xref: /titanic_52/usr/src/uts/common/fs/zfs/dsl_dataset.c (revision bd2ee4f4d736b3a98de7cb84206a8cd8d65ccdda)
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 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/dmu_objset.h>
27 #include <sys/dsl_dataset.h>
28 #include <sys/dsl_dir.h>
29 #include <sys/dsl_prop.h>
30 #include <sys/dsl_synctask.h>
31 #include <sys/dmu_traverse.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/arc.h>
34 #include <sys/zio.h>
35 #include <sys/zap.h>
36 #include <sys/unique.h>
37 #include <sys/zfs_context.h>
38 #include <sys/zfs_ioctl.h>
39 #include <sys/spa.h>
40 #include <sys/zfs_znode.h>
41 #include <sys/sunddi.h>
42 #include <sys/zvol.h>
43 
44 static char *dsl_reaper = "the grim reaper";
45 
46 static dsl_checkfunc_t dsl_dataset_destroy_begin_check;
47 static dsl_syncfunc_t dsl_dataset_destroy_begin_sync;
48 static dsl_checkfunc_t dsl_dataset_rollback_check;
49 static dsl_syncfunc_t dsl_dataset_rollback_sync;
50 static dsl_syncfunc_t dsl_dataset_set_reservation_sync;
51 
52 #define	DS_REF_MAX	(1ULL << 62)
53 
54 #define	DSL_DEADLIST_BLOCKSIZE	SPA_MAXBLOCKSIZE
55 
56 #define	DSL_DATASET_IS_DESTROYED(ds)	((ds)->ds_owner == dsl_reaper)
57 
58 
59 /*
60  * Figure out how much of this delta should be propogated to the dsl_dir
61  * layer.  If there's a refreservation, that space has already been
62  * partially accounted for in our ancestors.
63  */
64 static int64_t
65 parent_delta(dsl_dataset_t *ds, int64_t delta)
66 {
67 	uint64_t old_bytes, new_bytes;
68 
69 	if (ds->ds_reserved == 0)
70 		return (delta);
71 
72 	old_bytes = MAX(ds->ds_phys->ds_unique_bytes, ds->ds_reserved);
73 	new_bytes = MAX(ds->ds_phys->ds_unique_bytes + delta, ds->ds_reserved);
74 
75 	ASSERT3U(ABS((int64_t)(new_bytes - old_bytes)), <=, ABS(delta));
76 	return (new_bytes - old_bytes);
77 }
78 
79 void
80 dsl_dataset_block_born(dsl_dataset_t *ds, blkptr_t *bp, dmu_tx_t *tx)
81 {
82 	int used = bp_get_dasize(tx->tx_pool->dp_spa, bp);
83 	int compressed = BP_GET_PSIZE(bp);
84 	int uncompressed = BP_GET_UCSIZE(bp);
85 	int64_t delta;
86 
87 	dprintf_bp(bp, "born, ds=%p\n", ds);
88 
89 	ASSERT(dmu_tx_is_syncing(tx));
90 	/* It could have been compressed away to nothing */
91 	if (BP_IS_HOLE(bp))
92 		return;
93 	ASSERT(BP_GET_TYPE(bp) != DMU_OT_NONE);
94 	ASSERT3U(BP_GET_TYPE(bp), <, DMU_OT_NUMTYPES);
95 	if (ds == NULL) {
96 		/*
97 		 * Account for the meta-objset space in its placeholder
98 		 * dsl_dir.
99 		 */
100 		ASSERT3U(compressed, ==, uncompressed); /* it's all metadata */
101 		dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
102 		    used, compressed, uncompressed, tx);
103 		dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
104 		return;
105 	}
106 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
107 	mutex_enter(&ds->ds_dir->dd_lock);
108 	mutex_enter(&ds->ds_lock);
109 	delta = parent_delta(ds, used);
110 	ds->ds_phys->ds_used_bytes += used;
111 	ds->ds_phys->ds_compressed_bytes += compressed;
112 	ds->ds_phys->ds_uncompressed_bytes += uncompressed;
113 	ds->ds_phys->ds_unique_bytes += used;
114 	mutex_exit(&ds->ds_lock);
115 	dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD, delta,
116 	    compressed, uncompressed, tx);
117 	dsl_dir_transfer_space(ds->ds_dir, used - delta,
118 	    DD_USED_REFRSRV, DD_USED_HEAD, tx);
119 	mutex_exit(&ds->ds_dir->dd_lock);
120 }
121 
122 int
123 dsl_dataset_block_kill(dsl_dataset_t *ds, blkptr_t *bp, zio_t *pio,
124     dmu_tx_t *tx)
125 {
126 	int used = bp_get_dasize(tx->tx_pool->dp_spa, bp);
127 	int compressed = BP_GET_PSIZE(bp);
128 	int uncompressed = BP_GET_UCSIZE(bp);
129 
130 	ASSERT(pio != NULL);
131 	ASSERT(dmu_tx_is_syncing(tx));
132 	/* No block pointer => nothing to free */
133 	if (BP_IS_HOLE(bp))
134 		return (0);
135 
136 	ASSERT(used > 0);
137 	if (ds == NULL) {
138 		int err;
139 		/*
140 		 * Account for the meta-objset space in its placeholder
141 		 * dataset.
142 		 */
143 		err = dsl_free(pio, tx->tx_pool,
144 		    tx->tx_txg, bp, NULL, NULL, ARC_NOWAIT);
145 		ASSERT(err == 0);
146 
147 		dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir, DD_USED_HEAD,
148 		    -used, -compressed, -uncompressed, tx);
149 		dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
150 		return (used);
151 	}
152 	ASSERT3P(tx->tx_pool, ==, ds->ds_dir->dd_pool);
153 
154 	ASSERT(!dsl_dataset_is_snapshot(ds));
155 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
156 
157 	if (bp->blk_birth > ds->ds_phys->ds_prev_snap_txg) {
158 		int err;
159 		int64_t delta;
160 
161 		dprintf_bp(bp, "freeing: %s", "");
162 		err = dsl_free(pio, tx->tx_pool,
163 		    tx->tx_txg, bp, NULL, NULL, ARC_NOWAIT);
164 		ASSERT(err == 0);
165 
166 		mutex_enter(&ds->ds_dir->dd_lock);
167 		mutex_enter(&ds->ds_lock);
168 		ASSERT(ds->ds_phys->ds_unique_bytes >= used ||
169 		    !DS_UNIQUE_IS_ACCURATE(ds));
170 		delta = parent_delta(ds, -used);
171 		ds->ds_phys->ds_unique_bytes -= used;
172 		mutex_exit(&ds->ds_lock);
173 		dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD,
174 		    delta, -compressed, -uncompressed, tx);
175 		dsl_dir_transfer_space(ds->ds_dir, -used - delta,
176 		    DD_USED_REFRSRV, DD_USED_HEAD, tx);
177 		mutex_exit(&ds->ds_dir->dd_lock);
178 	} else {
179 		dprintf_bp(bp, "putting on dead list: %s", "");
180 		VERIFY(0 == bplist_enqueue(&ds->ds_deadlist, bp, tx));
181 		ASSERT3U(ds->ds_prev->ds_object, ==,
182 		    ds->ds_phys->ds_prev_snap_obj);
183 		ASSERT(ds->ds_prev->ds_phys->ds_num_children > 0);
184 		/* if (bp->blk_birth > prev prev snap txg) prev unique += bs */
185 		if (ds->ds_prev->ds_phys->ds_next_snap_obj ==
186 		    ds->ds_object && bp->blk_birth >
187 		    ds->ds_prev->ds_phys->ds_prev_snap_txg) {
188 			dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
189 			mutex_enter(&ds->ds_prev->ds_lock);
190 			ds->ds_prev->ds_phys->ds_unique_bytes += used;
191 			mutex_exit(&ds->ds_prev->ds_lock);
192 		}
193 		if (bp->blk_birth > ds->ds_origin_txg) {
194 			dsl_dir_transfer_space(ds->ds_dir, used,
195 			    DD_USED_HEAD, DD_USED_SNAP, tx);
196 		}
197 	}
198 	mutex_enter(&ds->ds_lock);
199 	ASSERT3U(ds->ds_phys->ds_used_bytes, >=, used);
200 	ds->ds_phys->ds_used_bytes -= used;
201 	ASSERT3U(ds->ds_phys->ds_compressed_bytes, >=, compressed);
202 	ds->ds_phys->ds_compressed_bytes -= compressed;
203 	ASSERT3U(ds->ds_phys->ds_uncompressed_bytes, >=, uncompressed);
204 	ds->ds_phys->ds_uncompressed_bytes -= uncompressed;
205 	mutex_exit(&ds->ds_lock);
206 
207 	return (used);
208 }
209 
210 uint64_t
211 dsl_dataset_prev_snap_txg(dsl_dataset_t *ds)
212 {
213 	uint64_t trysnap = 0;
214 
215 	if (ds == NULL)
216 		return (0);
217 	/*
218 	 * The snapshot creation could fail, but that would cause an
219 	 * incorrect FALSE return, which would only result in an
220 	 * overestimation of the amount of space that an operation would
221 	 * consume, which is OK.
222 	 *
223 	 * There's also a small window where we could miss a pending
224 	 * snapshot, because we could set the sync task in the quiescing
225 	 * phase.  So this should only be used as a guess.
226 	 */
227 	if (ds->ds_trysnap_txg >
228 	    spa_last_synced_txg(ds->ds_dir->dd_pool->dp_spa))
229 		trysnap = ds->ds_trysnap_txg;
230 	return (MAX(ds->ds_phys->ds_prev_snap_txg, trysnap));
231 }
232 
233 boolean_t
234 dsl_dataset_block_freeable(dsl_dataset_t *ds, uint64_t blk_birth)
235 {
236 	return (blk_birth > dsl_dataset_prev_snap_txg(ds));
237 }
238 
239 /* ARGSUSED */
240 static void
241 dsl_dataset_evict(dmu_buf_t *db, void *dsv)
242 {
243 	dsl_dataset_t *ds = dsv;
244 
245 	ASSERT(ds->ds_owner == NULL || DSL_DATASET_IS_DESTROYED(ds));
246 
247 	dprintf_ds(ds, "evicting %s\n", "");
248 
249 	unique_remove(ds->ds_fsid_guid);
250 
251 	if (ds->ds_user_ptr != NULL)
252 		ds->ds_user_evict_func(ds, ds->ds_user_ptr);
253 
254 	if (ds->ds_prev) {
255 		dsl_dataset_drop_ref(ds->ds_prev, ds);
256 		ds->ds_prev = NULL;
257 	}
258 
259 	bplist_close(&ds->ds_deadlist);
260 	if (ds->ds_dir)
261 		dsl_dir_close(ds->ds_dir, ds);
262 
263 	ASSERT(!list_link_active(&ds->ds_synced_link));
264 
265 	mutex_destroy(&ds->ds_lock);
266 	mutex_destroy(&ds->ds_recvlock);
267 	mutex_destroy(&ds->ds_opening_lock);
268 	mutex_destroy(&ds->ds_deadlist.bpl_lock);
269 	rw_destroy(&ds->ds_rwlock);
270 	cv_destroy(&ds->ds_exclusive_cv);
271 
272 	kmem_free(ds, sizeof (dsl_dataset_t));
273 }
274 
275 static int
276 dsl_dataset_get_snapname(dsl_dataset_t *ds)
277 {
278 	dsl_dataset_phys_t *headphys;
279 	int err;
280 	dmu_buf_t *headdbuf;
281 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
282 	objset_t *mos = dp->dp_meta_objset;
283 
284 	if (ds->ds_snapname[0])
285 		return (0);
286 	if (ds->ds_phys->ds_next_snap_obj == 0)
287 		return (0);
288 
289 	err = dmu_bonus_hold(mos, ds->ds_dir->dd_phys->dd_head_dataset_obj,
290 	    FTAG, &headdbuf);
291 	if (err)
292 		return (err);
293 	headphys = headdbuf->db_data;
294 	err = zap_value_search(dp->dp_meta_objset,
295 	    headphys->ds_snapnames_zapobj, ds->ds_object, 0, ds->ds_snapname);
296 	dmu_buf_rele(headdbuf, FTAG);
297 	return (err);
298 }
299 
300 static int
301 dsl_dataset_snap_lookup(dsl_dataset_t *ds, const char *name, uint64_t *value)
302 {
303 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
304 	uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
305 	matchtype_t mt;
306 	int err;
307 
308 	if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
309 		mt = MT_FIRST;
310 	else
311 		mt = MT_EXACT;
312 
313 	err = zap_lookup_norm(mos, snapobj, name, 8, 1,
314 	    value, mt, NULL, 0, NULL);
315 	if (err == ENOTSUP && mt == MT_FIRST)
316 		err = zap_lookup(mos, snapobj, name, 8, 1, value);
317 	return (err);
318 }
319 
320 static int
321 dsl_dataset_snap_remove(dsl_dataset_t *ds, char *name, dmu_tx_t *tx)
322 {
323 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
324 	uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
325 	matchtype_t mt;
326 	int err;
327 
328 	if (ds->ds_phys->ds_flags & DS_FLAG_CI_DATASET)
329 		mt = MT_FIRST;
330 	else
331 		mt = MT_EXACT;
332 
333 	err = zap_remove_norm(mos, snapobj, name, mt, tx);
334 	if (err == ENOTSUP && mt == MT_FIRST)
335 		err = zap_remove(mos, snapobj, name, tx);
336 	return (err);
337 }
338 
339 static int
340 dsl_dataset_get_ref(dsl_pool_t *dp, uint64_t dsobj, void *tag,
341     dsl_dataset_t **dsp)
342 {
343 	objset_t *mos = dp->dp_meta_objset;
344 	dmu_buf_t *dbuf;
345 	dsl_dataset_t *ds;
346 	int err;
347 
348 	ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
349 	    dsl_pool_sync_context(dp));
350 
351 	err = dmu_bonus_hold(mos, dsobj, tag, &dbuf);
352 	if (err)
353 		return (err);
354 	ds = dmu_buf_get_user(dbuf);
355 	if (ds == NULL) {
356 		dsl_dataset_t *winner;
357 
358 		ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_SLEEP);
359 		ds->ds_dbuf = dbuf;
360 		ds->ds_object = dsobj;
361 		ds->ds_phys = dbuf->db_data;
362 
363 		mutex_init(&ds->ds_lock, NULL, MUTEX_DEFAULT, NULL);
364 		mutex_init(&ds->ds_recvlock, NULL, MUTEX_DEFAULT, NULL);
365 		mutex_init(&ds->ds_opening_lock, NULL, MUTEX_DEFAULT, NULL);
366 		mutex_init(&ds->ds_deadlist.bpl_lock, NULL, MUTEX_DEFAULT,
367 		    NULL);
368 		rw_init(&ds->ds_rwlock, 0, 0, 0);
369 		cv_init(&ds->ds_exclusive_cv, NULL, CV_DEFAULT, NULL);
370 
371 		err = bplist_open(&ds->ds_deadlist,
372 		    mos, ds->ds_phys->ds_deadlist_obj);
373 		if (err == 0) {
374 			err = dsl_dir_open_obj(dp,
375 			    ds->ds_phys->ds_dir_obj, NULL, ds, &ds->ds_dir);
376 		}
377 		if (err) {
378 			/*
379 			 * we don't really need to close the blist if we
380 			 * just opened it.
381 			 */
382 			mutex_destroy(&ds->ds_lock);
383 			mutex_destroy(&ds->ds_recvlock);
384 			mutex_destroy(&ds->ds_opening_lock);
385 			mutex_destroy(&ds->ds_deadlist.bpl_lock);
386 			rw_destroy(&ds->ds_rwlock);
387 			cv_destroy(&ds->ds_exclusive_cv);
388 			kmem_free(ds, sizeof (dsl_dataset_t));
389 			dmu_buf_rele(dbuf, tag);
390 			return (err);
391 		}
392 
393 		if (!dsl_dataset_is_snapshot(ds)) {
394 			ds->ds_snapname[0] = '\0';
395 			if (ds->ds_phys->ds_prev_snap_obj) {
396 				err = dsl_dataset_get_ref(dp,
397 				    ds->ds_phys->ds_prev_snap_obj,
398 				    ds, &ds->ds_prev);
399 			}
400 
401 			if (err == 0 && dsl_dir_is_clone(ds->ds_dir)) {
402 				dsl_dataset_t *origin;
403 
404 				err = dsl_dataset_hold_obj(dp,
405 				    ds->ds_dir->dd_phys->dd_origin_obj,
406 				    FTAG, &origin);
407 				if (err == 0) {
408 					ds->ds_origin_txg =
409 					    origin->ds_phys->ds_creation_txg;
410 					dsl_dataset_rele(origin, FTAG);
411 				}
412 			}
413 		} else {
414 			if (zfs_flags & ZFS_DEBUG_SNAPNAMES)
415 				err = dsl_dataset_get_snapname(ds);
416 			if (err == 0 && ds->ds_phys->ds_userrefs_obj != 0) {
417 				err = zap_count(
418 				    ds->ds_dir->dd_pool->dp_meta_objset,
419 				    ds->ds_phys->ds_userrefs_obj,
420 				    &ds->ds_userrefs);
421 			}
422 		}
423 
424 		if (err == 0 && !dsl_dataset_is_snapshot(ds)) {
425 			/*
426 			 * In sync context, we're called with either no lock
427 			 * or with the write lock.  If we're not syncing,
428 			 * we're always called with the read lock held.
429 			 */
430 			boolean_t need_lock =
431 			    !RW_WRITE_HELD(&dp->dp_config_rwlock) &&
432 			    dsl_pool_sync_context(dp);
433 
434 			if (need_lock)
435 				rw_enter(&dp->dp_config_rwlock, RW_READER);
436 
437 			err = dsl_prop_get_ds(ds,
438 			    "refreservation", sizeof (uint64_t), 1,
439 			    &ds->ds_reserved, NULL);
440 			if (err == 0) {
441 				err = dsl_prop_get_ds(ds,
442 				    "refquota", sizeof (uint64_t), 1,
443 				    &ds->ds_quota, NULL);
444 			}
445 
446 			if (need_lock)
447 				rw_exit(&dp->dp_config_rwlock);
448 		} else {
449 			ds->ds_reserved = ds->ds_quota = 0;
450 		}
451 
452 		if (err == 0) {
453 			winner = dmu_buf_set_user_ie(dbuf, ds, &ds->ds_phys,
454 			    dsl_dataset_evict);
455 		}
456 		if (err || winner) {
457 			bplist_close(&ds->ds_deadlist);
458 			if (ds->ds_prev)
459 				dsl_dataset_drop_ref(ds->ds_prev, ds);
460 			dsl_dir_close(ds->ds_dir, ds);
461 			mutex_destroy(&ds->ds_lock);
462 			mutex_destroy(&ds->ds_recvlock);
463 			mutex_destroy(&ds->ds_opening_lock);
464 			mutex_destroy(&ds->ds_deadlist.bpl_lock);
465 			rw_destroy(&ds->ds_rwlock);
466 			cv_destroy(&ds->ds_exclusive_cv);
467 			kmem_free(ds, sizeof (dsl_dataset_t));
468 			if (err) {
469 				dmu_buf_rele(dbuf, tag);
470 				return (err);
471 			}
472 			ds = winner;
473 		} else {
474 			ds->ds_fsid_guid =
475 			    unique_insert(ds->ds_phys->ds_fsid_guid);
476 		}
477 	}
478 	ASSERT3P(ds->ds_dbuf, ==, dbuf);
479 	ASSERT3P(ds->ds_phys, ==, dbuf->db_data);
480 	ASSERT(ds->ds_phys->ds_prev_snap_obj != 0 ||
481 	    spa_version(dp->dp_spa) < SPA_VERSION_ORIGIN ||
482 	    dp->dp_origin_snap == NULL || ds == dp->dp_origin_snap);
483 	mutex_enter(&ds->ds_lock);
484 	if (!dsl_pool_sync_context(dp) && DSL_DATASET_IS_DESTROYED(ds)) {
485 		mutex_exit(&ds->ds_lock);
486 		dmu_buf_rele(ds->ds_dbuf, tag);
487 		return (ENOENT);
488 	}
489 	mutex_exit(&ds->ds_lock);
490 	*dsp = ds;
491 	return (0);
492 }
493 
494 static int
495 dsl_dataset_hold_ref(dsl_dataset_t *ds, void *tag)
496 {
497 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
498 
499 	/*
500 	 * In syncing context we don't want the rwlock lock: there
501 	 * may be an existing writer waiting for sync phase to
502 	 * finish.  We don't need to worry about such writers, since
503 	 * sync phase is single-threaded, so the writer can't be
504 	 * doing anything while we are active.
505 	 */
506 	if (dsl_pool_sync_context(dp)) {
507 		ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
508 		return (0);
509 	}
510 
511 	/*
512 	 * Normal users will hold the ds_rwlock as a READER until they
513 	 * are finished (i.e., call dsl_dataset_rele()).  "Owners" will
514 	 * drop their READER lock after they set the ds_owner field.
515 	 *
516 	 * If the dataset is being destroyed, the destroy thread will
517 	 * obtain a WRITER lock for exclusive access after it's done its
518 	 * open-context work and then change the ds_owner to
519 	 * dsl_reaper once destruction is assured.  So threads
520 	 * may block here temporarily, until the "destructability" of
521 	 * the dataset is determined.
522 	 */
523 	ASSERT(!RW_WRITE_HELD(&dp->dp_config_rwlock));
524 	mutex_enter(&ds->ds_lock);
525 	while (!rw_tryenter(&ds->ds_rwlock, RW_READER)) {
526 		rw_exit(&dp->dp_config_rwlock);
527 		cv_wait(&ds->ds_exclusive_cv, &ds->ds_lock);
528 		if (DSL_DATASET_IS_DESTROYED(ds)) {
529 			mutex_exit(&ds->ds_lock);
530 			dsl_dataset_drop_ref(ds, tag);
531 			rw_enter(&dp->dp_config_rwlock, RW_READER);
532 			return (ENOENT);
533 		}
534 		rw_enter(&dp->dp_config_rwlock, RW_READER);
535 	}
536 	mutex_exit(&ds->ds_lock);
537 	return (0);
538 }
539 
540 int
541 dsl_dataset_hold_obj(dsl_pool_t *dp, uint64_t dsobj, void *tag,
542     dsl_dataset_t **dsp)
543 {
544 	int err = dsl_dataset_get_ref(dp, dsobj, tag, dsp);
545 
546 	if (err)
547 		return (err);
548 	return (dsl_dataset_hold_ref(*dsp, tag));
549 }
550 
551 int
552 dsl_dataset_own_obj(dsl_pool_t *dp, uint64_t dsobj, int flags, void *owner,
553     dsl_dataset_t **dsp)
554 {
555 	int err = dsl_dataset_hold_obj(dp, dsobj, owner, dsp);
556 
557 	ASSERT(DS_MODE_TYPE(flags) != DS_MODE_USER);
558 
559 	if (err)
560 		return (err);
561 	if (!dsl_dataset_tryown(*dsp, DS_MODE_IS_INCONSISTENT(flags), owner)) {
562 		dsl_dataset_rele(*dsp, owner);
563 		*dsp = NULL;
564 		return (EBUSY);
565 	}
566 	return (0);
567 }
568 
569 int
570 dsl_dataset_hold(const char *name, void *tag, dsl_dataset_t **dsp)
571 {
572 	dsl_dir_t *dd;
573 	dsl_pool_t *dp;
574 	const char *snapname;
575 	uint64_t obj;
576 	int err = 0;
577 
578 	err = dsl_dir_open_spa(NULL, name, FTAG, &dd, &snapname);
579 	if (err)
580 		return (err);
581 
582 	dp = dd->dd_pool;
583 	obj = dd->dd_phys->dd_head_dataset_obj;
584 	rw_enter(&dp->dp_config_rwlock, RW_READER);
585 	if (obj)
586 		err = dsl_dataset_get_ref(dp, obj, tag, dsp);
587 	else
588 		err = ENOENT;
589 	if (err)
590 		goto out;
591 
592 	err = dsl_dataset_hold_ref(*dsp, tag);
593 
594 	/* we may be looking for a snapshot */
595 	if (err == 0 && snapname != NULL) {
596 		dsl_dataset_t *ds = NULL;
597 
598 		if (*snapname++ != '@') {
599 			dsl_dataset_rele(*dsp, tag);
600 			err = ENOENT;
601 			goto out;
602 		}
603 
604 		dprintf("looking for snapshot '%s'\n", snapname);
605 		err = dsl_dataset_snap_lookup(*dsp, snapname, &obj);
606 		if (err == 0)
607 			err = dsl_dataset_get_ref(dp, obj, tag, &ds);
608 		dsl_dataset_rele(*dsp, tag);
609 
610 		ASSERT3U((err == 0), ==, (ds != NULL));
611 
612 		if (ds) {
613 			mutex_enter(&ds->ds_lock);
614 			if (ds->ds_snapname[0] == 0)
615 				(void) strlcpy(ds->ds_snapname, snapname,
616 				    sizeof (ds->ds_snapname));
617 			mutex_exit(&ds->ds_lock);
618 			err = dsl_dataset_hold_ref(ds, tag);
619 			*dsp = err ? NULL : ds;
620 		}
621 	}
622 out:
623 	rw_exit(&dp->dp_config_rwlock);
624 	dsl_dir_close(dd, FTAG);
625 	return (err);
626 }
627 
628 int
629 dsl_dataset_own(const char *name, int flags, void *owner, dsl_dataset_t **dsp)
630 {
631 	int err = dsl_dataset_hold(name, owner, dsp);
632 	if (err)
633 		return (err);
634 	if ((*dsp)->ds_phys->ds_num_children > 0 &&
635 	    !DS_MODE_IS_READONLY(flags)) {
636 		dsl_dataset_rele(*dsp, owner);
637 		return (EROFS);
638 	}
639 	if (!dsl_dataset_tryown(*dsp, DS_MODE_IS_INCONSISTENT(flags), owner)) {
640 		dsl_dataset_rele(*dsp, owner);
641 		return (EBUSY);
642 	}
643 	return (0);
644 }
645 
646 void
647 dsl_dataset_name(dsl_dataset_t *ds, char *name)
648 {
649 	if (ds == NULL) {
650 		(void) strcpy(name, "mos");
651 	} else {
652 		dsl_dir_name(ds->ds_dir, name);
653 		VERIFY(0 == dsl_dataset_get_snapname(ds));
654 		if (ds->ds_snapname[0]) {
655 			(void) strcat(name, "@");
656 			/*
657 			 * We use a "recursive" mutex so that we
658 			 * can call dprintf_ds() with ds_lock held.
659 			 */
660 			if (!MUTEX_HELD(&ds->ds_lock)) {
661 				mutex_enter(&ds->ds_lock);
662 				(void) strcat(name, ds->ds_snapname);
663 				mutex_exit(&ds->ds_lock);
664 			} else {
665 				(void) strcat(name, ds->ds_snapname);
666 			}
667 		}
668 	}
669 }
670 
671 static int
672 dsl_dataset_namelen(dsl_dataset_t *ds)
673 {
674 	int result;
675 
676 	if (ds == NULL) {
677 		result = 3;	/* "mos" */
678 	} else {
679 		result = dsl_dir_namelen(ds->ds_dir);
680 		VERIFY(0 == dsl_dataset_get_snapname(ds));
681 		if (ds->ds_snapname[0]) {
682 			++result;	/* adding one for the @-sign */
683 			if (!MUTEX_HELD(&ds->ds_lock)) {
684 				mutex_enter(&ds->ds_lock);
685 				result += strlen(ds->ds_snapname);
686 				mutex_exit(&ds->ds_lock);
687 			} else {
688 				result += strlen(ds->ds_snapname);
689 			}
690 		}
691 	}
692 
693 	return (result);
694 }
695 
696 void
697 dsl_dataset_drop_ref(dsl_dataset_t *ds, void *tag)
698 {
699 	dmu_buf_rele(ds->ds_dbuf, tag);
700 }
701 
702 void
703 dsl_dataset_rele(dsl_dataset_t *ds, void *tag)
704 {
705 	if (!dsl_pool_sync_context(ds->ds_dir->dd_pool)) {
706 		rw_exit(&ds->ds_rwlock);
707 	}
708 	dsl_dataset_drop_ref(ds, tag);
709 }
710 
711 void
712 dsl_dataset_disown(dsl_dataset_t *ds, void *owner)
713 {
714 	ASSERT((ds->ds_owner == owner && ds->ds_dbuf) ||
715 	    (DSL_DATASET_IS_DESTROYED(ds) && ds->ds_dbuf == NULL));
716 
717 	mutex_enter(&ds->ds_lock);
718 	ds->ds_owner = NULL;
719 	if (RW_WRITE_HELD(&ds->ds_rwlock)) {
720 		rw_exit(&ds->ds_rwlock);
721 		cv_broadcast(&ds->ds_exclusive_cv);
722 	}
723 	mutex_exit(&ds->ds_lock);
724 	if (ds->ds_dbuf)
725 		dsl_dataset_drop_ref(ds, owner);
726 	else
727 		dsl_dataset_evict(ds->ds_dbuf, ds);
728 }
729 
730 boolean_t
731 dsl_dataset_tryown(dsl_dataset_t *ds, boolean_t inconsistentok, void *owner)
732 {
733 	boolean_t gotit = FALSE;
734 
735 	mutex_enter(&ds->ds_lock);
736 	if (ds->ds_owner == NULL &&
737 	    (!DS_IS_INCONSISTENT(ds) || inconsistentok)) {
738 		ds->ds_owner = owner;
739 		if (!dsl_pool_sync_context(ds->ds_dir->dd_pool))
740 			rw_exit(&ds->ds_rwlock);
741 		gotit = TRUE;
742 	}
743 	mutex_exit(&ds->ds_lock);
744 	return (gotit);
745 }
746 
747 void
748 dsl_dataset_make_exclusive(dsl_dataset_t *ds, void *owner)
749 {
750 	ASSERT3P(owner, ==, ds->ds_owner);
751 	if (!RW_WRITE_HELD(&ds->ds_rwlock))
752 		rw_enter(&ds->ds_rwlock, RW_WRITER);
753 }
754 
755 uint64_t
756 dsl_dataset_create_sync_dd(dsl_dir_t *dd, dsl_dataset_t *origin,
757     uint64_t flags, dmu_tx_t *tx)
758 {
759 	dsl_pool_t *dp = dd->dd_pool;
760 	dmu_buf_t *dbuf;
761 	dsl_dataset_phys_t *dsphys;
762 	uint64_t dsobj;
763 	objset_t *mos = dp->dp_meta_objset;
764 
765 	if (origin == NULL)
766 		origin = dp->dp_origin_snap;
767 
768 	ASSERT(origin == NULL || origin->ds_dir->dd_pool == dp);
769 	ASSERT(origin == NULL || origin->ds_phys->ds_num_children > 0);
770 	ASSERT(dmu_tx_is_syncing(tx));
771 	ASSERT(dd->dd_phys->dd_head_dataset_obj == 0);
772 
773 	dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
774 	    DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
775 	VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
776 	dmu_buf_will_dirty(dbuf, tx);
777 	dsphys = dbuf->db_data;
778 	bzero(dsphys, sizeof (dsl_dataset_phys_t));
779 	dsphys->ds_dir_obj = dd->dd_object;
780 	dsphys->ds_flags = flags;
781 	dsphys->ds_fsid_guid = unique_create();
782 	(void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
783 	    sizeof (dsphys->ds_guid));
784 	dsphys->ds_snapnames_zapobj =
785 	    zap_create_norm(mos, U8_TEXTPREP_TOUPPER, DMU_OT_DSL_DS_SNAP_MAP,
786 	    DMU_OT_NONE, 0, tx);
787 	dsphys->ds_creation_time = gethrestime_sec();
788 	dsphys->ds_creation_txg = tx->tx_txg == TXG_INITIAL ? 1 : tx->tx_txg;
789 	dsphys->ds_deadlist_obj =
790 	    bplist_create(mos, DSL_DEADLIST_BLOCKSIZE, tx);
791 
792 	if (origin) {
793 		dsphys->ds_prev_snap_obj = origin->ds_object;
794 		dsphys->ds_prev_snap_txg =
795 		    origin->ds_phys->ds_creation_txg;
796 		dsphys->ds_used_bytes =
797 		    origin->ds_phys->ds_used_bytes;
798 		dsphys->ds_compressed_bytes =
799 		    origin->ds_phys->ds_compressed_bytes;
800 		dsphys->ds_uncompressed_bytes =
801 		    origin->ds_phys->ds_uncompressed_bytes;
802 		dsphys->ds_bp = origin->ds_phys->ds_bp;
803 		dsphys->ds_flags |= origin->ds_phys->ds_flags;
804 
805 		dmu_buf_will_dirty(origin->ds_dbuf, tx);
806 		origin->ds_phys->ds_num_children++;
807 
808 		if (spa_version(dp->dp_spa) >= SPA_VERSION_NEXT_CLONES) {
809 			if (origin->ds_phys->ds_next_clones_obj == 0) {
810 				origin->ds_phys->ds_next_clones_obj =
811 				    zap_create(mos,
812 				    DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
813 			}
814 			VERIFY(0 == zap_add_int(mos,
815 			    origin->ds_phys->ds_next_clones_obj,
816 			    dsobj, tx));
817 		}
818 
819 		dmu_buf_will_dirty(dd->dd_dbuf, tx);
820 		dd->dd_phys->dd_origin_obj = origin->ds_object;
821 	}
822 
823 	if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
824 		dsphys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
825 
826 	dmu_buf_rele(dbuf, FTAG);
827 
828 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
829 	dd->dd_phys->dd_head_dataset_obj = dsobj;
830 
831 	return (dsobj);
832 }
833 
834 uint64_t
835 dsl_dataset_create_sync(dsl_dir_t *pdd, const char *lastname,
836     dsl_dataset_t *origin, uint64_t flags, cred_t *cr, dmu_tx_t *tx)
837 {
838 	dsl_pool_t *dp = pdd->dd_pool;
839 	uint64_t dsobj, ddobj;
840 	dsl_dir_t *dd;
841 
842 	ASSERT(lastname[0] != '@');
843 
844 	ddobj = dsl_dir_create_sync(dp, pdd, lastname, tx);
845 	VERIFY(0 == dsl_dir_open_obj(dp, ddobj, lastname, FTAG, &dd));
846 
847 	dsobj = dsl_dataset_create_sync_dd(dd, origin, flags, tx);
848 
849 	dsl_deleg_set_create_perms(dd, tx, cr);
850 
851 	dsl_dir_close(dd, FTAG);
852 
853 	return (dsobj);
854 }
855 
856 struct destroyarg {
857 	dsl_sync_task_group_t *dstg;
858 	char *snapname;
859 	char *failed;
860 	boolean_t defer;
861 };
862 
863 static int
864 dsl_snapshot_destroy_one(char *name, void *arg)
865 {
866 	struct destroyarg *da = arg;
867 	dsl_dataset_t *ds;
868 	int err;
869 	char *dsname;
870 	size_t buflen;
871 
872 	/* alloc a buffer to hold name@snapname, plus the terminating NULL */
873 	buflen = strlen(name) + strlen(da->snapname) + 2;
874 	dsname = kmem_alloc(buflen, KM_SLEEP);
875 	(void) snprintf(dsname, buflen, "%s@%s", name, da->snapname);
876 	err = dsl_dataset_own(dsname, DS_MODE_READONLY | DS_MODE_INCONSISTENT,
877 	    da->dstg, &ds);
878 	kmem_free(dsname, buflen);
879 	if (err == 0) {
880 		struct dsl_ds_destroyarg *dsda;
881 
882 		dsl_dataset_make_exclusive(ds, da->dstg);
883 		if (ds->ds_user_ptr) {
884 			ds->ds_user_evict_func(ds, ds->ds_user_ptr);
885 			ds->ds_user_ptr = NULL;
886 		}
887 		dsda = kmem_zalloc(sizeof (struct dsl_ds_destroyarg), KM_SLEEP);
888 		dsda->ds = ds;
889 		dsda->defer = da->defer;
890 		dsl_sync_task_create(da->dstg, dsl_dataset_destroy_check,
891 		    dsl_dataset_destroy_sync, dsda, da->dstg, 0);
892 	} else if (err == ENOENT) {
893 		err = 0;
894 	} else {
895 		(void) strcpy(da->failed, name);
896 	}
897 	return (err);
898 }
899 
900 /*
901  * Destroy 'snapname' in all descendants of 'fsname'.
902  */
903 #pragma weak dmu_snapshots_destroy = dsl_snapshots_destroy
904 int
905 dsl_snapshots_destroy(char *fsname, char *snapname, boolean_t defer)
906 {
907 	int err;
908 	struct destroyarg da;
909 	dsl_sync_task_t *dst;
910 	spa_t *spa;
911 
912 	err = spa_open(fsname, &spa, FTAG);
913 	if (err)
914 		return (err);
915 	da.dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
916 	da.snapname = snapname;
917 	da.failed = fsname;
918 	da.defer = defer;
919 
920 	err = dmu_objset_find(fsname,
921 	    dsl_snapshot_destroy_one, &da, DS_FIND_CHILDREN);
922 
923 	if (err == 0)
924 		err = dsl_sync_task_group_wait(da.dstg);
925 
926 	for (dst = list_head(&da.dstg->dstg_tasks); dst;
927 	    dst = list_next(&da.dstg->dstg_tasks, dst)) {
928 		struct dsl_ds_destroyarg *dsda = dst->dst_arg1;
929 		dsl_dataset_t *ds = dsda->ds;
930 
931 		/*
932 		 * Return the file system name that triggered the error
933 		 */
934 		if (dst->dst_err) {
935 			dsl_dataset_name(ds, fsname);
936 			*strchr(fsname, '@') = '\0';
937 		}
938 		ASSERT3P(dsda->rm_origin, ==, NULL);
939 		dsl_dataset_disown(ds, da.dstg);
940 		kmem_free(dsda, sizeof (struct dsl_ds_destroyarg));
941 	}
942 
943 	dsl_sync_task_group_destroy(da.dstg);
944 	spa_close(spa, FTAG);
945 	return (err);
946 }
947 
948 static boolean_t
949 dsl_dataset_might_destroy_origin(dsl_dataset_t *ds)
950 {
951 	boolean_t might_destroy = B_FALSE;
952 
953 	mutex_enter(&ds->ds_lock);
954 	if (ds->ds_phys->ds_num_children == 2 && ds->ds_userrefs == 0 &&
955 	    DS_IS_DEFER_DESTROY(ds))
956 		might_destroy = B_TRUE;
957 	mutex_exit(&ds->ds_lock);
958 
959 	return (might_destroy);
960 }
961 
962 #ifdef _KERNEL
963 static int
964 dsl_dataset_zvol_cleanup(dsl_dataset_t *ds, const char *name)
965 {
966 	int error;
967 	objset_t *os;
968 
969 	error = dmu_objset_open_ds(ds, DMU_OST_ANY, &os);
970 	if (error)
971 		return (error);
972 
973 	if (dmu_objset_type(os) == DMU_OST_ZVOL)
974 		error = zvol_remove_minor(name);
975 	dmu_objset_close(os);
976 
977 	return (error);
978 }
979 #endif
980 
981 /*
982  * If we're removing a clone, and these three conditions are true:
983  *	1) the clone's origin has no other children
984  *	2) the clone's origin has no user references
985  *	3) the clone's origin has been marked for deferred destruction
986  * Then, prepare to remove the origin as part of this sync task group.
987  */
988 static int
989 dsl_dataset_origin_rm_prep(struct dsl_ds_destroyarg *dsda, void *tag)
990 {
991 	dsl_dataset_t *ds = dsda->ds;
992 	dsl_dataset_t *origin = ds->ds_prev;
993 
994 	if (dsl_dataset_might_destroy_origin(origin)) {
995 		char *name;
996 		int namelen;
997 		int error;
998 
999 		namelen = dsl_dataset_namelen(origin) + 1;
1000 		name = kmem_alloc(namelen, KM_SLEEP);
1001 		dsl_dataset_name(origin, name);
1002 #ifdef _KERNEL
1003 		error = zfs_unmount_snap(name, NULL);
1004 		if (error) {
1005 			kmem_free(name, namelen);
1006 			return (error);
1007 		}
1008 		error = dsl_dataset_zvol_cleanup(origin, name);
1009 		if (error) {
1010 			kmem_free(name, namelen);
1011 			return (error);
1012 		}
1013 #endif
1014 		error = dsl_dataset_own(name,
1015 		    DS_MODE_READONLY | DS_MODE_INCONSISTENT,
1016 		    tag, &origin);
1017 		kmem_free(name, namelen);
1018 		if (error)
1019 			return (error);
1020 		dsda->rm_origin = origin;
1021 		dsl_dataset_make_exclusive(origin, tag);
1022 	}
1023 
1024 	return (0);
1025 }
1026 
1027 /*
1028  * ds must be opened as OWNER.  On return (whether successful or not),
1029  * ds will be closed and caller can no longer dereference it.
1030  */
1031 int
1032 dsl_dataset_destroy(dsl_dataset_t *ds, void *tag, boolean_t defer)
1033 {
1034 	int err;
1035 	dsl_sync_task_group_t *dstg;
1036 	objset_t *os;
1037 	dsl_dir_t *dd;
1038 	uint64_t obj;
1039 	struct dsl_ds_destroyarg dsda = {0};
1040 
1041 	dsda.ds = ds;
1042 
1043 	if (dsl_dataset_is_snapshot(ds)) {
1044 		/* Destroying a snapshot is simpler */
1045 		dsl_dataset_make_exclusive(ds, tag);
1046 
1047 		if (ds->ds_user_ptr) {
1048 			ds->ds_user_evict_func(ds, ds->ds_user_ptr);
1049 			ds->ds_user_ptr = NULL;
1050 		}
1051 		/* NOTE: defer is always B_FALSE for non-snapshots */
1052 		dsda.defer = defer;
1053 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1054 		    dsl_dataset_destroy_check, dsl_dataset_destroy_sync,
1055 		    &dsda, tag, 0);
1056 		ASSERT3P(dsda.rm_origin, ==, NULL);
1057 		goto out;
1058 	}
1059 
1060 	dd = ds->ds_dir;
1061 
1062 	/*
1063 	 * Check for errors and mark this ds as inconsistent, in
1064 	 * case we crash while freeing the objects.
1065 	 */
1066 	err = dsl_sync_task_do(dd->dd_pool, dsl_dataset_destroy_begin_check,
1067 	    dsl_dataset_destroy_begin_sync, ds, NULL, 0);
1068 	if (err)
1069 		goto out;
1070 
1071 	err = dmu_objset_open_ds(ds, DMU_OST_ANY, &os);
1072 	if (err)
1073 		goto out;
1074 
1075 	/*
1076 	 * remove the objects in open context, so that we won't
1077 	 * have too much to do in syncing context.
1078 	 */
1079 	for (obj = 0; err == 0; err = dmu_object_next(os, &obj, FALSE,
1080 	    ds->ds_phys->ds_prev_snap_txg)) {
1081 		/*
1082 		 * Ignore errors, if there is not enough disk space
1083 		 * we will deal with it in dsl_dataset_destroy_sync().
1084 		 */
1085 		(void) dmu_free_object(os, obj);
1086 	}
1087 
1088 	/*
1089 	 * We need to sync out all in-flight IO before we try to evict
1090 	 * (the dataset evict func is trying to clear the cached entries
1091 	 * for this dataset in the ARC).
1092 	 */
1093 	txg_wait_synced(dd->dd_pool, 0);
1094 
1095 	/*
1096 	 * If we managed to free all the objects in open
1097 	 * context, the user space accounting should be zero.
1098 	 */
1099 	if (ds->ds_phys->ds_bp.blk_fill == 0 &&
1100 	    dmu_objset_userused_enabled(os->os)) {
1101 		uint64_t count;
1102 
1103 		ASSERT(zap_count(os, DMU_USERUSED_OBJECT, &count) != 0 ||
1104 		    count == 0);
1105 		ASSERT(zap_count(os, DMU_GROUPUSED_OBJECT, &count) != 0 ||
1106 		    count == 0);
1107 	}
1108 
1109 	dmu_objset_close(os);
1110 	if (err != ESRCH)
1111 		goto out;
1112 
1113 	rw_enter(&dd->dd_pool->dp_config_rwlock, RW_READER);
1114 	err = dsl_dir_open_obj(dd->dd_pool, dd->dd_object, NULL, FTAG, &dd);
1115 	rw_exit(&dd->dd_pool->dp_config_rwlock);
1116 
1117 	if (err)
1118 		goto out;
1119 
1120 	if (ds->ds_user_ptr) {
1121 		/*
1122 		 * We need to sync out all in-flight IO before we try
1123 		 * to evict (the dataset evict func is trying to clear
1124 		 * the cached entries for this dataset in the ARC).
1125 		 */
1126 		txg_wait_synced(dd->dd_pool, 0);
1127 	}
1128 
1129 	/*
1130 	 * Blow away the dsl_dir + head dataset.
1131 	 */
1132 	dsl_dataset_make_exclusive(ds, tag);
1133 	if (ds->ds_user_ptr) {
1134 		ds->ds_user_evict_func(ds, ds->ds_user_ptr);
1135 		ds->ds_user_ptr = NULL;
1136 	}
1137 
1138 	/*
1139 	 * If we're removing a clone, we might also need to remove its
1140 	 * origin.
1141 	 */
1142 	do {
1143 		dsda.need_prep = B_FALSE;
1144 		if (dsl_dir_is_clone(dd)) {
1145 			err = dsl_dataset_origin_rm_prep(&dsda, tag);
1146 			if (err) {
1147 				dsl_dir_close(dd, FTAG);
1148 				goto out;
1149 			}
1150 		}
1151 
1152 		dstg = dsl_sync_task_group_create(ds->ds_dir->dd_pool);
1153 		dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
1154 		    dsl_dataset_destroy_sync, &dsda, tag, 0);
1155 		dsl_sync_task_create(dstg, dsl_dir_destroy_check,
1156 		    dsl_dir_destroy_sync, dd, FTAG, 0);
1157 		err = dsl_sync_task_group_wait(dstg);
1158 		dsl_sync_task_group_destroy(dstg);
1159 
1160 		/*
1161 		 * We could be racing against 'zfs release' or 'zfs destroy -d'
1162 		 * on the origin snap, in which case we can get EBUSY if we
1163 		 * needed to destroy the origin snap but were not ready to
1164 		 * do so.
1165 		 */
1166 		if (dsda.need_prep) {
1167 			ASSERT(err == EBUSY);
1168 			ASSERT(dsl_dir_is_clone(dd));
1169 			ASSERT(dsda.rm_origin == NULL);
1170 		}
1171 	} while (dsda.need_prep);
1172 
1173 	if (dsda.rm_origin != NULL)
1174 		dsl_dataset_disown(dsda.rm_origin, tag);
1175 
1176 	/* if it is successful, dsl_dir_destroy_sync will close the dd */
1177 	if (err)
1178 		dsl_dir_close(dd, FTAG);
1179 out:
1180 	dsl_dataset_disown(ds, tag);
1181 	return (err);
1182 }
1183 
1184 int
1185 dsl_dataset_rollback(dsl_dataset_t *ds, dmu_objset_type_t ost)
1186 {
1187 	int err;
1188 
1189 	ASSERT(ds->ds_owner);
1190 
1191 	dsl_dataset_make_exclusive(ds, ds->ds_owner);
1192 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1193 	    dsl_dataset_rollback_check, dsl_dataset_rollback_sync,
1194 	    ds, &ost, 0);
1195 	/* drop exclusive access */
1196 	mutex_enter(&ds->ds_lock);
1197 	rw_exit(&ds->ds_rwlock);
1198 	cv_broadcast(&ds->ds_exclusive_cv);
1199 	mutex_exit(&ds->ds_lock);
1200 	return (err);
1201 }
1202 
1203 void *
1204 dsl_dataset_set_user_ptr(dsl_dataset_t *ds,
1205     void *p, dsl_dataset_evict_func_t func)
1206 {
1207 	void *old;
1208 
1209 	mutex_enter(&ds->ds_lock);
1210 	old = ds->ds_user_ptr;
1211 	if (old == NULL) {
1212 		ds->ds_user_ptr = p;
1213 		ds->ds_user_evict_func = func;
1214 	}
1215 	mutex_exit(&ds->ds_lock);
1216 	return (old);
1217 }
1218 
1219 void *
1220 dsl_dataset_get_user_ptr(dsl_dataset_t *ds)
1221 {
1222 	return (ds->ds_user_ptr);
1223 }
1224 
1225 blkptr_t *
1226 dsl_dataset_get_blkptr(dsl_dataset_t *ds)
1227 {
1228 	return (&ds->ds_phys->ds_bp);
1229 }
1230 
1231 void
1232 dsl_dataset_set_blkptr(dsl_dataset_t *ds, blkptr_t *bp, dmu_tx_t *tx)
1233 {
1234 	ASSERT(dmu_tx_is_syncing(tx));
1235 	/* If it's the meta-objset, set dp_meta_rootbp */
1236 	if (ds == NULL) {
1237 		tx->tx_pool->dp_meta_rootbp = *bp;
1238 	} else {
1239 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
1240 		ds->ds_phys->ds_bp = *bp;
1241 	}
1242 }
1243 
1244 spa_t *
1245 dsl_dataset_get_spa(dsl_dataset_t *ds)
1246 {
1247 	return (ds->ds_dir->dd_pool->dp_spa);
1248 }
1249 
1250 void
1251 dsl_dataset_dirty(dsl_dataset_t *ds, dmu_tx_t *tx)
1252 {
1253 	dsl_pool_t *dp;
1254 
1255 	if (ds == NULL) /* this is the meta-objset */
1256 		return;
1257 
1258 	ASSERT(ds->ds_user_ptr != NULL);
1259 
1260 	if (ds->ds_phys->ds_next_snap_obj != 0)
1261 		panic("dirtying snapshot!");
1262 
1263 	dp = ds->ds_dir->dd_pool;
1264 
1265 	if (txg_list_add(&dp->dp_dirty_datasets, ds, tx->tx_txg) == 0) {
1266 		/* up the hold count until we can be written out */
1267 		dmu_buf_add_ref(ds->ds_dbuf, ds);
1268 	}
1269 }
1270 
1271 /*
1272  * The unique space in the head dataset can be calculated by subtracting
1273  * the space used in the most recent snapshot, that is still being used
1274  * in this file system, from the space currently in use.  To figure out
1275  * the space in the most recent snapshot still in use, we need to take
1276  * the total space used in the snapshot and subtract out the space that
1277  * has been freed up since the snapshot was taken.
1278  */
1279 static void
1280 dsl_dataset_recalc_head_uniq(dsl_dataset_t *ds)
1281 {
1282 	uint64_t mrs_used;
1283 	uint64_t dlused, dlcomp, dluncomp;
1284 
1285 	ASSERT(ds->ds_object == ds->ds_dir->dd_phys->dd_head_dataset_obj);
1286 
1287 	if (ds->ds_phys->ds_prev_snap_obj != 0)
1288 		mrs_used = ds->ds_prev->ds_phys->ds_used_bytes;
1289 	else
1290 		mrs_used = 0;
1291 
1292 	VERIFY(0 == bplist_space(&ds->ds_deadlist, &dlused, &dlcomp,
1293 	    &dluncomp));
1294 
1295 	ASSERT3U(dlused, <=, mrs_used);
1296 	ds->ds_phys->ds_unique_bytes =
1297 	    ds->ds_phys->ds_used_bytes - (mrs_used - dlused);
1298 
1299 	if (!DS_UNIQUE_IS_ACCURATE(ds) &&
1300 	    spa_version(ds->ds_dir->dd_pool->dp_spa) >=
1301 	    SPA_VERSION_UNIQUE_ACCURATE)
1302 		ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
1303 }
1304 
1305 static uint64_t
1306 dsl_dataset_unique(dsl_dataset_t *ds)
1307 {
1308 	if (!DS_UNIQUE_IS_ACCURATE(ds) && !dsl_dataset_is_snapshot(ds))
1309 		dsl_dataset_recalc_head_uniq(ds);
1310 
1311 	return (ds->ds_phys->ds_unique_bytes);
1312 }
1313 
1314 struct killarg {
1315 	dsl_dataset_t *ds;
1316 	zio_t *zio;
1317 	dmu_tx_t *tx;
1318 };
1319 
1320 /* ARGSUSED */
1321 static int
1322 kill_blkptr(spa_t *spa, blkptr_t *bp, const zbookmark_t *zb,
1323     const dnode_phys_t *dnp, void *arg)
1324 {
1325 	struct killarg *ka = arg;
1326 
1327 	if (bp == NULL)
1328 		return (0);
1329 
1330 	if ((zb->zb_level == -1ULL && zb->zb_blkid != 0) ||
1331 	    (zb->zb_object != 0 && dnp == NULL)) {
1332 		/*
1333 		 * It's a block in the intent log.  It has no
1334 		 * accounting, so just free it.
1335 		 */
1336 		VERIFY3U(0, ==, dsl_free(ka->zio, ka->tx->tx_pool,
1337 		    ka->tx->tx_txg, bp, NULL, NULL, ARC_NOWAIT));
1338 	} else {
1339 		ASSERT3U(bp->blk_birth, >, ka->ds->ds_phys->ds_prev_snap_txg);
1340 		(void) dsl_dataset_block_kill(ka->ds, bp, ka->zio, ka->tx);
1341 	}
1342 
1343 	return (0);
1344 }
1345 
1346 /* ARGSUSED */
1347 static int
1348 dsl_dataset_rollback_check(void *arg1, void *arg2, dmu_tx_t *tx)
1349 {
1350 	dsl_dataset_t *ds = arg1;
1351 	dmu_objset_type_t *ost = arg2;
1352 
1353 	/*
1354 	 * We can only roll back to emptyness if it is a ZPL objset.
1355 	 */
1356 	if (*ost != DMU_OST_ZFS &&
1357 	    ds->ds_phys->ds_prev_snap_txg < TXG_INITIAL)
1358 		return (EINVAL);
1359 
1360 	/*
1361 	 * This must not be a snapshot.
1362 	 */
1363 	if (ds->ds_phys->ds_next_snap_obj != 0)
1364 		return (EINVAL);
1365 
1366 	/*
1367 	 * If we made changes this txg, traverse_dataset won't find
1368 	 * them.  Try again.
1369 	 */
1370 	if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
1371 		return (EAGAIN);
1372 
1373 	return (0);
1374 }
1375 
1376 /* ARGSUSED */
1377 static void
1378 dsl_dataset_rollback_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
1379 {
1380 	dsl_dataset_t *ds = arg1;
1381 	dmu_objset_type_t *ost = arg2;
1382 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1383 
1384 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1385 
1386 	if (ds->ds_user_ptr != NULL) {
1387 		/*
1388 		 * We need to make sure that the objset_impl_t is reopened after
1389 		 * we do the rollback, otherwise it will have the wrong
1390 		 * objset_phys_t.  Normally this would happen when this
1391 		 * dataset-open is closed, thus causing the
1392 		 * dataset to be immediately evicted.  But when doing "zfs recv
1393 		 * -F", we reopen the objset before that, so that there is no
1394 		 * window where the dataset is closed and inconsistent.
1395 		 */
1396 		ds->ds_user_evict_func(ds, ds->ds_user_ptr);
1397 		ds->ds_user_ptr = NULL;
1398 	}
1399 
1400 	/* Transfer space that was freed since last snap back to the head. */
1401 	{
1402 		uint64_t used;
1403 
1404 		VERIFY(0 == bplist_space_birthrange(&ds->ds_deadlist,
1405 		    ds->ds_origin_txg, UINT64_MAX, &used));
1406 		dsl_dir_transfer_space(ds->ds_dir, used,
1407 		    DD_USED_SNAP, DD_USED_HEAD, tx);
1408 	}
1409 
1410 	/* Zero out the deadlist. */
1411 	bplist_close(&ds->ds_deadlist);
1412 	bplist_destroy(mos, ds->ds_phys->ds_deadlist_obj, tx);
1413 	ds->ds_phys->ds_deadlist_obj =
1414 	    bplist_create(mos, DSL_DEADLIST_BLOCKSIZE, tx);
1415 	VERIFY(0 == bplist_open(&ds->ds_deadlist, mos,
1416 	    ds->ds_phys->ds_deadlist_obj));
1417 
1418 	{
1419 		/*
1420 		 * Free blkptrs that we gave birth to - this covers
1421 		 * claimed but not played log blocks too.
1422 		 */
1423 		zio_t *zio;
1424 		struct killarg ka;
1425 
1426 		zio = zio_root(tx->tx_pool->dp_spa, NULL, NULL,
1427 		    ZIO_FLAG_MUSTSUCCEED);
1428 		ka.ds = ds;
1429 		ka.zio = zio;
1430 		ka.tx = tx;
1431 		(void) traverse_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
1432 		    TRAVERSE_POST, kill_blkptr, &ka);
1433 		(void) zio_wait(zio);
1434 	}
1435 
1436 	ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) || ds->ds_phys->ds_unique_bytes == 0);
1437 
1438 	if (ds->ds_prev && ds->ds_prev != ds->ds_dir->dd_pool->dp_origin_snap) {
1439 		/* Change our contents to that of the prev snapshot */
1440 
1441 		ASSERT3U(ds->ds_prev->ds_object, ==,
1442 		    ds->ds_phys->ds_prev_snap_obj);
1443 		ASSERT3U(ds->ds_phys->ds_used_bytes, <=,
1444 		    ds->ds_prev->ds_phys->ds_used_bytes);
1445 
1446 		ds->ds_phys->ds_bp = ds->ds_prev->ds_phys->ds_bp;
1447 		ds->ds_phys->ds_used_bytes =
1448 		    ds->ds_prev->ds_phys->ds_used_bytes;
1449 		ds->ds_phys->ds_compressed_bytes =
1450 		    ds->ds_prev->ds_phys->ds_compressed_bytes;
1451 		ds->ds_phys->ds_uncompressed_bytes =
1452 		    ds->ds_prev->ds_phys->ds_uncompressed_bytes;
1453 		ds->ds_phys->ds_flags = ds->ds_prev->ds_phys->ds_flags;
1454 
1455 		if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
1456 			dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
1457 			ds->ds_prev->ds_phys->ds_unique_bytes = 0;
1458 		}
1459 	} else {
1460 		objset_impl_t *osi;
1461 
1462 		ASSERT(*ost != DMU_OST_ZVOL);
1463 		ASSERT3U(ds->ds_phys->ds_used_bytes, ==, 0);
1464 		ASSERT3U(ds->ds_phys->ds_compressed_bytes, ==, 0);
1465 		ASSERT3U(ds->ds_phys->ds_uncompressed_bytes, ==, 0);
1466 
1467 		bzero(&ds->ds_phys->ds_bp, sizeof (blkptr_t));
1468 		ds->ds_phys->ds_flags = 0;
1469 		ds->ds_phys->ds_unique_bytes = 0;
1470 		if (spa_version(ds->ds_dir->dd_pool->dp_spa) >=
1471 		    SPA_VERSION_UNIQUE_ACCURATE)
1472 			ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
1473 
1474 		osi = dmu_objset_create_impl(ds->ds_dir->dd_pool->dp_spa, ds,
1475 		    &ds->ds_phys->ds_bp, *ost, tx);
1476 #ifdef _KERNEL
1477 		zfs_create_fs(&osi->os, kcred, NULL, tx);
1478 #endif
1479 	}
1480 
1481 	spa_history_internal_log(LOG_DS_ROLLBACK, ds->ds_dir->dd_pool->dp_spa,
1482 	    tx, cr, "dataset = %llu", ds->ds_object);
1483 }
1484 
1485 /* ARGSUSED */
1486 static int
1487 dsl_dataset_destroy_begin_check(void *arg1, void *arg2, dmu_tx_t *tx)
1488 {
1489 	dsl_dataset_t *ds = arg1;
1490 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1491 	uint64_t count;
1492 	int err;
1493 
1494 	/*
1495 	 * Can't delete a head dataset if there are snapshots of it.
1496 	 * (Except if the only snapshots are from the branch we cloned
1497 	 * from.)
1498 	 */
1499 	if (ds->ds_prev != NULL &&
1500 	    ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1501 		return (EINVAL);
1502 
1503 	/*
1504 	 * This is really a dsl_dir thing, but check it here so that
1505 	 * we'll be less likely to leave this dataset inconsistent &
1506 	 * nearly destroyed.
1507 	 */
1508 	err = zap_count(mos, ds->ds_dir->dd_phys->dd_child_dir_zapobj, &count);
1509 	if (err)
1510 		return (err);
1511 	if (count != 0)
1512 		return (EEXIST);
1513 
1514 	return (0);
1515 }
1516 
1517 /* ARGSUSED */
1518 static void
1519 dsl_dataset_destroy_begin_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
1520 {
1521 	dsl_dataset_t *ds = arg1;
1522 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1523 
1524 	/* Mark it as inconsistent on-disk, in case we crash */
1525 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1526 	ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
1527 
1528 	spa_history_internal_log(LOG_DS_DESTROY_BEGIN, dp->dp_spa, tx,
1529 	    cr, "dataset = %llu", ds->ds_object);
1530 }
1531 
1532 static int
1533 dsl_dataset_origin_check(struct dsl_ds_destroyarg *dsda, void *tag,
1534     dmu_tx_t *tx)
1535 {
1536 	dsl_dataset_t *ds = dsda->ds;
1537 	dsl_dataset_t *ds_prev = ds->ds_prev;
1538 
1539 	if (dsl_dataset_might_destroy_origin(ds_prev)) {
1540 		struct dsl_ds_destroyarg ndsda = {0};
1541 
1542 		/*
1543 		 * If we're not prepared to remove the origin, don't remove
1544 		 * the clone either.
1545 		 */
1546 		if (dsda->rm_origin == NULL) {
1547 			dsda->need_prep = B_TRUE;
1548 			return (EBUSY);
1549 		}
1550 
1551 		ndsda.ds = ds_prev;
1552 		ndsda.is_origin_rm = B_TRUE;
1553 		return (dsl_dataset_destroy_check(&ndsda, tag, tx));
1554 	}
1555 
1556 	/*
1557 	 * If we're not going to remove the origin after all,
1558 	 * undo the open context setup.
1559 	 */
1560 	if (dsda->rm_origin != NULL) {
1561 		dsl_dataset_disown(dsda->rm_origin, tag);
1562 		dsda->rm_origin = NULL;
1563 	}
1564 
1565 	return (0);
1566 }
1567 
1568 /* ARGSUSED */
1569 int
1570 dsl_dataset_destroy_check(void *arg1, void *arg2, dmu_tx_t *tx)
1571 {
1572 	struct dsl_ds_destroyarg *dsda = arg1;
1573 	dsl_dataset_t *ds = dsda->ds;
1574 
1575 	/* we have an owner hold, so noone else can destroy us */
1576 	ASSERT(!DSL_DATASET_IS_DESTROYED(ds));
1577 
1578 	/*
1579 	 * Only allow deferred destroy on pools that support it.
1580 	 * NOTE: deferred destroy is only supported on snapshots.
1581 	 */
1582 	if (dsda->defer) {
1583 		if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
1584 		    SPA_VERSION_USERREFS)
1585 			return (ENOTSUP);
1586 		ASSERT(dsl_dataset_is_snapshot(ds));
1587 		return (0);
1588 	}
1589 
1590 	/*
1591 	 * Can't delete a head dataset if there are snapshots of it.
1592 	 * (Except if the only snapshots are from the branch we cloned
1593 	 * from.)
1594 	 */
1595 	if (ds->ds_prev != NULL &&
1596 	    ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1597 		return (EINVAL);
1598 
1599 	/*
1600 	 * If we made changes this txg, traverse_dsl_dataset won't find
1601 	 * them.  Try again.
1602 	 */
1603 	if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
1604 		return (EAGAIN);
1605 
1606 	if (dsl_dataset_is_snapshot(ds)) {
1607 		/*
1608 		 * If this snapshot has an elevated user reference count,
1609 		 * we can't destroy it yet.
1610 		 */
1611 		if (ds->ds_userrefs > 0 && !dsda->releasing)
1612 			return (EBUSY);
1613 
1614 		mutex_enter(&ds->ds_lock);
1615 		/*
1616 		 * Can't delete a branch point. However, if we're destroying
1617 		 * a clone and removing its origin due to it having a user
1618 		 * hold count of 0 and having been marked for deferred destroy,
1619 		 * it's OK for the origin to have a single clone.
1620 		 */
1621 		if (ds->ds_phys->ds_num_children >
1622 		    (dsda->is_origin_rm ? 2 : 1)) {
1623 			mutex_exit(&ds->ds_lock);
1624 			return (EEXIST);
1625 		}
1626 		mutex_exit(&ds->ds_lock);
1627 	} else if (dsl_dir_is_clone(ds->ds_dir)) {
1628 		return (dsl_dataset_origin_check(dsda, arg2, tx));
1629 	}
1630 
1631 	/* XXX we should do some i/o error checking... */
1632 	return (0);
1633 }
1634 
1635 struct refsarg {
1636 	kmutex_t lock;
1637 	boolean_t gone;
1638 	kcondvar_t cv;
1639 };
1640 
1641 /* ARGSUSED */
1642 static void
1643 dsl_dataset_refs_gone(dmu_buf_t *db, void *argv)
1644 {
1645 	struct refsarg *arg = argv;
1646 
1647 	mutex_enter(&arg->lock);
1648 	arg->gone = TRUE;
1649 	cv_signal(&arg->cv);
1650 	mutex_exit(&arg->lock);
1651 }
1652 
1653 static void
1654 dsl_dataset_drain_refs(dsl_dataset_t *ds, void *tag)
1655 {
1656 	struct refsarg arg;
1657 
1658 	mutex_init(&arg.lock, NULL, MUTEX_DEFAULT, NULL);
1659 	cv_init(&arg.cv, NULL, CV_DEFAULT, NULL);
1660 	arg.gone = FALSE;
1661 	(void) dmu_buf_update_user(ds->ds_dbuf, ds, &arg, &ds->ds_phys,
1662 	    dsl_dataset_refs_gone);
1663 	dmu_buf_rele(ds->ds_dbuf, tag);
1664 	mutex_enter(&arg.lock);
1665 	while (!arg.gone)
1666 		cv_wait(&arg.cv, &arg.lock);
1667 	ASSERT(arg.gone);
1668 	mutex_exit(&arg.lock);
1669 	ds->ds_dbuf = NULL;
1670 	ds->ds_phys = NULL;
1671 	mutex_destroy(&arg.lock);
1672 	cv_destroy(&arg.cv);
1673 }
1674 
1675 void
1676 dsl_dataset_destroy_sync(void *arg1, void *tag, cred_t *cr, dmu_tx_t *tx)
1677 {
1678 	struct dsl_ds_destroyarg *dsda = arg1;
1679 	dsl_dataset_t *ds = dsda->ds;
1680 	zio_t *zio;
1681 	int err;
1682 	int after_branch_point = FALSE;
1683 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1684 	objset_t *mos = dp->dp_meta_objset;
1685 	dsl_dataset_t *ds_prev = NULL;
1686 	uint64_t obj;
1687 
1688 	ASSERT(ds->ds_owner);
1689 	ASSERT(dsda->defer || ds->ds_phys->ds_num_children <= 1);
1690 	ASSERT(ds->ds_prev == NULL ||
1691 	    ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
1692 	ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
1693 
1694 	if (dsda->defer) {
1695 		ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1696 		if (ds->ds_userrefs > 0 || ds->ds_phys->ds_num_children > 1) {
1697 			dmu_buf_will_dirty(ds->ds_dbuf, tx);
1698 			ds->ds_phys->ds_flags |= DS_FLAG_DEFER_DESTROY;
1699 			return;
1700 		}
1701 	}
1702 
1703 	/* signal any waiters that this dataset is going away */
1704 	mutex_enter(&ds->ds_lock);
1705 	ds->ds_owner = dsl_reaper;
1706 	cv_broadcast(&ds->ds_exclusive_cv);
1707 	mutex_exit(&ds->ds_lock);
1708 
1709 	/* Remove our reservation */
1710 	if (ds->ds_reserved != 0) {
1711 		uint64_t val = 0;
1712 		dsl_dataset_set_reservation_sync(ds, &val, cr, tx);
1713 		ASSERT3U(ds->ds_reserved, ==, 0);
1714 	}
1715 
1716 	ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1717 
1718 	dsl_pool_ds_destroyed(ds, tx);
1719 
1720 	obj = ds->ds_object;
1721 
1722 	if (ds->ds_phys->ds_prev_snap_obj != 0) {
1723 		if (ds->ds_prev) {
1724 			ds_prev = ds->ds_prev;
1725 		} else {
1726 			VERIFY(0 == dsl_dataset_hold_obj(dp,
1727 			    ds->ds_phys->ds_prev_snap_obj, FTAG, &ds_prev));
1728 		}
1729 		after_branch_point =
1730 		    (ds_prev->ds_phys->ds_next_snap_obj != obj);
1731 
1732 		dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
1733 		if (after_branch_point &&
1734 		    ds_prev->ds_phys->ds_next_clones_obj != 0) {
1735 			VERIFY3U(0, ==, zap_remove_int(mos,
1736 			    ds_prev->ds_phys->ds_next_clones_obj, obj, tx));
1737 			if (ds->ds_phys->ds_next_snap_obj != 0) {
1738 				VERIFY(0 == zap_add_int(mos,
1739 				    ds_prev->ds_phys->ds_next_clones_obj,
1740 				    ds->ds_phys->ds_next_snap_obj, tx));
1741 			}
1742 		}
1743 		if (after_branch_point &&
1744 		    ds->ds_phys->ds_next_snap_obj == 0) {
1745 			/* This clone is toast. */
1746 			ASSERT(ds_prev->ds_phys->ds_num_children > 1);
1747 			ds_prev->ds_phys->ds_num_children--;
1748 
1749 			/*
1750 			 * If the clone's origin has no other clones, no
1751 			 * user holds, and has been marked for deferred
1752 			 * deletion, then we should have done the necessary
1753 			 * destroy setup for it.
1754 			 */
1755 			if (ds_prev->ds_phys->ds_num_children == 1 &&
1756 			    ds_prev->ds_userrefs == 0 &&
1757 			    DS_IS_DEFER_DESTROY(ds_prev)) {
1758 				ASSERT3P(dsda->rm_origin, !=, NULL);
1759 			} else {
1760 				ASSERT3P(dsda->rm_origin, ==, NULL);
1761 			}
1762 		} else if (!after_branch_point) {
1763 			ds_prev->ds_phys->ds_next_snap_obj =
1764 			    ds->ds_phys->ds_next_snap_obj;
1765 		}
1766 	}
1767 
1768 	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
1769 
1770 	if (ds->ds_phys->ds_next_snap_obj != 0) {
1771 		blkptr_t bp;
1772 		dsl_dataset_t *ds_next;
1773 		uint64_t itor = 0;
1774 		uint64_t old_unique;
1775 		int64_t used = 0, compressed = 0, uncompressed = 0;
1776 
1777 		VERIFY(0 == dsl_dataset_hold_obj(dp,
1778 		    ds->ds_phys->ds_next_snap_obj, FTAG, &ds_next));
1779 		ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
1780 
1781 		old_unique = dsl_dataset_unique(ds_next);
1782 
1783 		dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
1784 		ds_next->ds_phys->ds_prev_snap_obj =
1785 		    ds->ds_phys->ds_prev_snap_obj;
1786 		ds_next->ds_phys->ds_prev_snap_txg =
1787 		    ds->ds_phys->ds_prev_snap_txg;
1788 		ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1789 		    ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
1790 
1791 		/*
1792 		 * Transfer to our deadlist (which will become next's
1793 		 * new deadlist) any entries from next's current
1794 		 * deadlist which were born before prev, and free the
1795 		 * other entries.
1796 		 *
1797 		 * XXX we're doing this long task with the config lock held
1798 		 */
1799 		while (bplist_iterate(&ds_next->ds_deadlist, &itor, &bp) == 0) {
1800 			if (bp.blk_birth <= ds->ds_phys->ds_prev_snap_txg) {
1801 				VERIFY(0 == bplist_enqueue(&ds->ds_deadlist,
1802 				    &bp, tx));
1803 				if (ds_prev && !after_branch_point &&
1804 				    bp.blk_birth >
1805 				    ds_prev->ds_phys->ds_prev_snap_txg) {
1806 					ds_prev->ds_phys->ds_unique_bytes +=
1807 					    bp_get_dasize(dp->dp_spa, &bp);
1808 				}
1809 			} else {
1810 				used += bp_get_dasize(dp->dp_spa, &bp);
1811 				compressed += BP_GET_PSIZE(&bp);
1812 				uncompressed += BP_GET_UCSIZE(&bp);
1813 				/* XXX check return value? */
1814 				(void) dsl_free(zio, dp, tx->tx_txg,
1815 				    &bp, NULL, NULL, ARC_NOWAIT);
1816 			}
1817 		}
1818 
1819 		ASSERT3U(used, ==, ds->ds_phys->ds_unique_bytes);
1820 
1821 		/* change snapused */
1822 		dsl_dir_diduse_space(ds->ds_dir, DD_USED_SNAP,
1823 		    -used, -compressed, -uncompressed, tx);
1824 
1825 		/* free next's deadlist */
1826 		bplist_close(&ds_next->ds_deadlist);
1827 		bplist_destroy(mos, ds_next->ds_phys->ds_deadlist_obj, tx);
1828 
1829 		/* set next's deadlist to our deadlist */
1830 		bplist_close(&ds->ds_deadlist);
1831 		ds_next->ds_phys->ds_deadlist_obj =
1832 		    ds->ds_phys->ds_deadlist_obj;
1833 		VERIFY(0 == bplist_open(&ds_next->ds_deadlist, mos,
1834 		    ds_next->ds_phys->ds_deadlist_obj));
1835 		ds->ds_phys->ds_deadlist_obj = 0;
1836 
1837 		if (ds_next->ds_phys->ds_next_snap_obj != 0) {
1838 			/*
1839 			 * Update next's unique to include blocks which
1840 			 * were previously shared by only this snapshot
1841 			 * and it.  Those blocks will be born after the
1842 			 * prev snap and before this snap, and will have
1843 			 * died after the next snap and before the one
1844 			 * after that (ie. be on the snap after next's
1845 			 * deadlist).
1846 			 *
1847 			 * XXX we're doing this long task with the
1848 			 * config lock held
1849 			 */
1850 			dsl_dataset_t *ds_after_next;
1851 			uint64_t space;
1852 
1853 			VERIFY(0 == dsl_dataset_hold_obj(dp,
1854 			    ds_next->ds_phys->ds_next_snap_obj,
1855 			    FTAG, &ds_after_next));
1856 
1857 			VERIFY(0 ==
1858 			    bplist_space_birthrange(&ds_after_next->ds_deadlist,
1859 			    ds->ds_phys->ds_prev_snap_txg,
1860 			    ds->ds_phys->ds_creation_txg, &space));
1861 			ds_next->ds_phys->ds_unique_bytes += space;
1862 
1863 			dsl_dataset_rele(ds_after_next, FTAG);
1864 			ASSERT3P(ds_next->ds_prev, ==, NULL);
1865 		} else {
1866 			ASSERT3P(ds_next->ds_prev, ==, ds);
1867 			dsl_dataset_drop_ref(ds_next->ds_prev, ds_next);
1868 			ds_next->ds_prev = NULL;
1869 			if (ds_prev) {
1870 				VERIFY(0 == dsl_dataset_get_ref(dp,
1871 				    ds->ds_phys->ds_prev_snap_obj,
1872 				    ds_next, &ds_next->ds_prev));
1873 			}
1874 
1875 			dsl_dataset_recalc_head_uniq(ds_next);
1876 
1877 			/*
1878 			 * Reduce the amount of our unconsmed refreservation
1879 			 * being charged to our parent by the amount of
1880 			 * new unique data we have gained.
1881 			 */
1882 			if (old_unique < ds_next->ds_reserved) {
1883 				int64_t mrsdelta;
1884 				uint64_t new_unique =
1885 				    ds_next->ds_phys->ds_unique_bytes;
1886 
1887 				ASSERT(old_unique <= new_unique);
1888 				mrsdelta = MIN(new_unique - old_unique,
1889 				    ds_next->ds_reserved - old_unique);
1890 				dsl_dir_diduse_space(ds->ds_dir,
1891 				    DD_USED_REFRSRV, -mrsdelta, 0, 0, tx);
1892 			}
1893 		}
1894 		dsl_dataset_rele(ds_next, FTAG);
1895 	} else {
1896 		/*
1897 		 * There's no next snapshot, so this is a head dataset.
1898 		 * Destroy the deadlist.  Unless it's a clone, the
1899 		 * deadlist should be empty.  (If it's a clone, it's
1900 		 * safe to ignore the deadlist contents.)
1901 		 */
1902 		struct killarg ka;
1903 
1904 		ASSERT(after_branch_point || bplist_empty(&ds->ds_deadlist));
1905 		bplist_close(&ds->ds_deadlist);
1906 		bplist_destroy(mos, ds->ds_phys->ds_deadlist_obj, tx);
1907 		ds->ds_phys->ds_deadlist_obj = 0;
1908 
1909 		/*
1910 		 * Free everything that we point to (that's born after
1911 		 * the previous snapshot, if we are a clone)
1912 		 *
1913 		 * NB: this should be very quick, because we already
1914 		 * freed all the objects in open context.
1915 		 */
1916 		ka.ds = ds;
1917 		ka.zio = zio;
1918 		ka.tx = tx;
1919 		err = traverse_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
1920 		    TRAVERSE_POST, kill_blkptr, &ka);
1921 		ASSERT3U(err, ==, 0);
1922 		ASSERT(!DS_UNIQUE_IS_ACCURATE(ds) ||
1923 		    ds->ds_phys->ds_unique_bytes == 0);
1924 	}
1925 
1926 	err = zio_wait(zio);
1927 	ASSERT3U(err, ==, 0);
1928 
1929 	if (ds->ds_dir->dd_phys->dd_head_dataset_obj == ds->ds_object) {
1930 		/* Erase the link in the dir */
1931 		dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
1932 		ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
1933 		ASSERT(ds->ds_phys->ds_snapnames_zapobj != 0);
1934 		err = zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx);
1935 		ASSERT(err == 0);
1936 	} else {
1937 		/* remove from snapshot namespace */
1938 		dsl_dataset_t *ds_head;
1939 		ASSERT(ds->ds_phys->ds_snapnames_zapobj == 0);
1940 		VERIFY(0 == dsl_dataset_hold_obj(dp,
1941 		    ds->ds_dir->dd_phys->dd_head_dataset_obj, FTAG, &ds_head));
1942 		VERIFY(0 == dsl_dataset_get_snapname(ds));
1943 #ifdef ZFS_DEBUG
1944 		{
1945 			uint64_t val;
1946 
1947 			err = dsl_dataset_snap_lookup(ds_head,
1948 			    ds->ds_snapname, &val);
1949 			ASSERT3U(err, ==, 0);
1950 			ASSERT3U(val, ==, obj);
1951 		}
1952 #endif
1953 		err = dsl_dataset_snap_remove(ds_head, ds->ds_snapname, tx);
1954 		ASSERT(err == 0);
1955 		dsl_dataset_rele(ds_head, FTAG);
1956 	}
1957 
1958 	if (ds_prev && ds->ds_prev != ds_prev)
1959 		dsl_dataset_rele(ds_prev, FTAG);
1960 
1961 	spa_prop_clear_bootfs(dp->dp_spa, ds->ds_object, tx);
1962 	spa_history_internal_log(LOG_DS_DESTROY, dp->dp_spa, tx,
1963 	    cr, "dataset = %llu", ds->ds_object);
1964 
1965 	if (ds->ds_phys->ds_next_clones_obj != 0) {
1966 		uint64_t count;
1967 		ASSERT(0 == zap_count(mos,
1968 		    ds->ds_phys->ds_next_clones_obj, &count) && count == 0);
1969 		VERIFY(0 == dmu_object_free(mos,
1970 		    ds->ds_phys->ds_next_clones_obj, tx));
1971 	}
1972 	if (ds->ds_phys->ds_props_obj != 0)
1973 		VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_props_obj, tx));
1974 	if (ds->ds_phys->ds_userrefs_obj != 0)
1975 		VERIFY(0 == zap_destroy(mos, ds->ds_phys->ds_userrefs_obj, tx));
1976 	dsl_dir_close(ds->ds_dir, ds);
1977 	ds->ds_dir = NULL;
1978 	dsl_dataset_drain_refs(ds, tag);
1979 	VERIFY(0 == dmu_object_free(mos, obj, tx));
1980 
1981 	if (dsda->rm_origin) {
1982 		/*
1983 		 * Remove the origin of the clone we just destroyed.
1984 		 */
1985 		dsl_dataset_t *origin = ds->ds_prev;
1986 		struct dsl_ds_destroyarg ndsda = {0};
1987 
1988 		ASSERT3P(origin, ==, dsda->rm_origin);
1989 		if (origin->ds_user_ptr) {
1990 			origin->ds_user_evict_func(origin, origin->ds_user_ptr);
1991 			origin->ds_user_ptr = NULL;
1992 		}
1993 
1994 		dsl_dataset_rele(origin, tag);
1995 		ds->ds_prev = NULL;
1996 
1997 		ndsda.ds = origin;
1998 		dsl_dataset_destroy_sync(&ndsda, tag, cr, tx);
1999 	}
2000 }
2001 
2002 static int
2003 dsl_dataset_snapshot_reserve_space(dsl_dataset_t *ds, dmu_tx_t *tx)
2004 {
2005 	uint64_t asize;
2006 
2007 	if (!dmu_tx_is_syncing(tx))
2008 		return (0);
2009 
2010 	/*
2011 	 * If there's an fs-only reservation, any blocks that might become
2012 	 * owned by the snapshot dataset must be accommodated by space
2013 	 * outside of the reservation.
2014 	 */
2015 	asize = MIN(dsl_dataset_unique(ds), ds->ds_reserved);
2016 	if (asize > dsl_dir_space_available(ds->ds_dir, NULL, 0, FALSE))
2017 		return (ENOSPC);
2018 
2019 	/*
2020 	 * Propogate any reserved space for this snapshot to other
2021 	 * snapshot checks in this sync group.
2022 	 */
2023 	if (asize > 0)
2024 		dsl_dir_willuse_space(ds->ds_dir, asize, tx);
2025 
2026 	return (0);
2027 }
2028 
2029 /* ARGSUSED */
2030 int
2031 dsl_dataset_snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
2032 {
2033 	dsl_dataset_t *ds = arg1;
2034 	const char *snapname = arg2;
2035 	int err;
2036 	uint64_t value;
2037 
2038 	/*
2039 	 * We don't allow multiple snapshots of the same txg.  If there
2040 	 * is already one, try again.
2041 	 */
2042 	if (ds->ds_phys->ds_prev_snap_txg >= tx->tx_txg)
2043 		return (EAGAIN);
2044 
2045 	/*
2046 	 * Check for conflicting name snapshot name.
2047 	 */
2048 	err = dsl_dataset_snap_lookup(ds, snapname, &value);
2049 	if (err == 0)
2050 		return (EEXIST);
2051 	if (err != ENOENT)
2052 		return (err);
2053 
2054 	/*
2055 	 * Check that the dataset's name is not too long.  Name consists
2056 	 * of the dataset's length + 1 for the @-sign + snapshot name's length
2057 	 */
2058 	if (dsl_dataset_namelen(ds) + 1 + strlen(snapname) >= MAXNAMELEN)
2059 		return (ENAMETOOLONG);
2060 
2061 	err = dsl_dataset_snapshot_reserve_space(ds, tx);
2062 	if (err)
2063 		return (err);
2064 
2065 	ds->ds_trysnap_txg = tx->tx_txg;
2066 	return (0);
2067 }
2068 
2069 void
2070 dsl_dataset_snapshot_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
2071 {
2072 	dsl_dataset_t *ds = arg1;
2073 	const char *snapname = arg2;
2074 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
2075 	dmu_buf_t *dbuf;
2076 	dsl_dataset_phys_t *dsphys;
2077 	uint64_t dsobj, crtxg;
2078 	objset_t *mos = dp->dp_meta_objset;
2079 	int err;
2080 
2081 	ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
2082 
2083 	/*
2084 	 * The origin's ds_creation_txg has to be < TXG_INITIAL
2085 	 */
2086 	if (strcmp(snapname, ORIGIN_DIR_NAME) == 0)
2087 		crtxg = 1;
2088 	else
2089 		crtxg = tx->tx_txg;
2090 
2091 	dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
2092 	    DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
2093 	VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
2094 	dmu_buf_will_dirty(dbuf, tx);
2095 	dsphys = dbuf->db_data;
2096 	bzero(dsphys, sizeof (dsl_dataset_phys_t));
2097 	dsphys->ds_dir_obj = ds->ds_dir->dd_object;
2098 	dsphys->ds_fsid_guid = unique_create();
2099 	(void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
2100 	    sizeof (dsphys->ds_guid));
2101 	dsphys->ds_prev_snap_obj = ds->ds_phys->ds_prev_snap_obj;
2102 	dsphys->ds_prev_snap_txg = ds->ds_phys->ds_prev_snap_txg;
2103 	dsphys->ds_next_snap_obj = ds->ds_object;
2104 	dsphys->ds_num_children = 1;
2105 	dsphys->ds_creation_time = gethrestime_sec();
2106 	dsphys->ds_creation_txg = crtxg;
2107 	dsphys->ds_deadlist_obj = ds->ds_phys->ds_deadlist_obj;
2108 	dsphys->ds_used_bytes = ds->ds_phys->ds_used_bytes;
2109 	dsphys->ds_compressed_bytes = ds->ds_phys->ds_compressed_bytes;
2110 	dsphys->ds_uncompressed_bytes = ds->ds_phys->ds_uncompressed_bytes;
2111 	dsphys->ds_flags = ds->ds_phys->ds_flags;
2112 	dsphys->ds_bp = ds->ds_phys->ds_bp;
2113 	dmu_buf_rele(dbuf, FTAG);
2114 
2115 	ASSERT3U(ds->ds_prev != 0, ==, ds->ds_phys->ds_prev_snap_obj != 0);
2116 	if (ds->ds_prev) {
2117 		uint64_t next_clones_obj =
2118 		    ds->ds_prev->ds_phys->ds_next_clones_obj;
2119 		ASSERT(ds->ds_prev->ds_phys->ds_next_snap_obj ==
2120 		    ds->ds_object ||
2121 		    ds->ds_prev->ds_phys->ds_num_children > 1);
2122 		if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
2123 			dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
2124 			ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
2125 			    ds->ds_prev->ds_phys->ds_creation_txg);
2126 			ds->ds_prev->ds_phys->ds_next_snap_obj = dsobj;
2127 		} else if (next_clones_obj != 0) {
2128 			VERIFY3U(0, ==, zap_remove_int(mos,
2129 			    next_clones_obj, dsphys->ds_next_snap_obj, tx));
2130 			VERIFY3U(0, ==, zap_add_int(mos,
2131 			    next_clones_obj, dsobj, tx));
2132 		}
2133 	}
2134 
2135 	/*
2136 	 * If we have a reference-reservation on this dataset, we will
2137 	 * need to increase the amount of refreservation being charged
2138 	 * since our unique space is going to zero.
2139 	 */
2140 	if (ds->ds_reserved) {
2141 		int64_t add = MIN(dsl_dataset_unique(ds), ds->ds_reserved);
2142 		dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV,
2143 		    add, 0, 0, tx);
2144 	}
2145 
2146 	bplist_close(&ds->ds_deadlist);
2147 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
2148 	ASSERT3U(ds->ds_phys->ds_prev_snap_txg, <, tx->tx_txg);
2149 	ds->ds_phys->ds_prev_snap_obj = dsobj;
2150 	ds->ds_phys->ds_prev_snap_txg = crtxg;
2151 	ds->ds_phys->ds_unique_bytes = 0;
2152 	if (spa_version(dp->dp_spa) >= SPA_VERSION_UNIQUE_ACCURATE)
2153 		ds->ds_phys->ds_flags |= DS_FLAG_UNIQUE_ACCURATE;
2154 	ds->ds_phys->ds_deadlist_obj =
2155 	    bplist_create(mos, DSL_DEADLIST_BLOCKSIZE, tx);
2156 	VERIFY(0 == bplist_open(&ds->ds_deadlist, mos,
2157 	    ds->ds_phys->ds_deadlist_obj));
2158 
2159 	dprintf("snap '%s' -> obj %llu\n", snapname, dsobj);
2160 	err = zap_add(mos, ds->ds_phys->ds_snapnames_zapobj,
2161 	    snapname, 8, 1, &dsobj, tx);
2162 	ASSERT(err == 0);
2163 
2164 	if (ds->ds_prev)
2165 		dsl_dataset_drop_ref(ds->ds_prev, ds);
2166 	VERIFY(0 == dsl_dataset_get_ref(dp,
2167 	    ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
2168 
2169 	dsl_pool_ds_snapshotted(ds, tx);
2170 
2171 	spa_history_internal_log(LOG_DS_SNAPSHOT, dp->dp_spa, tx, cr,
2172 	    "dataset = %llu", dsobj);
2173 }
2174 
2175 void
2176 dsl_dataset_sync(dsl_dataset_t *ds, zio_t *zio, dmu_tx_t *tx)
2177 {
2178 	ASSERT(dmu_tx_is_syncing(tx));
2179 	ASSERT(ds->ds_user_ptr != NULL);
2180 	ASSERT(ds->ds_phys->ds_next_snap_obj == 0);
2181 
2182 	/*
2183 	 * in case we had to change ds_fsid_guid when we opened it,
2184 	 * sync it out now.
2185 	 */
2186 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
2187 	ds->ds_phys->ds_fsid_guid = ds->ds_fsid_guid;
2188 
2189 	dsl_dir_dirty(ds->ds_dir, tx);
2190 	dmu_objset_sync(ds->ds_user_ptr, zio, tx);
2191 }
2192 
2193 void
2194 dsl_dataset_stats(dsl_dataset_t *ds, nvlist_t *nv)
2195 {
2196 	uint64_t refd, avail, uobjs, aobjs;
2197 
2198 	dsl_dir_stats(ds->ds_dir, nv);
2199 
2200 	dsl_dataset_space(ds, &refd, &avail, &uobjs, &aobjs);
2201 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_AVAILABLE, avail);
2202 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFERENCED, refd);
2203 
2204 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATION,
2205 	    ds->ds_phys->ds_creation_time);
2206 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATETXG,
2207 	    ds->ds_phys->ds_creation_txg);
2208 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFQUOTA,
2209 	    ds->ds_quota);
2210 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFRESERVATION,
2211 	    ds->ds_reserved);
2212 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_GUID,
2213 	    ds->ds_phys->ds_guid);
2214 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USERREFS, ds->ds_userrefs);
2215 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_DEFER_DESTROY,
2216 	    DS_IS_DEFER_DESTROY(ds) ? 1 : 0);
2217 
2218 	if (ds->ds_phys->ds_next_snap_obj) {
2219 		/*
2220 		 * This is a snapshot; override the dd's space used with
2221 		 * our unique space and compression ratio.
2222 		 */
2223 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USED,
2224 		    ds->ds_phys->ds_unique_bytes);
2225 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_COMPRESSRATIO,
2226 		    ds->ds_phys->ds_compressed_bytes == 0 ? 100 :
2227 		    (ds->ds_phys->ds_uncompressed_bytes * 100 /
2228 		    ds->ds_phys->ds_compressed_bytes));
2229 	}
2230 }
2231 
2232 void
2233 dsl_dataset_fast_stat(dsl_dataset_t *ds, dmu_objset_stats_t *stat)
2234 {
2235 	stat->dds_creation_txg = ds->ds_phys->ds_creation_txg;
2236 	stat->dds_inconsistent = ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT;
2237 	stat->dds_guid = ds->ds_phys->ds_guid;
2238 	if (ds->ds_phys->ds_next_snap_obj) {
2239 		stat->dds_is_snapshot = B_TRUE;
2240 		stat->dds_num_clones = ds->ds_phys->ds_num_children - 1;
2241 	} else {
2242 		stat->dds_is_snapshot = B_FALSE;
2243 		stat->dds_num_clones = 0;
2244 	}
2245 
2246 	/* clone origin is really a dsl_dir thing... */
2247 	rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
2248 	if (dsl_dir_is_clone(ds->ds_dir)) {
2249 		dsl_dataset_t *ods;
2250 
2251 		VERIFY(0 == dsl_dataset_get_ref(ds->ds_dir->dd_pool,
2252 		    ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &ods));
2253 		dsl_dataset_name(ods, stat->dds_origin);
2254 		dsl_dataset_drop_ref(ods, FTAG);
2255 	} else {
2256 		stat->dds_origin[0] = '\0';
2257 	}
2258 	rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
2259 }
2260 
2261 uint64_t
2262 dsl_dataset_fsid_guid(dsl_dataset_t *ds)
2263 {
2264 	return (ds->ds_fsid_guid);
2265 }
2266 
2267 void
2268 dsl_dataset_space(dsl_dataset_t *ds,
2269     uint64_t *refdbytesp, uint64_t *availbytesp,
2270     uint64_t *usedobjsp, uint64_t *availobjsp)
2271 {
2272 	*refdbytesp = ds->ds_phys->ds_used_bytes;
2273 	*availbytesp = dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE);
2274 	if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes)
2275 		*availbytesp += ds->ds_reserved - ds->ds_phys->ds_unique_bytes;
2276 	if (ds->ds_quota != 0) {
2277 		/*
2278 		 * Adjust available bytes according to refquota
2279 		 */
2280 		if (*refdbytesp < ds->ds_quota)
2281 			*availbytesp = MIN(*availbytesp,
2282 			    ds->ds_quota - *refdbytesp);
2283 		else
2284 			*availbytesp = 0;
2285 	}
2286 	*usedobjsp = ds->ds_phys->ds_bp.blk_fill;
2287 	*availobjsp = DN_MAX_OBJECT - *usedobjsp;
2288 }
2289 
2290 boolean_t
2291 dsl_dataset_modified_since_lastsnap(dsl_dataset_t *ds)
2292 {
2293 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
2294 
2295 	ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
2296 	    dsl_pool_sync_context(dp));
2297 	if (ds->ds_prev == NULL)
2298 		return (B_FALSE);
2299 	if (ds->ds_phys->ds_bp.blk_birth >
2300 	    ds->ds_prev->ds_phys->ds_creation_txg)
2301 		return (B_TRUE);
2302 	return (B_FALSE);
2303 }
2304 
2305 /* ARGSUSED */
2306 static int
2307 dsl_dataset_snapshot_rename_check(void *arg1, void *arg2, dmu_tx_t *tx)
2308 {
2309 	dsl_dataset_t *ds = arg1;
2310 	char *newsnapname = arg2;
2311 	dsl_dir_t *dd = ds->ds_dir;
2312 	dsl_dataset_t *hds;
2313 	uint64_t val;
2314 	int err;
2315 
2316 	err = dsl_dataset_hold_obj(dd->dd_pool,
2317 	    dd->dd_phys->dd_head_dataset_obj, FTAG, &hds);
2318 	if (err)
2319 		return (err);
2320 
2321 	/* new name better not be in use */
2322 	err = dsl_dataset_snap_lookup(hds, newsnapname, &val);
2323 	dsl_dataset_rele(hds, FTAG);
2324 
2325 	if (err == 0)
2326 		err = EEXIST;
2327 	else if (err == ENOENT)
2328 		err = 0;
2329 
2330 	/* dataset name + 1 for the "@" + the new snapshot name must fit */
2331 	if (dsl_dir_namelen(ds->ds_dir) + 1 + strlen(newsnapname) >= MAXNAMELEN)
2332 		err = ENAMETOOLONG;
2333 
2334 	return (err);
2335 }
2336 
2337 static void
2338 dsl_dataset_snapshot_rename_sync(void *arg1, void *arg2,
2339     cred_t *cr, dmu_tx_t *tx)
2340 {
2341 	dsl_dataset_t *ds = arg1;
2342 	const char *newsnapname = arg2;
2343 	dsl_dir_t *dd = ds->ds_dir;
2344 	objset_t *mos = dd->dd_pool->dp_meta_objset;
2345 	dsl_dataset_t *hds;
2346 	int err;
2347 
2348 	ASSERT(ds->ds_phys->ds_next_snap_obj != 0);
2349 
2350 	VERIFY(0 == dsl_dataset_hold_obj(dd->dd_pool,
2351 	    dd->dd_phys->dd_head_dataset_obj, FTAG, &hds));
2352 
2353 	VERIFY(0 == dsl_dataset_get_snapname(ds));
2354 	err = dsl_dataset_snap_remove(hds, ds->ds_snapname, tx);
2355 	ASSERT3U(err, ==, 0);
2356 	mutex_enter(&ds->ds_lock);
2357 	(void) strcpy(ds->ds_snapname, newsnapname);
2358 	mutex_exit(&ds->ds_lock);
2359 	err = zap_add(mos, hds->ds_phys->ds_snapnames_zapobj,
2360 	    ds->ds_snapname, 8, 1, &ds->ds_object, tx);
2361 	ASSERT3U(err, ==, 0);
2362 
2363 	spa_history_internal_log(LOG_DS_RENAME, dd->dd_pool->dp_spa, tx,
2364 	    cr, "dataset = %llu", ds->ds_object);
2365 	dsl_dataset_rele(hds, FTAG);
2366 }
2367 
2368 struct renamesnaparg {
2369 	dsl_sync_task_group_t *dstg;
2370 	char failed[MAXPATHLEN];
2371 	char *oldsnap;
2372 	char *newsnap;
2373 };
2374 
2375 static int
2376 dsl_snapshot_rename_one(char *name, void *arg)
2377 {
2378 	struct renamesnaparg *ra = arg;
2379 	dsl_dataset_t *ds = NULL;
2380 	char *cp;
2381 	int err;
2382 
2383 	cp = name + strlen(name);
2384 	*cp = '@';
2385 	(void) strcpy(cp + 1, ra->oldsnap);
2386 
2387 	/*
2388 	 * For recursive snapshot renames the parent won't be changing
2389 	 * so we just pass name for both the to/from argument.
2390 	 */
2391 	err = zfs_secpolicy_rename_perms(name, name, CRED());
2392 	if (err == ENOENT) {
2393 		return (0);
2394 	} else if (err) {
2395 		(void) strcpy(ra->failed, name);
2396 		return (err);
2397 	}
2398 
2399 #ifdef _KERNEL
2400 	/*
2401 	 * For all filesystems undergoing rename, we'll need to unmount it.
2402 	 */
2403 	(void) zfs_unmount_snap(name, NULL);
2404 #endif
2405 	err = dsl_dataset_hold(name, ra->dstg, &ds);
2406 	*cp = '\0';
2407 	if (err == ENOENT) {
2408 		return (0);
2409 	} else if (err) {
2410 		(void) strcpy(ra->failed, name);
2411 		return (err);
2412 	}
2413 
2414 	dsl_sync_task_create(ra->dstg, dsl_dataset_snapshot_rename_check,
2415 	    dsl_dataset_snapshot_rename_sync, ds, ra->newsnap, 0);
2416 
2417 	return (0);
2418 }
2419 
2420 static int
2421 dsl_recursive_rename(char *oldname, const char *newname)
2422 {
2423 	int err;
2424 	struct renamesnaparg *ra;
2425 	dsl_sync_task_t *dst;
2426 	spa_t *spa;
2427 	char *cp, *fsname = spa_strdup(oldname);
2428 	int len = strlen(oldname);
2429 
2430 	/* truncate the snapshot name to get the fsname */
2431 	cp = strchr(fsname, '@');
2432 	*cp = '\0';
2433 
2434 	err = spa_open(fsname, &spa, FTAG);
2435 	if (err) {
2436 		kmem_free(fsname, len + 1);
2437 		return (err);
2438 	}
2439 	ra = kmem_alloc(sizeof (struct renamesnaparg), KM_SLEEP);
2440 	ra->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
2441 
2442 	ra->oldsnap = strchr(oldname, '@') + 1;
2443 	ra->newsnap = strchr(newname, '@') + 1;
2444 	*ra->failed = '\0';
2445 
2446 	err = dmu_objset_find(fsname, dsl_snapshot_rename_one, ra,
2447 	    DS_FIND_CHILDREN);
2448 	kmem_free(fsname, len + 1);
2449 
2450 	if (err == 0) {
2451 		err = dsl_sync_task_group_wait(ra->dstg);
2452 	}
2453 
2454 	for (dst = list_head(&ra->dstg->dstg_tasks); dst;
2455 	    dst = list_next(&ra->dstg->dstg_tasks, dst)) {
2456 		dsl_dataset_t *ds = dst->dst_arg1;
2457 		if (dst->dst_err) {
2458 			dsl_dir_name(ds->ds_dir, ra->failed);
2459 			(void) strcat(ra->failed, "@");
2460 			(void) strcat(ra->failed, ra->newsnap);
2461 		}
2462 		dsl_dataset_rele(ds, ra->dstg);
2463 	}
2464 
2465 	if (err)
2466 		(void) strcpy(oldname, ra->failed);
2467 
2468 	dsl_sync_task_group_destroy(ra->dstg);
2469 	kmem_free(ra, sizeof (struct renamesnaparg));
2470 	spa_close(spa, FTAG);
2471 	return (err);
2472 }
2473 
2474 static int
2475 dsl_valid_rename(char *oldname, void *arg)
2476 {
2477 	int delta = *(int *)arg;
2478 
2479 	if (strlen(oldname) + delta >= MAXNAMELEN)
2480 		return (ENAMETOOLONG);
2481 
2482 	return (0);
2483 }
2484 
2485 #pragma weak dmu_objset_rename = dsl_dataset_rename
2486 int
2487 dsl_dataset_rename(char *oldname, const char *newname, boolean_t recursive)
2488 {
2489 	dsl_dir_t *dd;
2490 	dsl_dataset_t *ds;
2491 	const char *tail;
2492 	int err;
2493 
2494 	err = dsl_dir_open(oldname, FTAG, &dd, &tail);
2495 	if (err)
2496 		return (err);
2497 	/*
2498 	 * If there are more than 2 references there may be holds
2499 	 * hanging around that haven't been cleared out yet.
2500 	 */
2501 	if (dmu_buf_refcount(dd->dd_dbuf) > 2)
2502 		txg_wait_synced(dd->dd_pool, 0);
2503 	if (tail == NULL) {
2504 		int delta = strlen(newname) - strlen(oldname);
2505 
2506 		/* if we're growing, validate child name lengths */
2507 		if (delta > 0)
2508 			err = dmu_objset_find(oldname, dsl_valid_rename,
2509 			    &delta, DS_FIND_CHILDREN | DS_FIND_SNAPSHOTS);
2510 
2511 		if (!err)
2512 			err = dsl_dir_rename(dd, newname);
2513 		dsl_dir_close(dd, FTAG);
2514 		return (err);
2515 	}
2516 	if (tail[0] != '@') {
2517 		/* the name ended in a nonexistant component */
2518 		dsl_dir_close(dd, FTAG);
2519 		return (ENOENT);
2520 	}
2521 
2522 	dsl_dir_close(dd, FTAG);
2523 
2524 	/* new name must be snapshot in same filesystem */
2525 	tail = strchr(newname, '@');
2526 	if (tail == NULL)
2527 		return (EINVAL);
2528 	tail++;
2529 	if (strncmp(oldname, newname, tail - newname) != 0)
2530 		return (EXDEV);
2531 
2532 	if (recursive) {
2533 		err = dsl_recursive_rename(oldname, newname);
2534 	} else {
2535 		err = dsl_dataset_hold(oldname, FTAG, &ds);
2536 		if (err)
2537 			return (err);
2538 
2539 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
2540 		    dsl_dataset_snapshot_rename_check,
2541 		    dsl_dataset_snapshot_rename_sync, ds, (char *)tail, 1);
2542 
2543 		dsl_dataset_rele(ds, FTAG);
2544 	}
2545 
2546 	return (err);
2547 }
2548 
2549 struct promotenode {
2550 	list_node_t link;
2551 	dsl_dataset_t *ds;
2552 };
2553 
2554 struct promotearg {
2555 	list_t shared_snaps, origin_snaps, clone_snaps;
2556 	dsl_dataset_t *origin_origin, *origin_head;
2557 	uint64_t used, comp, uncomp, unique, cloneusedsnap, originusedsnap;
2558 };
2559 
2560 static int snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep);
2561 
2562 /* ARGSUSED */
2563 static int
2564 dsl_dataset_promote_check(void *arg1, void *arg2, dmu_tx_t *tx)
2565 {
2566 	dsl_dataset_t *hds = arg1;
2567 	struct promotearg *pa = arg2;
2568 	struct promotenode *snap = list_head(&pa->shared_snaps);
2569 	dsl_dataset_t *origin_ds = snap->ds;
2570 	int err;
2571 
2572 	/* Check that it is a real clone */
2573 	if (!dsl_dir_is_clone(hds->ds_dir))
2574 		return (EINVAL);
2575 
2576 	/* Since this is so expensive, don't do the preliminary check */
2577 	if (!dmu_tx_is_syncing(tx))
2578 		return (0);
2579 
2580 	if (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE)
2581 		return (EXDEV);
2582 
2583 	/* compute origin's new unique space */
2584 	snap = list_tail(&pa->clone_snaps);
2585 	ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2586 	err = bplist_space_birthrange(&snap->ds->ds_deadlist,
2587 	    origin_ds->ds_phys->ds_prev_snap_txg, UINT64_MAX, &pa->unique);
2588 	if (err)
2589 		return (err);
2590 
2591 	/*
2592 	 * Walk the snapshots that we are moving
2593 	 *
2594 	 * Compute space to transfer.  Consider the incremental changes
2595 	 * to used for each snapshot:
2596 	 * (my used) = (prev's used) + (blocks born) - (blocks killed)
2597 	 * So each snapshot gave birth to:
2598 	 * (blocks born) = (my used) - (prev's used) + (blocks killed)
2599 	 * So a sequence would look like:
2600 	 * (uN - u(N-1) + kN) + ... + (u1 - u0 + k1) + (u0 - 0 + k0)
2601 	 * Which simplifies to:
2602 	 * uN + kN + kN-1 + ... + k1 + k0
2603 	 * Note however, if we stop before we reach the ORIGIN we get:
2604 	 * uN + kN + kN-1 + ... + kM - uM-1
2605 	 */
2606 	pa->used = origin_ds->ds_phys->ds_used_bytes;
2607 	pa->comp = origin_ds->ds_phys->ds_compressed_bytes;
2608 	pa->uncomp = origin_ds->ds_phys->ds_uncompressed_bytes;
2609 	for (snap = list_head(&pa->shared_snaps); snap;
2610 	    snap = list_next(&pa->shared_snaps, snap)) {
2611 		uint64_t val, dlused, dlcomp, dluncomp;
2612 		dsl_dataset_t *ds = snap->ds;
2613 
2614 		/* Check that the snapshot name does not conflict */
2615 		VERIFY(0 == dsl_dataset_get_snapname(ds));
2616 		err = dsl_dataset_snap_lookup(hds, ds->ds_snapname, &val);
2617 		if (err == 0)
2618 			return (EEXIST);
2619 		if (err != ENOENT)
2620 			return (err);
2621 
2622 		/* The very first snapshot does not have a deadlist */
2623 		if (ds->ds_phys->ds_prev_snap_obj == 0)
2624 			continue;
2625 
2626 		if (err = bplist_space(&ds->ds_deadlist,
2627 		    &dlused, &dlcomp, &dluncomp))
2628 			return (err);
2629 		pa->used += dlused;
2630 		pa->comp += dlcomp;
2631 		pa->uncomp += dluncomp;
2632 	}
2633 
2634 	/*
2635 	 * If we are a clone of a clone then we never reached ORIGIN,
2636 	 * so we need to subtract out the clone origin's used space.
2637 	 */
2638 	if (pa->origin_origin) {
2639 		pa->used -= pa->origin_origin->ds_phys->ds_used_bytes;
2640 		pa->comp -= pa->origin_origin->ds_phys->ds_compressed_bytes;
2641 		pa->uncomp -= pa->origin_origin->ds_phys->ds_uncompressed_bytes;
2642 	}
2643 
2644 	/* Check that there is enough space here */
2645 	err = dsl_dir_transfer_possible(origin_ds->ds_dir, hds->ds_dir,
2646 	    pa->used);
2647 	if (err)
2648 		return (err);
2649 
2650 	/*
2651 	 * Compute the amounts of space that will be used by snapshots
2652 	 * after the promotion (for both origin and clone).  For each,
2653 	 * it is the amount of space that will be on all of their
2654 	 * deadlists (that was not born before their new origin).
2655 	 */
2656 	if (hds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2657 		uint64_t space;
2658 
2659 		/*
2660 		 * Note, typically this will not be a clone of a clone,
2661 		 * so snap->ds->ds_origin_txg will be < TXG_INITIAL, so
2662 		 * these snaplist_space() -> bplist_space_birthrange()
2663 		 * calls will be fast because they do not have to
2664 		 * iterate over all bps.
2665 		 */
2666 		snap = list_head(&pa->origin_snaps);
2667 		err = snaplist_space(&pa->shared_snaps,
2668 		    snap->ds->ds_origin_txg, &pa->cloneusedsnap);
2669 		if (err)
2670 			return (err);
2671 
2672 		err = snaplist_space(&pa->clone_snaps,
2673 		    snap->ds->ds_origin_txg, &space);
2674 		if (err)
2675 			return (err);
2676 		pa->cloneusedsnap += space;
2677 	}
2678 	if (origin_ds->ds_dir->dd_phys->dd_flags & DD_FLAG_USED_BREAKDOWN) {
2679 		err = snaplist_space(&pa->origin_snaps,
2680 		    origin_ds->ds_phys->ds_creation_txg, &pa->originusedsnap);
2681 		if (err)
2682 			return (err);
2683 	}
2684 
2685 	return (0);
2686 }
2687 
2688 static void
2689 dsl_dataset_promote_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
2690 {
2691 	dsl_dataset_t *hds = arg1;
2692 	struct promotearg *pa = arg2;
2693 	struct promotenode *snap = list_head(&pa->shared_snaps);
2694 	dsl_dataset_t *origin_ds = snap->ds;
2695 	dsl_dataset_t *origin_head;
2696 	dsl_dir_t *dd = hds->ds_dir;
2697 	dsl_pool_t *dp = hds->ds_dir->dd_pool;
2698 	dsl_dir_t *odd = NULL;
2699 	uint64_t oldnext_obj;
2700 	int64_t delta;
2701 
2702 	ASSERT(0 == (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE));
2703 
2704 	snap = list_head(&pa->origin_snaps);
2705 	origin_head = snap->ds;
2706 
2707 	/*
2708 	 * We need to explicitly open odd, since origin_ds's dd will be
2709 	 * changing.
2710 	 */
2711 	VERIFY(0 == dsl_dir_open_obj(dp, origin_ds->ds_dir->dd_object,
2712 	    NULL, FTAG, &odd));
2713 
2714 	/* change origin's next snap */
2715 	dmu_buf_will_dirty(origin_ds->ds_dbuf, tx);
2716 	oldnext_obj = origin_ds->ds_phys->ds_next_snap_obj;
2717 	snap = list_tail(&pa->clone_snaps);
2718 	ASSERT3U(snap->ds->ds_phys->ds_prev_snap_obj, ==, origin_ds->ds_object);
2719 	origin_ds->ds_phys->ds_next_snap_obj = snap->ds->ds_object;
2720 
2721 	/* change the origin's next clone */
2722 	if (origin_ds->ds_phys->ds_next_clones_obj) {
2723 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2724 		    origin_ds->ds_phys->ds_next_clones_obj,
2725 		    origin_ds->ds_phys->ds_next_snap_obj, tx));
2726 		VERIFY3U(0, ==, zap_add_int(dp->dp_meta_objset,
2727 		    origin_ds->ds_phys->ds_next_clones_obj,
2728 		    oldnext_obj, tx));
2729 	}
2730 
2731 	/* change origin */
2732 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
2733 	ASSERT3U(dd->dd_phys->dd_origin_obj, ==, origin_ds->ds_object);
2734 	dd->dd_phys->dd_origin_obj = odd->dd_phys->dd_origin_obj;
2735 	hds->ds_origin_txg = origin_head->ds_origin_txg;
2736 	dmu_buf_will_dirty(odd->dd_dbuf, tx);
2737 	odd->dd_phys->dd_origin_obj = origin_ds->ds_object;
2738 	origin_head->ds_origin_txg = origin_ds->ds_phys->ds_creation_txg;
2739 
2740 	/* move snapshots to this dir */
2741 	for (snap = list_head(&pa->shared_snaps); snap;
2742 	    snap = list_next(&pa->shared_snaps, snap)) {
2743 		dsl_dataset_t *ds = snap->ds;
2744 
2745 		/* unregister props as dsl_dir is changing */
2746 		if (ds->ds_user_ptr) {
2747 			ds->ds_user_evict_func(ds, ds->ds_user_ptr);
2748 			ds->ds_user_ptr = NULL;
2749 		}
2750 		/* move snap name entry */
2751 		VERIFY(0 == dsl_dataset_get_snapname(ds));
2752 		VERIFY(0 == dsl_dataset_snap_remove(origin_head,
2753 		    ds->ds_snapname, tx));
2754 		VERIFY(0 == zap_add(dp->dp_meta_objset,
2755 		    hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
2756 		    8, 1, &ds->ds_object, tx));
2757 		/* change containing dsl_dir */
2758 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
2759 		ASSERT3U(ds->ds_phys->ds_dir_obj, ==, odd->dd_object);
2760 		ds->ds_phys->ds_dir_obj = dd->dd_object;
2761 		ASSERT3P(ds->ds_dir, ==, odd);
2762 		dsl_dir_close(ds->ds_dir, ds);
2763 		VERIFY(0 == dsl_dir_open_obj(dp, dd->dd_object,
2764 		    NULL, ds, &ds->ds_dir));
2765 
2766 		ASSERT3U(dsl_prop_numcb(ds), ==, 0);
2767 	}
2768 
2769 	/*
2770 	 * Change space accounting.
2771 	 * Note, pa->*usedsnap and dd_used_breakdown[SNAP] will either
2772 	 * both be valid, or both be 0 (resulting in delta == 0).  This
2773 	 * is true for each of {clone,origin} independently.
2774 	 */
2775 
2776 	delta = pa->cloneusedsnap -
2777 	    dd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2778 	ASSERT3S(delta, >=, 0);
2779 	ASSERT3U(pa->used, >=, delta);
2780 	dsl_dir_diduse_space(dd, DD_USED_SNAP, delta, 0, 0, tx);
2781 	dsl_dir_diduse_space(dd, DD_USED_HEAD,
2782 	    pa->used - delta, pa->comp, pa->uncomp, tx);
2783 
2784 	delta = pa->originusedsnap -
2785 	    odd->dd_phys->dd_used_breakdown[DD_USED_SNAP];
2786 	ASSERT3S(delta, <=, 0);
2787 	ASSERT3U(pa->used, >=, -delta);
2788 	dsl_dir_diduse_space(odd, DD_USED_SNAP, delta, 0, 0, tx);
2789 	dsl_dir_diduse_space(odd, DD_USED_HEAD,
2790 	    -pa->used - delta, -pa->comp, -pa->uncomp, tx);
2791 
2792 	origin_ds->ds_phys->ds_unique_bytes = pa->unique;
2793 
2794 	/* log history record */
2795 	spa_history_internal_log(LOG_DS_PROMOTE, dd->dd_pool->dp_spa, tx,
2796 	    cr, "dataset = %llu", hds->ds_object);
2797 
2798 	dsl_dir_close(odd, FTAG);
2799 }
2800 
2801 static char *snaplist_tag = "snaplist";
2802 /*
2803  * Make a list of dsl_dataset_t's for the snapshots between first_obj
2804  * (exclusive) and last_obj (inclusive).  The list will be in reverse
2805  * order (last_obj will be the list_head()).  If first_obj == 0, do all
2806  * snapshots back to this dataset's origin.
2807  */
2808 static int
2809 snaplist_make(dsl_pool_t *dp, boolean_t own,
2810     uint64_t first_obj, uint64_t last_obj, list_t *l)
2811 {
2812 	uint64_t obj = last_obj;
2813 
2814 	ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock));
2815 
2816 	list_create(l, sizeof (struct promotenode),
2817 	    offsetof(struct promotenode, link));
2818 
2819 	while (obj != first_obj) {
2820 		dsl_dataset_t *ds;
2821 		struct promotenode *snap;
2822 		int err;
2823 
2824 		if (own) {
2825 			err = dsl_dataset_own_obj(dp, obj,
2826 			    0, snaplist_tag, &ds);
2827 			if (err == 0)
2828 				dsl_dataset_make_exclusive(ds, snaplist_tag);
2829 		} else {
2830 			err = dsl_dataset_hold_obj(dp, obj, snaplist_tag, &ds);
2831 		}
2832 		if (err == ENOENT) {
2833 			/* lost race with snapshot destroy */
2834 			struct promotenode *last = list_tail(l);
2835 			ASSERT(obj != last->ds->ds_phys->ds_prev_snap_obj);
2836 			obj = last->ds->ds_phys->ds_prev_snap_obj;
2837 			continue;
2838 		} else if (err) {
2839 			return (err);
2840 		}
2841 
2842 		if (first_obj == 0)
2843 			first_obj = ds->ds_dir->dd_phys->dd_origin_obj;
2844 
2845 		snap = kmem_alloc(sizeof (struct promotenode), KM_SLEEP);
2846 		snap->ds = ds;
2847 		list_insert_tail(l, snap);
2848 		obj = ds->ds_phys->ds_prev_snap_obj;
2849 	}
2850 
2851 	return (0);
2852 }
2853 
2854 static int
2855 snaplist_space(list_t *l, uint64_t mintxg, uint64_t *spacep)
2856 {
2857 	struct promotenode *snap;
2858 
2859 	*spacep = 0;
2860 	for (snap = list_head(l); snap; snap = list_next(l, snap)) {
2861 		uint64_t used;
2862 		int err = bplist_space_birthrange(&snap->ds->ds_deadlist,
2863 		    mintxg, UINT64_MAX, &used);
2864 		if (err)
2865 			return (err);
2866 		*spacep += used;
2867 	}
2868 	return (0);
2869 }
2870 
2871 static void
2872 snaplist_destroy(list_t *l, boolean_t own)
2873 {
2874 	struct promotenode *snap;
2875 
2876 	if (!l || !list_link_active(&l->list_head))
2877 		return;
2878 
2879 	while ((snap = list_tail(l)) != NULL) {
2880 		list_remove(l, snap);
2881 		if (own)
2882 			dsl_dataset_disown(snap->ds, snaplist_tag);
2883 		else
2884 			dsl_dataset_rele(snap->ds, snaplist_tag);
2885 		kmem_free(snap, sizeof (struct promotenode));
2886 	}
2887 	list_destroy(l);
2888 }
2889 
2890 /*
2891  * Promote a clone.  Nomenclature note:
2892  * "clone" or "cds": the original clone which is being promoted
2893  * "origin" or "ods": the snapshot which is originally clone's origin
2894  * "origin head" or "ohds": the dataset which is the head
2895  * (filesystem/volume) for the origin
2896  * "origin origin": the origin of the origin's filesystem (typically
2897  * NULL, indicating that the clone is not a clone of a clone).
2898  */
2899 int
2900 dsl_dataset_promote(const char *name)
2901 {
2902 	dsl_dataset_t *ds;
2903 	dsl_dir_t *dd;
2904 	dsl_pool_t *dp;
2905 	dmu_object_info_t doi;
2906 	struct promotearg pa = { 0 };
2907 	struct promotenode *snap;
2908 	int err;
2909 
2910 	err = dsl_dataset_hold(name, FTAG, &ds);
2911 	if (err)
2912 		return (err);
2913 	dd = ds->ds_dir;
2914 	dp = dd->dd_pool;
2915 
2916 	err = dmu_object_info(dp->dp_meta_objset,
2917 	    ds->ds_phys->ds_snapnames_zapobj, &doi);
2918 	if (err) {
2919 		dsl_dataset_rele(ds, FTAG);
2920 		return (err);
2921 	}
2922 
2923 	if (dsl_dataset_is_snapshot(ds) || dd->dd_phys->dd_origin_obj == 0) {
2924 		dsl_dataset_rele(ds, FTAG);
2925 		return (EINVAL);
2926 	}
2927 
2928 	/*
2929 	 * We are going to inherit all the snapshots taken before our
2930 	 * origin (i.e., our new origin will be our parent's origin).
2931 	 * Take ownership of them so that we can rename them into our
2932 	 * namespace.
2933 	 */
2934 	rw_enter(&dp->dp_config_rwlock, RW_READER);
2935 
2936 	err = snaplist_make(dp, B_TRUE, 0, dd->dd_phys->dd_origin_obj,
2937 	    &pa.shared_snaps);
2938 	if (err != 0)
2939 		goto out;
2940 
2941 	err = snaplist_make(dp, B_FALSE, 0, ds->ds_object, &pa.clone_snaps);
2942 	if (err != 0)
2943 		goto out;
2944 
2945 	snap = list_head(&pa.shared_snaps);
2946 	ASSERT3U(snap->ds->ds_object, ==, dd->dd_phys->dd_origin_obj);
2947 	err = snaplist_make(dp, B_FALSE, dd->dd_phys->dd_origin_obj,
2948 	    snap->ds->ds_dir->dd_phys->dd_head_dataset_obj, &pa.origin_snaps);
2949 	if (err != 0)
2950 		goto out;
2951 
2952 	if (dsl_dir_is_clone(snap->ds->ds_dir)) {
2953 		err = dsl_dataset_own_obj(dp,
2954 		    snap->ds->ds_dir->dd_phys->dd_origin_obj,
2955 		    0, FTAG, &pa.origin_origin);
2956 		if (err != 0)
2957 			goto out;
2958 	}
2959 
2960 out:
2961 	rw_exit(&dp->dp_config_rwlock);
2962 
2963 	/*
2964 	 * Add in 128x the snapnames zapobj size, since we will be moving
2965 	 * a bunch of snapnames to the promoted ds, and dirtying their
2966 	 * bonus buffers.
2967 	 */
2968 	if (err == 0) {
2969 		err = dsl_sync_task_do(dp, dsl_dataset_promote_check,
2970 		    dsl_dataset_promote_sync, ds, &pa,
2971 		    2 + 2 * doi.doi_physical_blks);
2972 	}
2973 
2974 	snaplist_destroy(&pa.shared_snaps, B_TRUE);
2975 	snaplist_destroy(&pa.clone_snaps, B_FALSE);
2976 	snaplist_destroy(&pa.origin_snaps, B_FALSE);
2977 	if (pa.origin_origin)
2978 		dsl_dataset_disown(pa.origin_origin, FTAG);
2979 	dsl_dataset_rele(ds, FTAG);
2980 	return (err);
2981 }
2982 
2983 struct cloneswaparg {
2984 	dsl_dataset_t *cds; /* clone dataset */
2985 	dsl_dataset_t *ohds; /* origin's head dataset */
2986 	boolean_t force;
2987 	int64_t unused_refres_delta; /* change in unconsumed refreservation */
2988 };
2989 
2990 /* ARGSUSED */
2991 static int
2992 dsl_dataset_clone_swap_check(void *arg1, void *arg2, dmu_tx_t *tx)
2993 {
2994 	struct cloneswaparg *csa = arg1;
2995 
2996 	/* they should both be heads */
2997 	if (dsl_dataset_is_snapshot(csa->cds) ||
2998 	    dsl_dataset_is_snapshot(csa->ohds))
2999 		return (EINVAL);
3000 
3001 	/* the branch point should be just before them */
3002 	if (csa->cds->ds_prev != csa->ohds->ds_prev)
3003 		return (EINVAL);
3004 
3005 	/* cds should be the clone */
3006 	if (csa->cds->ds_prev->ds_phys->ds_next_snap_obj !=
3007 	    csa->ohds->ds_object)
3008 		return (EINVAL);
3009 
3010 	/* the clone should be a child of the origin */
3011 	if (csa->cds->ds_dir->dd_parent != csa->ohds->ds_dir)
3012 		return (EINVAL);
3013 
3014 	/* ohds shouldn't be modified unless 'force' */
3015 	if (!csa->force && dsl_dataset_modified_since_lastsnap(csa->ohds))
3016 		return (ETXTBSY);
3017 
3018 	/* adjust amount of any unconsumed refreservation */
3019 	csa->unused_refres_delta =
3020 	    (int64_t)MIN(csa->ohds->ds_reserved,
3021 	    csa->ohds->ds_phys->ds_unique_bytes) -
3022 	    (int64_t)MIN(csa->ohds->ds_reserved,
3023 	    csa->cds->ds_phys->ds_unique_bytes);
3024 
3025 	if (csa->unused_refres_delta > 0 &&
3026 	    csa->unused_refres_delta >
3027 	    dsl_dir_space_available(csa->ohds->ds_dir, NULL, 0, TRUE))
3028 		return (ENOSPC);
3029 
3030 	return (0);
3031 }
3032 
3033 /* ARGSUSED */
3034 static void
3035 dsl_dataset_clone_swap_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
3036 {
3037 	struct cloneswaparg *csa = arg1;
3038 	dsl_pool_t *dp = csa->cds->ds_dir->dd_pool;
3039 
3040 	ASSERT(csa->cds->ds_reserved == 0);
3041 	ASSERT(csa->cds->ds_quota == csa->ohds->ds_quota);
3042 
3043 	dmu_buf_will_dirty(csa->cds->ds_dbuf, tx);
3044 	dmu_buf_will_dirty(csa->ohds->ds_dbuf, tx);
3045 	dmu_buf_will_dirty(csa->cds->ds_prev->ds_dbuf, tx);
3046 
3047 	if (csa->cds->ds_user_ptr != NULL) {
3048 		csa->cds->ds_user_evict_func(csa->cds, csa->cds->ds_user_ptr);
3049 		csa->cds->ds_user_ptr = NULL;
3050 	}
3051 
3052 	if (csa->ohds->ds_user_ptr != NULL) {
3053 		csa->ohds->ds_user_evict_func(csa->ohds,
3054 		    csa->ohds->ds_user_ptr);
3055 		csa->ohds->ds_user_ptr = NULL;
3056 	}
3057 
3058 	/* reset origin's unique bytes */
3059 	VERIFY(0 == bplist_space_birthrange(&csa->cds->ds_deadlist,
3060 	    csa->cds->ds_prev->ds_phys->ds_prev_snap_txg, UINT64_MAX,
3061 	    &csa->cds->ds_prev->ds_phys->ds_unique_bytes));
3062 
3063 	/* swap blkptrs */
3064 	{
3065 		blkptr_t tmp;
3066 		tmp = csa->ohds->ds_phys->ds_bp;
3067 		csa->ohds->ds_phys->ds_bp = csa->cds->ds_phys->ds_bp;
3068 		csa->cds->ds_phys->ds_bp = tmp;
3069 	}
3070 
3071 	/* set dd_*_bytes */
3072 	{
3073 		int64_t dused, dcomp, duncomp;
3074 		uint64_t cdl_used, cdl_comp, cdl_uncomp;
3075 		uint64_t odl_used, odl_comp, odl_uncomp;
3076 
3077 		ASSERT3U(csa->cds->ds_dir->dd_phys->
3078 		    dd_used_breakdown[DD_USED_SNAP], ==, 0);
3079 
3080 		VERIFY(0 == bplist_space(&csa->cds->ds_deadlist, &cdl_used,
3081 		    &cdl_comp, &cdl_uncomp));
3082 		VERIFY(0 == bplist_space(&csa->ohds->ds_deadlist, &odl_used,
3083 		    &odl_comp, &odl_uncomp));
3084 
3085 		dused = csa->cds->ds_phys->ds_used_bytes + cdl_used -
3086 		    (csa->ohds->ds_phys->ds_used_bytes + odl_used);
3087 		dcomp = csa->cds->ds_phys->ds_compressed_bytes + cdl_comp -
3088 		    (csa->ohds->ds_phys->ds_compressed_bytes + odl_comp);
3089 		duncomp = csa->cds->ds_phys->ds_uncompressed_bytes +
3090 		    cdl_uncomp -
3091 		    (csa->ohds->ds_phys->ds_uncompressed_bytes + odl_uncomp);
3092 
3093 		dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_HEAD,
3094 		    dused, dcomp, duncomp, tx);
3095 		dsl_dir_diduse_space(csa->cds->ds_dir, DD_USED_HEAD,
3096 		    -dused, -dcomp, -duncomp, tx);
3097 
3098 		/*
3099 		 * The difference in the space used by snapshots is the
3100 		 * difference in snapshot space due to the head's
3101 		 * deadlist (since that's the only thing that's
3102 		 * changing that affects the snapused).
3103 		 */
3104 		VERIFY(0 == bplist_space_birthrange(&csa->cds->ds_deadlist,
3105 		    csa->ohds->ds_origin_txg, UINT64_MAX, &cdl_used));
3106 		VERIFY(0 == bplist_space_birthrange(&csa->ohds->ds_deadlist,
3107 		    csa->ohds->ds_origin_txg, UINT64_MAX, &odl_used));
3108 		dsl_dir_transfer_space(csa->ohds->ds_dir, cdl_used - odl_used,
3109 		    DD_USED_HEAD, DD_USED_SNAP, tx);
3110 	}
3111 
3112 #define	SWITCH64(x, y) \
3113 	{ \
3114 		uint64_t __tmp = (x); \
3115 		(x) = (y); \
3116 		(y) = __tmp; \
3117 	}
3118 
3119 	/* swap ds_*_bytes */
3120 	SWITCH64(csa->ohds->ds_phys->ds_used_bytes,
3121 	    csa->cds->ds_phys->ds_used_bytes);
3122 	SWITCH64(csa->ohds->ds_phys->ds_compressed_bytes,
3123 	    csa->cds->ds_phys->ds_compressed_bytes);
3124 	SWITCH64(csa->ohds->ds_phys->ds_uncompressed_bytes,
3125 	    csa->cds->ds_phys->ds_uncompressed_bytes);
3126 	SWITCH64(csa->ohds->ds_phys->ds_unique_bytes,
3127 	    csa->cds->ds_phys->ds_unique_bytes);
3128 
3129 	/* apply any parent delta for change in unconsumed refreservation */
3130 	dsl_dir_diduse_space(csa->ohds->ds_dir, DD_USED_REFRSRV,
3131 	    csa->unused_refres_delta, 0, 0, tx);
3132 
3133 	/* swap deadlists */
3134 	bplist_close(&csa->cds->ds_deadlist);
3135 	bplist_close(&csa->ohds->ds_deadlist);
3136 	SWITCH64(csa->ohds->ds_phys->ds_deadlist_obj,
3137 	    csa->cds->ds_phys->ds_deadlist_obj);
3138 	VERIFY(0 == bplist_open(&csa->cds->ds_deadlist, dp->dp_meta_objset,
3139 	    csa->cds->ds_phys->ds_deadlist_obj));
3140 	VERIFY(0 == bplist_open(&csa->ohds->ds_deadlist, dp->dp_meta_objset,
3141 	    csa->ohds->ds_phys->ds_deadlist_obj));
3142 
3143 	dsl_pool_ds_clone_swapped(csa->ohds, csa->cds, tx);
3144 }
3145 
3146 /*
3147  * Swap 'clone' with its origin head file system.  Used at the end
3148  * of "online recv" to swizzle the file system to the new version.
3149  */
3150 int
3151 dsl_dataset_clone_swap(dsl_dataset_t *clone, dsl_dataset_t *origin_head,
3152     boolean_t force)
3153 {
3154 	struct cloneswaparg csa;
3155 	int error;
3156 
3157 	ASSERT(clone->ds_owner);
3158 	ASSERT(origin_head->ds_owner);
3159 retry:
3160 	/* Need exclusive access for the swap */
3161 	rw_enter(&clone->ds_rwlock, RW_WRITER);
3162 	if (!rw_tryenter(&origin_head->ds_rwlock, RW_WRITER)) {
3163 		rw_exit(&clone->ds_rwlock);
3164 		rw_enter(&origin_head->ds_rwlock, RW_WRITER);
3165 		if (!rw_tryenter(&clone->ds_rwlock, RW_WRITER)) {
3166 			rw_exit(&origin_head->ds_rwlock);
3167 			goto retry;
3168 		}
3169 	}
3170 	csa.cds = clone;
3171 	csa.ohds = origin_head;
3172 	csa.force = force;
3173 	error = dsl_sync_task_do(clone->ds_dir->dd_pool,
3174 	    dsl_dataset_clone_swap_check,
3175 	    dsl_dataset_clone_swap_sync, &csa, NULL, 9);
3176 	return (error);
3177 }
3178 
3179 /*
3180  * Given a pool name and a dataset object number in that pool,
3181  * return the name of that dataset.
3182  */
3183 int
3184 dsl_dsobj_to_dsname(char *pname, uint64_t obj, char *buf)
3185 {
3186 	spa_t *spa;
3187 	dsl_pool_t *dp;
3188 	dsl_dataset_t *ds;
3189 	int error;
3190 
3191 	if ((error = spa_open(pname, &spa, FTAG)) != 0)
3192 		return (error);
3193 	dp = spa_get_dsl(spa);
3194 	rw_enter(&dp->dp_config_rwlock, RW_READER);
3195 	if ((error = dsl_dataset_hold_obj(dp, obj, FTAG, &ds)) == 0) {
3196 		dsl_dataset_name(ds, buf);
3197 		dsl_dataset_rele(ds, FTAG);
3198 	}
3199 	rw_exit(&dp->dp_config_rwlock);
3200 	spa_close(spa, FTAG);
3201 
3202 	return (error);
3203 }
3204 
3205 int
3206 dsl_dataset_check_quota(dsl_dataset_t *ds, boolean_t check_quota,
3207     uint64_t asize, uint64_t inflight, uint64_t *used, uint64_t *ref_rsrv)
3208 {
3209 	int error = 0;
3210 
3211 	ASSERT3S(asize, >, 0);
3212 
3213 	/*
3214 	 * *ref_rsrv is the portion of asize that will come from any
3215 	 * unconsumed refreservation space.
3216 	 */
3217 	*ref_rsrv = 0;
3218 
3219 	mutex_enter(&ds->ds_lock);
3220 	/*
3221 	 * Make a space adjustment for reserved bytes.
3222 	 */
3223 	if (ds->ds_reserved > ds->ds_phys->ds_unique_bytes) {
3224 		ASSERT3U(*used, >=,
3225 		    ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3226 		*used -= (ds->ds_reserved - ds->ds_phys->ds_unique_bytes);
3227 		*ref_rsrv =
3228 		    asize - MIN(asize, parent_delta(ds, asize + inflight));
3229 	}
3230 
3231 	if (!check_quota || ds->ds_quota == 0) {
3232 		mutex_exit(&ds->ds_lock);
3233 		return (0);
3234 	}
3235 	/*
3236 	 * If they are requesting more space, and our current estimate
3237 	 * is over quota, they get to try again unless the actual
3238 	 * on-disk is over quota and there are no pending changes (which
3239 	 * may free up space for us).
3240 	 */
3241 	if (ds->ds_phys->ds_used_bytes + inflight >= ds->ds_quota) {
3242 		if (inflight > 0 || ds->ds_phys->ds_used_bytes < ds->ds_quota)
3243 			error = ERESTART;
3244 		else
3245 			error = EDQUOT;
3246 	}
3247 	mutex_exit(&ds->ds_lock);
3248 
3249 	return (error);
3250 }
3251 
3252 /* ARGSUSED */
3253 static int
3254 dsl_dataset_set_quota_check(void *arg1, void *arg2, dmu_tx_t *tx)
3255 {
3256 	dsl_dataset_t *ds = arg1;
3257 	uint64_t *quotap = arg2;
3258 	uint64_t new_quota = *quotap;
3259 
3260 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_REFQUOTA)
3261 		return (ENOTSUP);
3262 
3263 	if (new_quota == 0)
3264 		return (0);
3265 
3266 	if (new_quota < ds->ds_phys->ds_used_bytes ||
3267 	    new_quota < ds->ds_reserved)
3268 		return (ENOSPC);
3269 
3270 	return (0);
3271 }
3272 
3273 /* ARGSUSED */
3274 void
3275 dsl_dataset_set_quota_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
3276 {
3277 	dsl_dataset_t *ds = arg1;
3278 	uint64_t *quotap = arg2;
3279 	uint64_t new_quota = *quotap;
3280 
3281 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
3282 
3283 	ds->ds_quota = new_quota;
3284 
3285 	dsl_dir_prop_set_uint64_sync(ds->ds_dir, "refquota", new_quota, cr, tx);
3286 
3287 	spa_history_internal_log(LOG_DS_REFQUOTA, ds->ds_dir->dd_pool->dp_spa,
3288 	    tx, cr, "%lld dataset = %llu ",
3289 	    (longlong_t)new_quota, ds->ds_object);
3290 }
3291 
3292 int
3293 dsl_dataset_set_quota(const char *dsname, uint64_t quota)
3294 {
3295 	dsl_dataset_t *ds;
3296 	int err;
3297 
3298 	err = dsl_dataset_hold(dsname, FTAG, &ds);
3299 	if (err)
3300 		return (err);
3301 
3302 	if (quota != ds->ds_quota) {
3303 		/*
3304 		 * If someone removes a file, then tries to set the quota, we
3305 		 * want to make sure the file freeing takes effect.
3306 		 */
3307 		txg_wait_open(ds->ds_dir->dd_pool, 0);
3308 
3309 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3310 		    dsl_dataset_set_quota_check, dsl_dataset_set_quota_sync,
3311 		    ds, &quota, 0);
3312 	}
3313 	dsl_dataset_rele(ds, FTAG);
3314 	return (err);
3315 }
3316 
3317 static int
3318 dsl_dataset_set_reservation_check(void *arg1, void *arg2, dmu_tx_t *tx)
3319 {
3320 	dsl_dataset_t *ds = arg1;
3321 	uint64_t *reservationp = arg2;
3322 	uint64_t new_reservation = *reservationp;
3323 	uint64_t unique;
3324 
3325 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) <
3326 	    SPA_VERSION_REFRESERVATION)
3327 		return (ENOTSUP);
3328 
3329 	if (dsl_dataset_is_snapshot(ds))
3330 		return (EINVAL);
3331 
3332 	/*
3333 	 * If we are doing the preliminary check in open context, the
3334 	 * space estimates may be inaccurate.
3335 	 */
3336 	if (!dmu_tx_is_syncing(tx))
3337 		return (0);
3338 
3339 	mutex_enter(&ds->ds_lock);
3340 	unique = dsl_dataset_unique(ds);
3341 	mutex_exit(&ds->ds_lock);
3342 
3343 	if (MAX(unique, new_reservation) > MAX(unique, ds->ds_reserved)) {
3344 		uint64_t delta = MAX(unique, new_reservation) -
3345 		    MAX(unique, ds->ds_reserved);
3346 
3347 		if (delta > dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE))
3348 			return (ENOSPC);
3349 		if (ds->ds_quota > 0 &&
3350 		    new_reservation > ds->ds_quota)
3351 			return (ENOSPC);
3352 	}
3353 
3354 	return (0);
3355 }
3356 
3357 /* ARGSUSED */
3358 static void
3359 dsl_dataset_set_reservation_sync(void *arg1, void *arg2, cred_t *cr,
3360     dmu_tx_t *tx)
3361 {
3362 	dsl_dataset_t *ds = arg1;
3363 	uint64_t *reservationp = arg2;
3364 	uint64_t new_reservation = *reservationp;
3365 	uint64_t unique;
3366 	int64_t delta;
3367 
3368 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
3369 
3370 	mutex_enter(&ds->ds_dir->dd_lock);
3371 	mutex_enter(&ds->ds_lock);
3372 	unique = dsl_dataset_unique(ds);
3373 	delta = MAX(0, (int64_t)(new_reservation - unique)) -
3374 	    MAX(0, (int64_t)(ds->ds_reserved - unique));
3375 	ds->ds_reserved = new_reservation;
3376 	mutex_exit(&ds->ds_lock);
3377 
3378 	dsl_dir_diduse_space(ds->ds_dir, DD_USED_REFRSRV, delta, 0, 0, tx);
3379 	mutex_exit(&ds->ds_dir->dd_lock);
3380 	dsl_dir_prop_set_uint64_sync(ds->ds_dir, "refreservation",
3381 	    new_reservation, cr, tx);
3382 
3383 	spa_history_internal_log(LOG_DS_REFRESERV,
3384 	    ds->ds_dir->dd_pool->dp_spa, tx, cr, "%lld dataset = %llu",
3385 	    (longlong_t)new_reservation, ds->ds_object);
3386 }
3387 
3388 int
3389 dsl_dataset_set_reservation(const char *dsname, uint64_t reservation)
3390 {
3391 	dsl_dataset_t *ds;
3392 	int err;
3393 
3394 	err = dsl_dataset_hold(dsname, FTAG, &ds);
3395 	if (err)
3396 		return (err);
3397 
3398 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
3399 	    dsl_dataset_set_reservation_check,
3400 	    dsl_dataset_set_reservation_sync, ds, &reservation, 0);
3401 	dsl_dataset_rele(ds, FTAG);
3402 	return (err);
3403 }
3404 
3405 static int
3406 dsl_dataset_user_hold_check(void *arg1, void *arg2, dmu_tx_t *tx)
3407 {
3408 	dsl_dataset_t *ds = arg1;
3409 	char *htag = arg2;
3410 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3411 	int error = 0;
3412 
3413 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3414 		return (ENOTSUP);
3415 
3416 	if (!dsl_dataset_is_snapshot(ds))
3417 		return (EINVAL);
3418 
3419 	if (strlen(htag) >= ZAP_MAXNAMELEN)
3420 		return (ENAMETOOLONG);
3421 
3422 	/* tags must be unique */
3423 	mutex_enter(&ds->ds_lock);
3424 	if (ds->ds_phys->ds_userrefs_obj) {
3425 		error = zap_lookup(mos, ds->ds_phys->ds_userrefs_obj, htag,
3426 		    8, 1, tx);
3427 		if (error == 0)
3428 			error = EEXIST;
3429 		else if (error == ENOENT)
3430 			error = 0;
3431 	}
3432 	mutex_exit(&ds->ds_lock);
3433 
3434 	return (error);
3435 }
3436 
3437 static void
3438 dsl_dataset_user_hold_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
3439 {
3440 	dsl_dataset_t *ds = arg1;
3441 	char *htag = arg2;
3442 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3443 	time_t now = gethrestime_sec();
3444 	uint64_t zapobj;
3445 
3446 	mutex_enter(&ds->ds_lock);
3447 	if (ds->ds_phys->ds_userrefs_obj == 0) {
3448 		/*
3449 		 * This is the first user hold for this dataset.  Create
3450 		 * the userrefs zap object.
3451 		 */
3452 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
3453 		zapobj = ds->ds_phys->ds_userrefs_obj =
3454 		    zap_create(mos, DMU_OT_USERREFS, DMU_OT_NONE, 0, tx);
3455 	} else {
3456 		zapobj = ds->ds_phys->ds_userrefs_obj;
3457 	}
3458 	ds->ds_userrefs++;
3459 	mutex_exit(&ds->ds_lock);
3460 
3461 	VERIFY(0 == zap_add(mos, zapobj, htag, 8, 1, &now, tx));
3462 
3463 	spa_history_internal_log(LOG_DS_USER_HOLD,
3464 	    ds->ds_dir->dd_pool->dp_spa, tx, cr, "<%s> dataset = %llu",
3465 	    htag, ds->ds_object);
3466 }
3467 
3468 struct dsl_ds_holdarg {
3469 	dsl_sync_task_group_t *dstg;
3470 	char *htag;
3471 	char *snapname;
3472 	boolean_t recursive;
3473 	char failed[MAXPATHLEN];
3474 };
3475 
3476 static int
3477 dsl_dataset_user_hold_one(char *dsname, void *arg)
3478 {
3479 	struct dsl_ds_holdarg *ha = arg;
3480 	dsl_dataset_t *ds;
3481 	int error;
3482 	char *name;
3483 	size_t buflen;
3484 
3485 	/* alloc a buffer to hold dsname@snapname plus terminating NULL */
3486 	buflen = strlen(dsname) + strlen(ha->snapname) + 2;
3487 	name = kmem_alloc(buflen, KM_SLEEP);
3488 	(void) snprintf(name, buflen, "%s@%s", dsname, ha->snapname);
3489 	error = dsl_dataset_hold(name, ha->dstg, &ds);
3490 	kmem_free(name, buflen);
3491 	if (error == 0) {
3492 		dsl_sync_task_create(ha->dstg, dsl_dataset_user_hold_check,
3493 		    dsl_dataset_user_hold_sync, ds, ha->htag, 0);
3494 	} else if (error == ENOENT && ha->recursive) {
3495 		error = 0;
3496 	} else {
3497 		(void) strcpy(ha->failed, dsname);
3498 	}
3499 	return (error);
3500 }
3501 
3502 int
3503 dsl_dataset_user_hold(char *dsname, char *snapname, char *htag,
3504     boolean_t recursive)
3505 {
3506 	struct dsl_ds_holdarg *ha;
3507 	dsl_sync_task_t *dst;
3508 	spa_t *spa;
3509 	int error;
3510 
3511 	ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3512 
3513 	(void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3514 
3515 	error = spa_open(dsname, &spa, FTAG);
3516 	if (error) {
3517 		kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3518 		return (error);
3519 	}
3520 
3521 	ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3522 	ha->htag = htag;
3523 	ha->snapname = snapname;
3524 	ha->recursive = recursive;
3525 	if (recursive) {
3526 		error = dmu_objset_find(dsname, dsl_dataset_user_hold_one,
3527 		    ha, DS_FIND_CHILDREN);
3528 	} else {
3529 		error = dsl_dataset_user_hold_one(dsname, ha);
3530 	}
3531 	if (error == 0)
3532 		error = dsl_sync_task_group_wait(ha->dstg);
3533 
3534 	for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3535 	    dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3536 		dsl_dataset_t *ds = dst->dst_arg1;
3537 
3538 		if (dst->dst_err) {
3539 			dsl_dataset_name(ds, ha->failed);
3540 			*strchr(ha->failed, '@') = '\0';
3541 		}
3542 		dsl_dataset_rele(ds, ha->dstg);
3543 	}
3544 
3545 	if (error)
3546 		(void) strcpy(dsname, ha->failed);
3547 
3548 	dsl_sync_task_group_destroy(ha->dstg);
3549 	kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3550 	spa_close(spa, FTAG);
3551 	return (error);
3552 }
3553 
3554 struct dsl_ds_releasearg {
3555 	dsl_dataset_t *ds;
3556 	const char *htag;
3557 	boolean_t own;		/* do we own or just hold ds? */
3558 };
3559 
3560 static int
3561 dsl_dataset_release_might_destroy(dsl_dataset_t *ds, const char *htag,
3562     boolean_t *might_destroy)
3563 {
3564 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3565 	uint64_t zapobj;
3566 	uint64_t tmp;
3567 	int error;
3568 
3569 	*might_destroy = B_FALSE;
3570 
3571 	mutex_enter(&ds->ds_lock);
3572 	zapobj = ds->ds_phys->ds_userrefs_obj;
3573 	if (zapobj == 0) {
3574 		/* The tag can't possibly exist */
3575 		mutex_exit(&ds->ds_lock);
3576 		return (ESRCH);
3577 	}
3578 
3579 	/* Make sure the tag exists */
3580 	error = zap_lookup(mos, zapobj, htag, 8, 1, &tmp);
3581 	if (error) {
3582 		mutex_exit(&ds->ds_lock);
3583 		if (error == ENOENT)
3584 			error = ESRCH;
3585 		return (error);
3586 	}
3587 
3588 	if (ds->ds_userrefs == 1 && ds->ds_phys->ds_num_children == 1 &&
3589 	    DS_IS_DEFER_DESTROY(ds))
3590 		*might_destroy = B_TRUE;
3591 
3592 	mutex_exit(&ds->ds_lock);
3593 	return (0);
3594 }
3595 
3596 static int
3597 dsl_dataset_user_release_check(void *arg1, void *tag, dmu_tx_t *tx)
3598 {
3599 	struct dsl_ds_releasearg *ra = arg1;
3600 	dsl_dataset_t *ds = ra->ds;
3601 	boolean_t might_destroy;
3602 	int error;
3603 
3604 	if (spa_version(ds->ds_dir->dd_pool->dp_spa) < SPA_VERSION_USERREFS)
3605 		return (ENOTSUP);
3606 
3607 	error = dsl_dataset_release_might_destroy(ds, ra->htag, &might_destroy);
3608 	if (error)
3609 		return (error);
3610 
3611 	if (might_destroy) {
3612 		struct dsl_ds_destroyarg dsda = {0};
3613 
3614 		if (dmu_tx_is_syncing(tx)) {
3615 			/*
3616 			 * If we're not prepared to remove the snapshot,
3617 			 * we can't allow the release to happen right now.
3618 			 */
3619 			if (!ra->own)
3620 				return (EBUSY);
3621 			if (ds->ds_user_ptr) {
3622 				ds->ds_user_evict_func(ds, ds->ds_user_ptr);
3623 				ds->ds_user_ptr = NULL;
3624 			}
3625 		}
3626 		dsda.ds = ds;
3627 		dsda.releasing = B_TRUE;
3628 		return (dsl_dataset_destroy_check(&dsda, tag, tx));
3629 	}
3630 
3631 	return (0);
3632 }
3633 
3634 static void
3635 dsl_dataset_user_release_sync(void *arg1, void *tag, cred_t *cr, dmu_tx_t *tx)
3636 {
3637 	struct dsl_ds_releasearg *ra = arg1;
3638 	dsl_dataset_t *ds = ra->ds;
3639 	spa_t *spa = ds->ds_dir->dd_pool->dp_spa;
3640 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
3641 	uint64_t zapobj;
3642 	uint64_t dsobj = ds->ds_object;
3643 	uint64_t refs;
3644 
3645 	mutex_enter(&ds->ds_lock);
3646 	ds->ds_userrefs--;
3647 	refs = ds->ds_userrefs;
3648 	mutex_exit(&ds->ds_lock);
3649 	zapobj = ds->ds_phys->ds_userrefs_obj;
3650 	VERIFY(0 == zap_remove(mos, zapobj, ra->htag, tx));
3651 	if (ds->ds_userrefs == 0 && ds->ds_phys->ds_num_children == 1 &&
3652 	    DS_IS_DEFER_DESTROY(ds)) {
3653 		struct dsl_ds_destroyarg dsda = {0};
3654 
3655 		ASSERT(ra->own);
3656 		dsda.ds = ds;
3657 		dsda.releasing = B_TRUE;
3658 		/* We already did the destroy_check */
3659 		dsl_dataset_destroy_sync(&dsda, tag, cr, tx);
3660 	}
3661 
3662 	spa_history_internal_log(LOG_DS_USER_RELEASE,
3663 	    spa, tx, cr, "<%s> %lld dataset = %llu",
3664 	    ra->htag, (longlong_t)refs, dsobj);
3665 }
3666 
3667 static int
3668 dsl_dataset_user_release_one(char *dsname, void *arg)
3669 {
3670 	struct dsl_ds_holdarg *ha = arg;
3671 	struct dsl_ds_releasearg *ra;
3672 	dsl_dataset_t *ds;
3673 	int error;
3674 	void *dtag = ha->dstg;
3675 	char *name;
3676 	size_t buflen;
3677 	boolean_t own = B_FALSE;
3678 	boolean_t might_destroy;
3679 
3680 	if (strlen(ha->htag) >= ZAP_MAXNAMELEN)
3681 		return (ENAMETOOLONG);
3682 
3683 	/* alloc a buffer to hold dsname@snapname, plus the terminating NULL */
3684 	buflen = strlen(dsname) + strlen(ha->snapname) + 2;
3685 	name = kmem_alloc(buflen, KM_SLEEP);
3686 	(void) snprintf(name, buflen, "%s@%s", dsname, ha->snapname);
3687 	error = dsl_dataset_hold(name, dtag, &ds);
3688 	kmem_free(name, buflen);
3689 	if (error == ENOENT && ha->recursive)
3690 		return (0);
3691 	(void) strcpy(ha->failed, dsname);
3692 	if (error)
3693 		return (error);
3694 
3695 	ASSERT(dsl_dataset_is_snapshot(ds));
3696 
3697 	error = dsl_dataset_release_might_destroy(ds, ha->htag, &might_destroy);
3698 	if (error) {
3699 		dsl_dataset_rele(ds, dtag);
3700 		return (error);
3701 	}
3702 
3703 	if (might_destroy) {
3704 #ifdef _KERNEL
3705 		error = zfs_unmount_snap(name, NULL);
3706 		if (error) {
3707 			dsl_dataset_rele(ds, dtag);
3708 			return (error);
3709 		}
3710 		error = dsl_dataset_zvol_cleanup(ds, name);
3711 		if (error) {
3712 			dsl_dataset_rele(ds, dtag);
3713 			return (error);
3714 		}
3715 #endif
3716 		if (!dsl_dataset_tryown(ds,
3717 		    DS_MODE_READONLY | DS_MODE_INCONSISTENT, dtag)) {
3718 			dsl_dataset_rele(ds, dtag);
3719 			return (EBUSY);
3720 		} else {
3721 			own = B_TRUE;
3722 			dsl_dataset_make_exclusive(ds, dtag);
3723 		}
3724 	}
3725 
3726 	ra = kmem_alloc(sizeof (struct dsl_ds_releasearg), KM_SLEEP);
3727 	ra->ds = ds;
3728 	ra->htag = ha->htag;
3729 	ra->own = own;
3730 	dsl_sync_task_create(ha->dstg, dsl_dataset_user_release_check,
3731 	    dsl_dataset_user_release_sync, ra, dtag, 0);
3732 
3733 	return (0);
3734 }
3735 
3736 int
3737 dsl_dataset_user_release(char *dsname, char *snapname, char *htag,
3738     boolean_t recursive)
3739 {
3740 	struct dsl_ds_holdarg *ha;
3741 	dsl_sync_task_t *dst;
3742 	spa_t *spa;
3743 	int error;
3744 
3745 	ha = kmem_zalloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
3746 
3747 	(void) strlcpy(ha->failed, dsname, sizeof (ha->failed));
3748 
3749 	error = spa_open(dsname, &spa, FTAG);
3750 	if (error) {
3751 		kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3752 		return (error);
3753 	}
3754 
3755 	ha->dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
3756 	ha->htag = htag;
3757 	ha->snapname = snapname;
3758 	ha->recursive = recursive;
3759 	if (recursive) {
3760 		error = dmu_objset_find(dsname, dsl_dataset_user_release_one,
3761 		    ha, DS_FIND_CHILDREN);
3762 	} else {
3763 		error = dsl_dataset_user_release_one(dsname, ha);
3764 	}
3765 	if (error == 0)
3766 		error = dsl_sync_task_group_wait(ha->dstg);
3767 
3768 	for (dst = list_head(&ha->dstg->dstg_tasks); dst;
3769 	    dst = list_next(&ha->dstg->dstg_tasks, dst)) {
3770 		struct dsl_ds_releasearg *ra = dst->dst_arg1;
3771 		dsl_dataset_t *ds = ra->ds;
3772 
3773 		if (dst->dst_err)
3774 			dsl_dataset_name(ds, ha->failed);
3775 
3776 		if (ra->own)
3777 			dsl_dataset_disown(ds, ha->dstg);
3778 		else
3779 			dsl_dataset_rele(ds, ha->dstg);
3780 
3781 		kmem_free(ra, sizeof (struct dsl_ds_releasearg));
3782 	}
3783 
3784 	if (error)
3785 		(void) strcpy(dsname, ha->failed);
3786 
3787 	dsl_sync_task_group_destroy(ha->dstg);
3788 	kmem_free(ha, sizeof (struct dsl_ds_holdarg));
3789 	spa_close(spa, FTAG);
3790 	return (error);
3791 }
3792 
3793 int
3794 dsl_dataset_get_holds(const char *dsname, nvlist_t **nvp)
3795 {
3796 	dsl_dataset_t *ds;
3797 	int err;
3798 
3799 	err = dsl_dataset_hold(dsname, FTAG, &ds);
3800 	if (err)
3801 		return (err);
3802 
3803 	VERIFY(0 == nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP));
3804 	if (ds->ds_phys->ds_userrefs_obj != 0) {
3805 		zap_attribute_t *za;
3806 		zap_cursor_t zc;
3807 
3808 		za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
3809 		for (zap_cursor_init(&zc, ds->ds_dir->dd_pool->dp_meta_objset,
3810 		    ds->ds_phys->ds_userrefs_obj);
3811 		    zap_cursor_retrieve(&zc, za) == 0;
3812 		    zap_cursor_advance(&zc)) {
3813 			VERIFY(0 == nvlist_add_uint64(*nvp, za->za_name,
3814 			    za->za_first_integer));
3815 		}
3816 		zap_cursor_fini(&zc);
3817 		kmem_free(za, sizeof (zap_attribute_t));
3818 	}
3819 	dsl_dataset_rele(ds, FTAG);
3820 	return (0);
3821 }
3822