xref: /titanic_52/usr/src/uts/common/fs/zfs/dsl_dataset.c (revision 6be61d4ea129a94bdfe33533b2bc265d4447f05c)
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 /* ARGSUSED */
1379 int
1380 dsl_dataset_destroy_check(void *arg1, void *arg2, dmu_tx_t *tx)
1381 {
1382 	struct dsl_ds_destroyarg *dsda = arg1;
1383 	dsl_dataset_t *ds = dsda->ds;
1384 
1385 	/* we have an owner hold, so noone else can destroy us */
1386 	ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
1387 
1388 	/*
1389 	 * Only allow deferred destroy on pools that support it.
1390 	 * NOTE: deferred destroy is only supported on snapshots.
1391 	 */
1392 	if (dsda->defer) {
1393 		if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
1394 		    SPA_VERSION_USERREFS)
1395 			return (ENOTSUP);
1396 		ASSERT(dsl_dataset_is_snapshot(ds));
1397 		return (0);
1398 	}
1399 
1400 	/*
1401 	 * Can't delete a head dataset if there are snapshots of it.
1402 	 * (Except if the only snapshots are from the branch we cloned
1403 	 * from.)
1404 	 */
1405 	if (ds->ds_prev != NULL &&
1406 	    ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1407 		return (EBUSY);
1408 
1409 	/*
1410 	 * If we made changes this txg, traverse_dsl_dataset won't find
1411 	 * them.  Try again.
1412 	 */
1413 	if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
1414 		return (EAGAIN);
1415 
1416 	if (dsl_dataset_is_snapshot(ds)) {
1417 		/*
1418 		 * If this snapshot has an elevated user reference count,
1419 		 * we can't destroy it yet.
1420 		 */
1421 		if (ds->ds_userrefs > 0 && !dsda->releasing)
1422 			return (EBUSY);
1423 
1424 		mutex_enter(&ds->ds_lock);
1425 		/*
1426 		 * Can't delete a branch point. However, if we're destroying
1427 		 * a clone and removing its origin due to it having a user
1428 		 * hold count of 0 and having been marked for deferred destroy,
1429 		 * it's OK for the origin to have a single clone.
1430 		 */
1431 		if (ds->ds_phys->ds_num_children >
1432 		    (dsda->is_origin_rm ? 2 : 1)) {
1433 			mutex_exit(&ds->ds_lock);
1434 			return (EEXIST);
1435 		}
1436 		mutex_exit(&ds->ds_lock);
1437 	} else if (dsl_dir_is_clone(ds->ds_dir)) {
1438 		return (dsl_dataset_origin_check(dsda, arg2, tx));
1439 	}
1440 
1441 	/* XXX we should do some i/o error checking... */
1442 	return (0);
1443 }
1444 
1445 struct refsarg {
1446 	kmutex_t lock;
1447 	boolean_t gone;
1448 	kcondvar_t cv;
1449 };
1450 
1451 /* ARGSUSED */
1452 static void
1453 dsl_dataset_refs_gone(dmu_buf_t *db, void *argv)
1454 {
1455 	struct refsarg *arg = argv;
1456 
1457 	mutex_enter(&arg->lock);
1458 	arg->gone = TRUE;
1459 	cv_signal(&arg->cv);
1460 	mutex_exit(&arg->lock);
1461 }
1462 
1463 static void
1464 dsl_dataset_drain_refs(dsl_dataset_t *ds, void *tag)
1465 {
1466 	struct refsarg arg;
1467 
1468 	mutex_init(&arg.lock, NULL, MUTEX_DEFAULT, NULL);
1469 	cv_init(&arg.cv, NULL, CV_DEFAULT, NULL);
1470 	arg.gone = FALSE;
1471 	(void) dmu_buf_update_user(ds->ds_dbuf, ds, &arg, &ds->ds_phys,
1472 	    dsl_dataset_refs_gone);
1473 	dmu_buf_rele(ds->ds_dbuf, tag);
1474 	mutex_enter(&arg.lock);
1475 	while (!arg.gone)
1476 		cv_wait(&arg.cv, &arg.lock);
1477 	ASSERT(arg.gone);
1478 	mutex_exit(&arg.lock);
1479 	ds->ds_dbuf = NULL;
1480 	ds->ds_phys = NULL;
1481 	mutex_destroy(&arg.lock);
1482 	cv_destroy(&arg.cv);
1483 }
1484 
1485 static void
1486 remove_from_next_clones(dsl_dataset_t *ds, uint64_t obj, dmu_tx_t *tx)
1487 {
1488 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1489 	uint64_t count;
1490 	int err;
1491 
1492 	ASSERT(ds->ds_phys->ds_num_children >= 2);
1493 	err = zap_remove_int(mos, ds->ds_phys->ds_next_clones_obj, obj, tx);
1494 	/*
1495 	 * The err should not be ENOENT, but a bug in a previous version
1496 	 * of the code could cause upgrade_clones_cb() to not set
1497 	 * ds_next_snap_obj when it should, leading to a missing entry.
1498 	 * If we knew that the pool was created after
1499 	 * SPA_VERSION_NEXT_CLONES, we could assert that it isn't
1500 	 * ENOENT.  However, at least we can check that we don't have
1501 	 * too many entries in the next_clones_obj even after failing to
1502 	 * remove this one.
1503 	 */
1504 	if (err != ENOENT) {
1505 		VERIFY3U(err, ==, 0);
1506 	}
1507 	ASSERT3U(0, ==, zap_count(mos, ds->ds_phys->ds_next_clones_obj,
1508 	    &count));
1509 	ASSERT3U(count, <=, ds->ds_phys->ds_num_children - 2);
1510 }
1511 
1512 static void
1513 dsl_dataset_remove_clones_key(dsl_dataset_t *ds, uint64_t mintxg, dmu_tx_t *tx)
1514 {
1515 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1516 	zap_cursor_t zc;
1517 	zap_attribute_t za;
1518 
1519 	/*
1520 	 * If it is the old version, dd_clones doesn't exist so we can't
1521 	 * find the clones, but deadlist_remove_key() is a no-op so it
1522 	 * doesn't matter.
1523 	 */
1524 	if (ds->ds_dir->dd_phys->dd_clones == 0)
1525 		return;
1526 
1527 	for (zap_cursor_init(&zc, mos, ds->ds_dir->dd_phys->dd_clones);
1528 	    zap_cursor_retrieve(&zc, &za) == 0;
1529 	    zap_cursor_advance(&zc)) {
1530 		dsl_dataset_t *clone;
1531 
1532 		VERIFY3U(0, ==, dsl_dataset_hold_obj(ds->ds_dir->dd_pool,
1533 		    za.za_first_integer, FTAG, &clone));
1534 		if (clone->ds_dir->dd_origin_txg > mintxg) {
1535 			dsl_deadlist_remove_key(&clone->ds_deadlist,
1536 			    mintxg, tx);
1537 			dsl_dataset_remove_clones_key(clone, mintxg, tx);
1538 		}
1539 		dsl_dataset_rele(clone, FTAG);
1540 	}
1541 	zap_cursor_fini(&zc);
1542 }
1543 
1544 struct process_old_arg {
1545 	dsl_dataset_t *ds;
1546 	dsl_dataset_t *ds_prev;
1547 	boolean_t after_branch_point;
1548 	zio_t *pio;
1549 	uint64_t used, comp, uncomp;
1550 };
1551 
1552 static int
1553 process_old_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1554 {
1555 	struct process_old_arg *poa = arg;
1556 	dsl_pool_t *dp = poa->ds->ds_dir->dd_pool;
1557 
1558 	if (bp->blk_birth <= poa->ds->ds_phys->ds_prev_snap_txg) {
1559 		dsl_deadlist_insert(&poa->ds->ds_deadlist, bp, tx);
1560 		if (poa->ds_prev && !poa->after_branch_point &&
1561 		    bp->blk_birth >
1562 		    poa->ds_prev->ds_phys->ds_prev_snap_txg) {
1563 			poa->ds_prev->ds_phys->ds_unique_bytes +=
1564 			    bp_get_dsize_sync(dp->dp_spa, bp);
1565 		}
1566 	} else {
1567 		poa->used += bp_get_dsize_sync(dp->dp_spa, bp);
1568 		poa->comp += BP_GET_PSIZE(bp);
1569 		poa->uncomp += BP_GET_UCSIZE(bp);
1570 		dsl_free_sync(poa->pio, dp, tx->tx_txg, bp);
1571 	}
1572 	return (0);
1573 }
1574 
1575 static void
1576 process_old_deadlist(dsl_dataset_t *ds, dsl_dataset_t *ds_prev,
1577     dsl_dataset_t *ds_next, boolean_t after_branch_point, dmu_tx_t *tx)
1578 {
1579 	struct process_old_arg poa = { 0 };
1580 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1581 	objset_t *mos = dp->dp_meta_objset;
1582 
1583 	ASSERT(ds->ds_deadlist.dl_oldfmt);
1584 	ASSERT(ds_next->ds_deadlist.dl_oldfmt);
1585 
1586 	poa.ds = ds;
1587 	poa.ds_prev = ds_prev;
1588 	poa.after_branch_point = after_branch_point;
1589 	poa.pio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
1590 	VERIFY3U(0, ==, bpobj_iterate(&ds_next->ds_deadlist.dl_bpobj,
1591 	    process_old_cb, &poa, tx));
1592 	VERIFY3U(zio_wait(poa.pio), ==, 0);
1593 	ASSERT3U(poa.used, ==, ds->ds_phys->ds_unique_bytes);
1594 
1595 	/* change snapused */
1596 	dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1597 	    -poa.used, -poa.comp, -poa.uncomp, tx);
1598 
1599 	/* swap next's deadlist to our deadlist */
1600 	dsl_deadlist_close(&ds->ds_deadlist);
1601 	dsl_deadlist_close(&ds_next->ds_deadlist);
1602 	SWITCH64(ds_next->ds_phys->ds_deadlist_obj,
1603 	    ds->ds_phys->ds_deadlist_obj);
1604 	dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
1605 	dsl_deadlist_open(&ds_next->ds_deadlist, mos,
1606 	    ds_next->ds_phys->ds_deadlist_obj);
1607 }
1608 
1609 void
1610 dsl_dataset_destroy_sync(void *arg1, void *tag, dmu_tx_t *tx)
1611 {
1612 	struct dsl_ds_destroyarg *dsda = arg1;
1613 	dsl_dataset_t *ds = dsda->ds;
1614 	int err;
1615 	int after_branch_point = FALSE;
1616 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1617 	objset_t *mos = dp->dp_meta_objset;
1618 	dsl_dataset_t *ds_prev = NULL;
1619 	uint64_t obj;
1620 
1621 	ASSERT(ds->ds_owner);
1622 	ASSERT(dsda->defer || ds->ds_phys->ds_num_children <= 1);
1623 	ASSERT(ds->ds_prev == NULL ||
1624 	    ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
1625 	ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
1626 
1627 	if (dsda->defer) {
1628 		ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1629 		if (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1) {
1630 			dmu_buf_will_dirty(ds->ds_dbuf, tx);
1631 			ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
1632 			return;
1633 		}
1634 	}
1635 
1636 	/* signal any waiters that this dataset is going away */
1637 	mutex_enter(&ds->ds_lock);
1638 	ds->ds_owner = dsl_reaper;
1639 	cv_broadcast(&ds->ds_exclusive_cv);
1640 	mutex_exit(&ds->ds_lock);
1641 
1642 	/* Remove our reservation */
1643 	if (ds->ds_reserved != 0) {
1644 		dsl_prop_setarg_t psa;
1645 		uint64_t value = 0;
1646 
1647 		dsl_prop_setarg_init_uint64(&psa, "refreservation",
1648 		    (ZPROP_SRC_NONE | ZPROP_SRC_LOCAL | ZPROP_SRC_RECEIVED),
1649 		    &value);
1650 		psa.psa_effective_value = 0;	/* predict default value */
1651 
1652 		dsl_dataset_set_reservation_sync(ds, &psa, tx);
1653 		ASSERT3U(ds->ds_reserved, ==, 0);
1654 	}
1655 
1656 	ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1657 
1658 	dsl_scan_ds_destroyed(ds, tx);
1659 
1660 	obj = ds->ds_object;
1661 
1662 	if (ds->ds_phys->ds_prev_snap_obj != 0) {
1663 		if (ds->ds_prev) {
1664 			ds_prev = ds->ds_prev;
1665 		} else {
1666 			VERIFY(0 == dsl_dataset_hold_obj(dp,
1667 			    ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
1668 		}
1669 		after_branch_point =
1670 		    (ds_prev->ds_phys->ds_next_snap_obj != obj);
1671 
1672 		dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
1673 		if (after_branch_point &&
1674 		    ds_prev->ds_phys->ds_next_clones_obj != 0) {
1675 			remove_from_next_clones(ds_prev, obj, tx);
1676 			if (ds->ds_phys->ds_next_snap_obj != 0) {
1677 				VERIFY(0 == zap_add_int(mos,
1678 				    ds_prev->ds_phys->ds_next_clones_obj,
1679 				    ds->ds_phys->ds_next_snap_obj, tx));
1680 			}
1681 		}
1682 		if (after_branch_point &&
1683 		    ds->ds_phys->ds_next_snap_obj == 0) {
1684 			/* This clone is toast. */
1685 			ASSERT(ds_prev->ds_phys->ds_num_children > 1);
1686 			ds_prev->ds_phys->ds_num_children--;
1687 
1688 			/*
1689 			 * If the clone's origin has no other clones, no
1690 			 * user holds, and has been marked for deferred
1691 			 * deletion, then we should have done the necessary
1692 			 * destroy setup for it.
1693 			 */
1694 			if (ds_prev->ds_phys->ds_num_children == 1 &&
1695 			    ds_prev->ds_userrefs == 0 &&
1696 			    DS_IS_DEFER_DESTROY(ds_prev)) {
1697 				ASSERT3P(dsda->rm_origin, !=, NULL);
1698 			} else {
1699 				ASSERT3P(dsda->rm_origin, ==, NULL);
1700 			}
1701 		} else if (!after_branch_point) {
1702 			ds_prev->ds_phys->ds_next_snap_obj =
1703 			    ds->ds_phys->ds_next_snap_obj;
1704 		}
1705 	}
1706 
1707 	if (dsl_dataset_is_snapshot(ds)) {
1708 		dsl_dataset_t *ds_next;
1709 		uint64_t old_unique;
1710 		uint64_t used = 0, comp = 0, uncomp = 0;
1711 
1712 		VERIFY(0 == dsl_dataset_hold_obj(dp,
1713 		    ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
1714 		ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
1715 
1716 		old_unique = ds_next->ds_phys->ds_unique_bytes;
1717 
1718 		dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
1719 		ds_next->ds_phys->ds_prev_snap_obj =
1720 		    ds->ds_phys->ds_prev_snap_obj;
1721 		ds_next->ds_phys->ds_prev_snap_txg =
1722 		    ds->ds_phys->ds_prev_snap_txg;
1723 		ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1724 		    ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
1725 
1726 
1727 		if (ds_next->ds_deadlist.dl_oldfmt) {
1728 			process_old_deadlist(ds, ds_prev, ds_next,
1729 			    after_branch_point, tx);
1730 		} else {
1731 			/* Adjust prev's unique space. */
1732 			if (ds_prev && !after_branch_point) {
1733 				dsl_deadlist_space_range(&ds_next->ds_deadlist,
1734 				    ds_prev->ds_phys->ds_prev_snap_txg,
1735 				    ds->ds_phys->ds_prev_snap_txg,
1736 				    &used, &comp, &uncomp);
1737 				ds_prev->ds_phys->ds_unique_bytes += used;
1738 			}
1739 
1740 			/* Adjust snapused. */
1741 			dsl_deadlist_space_range(&ds_next->ds_deadlist,
1742 			    ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
1743 			    &used, &comp, &uncomp);
1744 			dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1745 			    -used, -comp, -uncomp, tx);
1746 
1747 			/* Move blocks to be freed to pool's free list. */
1748 			dsl_deadlist_move_bpobj(&ds_next->ds_deadlist,
1749 			    &dp->dp_free_bpobj, ds->ds_phys->ds_prev_snap_txg,
1750 			    tx);
1751 			dsl_dir_diduse_space(tx->tx_pool->dp_free_dir,
1752 			    DD_USED_HEAD, used, comp, uncomp, tx);
1753 			dsl_dir_dirty(tx->tx_pool->dp_free_dir, tx);
1754 
1755 			/* Merge our deadlist into next's and free it. */
1756 			dsl_deadlist_merge(&ds_next->ds_deadlist,
1757 			    ds->ds_phys->ds_deadlist_obj, tx);
1758 		}
1759 		dsl_deadlist_close(&ds->ds_deadlist);
1760 		dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1761 
1762 		/* Collapse range in clone heads */
1763 		dsl_dataset_remove_clones_key(ds,
1764 		    ds->ds_phys->ds_creation_txg, tx);
1765 
1766 		if (dsl_dataset_is_snapshot(ds_next)) {
1767 			dsl_dataset_t *ds_nextnext;
1768 
1769 			/*
1770 			 * Update next's unique to include blocks which
1771 			 * were previously shared by only this snapshot
1772 			 * and it.  Those blocks will be born after the
1773 			 * prev snap and before this snap, and will have
1774 			 * died after the next snap and before the one
1775 			 * after that (ie. be on the snap after next's
1776 			 * deadlist).
1777 			 */
1778 			VERIFY(0 == dsl_dataset_hold_obj(dp,
1779 			    ds_next->ds_phys->ds_next_snap_obj,
1780 			    FTAG, &ds_nextnext));
1781 			dsl_deadlist_space_range(&ds_nextnext->ds_deadlist,
1782 			    ds->ds_phys->ds_prev_snap_txg,
1783 			    ds->ds_phys->ds_creation_txg,
1784 			    &used, &comp, &uncomp);
1785 			ds_next->ds_phys->ds_unique_bytes += used;
1786 			dsl_dataset_rele(ds_nextnext, FTAG);
1787 			ASSERT3P(ds_next->ds_prev, ==, NULL);
1788 
1789 			/* Collapse range in this head. */
1790 			dsl_dataset_t *hds;
1791 			VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
1792 			    ds->ds_dir->dd_phys->dd_head_dataset_obj,
1793 			    FTAG, &hds));
1794 			dsl_deadlist_remove_key(&hds->ds_deadlist,
1795 			    ds->ds_phys->ds_creation_txg, tx);
1796 			dsl_dataset_rele(hds, FTAG);
1797 
1798 		} else {
1799 			ASSERT3P(ds_next->ds_prev, ==, ds);
1800 			dsl_dataset_drop_ref(ds_next->ds_prev, ds_next);
1801 			ds_next->ds_prev = NULL;
1802 			if (ds_prev) {
1803 				VERIFY(0 == dsl_dataset_get_ref(dp,
1804 				    ds->ds_phys->ds_prev_snap_obj,
1805 				    ds_next, &ds_next->ds_prev));
1806 			}
1807 
1808 			dsl_dataset_recalc_head_uniq(ds_next);
1809 
1810 			/*
1811 			 * Reduce the amount of our unconsmed refreservation
1812 			 * being charged to our parent by the amount of
1813 			 * new unique data we have gained.
1814 			 */
1815 			if (old_unique < ds_next->ds_reserved) {
1816 				int64_t mrsdelta;
1817 				uint64_t new_unique =
1818 				    ds_next->ds_phys->ds_unique_bytes;
1819 
1820 				ASSERT(old_unique <= new_unique);
1821 				mrsdelta = MIN(new_unique - old_unique,
1822 				    ds_next->ds_reserved - old_unique);
1823 				dsl_dir_diduse_space(ds->ds_dir,
1824 				    DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
1825 			}
1826 		}
1827 		dsl_dataset_rele(ds_next, FTAG);
1828 	} else {
1829 		/*
1830 		 * There's no next snapshot, so this is a head dataset.
1831 		 * Destroy the deadlist.  Unless it's a clone, the
1832 		 * deadlist should be empty.  (If it's a clone, it's
1833 		 * safe to ignore the deadlist contents.)
1834 		 */
1835 		struct killarg ka;
1836 
1837 		dsl_deadlist_close(&ds->ds_deadlist);
1838 		dsl_deadlist_free(mos, ds->ds_phys->ds_deadlist_obj, tx);
1839 		ds->ds_phys->ds_deadlist_obj = 0;
1840 
1841 		/*
1842 		 * Free everything that we point to (that's born after
1843 		 * the previous snapshot, if we are a clone)
1844 		 *
1845 		 * NB: this should be very quick, because we already
1846 		 * freed all the objects in open context.
1847 		 */
1848 		ka.ds = ds;
1849 		ka.tx = tx;
1850 		err = traverse_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
1851 		    TRAVERSE_POST, kill_blkptr, &ka);
1852 		ASSERT3U(err, ==, 0);
1853 		ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
1854 		    ds->ds_phys->ds_unique_bytes == 0);
1855 
1856 		if (ds->ds_prev != NULL) {
1857 			if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
1858 				VERIFY3U(0, ==, zap_remove_int(mos,
1859 				    ds->ds_prev->ds_dir->dd_phys->dd_clones,
1860 				    ds->ds_object, tx));
1861 			}
1862 			dsl_dataset_rele(ds->ds_prev, ds);
1863 			ds->ds_prev = ds_prev = NULL;
1864 		}
1865 	}
1866 
1867 	/*
1868 	 * This must be done after the dsl_traverse(), because it will
1869 	 * re-open the objset.
1870 	 */
1871 	if (ds->ds_objset) {
1872 		dmu_objset_evict(ds->ds_objset);
1873 		ds->ds_objset = NULL;
1874 	}
1875 
1876 	if (ds->ds_dir->dd_phys->dd_head_dataset_obj == ds->ds_object) {
1877 		/* Erase the link in the dir */
1878 		dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
1879 		ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
1880 		ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
1881 		err = zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx);
1882 		ASSERT(err == 0);
1883 	} else {
1884 		/* remove from snapshot namespace */
1885 		dsl_dataset_t *ds_head;
1886 		ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
1887 		VERIFY(0 == dsl_dataset_hold_obj(dp,
1888 		    ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
1889 		VERIFY(0 == dsl_dataset_get_snapname(ds));
1890 #ifdef ZFS_DEBUG
1891 		{
1892 			uint64_t val;
1893 
1894 			err = dsl_dataset_snap_lookup(ds_head,
1895 			    ds->ds_snapname, &val);
1896 			ASSERT3U(err, ==, 0);
1897 			ASSERT3U(val, ==, obj);
1898 		}
1899 #endif
1900 		err = dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx);
1901 		ASSERT(err == 0);
1902 		dsl_dataset_rele(ds_head, FTAG);
1903 	}
1904 
1905 	if (ds_prev && ds->ds_prev != ds_prev)
1906 		dsl_dataset_rele(ds_prev, FTAG);
1907 
1908 	spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
1909 	spa_history_log_internal(LOG_DS_DESTROY, dp->dp_spa, tx,
1910 	    "dataset = %llu", ds->ds_object);
1911 
1912 	if (ds->ds_phys->ds_next_clones_obj != 0) {
1913 		uint64_t count;
1914 		ASSERT(0 == zap_count(mos,
1915 		    ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
1916 		VERIFY(0 == dmu_object_free(mos,
1917 		    ds->ds_phys->ds_next_clones_obj, tx));
1918 	}
1919 	if (ds->ds_phys->ds_props_obj != 0)
1920 		VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
1921 	if (ds->ds_phys->ds_userrefs_obj != 0)
1922 		VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
1923 	dsl_dir_close(ds->ds_dir, ds);
1924 	ds->ds_dir = NULL;
1925 	dsl_dataset_drain_refs(ds, tag);
1926 	VERIFY(0 == dmu_object_free(mos, obj, tx));
1927 
1928 	if (dsda->rm_origin) {
1929 		/*
1930 		 * Remove the origin of the clone we just destroyed.
1931 		 */
1932 		struct dsl_ds_destroyarg ndsda = {0};
1933 
1934 		ndsda.ds = dsda->rm_origin;
1935 		dsl_dataset_destroy_sync(&ndsda, tag, tx);
1936 	}
1937 }
1938 
1939 static int
1940 dsl_dataset_snapshot_reserve_space(dsl_dataset_t *ds, dmu_tx_t *tx)
1941 {
1942 	uint64_t asize;
1943 
1944 	if (!dmu_tx_is_syncing(tx))
1945 		return (0);
1946 
1947 	/*
1948 	 * If there's an fs-only reservation, any blocks that might become
1949 	 * owned by the snapshot dataset must be accommodated by space
1950 	 * outside of the reservation.
1951 	 */
1952 	ASSERT(ds->ds_reserved == 0 || DS_UNIQUE_IS_ACCURATE(ds));
1953 	asize = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
1954 	if (asize > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
1955 		return (ENOSPC);
1956 
1957 	/*
1958 	 * Propogate any reserved space for this snapshot to other
1959 	 * snapshot checks in this sync group.
1960 	 */
1961 	if (asize > 0)
1962 		dsl_dir_willuse_space(ds->ds_dir, asize, tx);
1963 
1964 	return (0);
1965 }
1966 
1967 int
1968 dsl_dataset_snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
1969 {
1970 	dsl_dataset_t *ds = arg1;
1971 	const char *snapname = arg2;
1972 	int err;
1973 	uint64_t value;
1974 
1975 	/*
1976 	 * We don't allow multiple snapshots of the same txg.  If there
1977 	 * is already one, try again.
1978 	 */
1979 	if (ds->ds_phys->ds_prev_snap_txg >= tx->tx_txg)
1980 		return (EAGAIN);
1981 
1982 	/*
1983 	 * Check for conflicting name snapshot name.
1984 	 */
1985 	err = dsl_dataset_snap_lookup(ds, snapname, &value);
1986 	if (err == 0)
1987 		return (EEXIST);
1988 	if (err != ENOENT)
1989 		return (err);
1990 
1991 	/*
1992 	 * Check that the dataset's name is not too long.  Name consists
1993 	 * of the dataset's length + 1 for the @-sign + snapshot name's length
1994 	 */
1995 	if (dsl_dataset_namelen(ds) + 1 + strlen(snapname) >= MAXNAMELEN)
1996 		return (ENAMETOOLONG);
1997 
1998 	err = dsl_dataset_snapshot_reserve_space(ds, tx);
1999 	if (err)
2000 		return (err);
2001 
2002 	ds->ds_trysnap_txg = tx->tx_txg;
2003 	return (0);
2004 }
2005 
2006 void
2007 dsl_dataset_snapshot_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2008 {
2009 	dsl_dataset_t *ds = arg1;
2010 	const char *snapname = arg2;
2011 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
2012 	dmu_buf_t *dbuf;
2013 	dsl_dataset_phys_t *dsphys;
2014 	uint64_t dsobj, crtxg;
2015 	objset_t *mos = dp->dp_meta_objset;
2016 	int err;
2017 
2018 	ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
2019 
2020 	/*
2021 	 * The origin's ds_creation_txg has to be < TXG_INITIAL
2022 	 */
2023 	if (strcmp(snapname, ORIGIN_DIR_NAME) == 0)
2024 		crtxg = 1;
2025 	else
2026 		crtxg = tx->tx_txg;
2027 
2028 	dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
2029 	    DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
2030 	VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
2031 	dmu_buf_will_dirty(dbuf, tx);
2032 	dsphys = dbuf->db_data;
2033 	bzero(dsphys, sizeof (dsl_dataset_phys_t));
2034 	dsphys->ds_dir_obj = ds->ds_dir->dd_object;
2035 	dsphys->ds_fsid_guid = unique_create();
2036 	(void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
2037 	    sizeof (dsphys->ds_guid));
2038 	dsphys->ds_prev_snap_obj = ds->ds_phys->ds_prev_snap_obj;
2039 	dsphys->ds_prev_snap_txg = ds->ds_phys->ds_prev_snap_txg;
2040 	dsphys->ds_next_snap_obj = ds->ds_object;
2041 	dsphys->ds_num_children = 1;
2042 	dsphys->ds_creation_time = gethrestime_sec();
2043 	dsphys->ds_creation_txg = crtxg;
2044 	dsphys->ds_deadlist_obj = ds->ds_phys->ds_deadlist_obj;
2045 	dsphys->ds_used_bytes = ds->ds_phys->ds_used_bytes;
2046 	dsphys->ds_compressed_bytes = ds->ds_phys->ds_compressed_bytes;
2047 	dsphys->ds_uncompressed_bytes = ds->ds_phys->ds_uncompressed_bytes;
2048 	dsphys->ds_flags = ds->ds_phys->ds_flags;
2049 	dsphys->ds_bp = ds->ds_phys->ds_bp;
2050 	dmu_buf_rele(dbuf, FTAG);
2051 
2052 	ASSERT3U(ds->ds_prev != 0, ==, ds->ds_phys->ds_prev_snap_obj != 0);
2053 	if (ds->ds_prev) {
2054 		uint64_t next_clones_obj =
2055 		    ds->ds_prev->ds_phys->ds_next_clones_obj;
2056 		ASSERT(ds->ds_prev->ds_phys->ds_next_snap_obj ==
2057 		    ds->ds_object ||
2058 		    ds->ds_prev->ds_phys->ds_num_children > 1);
2059 		if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
2060 			dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2061 			ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
2062 			    ds->ds_prev->ds_phys->ds_creation_txg);
2063 			ds->ds_prev->ds_phys->ds_next_snap_obj = dsobj;
2064 		} else if (next_clones_obj != 0) {
2065 			remove_from_next_clones(ds->ds_prev,
2066 			    dsphys->ds_next_snap_obj, tx);
2067 			VERIFY3U(0, ==, zap_add_int(mos,
2068 			    next_clones_obj, dsobj, tx));
2069 		}
2070 	}
2071 
2072 	/*
2073 	 * If we have a reference-reservation on this dataset, we will
2074 	 * need to increase the amount of refreservation being charged
2075 	 * since our unique space is going to zero.
2076 	 */
2077 	if (ds->ds_reserved) {
2078 		int64_t delta;
2079 		ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
2080 		delta = MIN(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
2081 		dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV,
2082 		    delta, 0, 0, tx);
2083 	}
2084 
2085 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
2086 	zfs_dbgmsg("taking snapshot %s@%s/%llu; newkey=%llu",
2087 	    ds->ds_dir->dd_myname, snapname, dsobj,
2088 	    ds->ds_phys->ds_prev_snap_txg);
2089 	ds->ds_phys->ds_deadlist_obj = dsl_deadlist_clone(&ds->ds_deadlist,
2090 	    UINT64_MAX, ds->ds_phys->ds_prev_snap_obj, tx);
2091 	dsl_deadlist_close(&ds->ds_deadlist);
2092 	dsl_deadlist_open(&ds->ds_deadlist, mos, ds->ds_phys->ds_deadlist_obj);
2093 	dsl_deadlist_add_key(&ds->ds_deadlist,
2094 	    ds->ds_phys->ds_prev_snap_txg, tx);
2095 
2096 	ASSERT3U(ds->ds_phys->ds_prev_snap_txg, <, tx->tx_txg);
2097 	ds->ds_phys->ds_prev_snap_obj = dsobj;
2098 	ds->ds_phys->ds_prev_snap_txg = crtxg;
2099 	ds->ds_phys->ds_unique_bytes = 0;
2100 	if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
2101 		ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
2102 
2103 	err = zap_add(mos, ds->ds_phys->ds_snapnames_zapobj,
2104 	    snapname, 8, 1, &dsobj, tx);
2105 	ASSERT(err == 0);
2106 
2107 	if (ds->ds_prev)
2108 		dsl_dataset_drop_ref(ds->ds_prev, ds);
2109 	VERIFY(0 == dsl_dataset_get_ref(dp,
2110 	    ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
2111 
2112 	dsl_scan_ds_snapshotted(ds, tx);
2113 
2114 	dsl_dir_snap_cmtime_update(ds->ds_dir);
2115 
2116 	spa_history_log_internal(LOG_DS_SNAPSHOT, dp->dp_spa, tx,
2117 	    "dataset = %llu", dsobj);
2118 }
2119 
2120 void
2121 dsl_dataset_sync(dsl_dataset_t *ds, zio_t *zio, dmu_tx_t *tx)
2122 {
2123 	ASSERT(dmu_tx_is_syncing(tx));
2124 	ASSERT(ds->ds_objset != NULL);
2125 	ASSERT(ds->ds_phys->ds_next_snap_obj == 0);
2126 
2127 	/*
2128 	 * in case we had to change ds_fsid_guid when we opened it,
2129 	 * sync it out now.
2130 	 */
2131 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
2132 	ds->ds_phys->ds_fsid_guid = ds->ds_fsid_guid;
2133 
2134 	dsl_dir_dirty(ds->ds_dir, tx);
2135 	dmu_objset_sync(ds->ds_objset, zio, tx);
2136 }
2137 
2138 void
2139 dsl_dataset_stats(dsl_dataset_t *ds, nvlist_t *nv)
2140 {
2141 	uint64_t refd, avail, uobjs, aobjs;
2142 
2143 	dsl_dir_stats(ds->ds_dir, nv);
2144 
2145 	dsl_dataset_space(ds, &refd, &avail, &uobjs, &aobjs);
2146 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_AVAILABLE, avail);
2147 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFERENCED, refd);
2148 
2149 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATION,
2150 	    ds->ds_phys->ds_creation_time);
2151 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATETXG,
2152 	    ds->ds_phys->ds_creation_txg);
2153 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFQUOTA,
2154 	    ds->ds_quota);
2155 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRESERVATION,
2156 	    ds->ds_reserved);
2157 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_GUID,
2158 	    ds->ds_phys->ds_guid);
2159 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_UNIQUE,
2160 	    ds->ds_phys->ds_unique_bytes);
2161 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_OBJSETID,
2162 	    ds->ds_object);
2163 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USERREFS,
2164 	    ds->ds_userrefs);
2165 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_DEFER_DESTROY,
2166 	    DS_IS_DEFER_DESTROY(ds) ? 1 : 0);
2167 
2168 	if (ds->ds_phys->ds_next_snap_obj) {
2169 		/*
2170 		 * This is a snapshot; override the dd's space used with
2171 		 * our unique space and compression ratio.
2172 		 */
2173 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USED,
2174 		    ds->ds_phys->ds_unique_bytes);
2175 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_COMPRESSRATIO,
2176 		    ds->ds_phys->ds_compressed_bytes == 0 ? 100 :
2177 		    (ds->ds_phys->ds_uncompressed_bytes * 100 /
2178 		    ds->ds_phys->ds_compressed_bytes));
2179 	}
2180 }
2181 
2182 void
2183 dsl_dataset_fast_stat(dsl_dataset_t *ds, dmu_objset_stats_t *stat)
2184 {
2185 	stat->dds_creation_txg = ds->ds_phys->ds_creation_txg;
2186 	stat->dds_inconsistent = ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT;
2187 	stat->dds_guid = ds->ds_phys->ds_guid;
2188 	if (ds->ds_phys->ds_next_snap_obj) {
2189 		stat->dds_is_snapshot = B_TRUE;
2190 		stat->dds_num_clones = ds->ds_phys->ds_num_children - 1;
2191 	} else {
2192 		stat->dds_is_snapshot = B_FALSE;
2193 		stat->dds_num_clones = 0;
2194 	}
2195 
2196 	/* clone origin is really a dsl_dir thing... */
2197 	rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2198 	if (dsl_dir_is_clone(ds->ds_dir)) {
2199 		dsl_dataset_t *ods;
2200 
2201 		VERIFY(0 == dsl_dataset_get_ref(ds->ds_dir->dd_pool,
2202 		    ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &ods));
2203 		dsl_dataset_name(ods, stat->dds_origin);
2204 		dsl_dataset_drop_ref(ods, FTAG);
2205 	} else {
2206 		stat->dds_origin[0] = '\0';
2207 	}
2208 	rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2209 }
2210 
2211 uint64_t
2212 dsl_dataset_fsid_guid(dsl_dataset_t *ds)
2213 {
2214 	return (ds->ds_fsid_guid);
2215 }
2216 
2217 void
2218 dsl_dataset_space(dsl_dataset_t *ds,
2219     uint64_t *refdbytesp, uint64_t *availbytesp,
2220     uint64_t *usedobjsp, uint64_t *availobjsp)
2221 {
2222 	*refdbytesp = ds->ds_phys->ds_used_bytes;
2223 	*availbytesp = dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE);
2224 	if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes)
2225 		*availbytesp += ds->ds_reserved - ds->ds_phys->ds_unique_bytes;
2226 	if (ds->ds_quota != 0) {
2227 		/*
2228 		 * Adjust available bytes according to refquota
2229 		 */
2230 		if (*refdbytesp < ds->ds_quota)
2231 			*availbytesp = MIN(*availbytesp,
2232 			    ds->ds_quota - *refdbytesp);
2233 		else
2234 			*availbytesp = 0;
2235 	}
2236 	*usedobjsp = ds->ds_phys->ds_bp.blk_fill;
2237 	*availobjsp = DN_MAX_OBJECT - *usedobjsp;
2238 }
2239 
2240 boolean_t
2241 dsl_dataset_modified_since_lastsnap(dsl_dataset_t *ds)
2242 {
2243 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
2244 
2245 	ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
2246 	    dsl_pool_sync_context(dp));
2247 	if (ds->ds_prev == NULL)
2248 		return (B_FALSE);
2249 	if (ds->ds_phys->ds_bp.blk_birth >
2250 	    ds->ds_prev->ds_phys->ds_creation_txg) {
2251 		objset_t *os, *os_prev;
2252 		/*
2253 		 * It may be that only the ZIL differs, because it was
2254 		 * reset in the head.  Don't count that as being
2255 		 * modified.
2256 		 */
2257 		if (dmu_objset_from_ds(ds, &os) != 0)
2258 			return (B_TRUE);
2259 		if (dmu_objset_from_ds(ds->ds_prev, &os_prev) != 0)
2260 			return (B_TRUE);
2261 		return (bcmp(&os->os_phys->os_meta_dnode,
2262 		    &os_prev->os_phys->os_meta_dnode,
2263 		    sizeof (os->os_phys->os_meta_dnode)) != 0);
2264 	}
2265 	return (B_FALSE);
2266 }
2267 
2268 /* ARGSUSED */
2269 static int
2270 dsl_dataset_snapshot_rename_check(void *arg1, void *arg2, dmu_tx_t *tx)
2271 {
2272 	dsl_dataset_t *ds = arg1;
2273 	char *newsnapname = arg2;
2274 	dsl_dir_t *dd = ds->ds_dir;
2275 	dsl_dataset_t *hds;
2276 	uint64_t val;
2277 	int err;
2278 
2279 	err = dsl_dataset_hold_obj(dd->dd_pool,
2280 	    dd->dd_phys->dd_head_dataset_obj, FTAG, &hds);
2281 	if (err)
2282 		return (err);
2283 
2284 	/* new name better not be in use */
2285 	err = dsl_dataset_snap_lookup(hds, newsnapname, &val);
2286 	dsl_dataset_rele(hds, FTAG);
2287 
2288 	if (err == 0)
2289 		err = EEXIST;
2290 	else if (err == ENOENT)
2291 		err = 0;
2292 
2293 	/* dataset name + 1 for the "@" + the new snapshot name must fit */
2294 	if (dsl_dir_namelen(ds->ds_dir) + 1 + strlen(newsnapname) >= MAXNAMELEN)
2295 		err = ENAMETOOLONG;
2296 
2297 	return (err);
2298 }
2299 
2300 static void
2301 dsl_dataset_snapshot_rename_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2302 {
2303 	dsl_dataset_t *ds = arg1;
2304 	const char *newsnapname = arg2;
2305 	dsl_dir_t *dd = ds->ds_dir;
2306 	objset_t *mos = dd->dd_pool->dp_meta_objset;
2307 	dsl_dataset_t *hds;
2308 	int err;
2309 
2310 	ASSERT(ds->ds_phys->ds_next_snap_obj != 0);
2311 
2312 	VERIFY(0 == dsl_dataset_hold_obj(dd->dd_pool,
2313 	    dd->dd_phys->dd_head_dataset_obj, FTAG, &hds));
2314 
2315 	VERIFY(0 == dsl_dataset_get_snapname(ds));
2316 	err = dsl_dataset_snap_remove(hds, ds->ds_snapname, tx);
2317 	ASSERT3U(err, ==, 0);
2318 	mutex_enter(&ds->ds_lock);
2319 	(void) strcpy(ds->ds_snapname, newsnapname);
2320 	mutex_exit(&ds->ds_lock);
2321 	err = zap_add(mos, hds->ds_phys->ds_snapnames_zapobj,
2322 	    ds->ds_snapname, 8, 1, &ds->ds_object, tx);
2323 	ASSERT3U(err, ==, 0);
2324 
2325 	spa_history_log_internal(LOG_DS_RENAME, dd->dd_pool->dp_spa, tx,
2326 	    "dataset = %llu", ds->ds_object);
2327 	dsl_dataset_rele(hds, FTAG);
2328 }
2329 
2330 struct renamesnaparg {
2331 	dsl_sync_task_group_t *dstg;
2332 	char failed[MAXPATHLEN];
2333 	char *oldsnap;
2334 	char *newsnap;
2335 };
2336 
2337 static int
2338 dsl_snapshot_rename_one(const char *name, void *arg)
2339 {
2340 	struct renamesnaparg *ra = arg;
2341 	dsl_dataset_t *ds = NULL;
2342 	char *snapname;
2343 	int err;
2344 
2345 	snapname = kmem_asprintf("%s@%s", name, ra->oldsnap);
2346 	(void) strlcpy(ra->failed, snapname, sizeof (ra->failed));
2347 
2348 	/*
2349 	 * For recursive snapshot renames the parent won't be changing
2350 	 * so we just pass name for both the to/from argument.
2351 	 */
2352 	err = zfs_secpolicy_rename_perms(snapname, snapname, CRED());
2353 	if (err != 0) {
2354 		strfree(snapname);
2355 		return (err == ENOENT ? 0 : err);
2356 	}
2357 
2358 #ifdef _KERNEL
2359 	/*
2360 	 * For all filesystems undergoing rename, we'll need to unmount it.
2361 	 */
2362 	(void) zfs_unmount_snap(snapname, NULL);
2363 #endif
2364 	err = dsl_dataset_hold(snapname, ra->dstg, &ds);
2365 	strfree(snapname);
2366 	if (err != 0)
2367 		return (err == ENOENT ? 0 : err);
2368 
2369 	dsl_sync_task_create(ra->dstg, dsl_dataset_snapshot_rename_check,
2370 	    dsl_dataset_snapshot_rename_sync, ds, ra->newsnap, 0);
2371 
2372 	return (0);
2373 }
2374 
2375 static int
2376 dsl_recursive_rename(char *oldname, const char *newname)
2377 {
2378 	int err;
2379 	struct renamesnaparg *ra;
2380 	dsl_sync_task_t *dst;
2381 	spa_t *spa;
2382 	char *cp, *fsname = spa_strdup(oldname);
2383 	int len = strlen(oldname) + 1;
2384 
2385 	/* truncate the snapshot name to get the fsname */
2386 	cp = strchr(fsname, '@');
2387 	*cp = '\0';
2388 
2389 	err = spa_open(fsname, &spa, FTAG);
2390 	if (err) {
2391 		kmem_free(fsname, len);
2392 		return (err);
2393 	}
2394 	ra = kmem_alloc(sizeof (struct renamesnaparg), KM_SLEEP);
2395 	ra->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
2396 
2397 	ra->oldsnap = strchr(oldname, '@') + 1;
2398 	ra->newsnap = strchr(newname, '@') + 1;
2399 	*ra->failed = '\0';
2400 
2401 	err = dmu_objset_find(fsname, dsl_snapshot_rename_one, ra,
2402 	    DS_FIND_CHILDREN);
2403 	kmem_free(fsname, len);
2404 
2405 	if (err == 0) {
2406 		err = dsl_sync_task_group_wait(ra->dstg);
2407 	}
2408 
2409 	for (dst = list_head(&ra->dstg->dstg_tasks); dst;
2410 	    dst = list_next(&ra->dstg->dstg_tasks, dst)) {
2411 		dsl_dataset_t *ds = dst->dst_arg1;
2412 		if (dst->dst_err) {
2413 			dsl_dir_name(ds->ds_dir, ra->failed);
2414 			(void) strlcat(ra->failed, "@", sizeof (ra->failed));
2415 			(void) strlcat(ra->failed, ra->newsnap,
2416 			    sizeof (ra->failed));
2417 		}
2418 		dsl_dataset_rele(ds, ra->dstg);
2419 	}
2420 
2421 	if (err)
2422 		(void) strlcpy(oldname, ra->failed, sizeof (ra->failed));
2423 
2424 	dsl_sync_task_group_destroy(ra->dstg);
2425 	kmem_free(ra, sizeof (struct renamesnaparg));
2426 	spa_close(spa, FTAG);
2427 	return (err);
2428 }
2429 
2430 static int
2431 dsl_valid_rename(const char *oldname, void *arg)
2432 {
2433 	int delta = *(int *)arg;
2434 
2435 	if (strlen(oldname) + delta >= MAXNAMELEN)
2436 		return (ENAMETOOLONG);
2437 
2438 	return (0);
2439 }
2440 
2441 #pragma weak dmu_objset_rename = dsl_dataset_rename
2442 int
2443 dsl_dataset_rename(char *oldname, const char *newname, boolean_t recursive)
2444 {
2445 	dsl_dir_t *dd;
2446 	dsl_dataset_t *ds;
2447 	const char *tail;
2448 	int err;
2449 
2450 	err = dsl_dir_open(oldname, FTAG, &dd, &tail);
2451 	if (err)
2452 		return (err);
2453 
2454 	if (tail == NULL) {
2455 		int delta = strlen(newname) - strlen(oldname);
2456 
2457 		/* if we're growing, validate child name lengths */
2458 		if (delta > 0)
2459 			err = dmu_objset_find(oldname, dsl_valid_rename,
2460 			    &delta, DS_FIND_CHILDREN | DS_FIND_SNAPSHOTS);
2461 
2462 		if (err == 0)
2463 			err = dsl_dir_rename(dd, newname);
2464 		dsl_dir_close(dd, FTAG);
2465 		return (err);
2466 	}
2467 
2468 	if (tail[0] != '@') {
2469 		/* the name ended in a nonexistent component */
2470 		dsl_dir_close(dd, FTAG);
2471 		return (ENOENT);
2472 	}
2473 
2474 	dsl_dir_close(dd, FTAG);
2475 
2476 	/* new name must be snapshot in same filesystem */
2477 	tail = strchr(newname, '@');
2478 	if (tail == NULL)
2479 		return (EINVAL);
2480 	tail++;
2481 	if (strncmp(oldname, newname, tail - newname) != 0)
2482 		return (EXDEV);
2483 
2484 	if (recursive) {
2485 		err = dsl_recursive_rename(oldname, newname);
2486 	} else {
2487 		err = dsl_dataset_hold(oldname, FTAG, &ds);
2488 		if (err)
2489 			return (err);
2490 
2491 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
2492 		    dsl_dataset_snapshot_rename_check,
2493 		    dsl_dataset_snapshot_rename_sync, ds, (char *)tail, 1);
2494 
2495 		dsl_dataset_rele(ds, FTAG);
2496 	}
2497 
2498 	return (err);
2499 }
2500 
2501 struct promotenode {
2502 	list_node_t link;
2503 	dsl_dataset_t *ds;
2504 };
2505 
2506 struct promotearg {
2507 	list_t shared_snaps, origin_snaps, clone_snaps;
2508 	dsl_dataset_t *origin_origin;
2509 	uint64_t used, comp, uncomp, unique, cloneusedsnap, originusedsnap;
2510 	char *err_ds;
2511 };
2512 
2513 static int snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep);
2514 static boolean_t snaplist_unstable(list_t *l);
2515 
2516 static int
2517 dsl_dataset_promote_check(void *arg1, void *arg2, dmu_tx_t *tx)
2518 {
2519 	dsl_dataset_t *hds = arg1;
2520 	struct promotearg *pa = arg2;
2521 	struct promotenode *snap = list_head(&pa->shared_snaps);
2522 	dsl_dataset_t *origin_ds = snap->ds;
2523 	int err;
2524 	uint64_t unused;
2525 
2526 	/* Check that it is a real clone */
2527 	if (!dsl_dir_is_clone(hds->ds_dir))
2528 		return (EINVAL);
2529 
2530 	/* Since this is so expensive, don't do the preliminary check */
2531 	if (!dmu_tx_is_syncing(tx))
2532 		return (0);
2533 
2534 	if (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE)
2535 		return (EXDEV);
2536 
2537 	/* compute origin's new unique space */
2538 	snap = list_tail(&pa->clone_snaps);
2539 	ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2540 	dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2541 	    origin_ds->ds_phys->ds_prev_snap_txg, UINT64_MAX,
2542 	    &pa->unique, &unused, &unused);
2543 
2544 	/*
2545 	 * Walk the snapshots that we are moving
2546 	 *
2547 	 * Compute space to transfer.  Consider the incremental changes
2548 	 * to used for each snapshot:
2549 	 * (my used) = (prev's used) + (blocks born) - (blocks killed)
2550 	 * So each snapshot gave birth to:
2551 	 * (blocks born) = (my used) - (prev's used) + (blocks killed)
2552 	 * So a sequence would look like:
2553 	 * (uN - u(N-1) + kN) + ... + (u1 - u0 + k1) + (u0 - 0 + k0)
2554 	 * Which simplifies to:
2555 	 * uN + kN + kN-1 + ... + k1 + k0
2556 	 * Note however, if we stop before we reach the ORIGIN we get:
2557 	 * uN + kN + kN-1 + ... + kM - uM-1
2558 	 */
2559 	pa->used = origin_ds->ds_phys->ds_used_bytes;
2560 	pa->comp = origin_ds->ds_phys->ds_compressed_bytes;
2561 	pa->uncomp = origin_ds->ds_phys->ds_uncompressed_bytes;
2562 	for (snap = list_head(&pa->shared_snaps); snap;
2563 	    snap = list_next(&pa->shared_snaps, snap)) {
2564 		uint64_t val, dlused, dlcomp, dluncomp;
2565 		dsl_dataset_t *ds = snap->ds;
2566 
2567 		/* Check that the snapshot name does not conflict */
2568 		VERIFY(0 == dsl_dataset_get_snapname(ds));
2569 		err = dsl_dataset_snap_lookup(hds, ds->ds_snapname, &val);
2570 		if (err == 0) {
2571 			err = EEXIST;
2572 			goto out;
2573 		}
2574 		if (err != ENOENT)
2575 			goto out;
2576 
2577 		/* The very first snapshot does not have a deadlist */
2578 		if (ds->ds_phys->ds_prev_snap_obj == 0)
2579 			continue;
2580 
2581 		dsl_deadlist_space(&ds->ds_deadlist,
2582 		    &dlused, &dlcomp, &dluncomp);
2583 		pa->used += dlused;
2584 		pa->comp += dlcomp;
2585 		pa->uncomp += dluncomp;
2586 	}
2587 
2588 	/*
2589 	 * If we are a clone of a clone then we never reached ORIGIN,
2590 	 * so we need to subtract out the clone origin's used space.
2591 	 */
2592 	if (pa->origin_origin) {
2593 		pa->used -= pa->origin_origin->ds_phys->ds_used_bytes;
2594 		pa->comp -= pa->origin_origin->ds_phys->ds_compressed_bytes;
2595 		pa->uncomp -= pa->origin_origin->ds_phys->ds_uncompressed_bytes;
2596 	}
2597 
2598 	/* Check that there is enough space here */
2599 	err = dsl_dir_transfer_possible(origin_ds->ds_dir, hds->ds_dir,
2600 	    pa->used);
2601 	if (err)
2602 		return (err);
2603 
2604 	/*
2605 	 * Compute the amounts of space that will be used by snapshots
2606 	 * after the promotion (for both origin and clone).  For each,
2607 	 * it is the amount of space that will be on all of their
2608 	 * deadlists (that was not born before their new origin).
2609 	 */
2610 	if (hds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2611 		uint64_t space;
2612 
2613 		/*
2614 		 * Note, typically this will not be a clone of a clone,
2615 		 * so dd_origin_txg will be < TXG_INITIAL, so
2616 		 * these snaplist_space() -> dsl_deadlist_space_range()
2617 		 * calls will be fast because they do not have to
2618 		 * iterate over all bps.
2619 		 */
2620 		snap = list_head(&pa->origin_snaps);
2621 		err = snaplist_space(&pa->shared_snaps,
2622 		    snap->ds->ds_dir->dd_origin_txg, &pa->cloneusedsnap);
2623 		if (err)
2624 			return (err);
2625 
2626 		err = snaplist_space(&pa->clone_snaps,
2627 		    snap->ds->ds_dir->dd_origin_txg, &space);
2628 		if (err)
2629 			return (err);
2630 		pa->cloneusedsnap += space;
2631 	}
2632 	if (origin_ds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2633 		err = snaplist_space(&pa->origin_snaps,
2634 		    origin_ds->ds_phys->ds_creation_txg, &pa->originusedsnap);
2635 		if (err)
2636 			return (err);
2637 	}
2638 
2639 	return (0);
2640 out:
2641 	pa->err_ds =  snap->ds->ds_snapname;
2642 	return (err);
2643 }
2644 
2645 static void
2646 dsl_dataset_promote_sync(void *arg1, void *arg2, dmu_tx_t *tx)
2647 {
2648 	dsl_dataset_t *hds = arg1;
2649 	struct promotearg *pa = arg2;
2650 	struct promotenode *snap = list_head(&pa->shared_snaps);
2651 	dsl_dataset_t *origin_ds = snap->ds;
2652 	dsl_dataset_t *origin_head;
2653 	dsl_dir_t *dd = hds->ds_dir;
2654 	dsl_pool_t *dp = hds->ds_dir->dd_pool;
2655 	dsl_dir_t *odd = NULL;
2656 	uint64_t oldnext_obj;
2657 	int64_t delta;
2658 
2659 	ASSERT(0 == (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE));
2660 
2661 	snap = list_head(&pa->origin_snaps);
2662 	origin_head = snap->ds;
2663 
2664 	/*
2665 	 * We need to explicitly open odd, since origin_ds's dd will be
2666 	 * changing.
2667 	 */
2668 	VERIFY(0 == dsl_dir_open_obj(dp, origin_ds->ds_dir->dd_object,
2669 	    NULL, FTAG, &odd));
2670 
2671 	/* change origin's next snap */
2672 	dmu_buf_will_dirty(origin_ds->ds_dbuf, tx);
2673 	oldnext_obj = origin_ds->ds_phys->ds_next_snap_obj;
2674 	snap = list_tail(&pa->clone_snaps);
2675 	ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2676 	origin_ds->ds_phys->ds_next_snap_obj = snap->ds->ds_object;
2677 
2678 	/* change the origin's next clone */
2679 	if (origin_ds->ds_phys->ds_next_clones_obj) {
2680 		remove_from_next_clones(origin_ds, snap->ds->ds_object, tx);
2681 		VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2682 		    origin_ds->ds_phys->ds_next_clones_obj,
2683 		    oldnext_obj, tx));
2684 	}
2685 
2686 	/* change origin */
2687 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
2688 	ASSERT3U(dd->dd_phys->dd_origin_obj, ==, origin_ds->ds_object);
2689 	dd->dd_phys->dd_origin_obj = odd->dd_phys->dd_origin_obj;
2690 	dd->dd_origin_txg = origin_head->ds_dir->dd_origin_txg;
2691 	dmu_buf_will_dirty(odd->dd_dbuf, tx);
2692 	odd->dd_phys->dd_origin_obj = origin_ds->ds_object;
2693 	origin_head->ds_dir->dd_origin_txg =
2694 	    origin_ds->ds_phys->ds_creation_txg;
2695 
2696 	/* change dd_clone entries */
2697 	if (spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2698 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2699 		    odd->dd_phys->dd_clones, hds->ds_object, tx));
2700 		VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2701 		    pa->origin_origin->ds_dir->dd_phys->dd_clones,
2702 		    hds->ds_object, tx));
2703 
2704 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2705 		    pa->origin_origin->ds_dir->dd_phys->dd_clones,
2706 		    origin_head->ds_object, tx));
2707 		if (dd->dd_phys->dd_clones == 0) {
2708 			dd->dd_phys->dd_clones = zap_create(dp->dp_meta_objset,
2709 			    DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
2710 		}
2711 		VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2712 		    dd->dd_phys->dd_clones, origin_head->ds_object, tx));
2713 
2714 	}
2715 
2716 	/* move snapshots to this dir */
2717 	for (snap = list_head(&pa->shared_snaps); snap;
2718 	    snap = list_next(&pa->shared_snaps, snap)) {
2719 		dsl_dataset_t *ds = snap->ds;
2720 
2721 		/* unregister props as dsl_dir is changing */
2722 		if (ds->ds_objset) {
2723 			dmu_objset_evict(ds->ds_objset);
2724 			ds->ds_objset = NULL;
2725 		}
2726 		/* move snap name entry */
2727 		VERIFY(0 == dsl_dataset_get_snapname(ds));
2728 		VERIFY(0 == dsl_dataset_snap_remove(origin_head,
2729 		    ds->ds_snapname, tx));
2730 		VERIFY(0 == zap_add(dp->dp_meta_objset,
2731 		    hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
2732 		    8, 1, &ds->ds_object, tx));
2733 
2734 		/* change containing dsl_dir */
2735 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
2736 		ASSERT3U(ds->ds_phys->ds_dir_obj, ==, odd->dd_object);
2737 		ds->ds_phys->ds_dir_obj = dd->dd_object;
2738 		ASSERT3P(ds->ds_dir, ==, odd);
2739 		dsl_dir_close(ds->ds_dir, ds);
2740 		VERIFY(0 == dsl_dir_open_obj(dp, dd->dd_object,
2741 		    NULL, ds, &ds->ds_dir));
2742 
2743 		/* move any clone references */
2744 		if (ds->ds_phys->ds_next_clones_obj &&
2745 		    spa_version(dp->dp_spa) >= SPA_VERSION_DIR_CLONES) {
2746 			zap_cursor_t zc;
2747 			zap_attribute_t za;
2748 
2749 			for (zap_cursor_init(&zc, dp->dp_meta_objset,
2750 			    ds->ds_phys->ds_next_clones_obj);
2751 			    zap_cursor_retrieve(&zc, &za) == 0;
2752 			    zap_cursor_advance(&zc)) {
2753 				dsl_dataset_t *cnds;
2754 				uint64_t o;
2755 
2756 				if (za.za_first_integer == oldnext_obj) {
2757 					/*
2758 					 * We've already moved the
2759 					 * origin's reference.
2760 					 */
2761 					continue;
2762 				}
2763 
2764 				VERIFY3U(0, ==, dsl_dataset_hold_obj(dp,
2765 				    za.za_first_integer, FTAG, &cnds));
2766 				o = cnds->ds_dir->dd_phys->dd_head_dataset_obj;
2767 
2768 				VERIFY3U(zap_remove_int(dp->dp_meta_objset,
2769 				    odd->dd_phys->dd_clones, o, tx), ==, 0);
2770 				VERIFY3U(zap_add_int(dp->dp_meta_objset,
2771 				    dd->dd_phys->dd_clones, o, tx), ==, 0);
2772 				dsl_dataset_rele(cnds, FTAG);
2773 			}
2774 			zap_cursor_fini(&zc);
2775 		}
2776 
2777 		ASSERT3U(dsl_prop_numcb(ds), ==, 0);
2778 	}
2779 
2780 	/*
2781 	 * Change space accounting.
2782 	 * Note, pa->*usedsnap and dd_used_breakdown[SNAP] will either
2783 	 * both be valid, or both be 0 (resulting in delta == 0).  This
2784 	 * is true for each of {clone,origin} independently.
2785 	 */
2786 
2787 	delta = pa->cloneusedsnap -
2788 	    dd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2789 	ASSERT3S(delta, >=, 0);
2790 	ASSERT3U(pa->used, >=, delta);
2791 	dsl_dir_diduse_space(dd, DD_USED_SNAP, delta, 0, 0, tx);
2792 	dsl_dir_diduse_space(dd, DD_USED_HEAD,
2793 	    pa->used - delta, pa->comp, pa->uncomp, tx);
2794 
2795 	delta = pa->originusedsnap -
2796 	    odd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2797 	ASSERT3S(delta, <=, 0);
2798 	ASSERT3U(pa->used, >=, -delta);
2799 	dsl_dir_diduse_space(odd, DD_USED_SNAP, delta, 0, 0, tx);
2800 	dsl_dir_diduse_space(odd, DD_USED_HEAD,
2801 	    -pa->used - delta, -pa->comp, -pa->uncomp, tx);
2802 
2803 	origin_ds->ds_phys->ds_unique_bytes = pa->unique;
2804 
2805 	/* log history record */
2806 	spa_history_log_internal(LOG_DS_PROMOTE, dd->dd_pool->dp_spa, tx,
2807 	    "dataset = %llu", hds->ds_object);
2808 
2809 	dsl_dir_close(odd, FTAG);
2810 }
2811 
2812 static char *snaplist_tag = "snaplist";
2813 /*
2814  * Make a list of dsl_dataset_t's for the snapshots between first_obj
2815  * (exclusive) and last_obj (inclusive).  The list will be in reverse
2816  * order (last_obj will be the list_head()).  If first_obj == 0, do all
2817  * snapshots back to this dataset's origin.
2818  */
2819 static int
2820 snaplist_make(dsl_pool_t *dp, boolean_t own,
2821     uint64_t first_obj, uint64_t last_obj, list_t *l)
2822 {
2823 	uint64_t obj = last_obj;
2824 
2825 	ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock));
2826 
2827 	list_create(l, sizeof (struct promotenode),
2828 	    offsetof(struct promotenode, link));
2829 
2830 	while (obj != first_obj) {
2831 		dsl_dataset_t *ds;
2832 		struct promotenode *snap;
2833 		int err;
2834 
2835 		if (own) {
2836 			err = dsl_dataset_own_obj(dp, obj,
2837 			    0, snaplist_tag, &ds);
2838 			if (err == 0)
2839 				dsl_dataset_make_exclusive(ds, snaplist_tag);
2840 		} else {
2841 			err = dsl_dataset_hold_obj(dp, obj, snaplist_tag, &ds);
2842 		}
2843 		if (err == ENOENT) {
2844 			/* lost race with snapshot destroy */
2845 			struct promotenode *last = list_tail(l);
2846 			ASSERT(obj != last->ds->ds_phys->ds_prev_snap_obj);
2847 			obj = last->ds->ds_phys->ds_prev_snap_obj;
2848 			continue;
2849 		} else if (err) {
2850 			return (err);
2851 		}
2852 
2853 		if (first_obj == 0)
2854 			first_obj = ds->ds_dir->dd_phys->dd_origin_obj;
2855 
2856 		snap = kmem_alloc(sizeof (struct promotenode), KM_SLEEP);
2857 		snap->ds = ds;
2858 		list_insert_tail(l, snap);
2859 		obj = ds->ds_phys->ds_prev_snap_obj;
2860 	}
2861 
2862 	return (0);
2863 }
2864 
2865 static int
2866 snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep)
2867 {
2868 	struct promotenode *snap;
2869 
2870 	*spacep = 0;
2871 	for (snap = list_head(l); snap; snap = list_next(l, snap)) {
2872 		uint64_t used, comp, uncomp;
2873 		dsl_deadlist_space_range(&snap->ds->ds_deadlist,
2874 		    mintxg, UINT64_MAX, &used, &comp, &uncomp);
2875 		*spacep += used;
2876 	}
2877 	return (0);
2878 }
2879 
2880 static void
2881 snaplist_destroy(list_t *l, boolean_t own)
2882 {
2883 	struct promotenode *snap;
2884 
2885 	if (!l || !list_link_active(&l->list_head))
2886 		return;
2887 
2888 	while ((snap = list_tail(l)) != NULL) {
2889 		list_remove(l, snap);
2890 		if (own)
2891 			dsl_dataset_disown(snap->ds, snaplist_tag);
2892 		else
2893 			dsl_dataset_rele(snap->ds, snaplist_tag);
2894 		kmem_free(snap, sizeof (struct promotenode));
2895 	}
2896 	list_destroy(l);
2897 }
2898 
2899 /*
2900  * Promote a clone.  Nomenclature note:
2901  * "clone" or "cds": the original clone which is being promoted
2902  * "origin" or "ods": the snapshot which is originally clone's origin
2903  * "origin head" or "ohds": the dataset which is the head
2904  * (filesystem/volume) for the origin
2905  * "origin origin": the origin of the origin's filesystem (typically
2906  * NULL, indicating that the clone is not a clone of a clone).
2907  */
2908 int
2909 dsl_dataset_promote(const char *name, char *conflsnap)
2910 {
2911 	dsl_dataset_t *ds;
2912 	dsl_dir_t *dd;
2913 	dsl_pool_t *dp;
2914 	dmu_object_info_t doi;
2915 	struct promotearg pa = { 0 };
2916 	struct promotenode *snap;
2917 	int err;
2918 
2919 	err = dsl_dataset_hold(name, FTAG, &ds);
2920 	if (err)
2921 		return (err);
2922 	dd = ds->ds_dir;
2923 	dp = dd->dd_pool;
2924 
2925 	err = dmu_object_info(dp->dp_meta_objset,
2926 	    ds->ds_phys->ds_snapnames_zapobj, &doi);
2927 	if (err) {
2928 		dsl_dataset_rele(ds, FTAG);
2929 		return (err);
2930 	}
2931 
2932 	if (dsl_dataset_is_snapshot(ds) || dd->dd_phys->dd_origin_obj == 0) {
2933 		dsl_dataset_rele(ds, FTAG);
2934 		return (EINVAL);
2935 	}
2936 
2937 	/*
2938 	 * We are going to inherit all the snapshots taken before our
2939 	 * origin (i.e., our new origin will be our parent's origin).
2940 	 * Take ownership of them so that we can rename them into our
2941 	 * namespace.
2942 	 */
2943 	rw_enter(&dp->dp_config_rwlock, RW_READER);
2944 
2945 	err = snaplist_make(dp, B_TRUE, 0, dd->dd_phys->dd_origin_obj,
2946 	    &pa.shared_snaps);
2947 	if (err != 0)
2948 		goto out;
2949 
2950 	err = snaplist_make(dp, B_FALSE, 0, ds->ds_object, &pa.clone_snaps);
2951 	if (err != 0)
2952 		goto out;
2953 
2954 	snap = list_head(&pa.shared_snaps);
2955 	ASSERT3U(snap->ds->ds_object, ==, dd->dd_phys->dd_origin_obj);
2956 	err = snaplist_make(dp, B_FALSE, dd->dd_phys->dd_origin_obj,
2957 	    snap->ds->ds_dir->dd_phys->dd_head_dataset_obj, &pa.origin_snaps);
2958 	if (err != 0)
2959 		goto out;
2960 
2961 	if (snap->ds->ds_dir->dd_phys->dd_origin_obj != 0) {
2962 		err = dsl_dataset_hold_obj(dp,
2963 		    snap->ds->ds_dir->dd_phys->dd_origin_obj,
2964 		    FTAG, &pa.origin_origin);
2965 		if (err != 0)
2966 			goto out;
2967 	}
2968 
2969 out:
2970 	rw_exit(&dp->dp_config_rwlock);
2971 
2972 	/*
2973 	 * Add in 128x the snapnames zapobj size, since we will be moving
2974 	 * a bunch of snapnames to the promoted ds, and dirtying their
2975 	 * bonus buffers.
2976 	 */
2977 	if (err == 0) {
2978 		err = dsl_sync_task_do(dp, dsl_dataset_promote_check,
2979 		    dsl_dataset_promote_sync, ds, &pa,
2980 		    2 + 2 * doi.doi_physical_blocks_512);
2981 		if (err && pa.err_ds && conflsnap)
2982 			(void) strncpy(conflsnap, pa.err_ds, MAXNAMELEN);
2983 	}
2984 
2985 	snaplist_destroy(&pa.shared_snaps, B_TRUE);
2986 	snaplist_destroy(&pa.clone_snaps, B_FALSE);
2987 	snaplist_destroy(&pa.origin_snaps, B_FALSE);
2988 	if (pa.origin_origin)
2989 		dsl_dataset_rele(pa.origin_origin, FTAG);
2990 	dsl_dataset_rele(ds, FTAG);
2991 	return (err);
2992 }
2993 
2994 struct cloneswaparg {
2995 	dsl_dataset_t *cds; /* clone dataset */
2996 	dsl_dataset_t *ohds; /* origin's head dataset */
2997 	boolean_t force;
2998 	int64_t unused_refres_delta; /* change in unconsumed refreservation */
2999 };
3000 
3001 /* ARGSUSED */
3002 static int
3003 dsl_dataset_clone_swap_check(void *arg1, void *arg2, dmu_tx_t *tx)
3004 {
3005 	struct cloneswaparg *csa = arg1;
3006 
3007 	/* they should both be heads */
3008 	if (dsl_dataset_is_snapshot(csa->cds) ||
3009 	    dsl_dataset_is_snapshot(csa->ohds))
3010 		return (EINVAL);
3011 
3012 	/* the branch point should be just before them */
3013 	if (csa->cds->ds_prev != csa->ohds->ds_prev)
3014 		return (EINVAL);
3015 
3016 	/* cds should be the clone (unless they are unrelated) */
3017 	if (csa->cds->ds_prev != NULL &&
3018 	    csa->cds->ds_prev != csa->cds->ds_dir->dd_pool->dp_origin_snap &&
3019 	    csa->ohds->ds_object !=
3020 	    csa->cds->ds_prev->ds_phys->ds_next_snap_obj)
3021 		return (EINVAL);
3022 
3023 	/* the clone should be a child of the origin */
3024 	if (csa->cds->ds_dir->dd_parent != csa->ohds->ds_dir)
3025 		return (EINVAL);
3026 
3027 	/* ohds shouldn't be modified unless 'force' */
3028 	if (!csa->force && dsl_dataset_modified_since_lastsnap(csa->ohds))
3029 		return (ETXTBSY);
3030 
3031 	/* adjust amount of any unconsumed refreservation */
3032 	csa->unused_refres_delta =
3033 	    (int64_t)MIN(csa->ohds->ds_reserved,
3034 	    csa->ohds->ds_phys->ds_unique_bytes) -
3035 	    (int64_t)MIN(csa->ohds->ds_reserved,
3036 	    csa->cds->ds_phys->ds_unique_bytes);
3037 
3038 	if (csa->unused_refres_delta > 0 &&
3039 	    csa->unused_refres_delta >
3040 	    dsl_dir_space_available(csa->ohds->ds_dir, NULL, 0, TRUE))
3041 		return (ENOSPC);
3042 
3043 	if (csa->ohds->ds_quota != 0 &&
3044 	    csa->cds->ds_phys->ds_unique_bytes > csa->ohds->ds_quota)
3045 		return (EDQUOT);
3046 
3047 	return (0);
3048 }
3049 
3050 /* ARGSUSED */
3051 static void
3052 dsl_dataset_clone_swap_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3053 {
3054 	struct cloneswaparg *csa = arg1;
3055 	dsl_pool_t *dp = csa->cds->ds_dir->dd_pool;
3056 
3057 	ASSERT(csa->cds->ds_reserved == 0);
3058 	ASSERT(csa->ohds->ds_quota == 0 ||
3059 	    csa->cds->ds_phys->ds_unique_bytes <= csa->ohds->ds_quota);
3060 
3061 	dmu_buf_will_dirty(csa->cds->ds_dbuf, tx);
3062 	dmu_buf_will_dirty(csa->ohds->ds_dbuf, tx);
3063 
3064 	if (csa->cds->ds_objset != NULL) {
3065 		dmu_objset_evict(csa->cds->ds_objset);
3066 		csa->cds->ds_objset = NULL;
3067 	}
3068 
3069 	if (csa->ohds->ds_objset != NULL) {
3070 		dmu_objset_evict(csa->ohds->ds_objset);
3071 		csa->ohds->ds_objset = NULL;
3072 	}
3073 
3074 	/*
3075 	 * Reset origin's unique bytes, if it exists.
3076 	 */
3077 	if (csa->cds->ds_prev) {
3078 		dsl_dataset_t *origin = csa->cds->ds_prev;
3079 		uint64_t comp, uncomp;
3080 
3081 		dmu_buf_will_dirty(origin->ds_dbuf, tx);
3082 		dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3083 		    origin->ds_phys->ds_prev_snap_txg, UINT64_MAX,
3084 		    &origin->ds_phys->ds_unique_bytes, &comp, &uncomp);
3085 	}
3086 
3087 	/* swap blkptrs */
3088 	{
3089 		blkptr_t tmp;
3090 		tmp = csa->ohds->ds_phys->ds_bp;
3091 		csa->ohds->ds_phys->ds_bp = csa->cds->ds_phys->ds_bp;
3092 		csa->cds->ds_phys->ds_bp = tmp;
3093 	}
3094 
3095 	/* set dd_*_bytes */
3096 	{
3097 		int64_t dused, dcomp, duncomp;
3098 		uint64_t cdl_used, cdl_comp, cdl_uncomp;
3099 		uint64_t odl_used, odl_comp, odl_uncomp;
3100 
3101 		ASSERT3U(csa->cds->ds_dir->dd_phys->
3102 		    dd_used_breakdown[DD_USED_SNAP], ==, 0);
3103 
3104 		dsl_deadlist_space(&csa->cds->ds_deadlist,
3105 		    &cdl_used, &cdl_comp, &cdl_uncomp);
3106 		dsl_deadlist_space(&csa->ohds->ds_deadlist,
3107 		    &odl_used, &odl_comp, &odl_uncomp);
3108 
3109 		dused = csa->cds->ds_phys->ds_used_bytes + cdl_used -
3110 		    (csa->ohds->ds_phys->ds_used_bytes + odl_used);
3111 		dcomp = csa->cds->ds_phys->ds_compressed_bytes + cdl_comp -
3112 		    (csa->ohds->ds_phys->ds_compressed_bytes + odl_comp);
3113 		duncomp = csa->cds->ds_phys->ds_uncompressed_bytes +
3114 		    cdl_uncomp -
3115 		    (csa->ohds->ds_phys->ds_uncompressed_bytes + odl_uncomp);
3116 
3117 		dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_HEAD,
3118 		    dused, dcomp, duncomp, tx);
3119 		dsl_dir_diduse_space(csa->cds->ds_dir, DD_USED_HEAD,
3120 		    -dused, -dcomp, -duncomp, tx);
3121 
3122 		/*
3123 		 * The difference in the space used by snapshots is the
3124 		 * difference in snapshot space due to the head's
3125 		 * deadlist (since that's the only thing that's
3126 		 * changing that affects the snapused).
3127 		 */
3128 		dsl_deadlist_space_range(&csa->cds->ds_deadlist,
3129 		    csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3130 		    &cdl_used, &cdl_comp, &cdl_uncomp);
3131 		dsl_deadlist_space_range(&csa->ohds->ds_deadlist,
3132 		    csa->ohds->ds_dir->dd_origin_txg, UINT64_MAX,
3133 		    &odl_used, &odl_comp, &odl_uncomp);
3134 		dsl_dir_transfer_space(csa->ohds->ds_dir, cdl_used - odl_used,
3135 		    DD_USED_HEAD, DD_USED_SNAP, tx);
3136 	}
3137 
3138 	/* swap ds_*_bytes */
3139 	SWITCH64(csa->ohds->ds_phys->ds_used_bytes,
3140 	    csa->cds->ds_phys->ds_used_bytes);
3141 	SWITCH64(csa->ohds->ds_phys->ds_compressed_bytes,
3142 	    csa->cds->ds_phys->ds_compressed_bytes);
3143 	SWITCH64(csa->ohds->ds_phys->ds_uncompressed_bytes,
3144 	    csa->cds->ds_phys->ds_uncompressed_bytes);
3145 	SWITCH64(csa->ohds->ds_phys->ds_unique_bytes,
3146 	    csa->cds->ds_phys->ds_unique_bytes);
3147 
3148 	/* apply any parent delta for change in unconsumed refreservation */
3149 	dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_REFRSRV,
3150 	    csa->unused_refres_delta, 0, 0, tx);
3151 
3152 	/*
3153 	 * Swap deadlists.
3154 	 */
3155 	dsl_deadlist_close(&csa->cds->ds_deadlist);
3156 	dsl_deadlist_close(&csa->ohds->ds_deadlist);
3157 	SWITCH64(csa->ohds->ds_phys->ds_deadlist_obj,
3158 	    csa->cds->ds_phys->ds_deadlist_obj);
3159 	dsl_deadlist_open(&csa->cds->ds_deadlist, dp->dp_meta_objset,
3160 	    csa->cds->ds_phys->ds_deadlist_obj);
3161 	dsl_deadlist_open(&csa->ohds->ds_deadlist, dp->dp_meta_objset,
3162 	    csa->ohds->ds_phys->ds_deadlist_obj);
3163 
3164 	dsl_scan_ds_clone_swapped(csa->ohds, csa->cds, tx);
3165 }
3166 
3167 /*
3168  * Swap 'clone' with its origin head datasets.  Used at the end of "zfs
3169  * recv" into an existing fs to swizzle the file system to the new
3170  * version, and by "zfs rollback".  Can also be used to swap two
3171  * independent head datasets if neither has any snapshots.
3172  */
3173 int
3174 dsl_dataset_clone_swap(dsl_dataset_t *clone, dsl_dataset_t *origin_head,
3175     boolean_t force)
3176 {
3177 	struct cloneswaparg csa;
3178 	int error;
3179 
3180 	ASSERT(clone->ds_owner);
3181 	ASSERT(origin_head->ds_owner);
3182 retry:
3183 	/*
3184 	 * Need exclusive access for the swap. If we're swapping these
3185 	 * datasets back after an error, we already hold the locks.
3186 	 */
3187 	if (!RW_WRITE_HELD(&clone->ds_rwlock))
3188 		rw_enter(&clone->ds_rwlock, RW_WRITER);
3189 	if (!RW_WRITE_HELD(&origin_head->ds_rwlock) &&
3190 	    !rw_tryenter(&origin_head->ds_rwlock, RW_WRITER)) {
3191 		rw_exit(&clone->ds_rwlock);
3192 		rw_enter(&origin_head->ds_rwlock, RW_WRITER);
3193 		if (!rw_tryenter(&clone->ds_rwlock, RW_WRITER)) {
3194 			rw_exit(&origin_head->ds_rwlock);
3195 			goto retry;
3196 		}
3197 	}
3198 	csa.cds = clone;
3199 	csa.ohds = origin_head;
3200 	csa.force = force;
3201 	error = dsl_sync_task_do(clone->ds_dir->dd_pool,
3202 	    dsl_dataset_clone_swap_check,
3203 	    dsl_dataset_clone_swap_sync, &csa, NULL, 9);
3204 	return (error);
3205 }
3206 
3207 /*
3208  * Given a pool name and a dataset object number in that pool,
3209  * return the name of that dataset.
3210  */
3211 int
3212 dsl_dsobj_to_dsname(char *pname, uint64_t obj, char *buf)
3213 {
3214 	spa_t *spa;
3215 	dsl_pool_t *dp;
3216 	dsl_dataset_t *ds;
3217 	int error;
3218 
3219 	if ((error = spa_open(pname, &spa, FTAG)) != 0)
3220 		return (error);
3221 	dp = spa_get_dsl(spa);
3222 	rw_enter(&dp->dp_config_rwlock, RW_READER);
3223 	if ((error = dsl_dataset_hold_obj(dp, obj, FTAG, &ds)) == 0) {
3224 		dsl_dataset_name(ds, buf);
3225 		dsl_dataset_rele(ds, FTAG);
3226 	}
3227 	rw_exit(&dp->dp_config_rwlock);
3228 	spa_close(spa, FTAG);
3229 
3230 	return (error);
3231 }
3232 
3233 int
3234 dsl_dataset_check_quota(dsl_dataset_t *ds, boolean_t check_quota,
3235     uint64_t asize, uint64_t inflight, uint64_t *used, uint64_t *ref_rsrv)
3236 {
3237 	int error = 0;
3238 
3239 	ASSERT3S(asize, >, 0);
3240 
3241 	/*
3242 	 * *ref_rsrv is the portion of asize that will come from any
3243 	 * unconsumed refreservation space.
3244 	 */
3245 	*ref_rsrv = 0;
3246 
3247 	mutex_enter(&ds->ds_lock);
3248 	/*
3249 	 * Make a space adjustment for reserved bytes.
3250 	 */
3251 	if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes) {
3252 		ASSERT3U(*used, >=,
3253 		    ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3254 		*used -= (ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3255 		*ref_rsrv =
3256 		    asize - MIN(asize, parent_delta(ds, asize + inflight));
3257 	}
3258 
3259 	if (!check_quota || ds->ds_quota == 0) {
3260 		mutex_exit(&ds->ds_lock);
3261 		return (0);
3262 	}
3263 	/*
3264 	 * If they are requesting more space, and our current estimate
3265 	 * is over quota, they get to try again unless the actual
3266 	 * on-disk is over quota and there are no pending changes (which
3267 	 * may free up space for us).
3268 	 */
3269 	if (ds->ds_phys->ds_used_bytes + inflight >= ds->ds_quota) {
3270 		if (inflight > 0 || ds->ds_phys->ds_used_bytes < ds->ds_quota)
3271 			error = ERESTART;
3272 		else
3273 			error = EDQUOT;
3274 	}
3275 	mutex_exit(&ds->ds_lock);
3276 
3277 	return (error);
3278 }
3279 
3280 /* ARGSUSED */
3281 static int
3282 dsl_dataset_set_quota_check(void *arg1, void *arg2, dmu_tx_t *tx)
3283 {
3284 	dsl_dataset_t *ds = arg1;
3285 	dsl_prop_setarg_t *psa = arg2;
3286 	int err;
3287 
3288 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_REFQUOTA)
3289 		return (ENOTSUP);
3290 
3291 	if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3292 		return (err);
3293 
3294 	if (psa->psa_effective_value == 0)
3295 		return (0);
3296 
3297 	if (psa->psa_effective_value < ds->ds_phys->ds_used_bytes ||
3298 	    psa->psa_effective_value < ds->ds_reserved)
3299 		return (ENOSPC);
3300 
3301 	return (0);
3302 }
3303 
3304 extern void dsl_prop_set_sync(void *, void *, dmu_tx_t *);
3305 
3306 void
3307 dsl_dataset_set_quota_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3308 {
3309 	dsl_dataset_t *ds = arg1;
3310 	dsl_prop_setarg_t *psa = arg2;
3311 	uint64_t effective_value = psa->psa_effective_value;
3312 
3313 	dsl_prop_set_sync(ds, psa, tx);
3314 	DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3315 
3316 	if (ds->ds_quota != effective_value) {
3317 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
3318 		ds->ds_quota = effective_value;
3319 
3320 		spa_history_log_internal(LOG_DS_REFQUOTA,
3321 		    ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu ",
3322 		    (longlong_t)ds->ds_quota, ds->ds_object);
3323 	}
3324 }
3325 
3326 int
3327 dsl_dataset_set_quota(const char *dsname, zprop_source_t source, uint64_t quota)
3328 {
3329 	dsl_dataset_t *ds;
3330 	dsl_prop_setarg_t psa;
3331 	int err;
3332 
3333 	dsl_prop_setarg_init_uint64(&psa, "refquota", source, &quota);
3334 
3335 	err = dsl_dataset_hold(dsname, FTAG, &ds);
3336 	if (err)
3337 		return (err);
3338 
3339 	/*
3340 	 * If someone removes a file, then tries to set the quota, we
3341 	 * want to make sure the file freeing takes effect.
3342 	 */
3343 	txg_wait_open(ds->ds_dir->dd_pool, 0);
3344 
3345 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3346 	    dsl_dataset_set_quota_check, dsl_dataset_set_quota_sync,
3347 	    ds, &psa, 0);
3348 
3349 	dsl_dataset_rele(ds, FTAG);
3350 	return (err);
3351 }
3352 
3353 static int
3354 dsl_dataset_set_reservation_check(void *arg1, void *arg2, dmu_tx_t *tx)
3355 {
3356 	dsl_dataset_t *ds = arg1;
3357 	dsl_prop_setarg_t *psa = arg2;
3358 	uint64_t effective_value;
3359 	uint64_t unique;
3360 	int err;
3361 
3362 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
3363 	    SPA_VERSION_REFRESERVATION)
3364 		return (ENOTSUP);
3365 
3366 	if (dsl_dataset_is_snapshot(ds))
3367 		return (EINVAL);
3368 
3369 	if ((err = dsl_prop_predict_sync(ds->ds_dir, psa)) != 0)
3370 		return (err);
3371 
3372 	effective_value = psa->psa_effective_value;
3373 
3374 	/*
3375 	 * If we are doing the preliminary check in open context, the
3376 	 * space estimates may be inaccurate.
3377 	 */
3378 	if (!dmu_tx_is_syncing(tx))
3379 		return (0);
3380 
3381 	mutex_enter(&ds->ds_lock);
3382 	if (!DS_UNIQUE_IS_ACCURATE(ds))
3383 		dsl_dataset_recalc_head_uniq(ds);
3384 	unique = ds->ds_phys->ds_unique_bytes;
3385 	mutex_exit(&ds->ds_lock);
3386 
3387 	if (MAX(unique, effective_value) > MAX(unique, ds->ds_reserved)) {
3388 		uint64_t delta = MAX(unique, effective_value) -
3389 		    MAX(unique, ds->ds_reserved);
3390 
3391 		if (delta > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
3392 			return (ENOSPC);
3393 		if (ds->ds_quota > 0 &&
3394 		    effective_value > ds->ds_quota)
3395 			return (ENOSPC);
3396 	}
3397 
3398 	return (0);
3399 }
3400 
3401 static void
3402 dsl_dataset_set_reservation_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3403 {
3404 	dsl_dataset_t *ds = arg1;
3405 	dsl_prop_setarg_t *psa = arg2;
3406 	uint64_t effective_value = psa->psa_effective_value;
3407 	uint64_t unique;
3408 	int64_t delta;
3409 
3410 	dsl_prop_set_sync(ds, psa, tx);
3411 	DSL_PROP_CHECK_PREDICTION(ds->ds_dir, psa);
3412 
3413 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
3414 
3415 	mutex_enter(&ds->ds_dir->dd_lock);
3416 	mutex_enter(&ds->ds_lock);
3417 	ASSERT(DS_UNIQUE_IS_ACCURATE(ds));
3418 	unique = ds->ds_phys->ds_unique_bytes;
3419 	delta = MAX(0, (int64_t)(effective_value - unique)) -
3420 	    MAX(0, (int64_t)(ds->ds_reserved - unique));
3421 	ds->ds_reserved = effective_value;
3422 	mutex_exit(&ds->ds_lock);
3423 
3424 	dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV, delta, 0, 0, tx);
3425 	mutex_exit(&ds->ds_dir->dd_lock);
3426 
3427 	spa_history_log_internal(LOG_DS_REFRESERV,
3428 	    ds->ds_dir->dd_pool->dp_spa, tx, "%lld dataset = %llu",
3429 	    (longlong_t)effective_value, ds->ds_object);
3430 }
3431 
3432 int
3433 dsl_dataset_set_reservation(const char *dsname, zprop_source_t source,
3434     uint64_t reservation)
3435 {
3436 	dsl_dataset_t *ds;
3437 	dsl_prop_setarg_t psa;
3438 	int err;
3439 
3440 	dsl_prop_setarg_init_uint64(&psa, "refreservation", source,
3441 	    &reservation);
3442 
3443 	err = dsl_dataset_hold(dsname, FTAG, &ds);
3444 	if (err)
3445 		return (err);
3446 
3447 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3448 	    dsl_dataset_set_reservation_check,
3449 	    dsl_dataset_set_reservation_sync, ds, &psa, 0);
3450 
3451 	dsl_dataset_rele(ds, FTAG);
3452 	return (err);
3453 }
3454 
3455 struct dsl_ds_holdarg {
3456 	dsl_sync_task_group_t *dstg;
3457 	char *htag;
3458 	char *snapname;
3459 	boolean_t recursive;
3460 	boolean_t gotone;
3461 	boolean_t temphold;
3462 	char failed[MAXPATHLEN];
3463 };
3464 
3465 typedef struct zfs_hold_cleanup_arg {
3466 	dsl_pool_t *dp;
3467 	uint64_t dsobj;
3468 	char htag[MAXNAMELEN];
3469 } zfs_hold_cleanup_arg_t;
3470 
3471 static void
3472 dsl_dataset_user_release_onexit(void *arg)
3473 {
3474 	zfs_hold_cleanup_arg_t *ca = arg;
3475 
3476 	(void) dsl_dataset_user_release_tmp(ca->dp, ca->dsobj, ca->htag,
3477 	    B_TRUE);
3478 	kmem_free(ca, sizeof (zfs_hold_cleanup_arg_t));
3479 }
3480 
3481 void
3482 dsl_register_onexit_hold_cleanup(dsl_dataset_t *ds, const char *htag,
3483     minor_t minor)
3484 {
3485 	zfs_hold_cleanup_arg_t *ca;
3486 
3487 	ca = kmem_alloc(sizeof (zfs_hold_cleanup_arg_t), KM_SLEEP);
3488 	ca->dp = ds->ds_dir->dd_pool;
3489 	ca->dsobj = ds->ds_object;
3490 	(void) strlcpy(ca->htag, htag, sizeof (ca->htag));
3491 	VERIFY3U(0, ==, zfs_onexit_add_cb(minor,
3492 	    dsl_dataset_user_release_onexit, ca, NULL));
3493 }
3494 
3495 /*
3496  * The max length of a temporary tag prefix is the number of hex digits
3497  * required to express UINT64_MAX plus one for the hyphen.
3498  */
3499 #define	MAX_TAG_PREFIX_LEN	17
3500 
3501 static int
3502 dsl_dataset_user_hold_check(void *arg1, void *arg2, dmu_tx_t *tx)
3503 {
3504 	dsl_dataset_t *ds = arg1;
3505 	struct dsl_ds_holdarg *ha = arg2;
3506 	char *htag = ha->htag;
3507 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3508 	int error = 0;
3509 
3510 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3511 		return (ENOTSUP);
3512 
3513 	if (!dsl_dataset_is_snapshot(ds))
3514 		return (EINVAL);
3515 
3516 	/* tags must be unique */
3517 	mutex_enter(&ds->ds_lock);
3518 	if (ds->ds_phys->ds_userrefs_obj) {
3519 		error = zap_lookup(mos, ds->ds_phys->ds_userrefs_obj, htag,
3520 		    8, 1, tx);
3521 		if (error == 0)
3522 			error = EEXIST;
3523 		else if (error == ENOENT)
3524 			error = 0;
3525 	}
3526 	mutex_exit(&ds->ds_lock);
3527 
3528 	if (error == 0 && ha->temphold &&
3529 	    strlen(htag) + MAX_TAG_PREFIX_LEN >= MAXNAMELEN)
3530 		error = E2BIG;
3531 
3532 	return (error);
3533 }
3534 
3535 static void
3536 dsl_dataset_user_hold_sync(void *arg1, void *arg2, dmu_tx_t *tx)
3537 {
3538 	dsl_dataset_t *ds = arg1;
3539 	struct dsl_ds_holdarg *ha = arg2;
3540 	char *htag = ha->htag;
3541 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
3542 	objset_t *mos = dp->dp_meta_objset;
3543 	uint64_t now = gethrestime_sec();
3544 	uint64_t zapobj;
3545 
3546 	mutex_enter(&ds->ds_lock);
3547 	if (ds->ds_phys->ds_userrefs_obj == 0) {
3548 		/*
3549 		 * This is the first user hold for this dataset.  Create
3550 		 * the userrefs zap object.
3551 		 */
3552 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
3553 		zapobj = ds->ds_phys->ds_userrefs_obj =
3554 		    zap_create(mos, DMU_OT_USERREFS, DMU_OT_NONE, 0, tx);
3555 	} else {
3556 		zapobj = ds->ds_phys->ds_userrefs_obj;
3557 	}
3558 	ds->ds_userrefs++;
3559 	mutex_exit(&ds->ds_lock);
3560 
3561 	VERIFY(0 == zap_add(mos, zapobj, htag, 8, 1, &now, tx));
3562 
3563 	if (ha->temphold) {
3564 		VERIFY(0 == dsl_pool_user_hold(dp, ds->ds_object,
3565 		    htag, &now, tx));
3566 	}
3567 
3568 	spa_history_log_internal(LOG_DS_USER_HOLD,
3569 	    dp->dp_spa, tx, "<%s> temp = %d dataset = %llu", htag,
3570 	    (int)ha->temphold, ds->ds_object);
3571 }
3572 
3573 static int
3574 dsl_dataset_user_hold_one(const char *dsname, void *arg)
3575 {
3576 	struct dsl_ds_holdarg *ha = arg;
3577 	dsl_dataset_t *ds;
3578 	int error;
3579 	char *name;
3580 
3581 	/* alloc a buffer to hold dsname@snapname plus terminating NULL */
3582 	name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3583 	error = dsl_dataset_hold(name, ha->dstg, &ds);
3584 	strfree(name);
3585 	if (error == 0) {
3586 		ha->gotone = B_TRUE;
3587 		dsl_sync_task_create(ha->dstg, dsl_dataset_user_hold_check,
3588 		    dsl_dataset_user_hold_sync, ds, ha, 0);
3589 	} else if (error == ENOENT && ha->recursive) {
3590 		error = 0;
3591 	} else {
3592 		(void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3593 	}
3594 	return (error);
3595 }
3596 
3597 int
3598 dsl_dataset_user_hold_for_send(dsl_dataset_t *ds, char *htag,
3599     boolean_t temphold)
3600 {
3601 	struct dsl_ds_holdarg *ha;
3602 	int error;
3603 
3604 	ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3605 	ha->htag = htag;
3606 	ha->temphold = temphold;
3607 	error = dsl_sync_task_do(ds->ds_dir->dd_pool,
3608 	    dsl_dataset_user_hold_check, dsl_dataset_user_hold_sync,
3609 	    ds, ha, 0);
3610 	kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3611 
3612 	return (error);
3613 }
3614 
3615 int
3616 dsl_dataset_user_hold(char *dsname, char *snapname, char *htag,
3617     boolean_t recursive, boolean_t temphold, int cleanup_fd)
3618 {
3619 	struct dsl_ds_holdarg *ha;
3620 	dsl_sync_task_t *dst;
3621 	spa_t *spa;
3622 	int error;
3623 	minor_t minor = 0;
3624 
3625 	if (cleanup_fd != -1) {
3626 		/* Currently we only support cleanup-on-exit of tempholds. */
3627 		if (!temphold)
3628 			return (EINVAL);
3629 		error = zfs_onexit_fd_hold(cleanup_fd, &minor);
3630 		if (error)
3631 			return (error);
3632 	}
3633 
3634 	ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3635 
3636 	(void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3637 
3638 	error = spa_open(dsname, &spa, FTAG);
3639 	if (error) {
3640 		kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3641 		if (cleanup_fd != -1)
3642 			zfs_onexit_fd_rele(cleanup_fd);
3643 		return (error);
3644 	}
3645 
3646 	ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3647 	ha->htag = htag;
3648 	ha->snapname = snapname;
3649 	ha->recursive = recursive;
3650 	ha->temphold = temphold;
3651 
3652 	if (recursive) {
3653 		error = dmu_objset_find(dsname, dsl_dataset_user_hold_one,
3654 		    ha, DS_FIND_CHILDREN);
3655 	} else {
3656 		error = dsl_dataset_user_hold_one(dsname, ha);
3657 	}
3658 	if (error == 0)
3659 		error = dsl_sync_task_group_wait(ha->dstg);
3660 
3661 	for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3662 	    dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3663 		dsl_dataset_t *ds = dst->dst_arg1;
3664 
3665 		if (dst->dst_err) {
3666 			dsl_dataset_name(ds, ha->failed);
3667 			*strchr(ha->failed, '@') = '\0';
3668 		} else if (error == 0 && minor != 0 && temphold) {
3669 			/*
3670 			 * If this hold is to be released upon process exit,
3671 			 * register that action now.
3672 			 */
3673 			dsl_register_onexit_hold_cleanup(ds, htag, minor);
3674 		}
3675 		dsl_dataset_rele(ds, ha->dstg);
3676 	}
3677 
3678 	if (error == 0 && recursive && !ha->gotone)
3679 		error = ENOENT;
3680 
3681 	if (error)
3682 		(void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3683 
3684 	dsl_sync_task_group_destroy(ha->dstg);
3685 
3686 	kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3687 	spa_close(spa, FTAG);
3688 	if (cleanup_fd != -1)
3689 		zfs_onexit_fd_rele(cleanup_fd);
3690 	return (error);
3691 }
3692 
3693 struct dsl_ds_releasearg {
3694 	dsl_dataset_t *ds;
3695 	const char *htag;
3696 	boolean_t own;		/* do we own or just hold ds? */
3697 };
3698 
3699 static int
3700 dsl_dataset_release_might_destroy(dsl_dataset_t *ds, const char *htag,
3701     boolean_t *might_destroy)
3702 {
3703 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3704 	uint64_t zapobj;
3705 	uint64_t tmp;
3706 	int error;
3707 
3708 	*might_destroy = B_FALSE;
3709 
3710 	mutex_enter(&ds->ds_lock);
3711 	zapobj = ds->ds_phys->ds_userrefs_obj;
3712 	if (zapobj == 0) {
3713 		/* The tag can't possibly exist */
3714 		mutex_exit(&ds->ds_lock);
3715 		return (ESRCH);
3716 	}
3717 
3718 	/* Make sure the tag exists */
3719 	error = zap_lookup(mos, zapobj, htag, 8, 1, &tmp);
3720 	if (error) {
3721 		mutex_exit(&ds->ds_lock);
3722 		if (error == ENOENT)
3723 			error = ESRCH;
3724 		return (error);
3725 	}
3726 
3727 	if (ds->ds_userrefs == 1 && ds->ds_phys->ds_num_children == 1 &&
3728 	    DS_IS_DEFER_DESTROY(ds))
3729 		*might_destroy = B_TRUE;
3730 
3731 	mutex_exit(&ds->ds_lock);
3732 	return (0);
3733 }
3734 
3735 static int
3736 dsl_dataset_user_release_check(void *arg1, void *tag, dmu_tx_t *tx)
3737 {
3738 	struct dsl_ds_releasearg *ra = arg1;
3739 	dsl_dataset_t *ds = ra->ds;
3740 	boolean_t might_destroy;
3741 	int error;
3742 
3743 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3744 		return (ENOTSUP);
3745 
3746 	error = dsl_dataset_release_might_destroy(ds, ra->htag, &might_destroy);
3747 	if (error)
3748 		return (error);
3749 
3750 	if (might_destroy) {
3751 		struct dsl_ds_destroyarg dsda = {0};
3752 
3753 		if (dmu_tx_is_syncing(tx)) {
3754 			/*
3755 			 * If we're not prepared to remove the snapshot,
3756 			 * we can't allow the release to happen right now.
3757 			 */
3758 			if (!ra->own)
3759 				return (EBUSY);
3760 		}
3761 		dsda.ds = ds;
3762 		dsda.releasing = B_TRUE;
3763 		return (dsl_dataset_destroy_check(&dsda, tag, tx));
3764 	}
3765 
3766 	return (0);
3767 }
3768 
3769 static void
3770 dsl_dataset_user_release_sync(void *arg1, void *tag, dmu_tx_t *tx)
3771 {
3772 	struct dsl_ds_releasearg *ra = arg1;
3773 	dsl_dataset_t *ds = ra->ds;
3774 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
3775 	objset_t *mos = dp->dp_meta_objset;
3776 	uint64_t zapobj;
3777 	uint64_t dsobj = ds->ds_object;
3778 	uint64_t refs;
3779 	int error;
3780 
3781 	mutex_enter(&ds->ds_lock);
3782 	ds->ds_userrefs--;
3783 	refs = ds->ds_userrefs;
3784 	mutex_exit(&ds->ds_lock);
3785 	error = dsl_pool_user_release(dp, ds->ds_object, ra->htag, tx);
3786 	VERIFY(error == 0 || error == ENOENT);
3787 	zapobj = ds->ds_phys->ds_userrefs_obj;
3788 	VERIFY(0 == zap_remove(mos, zapobj, ra->htag, tx));
3789 	if (ds->ds_userrefs == 0 && ds->ds_phys->ds_num_children == 1 &&
3790 	    DS_IS_DEFER_DESTROY(ds)) {
3791 		struct dsl_ds_destroyarg dsda = {0};
3792 
3793 		ASSERT(ra->own);
3794 		dsda.ds = ds;
3795 		dsda.releasing = B_TRUE;
3796 		/* We already did the destroy_check */
3797 		dsl_dataset_destroy_sync(&dsda, tag, tx);
3798 	}
3799 
3800 	spa_history_log_internal(LOG_DS_USER_RELEASE,
3801 	    dp->dp_spa, tx, "<%s> %lld dataset = %llu",
3802 	    ra->htag, (longlong_t)refs, dsobj);
3803 }
3804 
3805 static int
3806 dsl_dataset_user_release_one(const char *dsname, void *arg)
3807 {
3808 	struct dsl_ds_holdarg *ha = arg;
3809 	struct dsl_ds_releasearg *ra;
3810 	dsl_dataset_t *ds;
3811 	int error;
3812 	void *dtag = ha->dstg;
3813 	char *name;
3814 	boolean_t own = B_FALSE;
3815 	boolean_t might_destroy;
3816 
3817 	/* alloc a buffer to hold dsname@snapname, plus the terminating NULL */
3818 	name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3819 	error = dsl_dataset_hold(name, dtag, &ds);
3820 	strfree(name);
3821 	if (error == ENOENT && ha->recursive)
3822 		return (0);
3823 	(void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3824 	if (error)
3825 		return (error);
3826 
3827 	ha->gotone = B_TRUE;
3828 
3829 	ASSERT(dsl_dataset_is_snapshot(ds));
3830 
3831 	error = dsl_dataset_release_might_destroy(ds, ha->htag, &might_destroy);
3832 	if (error) {
3833 		dsl_dataset_rele(ds, dtag);
3834 		return (error);
3835 	}
3836 
3837 	if (might_destroy) {
3838 #ifdef _KERNEL
3839 		name = kmem_asprintf("%s@%s", dsname, ha->snapname);
3840 		error = zfs_unmount_snap(name, NULL);
3841 		strfree(name);
3842 		if (error) {
3843 			dsl_dataset_rele(ds, dtag);
3844 			return (error);
3845 		}
3846 #endif
3847 		if (!dsl_dataset_tryown(ds, B_TRUE, dtag)) {
3848 			dsl_dataset_rele(ds, dtag);
3849 			return (EBUSY);
3850 		} else {
3851 			own = B_TRUE;
3852 			dsl_dataset_make_exclusive(ds, dtag);
3853 		}
3854 	}
3855 
3856 	ra = kmem_alloc(sizeof (struct dsl_ds_releasearg), KM_SLEEP);
3857 	ra->ds = ds;
3858 	ra->htag = ha->htag;
3859 	ra->own = own;
3860 	dsl_sync_task_create(ha->dstg, dsl_dataset_user_release_check,
3861 	    dsl_dataset_user_release_sync, ra, dtag, 0);
3862 
3863 	return (0);
3864 }
3865 
3866 int
3867 dsl_dataset_user_release(char *dsname, char *snapname, char *htag,
3868     boolean_t recursive)
3869 {
3870 	struct dsl_ds_holdarg *ha;
3871 	dsl_sync_task_t *dst;
3872 	spa_t *spa;
3873 	int error;
3874 
3875 top:
3876 	ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3877 
3878 	(void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3879 
3880 	error = spa_open(dsname, &spa, FTAG);
3881 	if (error) {
3882 		kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3883 		return (error);
3884 	}
3885 
3886 	ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3887 	ha->htag = htag;
3888 	ha->snapname = snapname;
3889 	ha->recursive = recursive;
3890 	if (recursive) {
3891 		error = dmu_objset_find(dsname, dsl_dataset_user_release_one,
3892 		    ha, DS_FIND_CHILDREN);
3893 	} else {
3894 		error = dsl_dataset_user_release_one(dsname, ha);
3895 	}
3896 	if (error == 0)
3897 		error = dsl_sync_task_group_wait(ha->dstg);
3898 
3899 	for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3900 	    dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3901 		struct dsl_ds_releasearg *ra = dst->dst_arg1;
3902 		dsl_dataset_t *ds = ra->ds;
3903 
3904 		if (dst->dst_err)
3905 			dsl_dataset_name(ds, ha->failed);
3906 
3907 		if (ra->own)
3908 			dsl_dataset_disown(ds, ha->dstg);
3909 		else
3910 			dsl_dataset_rele(ds, ha->dstg);
3911 
3912 		kmem_free(ra, sizeof (struct dsl_ds_releasearg));
3913 	}
3914 
3915 	if (error == 0 && recursive && !ha->gotone)
3916 		error = ENOENT;
3917 
3918 	if (error && error != EBUSY)
3919 		(void) strlcpy(dsname, ha->failed, sizeof (ha->failed));
3920 
3921 	dsl_sync_task_group_destroy(ha->dstg);
3922 	kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3923 	spa_close(spa, FTAG);
3924 
3925 	/*
3926 	 * We can get EBUSY if we were racing with deferred destroy and
3927 	 * dsl_dataset_user_release_check() hadn't done the necessary
3928 	 * open context setup.  We can also get EBUSY if we're racing
3929 	 * with destroy and that thread is the ds_owner.  Either way
3930 	 * the busy condition should be transient, and we should retry
3931 	 * the release operation.
3932 	 */
3933 	if (error == EBUSY)
3934 		goto top;
3935 
3936 	return (error);
3937 }
3938 
3939 /*
3940  * Called at spa_load time (with retry == B_FALSE) to release a stale
3941  * temporary user hold. Also called by the onexit code (with retry == B_TRUE).
3942  */
3943 int
3944 dsl_dataset_user_release_tmp(dsl_pool_t *dp, uint64_t dsobj, char *htag,
3945     boolean_t retry)
3946 {
3947 	dsl_dataset_t *ds;
3948 	char *snap;
3949 	char *name;
3950 	int namelen;
3951 	int error;
3952 
3953 	do {
3954 		rw_enter(&dp->dp_config_rwlock, RW_READER);
3955 		error = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds);
3956 		rw_exit(&dp->dp_config_rwlock);
3957 		if (error)
3958 			return (error);
3959 		namelen = dsl_dataset_namelen(ds)+1;
3960 		name = kmem_alloc(namelen, KM_SLEEP);
3961 		dsl_dataset_name(ds, name);
3962 		dsl_dataset_rele(ds, FTAG);
3963 
3964 		snap = strchr(name, '@');
3965 		*snap = '\0';
3966 		++snap;
3967 		error = dsl_dataset_user_release(name, snap, htag, B_FALSE);
3968 		kmem_free(name, namelen);
3969 
3970 		/*
3971 		 * The object can't have been destroyed because we have a hold,
3972 		 * but it might have been renamed, resulting in ENOENT.  Retry
3973 		 * if we've been requested to do so.
3974 		 *
3975 		 * It would be nice if we could use the dsobj all the way
3976 		 * through and avoid ENOENT entirely.  But we might need to
3977 		 * unmount the snapshot, and there's currently no way to lookup
3978 		 * a vfsp using a ZFS object id.
3979 		 */
3980 	} while ((error == ENOENT) && retry);
3981 
3982 	return (error);
3983 }
3984 
3985 int
3986 dsl_dataset_get_holds(const char *dsname, nvlist_t **nvp)
3987 {
3988 	dsl_dataset_t *ds;
3989 	int err;
3990 
3991 	err = dsl_dataset_hold(dsname, FTAG, &ds);
3992 	if (err)
3993 		return (err);
3994 
3995 	VERIFY(0 == nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP));
3996 	if (ds->ds_phys->ds_userrefs_obj != 0) {
3997 		zap_attribute_t *za;
3998 		zap_cursor_t zc;
3999 
4000 		za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
4001 		for (zap_cursor_init(&zc, ds->ds_dir->dd_pool->dp_meta_objset,
4002 		    ds->ds_phys->ds_userrefs_obj);
4003 		    zap_cursor_retrieve(&zc, za) == 0;
4004 		    zap_cursor_advance(&zc)) {
4005 			VERIFY(0 == nvlist_add_uint64(*nvp, za->za_name,
4006 			    za->za_first_integer));
4007 		}
4008 		zap_cursor_fini(&zc);
4009 		kmem_free(za, sizeof (zap_attribute_t));
4010 	}
4011 	dsl_dataset_rele(ds, FTAG);
4012 	return (0);
4013 }
4014 
4015 /*
4016  * Note, this fuction is used as the callback for dmu_objset_find().  We
4017  * always return 0 so that we will continue to find and process
4018  * inconsistent datasets, even if we encounter an error trying to
4019  * process one of them.
4020  */
4021 /* ARGSUSED */
4022 int
4023 dsl_destroy_inconsistent(const char *dsname, void *arg)
4024 {
4025 	dsl_dataset_t *ds;
4026 
4027 	if (dsl_dataset_own(dsname, B_TRUE, FTAG, &ds) == 0) {
4028 		if (DS_IS_INCONSISTENT(ds))
4029 			(void) dsl_dataset_destroy(ds, FTAG, B_FALSE);
4030 		else
4031 			dsl_dataset_disown(ds, FTAG);
4032 	}
4033 	return (0);
4034 }
4035