuf, tx);
2699 	odd->dd_phys->dd_origin_obj = origin_ds->ds_object;
2700 	origin_head->ds_dir->dd_origin_txg =
2701 	    origin_ds->ds_phys->ds_creation_txg;
2702 
2703 	/* change dd_clone entries */
2704 	if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2705 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2706 		    odd->dd_phys->dd_clones, hds->ds_object, tx));
2707 		VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2708 		    pa->origin_origin->ds_dir->dd_phys->dd_clones,
2709 		    hds->ds_object, tx));
2710 
2711 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2712 		    pa->origin_origin->ds_dir->dd_phys->dd_clones,
2713 		    origin_head->ds_object, tx));
2714 		if (dd->dd_phys->dd_clones == 0) {
2715 			dd->dd_phys->dd_clones = zap_create(dp->dp_meta_objset,
2716 			    DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
2717 		}
2718 		VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2719 		    dd->dd_phys->dd_clones, origin_head->ds_object, tx));
2720 
2721 	}
2722 
2723 	/* move snapshots to this dir */
2724 	for (snap = list_head(&pa->shared_snaps); snap;
2725 	    snap = list_next(&pa->shared_snaps, snap)) {
2726 		dsl_dataset_t *ds = snap->ds;
2727 
2728 		/* unregister props as dsl_dir is changing */
2729 		if (ds->ds_objset) {
2730 			dmu_objset_evict(ds->ds_objset);
2731 			ds->ds_objset = NULL;
2732 		}
2733 		/* move snap name entry */
2734 		VERIFY(0 == dsl_dataset_get_snapname(ds));
2735 		VERIFY(0 == dsl_dataset_snap_remove(origin_head,
2736 		    ds->ds_snapname, tx));
2737 		VERIFY(0 == zap_add(dp->dp_meta_objset,
2738 		    hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
2739 		    8, 1, &ds->ds_object, tx));
2740 
2741 		/* change containing dsl_dir */
2742 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
2743 		ASSERT3U(ds->ds_phys->ds_dir_obj, ==, odd->dd_object);
2744 		ds->ds_phys->ds_dir_obj = dd->dd_object;
2745 		ASSERT3P(ds->ds_dir, ==, odd);
2746 		dsl_dir_close(ds->ds_dir, ds);
2747 		VERIFY(0 == dsl_dir_open_obj(dp, dd->dd_object,
2748 		    NULL, ds, &ds->ds_dir));
2749 
2750 		/* move any clone references */
2751 		if (ds->ds_phys->ds_next_clones_obj &&
2752 		    spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2753 			zap_cursor_t zc;
2754 			zap_attribute_t za;
2755 
2756 			for (zap_cursor_init(&zc, dp->dp_meta_objset,
2757 			    ds->ds_phys->ds_next_clones_obj);
2758 			    zap_cursor_retrieve(&zc, &za) == 0;
2759 			    zap_cursor_advance(&zc)) {
2760 				dsl_dataset_t *cnds;
2761 				uint64_t o;
2762 
2763 				if (za.za_first_integer == oldnext_obj) {
2764 					/*
2765 					 * We've already moved the
2766 					 * origin's reference.
2767 					 */
2768 					continue;
2769 				}
2770 
2771 				VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
2772 				    za.za_first_integer, FTAG, &cnds));
2773 				o = cnds->ds_dir->dd_phys->dd_head_dataset_obj;
2774 
2775 				VERIFY3U(zap_remove_int(dp->dp_meta_objset,
2776 				    odd->dd_phys->dd_clones, o, tx), ==, 0);
2777 				VERIFY3U(zap_add_int(dp->dp_meta_objset,
2778 				    dd->dd_phys->dd_clones, o, tx), ==, 0);
2779 				dsl_dataset_rele(cnds, FTAG);
2780 			}
2781 			zap_cursor_fini(&zc);
2782 		}
2783 
2784 		ASSERT3U(dsl_prop_numcb(ds), ==, 0);
2785 	}
2786 
2787 	/*
2788 	 * Change space accounting.
2789 	 * Note, pa->*usedsnap and dd_used_breakdown[SNAP] will either
2790 	 * both be valid, or both be 0 (resulting in delta == 0).  This
2791 	 * is true for each of {clone,origin} independently.
2792 	 */
2793 
2794 	delta = pa->cloneusedsnap -
2795 	    dd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2796 	ASSERT3S(delta, >=, 0);
2797 	ASSERT3U(pa->used, >=, delta);
2798 	dsl_dir_diduse_space(dd, DD_USED_SNAP, delta, 0, 0, tx);
2799 	dsl_dir_diduse_space(dd, DD_USED_HEAD,
2800 	    pa->used - delta, pa->comp, pa->uncomp, tx);
2801 
2802 	delta = pa->originusedsnap -
2803 	    odd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2804 	ASSERT3S(delta, <=, 0);
2805 	ASSERT3U(pa->used, >=, -delta);
2806 	dsl_dir_diduse_space(odd, DD_USED_SNAP, delta, 0, 0, tx);
2807 	dsl_dir_diduse_space(odd, DD_USED_HEAD,
2808 	    -pa->used - delta, -pa->comp, -pa->uncomp, tx);
2809 
2810 	origin_ds->ds_phys->ds_unique_bytes = pa->unique;
2811 
2812 	/* log history record */
2813 	spa_history_log_internal(LOG_DS_PROMOTE, dd->dd_pool->dp_spa, tx,
2814 	    "dataset = %llu", hds->ds_object);
2815 
2816 	dsl_dir_close(odd, FTAG);
2817 }
2818 
2819 static char *snaplist_tag = "snaplist";
2820 /*
2821  * Make a list of dsl_dataset_t's for the snapshots between first_obj
2822  * (exclusive) and last_obj (inclusive).  The list will be in reverse
2823  * order (last_obj will be the list_head()).  If first_obj == 0, do all
2824  * snapshots back to this dataset's origin.
2825  */
2826 static int
2827 snaplist_make(dsl_pool_t *dp, boolean_t own,
2828     uint64_t first_obj, uint64_t last_obj, list_t *l)
2829 {
2830 	uint64_t obj = last_obj;
2831 
2832 	ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock));
2833 
2834 	list_create(l, sizeof (struct promotenode),
2835 	    offsetof(struct promotenode, link));
2836 
2837 	while (obj != first_obj) {
2838 		dsl_dataset_t *ds;
2839 		struct promotenode *snap;
2840 		int err;
2841 
2842 		if (own) {
2843 			err = dsl_dataset_own_obj(dp, obj,
2844 			    0, snaplist_tag, &ds);
2845 			if (err == 0)
2846 				dsl_dataset_make_exclusive(ds, snaplist_tag);
2847 		} else {
2848 			err = dsl_dataset_hold_obj(dp, obj, snaplist_tag, &ds);
2849 		}
2850 		if (err == ENOENT) {
2851 			/* lost race with snapshot destroy */
2852 			struct promotenode *last = list_tail(l);
2853 			ASSERT(obj != last->ds->ds_phys->ds_prev_snap_obj);
2854 			obj = last->ds->ds_phys->ds_prev_snap_obj;
2855 			continue;
2856 		} else if (err) {
2857 			return (err);
2858 		}
2859 
2860 		if (first_obj == 0)
2861 			first_obj = ds->ds_dir->dd_phys->dd_origin_obj;
2862 
2863 		snap = kmem_alloc(sizeof (struct promotenode), KM_SLEEP);
2864 		snap->ds = ds;
2865 		list_insert_tail(l, snap);
2866 		obj = ds->ds_phys->ds_prev_snap_obj;
2867 	}
2868 
2869 	return (0);
2870 }
2871 
2872 static int
2873 snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep)
2874 {
2875 	struct promotenode *snap;
2876 
2877 	*spacep = 0;
2878 	for (snap = list_head(l); snap; snap = list_next(l, snap)) {
2879 		uint64_t used, comp, uncomp;
2880 		dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2881 		    mintxg, UINT64_MAX, &used, &comp, &uncomp);
2882 		*spacep += used;
2883 	}
2884 	return (0);
2885 }
2886 
2887 static void
2888 snaplist_destroy(list_t *l, boolean_t own)
2889 {
2890 	struct promotenode *snap;
2891 
2892 	if (!l || !list_link_active(&l->list_head))
2893 		return;
2894 
2895 	while ((snap = list_tail(l)) != NULL) {
2896 		list_remove(l, snap);
2897 		if (own)
2898 			dsl_dataset_disown(snap->ds, snaplist_tag);
2899 		else
2900 			dsl_dataset_rele(snap->ds, snaplist_tag);
2901 		kmem_free(snap, sizeof (struct promotenode));
2902 	}
2903 	list_destroy(l);
2904 }
2905 
2906 /*
2907  * Promote a clone.  Nomenclature note:
2908  * "clone" or "cds": the original clone which is being promoted
2909  * "origin" or "ods": the snapshot which is originally clone's origin
2910  * "origin head" or "ohds": the dataset which is the head
2911  * (filesystem/volume) for the origin
2912  * "origin origin": the origin of the origin's filesystem (typically
2913  * NULL, indicating that the clone is not a clone of a clone).
2914  */
2915 int
2916 dsl_dataset_promote(const char *name, char *conflsnap)
2917 {
2918 	dsl_dataset_t *ds;
2919 	dsl_dir_t *dd;
2920 	dsl_pool_t *dp;
2921 	dmu_object_info_t doi;
2922 	struct promotearg pa = { 0 };
2923 	struct promotenode *snap;
2924 	int err;
2925 
2926 	err = dsl_dataset_hold(name, FTAG, &ds);
2927 	if (err)
2928 		return (err);
2929 	dd = ds->ds_dir;
2930 	dp = dd->dd_pool;
2931 
2932 	err = dmu_object_info(dp->dp_meta_objset,
2933 	    ds->ds_phys->ds_snapnames_zapobj, &doi);
2934 	if (err) {
2935 		dsl_dataset_rele(ds, FTAG);
2936 		return (err);
2937 	}
2938 
2939 	if (dsl_dataset_is_snapshot(ds) || dd->dd_phys->dd_origin_obj == 0) {
2940 		dsl_dataset_rele(ds, FTAG);
2941 		return (EINVAL);
2942 	}
2943 
2944 	/*
2945 	 * We are going to inherit all the snapshots taken before our
2946 	 * origin (i.e., our new origin will be our parent's origin).
2947 	 * Take ownership of them so that we can rename them into our
2948 	 * namespace.
2949 	 */
2950 	rw_enter(&dp->dp_config_rwlock, RW_READER);
2951 
2952 	err = snaplist_make(dp, B_TRUE, 0, dd->dd_phys->dd_origin_obj,
2953 	    &pa.shared_snaps);
2954 	if (err != 0)
2955 		goto out;
2956 
2957 	err = snaplist_make(dp, B_FALSE, 0, ds->ds_object, &pa.clone_snaps);
2958 	if (err != 0)
2959 		goto out;
2960 
2961 	snap = list_head(&pa.shared_snaps);
2962 	ASSERT3U(snap->ds->ds_object, ==, dd->dd_phys->dd_origin_obj);
2963 	err = snaplist_make(dp, B_FALSE, dd->dd_phys->dd_origin_obj,
2964 	    snap->ds->ds_dir->dd_phys->dd_head_dataset_obj, &pa.origin_snaps);
2965 	if (err != 0)
2966 		goto out;
2967 
2968 	if (snap->ds->ds_dir->dd_phys->dd_origin_obj != 0) {
2969 		err = dsl_dataset_hold_obj(dp,
2970 		    snap->ds->ds_dir->dd_phys->dd_origin_obj,
2971 		    FTAG, &pa.origin_origin);
2972 		if (err != 0)
2973 			goto out;
2974 	}
2975 
2976 out:
2977 	rw_exit(&dp->dp_config_rwlock);
2978 
2979 	/*
2980 	 * Add in 128x the snapnames zapobj size, since we will be moving
2981 	 * a bunch of snapnames to the promoted ds, and dirtying their
2982 	 * bonus buffers.
2983 	 */
2984 	if (err == 0) {
2985 		err = dsl_sync_task_do(dp, dsl_dataset_promote_check,
2986 		    dsl_dataset_promote_sync, ds, &pa,
2987 		    2 + 2 * doi.doi_physical_blocks_512);
2988 		if (err && pa.err_ds && conflsnap)
2989 			(void) strncpy(conflsnap, pa.err_ds, MAXNAMELEN);
2990 	}
2991 
2992 	snaplist_destroy(&pa.shared_snaps, B_TRUE);
2993 	snaplist_destroy(&pa.clone_snaps, B_FALSE);
2994 	snaplist_destroy(&pa.origin_snaps, B_FALSE);
2995 	if (pa.origin_origin)
2996 		dsl_dataset_rele(pa.origin_origin, FTAG);
2997 	dsl_dataset_rele(ds, FTAG);
2998 	return (err);
2999 }
3000 
3001 struct cloneswaparg {
3002 	dsl_dataset_t *cds; /* clone dataset */
3003 	dsl_dataset_t *ohds; /* origin's head dataset */
3004 	boolean_t force;
3005 	int64_t unused_refres_delta; /* change in unconsumed refreservation */
3006 };
3007 
3008 /* ARGSUSED */
3009 static int
3010 dsl_dataset_clone_swap_check(void *arg1, void *arg2, dmu_tx_t *tx)
3011 {
3012 	struct cloneswaparg *csa = arg1;
3013 
3014 	/* they should both be heads */
3015 	if (dsl_dataset_is_snapshot(csa->cds) ||
3016 	    dsl_dataset_is_snapshot(csa->ohds))
3017 		return (EINVAL);
3018 
3019 	/* the branch point should be just before them */
3020 	if (csa->cds->ds_prev != csa->ohds->ds_prev)
3021 		return (EINVAL);
3022 
3023 	/* cds should be the clone (unless they are unrelated) */
3024 	if (csa->cds->ds_prev != NULL &&
3025 	    csa->cds->ds_prev != csa->cds->ds_dir->dd_pool->dp_origin_snap &&
3026 	    csa->ohds->ds_object !=
3027 	    csa->cds->ds_prev->ds_phys->ds_next_snap_obj)
3028 		return (EINVAL);
3029 
3030 	/* the clone should be a child of the origin */
3031 	if (csa->cds->ds_dir->dd_parent != csa->ohds->ds_dir)
3032 		return (EINVAL);
3033 
3034 	/* ohds shouldn't be modified unless 'force' */
3035 	if (!csa->force && dsl_dataset_modified_since_lastsnap(csa->ohds))
3036 		return (ETXTBSY);
3037 
3038 	/* adjust amount of any unconsumed refreservation */
3039 	csa->unused_refres_delta =
3040 	    (int64_t)MIN(csa->ohds->ds_reserved,
3041 	    csa->ohds->ds_phys->ds_unique_bytes) -
3042 	    (int64_t)MIN(csa->ohds->ds_reserved,
3043 	    csa->cds->ds_phys->ds_unique_bytes);
3044 
3045 	if (csa->unused_refres_delta > 0 &&
3046 	    csa->unused_refres_delta >
3047 	    dsl_dir_space_available(csa->ohds->ds_dir, NULL, 0, TRUE))
3048 		return (ENOSPC);
3049 
3050 	if (csa->ohds->ds_quota != 0 &&
3051 	    csa->cds->ds_phys->ds_unique_bytes > csa->ohds->ds_quota)
3052 		return (EDQUOT);
3053 
3054 	return (0);
3055 }
3056 
3057 /* ARGSUSED */
3058 static void
3059 dsl_dataset_clone_swap_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3060 {
3061 	struct cloneswaparg *csa = arg1;
3062 	dsl_pool_t *dp = csa->cds->ds_dir->dd_pool;
3063 
3064 	ASSERT(csa->cds->ds_reserved == 0);
3065 	ASSERT(csa->ohds->ds_quota == 0 ||
3066 	    csa->cds->ds_phys->ds_unique_bytes <= csa->ohds->ds_quota);
3067 
3068 	dmu_buf_will_dirty(csa->cds->ds_dbuf, tx);
3069 	dmu_buf_will_dirty(csa->ohds->ds_dbuf, tx);
3070 
3071 	if (csa->cds->ds_objset != NULL) {
3072 		dmu_objset_evict(csa->cds->ds_objset);
3073 		csa->cds->ds_objset = NULL;
3074 	}
3075 
3076 	if (csa->ohds->ds_objset != NULL) {
3077 		dmu_objset_evict(csa->ohds->ds_objset);
3078 		csa->ohds->ds_objset = NULL;
3079 	}
3080 
3081 	/*
3082 	 * Reset origin's unique bytes, if it exists.
3083 	 */
3084 	if (csa->cds->ds_prev) {
3085 		dsl_dataset_t *origin = csa->cds->ds_prev;
3086 		uint64_t comp, uncomp;
3087 
3088 		dmu_buf_will_dirty(origin->ds_dbuf, tx);
3089 		dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3090 		    origin->ds_phys->ds_prev_snap_txg, UINT64_MAX,
3091 		    &origin->ds_phys->ds_unique_bytes, &comp, &uncomp);
3092 	}
3093 
3094 	/* swap blkptrs */
3095 	{
3096 		blkptr_t tmp;
3097 		tmp = csa->ohds->ds_phys->ds_bp;
3098 		csa->ohds->ds_phys->ds_bp = csa->cds->ds_phys->ds_bp;
3099 		csa->cds->ds_phys->ds_bp = tmp;
3100 	}
3101 
3102 	/* set dd_*_bytes */
3103 	{
3104 		int64_t dused, dcomp, duncomp;
3105 		uint64_t cdl_used, cdl_comp, cdl_uncomp;
3106 		uint64_t odl_used, odl_comp, odl_uncomp;
3107 
3108 		ASSERT3U(csa->cds->ds_dir->dd_phys->
3109 		    dd_used_breakdown[DD_USED_SNAP], ==, 0);
3110 
3111 		dsl_deadlist_space(&csa->cds->ds_deadlist,
3112 		    &cdl_used, &cdl_comp, &cdl_uncomp);
3113 		dsl_deadlist_space(&csa->ohds->ds_deadlist,
3114 		    &odl_used, &odl_comp, &odl_uncomp);
3115 
3116 		dused = csa->cds->ds_phys->ds_used_bytes + cdl_used -
3117 		    (csa->ohds->ds_phys->ds_used_bytes + odl_used);
3118 		dcomp = csa->cds->ds_phys->ds_compressed_bytes + cdl_comp -
3119 		    (csa->ohds->ds_phys->ds_compressed_bytes + odl_comp);
3120 		duncomp = csa->cds->ds_phys->ds_uncompressed_bytes +
3121 		    cdl_uncomp -
3122 		    (csa->ohds->ds_phys->ds_uncompressed_bytes + odl_uncomp);
3123 
3124 		dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_HEAD,
3125 		    dused, dcomp, duncomp, tx);
3126 		dsl_dir_diduse_space(csa->cds->ds_dir, DD_USED_HEAD,
3127 		    -dused, -dcomp, -duncomp, tx);
3128 
3129 		/*
3130 		 * The difference in the space used by snapshots is the
3131 		 * difference in snapshot space due to the head's
3132 		 * deadlist (since that's the only thing that's
3133 		 * changing that affects the snapused).
3134 		 */
3135 		dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3136 		    csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3137 		    &cdl_used, &cdl_comp, &cdl_uncomp);
3138 		dsl_deadlist_space_range(&csa->ohds->ds_deadlist,
3139 		    csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3140 		    &odl_used, &odl_comp, &odl_uncomp);
3141 		dsl_dir_transfer_space(csa->ohds->ds_dir, cdl_used - odl_used,
3142 		    DD_USED_HEAD, DD_USED_SNAP, tx);
3143 	}
3144 
3145 	/* swap ds_*_bytes */
3146 	SWITCH64(csa->ohds->ds_phys->ds_used_bytes,
3147 	    csa->cds->ds_phys->ds_used_bytes);
3148 	SWITCH64(csa->ohds->ds_phys->ds_compressed_bytes,
3149 	    csa->cds->ds_phys->ds_compressed_bytes);
3150 	SWITCH64(csa->ohds->ds_phys->ds_uncompressed_bytes,
3151 	    csa->cds->ds_phys->ds_uncompressed_bytes);
3152 	SWITCH64(csa->ohds->ds_phys->ds_unique_bytes,
3153 	    csa->cds->ds_phys->ds_unique_bytes);
3154 
3155 	/* apply any parent delta for change in unconsumed refreservation */
3156 	dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_REFRSRV,
3157 	    csa->unused_refres_delta, 0, 0, tx);
3158 
3159 	/*
3160 	 * Swap deadlists.
3161 	 */
3162 	dsl_deadlist_close(&csa->cds->ds_deadlist);
3163 	dsl_deadlist_close(&csa->ohds->ds_deadlist);
3164 	SWITCH64(csa->ohds->ds_phys->ds_deadlist_obj,
3165 	    csa->cds->ds_phys->ds_deadlist_obj);
3166 	dsl_deadlist_open(&csa->cds->ds_deadlist, dp->dp_meta_objset,
3167 	    csa->cds->ds_phys->ds_deadlist_obj);
3168 	dsl_deadlist_open(&csa->ohds->ds_deadlist, dp->dp_meta_objset,
3169 	    csa->ohds->ds_phys->ds_deadlist_obj);
3170 
3171 	dsl_scan_ds_clone_swapped(csa->ohds, csa->cds, tx);
3172 }
3173 
3174 /*
3175  * Swap 'clone' with its origin head datasets.  Used at the end of "zfs
3176  * recv" into an existing fs to swizzle the file system to the new
3177  * version, and by "zfs rollback".  Can also be used to swap two
3178  * independent head datasets if neither has any snapshots.
3179  */
3180 int
3181 dsl_dataset_clone_swap(dsl_dataset_t *clone, dsl_dataset_t *origin_head,
3182     boolean_t force)
3183 {
3184 	struct cloneswaparg csa;
3185 	int error;
3186 
3187 	ASSERT(clone->ds_owner);
3188 	ASSERT(origin_head->ds_owner);
3189 retry:
3190 	/*
3191 	 * Need exclusive access for the swap. If we're swapping these
3192 	 * datasets back after an error, we already hold the locks.
3193 	 */
3194 	if (!RW_WRITE_HELD(&clone->ds_rwlock))
3195 		rw_enter(&clone->ds_rwlock, RW_WRITER);
3196 	if (!RW_WRITE_HELD(&origin_head->ds_rwlock) &&
3197 	    !rw_tryenter(&origin_head->ds_rwlock, RW_WRITER)) {
3198 		rw_exit(&clone->ds_rwlock);
3199 		rw_enter(&origin_head->ds_rwlock, RW_WRITER);
3200 		if (!rw_tryenter(&clone->ds_rwlock, RW_WRITER)) {
3201 			rw_exit(&origin_head->ds_rwlock);
3202 			goto retry;
3203 		}
3204 	}
3205 	csa.cds = clone;
3206 	csa.ohds = origin_head;
3207 	csa.force = force;
3208 	error = dsl_sync_task_do(clone->ds_dir->dd_pool,
3209 	    dsl_dataset_clone_swap_check,
3210 	    dsl_dataset_clone_swap_sync, &csa, NULL, 9);
3211 	return (error);
3212 }
3213 
3214 /*
3215  * Given a pool name and a dataset object number in that pool,
3216  * return the name of that dataset.
3217  */
3218 int
3219 dsl_dsobj_to_dsname(char *pname, uint64_t obj, char *buf)
3220 {
3221 	spa_t *spa;
3222 	dsl_pool_t *dp;
3223 	dsl_dataset_t *ds;
3224 	int error;
3225 
3226 	if ((error = spa_open(pname, &spa, FTAG)) != 0)
3227 		return (error);
3228 	dp = spa_get_dsl(spa);
3229 	rw_enter(&dp->dp_config_rwlock, RW_READER);
3230 	if ((error = dsl_dataset_hold_obj(dp, obj, FTAG, &ds)) == 0) {
3231 		dsl_dataset_name(ds, buf);
3232 		dsl_dataset_rele(ds, FTAG);
3233 	}
3234 	rw_exit(&dp->dp_config_rwlock);
3235 	spa_close(spa, FTAG);
3236 
3237 	return (error);
3238 }
3239 
3240 int
3241 dsl_dataset_check_quota(dsl_dataset_t *ds, boolean_t check_quota,
3242     uint64_t asize, uint64_t inflight, uint64_t *used, uint64_t *ref_rsrv)
3243 {
3244 	int error = 0;
3245 
3246 	ASSERT3S(asize, >, 0);
3247 
3248 	/*
3249 	 * *ref_rsrv is the portion of asize that will come from any
3250 	 * unconsumed refreservation space.
3251 	 */
3252 	*ref_rsrv = 0;
3253 
3254 	mutex_enter(&ds->ds_lock);
3255 	/*
3256 	 * Make a space adjustment for reserved bytes.
3257 	 */
3258 	if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes) {
3259 		ASSERT3U(*used, >=,
3260 		    ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3261 		*used -= (ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3262 		*ref_rsrv =
3263 		    asize - MIN(asize, parent_delta(ds, asize + inflight));
3264 	}
3265 
3266 	if (!check_quota || ds->ds_quota == 0) {
3267 		mutex_exit(&ds->ds_lock);
3268 		return (0);
3269 	}
3270 	/*
3271 	 * If they are requesting more space, and our current estimate
3272 	 * is over quota, they get to try again unless the actual
3273 	 * on-disk is over quota and there are no pending changes (which
3274 	 * may free up space for us).
3275 	 */
3276 	if (ds->ds_phys->ds_used_bytes + inflight >= ds->ds_quota) {
3277 		if (inflight > 0 || ds->ds_phys->ds_used_bytes < ds->ds_quota)
3278 			error = ERESTART;
3279 		else
3280 			error = EDQUOT;
3281 	}
3282 	mutex_exit(&ds->ds_lock);
3283 
3284 	return (error);
3285 }
3286 
3287 /* ARGSUSED */
3288 static int
3289 dsl_dataset_set_quota_check(void *arg1, void *arg2, dmu_tx_t *tx)
3290 {
3291 	dsl_dataset_t *ds = arg1;
3292 	dsl_prop_setarg_t *psa = arg2;
3293 	int err;
3294 
3295 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_REFQUOTA)
3296 		return (ENOTSUP);
3297 
3298 	if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3299 		return (err);
3300 
3301 	if (psa->psa_effective_value == 0)
3302 		return (0);
3303 
3304 	if (psa->psa_effective_value < ds->ds_phys->ds_used_bytes ||
3305 	    psa->psa_effective_value < ds->ds_reserved)
3306 		return (ENOSPC);
3307 
3308 	return (0);
3309 }
3310 
3311 extern void dsl_prop_set_sync(void *, void *, dmu_tx_t *);
3312 
3313 void
3314 dsl_dataset_set_quota_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3315 {
3316 	dsl_dataset_t *ds = arg1;
3317 	dsl_prop_setarg_t *psa = arg2;
3318 	uint64_t effective_value = psa->psa_effective_value;
3319 
3320 	dsl_prop_set_sync(ds, psa, tx);
3321 	DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3322 
3323 	if (ds->ds_quota != effective_value) {
3324 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
3325 		ds->ds_quota = effective_value;
3326 
3327 		spa_history_log_internal(LOG_DS_REFQUOTA,
3328 		    ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu ",
3329 		    (longlong_t)ds->ds_quota, ds->ds_object);
3330 	}
3331 }
3332 
3333 int
3334 dsl_dataset_set_quota(const char *dsname, zprop_source_t source, uint64_t quota)
3335 {
3336 	dsl_dataset_t *ds;
3337 	dsl_prop_setarg_t psa;
3338 	int err;
3339 
3340 	dsl_prop_setarg_init_uint64(&psa, "refquota", source, &quota);
3341 
3342 	err = dsl_dataset_hold(dsname, FTAG, &ds);
3343 	if (err)
3344 		return (err);
3345 
3346 	/*
3347 	 * If someone removes a file, then tries to set the quota, we
3348 	 * want to make sure the file freeing takes effect.
3349 	 */
3350 	txg_wait_open(ds->ds_dir->dd_pool, 0);
3351 
3352 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3353 	    dsl_dataset_set_quota_check, dsl_dataset_set_quota_sync,
3354 	    ds, &psa, 0);
3355 
3356 	dsl_dataset_rele(ds, FTAG);
3357 	return (err);
3358 }
3359 
3360 static int
3361 dsl_dataset_set_reservation_check(void *arg1, void *arg2, dmu_tx_t *tx)
3362 {
3363 	dsl_dataset_t *ds = arg1;
3364 	dsl_prop_setarg_t *psa = arg2;
3365 	uint64_t effective_value;
3366 	uint64_t unique;
3367 	int err;
3368 
3369 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
3370 	    SPA_VERSION_REFRESERVATION)
3371 		return (ENOTSUP);
3372 
3373 	if (dsl_dataset_is_snapshot(ds))
3374 		return (EINVAL);
3375 
3376 	if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3377 		return (err);
3378 
3379 	effective_value = psa->psa_effective_value;
3380 
3381 	/*
3382 	 * If we are doing the preliminary check in open context, the
3383 	 * space estimates may be inaccurate.
3384 	 */
3385 	if (!dmu_tx_is_syncing(tx))
3386 		return (0);
3387 
3388 	mutex_enter(&ds->ds_lock);
3389 	if (!DS_UNIQUE_IS_ACCURATE(ds))
3390 		dsl_dataset_recalc_head_uniq(ds);
3391 	unique = ds->ds_phys->ds_unique_bytes;
3392 	mutex_exit(&ds->ds_lock);
3393 
3394 	if (MAX(unique, effective_value) > MAX(unique, ds->ds_reserved)) {
3395 		uint64_t delta = MAX(unique, effective_value) -
3396 		    MAX(unique, ds->ds_reserved);
3397 
3398 		if (delta > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
3399 			return (ENOSPC);
3400 		if (ds->ds_quota > 0 &&
3401 		    effective_value > ds->ds_quota)
3402 			return (ENOSPC);
3403 	}
3404 
3405 	return (0);
3406 }
3407 
3408 static void
3409 dsl_dataset_set_reservation_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3410 {
3411 	dsl_dataset_t *ds = arg1;
3412 	dsl_prop_setarg_t *psa = arg2;
3413 	uint64_t effective_value = psa->psa_effective_value;
3414 	uint64_t unique;
3415 	int64_t delta;
3416 
3417 	dsl_prop_set_sync(ds, psa, tx);
3418 	DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3419 
3420 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
3421 
3422 	mutex_enter(&ds->ds_dir->dd_lock);
3423 	mutex_enter(&ds->ds_lock);
3424 	ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
3425 	unique = ds->ds_phys->ds_unique_bytes;
3426 	delta = MAX(0, (int64_t)(effective_value - unique)) -
3427 	    MAX(0, (int64_t)(ds->ds_reserved - unique));
3428 	ds->ds_reserved = effective_value;
3429 	mutex_exit(&ds->ds_lock);
3430 
3431 	dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV, delta, 0, 0, tx);
3432 	mutex_exit(&ds->ds_dir->dd_lock);
3433 
3434 	spa_history_log_internal(LOG_DS_REFRESERV,
3435 	    ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu",
3436 	    (longlong_t)effective_value, ds->ds_object);
3437 }
3438 
3439 int
3440 dsl_dataset_set_reservation(const char *dsname, zprop_source_t source,
3441     uint64_t reservation)
3442 {
3443 	dsl_dataset_t *ds;
3444 	dsl_prop_setarg_t psa;
3445 	int err;
3446 
3447 	dsl_prop_setarg_init_uint64(&psa, "refreservation", source,
3448 	    &reservation);
3449 
3450 	err = dsl_dataset_hold(dsname, FTAG, &ds);
3451 	if (err)
3452 		return (err);
3453 
3454 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3455 	    dsl_dataset_set_reservation_check,
3456 	    dsl_dataset_set_reservation_sync, ds, &psa, 0);
3457 
3458 	dsl_dataset_rele(ds, FTAG);
3459 	return (err);
3460 }
3461 
3462 typedef struct zfs_hold_cleanup_arg {
3463 	dsl_pool_t *dp;
3464 	uint64_t dsobj;
3465 	char htag[MAXNAMELEN];
3466 } zfs_hold_cleanup_arg_t;
3467 
3468 static void
3469 dsl_dataset_user_release_onexit(void *arg)
3470 {
3471 	zfs_hold_cleanup_arg_t *ca = arg;
3472 
3473 	(void) dsl_dataset_user_release_tmp(ca->dp, ca->dsobj, ca->htag,
3474 	    B_TRUE);
3475 	kmem_free(ca, sizeof (zfs_hold_cleanup_arg_t));
3476 }
3477 
3478 void
3479 dsl_register_onexit_hold_cleanup(dsl_dataset_t *ds, const char *htag,
3480     minor_t minor)
3481 {
3482 	zfs_hold_cleanup_arg_t *ca;
3483 
3484 	ca = kmem_alloc(sizeof (zfs_hold_cleanup_arg_t), KM_SLEEP);
3485 	ca->dp = ds->ds_dir->dd_pool;
3486 	ca->dsobj = ds->ds_object;
3487 	(void) strlcpy(ca->htag, htag, sizeof (ca->htag));
3488 	VERIFY3U(0, ==, zfs_onexit_add_cb(minor,
3489 	    dsl_dataset_user_release_onexit, ca, NULL));
3490 }
3491 
3492 /*
3493  * If you add new checks here, you may need to add
3494  * additional checks to the "temporary" case in
3495  * snapshot_check() in dmu_objset.c.
3496  */
3497 static int
3498 dsl_dataset_user_hold_check(void *arg1, void *arg2, dmu_tx_t *tx)
3499 {
3500 	dsl_dataset_t *ds = arg1;
3501 	struct dsl_ds_holdarg *ha = arg2;
3502 	char *htag = ha->htag;
3503 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3504 	int error = 0;
3505 
3506 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3507 		return (ENOTSUP);
3508 
3509 	if (!dsl_dataset_is_snapshot(ds))
3510 		return (EINVAL);
3511 
3512 	/* tags must be unique */
3513 	mutex_enter(&ds->ds_lock);
3514 	if (ds->ds_phys->ds_userrefs_obj) {
3515 		error = zap_lookup(mos, ds->ds_phys->ds_userrefs_obj, htag,
3516 		    8, 1, tx);
3517 		if (error == 0)
3518 			error = EEXIST;
3519 		else if (error == ENOENT)
3520 			error = 0;
3521 	}
3522 	mutex_exit(&ds->ds_lock);
3523 
3524 	if (error == 0 && ha->temphold &&
3525 	    strlen(htag) + MAX_TAG_PREFIX_LEN >= MAXNAMELEN)
3526 		error = E2BIG;
3527 
3528 	return (error);
3529 }
3530 
3531 void
3532 dsl_dataset_user_hold_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3533 {
3534 	dsl_dataset_t *ds = arg1;
3535 	struct dsl_ds_holdarg *ha = arg2;
3536 	char *htag = ha->htag;
3537 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
3538 	objset_t *mos = dp->dp_meta_objset;
3539 	uint64_t now = gethrestime_sec();
3540 	uint64_t zapobj;
3541 
3542 	mutex_enter(&ds->ds_lock);
3543 	if (ds->ds_phys->ds_userrefs_obj == 0) {
3544 		/*
3545 		 * This is the first user hold for this dataset.  Create
3546 		 * the userrefs zap object.
3547 		 */
3548 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
3549 		zapobj = ds->ds_phys->ds_userrefs_obj =
3550 		    zap_create(mos, DMU_OT_USERREFS, DMU_OT_NONE, 0, tx);
3551 	} else {
3552 		zapobj = ds->ds_phys->ds_userrefs_obj;
3553 	}
3554 	ds->ds_userrefs++;
3555 	mutex_exit(&ds->ds_lock);
3556 
3557 	VERIFY(0 == zap_add(mos, zapobj, htag, 8, 1, &now, tx));
3558 
3559 	if (ha->temphold) {
3560 		VERIFY(0 == dsl_pool_user_hold(dp, ds->ds_object,
3561 		    htag, &now, tx));
3562 	}
3563 
3564 	spa_history_log_internal(LOG_DS_USER_HOLD,
3565 	    dp->dp_spa, tx, "<%s> temp = %d dataset = %llu", htag,
3566 	    (int)ha->temphold, ds->ds_object);
3567 }
3568 
3569 static int
3570 dsl_dataset_user_hold_one(const char *dsname, void *arg)
3571 {
3572 	struct dsl_ds_holdarg *ha = arg;
3573 	dsl_dataset_t *ds;
3574 	int error;
3575 	char *name;
3576 
3577 	/* alloc a buffer to hold dsname@snapname plus terminating NULL */
3578 	name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3579 	error = dsl_dataset_hold(name, ha->dstg, &ds);
3580 	strfree(name);
3581 	if (error == 0) {
3582 		ha->gotone = B_TRUE;
3583 		dsl_sync_task_create(ha->dstg, dsl_dataset_user_hold_check,
3584 		    dsl_dataset_user_hold_sync, ds, ha, 0);
3585 	} else if (error == ENOENT && ha->recursive) {
3586 		error = 0;
3587 	} else {
3588 		(void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3589 	}
3590 	return (error);
3591 }
3592 
3593 int
3594 dsl_dataset_user_hold_for_send(dsl_dataset_t *ds, char *htag,
3595     boolean_t temphold)
3596 {
3597 	struct dsl_ds_holdarg *ha;
3598 	int error;
3599 
3600 	ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3601 	ha->htag = htag;
3602 	ha->temphold = temphold;
3603 	error = dsl_sync_task_do(ds->ds_dir->dd_pool,
3604 	    dsl_dataset_user_hold_check, dsl_dataset_user_hold_sync,
3605 	    ds, ha, 0);
3606 	kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3607 
3608 	return (error);
3609 }
3610 
3611 int
3612 dsl_dataset_user_hold(char *dsname, char *snapname, char *htag,
3613     boolean_t recursive, boolean_t temphold, int cleanup_fd)
3614 {
3615 	struct dsl_ds_holdarg *ha;
3616 	dsl_sync_task_t *dst;
3617 	spa_t *spa;
3618 	int error;
3619 	minor_t minor = 0;
3620 
3621 	if (cleanup_fd != -1) {
3622 		/* Currently we only support cleanup-on-exit of tempholds. */
3623 		if (!temphold)
3624 			return (EINVAL);
3625 		error = zfs_onexit_fd_hold(cleanup_fd, &minor);
3626 		if (error)
3627 			return (error);
3628 	}
3629 
3630 	ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3631 
3632 	(void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3633 
3634 	error = spa_open(dsname, &spa, FTAG);
3635 	if (error) {
3636 		kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3637 		if (cleanup_fd != -1)
3638 			zfs_onexit_fd_rele(cleanup_fd);
3639 		return (error);
3640 	}
3641 
3642 	ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3643 	ha->htag = htag;
3644 	ha->snapname = snapname;
3645 	ha->recursive = recursive;
3646 	ha->temphold = temphold;
3647 
3648 	if (recursive) {
3649 		error = dmu_objset_find(dsname, dsl_dataset_user_hold_one,
3650 		    ha, DS_FIND_CHILDREN);
3651 	} else {
3652 		error = dsl_dataset_user_hold_one(dsname, ha);
3653 	}
3654 	if (error == 0)
3655 		error = dsl_sync_task_group_wait(ha->dstg);
3656 
3657 	for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3658 	    dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3659 		dsl_dataset_t *ds = dst->dst_arg1;
3660 
3661 		if (dst->dst_err) {
3662 			dsl_dataset_name(ds, ha->failed);
3663 			*strchr(ha->failed, '@') = '\0';
3664 		} else if (error == 0 && minor != 0 && temphold) {
3665 			/*
3666 			 * If this hold is to be released upon process exit,
3667 			 * register that action now.
3668 			 */
3669 			dsl_register_onexit_hold_cleanup(ds, htag, minor);
3670 		}
3671 		dsl_dataset_rele(ds, ha->dstg);
3672 	}
3673 
3674 	if (error == 0 && recursive && !ha->gotone)
3675 		error = ENOENT;
3676 
3677 	if (error)
3678 		(void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3679 
3680 	dsl_sync_task_group_destroy(ha->dstg);
3681 
3682 	kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3683 	spa_close(spa, FTAG);
3684 	if (cleanup_fd != -1)
3685 		zfs_onexit_fd_rele(cleanup_fd);
3686 	return (error);
3687 }
3688 
3689 struct dsl_ds_releasearg {
3690 	dsl_dataset_t *ds;
3691 	const char *htag;
3692 	boolean_t own;		/* do we own or just hold ds? */
3693 };
3694 
3695 static int
3696 dsl_dataset_release_might_destroy(dsl_dataset_t *ds, const char *htag,
3697     boolean_t *might_destroy)
3698 {
3699 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3700 	uint64_t zapobj;
3701 	uint64_t tmp;
3702 	int error;
3703 
3704 	*might_destroy = B_FALSE;
3705 
3706 	mutex_enter(&ds->ds_lock);
3707 	zapobj = ds->ds_phys->ds_userrefs_obj;
3708 	if (zapobj == 0) {
3709 		/* The tag can't possibly exist */
3710 		mutex_exit(&ds->ds_lock);
3711 		return (ESRCH);
3712 	}
3713 
3714 	/* Make sure the tag exists */
3715 	error = zap_lookup(mos, zapobj, htag, 8, 1, &tmp);
3716 	if (error) {
3717 		mutex_exit(&ds->ds_lock);
3718 		if (error == ENOENT)
3719 			error = ESRCH;
3720 		return (error);
3721 	}
3722 
3723 	if (ds->ds_userrefs == 1 && ds->ds_phys->ds_num_children == 1 &&
3724 	    DS_IS_DEFER_DESTROY(ds))
3725 		*might_destroy = B_TRUE;
3726 
3727 	mutex_exit(&ds->ds_lock);
3728 	return (0);
3729 }
3730 
3731 static int
3732 dsl_dataset_user_release_check(void *arg1, void *tag, dmu_tx_t *tx)
3733 {
3734 	struct dsl_ds_releasearg *ra = arg1;
3735 	dsl_dataset_t *ds = ra->ds;
3736 	boolean_t might_destroy;
3737 	int error;
3738 
3739 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3740 		return (ENOTSUP);
3741 
3742 	error = dsl_dataset_release_might_destroy(ds, ra->htag, &might_destroy);
3743 	if (error)
3744 		return (error);
3745 
3746 	if (might_destroy) {
3747 		struct dsl_ds_destroyarg dsda = {0};
3748 
3749 		if (dmu_tx_is_syncing(tx)) {
3750 			/*
3751 			 * If we're not prepared to remove the snapshot,
3752 			 * we can't allow the release to happen right now.
3753 			 */
3754 			if (!ra->own)
3755 				return (EBUSY);
3756 		}
3757 		dsda.ds = ds;
3758 		dsda.releasing = B_TRUE;
3759 		return (dsl_dataset_destroy_check(&dsda, tag, tx));
3760 	}
3761 
3762 	return (0);
3763 }
3764 
3765 static void
3766 dsl_dataset_user_release_sync(void *arg1, void *tag, dmu_tx_t *tx)
3767 {
3768 	struct dsl_ds_releasearg *ra = arg1;
3769 	dsl_dataset_t *ds = ra->ds;
3770 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
3771 	objset_t *mos = dp->dp_meta_objset;
3772 	uint64_t zapobj;
3773 	uint64_t dsobj = ds->ds_object;
3774 	uint64_t refs;
3775 	int error;
3776 
3777 	mutex_enter(&ds->ds_lock);
3778 	ds->ds_userrefs--;
3779 	refs = ds->ds_userrefs;
3780 	mutex_exit(&ds->ds_lock);
3781 	error = dsl_pool_user_release(dp, ds->ds_object, ra->htag, tx);
3782 	VERIFY(error == 0 || error == ENOENT);
3783 	zapobj = ds->ds_phys->ds_userrefs_obj;
3784 	VERIFY(0 == zap_remove(mos, zapobj, ra->htag, tx));
3785 	if (ds->ds_userrefs == 0 && ds->ds_phys->ds_num_children == 1 &&
3786 	    DS_IS_DEFER_DESTROY(ds)) {
3787 		struct dsl_ds_destroyarg dsda = {0};
3788 
3789 		ASSERT(ra->own);
3790 		dsda.ds = ds;
3791 		dsda.releasing = B_TRUE;
3792 		/* We already did the destroy_check */
3793 		dsl_dataset_destroy_sync(&dsda, tag, tx);
3794 	}
3795 
3796 	spa_history_log_internal(LOG_DS_USER_RELEASE,
3797 	    dp->dp_spa, tx, "<%s> %lld dataset = %llu",
3798 	    ra->htag, (longlong_t)refs, dsobj);
3799 }
3800 
3801 static int
3802 dsl_dataset_user_release_one(const char *dsname, void *arg)
3803 {
3804 	struct dsl_ds_holdarg *ha = arg;
3805 	struct dsl_ds_releasearg *ra;
3806 	dsl_dataset_t *ds;
3807 	int error;
3808 	void *dtag = ha->dstg;
3809 	char *name;
3810 	boolean_t own = B_FALSE;
3811 	boolean_t might_destroy;
3812 
3813 	/* alloc a buffer to hold dsname@snapname, plus the terminating NULL */
3814 	name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3815 	error = dsl_dataset_hold(name, dtag, &ds);
3816 	strfree(name);
3817 	if (error == ENOENT && ha->recursive)
3818 		return (0);
3819 	(void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3820 	if (error)
3821 		return (error);
3822 
3823 	ha->gotone = B_TRUE;
3824 
3825 	ASSERT(dsl_dataset_is_snapshot(ds));
3826 
3827 	error = dsl_dataset_release_might_destroy(ds, ha->htag, &might_destroy);
3828 	if (error) {
3829 		dsl_dataset_rele(ds, dtag);
3830 		return (error);
3831 	}
3832 
3833 	if (might_destroy) {
3834 #ifdef _KERNEL
3835 		name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3836 		error = zfs_unmount_snap(name, NULL);
3837 		strfree(name);
3838 		if (error) {
3839 			dsl_dataset_rele(ds, dtag);
3840 			return (error);
3841 		}
3842 #endif
3843 		if (!dsl_dataset_tryown(ds, B_TRUE, dtag)) {
3844 			dsl_dataset_rele(ds, dtag);
3845 			return (EBUSY);
3846 		} else {
3847 			own = B_TRUE;
3848 			dsl_dataset_make_exclusive(ds, dtag);
3849 		}
3850 	}
3851 
3852 	ra = kmem_alloc(sizeof (struct dsl_ds_releasearg), KM_SLEEP);
3853 	ra->ds = ds;
3854 	ra->htag = ha->htag;
3855 	ra->own = own;
3856 	dsl_sync_task_create(ha->dstg, dsl_dataset_user_release_check,
3857 	    dsl_dataset_user_release_sync, ra, dtag, 0);
3858 
3859 	return (0);
3860 }
3861 
3862 int
3863 dsl_dataset_user_release(char *dsname, char *snapname, char *htag,
3864     boolean_t recursive)
3865 {
3866 	struct dsl_ds_holdarg *ha;
3867 	dsl_sync_task_t *dst;
3868 	spa_t *spa;
3869 	int error;
3870 
3871 top:
3872 	ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3873 
3874 	(void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3875 
3876 	error = spa_open(dsname, &spa, FTAG);
3877 	if (error) {
3878 		kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3879 		return (error);
3880 	}
3881 
3882 	ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3883 	ha->htag = htag;
3884 	ha->snapname = snapname;
3885 	ha->recursive = recursive;
3886 	if (recursive) {
3887 		error = dmu_objset_find(dsname, dsl_dataset_user_release_one,
3888 		    ha, DS_FIND_CHILDREN);
3889 	} else {
3890 		error = dsl_dataset_user_release_one(dsname, ha);
3891 	}
3892 	if (error == 0)
3893 		error = dsl_sync_task_group_wait(ha->dstg);
3894 
3895 	for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3896 	    dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3897 		struct dsl_ds_releasearg *ra = dst->dst_arg1;
3898 		dsl_dataset_t *ds = ra->ds;
3899 
3900 		if (dst->dst_err)
3901 			dsl_dataset_name(ds, ha->failed);
3902 
3903 		if (ra->own)
3904 			dsl_dataset_disown(ds, ha->dstg);
3905 		else
3906 			dsl_dataset_rele(ds, ha->dstg);
3907 
3908 		kmem_free(ra, sizeof (struct dsl_ds_releasearg));
3909 	}
3910 
3911 	if (error == 0 && recursive && !ha->gotone)
3912 		error = ENOENT;
3913 
3914 	if (error && error != EBUSY)
3915 		(void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3916 
3917 	dsl_sync_task_group_destroy(ha->dstg);
3918 	kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3919 	spa_close(spa, FTAG);
3920 
3921 	/*
3922 	 * We can get EBUSY if we were racing with deferred destroy and
3923 	 * dsl_dataset_user_release_check() hadn't done the necessary
3924 	 * open context setup.  We can also get EBUSY if we're racing
3925 	 * with destroy and that thread is the ds_owner.  Either way
3926 	 * the busy condition should be transient, and we should retry
3927 	 * the release operation.
3928 	 */
3929 	if (error == EBUSY)
3930 		goto top;
3931 
3932 	return (error);
3933 }
3934 
3935 /*
3936  * Called at spa_load time (with retry == B_FALSE) to release a stale
3937  * temporary user hold. Also called by the onexit code (with retry == B_TRUE).
3938  */
3939 int
3940 dsl_dataset_user_release_tmp(dsl_pool_t *dp, uint64_t dsobj, char *htag,
3941     boolean_t retry)
3942 {
3943 	dsl_dataset_t *ds;
3944 	char *snap;
3945 	char *name;
3946 	int namelen;
3947 	int error;
3948 
3949 	do {
3950 		rw_enter(&dp->dp_config_rwlock, RW_READER);
3951 		error = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
3952 		rw_exit(&dp->dp_config_rwlock);
3953 		if (error)
3954 			return (error);
3955 		namelen = dsl_dataset_namelen(ds)+1;
3956 		name = kmem_alloc(namelen, KM_SLEEP);
3957 		dsl_dataset_name(ds, name);
3958 		dsl_dataset_rele(ds, FTAG);
3959 
3960 		snap = strchr(name, '@');
3961 		*snap = '\0';
3962 		++snap;
3963 		error = dsl_dataset_user_release(name, snap, htag, B_FALSE);
3964 		kmem_free(name, namelen);
3965 
3966 		/*
3967 		 * The object can't have been destroyed because we have a hold,
3968 		 * but it might have been renamed, resulting in ENOENT.  Retry
3969 		 * if we've been requested to do so.
3970 		 *
3971 		 * It would be nice if we could use the dsobj all the way
3972 		 * through and avoid ENOENT entirely.  But we might need to
3973 		 * unmount the snapshot, and there's currently no way to lookup
3974 		 * a vfsp using a ZFS object id.
3975 		 */
3976 	} while ((error == ENOENT) && retry);
3977 
3978 	return (error);
3979 }
3980 
3981 int
3982 dsl_dataset_get_holds(const char *dsname, nvlist_t **nvp)
3983 {
3984 	dsl_dataset_t *ds;
3985 	int err;
3986 
3987 	err = dsl_dataset_hold(dsname, FTAG, &ds);
3988 	if (err)
3989 		return (err);
3990 
3991 	VERIFY(0 == nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP));
3992 	if (ds->ds_phys->ds_userrefs_obj != 0) {
3993 		zap_attribute_t *za;
3994 		zap_cursor_t zc;
3995 
3996 		za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
3997 		for (zap_cursor_init(&zc, ds->ds_dir->dd_pool->dp_meta_objset,
3998 		    ds->ds_phys->ds_userrefs_obj);
3999 		    zap_cursor_retrieve(&zc, za) == 0;
4000 		    zap_cursor_advance(&zc)) {
4001 			VERIFY(0 == nvlist_add_uint64(*nvp, za->za_name,
4002 			    za->za_first_integer));
4003 		}
4004 		zap_cursor_fini(&zc);
4005 		kmem_free(za, sizeof (zap_attribute_t));
4006 	}
4007 	dsl_dataset_rele(ds, FTAG);
4008 	return (0);
4009 }
4010 
4011 /*
4012  * Note, this fuction is used as the callback for dmu_objset_find().  We
4013  * always return 0 so that we will continue to find and process
4014  * inconsistent datasets, even if we encounter an error trying to
4015  * process one of them.
4016  */
4017 /* ARGSUSED */
4018 int
4019 dsl_destroy_inconsistent(const char *dsname, void *arg)
4020 {
4021 	dsl_dataset_t *ds;
4022 
4023 	if (dsl_dataset_own(dsname, B_TRUE, FTAG, &ds) == 0) {
4024 		if (DS_IS_INCONSISTENT(ds))
4025 			(void) dsl_dataset_destroy(ds, FTAG, B_FALSE);
4026 		else
4027 			dsl_dataset_disown(ds, FTAG);
4028 	}
4029 	return (0);
4030 }
4031