xref: /titanic_41/usr/src/uts/common/fs/zfs/dsl_dataset.c (revision cc52cd5b16ec3b8334f7db84aa4d148f1e4391d3)
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 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/dmu_objset.h>
29 #include <sys/dsl_dataset.h>
30 #include <sys/dsl_dir.h>
31 #include <sys/dsl_prop.h>
32 #include <sys/dsl_synctask.h>
33 #include <sys/dmu_traverse.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/arc.h>
36 #include <sys/zio.h>
37 #include <sys/zap.h>
38 #include <sys/unique.h>
39 #include <sys/zfs_context.h>
40 
41 static dsl_checkfunc_t dsl_dataset_destroy_begin_check;
42 static dsl_syncfunc_t dsl_dataset_destroy_begin_sync;
43 static dsl_checkfunc_t dsl_dataset_rollback_check;
44 static dsl_syncfunc_t dsl_dataset_rollback_sync;
45 static dsl_checkfunc_t dsl_dataset_destroy_check;
46 static dsl_syncfunc_t dsl_dataset_destroy_sync;
47 
48 #define	DS_REF_MAX	(1ULL << 62)
49 
50 #define	DSL_DEADLIST_BLOCKSIZE	SPA_MAXBLOCKSIZE
51 
52 /*
53  * We use weighted reference counts to express the various forms of exclusion
54  * between different open modes.  A STANDARD open is 1 point, an EXCLUSIVE open
55  * is DS_REF_MAX, and a PRIMARY open is little more than half of an EXCLUSIVE.
56  * This makes the exclusion logic simple: the total refcnt for all opens cannot
57  * exceed DS_REF_MAX.  For example, EXCLUSIVE opens are exclusive because their
58  * weight (DS_REF_MAX) consumes the entire refcnt space.  PRIMARY opens consume
59  * just over half of the refcnt space, so there can't be more than one, but it
60  * can peacefully coexist with any number of STANDARD opens.
61  */
62 static uint64_t ds_refcnt_weight[DS_MODE_LEVELS] = {
63 	0,			/* DS_MODE_NONE - invalid		*/
64 	1,			/* DS_MODE_STANDARD - unlimited number	*/
65 	(DS_REF_MAX >> 1) + 1,	/* DS_MODE_PRIMARY - only one of these	*/
66 	DS_REF_MAX		/* DS_MODE_EXCLUSIVE - no other opens	*/
67 };
68 
69 
70 void
71 dsl_dataset_block_born(dsl_dataset_t *ds, blkptr_t *bp, dmu_tx_t *tx)
72 {
73 	int used = bp_get_dasize(tx->tx_pool->dp_spa, bp);
74 	int compressed = BP_GET_PSIZE(bp);
75 	int uncompressed = BP_GET_UCSIZE(bp);
76 
77 	dprintf_bp(bp, "born, ds=%p\n", ds);
78 
79 	ASSERT(dmu_tx_is_syncing(tx));
80 	/* It could have been compressed away to nothing */
81 	if (BP_IS_HOLE(bp))
82 		return;
83 	ASSERT(BP_GET_TYPE(bp) != DMU_OT_NONE);
84 	ASSERT3U(BP_GET_TYPE(bp), <, DMU_OT_NUMTYPES);
85 	if (ds == NULL) {
86 		/*
87 		 * Account for the meta-objset space in its placeholder
88 		 * dsl_dir.
89 		 */
90 		ASSERT3U(compressed, ==, uncompressed); /* it's all metadata */
91 		dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir,
92 		    used, compressed, uncompressed, tx);
93 		dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
94 		return;
95 	}
96 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
97 	mutex_enter(&ds->ds_lock);
98 	ds->ds_phys->ds_used_bytes += used;
99 	ds->ds_phys->ds_compressed_bytes += compressed;
100 	ds->ds_phys->ds_uncompressed_bytes += uncompressed;
101 	ds->ds_phys->ds_unique_bytes += used;
102 	mutex_exit(&ds->ds_lock);
103 	dsl_dir_diduse_space(ds->ds_dir,
104 	    used, compressed, uncompressed, tx);
105 }
106 
107 void
108 dsl_dataset_block_kill(dsl_dataset_t *ds, blkptr_t *bp, dmu_tx_t *tx)
109 {
110 	int used = bp_get_dasize(tx->tx_pool->dp_spa, bp);
111 	int compressed = BP_GET_PSIZE(bp);
112 	int uncompressed = BP_GET_UCSIZE(bp);
113 
114 	ASSERT(dmu_tx_is_syncing(tx));
115 	if (BP_IS_HOLE(bp))
116 		return;
117 
118 	ASSERT(used > 0);
119 	if (ds == NULL) {
120 		/*
121 		 * Account for the meta-objset space in its placeholder
122 		 * dataset.
123 		 */
124 		/* XXX this can fail, what do we do when it does? */
125 		(void) arc_free(NULL, tx->tx_pool->dp_spa,
126 		    tx->tx_txg, bp, NULL, NULL, ARC_WAIT);
127 		bzero(bp, sizeof (blkptr_t));
128 
129 		dsl_dir_diduse_space(tx->tx_pool->dp_mos_dir,
130 		    -used, -compressed, -uncompressed, tx);
131 		dsl_dir_dirty(tx->tx_pool->dp_mos_dir, tx);
132 		return;
133 	}
134 	ASSERT3P(tx->tx_pool, ==, ds->ds_dir->dd_pool);
135 
136 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
137 
138 	if (bp->blk_birth > ds->ds_phys->ds_prev_snap_txg) {
139 		dprintf_bp(bp, "freeing: %s", "");
140 		/* XXX check return code? */
141 		(void) arc_free(NULL, tx->tx_pool->dp_spa,
142 		    tx->tx_txg, bp, NULL, NULL, ARC_WAIT);
143 
144 		mutex_enter(&ds->ds_lock);
145 		/* XXX unique_bytes is not accurate for head datasets */
146 		/* ASSERT3U(ds->ds_phys->ds_unique_bytes, >=, used); */
147 		ds->ds_phys->ds_unique_bytes -= used;
148 		mutex_exit(&ds->ds_lock);
149 		dsl_dir_diduse_space(ds->ds_dir,
150 		    -used, -compressed, -uncompressed, tx);
151 	} else {
152 		dprintf_bp(bp, "putting on dead list: %s", "");
153 		VERIFY(0 == bplist_enqueue(&ds->ds_deadlist, bp, tx));
154 		/* if (bp->blk_birth > prev prev snap txg) prev unique += bs */
155 		if (ds->ds_phys->ds_prev_snap_obj != 0) {
156 			ASSERT3U(ds->ds_prev->ds_object, ==,
157 			    ds->ds_phys->ds_prev_snap_obj);
158 			ASSERT(ds->ds_prev->ds_phys->ds_num_children > 0);
159 			if (ds->ds_prev->ds_phys->ds_next_snap_obj ==
160 			    ds->ds_object && bp->blk_birth >
161 			    ds->ds_prev->ds_phys->ds_prev_snap_txg) {
162 				dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
163 				mutex_enter(&ds->ds_prev->ds_lock);
164 				ds->ds_prev->ds_phys->ds_unique_bytes +=
165 				    used;
166 				mutex_exit(&ds->ds_prev->ds_lock);
167 			}
168 		}
169 	}
170 	bzero(bp, sizeof (blkptr_t));
171 	mutex_enter(&ds->ds_lock);
172 	ASSERT3U(ds->ds_phys->ds_used_bytes, >=, used);
173 	ds->ds_phys->ds_used_bytes -= used;
174 	ASSERT3U(ds->ds_phys->ds_compressed_bytes, >=, compressed);
175 	ds->ds_phys->ds_compressed_bytes -= compressed;
176 	ASSERT3U(ds->ds_phys->ds_uncompressed_bytes, >=, uncompressed);
177 	ds->ds_phys->ds_uncompressed_bytes -= uncompressed;
178 	mutex_exit(&ds->ds_lock);
179 }
180 
181 uint64_t
182 dsl_dataset_prev_snap_txg(dsl_dataset_t *ds)
183 {
184 	uint64_t trysnap = 0;
185 
186 	if (ds == NULL)
187 		return (0);
188 	/*
189 	 * The snapshot creation could fail, but that would cause an
190 	 * incorrect FALSE return, which would only result in an
191 	 * overestimation of the amount of space that an operation would
192 	 * consume, which is OK.
193 	 *
194 	 * There's also a small window where we could miss a pending
195 	 * snapshot, because we could set the sync task in the quiescing
196 	 * phase.  So this should only be used as a guess.
197 	 */
198 	if (ds->ds_trysnap_txg >
199 	    spa_last_synced_txg(ds->ds_dir->dd_pool->dp_spa))
200 		trysnap = ds->ds_trysnap_txg;
201 	return (MAX(ds->ds_phys->ds_prev_snap_txg, trysnap));
202 }
203 
204 int
205 dsl_dataset_block_freeable(dsl_dataset_t *ds, uint64_t blk_birth)
206 {
207 	return (blk_birth > dsl_dataset_prev_snap_txg(ds));
208 }
209 
210 /* ARGSUSED */
211 static void
212 dsl_dataset_evict(dmu_buf_t *db, void *dsv)
213 {
214 	dsl_dataset_t *ds = dsv;
215 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
216 
217 	/* open_refcount == DS_REF_MAX when deleting */
218 	ASSERT(ds->ds_open_refcount == 0 ||
219 	    ds->ds_open_refcount == DS_REF_MAX);
220 
221 	dprintf_ds(ds, "evicting %s\n", "");
222 
223 	unique_remove(ds->ds_phys->ds_fsid_guid);
224 
225 	if (ds->ds_user_ptr != NULL)
226 		ds->ds_user_evict_func(ds, ds->ds_user_ptr);
227 
228 	if (ds->ds_prev) {
229 		dsl_dataset_close(ds->ds_prev, DS_MODE_NONE, ds);
230 		ds->ds_prev = NULL;
231 	}
232 
233 	bplist_close(&ds->ds_deadlist);
234 	dsl_dir_close(ds->ds_dir, ds);
235 
236 	if (list_link_active(&ds->ds_synced_link))
237 		list_remove(&dp->dp_synced_objsets, ds);
238 
239 	mutex_destroy(&ds->ds_lock);
240 	mutex_destroy(&ds->ds_deadlist.bpl_lock);
241 
242 	kmem_free(ds, sizeof (dsl_dataset_t));
243 }
244 
245 static int
246 dsl_dataset_get_snapname(dsl_dataset_t *ds)
247 {
248 	dsl_dataset_phys_t *headphys;
249 	int err;
250 	dmu_buf_t *headdbuf;
251 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
252 	objset_t *mos = dp->dp_meta_objset;
253 
254 	if (ds->ds_snapname[0])
255 		return (0);
256 	if (ds->ds_phys->ds_next_snap_obj == 0)
257 		return (0);
258 
259 	err = dmu_bonus_hold(mos, ds->ds_dir->dd_phys->dd_head_dataset_obj,
260 	    FTAG, &headdbuf);
261 	if (err)
262 		return (err);
263 	headphys = headdbuf->db_data;
264 	err = zap_value_search(dp->dp_meta_objset,
265 	    headphys->ds_snapnames_zapobj, ds->ds_object, ds->ds_snapname);
266 	dmu_buf_rele(headdbuf, FTAG);
267 	return (err);
268 }
269 
270 int
271 dsl_dataset_open_obj(dsl_pool_t *dp, uint64_t dsobj, const char *snapname,
272     int mode, void *tag, dsl_dataset_t **dsp)
273 {
274 	uint64_t weight = ds_refcnt_weight[DS_MODE_LEVEL(mode)];
275 	objset_t *mos = dp->dp_meta_objset;
276 	dmu_buf_t *dbuf;
277 	dsl_dataset_t *ds;
278 	int err;
279 
280 	ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
281 	    dsl_pool_sync_context(dp));
282 
283 	err = dmu_bonus_hold(mos, dsobj, tag, &dbuf);
284 	if (err)
285 		return (err);
286 	ds = dmu_buf_get_user(dbuf);
287 	if (ds == NULL) {
288 		dsl_dataset_t *winner;
289 
290 		ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_SLEEP);
291 		ds->ds_dbuf = dbuf;
292 		ds->ds_object = dsobj;
293 		ds->ds_phys = dbuf->db_data;
294 
295 		mutex_init(&ds->ds_lock, NULL, MUTEX_DEFAULT, NULL);
296 		mutex_init(&ds->ds_deadlist.bpl_lock, NULL, MUTEX_DEFAULT,
297 		    NULL);
298 
299 		err = bplist_open(&ds->ds_deadlist,
300 		    mos, ds->ds_phys->ds_deadlist_obj);
301 		if (err == 0) {
302 			err = dsl_dir_open_obj(dp,
303 			    ds->ds_phys->ds_dir_obj, NULL, ds, &ds->ds_dir);
304 		}
305 		if (err) {
306 			/*
307 			 * we don't really need to close the blist if we
308 			 * just opened it.
309 			 */
310 			mutex_destroy(&ds->ds_lock);
311 			mutex_destroy(&ds->ds_deadlist.bpl_lock);
312 			kmem_free(ds, sizeof (dsl_dataset_t));
313 			dmu_buf_rele(dbuf, tag);
314 			return (err);
315 		}
316 
317 		if (ds->ds_dir->dd_phys->dd_head_dataset_obj == dsobj) {
318 			ds->ds_snapname[0] = '\0';
319 			if (ds->ds_phys->ds_prev_snap_obj) {
320 				err = dsl_dataset_open_obj(dp,
321 				    ds->ds_phys->ds_prev_snap_obj, NULL,
322 				    DS_MODE_NONE, ds, &ds->ds_prev);
323 			}
324 		} else {
325 			if (snapname) {
326 #ifdef ZFS_DEBUG
327 				dsl_dataset_phys_t *headphys;
328 				dmu_buf_t *headdbuf;
329 				err = dmu_bonus_hold(mos,
330 				    ds->ds_dir->dd_phys->dd_head_dataset_obj,
331 				    FTAG, &headdbuf);
332 				if (err == 0) {
333 					headphys = headdbuf->db_data;
334 					uint64_t foundobj;
335 					err = zap_lookup(dp->dp_meta_objset,
336 					    headphys->ds_snapnames_zapobj,
337 					    snapname, sizeof (foundobj), 1,
338 					    &foundobj);
339 					ASSERT3U(foundobj, ==, dsobj);
340 					dmu_buf_rele(headdbuf, FTAG);
341 				}
342 #endif
343 				(void) strcat(ds->ds_snapname, snapname);
344 			} else if (zfs_flags & ZFS_DEBUG_SNAPNAMES) {
345 				err = dsl_dataset_get_snapname(ds);
346 			}
347 		}
348 
349 		if (err == 0) {
350 			winner = dmu_buf_set_user_ie(dbuf, ds, &ds->ds_phys,
351 			    dsl_dataset_evict);
352 		}
353 		if (err || winner) {
354 			bplist_close(&ds->ds_deadlist);
355 			if (ds->ds_prev) {
356 				dsl_dataset_close(ds->ds_prev,
357 				    DS_MODE_NONE, ds);
358 			}
359 			dsl_dir_close(ds->ds_dir, ds);
360 			mutex_destroy(&ds->ds_lock);
361 			mutex_destroy(&ds->ds_deadlist.bpl_lock);
362 			kmem_free(ds, sizeof (dsl_dataset_t));
363 			if (err) {
364 				dmu_buf_rele(dbuf, tag);
365 				return (err);
366 			}
367 			ds = winner;
368 		} else {
369 			uint64_t new =
370 			    unique_insert(ds->ds_phys->ds_fsid_guid);
371 			if (new != ds->ds_phys->ds_fsid_guid) {
372 				/* XXX it won't necessarily be synced... */
373 				ds->ds_phys->ds_fsid_guid = new;
374 			}
375 		}
376 	}
377 	ASSERT3P(ds->ds_dbuf, ==, dbuf);
378 	ASSERT3P(ds->ds_phys, ==, dbuf->db_data);
379 
380 	mutex_enter(&ds->ds_lock);
381 	if ((DS_MODE_LEVEL(mode) == DS_MODE_PRIMARY &&
382 	    (ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT) &&
383 	    !DS_MODE_IS_INCONSISTENT(mode)) ||
384 	    (ds->ds_open_refcount + weight > DS_REF_MAX)) {
385 		mutex_exit(&ds->ds_lock);
386 		dsl_dataset_close(ds, DS_MODE_NONE, tag);
387 		return (EBUSY);
388 	}
389 	ds->ds_open_refcount += weight;
390 	mutex_exit(&ds->ds_lock);
391 
392 	*dsp = ds;
393 	return (0);
394 }
395 
396 int
397 dsl_dataset_open_spa(spa_t *spa, const char *name, int mode,
398     void *tag, dsl_dataset_t **dsp)
399 {
400 	dsl_dir_t *dd;
401 	dsl_pool_t *dp;
402 	const char *tail;
403 	uint64_t obj;
404 	dsl_dataset_t *ds = NULL;
405 	int err = 0;
406 
407 	err = dsl_dir_open_spa(spa, name, FTAG, &dd, &tail);
408 	if (err)
409 		return (err);
410 
411 	dp = dd->dd_pool;
412 	obj = dd->dd_phys->dd_head_dataset_obj;
413 	rw_enter(&dp->dp_config_rwlock, RW_READER);
414 	if (obj == 0) {
415 		/* A dataset with no associated objset */
416 		err = ENOENT;
417 		goto out;
418 	}
419 
420 	if (tail != NULL) {
421 		objset_t *mos = dp->dp_meta_objset;
422 
423 		err = dsl_dataset_open_obj(dp, obj, NULL,
424 		    DS_MODE_NONE, tag, &ds);
425 		if (err)
426 			goto out;
427 		obj = ds->ds_phys->ds_snapnames_zapobj;
428 		dsl_dataset_close(ds, DS_MODE_NONE, tag);
429 		ds = NULL;
430 
431 		if (tail[0] != '@') {
432 			err = ENOENT;
433 			goto out;
434 		}
435 		tail++;
436 
437 		/* Look for a snapshot */
438 		if (!DS_MODE_IS_READONLY(mode)) {
439 			err = EROFS;
440 			goto out;
441 		}
442 		dprintf("looking for snapshot '%s'\n", tail);
443 		err = zap_lookup(mos, obj, tail, 8, 1, &obj);
444 		if (err)
445 			goto out;
446 	}
447 	err = dsl_dataset_open_obj(dp, obj, tail, mode, tag, &ds);
448 
449 out:
450 	rw_exit(&dp->dp_config_rwlock);
451 	dsl_dir_close(dd, FTAG);
452 
453 	ASSERT3U((err == 0), ==, (ds != NULL));
454 	/* ASSERT(ds == NULL || strcmp(name, ds->ds_name) == 0); */
455 
456 	*dsp = ds;
457 	return (err);
458 }
459 
460 int
461 dsl_dataset_open(const char *name, int mode, void *tag, dsl_dataset_t **dsp)
462 {
463 	return (dsl_dataset_open_spa(NULL, name, mode, tag, dsp));
464 }
465 
466 void
467 dsl_dataset_name(dsl_dataset_t *ds, char *name)
468 {
469 	if (ds == NULL) {
470 		(void) strcpy(name, "mos");
471 	} else {
472 		dsl_dir_name(ds->ds_dir, name);
473 		VERIFY(0 == dsl_dataset_get_snapname(ds));
474 		if (ds->ds_snapname[0]) {
475 			(void) strcat(name, "@");
476 			if (!MUTEX_HELD(&ds->ds_lock)) {
477 				/*
478 				 * We use a "recursive" mutex so that we
479 				 * can call dprintf_ds() with ds_lock held.
480 				 */
481 				mutex_enter(&ds->ds_lock);
482 				(void) strcat(name, ds->ds_snapname);
483 				mutex_exit(&ds->ds_lock);
484 			} else {
485 				(void) strcat(name, ds->ds_snapname);
486 			}
487 		}
488 	}
489 }
490 
491 void
492 dsl_dataset_close(dsl_dataset_t *ds, int mode, void *tag)
493 {
494 	uint64_t weight = ds_refcnt_weight[DS_MODE_LEVEL(mode)];
495 	mutex_enter(&ds->ds_lock);
496 	ASSERT3U(ds->ds_open_refcount, >=, weight);
497 	ds->ds_open_refcount -= weight;
498 	dprintf_ds(ds, "closing mode %u refcount now 0x%llx\n",
499 	    mode, ds->ds_open_refcount);
500 	mutex_exit(&ds->ds_lock);
501 
502 	dmu_buf_rele(ds->ds_dbuf, tag);
503 }
504 
505 void
506 dsl_dataset_create_root(dsl_pool_t *dp, uint64_t *ddobjp, dmu_tx_t *tx)
507 {
508 	objset_t *mos = dp->dp_meta_objset;
509 	dmu_buf_t *dbuf;
510 	dsl_dataset_phys_t *dsphys;
511 	dsl_dataset_t *ds;
512 	uint64_t dsobj;
513 	dsl_dir_t *dd;
514 
515 	dsl_dir_create_root(mos, ddobjp, tx);
516 	VERIFY(0 == dsl_dir_open_obj(dp, *ddobjp, NULL, FTAG, &dd));
517 
518 	dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
519 	    DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
520 	VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
521 	dmu_buf_will_dirty(dbuf, tx);
522 	dsphys = dbuf->db_data;
523 	dsphys->ds_dir_obj = dd->dd_object;
524 	dsphys->ds_fsid_guid = unique_create();
525 	unique_remove(dsphys->ds_fsid_guid); /* it isn't open yet */
526 	(void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
527 	    sizeof (dsphys->ds_guid));
528 	dsphys->ds_snapnames_zapobj =
529 	    zap_create(mos, DMU_OT_DSL_DS_SNAP_MAP, DMU_OT_NONE, 0, tx);
530 	dsphys->ds_creation_time = gethrestime_sec();
531 	dsphys->ds_creation_txg = tx->tx_txg;
532 	dsphys->ds_deadlist_obj =
533 	    bplist_create(mos, DSL_DEADLIST_BLOCKSIZE, tx);
534 	dmu_buf_rele(dbuf, FTAG);
535 
536 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
537 	dd->dd_phys->dd_head_dataset_obj = dsobj;
538 	dsl_dir_close(dd, FTAG);
539 
540 	VERIFY(0 ==
541 	    dsl_dataset_open_obj(dp, dsobj, NULL, DS_MODE_NONE, FTAG, &ds));
542 	(void) dmu_objset_create_impl(dp->dp_spa, ds, DMU_OST_ZFS, tx);
543 	dsl_dataset_close(ds, DS_MODE_NONE, FTAG);
544 }
545 
546 uint64_t
547 dsl_dataset_create_sync(dsl_dir_t *pdd,
548     const char *lastname, dsl_dataset_t *clone_parent, dmu_tx_t *tx)
549 {
550 	dsl_pool_t *dp = pdd->dd_pool;
551 	dmu_buf_t *dbuf;
552 	dsl_dataset_phys_t *dsphys;
553 	uint64_t dsobj, ddobj;
554 	objset_t *mos = dp->dp_meta_objset;
555 	dsl_dir_t *dd;
556 
557 	ASSERT(clone_parent == NULL || clone_parent->ds_dir->dd_pool == dp);
558 	ASSERT(clone_parent == NULL ||
559 	    clone_parent->ds_phys->ds_num_children > 0);
560 	ASSERT(lastname[0] != '@');
561 	ASSERT(dmu_tx_is_syncing(tx));
562 
563 	ddobj = dsl_dir_create_sync(pdd, lastname, tx);
564 	VERIFY(0 == dsl_dir_open_obj(dp, ddobj, lastname, FTAG, &dd));
565 
566 	dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
567 	    DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
568 	VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
569 	dmu_buf_will_dirty(dbuf, tx);
570 	dsphys = dbuf->db_data;
571 	dsphys->ds_dir_obj = dd->dd_object;
572 	dsphys->ds_fsid_guid = unique_create();
573 	unique_remove(dsphys->ds_fsid_guid); /* it isn't open yet */
574 	(void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
575 	    sizeof (dsphys->ds_guid));
576 	dsphys->ds_snapnames_zapobj =
577 	    zap_create(mos, DMU_OT_DSL_DS_SNAP_MAP, DMU_OT_NONE, 0, tx);
578 	dsphys->ds_creation_time = gethrestime_sec();
579 	dsphys->ds_creation_txg = tx->tx_txg;
580 	dsphys->ds_deadlist_obj =
581 	    bplist_create(mos, DSL_DEADLIST_BLOCKSIZE, tx);
582 	if (clone_parent) {
583 		dsphys->ds_prev_snap_obj = clone_parent->ds_object;
584 		dsphys->ds_prev_snap_txg =
585 		    clone_parent->ds_phys->ds_creation_txg;
586 		dsphys->ds_used_bytes =
587 		    clone_parent->ds_phys->ds_used_bytes;
588 		dsphys->ds_compressed_bytes =
589 		    clone_parent->ds_phys->ds_compressed_bytes;
590 		dsphys->ds_uncompressed_bytes =
591 		    clone_parent->ds_phys->ds_uncompressed_bytes;
592 		dsphys->ds_bp = clone_parent->ds_phys->ds_bp;
593 
594 		dmu_buf_will_dirty(clone_parent->ds_dbuf, tx);
595 		clone_parent->ds_phys->ds_num_children++;
596 
597 		dmu_buf_will_dirty(dd->dd_dbuf, tx);
598 		dd->dd_phys->dd_clone_parent_obj = clone_parent->ds_object;
599 	}
600 	dmu_buf_rele(dbuf, FTAG);
601 
602 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
603 	dd->dd_phys->dd_head_dataset_obj = dsobj;
604 	dsl_dir_close(dd, FTAG);
605 
606 	return (dsobj);
607 }
608 
609 struct destroyarg {
610 	dsl_sync_task_group_t *dstg;
611 	char *snapname;
612 	void *tag;
613 	char *failed;
614 };
615 
616 static int
617 dsl_snapshot_destroy_one(char *name, void *arg)
618 {
619 	struct destroyarg *da = arg;
620 	dsl_dataset_t *ds;
621 	char *cp;
622 	int err;
623 
624 	(void) strcat(name, "@");
625 	(void) strcat(name, da->snapname);
626 	err = dsl_dataset_open(name,
627 	    DS_MODE_EXCLUSIVE | DS_MODE_READONLY | DS_MODE_INCONSISTENT,
628 	    da->tag, &ds);
629 	cp = strchr(name, '@');
630 	*cp = '\0';
631 	if (err == ENOENT)
632 		return (0);
633 	if (err) {
634 		(void) strcpy(da->failed, name);
635 		return (err);
636 	}
637 
638 	dsl_sync_task_create(da->dstg, dsl_dataset_destroy_check,
639 	    dsl_dataset_destroy_sync, ds, da->tag, 0);
640 	return (0);
641 }
642 
643 /*
644  * Destroy 'snapname' in all descendants of 'fsname'.
645  */
646 #pragma weak dmu_snapshots_destroy = dsl_snapshots_destroy
647 int
648 dsl_snapshots_destroy(char *fsname, char *snapname)
649 {
650 	int err;
651 	struct destroyarg da;
652 	dsl_sync_task_t *dst;
653 	spa_t *spa;
654 	char *cp;
655 
656 	cp = strchr(fsname, '/');
657 	if (cp) {
658 		*cp = '\0';
659 		err = spa_open(fsname, &spa, FTAG);
660 		*cp = '/';
661 	} else {
662 		err = spa_open(fsname, &spa, FTAG);
663 	}
664 	if (err)
665 		return (err);
666 	da.dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
667 	da.snapname = snapname;
668 	da.tag = FTAG;
669 	da.failed = fsname;
670 
671 	err = dmu_objset_find(fsname,
672 	    dsl_snapshot_destroy_one, &da, DS_FIND_CHILDREN);
673 
674 	if (err == 0)
675 		err = dsl_sync_task_group_wait(da.dstg);
676 
677 	for (dst = list_head(&da.dstg->dstg_tasks); dst;
678 	    dst = list_next(&da.dstg->dstg_tasks, dst)) {
679 		dsl_dataset_t *ds = dst->dst_arg1;
680 		if (dst->dst_err) {
681 			dsl_dataset_name(ds, fsname);
682 			cp = strchr(fsname, '@');
683 			*cp = '\0';
684 		}
685 		/*
686 		 * If it was successful, destroy_sync would have
687 		 * closed the ds
688 		 */
689 		if (err)
690 			dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
691 	}
692 
693 	dsl_sync_task_group_destroy(da.dstg);
694 	spa_close(spa, FTAG);
695 	return (err);
696 }
697 
698 int
699 dsl_dataset_destroy(const char *name)
700 {
701 	int err;
702 	dsl_sync_task_group_t *dstg;
703 	objset_t *os;
704 	dsl_dataset_t *ds;
705 	dsl_dir_t *dd;
706 	uint64_t obj;
707 
708 	if (strchr(name, '@')) {
709 		/* Destroying a snapshot is simpler */
710 		err = dsl_dataset_open(name,
711 		    DS_MODE_EXCLUSIVE | DS_MODE_READONLY | DS_MODE_INCONSISTENT,
712 		    FTAG, &ds);
713 		if (err)
714 			return (err);
715 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
716 		    dsl_dataset_destroy_check, dsl_dataset_destroy_sync,
717 		    ds, FTAG, 0);
718 		if (err)
719 			dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
720 		return (err);
721 	}
722 
723 	err = dmu_objset_open(name, DMU_OST_ANY,
724 	    DS_MODE_EXCLUSIVE | DS_MODE_INCONSISTENT, &os);
725 	if (err)
726 		return (err);
727 	ds = os->os->os_dsl_dataset;
728 	dd = ds->ds_dir;
729 
730 	/*
731 	 * Check for errors and mark this ds as inconsistent, in
732 	 * case we crash while freeing the objects.
733 	 */
734 	err = dsl_sync_task_do(dd->dd_pool, dsl_dataset_destroy_begin_check,
735 	    dsl_dataset_destroy_begin_sync, ds, NULL, 0);
736 	if (err) {
737 		dmu_objset_close(os);
738 		return (err);
739 	}
740 
741 	/*
742 	 * remove the objects in open context, so that we won't
743 	 * have too much to do in syncing context.
744 	 */
745 	for (obj = 0; err == 0; err = dmu_object_next(os, &obj, FALSE,
746 	    ds->ds_phys->ds_prev_snap_txg)) {
747 		dmu_tx_t *tx = dmu_tx_create(os);
748 		dmu_tx_hold_free(tx, obj, 0, DMU_OBJECT_END);
749 		dmu_tx_hold_bonus(tx, obj);
750 		err = dmu_tx_assign(tx, TXG_WAIT);
751 		if (err) {
752 			/*
753 			 * Perhaps there is not enough disk
754 			 * space.  Just deal with it from
755 			 * dsl_dataset_destroy_sync().
756 			 */
757 			dmu_tx_abort(tx);
758 			continue;
759 		}
760 		VERIFY(0 == dmu_object_free(os, obj, tx));
761 		dmu_tx_commit(tx);
762 	}
763 	/* Make sure it's not dirty before we finish destroying it. */
764 	txg_wait_synced(dd->dd_pool, 0);
765 
766 	dmu_objset_close(os);
767 	if (err != ESRCH)
768 		return (err);
769 
770 	err = dsl_dataset_open(name,
771 	    DS_MODE_EXCLUSIVE | DS_MODE_READONLY | DS_MODE_INCONSISTENT,
772 	    FTAG, &ds);
773 	if (err)
774 		return (err);
775 
776 	err = dsl_dir_open(name, FTAG, &dd, NULL);
777 	if (err) {
778 		dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
779 		return (err);
780 	}
781 
782 	/*
783 	 * Blow away the dsl_dir + head dataset.
784 	 */
785 	dstg = dsl_sync_task_group_create(ds->ds_dir->dd_pool);
786 	dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
787 	    dsl_dataset_destroy_sync, ds, FTAG, 0);
788 	dsl_sync_task_create(dstg, dsl_dir_destroy_check,
789 	    dsl_dir_destroy_sync, dd, FTAG, 0);
790 	err = dsl_sync_task_group_wait(dstg);
791 	dsl_sync_task_group_destroy(dstg);
792 	/* if it is successful, *destroy_sync will close the ds+dd */
793 	if (err) {
794 		dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
795 		dsl_dir_close(dd, FTAG);
796 	}
797 	return (err);
798 }
799 
800 int
801 dsl_dataset_rollback(dsl_dataset_t *ds)
802 {
803 	ASSERT3U(ds->ds_open_refcount, ==, DS_REF_MAX);
804 	return (dsl_sync_task_do(ds->ds_dir->dd_pool,
805 	    dsl_dataset_rollback_check, dsl_dataset_rollback_sync,
806 	    ds, NULL, 0));
807 }
808 
809 void *
810 dsl_dataset_set_user_ptr(dsl_dataset_t *ds,
811     void *p, dsl_dataset_evict_func_t func)
812 {
813 	void *old;
814 
815 	mutex_enter(&ds->ds_lock);
816 	old = ds->ds_user_ptr;
817 	if (old == NULL) {
818 		ds->ds_user_ptr = p;
819 		ds->ds_user_evict_func = func;
820 	}
821 	mutex_exit(&ds->ds_lock);
822 	return (old);
823 }
824 
825 void *
826 dsl_dataset_get_user_ptr(dsl_dataset_t *ds)
827 {
828 	return (ds->ds_user_ptr);
829 }
830 
831 
832 void
833 dsl_dataset_get_blkptr(dsl_dataset_t *ds, blkptr_t *bp)
834 {
835 	*bp = ds->ds_phys->ds_bp;
836 }
837 
838 void
839 dsl_dataset_set_blkptr(dsl_dataset_t *ds, blkptr_t *bp, dmu_tx_t *tx)
840 {
841 	ASSERT(dmu_tx_is_syncing(tx));
842 	/* If it's the meta-objset, set dp_meta_rootbp */
843 	if (ds == NULL) {
844 		tx->tx_pool->dp_meta_rootbp = *bp;
845 	} else {
846 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
847 		ds->ds_phys->ds_bp = *bp;
848 	}
849 }
850 
851 spa_t *
852 dsl_dataset_get_spa(dsl_dataset_t *ds)
853 {
854 	return (ds->ds_dir->dd_pool->dp_spa);
855 }
856 
857 void
858 dsl_dataset_dirty(dsl_dataset_t *ds, dmu_tx_t *tx)
859 {
860 	dsl_pool_t *dp;
861 
862 	if (ds == NULL) /* this is the meta-objset */
863 		return;
864 
865 	ASSERT(ds->ds_user_ptr != NULL);
866 
867 	if (ds->ds_phys->ds_next_snap_obj != 0)
868 		panic("dirtying snapshot!");
869 
870 	dp = ds->ds_dir->dd_pool;
871 
872 	if (txg_list_add(&dp->dp_dirty_datasets, ds, tx->tx_txg) == 0) {
873 		/* up the hold count until we can be written out */
874 		dmu_buf_add_ref(ds->ds_dbuf, ds);
875 	}
876 }
877 
878 struct killarg {
879 	uint64_t *usedp;
880 	uint64_t *compressedp;
881 	uint64_t *uncompressedp;
882 	zio_t *zio;
883 	dmu_tx_t *tx;
884 };
885 
886 static int
887 kill_blkptr(traverse_blk_cache_t *bc, spa_t *spa, void *arg)
888 {
889 	struct killarg *ka = arg;
890 	blkptr_t *bp = &bc->bc_blkptr;
891 
892 	ASSERT3U(bc->bc_errno, ==, 0);
893 
894 	/*
895 	 * Since this callback is not called concurrently, no lock is
896 	 * needed on the accounting values.
897 	 */
898 	*ka->usedp += bp_get_dasize(spa, bp);
899 	*ka->compressedp += BP_GET_PSIZE(bp);
900 	*ka->uncompressedp += BP_GET_UCSIZE(bp);
901 	/* XXX check for EIO? */
902 	(void) arc_free(ka->zio, spa, ka->tx->tx_txg, bp, NULL, NULL,
903 	    ARC_NOWAIT);
904 	return (0);
905 }
906 
907 /* ARGSUSED */
908 static int
909 dsl_dataset_rollback_check(void *arg1, void *arg2, dmu_tx_t *tx)
910 {
911 	dsl_dataset_t *ds = arg1;
912 
913 	/*
914 	 * There must be a previous snapshot.  I suppose we could roll
915 	 * it back to being empty (and re-initialize the upper (ZPL)
916 	 * layer).  But for now there's no way to do this via the user
917 	 * interface.
918 	 */
919 	if (ds->ds_phys->ds_prev_snap_txg == 0)
920 		return (EINVAL);
921 
922 	/*
923 	 * This must not be a snapshot.
924 	 */
925 	if (ds->ds_phys->ds_next_snap_obj != 0)
926 		return (EINVAL);
927 
928 	/*
929 	 * If we made changes this txg, traverse_dsl_dataset won't find
930 	 * them.  Try again.
931 	 */
932 	if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
933 		return (EAGAIN);
934 
935 	return (0);
936 }
937 
938 /* ARGSUSED */
939 static void
940 dsl_dataset_rollback_sync(void *arg1, void *arg2, dmu_tx_t *tx)
941 {
942 	dsl_dataset_t *ds = arg1;
943 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
944 
945 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
946 
947 	/* Zero out the deadlist. */
948 	bplist_close(&ds->ds_deadlist);
949 	bplist_destroy(mos, ds->ds_phys->ds_deadlist_obj, tx);
950 	ds->ds_phys->ds_deadlist_obj =
951 	    bplist_create(mos, DSL_DEADLIST_BLOCKSIZE, tx);
952 	VERIFY(0 == bplist_open(&ds->ds_deadlist, mos,
953 	    ds->ds_phys->ds_deadlist_obj));
954 
955 	{
956 		/* Free blkptrs that we gave birth to */
957 		zio_t *zio;
958 		uint64_t used = 0, compressed = 0, uncompressed = 0;
959 		struct killarg ka;
960 
961 		zio = zio_root(tx->tx_pool->dp_spa, NULL, NULL,
962 		    ZIO_FLAG_MUSTSUCCEED);
963 		ka.usedp = &used;
964 		ka.compressedp = &compressed;
965 		ka.uncompressedp = &uncompressed;
966 		ka.zio = zio;
967 		ka.tx = tx;
968 		(void) traverse_dsl_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
969 		    ADVANCE_POST, kill_blkptr, &ka);
970 		(void) zio_wait(zio);
971 
972 		dsl_dir_diduse_space(ds->ds_dir,
973 		    -used, -compressed, -uncompressed, tx);
974 	}
975 
976 	/* Change our contents to that of the prev snapshot */
977 	ASSERT3U(ds->ds_prev->ds_object, ==, ds->ds_phys->ds_prev_snap_obj);
978 	ds->ds_phys->ds_bp = ds->ds_prev->ds_phys->ds_bp;
979 	ds->ds_phys->ds_used_bytes = ds->ds_prev->ds_phys->ds_used_bytes;
980 	ds->ds_phys->ds_compressed_bytes =
981 	    ds->ds_prev->ds_phys->ds_compressed_bytes;
982 	ds->ds_phys->ds_uncompressed_bytes =
983 	    ds->ds_prev->ds_phys->ds_uncompressed_bytes;
984 	ds->ds_phys->ds_flags = ds->ds_prev->ds_phys->ds_flags;
985 	ds->ds_phys->ds_unique_bytes = 0;
986 
987 	if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
988 		dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
989 		ds->ds_prev->ds_phys->ds_unique_bytes = 0;
990 	}
991 }
992 
993 /* ARGSUSED */
994 static int
995 dsl_dataset_destroy_begin_check(void *arg1, void *arg2, dmu_tx_t *tx)
996 {
997 	dsl_dataset_t *ds = arg1;
998 
999 	/*
1000 	 * Can't delete a head dataset if there are snapshots of it.
1001 	 * (Except if the only snapshots are from the branch we cloned
1002 	 * from.)
1003 	 */
1004 	if (ds->ds_prev != NULL &&
1005 	    ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1006 		return (EINVAL);
1007 
1008 	return (0);
1009 }
1010 
1011 /* ARGSUSED */
1012 static void
1013 dsl_dataset_destroy_begin_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1014 {
1015 	dsl_dataset_t *ds = arg1;
1016 
1017 	/* Mark it as inconsistent on-disk, in case we crash */
1018 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1019 	ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
1020 }
1021 
1022 /* ARGSUSED */
1023 static int
1024 dsl_dataset_destroy_check(void *arg1, void *arg2, dmu_tx_t *tx)
1025 {
1026 	dsl_dataset_t *ds = arg1;
1027 
1028 	/* Can't delete a branch point. */
1029 	if (ds->ds_phys->ds_num_children > 1)
1030 		return (EEXIST);
1031 
1032 	/*
1033 	 * Can't delete a head dataset if there are snapshots of it.
1034 	 * (Except if the only snapshots are from the branch we cloned
1035 	 * from.)
1036 	 */
1037 	if (ds->ds_prev != NULL &&
1038 	    ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1039 		return (EINVAL);
1040 
1041 	/*
1042 	 * If we made changes this txg, traverse_dsl_dataset won't find
1043 	 * them.  Try again.
1044 	 */
1045 	if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
1046 		return (EAGAIN);
1047 
1048 	/* XXX we should do some i/o error checking... */
1049 	return (0);
1050 }
1051 
1052 static void
1053 dsl_dataset_destroy_sync(void *arg1, void *tag, dmu_tx_t *tx)
1054 {
1055 	dsl_dataset_t *ds = arg1;
1056 	uint64_t used = 0, compressed = 0, uncompressed = 0;
1057 	zio_t *zio;
1058 	int err;
1059 	int after_branch_point = FALSE;
1060 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1061 	objset_t *mos = dp->dp_meta_objset;
1062 	dsl_dataset_t *ds_prev = NULL;
1063 	uint64_t obj;
1064 
1065 	ASSERT3U(ds->ds_open_refcount, ==, DS_REF_MAX);
1066 	ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
1067 	ASSERT(ds->ds_prev == NULL ||
1068 	    ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
1069 	ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
1070 
1071 	ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1072 
1073 	obj = ds->ds_object;
1074 
1075 	if (ds->ds_phys->ds_prev_snap_obj != 0) {
1076 		if (ds->ds_prev) {
1077 			ds_prev = ds->ds_prev;
1078 		} else {
1079 			VERIFY(0 == dsl_dataset_open_obj(dp,
1080 			    ds->ds_phys->ds_prev_snap_obj, NULL,
1081 			    DS_MODE_NONE, FTAG, &ds_prev));
1082 		}
1083 		after_branch_point =
1084 		    (ds_prev->ds_phys->ds_next_snap_obj != obj);
1085 
1086 		dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
1087 		if (after_branch_point &&
1088 		    ds->ds_phys->ds_next_snap_obj == 0) {
1089 			/* This clone is toast. */
1090 			ASSERT(ds_prev->ds_phys->ds_num_children > 1);
1091 			ds_prev->ds_phys->ds_num_children--;
1092 		} else if (!after_branch_point) {
1093 			ds_prev->ds_phys->ds_next_snap_obj =
1094 			    ds->ds_phys->ds_next_snap_obj;
1095 		}
1096 	}
1097 
1098 	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
1099 
1100 	if (ds->ds_phys->ds_next_snap_obj != 0) {
1101 		blkptr_t bp;
1102 		dsl_dataset_t *ds_next;
1103 		uint64_t itor = 0;
1104 
1105 		spa_scrub_restart(dp->dp_spa, tx->tx_txg);
1106 
1107 		VERIFY(0 == dsl_dataset_open_obj(dp,
1108 		    ds->ds_phys->ds_next_snap_obj, NULL,
1109 		    DS_MODE_NONE, FTAG, &ds_next));
1110 		ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
1111 
1112 		dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
1113 		ds_next->ds_phys->ds_prev_snap_obj =
1114 		    ds->ds_phys->ds_prev_snap_obj;
1115 		ds_next->ds_phys->ds_prev_snap_txg =
1116 		    ds->ds_phys->ds_prev_snap_txg;
1117 		ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1118 		    ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
1119 
1120 		/*
1121 		 * Transfer to our deadlist (which will become next's
1122 		 * new deadlist) any entries from next's current
1123 		 * deadlist which were born before prev, and free the
1124 		 * other entries.
1125 		 *
1126 		 * XXX we're doing this long task with the config lock held
1127 		 */
1128 		while (bplist_iterate(&ds_next->ds_deadlist, &itor,
1129 		    &bp) == 0) {
1130 			if (bp.blk_birth <= ds->ds_phys->ds_prev_snap_txg) {
1131 				VERIFY(0 == bplist_enqueue(&ds->ds_deadlist,
1132 				    &bp, tx));
1133 				if (ds_prev && !after_branch_point &&
1134 				    bp.blk_birth >
1135 				    ds_prev->ds_phys->ds_prev_snap_txg) {
1136 					ds_prev->ds_phys->ds_unique_bytes +=
1137 					    bp_get_dasize(dp->dp_spa, &bp);
1138 				}
1139 			} else {
1140 				used += bp_get_dasize(dp->dp_spa, &bp);
1141 				compressed += BP_GET_PSIZE(&bp);
1142 				uncompressed += BP_GET_UCSIZE(&bp);
1143 				/* XXX check return value? */
1144 				(void) arc_free(zio, dp->dp_spa, tx->tx_txg,
1145 				    &bp, NULL, NULL, ARC_NOWAIT);
1146 			}
1147 		}
1148 
1149 		/* free next's deadlist */
1150 		bplist_close(&ds_next->ds_deadlist);
1151 		bplist_destroy(mos, ds_next->ds_phys->ds_deadlist_obj, tx);
1152 
1153 		/* set next's deadlist to our deadlist */
1154 		ds_next->ds_phys->ds_deadlist_obj =
1155 		    ds->ds_phys->ds_deadlist_obj;
1156 		VERIFY(0 == bplist_open(&ds_next->ds_deadlist, mos,
1157 		    ds_next->ds_phys->ds_deadlist_obj));
1158 		ds->ds_phys->ds_deadlist_obj = 0;
1159 
1160 		if (ds_next->ds_phys->ds_next_snap_obj != 0) {
1161 			/*
1162 			 * Update next's unique to include blocks which
1163 			 * were previously shared by only this snapshot
1164 			 * and it.  Those blocks will be born after the
1165 			 * prev snap and before this snap, and will have
1166 			 * died after the next snap and before the one
1167 			 * after that (ie. be on the snap after next's
1168 			 * deadlist).
1169 			 *
1170 			 * XXX we're doing this long task with the
1171 			 * config lock held
1172 			 */
1173 			dsl_dataset_t *ds_after_next;
1174 
1175 			VERIFY(0 == dsl_dataset_open_obj(dp,
1176 			    ds_next->ds_phys->ds_next_snap_obj, NULL,
1177 			    DS_MODE_NONE, FTAG, &ds_after_next));
1178 			itor = 0;
1179 			while (bplist_iterate(&ds_after_next->ds_deadlist,
1180 			    &itor, &bp) == 0) {
1181 				if (bp.blk_birth >
1182 				    ds->ds_phys->ds_prev_snap_txg &&
1183 				    bp.blk_birth <=
1184 				    ds->ds_phys->ds_creation_txg) {
1185 					ds_next->ds_phys->ds_unique_bytes +=
1186 					    bp_get_dasize(dp->dp_spa, &bp);
1187 				}
1188 			}
1189 
1190 			dsl_dataset_close(ds_after_next, DS_MODE_NONE, FTAG);
1191 			ASSERT3P(ds_next->ds_prev, ==, NULL);
1192 		} else {
1193 			/*
1194 			 * It would be nice to update the head dataset's
1195 			 * unique.  To do so we would have to traverse
1196 			 * it for blocks born after ds_prev, which is
1197 			 * pretty expensive just to maintain something
1198 			 * for debugging purposes.
1199 			 */
1200 			ASSERT3P(ds_next->ds_prev, ==, ds);
1201 			dsl_dataset_close(ds_next->ds_prev, DS_MODE_NONE,
1202 			    ds_next);
1203 			if (ds_prev) {
1204 				VERIFY(0 == dsl_dataset_open_obj(dp,
1205 				    ds->ds_phys->ds_prev_snap_obj, NULL,
1206 				    DS_MODE_NONE, ds_next, &ds_next->ds_prev));
1207 			} else {
1208 				ds_next->ds_prev = NULL;
1209 			}
1210 		}
1211 		dsl_dataset_close(ds_next, DS_MODE_NONE, FTAG);
1212 
1213 		/*
1214 		 * NB: unique_bytes is not accurate for head objsets
1215 		 * because we don't update it when we delete the most
1216 		 * recent snapshot -- see above comment.
1217 		 */
1218 		ASSERT3U(used, ==, ds->ds_phys->ds_unique_bytes);
1219 	} else {
1220 		/*
1221 		 * There's no next snapshot, so this is a head dataset.
1222 		 * Destroy the deadlist.  Unless it's a clone, the
1223 		 * deadlist should be empty.  (If it's a clone, it's
1224 		 * safe to ignore the deadlist contents.)
1225 		 */
1226 		struct killarg ka;
1227 
1228 		ASSERT(after_branch_point || bplist_empty(&ds->ds_deadlist));
1229 		bplist_close(&ds->ds_deadlist);
1230 		bplist_destroy(mos, ds->ds_phys->ds_deadlist_obj, tx);
1231 		ds->ds_phys->ds_deadlist_obj = 0;
1232 
1233 		/*
1234 		 * Free everything that we point to (that's born after
1235 		 * the previous snapshot, if we are a clone)
1236 		 *
1237 		 * XXX we're doing this long task with the config lock held
1238 		 */
1239 		ka.usedp = &used;
1240 		ka.compressedp = &compressed;
1241 		ka.uncompressedp = &uncompressed;
1242 		ka.zio = zio;
1243 		ka.tx = tx;
1244 		err = traverse_dsl_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
1245 		    ADVANCE_POST, kill_blkptr, &ka);
1246 		ASSERT3U(err, ==, 0);
1247 	}
1248 
1249 	err = zio_wait(zio);
1250 	ASSERT3U(err, ==, 0);
1251 
1252 	dsl_dir_diduse_space(ds->ds_dir, -used, -compressed, -uncompressed, tx);
1253 
1254 	if (ds->ds_phys->ds_snapnames_zapobj) {
1255 		err = zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx);
1256 		ASSERT(err == 0);
1257 	}
1258 
1259 	if (ds->ds_dir->dd_phys->dd_head_dataset_obj == ds->ds_object) {
1260 		/* Erase the link in the dataset */
1261 		dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
1262 		ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
1263 		/*
1264 		 * dsl_dir_sync_destroy() called us, they'll destroy
1265 		 * the dataset.
1266 		 */
1267 	} else {
1268 		/* remove from snapshot namespace */
1269 		dsl_dataset_t *ds_head;
1270 		VERIFY(0 == dsl_dataset_open_obj(dp,
1271 		    ds->ds_dir->dd_phys->dd_head_dataset_obj, NULL,
1272 		    DS_MODE_NONE, FTAG, &ds_head));
1273 		VERIFY(0 == dsl_dataset_get_snapname(ds));
1274 #ifdef ZFS_DEBUG
1275 		{
1276 			uint64_t val;
1277 			err = zap_lookup(mos,
1278 			    ds_head->ds_phys->ds_snapnames_zapobj,
1279 			    ds->ds_snapname, 8, 1, &val);
1280 			ASSERT3U(err, ==, 0);
1281 			ASSERT3U(val, ==, obj);
1282 		}
1283 #endif
1284 		err = zap_remove(mos, ds_head->ds_phys->ds_snapnames_zapobj,
1285 		    ds->ds_snapname, tx);
1286 		ASSERT(err == 0);
1287 		dsl_dataset_close(ds_head, DS_MODE_NONE, FTAG);
1288 	}
1289 
1290 	if (ds_prev && ds->ds_prev != ds_prev)
1291 		dsl_dataset_close(ds_prev, DS_MODE_NONE, FTAG);
1292 
1293 	dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, tag);
1294 	VERIFY(0 == dmu_object_free(mos, obj, tx));
1295 }
1296 
1297 /* ARGSUSED */
1298 int
1299 dsl_dataset_snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
1300 {
1301 	objset_t *os = arg1;
1302 	dsl_dataset_t *ds = os->os->os_dsl_dataset;
1303 	const char *snapname = arg2;
1304 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1305 	int err;
1306 	uint64_t value;
1307 
1308 	/*
1309 	 * We don't allow multiple snapshots of the same txg.  If there
1310 	 * is already one, try again.
1311 	 */
1312 	if (ds->ds_phys->ds_prev_snap_txg >= tx->tx_txg)
1313 		return (EAGAIN);
1314 
1315 	/*
1316 	 * Check for conflicting name snapshot name.
1317 	 */
1318 	err = zap_lookup(mos, ds->ds_phys->ds_snapnames_zapobj,
1319 	    snapname, 8, 1, &value);
1320 	if (err == 0)
1321 		return (EEXIST);
1322 	if (err != ENOENT)
1323 		return (err);
1324 
1325 	ds->ds_trysnap_txg = tx->tx_txg;
1326 	return (0);
1327 }
1328 
1329 void
1330 dsl_dataset_snapshot_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1331 {
1332 	objset_t *os = arg1;
1333 	dsl_dataset_t *ds = os->os->os_dsl_dataset;
1334 	const char *snapname = arg2;
1335 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1336 	dmu_buf_t *dbuf;
1337 	dsl_dataset_phys_t *dsphys;
1338 	uint64_t dsobj;
1339 	objset_t *mos = dp->dp_meta_objset;
1340 	int err;
1341 
1342 	spa_scrub_restart(dp->dp_spa, tx->tx_txg);
1343 	ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1344 
1345 	dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
1346 	    DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
1347 	VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
1348 	dmu_buf_will_dirty(dbuf, tx);
1349 	dsphys = dbuf->db_data;
1350 	dsphys->ds_dir_obj = ds->ds_dir->dd_object;
1351 	dsphys->ds_fsid_guid = unique_create();
1352 	unique_remove(dsphys->ds_fsid_guid); /* it isn't open yet */
1353 	(void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
1354 	    sizeof (dsphys->ds_guid));
1355 	dsphys->ds_prev_snap_obj = ds->ds_phys->ds_prev_snap_obj;
1356 	dsphys->ds_prev_snap_txg = ds->ds_phys->ds_prev_snap_txg;
1357 	dsphys->ds_next_snap_obj = ds->ds_object;
1358 	dsphys->ds_num_children = 1;
1359 	dsphys->ds_creation_time = gethrestime_sec();
1360 	dsphys->ds_creation_txg = tx->tx_txg;
1361 	dsphys->ds_deadlist_obj = ds->ds_phys->ds_deadlist_obj;
1362 	dsphys->ds_used_bytes = ds->ds_phys->ds_used_bytes;
1363 	dsphys->ds_compressed_bytes = ds->ds_phys->ds_compressed_bytes;
1364 	dsphys->ds_uncompressed_bytes = ds->ds_phys->ds_uncompressed_bytes;
1365 	dsphys->ds_flags = ds->ds_phys->ds_flags;
1366 	dsphys->ds_bp = ds->ds_phys->ds_bp;
1367 	dmu_buf_rele(dbuf, FTAG);
1368 
1369 	ASSERT3U(ds->ds_prev != 0, ==, ds->ds_phys->ds_prev_snap_obj != 0);
1370 	if (ds->ds_prev) {
1371 		ASSERT(ds->ds_prev->ds_phys->ds_next_snap_obj ==
1372 		    ds->ds_object ||
1373 		    ds->ds_prev->ds_phys->ds_num_children > 1);
1374 		if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
1375 			dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
1376 			ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1377 			    ds->ds_prev->ds_phys->ds_creation_txg);
1378 			ds->ds_prev->ds_phys->ds_next_snap_obj = dsobj;
1379 		}
1380 	}
1381 
1382 	bplist_close(&ds->ds_deadlist);
1383 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1384 	ASSERT3U(ds->ds_phys->ds_prev_snap_txg, <, dsphys->ds_creation_txg);
1385 	ds->ds_phys->ds_prev_snap_obj = dsobj;
1386 	ds->ds_phys->ds_prev_snap_txg = dsphys->ds_creation_txg;
1387 	ds->ds_phys->ds_unique_bytes = 0;
1388 	ds->ds_phys->ds_deadlist_obj =
1389 	    bplist_create(mos, DSL_DEADLIST_BLOCKSIZE, tx);
1390 	VERIFY(0 == bplist_open(&ds->ds_deadlist, mos,
1391 	    ds->ds_phys->ds_deadlist_obj));
1392 
1393 	dprintf("snap '%s' -> obj %llu\n", snapname, dsobj);
1394 	err = zap_add(mos, ds->ds_phys->ds_snapnames_zapobj,
1395 	    snapname, 8, 1, &dsobj, tx);
1396 	ASSERT(err == 0);
1397 
1398 	if (ds->ds_prev)
1399 		dsl_dataset_close(ds->ds_prev, DS_MODE_NONE, ds);
1400 	VERIFY(0 == dsl_dataset_open_obj(dp,
1401 	    ds->ds_phys->ds_prev_snap_obj, snapname,
1402 	    DS_MODE_NONE, ds, &ds->ds_prev));
1403 }
1404 
1405 void
1406 dsl_dataset_sync(dsl_dataset_t *ds, dmu_tx_t *tx)
1407 {
1408 	ASSERT(dmu_tx_is_syncing(tx));
1409 	ASSERT(ds->ds_user_ptr != NULL);
1410 	ASSERT(ds->ds_phys->ds_next_snap_obj == 0);
1411 
1412 	dmu_objset_sync(ds->ds_user_ptr, tx);
1413 	dsl_dir_dirty(ds->ds_dir, tx);
1414 	bplist_close(&ds->ds_deadlist);
1415 
1416 	dmu_buf_rele(ds->ds_dbuf, ds);
1417 }
1418 
1419 void
1420 dsl_dataset_stats(dsl_dataset_t *ds, nvlist_t *nv)
1421 {
1422 	dsl_dir_stats(ds->ds_dir, nv);
1423 
1424 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATION,
1425 	    ds->ds_phys->ds_creation_time);
1426 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_CREATETXG,
1427 	    ds->ds_phys->ds_creation_txg);
1428 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_REFERENCED,
1429 	    ds->ds_phys->ds_used_bytes);
1430 
1431 	if (ds->ds_phys->ds_next_snap_obj) {
1432 		/*
1433 		 * This is a snapshot; override the dd's space used with
1434 		 * our unique space and compression ratio.
1435 		 */
1436 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USED,
1437 		    ds->ds_phys->ds_unique_bytes);
1438 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_COMPRESSRATIO,
1439 		    ds->ds_phys->ds_compressed_bytes == 0 ? 100 :
1440 		    (ds->ds_phys->ds_uncompressed_bytes * 100 /
1441 		    ds->ds_phys->ds_compressed_bytes));
1442 	}
1443 }
1444 
1445 void
1446 dsl_dataset_fast_stat(dsl_dataset_t *ds, dmu_objset_stats_t *stat)
1447 {
1448 	stat->dds_creation_txg = ds->ds_phys->ds_creation_txg;
1449 	stat->dds_inconsistent = ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT;
1450 	if (ds->ds_phys->ds_next_snap_obj) {
1451 		stat->dds_is_snapshot = B_TRUE;
1452 		stat->dds_num_clones = ds->ds_phys->ds_num_children - 1;
1453 	}
1454 
1455 	/* clone origin is really a dsl_dir thing... */
1456 	if (ds->ds_dir->dd_phys->dd_clone_parent_obj) {
1457 		dsl_dataset_t *ods;
1458 
1459 		rw_enter(&ds->ds_dir->dd_pool->dp_config_rwlock, RW_READER);
1460 		VERIFY(0 == dsl_dataset_open_obj(ds->ds_dir->dd_pool,
1461 		    ds->ds_dir->dd_phys->dd_clone_parent_obj,
1462 		    NULL, DS_MODE_NONE, FTAG, &ods));
1463 		dsl_dataset_name(ods, stat->dds_clone_of);
1464 		dsl_dataset_close(ods, DS_MODE_NONE, FTAG);
1465 		rw_exit(&ds->ds_dir->dd_pool->dp_config_rwlock);
1466 	}
1467 }
1468 
1469 uint64_t
1470 dsl_dataset_fsid_guid(dsl_dataset_t *ds)
1471 {
1472 	return (ds->ds_phys->ds_fsid_guid);
1473 }
1474 
1475 void
1476 dsl_dataset_space(dsl_dataset_t *ds,
1477     uint64_t *refdbytesp, uint64_t *availbytesp,
1478     uint64_t *usedobjsp, uint64_t *availobjsp)
1479 {
1480 	*refdbytesp = ds->ds_phys->ds_used_bytes;
1481 	*availbytesp = dsl_dir_space_available(ds->ds_dir, NULL, 0, TRUE);
1482 	*usedobjsp = ds->ds_phys->ds_bp.blk_fill;
1483 	*availobjsp = DN_MAX_OBJECT - *usedobjsp;
1484 }
1485 
1486 /* ARGSUSED */
1487 static int
1488 dsl_dataset_snapshot_rename_check(void *arg1, void *arg2, dmu_tx_t *tx)
1489 {
1490 	dsl_dataset_t *ds = arg1;
1491 	char *newsnapname = arg2;
1492 	dsl_dir_t *dd = ds->ds_dir;
1493 	objset_t *mos = dd->dd_pool->dp_meta_objset;
1494 	dsl_dataset_t *hds;
1495 	uint64_t val;
1496 	int err;
1497 
1498 	err = dsl_dataset_open_obj(dd->dd_pool,
1499 	    dd->dd_phys->dd_head_dataset_obj, NULL, DS_MODE_NONE, FTAG, &hds);
1500 	if (err)
1501 		return (err);
1502 
1503 	/* new name better not be in use */
1504 	err = zap_lookup(mos, hds->ds_phys->ds_snapnames_zapobj,
1505 	    newsnapname, 8, 1, &val);
1506 	dsl_dataset_close(hds, DS_MODE_NONE, FTAG);
1507 
1508 	if (err == 0)
1509 		err = EEXIST;
1510 	else if (err == ENOENT)
1511 		err = 0;
1512 	return (err);
1513 }
1514 
1515 static void
1516 dsl_dataset_snapshot_rename_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1517 {
1518 	dsl_dataset_t *ds = arg1;
1519 	char *newsnapname = arg2;
1520 	dsl_dir_t *dd = ds->ds_dir;
1521 	objset_t *mos = dd->dd_pool->dp_meta_objset;
1522 	dsl_dataset_t *hds;
1523 	int err;
1524 
1525 	ASSERT(ds->ds_phys->ds_next_snap_obj != 0);
1526 
1527 	VERIFY(0 == dsl_dataset_open_obj(dd->dd_pool,
1528 	    dd->dd_phys->dd_head_dataset_obj, NULL, DS_MODE_NONE, FTAG, &hds));
1529 
1530 	VERIFY(0 == dsl_dataset_get_snapname(ds));
1531 	err = zap_remove(mos, hds->ds_phys->ds_snapnames_zapobj,
1532 	    ds->ds_snapname, tx);
1533 	ASSERT3U(err, ==, 0);
1534 	mutex_enter(&ds->ds_lock);
1535 	(void) strcpy(ds->ds_snapname, newsnapname);
1536 	mutex_exit(&ds->ds_lock);
1537 	err = zap_add(mos, hds->ds_phys->ds_snapnames_zapobj,
1538 	    ds->ds_snapname, 8, 1, &ds->ds_object, tx);
1539 	ASSERT3U(err, ==, 0);
1540 
1541 	dsl_dataset_close(hds, DS_MODE_NONE, FTAG);
1542 }
1543 
1544 #pragma weak dmu_objset_rename = dsl_dataset_rename
1545 int
1546 dsl_dataset_rename(const char *oldname, const char *newname)
1547 {
1548 	dsl_dir_t *dd;
1549 	dsl_dataset_t *ds;
1550 	const char *tail;
1551 	int err;
1552 
1553 	err = dsl_dir_open(oldname, FTAG, &dd, &tail);
1554 	if (err)
1555 		return (err);
1556 	if (tail == NULL) {
1557 		err = dsl_dir_rename(dd, newname);
1558 		dsl_dir_close(dd, FTAG);
1559 		return (err);
1560 	}
1561 	if (tail[0] != '@') {
1562 		/* the name ended in a nonexistant component */
1563 		dsl_dir_close(dd, FTAG);
1564 		return (ENOENT);
1565 	}
1566 
1567 	dsl_dir_close(dd, FTAG);
1568 
1569 	/* new name must be snapshot in same filesystem */
1570 	tail = strchr(newname, '@');
1571 	if (tail == NULL)
1572 		return (EINVAL);
1573 	tail++;
1574 	if (strncmp(oldname, newname, tail - newname) != 0)
1575 		return (EXDEV);
1576 
1577 	err = dsl_dataset_open(oldname,
1578 	    DS_MODE_READONLY | DS_MODE_STANDARD, FTAG, &ds);
1579 	if (err)
1580 		return (err);
1581 
1582 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1583 	    dsl_dataset_snapshot_rename_check,
1584 	    dsl_dataset_snapshot_rename_sync, ds, (char *)tail, 1);
1585 
1586 	dsl_dataset_close(ds, DS_MODE_STANDARD, FTAG);
1587 
1588 	return (err);
1589 }
1590 
1591 struct promotearg {
1592 	uint64_t used, comp, uncomp, unique;
1593 	uint64_t newnext_obj, snapnames_obj;
1594 };
1595 
1596 static int
1597 dsl_dataset_promote_check(void *arg1, void *arg2, dmu_tx_t *tx)
1598 {
1599 	dsl_dataset_t *hds = arg1;
1600 	struct promotearg *pa = arg2;
1601 	dsl_dir_t *dd = hds->ds_dir;
1602 	dsl_pool_t *dp = hds->ds_dir->dd_pool;
1603 	dsl_dir_t *pdd = NULL;
1604 	dsl_dataset_t *ds = NULL;
1605 	dsl_dataset_t *pivot_ds = NULL;
1606 	dsl_dataset_t *newnext_ds = NULL;
1607 	int err;
1608 	char *name = NULL;
1609 	uint64_t itor = 0;
1610 	blkptr_t bp;
1611 
1612 	bzero(pa, sizeof (*pa));
1613 
1614 	/* Check that it is a clone */
1615 	if (dd->dd_phys->dd_clone_parent_obj == 0)
1616 		return (EINVAL);
1617 
1618 	/* Since this is so expensive, don't do the preliminary check */
1619 	if (!dmu_tx_is_syncing(tx))
1620 		return (0);
1621 
1622 	if (err = dsl_dataset_open_obj(dp,
1623 	    dd->dd_phys->dd_clone_parent_obj,
1624 	    NULL, DS_MODE_EXCLUSIVE, FTAG, &pivot_ds))
1625 		goto out;
1626 	pdd = pivot_ds->ds_dir;
1627 
1628 	{
1629 		dsl_dataset_t *phds;
1630 		if (err = dsl_dataset_open_obj(dd->dd_pool,
1631 		    pdd->dd_phys->dd_head_dataset_obj,
1632 		    NULL, DS_MODE_NONE, FTAG, &phds))
1633 			goto out;
1634 		pa->snapnames_obj = phds->ds_phys->ds_snapnames_zapobj;
1635 		dsl_dataset_close(phds, DS_MODE_NONE, FTAG);
1636 	}
1637 
1638 	if (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE) {
1639 		err = EXDEV;
1640 		goto out;
1641 	}
1642 
1643 	/* find pivot point's new next ds */
1644 	VERIFY(0 == dsl_dataset_open_obj(dd->dd_pool, hds->ds_object,
1645 	    NULL, DS_MODE_NONE, FTAG, &newnext_ds));
1646 	while (newnext_ds->ds_phys->ds_prev_snap_obj != pivot_ds->ds_object) {
1647 		dsl_dataset_t *prev;
1648 
1649 		if (err = dsl_dataset_open_obj(dd->dd_pool,
1650 		    newnext_ds->ds_phys->ds_prev_snap_obj,
1651 		    NULL, DS_MODE_NONE, FTAG, &prev))
1652 			goto out;
1653 		dsl_dataset_close(newnext_ds, DS_MODE_NONE, FTAG);
1654 		newnext_ds = prev;
1655 	}
1656 	pa->newnext_obj = newnext_ds->ds_object;
1657 
1658 	/* compute pivot point's new unique space */
1659 	while ((err = bplist_iterate(&newnext_ds->ds_deadlist,
1660 	    &itor, &bp)) == 0) {
1661 		if (bp.blk_birth > pivot_ds->ds_phys->ds_prev_snap_txg)
1662 			pa->unique += bp_get_dasize(dd->dd_pool->dp_spa, &bp);
1663 	}
1664 	if (err != ENOENT)
1665 		goto out;
1666 
1667 	/* Walk the snapshots that we are moving */
1668 	name = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1669 	ds = pivot_ds;
1670 	/* CONSTCOND */
1671 	while (TRUE) {
1672 		uint64_t val, dlused, dlcomp, dluncomp;
1673 		dsl_dataset_t *prev;
1674 
1675 		/* Check that the snapshot name does not conflict */
1676 		dsl_dataset_name(ds, name);
1677 		err = zap_lookup(dd->dd_pool->dp_meta_objset,
1678 		    hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
1679 		    8, 1, &val);
1680 		if (err != ENOENT) {
1681 			if (err == 0)
1682 				err = EEXIST;
1683 			goto out;
1684 		}
1685 
1686 		/*
1687 		 * compute space to transfer.  Each snapshot gave birth to:
1688 		 * (my used) - (prev's used) + (deadlist's used)
1689 		 */
1690 		pa->used += ds->ds_phys->ds_used_bytes;
1691 		pa->comp += ds->ds_phys->ds_compressed_bytes;
1692 		pa->uncomp += ds->ds_phys->ds_uncompressed_bytes;
1693 
1694 		/* If we reach the first snapshot, we're done. */
1695 		if (ds->ds_phys->ds_prev_snap_obj == 0)
1696 			break;
1697 
1698 		if (err = bplist_space(&ds->ds_deadlist,
1699 		    &dlused, &dlcomp, &dluncomp))
1700 			goto out;
1701 		if (err = dsl_dataset_open_obj(dd->dd_pool,
1702 		    ds->ds_phys->ds_prev_snap_obj, NULL, DS_MODE_EXCLUSIVE,
1703 		    FTAG, &prev))
1704 			goto out;
1705 		pa->used += dlused - prev->ds_phys->ds_used_bytes;
1706 		pa->comp += dlcomp - prev->ds_phys->ds_compressed_bytes;
1707 		pa->uncomp += dluncomp - prev->ds_phys->ds_uncompressed_bytes;
1708 
1709 		/*
1710 		 * We could be a clone of a clone.  If we reach our
1711 		 * parent's branch point, we're done.
1712 		 */
1713 		if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1714 			dsl_dataset_close(prev, DS_MODE_EXCLUSIVE, FTAG);
1715 			break;
1716 		}
1717 		if (ds != pivot_ds)
1718 			dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
1719 		ds = prev;
1720 	}
1721 
1722 	/* Check that there is enough space here */
1723 	err = dsl_dir_transfer_possible(pdd, dd, pa->used);
1724 
1725 out:
1726 	if (ds && ds != pivot_ds)
1727 		dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
1728 	if (pivot_ds)
1729 		dsl_dataset_close(pivot_ds, DS_MODE_EXCLUSIVE, FTAG);
1730 	if (newnext_ds)
1731 		dsl_dataset_close(newnext_ds, DS_MODE_NONE, FTAG);
1732 	if (name)
1733 		kmem_free(name, MAXPATHLEN);
1734 	return (err);
1735 }
1736 
1737 static void
1738 dsl_dataset_promote_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1739 {
1740 	dsl_dataset_t *hds = arg1;
1741 	struct promotearg *pa = arg2;
1742 	dsl_dir_t *dd = hds->ds_dir;
1743 	dsl_pool_t *dp = hds->ds_dir->dd_pool;
1744 	dsl_dir_t *pdd = NULL;
1745 	dsl_dataset_t *ds, *pivot_ds;
1746 	char *name;
1747 
1748 	ASSERT(dd->dd_phys->dd_clone_parent_obj != 0);
1749 	ASSERT(0 == (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE));
1750 
1751 	VERIFY(0 == dsl_dataset_open_obj(dp,
1752 	    dd->dd_phys->dd_clone_parent_obj,
1753 	    NULL, DS_MODE_EXCLUSIVE, FTAG, &pivot_ds));
1754 	/*
1755 	 * We need to explicitly open pdd, since pivot_ds's pdd will be
1756 	 * changing.
1757 	 */
1758 	VERIFY(0 == dsl_dir_open_obj(dp, pivot_ds->ds_dir->dd_object,
1759 	    NULL, FTAG, &pdd));
1760 
1761 	/* move snapshots to this dir */
1762 	name = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1763 	ds = pivot_ds;
1764 	/* CONSTCOND */
1765 	while (TRUE) {
1766 		dsl_dataset_t *prev;
1767 
1768 		/* move snap name entry */
1769 		dsl_dataset_name(ds, name);
1770 		VERIFY(0 == zap_remove(dp->dp_meta_objset,
1771 		    pa->snapnames_obj, ds->ds_snapname, tx));
1772 		VERIFY(0 == zap_add(dp->dp_meta_objset,
1773 		    hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
1774 		    8, 1, &ds->ds_object, tx));
1775 
1776 		/* change containing dsl_dir */
1777 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
1778 		ASSERT3U(ds->ds_phys->ds_dir_obj, ==, pdd->dd_object);
1779 		ds->ds_phys->ds_dir_obj = dd->dd_object;
1780 		ASSERT3P(ds->ds_dir, ==, pdd);
1781 		dsl_dir_close(ds->ds_dir, ds);
1782 		VERIFY(0 == dsl_dir_open_obj(dp, dd->dd_object,
1783 		    NULL, ds, &ds->ds_dir));
1784 
1785 		ASSERT3U(dsl_prop_numcb(ds), ==, 0);
1786 
1787 		if (ds->ds_phys->ds_prev_snap_obj == 0)
1788 			break;
1789 
1790 		VERIFY(0 == dsl_dataset_open_obj(dp,
1791 		    ds->ds_phys->ds_prev_snap_obj, NULL, DS_MODE_EXCLUSIVE,
1792 		    FTAG, &prev));
1793 
1794 		if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1795 			dsl_dataset_close(prev, DS_MODE_EXCLUSIVE, FTAG);
1796 			break;
1797 		}
1798 		if (ds != pivot_ds)
1799 			dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
1800 		ds = prev;
1801 	}
1802 	if (ds != pivot_ds)
1803 		dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
1804 
1805 	/* change pivot point's next snap */
1806 	dmu_buf_will_dirty(pivot_ds->ds_dbuf, tx);
1807 	pivot_ds->ds_phys->ds_next_snap_obj = pa->newnext_obj;
1808 
1809 	/* change clone_parent-age */
1810 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
1811 	ASSERT3U(dd->dd_phys->dd_clone_parent_obj, ==, pivot_ds->ds_object);
1812 	dd->dd_phys->dd_clone_parent_obj = pdd->dd_phys->dd_clone_parent_obj;
1813 	dmu_buf_will_dirty(pdd->dd_dbuf, tx);
1814 	pdd->dd_phys->dd_clone_parent_obj = pivot_ds->ds_object;
1815 
1816 	/* change space accounting */
1817 	dsl_dir_diduse_space(pdd, -pa->used, -pa->comp, -pa->uncomp, tx);
1818 	dsl_dir_diduse_space(dd, pa->used, pa->comp, pa->uncomp, tx);
1819 	pivot_ds->ds_phys->ds_unique_bytes = pa->unique;
1820 
1821 	dsl_dir_close(pdd, FTAG);
1822 	dsl_dataset_close(pivot_ds, DS_MODE_EXCLUSIVE, FTAG);
1823 	kmem_free(name, MAXPATHLEN);
1824 }
1825 
1826 int
1827 dsl_dataset_promote(const char *name)
1828 {
1829 	dsl_dataset_t *ds;
1830 	int err;
1831 	dmu_object_info_t doi;
1832 	struct promotearg pa;
1833 
1834 	err = dsl_dataset_open(name, DS_MODE_NONE, FTAG, &ds);
1835 	if (err)
1836 		return (err);
1837 
1838 	err = dmu_object_info(ds->ds_dir->dd_pool->dp_meta_objset,
1839 	    ds->ds_phys->ds_snapnames_zapobj, &doi);
1840 	if (err) {
1841 		dsl_dataset_close(ds, DS_MODE_NONE, FTAG);
1842 		return (err);
1843 	}
1844 
1845 	/*
1846 	 * Add in 128x the snapnames zapobj size, since we will be moving
1847 	 * a bunch of snapnames to the promoted ds, and dirtying their
1848 	 * bonus buffers.
1849 	 */
1850 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1851 	    dsl_dataset_promote_check,
1852 	    dsl_dataset_promote_sync, ds, &pa, 2 + 2 * doi.doi_physical_blks);
1853 	dsl_dataset_close(ds, DS_MODE_NONE, FTAG);
1854 	return (err);
1855 }
1856