xref: /titanic_50/usr/src/uts/common/fs/zfs/dsl_dataset.c (revision 4088bb40326b75ef60834a6c2a92e29e25474b68)
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 2006 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	DOS_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 DOS_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 DOS_REF_MAX.  For example, EXCLUSIVE opens are exclusive because their
58  * weight (DOS_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,			/* DOS_MODE_NONE - invalid		*/
64 	1,			/* DOS_MODE_STANDARD - unlimited number	*/
65 	(DOS_REF_MAX >> 1) + 1,	/* DOS_MODE_PRIMARY - only one of these	*/
66 	DOS_REF_MAX		/* DOS_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 	if (ds == NULL)
185 		return (0);
186 	/*
187 	 * The snapshot creation could fail, but that would cause an
188 	 * incorrect FALSE return, which would only result in an
189 	 * overestimation of the amount of space that an operation would
190 	 * consume, which is OK.
191 	 *
192 	 * There's also a small window where we could miss a pending
193 	 * snapshot, because we could set the sync task in the quiescing
194 	 * phase.  So this should only be used as a guess.
195 	 */
196 	return (MAX(ds->ds_phys->ds_prev_snap_txg, ds->ds_trysnap_txg));
197 }
198 
199 int
200 dsl_dataset_block_freeable(dsl_dataset_t *ds, uint64_t blk_birth)
201 {
202 	return (blk_birth > dsl_dataset_prev_snap_txg(ds));
203 }
204 
205 /* ARGSUSED */
206 static void
207 dsl_dataset_evict(dmu_buf_t *db, void *dsv)
208 {
209 	dsl_dataset_t *ds = dsv;
210 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
211 
212 	/* open_refcount == DOS_REF_MAX when deleting */
213 	ASSERT(ds->ds_open_refcount == 0 ||
214 	    ds->ds_open_refcount == DOS_REF_MAX);
215 
216 	dprintf_ds(ds, "evicting %s\n", "");
217 
218 	unique_remove(ds->ds_phys->ds_fsid_guid);
219 
220 	if (ds->ds_user_ptr != NULL)
221 		ds->ds_user_evict_func(ds, ds->ds_user_ptr);
222 
223 	if (ds->ds_prev) {
224 		dsl_dataset_close(ds->ds_prev, DS_MODE_NONE, ds);
225 		ds->ds_prev = NULL;
226 	}
227 
228 	bplist_close(&ds->ds_deadlist);
229 	dsl_dir_close(ds->ds_dir, ds);
230 
231 	if (list_link_active(&ds->ds_synced_link))
232 		list_remove(&dp->dp_synced_objsets, ds);
233 
234 	kmem_free(ds, sizeof (dsl_dataset_t));
235 }
236 
237 static int
238 dsl_dataset_get_snapname(dsl_dataset_t *ds)
239 {
240 	dsl_dataset_phys_t *headphys;
241 	int err;
242 	dmu_buf_t *headdbuf;
243 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
244 	objset_t *mos = dp->dp_meta_objset;
245 
246 	if (ds->ds_snapname[0])
247 		return (0);
248 	if (ds->ds_phys->ds_next_snap_obj == 0)
249 		return (0);
250 
251 	err = dmu_bonus_hold(mos, ds->ds_dir->dd_phys->dd_head_dataset_obj,
252 	    FTAG, &headdbuf);
253 	if (err)
254 		return (err);
255 	headphys = headdbuf->db_data;
256 	err = zap_value_search(dp->dp_meta_objset,
257 	    headphys->ds_snapnames_zapobj, ds->ds_object, ds->ds_snapname);
258 	dmu_buf_rele(headdbuf, FTAG);
259 	return (err);
260 }
261 
262 int
263 dsl_dataset_open_obj(dsl_pool_t *dp, uint64_t dsobj, const char *snapname,
264     int mode, void *tag, dsl_dataset_t **dsp)
265 {
266 	uint64_t weight = ds_refcnt_weight[DS_MODE_LEVEL(mode)];
267 	objset_t *mos = dp->dp_meta_objset;
268 	dmu_buf_t *dbuf;
269 	dsl_dataset_t *ds;
270 	int err;
271 
272 	ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
273 	    dsl_pool_sync_context(dp));
274 
275 	err = dmu_bonus_hold(mos, dsobj, tag, &dbuf);
276 	if (err)
277 		return (err);
278 	ds = dmu_buf_get_user(dbuf);
279 	if (ds == NULL) {
280 		dsl_dataset_t *winner;
281 
282 		ds = kmem_zalloc(sizeof (dsl_dataset_t), KM_SLEEP);
283 		ds->ds_dbuf = dbuf;
284 		ds->ds_object = dsobj;
285 		ds->ds_phys = dbuf->db_data;
286 
287 		err = bplist_open(&ds->ds_deadlist,
288 		    mos, ds->ds_phys->ds_deadlist_obj);
289 		if (err == 0) {
290 			err = dsl_dir_open_obj(dp,
291 			    ds->ds_phys->ds_dir_obj, NULL, ds, &ds->ds_dir);
292 		}
293 		if (err) {
294 			/*
295 			 * we don't really need to close the blist if we
296 			 * just opened it.
297 			 */
298 			kmem_free(ds, sizeof (dsl_dataset_t));
299 			dmu_buf_rele(dbuf, tag);
300 			return (err);
301 		}
302 
303 		if (ds->ds_dir->dd_phys->dd_head_dataset_obj == dsobj) {
304 			ds->ds_snapname[0] = '\0';
305 			if (ds->ds_phys->ds_prev_snap_obj) {
306 				err = dsl_dataset_open_obj(dp,
307 				    ds->ds_phys->ds_prev_snap_obj, NULL,
308 				    DS_MODE_NONE, ds, &ds->ds_prev);
309 			}
310 		} else {
311 			if (snapname) {
312 #ifdef ZFS_DEBUG
313 				dsl_dataset_phys_t *headphys;
314 				dmu_buf_t *headdbuf;
315 				err = dmu_bonus_hold(mos,
316 				    ds->ds_dir->dd_phys->dd_head_dataset_obj,
317 				    FTAG, &headdbuf);
318 				if (err == 0) {
319 					headphys = headdbuf->db_data;
320 					uint64_t foundobj;
321 					err = zap_lookup(dp->dp_meta_objset,
322 					    headphys->ds_snapnames_zapobj,
323 					    snapname, sizeof (foundobj), 1,
324 					    &foundobj);
325 					ASSERT3U(foundobj, ==, dsobj);
326 					dmu_buf_rele(headdbuf, FTAG);
327 				}
328 #endif
329 				(void) strcat(ds->ds_snapname, snapname);
330 			} else if (zfs_flags & ZFS_DEBUG_SNAPNAMES) {
331 				err = dsl_dataset_get_snapname(ds);
332 			}
333 		}
334 
335 		if (err == 0) {
336 			winner = dmu_buf_set_user_ie(dbuf, ds, &ds->ds_phys,
337 			    dsl_dataset_evict);
338 		}
339 		if (err || winner) {
340 			bplist_close(&ds->ds_deadlist);
341 			if (ds->ds_prev) {
342 				dsl_dataset_close(ds->ds_prev,
343 				    DS_MODE_NONE, ds);
344 			}
345 			dsl_dir_close(ds->ds_dir, ds);
346 			kmem_free(ds, sizeof (dsl_dataset_t));
347 			if (err) {
348 				dmu_buf_rele(dbuf, tag);
349 				return (err);
350 			}
351 			ds = winner;
352 		} else {
353 			uint64_t new =
354 			    unique_insert(ds->ds_phys->ds_fsid_guid);
355 			if (new != ds->ds_phys->ds_fsid_guid) {
356 				/* XXX it won't necessarily be synced... */
357 				ds->ds_phys->ds_fsid_guid = new;
358 			}
359 		}
360 	}
361 	ASSERT3P(ds->ds_dbuf, ==, dbuf);
362 	ASSERT3P(ds->ds_phys, ==, dbuf->db_data);
363 
364 	mutex_enter(&ds->ds_lock);
365 	if ((DS_MODE_LEVEL(mode) == DS_MODE_PRIMARY &&
366 	    (ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT) &&
367 	    !DS_MODE_IS_INCONSISTENT(mode)) ||
368 	    (ds->ds_open_refcount + weight > DOS_REF_MAX)) {
369 		mutex_exit(&ds->ds_lock);
370 		dsl_dataset_close(ds, DS_MODE_NONE, tag);
371 		return (EBUSY);
372 	}
373 	ds->ds_open_refcount += weight;
374 	mutex_exit(&ds->ds_lock);
375 
376 	*dsp = ds;
377 	return (0);
378 }
379 
380 int
381 dsl_dataset_open_spa(spa_t *spa, const char *name, int mode,
382     void *tag, dsl_dataset_t **dsp)
383 {
384 	dsl_dir_t *dd;
385 	dsl_pool_t *dp;
386 	const char *tail;
387 	uint64_t obj;
388 	dsl_dataset_t *ds = NULL;
389 	int err = 0;
390 
391 	err = dsl_dir_open_spa(spa, name, FTAG, &dd, &tail);
392 	if (err)
393 		return (err);
394 
395 	dp = dd->dd_pool;
396 	obj = dd->dd_phys->dd_head_dataset_obj;
397 	rw_enter(&dp->dp_config_rwlock, RW_READER);
398 	if (obj == 0) {
399 		/* A dataset with no associated objset */
400 		err = ENOENT;
401 		goto out;
402 	}
403 
404 	if (tail != NULL) {
405 		objset_t *mos = dp->dp_meta_objset;
406 
407 		err = dsl_dataset_open_obj(dp, obj, NULL,
408 		    DS_MODE_NONE, tag, &ds);
409 		if (err)
410 			goto out;
411 		obj = ds->ds_phys->ds_snapnames_zapobj;
412 		dsl_dataset_close(ds, DS_MODE_NONE, tag);
413 		ds = NULL;
414 
415 		if (tail[0] != '@') {
416 			err = ENOENT;
417 			goto out;
418 		}
419 		tail++;
420 
421 		/* Look for a snapshot */
422 		if (!DS_MODE_IS_READONLY(mode)) {
423 			err = EROFS;
424 			goto out;
425 		}
426 		dprintf("looking for snapshot '%s'\n", tail);
427 		err = zap_lookup(mos, obj, tail, 8, 1, &obj);
428 		if (err)
429 			goto out;
430 	}
431 	err = dsl_dataset_open_obj(dp, obj, tail, mode, tag, &ds);
432 
433 out:
434 	rw_exit(&dp->dp_config_rwlock);
435 	dsl_dir_close(dd, FTAG);
436 
437 	ASSERT3U((err == 0), ==, (ds != NULL));
438 	/* ASSERT(ds == NULL || strcmp(name, ds->ds_name) == 0); */
439 
440 	*dsp = ds;
441 	return (err);
442 }
443 
444 int
445 dsl_dataset_open(const char *name, int mode, void *tag, dsl_dataset_t **dsp)
446 {
447 	return (dsl_dataset_open_spa(NULL, name, mode, tag, dsp));
448 }
449 
450 void
451 dsl_dataset_name(dsl_dataset_t *ds, char *name)
452 {
453 	if (ds == NULL) {
454 		(void) strcpy(name, "mos");
455 	} else {
456 		dsl_dir_name(ds->ds_dir, name);
457 		VERIFY(0 == dsl_dataset_get_snapname(ds));
458 		if (ds->ds_snapname[0]) {
459 			(void) strcat(name, "@");
460 			if (!MUTEX_HELD(&ds->ds_lock)) {
461 				/*
462 				 * We use a "recursive" mutex so that we
463 				 * can call dprintf_ds() with ds_lock held.
464 				 */
465 				mutex_enter(&ds->ds_lock);
466 				(void) strcat(name, ds->ds_snapname);
467 				mutex_exit(&ds->ds_lock);
468 			} else {
469 				(void) strcat(name, ds->ds_snapname);
470 			}
471 		}
472 	}
473 }
474 
475 void
476 dsl_dataset_close(dsl_dataset_t *ds, int mode, void *tag)
477 {
478 	uint64_t weight = ds_refcnt_weight[DS_MODE_LEVEL(mode)];
479 	mutex_enter(&ds->ds_lock);
480 	ASSERT3U(ds->ds_open_refcount, >=, weight);
481 	ds->ds_open_refcount -= weight;
482 	dprintf_ds(ds, "closing mode %u refcount now 0x%llx\n",
483 	    mode, ds->ds_open_refcount);
484 	mutex_exit(&ds->ds_lock);
485 
486 	dmu_buf_rele(ds->ds_dbuf, tag);
487 }
488 
489 void
490 dsl_dataset_create_root(dsl_pool_t *dp, uint64_t *ddobjp, dmu_tx_t *tx)
491 {
492 	objset_t *mos = dp->dp_meta_objset;
493 	dmu_buf_t *dbuf;
494 	dsl_dataset_phys_t *dsphys;
495 	dsl_dataset_t *ds;
496 	uint64_t dsobj;
497 	dsl_dir_t *dd;
498 
499 	dsl_dir_create_root(mos, ddobjp, tx);
500 	VERIFY(0 == dsl_dir_open_obj(dp, *ddobjp, NULL, FTAG, &dd));
501 
502 	dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
503 	    DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
504 	VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
505 	dmu_buf_will_dirty(dbuf, tx);
506 	dsphys = dbuf->db_data;
507 	dsphys->ds_dir_obj = dd->dd_object;
508 	dsphys->ds_fsid_guid = unique_create();
509 	unique_remove(dsphys->ds_fsid_guid); /* it isn't open yet */
510 	(void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
511 	    sizeof (dsphys->ds_guid));
512 	dsphys->ds_snapnames_zapobj =
513 	    zap_create(mos, DMU_OT_DSL_DS_SNAP_MAP, DMU_OT_NONE, 0, tx);
514 	dsphys->ds_creation_time = gethrestime_sec();
515 	dsphys->ds_creation_txg = tx->tx_txg;
516 	dsphys->ds_deadlist_obj =
517 	    bplist_create(mos, DSL_DEADLIST_BLOCKSIZE, tx);
518 	dmu_buf_rele(dbuf, FTAG);
519 
520 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
521 	dd->dd_phys->dd_head_dataset_obj = dsobj;
522 	dsl_dir_close(dd, FTAG);
523 
524 	VERIFY(0 ==
525 	    dsl_dataset_open_obj(dp, dsobj, NULL, DS_MODE_NONE, FTAG, &ds));
526 	(void) dmu_objset_create_impl(dp->dp_spa, ds, DMU_OST_ZFS, tx);
527 	dsl_dataset_close(ds, DS_MODE_NONE, FTAG);
528 }
529 
530 uint64_t
531 dsl_dataset_create_sync(dsl_dir_t *pdd,
532     const char *lastname, dsl_dataset_t *clone_parent, dmu_tx_t *tx)
533 {
534 	dsl_pool_t *dp = pdd->dd_pool;
535 	dmu_buf_t *dbuf;
536 	dsl_dataset_phys_t *dsphys;
537 	uint64_t dsobj, ddobj;
538 	objset_t *mos = dp->dp_meta_objset;
539 	dsl_dir_t *dd;
540 
541 	ASSERT(clone_parent == NULL || clone_parent->ds_dir->dd_pool == dp);
542 	ASSERT(clone_parent == NULL ||
543 	    clone_parent->ds_phys->ds_num_children > 0);
544 	ASSERT(lastname[0] != '@');
545 	ASSERT(dmu_tx_is_syncing(tx));
546 
547 	ddobj = dsl_dir_create_sync(pdd, lastname, tx);
548 	VERIFY(0 == dsl_dir_open_obj(dp, ddobj, lastname, FTAG, &dd));
549 
550 	dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
551 	    DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
552 	VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
553 	dmu_buf_will_dirty(dbuf, tx);
554 	dsphys = dbuf->db_data;
555 	dsphys->ds_dir_obj = dd->dd_object;
556 	dsphys->ds_fsid_guid = unique_create();
557 	unique_remove(dsphys->ds_fsid_guid); /* it isn't open yet */
558 	(void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
559 	    sizeof (dsphys->ds_guid));
560 	dsphys->ds_snapnames_zapobj =
561 	    zap_create(mos, DMU_OT_DSL_DS_SNAP_MAP, DMU_OT_NONE, 0, tx);
562 	dsphys->ds_creation_time = gethrestime_sec();
563 	dsphys->ds_creation_txg = tx->tx_txg;
564 	dsphys->ds_deadlist_obj =
565 	    bplist_create(mos, DSL_DEADLIST_BLOCKSIZE, tx);
566 	if (clone_parent) {
567 		dsphys->ds_prev_snap_obj = clone_parent->ds_object;
568 		dsphys->ds_prev_snap_txg =
569 		    clone_parent->ds_phys->ds_creation_txg;
570 		dsphys->ds_used_bytes =
571 		    clone_parent->ds_phys->ds_used_bytes;
572 		dsphys->ds_compressed_bytes =
573 		    clone_parent->ds_phys->ds_compressed_bytes;
574 		dsphys->ds_uncompressed_bytes =
575 		    clone_parent->ds_phys->ds_uncompressed_bytes;
576 		dsphys->ds_bp = clone_parent->ds_phys->ds_bp;
577 
578 		dmu_buf_will_dirty(clone_parent->ds_dbuf, tx);
579 		clone_parent->ds_phys->ds_num_children++;
580 
581 		dmu_buf_will_dirty(dd->dd_dbuf, tx);
582 		dd->dd_phys->dd_clone_parent_obj = clone_parent->ds_object;
583 	}
584 	dmu_buf_rele(dbuf, FTAG);
585 
586 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
587 	dd->dd_phys->dd_head_dataset_obj = dsobj;
588 	dsl_dir_close(dd, FTAG);
589 
590 	return (dsobj);
591 }
592 
593 struct destroyarg {
594 	dsl_sync_task_group_t *dstg;
595 	char *snapname;
596 	void *tag;
597 	char *failed;
598 };
599 
600 static int
601 dsl_snapshot_destroy_one(char *name, void *arg)
602 {
603 	struct destroyarg *da = arg;
604 	dsl_dataset_t *ds;
605 	char *cp;
606 	int err;
607 
608 	(void) strcat(name, "@");
609 	(void) strcat(name, da->snapname);
610 	err = dsl_dataset_open(name,
611 	    DS_MODE_EXCLUSIVE | DS_MODE_READONLY | DS_MODE_INCONSISTENT,
612 	    da->tag, &ds);
613 	cp = strchr(name, '@');
614 	*cp = '\0';
615 	if (err == ENOENT)
616 		return (0);
617 	if (err) {
618 		(void) strcpy(da->failed, name);
619 		return (err);
620 	}
621 
622 	dsl_sync_task_create(da->dstg, dsl_dataset_destroy_check,
623 	    dsl_dataset_destroy_sync, ds, da->tag, 0);
624 	return (0);
625 }
626 
627 /*
628  * Destroy 'snapname' in all descendants of 'fsname'.
629  */
630 #pragma weak dmu_snapshots_destroy = dsl_snapshots_destroy
631 int
632 dsl_snapshots_destroy(char *fsname, char *snapname)
633 {
634 	int err;
635 	struct destroyarg da;
636 	dsl_sync_task_t *dst;
637 	spa_t *spa;
638 	char *cp;
639 
640 	cp = strchr(fsname, '/');
641 	if (cp) {
642 		*cp = '\0';
643 		err = spa_open(fsname, &spa, FTAG);
644 		*cp = '/';
645 	} else {
646 		err = spa_open(fsname, &spa, FTAG);
647 	}
648 	if (err)
649 		return (err);
650 	da.dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
651 	da.snapname = snapname;
652 	da.tag = FTAG;
653 	da.failed = fsname;
654 
655 	err = dmu_objset_find(fsname,
656 	    dsl_snapshot_destroy_one, &da, DS_FIND_CHILDREN);
657 
658 	if (err == 0)
659 		err = dsl_sync_task_group_wait(da.dstg);
660 
661 	for (dst = list_head(&da.dstg->dstg_tasks); dst;
662 	    dst = list_next(&da.dstg->dstg_tasks, dst)) {
663 		dsl_dataset_t *ds = dst->dst_arg1;
664 		if (dst->dst_err) {
665 			dsl_dataset_name(ds, fsname);
666 			cp = strchr(fsname, '@');
667 			*cp = '\0';
668 		}
669 		/*
670 		 * If it was successful, destroy_sync would have
671 		 * closed the ds
672 		 */
673 		if (err)
674 			dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
675 	}
676 
677 	dsl_sync_task_group_destroy(da.dstg);
678 	spa_close(spa, FTAG);
679 	return (err);
680 }
681 
682 int
683 dsl_dataset_destroy(const char *name)
684 {
685 	int err;
686 	dsl_sync_task_group_t *dstg;
687 	objset_t *os;
688 	dsl_dataset_t *ds;
689 	dsl_dir_t *dd;
690 	uint64_t obj;
691 
692 	if (strchr(name, '@')) {
693 		/* Destroying a snapshot is simpler */
694 		err = dsl_dataset_open(name,
695 		    DS_MODE_EXCLUSIVE | DS_MODE_READONLY | DS_MODE_INCONSISTENT,
696 		    FTAG, &ds);
697 		if (err)
698 			return (err);
699 		err = dsl_sync_task_do(ds->ds_dir->dd_pool,
700 		    dsl_dataset_destroy_check, dsl_dataset_destroy_sync,
701 		    ds, FTAG, 0);
702 		if (err)
703 			dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
704 		return (err);
705 	}
706 
707 	err = dmu_objset_open(name, DMU_OST_ANY,
708 	    DS_MODE_EXCLUSIVE | DS_MODE_INCONSISTENT, &os);
709 	if (err)
710 		return (err);
711 	ds = os->os->os_dsl_dataset;
712 	dd = ds->ds_dir;
713 
714 	/*
715 	 * Check for errors and mark this ds as inconsistent, in
716 	 * case we crash while freeing the objects.
717 	 */
718 	err = dsl_sync_task_do(dd->dd_pool, dsl_dataset_destroy_begin_check,
719 	    dsl_dataset_destroy_begin_sync, ds, NULL, 0);
720 	if (err) {
721 		dmu_objset_close(os);
722 		return (err);
723 	}
724 
725 	/*
726 	 * remove the objects in open context, so that we won't
727 	 * have too much to do in syncing context.
728 	 */
729 	for (obj = 0; err == 0;
730 	    err = dmu_object_next(os, &obj, FALSE)) {
731 		dmu_tx_t *tx = dmu_tx_create(os);
732 		dmu_tx_hold_free(tx, obj, 0, DMU_OBJECT_END);
733 		dmu_tx_hold_bonus(tx, obj);
734 		err = dmu_tx_assign(tx, TXG_WAIT);
735 		if (err) {
736 			/*
737 			 * Perhaps there is not enough disk
738 			 * space.  Just deal with it from
739 			 * dsl_dataset_destroy_sync().
740 			 */
741 			dmu_tx_abort(tx);
742 			continue;
743 		}
744 		VERIFY(0 == dmu_object_free(os, obj, tx));
745 		dmu_tx_commit(tx);
746 	}
747 	/* Make sure it's not dirty before we finish destroying it. */
748 	txg_wait_synced(dd->dd_pool, 0);
749 
750 	dmu_objset_close(os);
751 	if (err != ESRCH)
752 		return (err);
753 
754 	err = dsl_dataset_open(name,
755 	    DS_MODE_EXCLUSIVE | DS_MODE_READONLY | DS_MODE_INCONSISTENT,
756 	    FTAG, &ds);
757 	if (err)
758 		return (err);
759 
760 	err = dsl_dir_open(name, FTAG, &dd, NULL);
761 	if (err) {
762 		dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
763 		return (err);
764 	}
765 
766 	/*
767 	 * Blow away the dsl_dir + head dataset.
768 	 */
769 	dstg = dsl_sync_task_group_create(ds->ds_dir->dd_pool);
770 	dsl_sync_task_create(dstg, dsl_dataset_destroy_check,
771 	    dsl_dataset_destroy_sync, ds, FTAG, 0);
772 	dsl_sync_task_create(dstg, dsl_dir_destroy_check,
773 	    dsl_dir_destroy_sync, dd, FTAG, 0);
774 	err = dsl_sync_task_group_wait(dstg);
775 	dsl_sync_task_group_destroy(dstg);
776 	/* if it is successful, *destroy_sync will close the ds+dd */
777 	if (err) {
778 		dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
779 		dsl_dir_close(dd, FTAG);
780 	}
781 	return (err);
782 }
783 
784 int
785 dsl_dataset_rollback(dsl_dataset_t *ds)
786 {
787 	ASSERT3U(ds->ds_open_refcount, ==, DOS_REF_MAX);
788 	return (dsl_sync_task_do(ds->ds_dir->dd_pool,
789 	    dsl_dataset_rollback_check, dsl_dataset_rollback_sync,
790 	    ds, NULL, 0));
791 }
792 
793 void *
794 dsl_dataset_set_user_ptr(dsl_dataset_t *ds,
795     void *p, dsl_dataset_evict_func_t func)
796 {
797 	void *old;
798 
799 	mutex_enter(&ds->ds_lock);
800 	old = ds->ds_user_ptr;
801 	if (old == NULL) {
802 		ds->ds_user_ptr = p;
803 		ds->ds_user_evict_func = func;
804 	}
805 	mutex_exit(&ds->ds_lock);
806 	return (old);
807 }
808 
809 void *
810 dsl_dataset_get_user_ptr(dsl_dataset_t *ds)
811 {
812 	return (ds->ds_user_ptr);
813 }
814 
815 
816 void
817 dsl_dataset_get_blkptr(dsl_dataset_t *ds, blkptr_t *bp)
818 {
819 	*bp = ds->ds_phys->ds_bp;
820 }
821 
822 void
823 dsl_dataset_set_blkptr(dsl_dataset_t *ds, blkptr_t *bp, dmu_tx_t *tx)
824 {
825 	ASSERT(dmu_tx_is_syncing(tx));
826 	/* If it's the meta-objset, set dp_meta_rootbp */
827 	if (ds == NULL) {
828 		tx->tx_pool->dp_meta_rootbp = *bp;
829 	} else {
830 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
831 		ds->ds_phys->ds_bp = *bp;
832 	}
833 }
834 
835 spa_t *
836 dsl_dataset_get_spa(dsl_dataset_t *ds)
837 {
838 	return (ds->ds_dir->dd_pool->dp_spa);
839 }
840 
841 void
842 dsl_dataset_dirty(dsl_dataset_t *ds, dmu_tx_t *tx)
843 {
844 	dsl_pool_t *dp;
845 
846 	if (ds == NULL) /* this is the meta-objset */
847 		return;
848 
849 	ASSERT(ds->ds_user_ptr != NULL);
850 	ASSERT(ds->ds_phys->ds_next_snap_obj == 0);
851 
852 	dp = ds->ds_dir->dd_pool;
853 
854 	if (txg_list_add(&dp->dp_dirty_datasets, ds, tx->tx_txg) == 0) {
855 		/* up the hold count until we can be written out */
856 		dmu_buf_add_ref(ds->ds_dbuf, ds);
857 	}
858 }
859 
860 struct killarg {
861 	uint64_t *usedp;
862 	uint64_t *compressedp;
863 	uint64_t *uncompressedp;
864 	zio_t *zio;
865 	dmu_tx_t *tx;
866 };
867 
868 static int
869 kill_blkptr(traverse_blk_cache_t *bc, spa_t *spa, void *arg)
870 {
871 	struct killarg *ka = arg;
872 	blkptr_t *bp = &bc->bc_blkptr;
873 
874 	ASSERT3U(bc->bc_errno, ==, 0);
875 
876 	/*
877 	 * Since this callback is not called concurrently, no lock is
878 	 * needed on the accounting values.
879 	 */
880 	*ka->usedp += bp_get_dasize(spa, bp);
881 	*ka->compressedp += BP_GET_PSIZE(bp);
882 	*ka->uncompressedp += BP_GET_UCSIZE(bp);
883 	/* XXX check for EIO? */
884 	(void) arc_free(ka->zio, spa, ka->tx->tx_txg, bp, NULL, NULL,
885 	    ARC_NOWAIT);
886 	return (0);
887 }
888 
889 /* ARGSUSED */
890 static int
891 dsl_dataset_rollback_check(void *arg1, void *arg2, dmu_tx_t *tx)
892 {
893 	dsl_dataset_t *ds = arg1;
894 
895 	/*
896 	 * There must be a previous snapshot.  I suppose we could roll
897 	 * it back to being empty (and re-initialize the upper (ZPL)
898 	 * layer).  But for now there's no way to do this via the user
899 	 * interface.
900 	 */
901 	if (ds->ds_phys->ds_prev_snap_txg == 0)
902 		return (EINVAL);
903 
904 	/*
905 	 * This must not be a snapshot.
906 	 */
907 	if (ds->ds_phys->ds_next_snap_obj != 0)
908 		return (EINVAL);
909 
910 	/*
911 	 * If we made changes this txg, traverse_dsl_dataset won't find
912 	 * them.  Try again.
913 	 */
914 	if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
915 		return (EAGAIN);
916 
917 	return (0);
918 }
919 
920 /* ARGSUSED */
921 static void
922 dsl_dataset_rollback_sync(void *arg1, void *arg2, dmu_tx_t *tx)
923 {
924 	dsl_dataset_t *ds = arg1;
925 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
926 
927 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
928 
929 	/* Zero out the deadlist. */
930 	bplist_close(&ds->ds_deadlist);
931 	bplist_destroy(mos, ds->ds_phys->ds_deadlist_obj, tx);
932 	ds->ds_phys->ds_deadlist_obj =
933 	    bplist_create(mos, DSL_DEADLIST_BLOCKSIZE, tx);
934 	VERIFY(0 == bplist_open(&ds->ds_deadlist, mos,
935 	    ds->ds_phys->ds_deadlist_obj));
936 
937 	{
938 		/* Free blkptrs that we gave birth to */
939 		zio_t *zio;
940 		uint64_t used = 0, compressed = 0, uncompressed = 0;
941 		struct killarg ka;
942 
943 		zio = zio_root(tx->tx_pool->dp_spa, NULL, NULL,
944 		    ZIO_FLAG_MUSTSUCCEED);
945 		ka.usedp = &used;
946 		ka.compressedp = &compressed;
947 		ka.uncompressedp = &uncompressed;
948 		ka.zio = zio;
949 		ka.tx = tx;
950 		(void) traverse_dsl_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
951 		    ADVANCE_POST, kill_blkptr, &ka);
952 		(void) zio_wait(zio);
953 
954 		dsl_dir_diduse_space(ds->ds_dir,
955 		    -used, -compressed, -uncompressed, tx);
956 	}
957 
958 	/* Change our contents to that of the prev snapshot */
959 	ASSERT3U(ds->ds_prev->ds_object, ==, ds->ds_phys->ds_prev_snap_obj);
960 	ds->ds_phys->ds_bp = ds->ds_prev->ds_phys->ds_bp;
961 	ds->ds_phys->ds_used_bytes = ds->ds_prev->ds_phys->ds_used_bytes;
962 	ds->ds_phys->ds_compressed_bytes =
963 	    ds->ds_prev->ds_phys->ds_compressed_bytes;
964 	ds->ds_phys->ds_uncompressed_bytes =
965 	    ds->ds_prev->ds_phys->ds_uncompressed_bytes;
966 	ds->ds_phys->ds_flags = ds->ds_prev->ds_phys->ds_flags;
967 	ds->ds_phys->ds_unique_bytes = 0;
968 
969 	if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
970 		dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
971 		ds->ds_prev->ds_phys->ds_unique_bytes = 0;
972 	}
973 }
974 
975 /* ARGSUSED */
976 static int
977 dsl_dataset_destroy_begin_check(void *arg1, void *arg2, dmu_tx_t *tx)
978 {
979 	dsl_dataset_t *ds = arg1;
980 
981 	/*
982 	 * Can't delete a head dataset if there are snapshots of it.
983 	 * (Except if the only snapshots are from the branch we cloned
984 	 * from.)
985 	 */
986 	if (ds->ds_prev != NULL &&
987 	    ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
988 		return (EINVAL);
989 
990 	return (0);
991 }
992 
993 /* ARGSUSED */
994 static void
995 dsl_dataset_destroy_begin_sync(void *arg1, void *arg2, dmu_tx_t *tx)
996 {
997 	dsl_dataset_t *ds = arg1;
998 
999 	/* Mark it as inconsistent on-disk, in case we crash */
1000 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1001 	ds->ds_phys->ds_flags |= DS_FLAG_INCONSISTENT;
1002 }
1003 
1004 /* ARGSUSED */
1005 static int
1006 dsl_dataset_destroy_check(void *arg1, void *arg2, dmu_tx_t *tx)
1007 {
1008 	dsl_dataset_t *ds = arg1;
1009 
1010 	/* Can't delete a branch point. */
1011 	if (ds->ds_phys->ds_num_children > 1)
1012 		return (EEXIST);
1013 
1014 	/*
1015 	 * Can't delete a head dataset if there are snapshots of it.
1016 	 * (Except if the only snapshots are from the branch we cloned
1017 	 * from.)
1018 	 */
1019 	if (ds->ds_prev != NULL &&
1020 	    ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object)
1021 		return (EINVAL);
1022 
1023 	/*
1024 	 * If we made changes this txg, traverse_dsl_dataset won't find
1025 	 * them.  Try again.
1026 	 */
1027 	if (ds->ds_phys->ds_bp.blk_birth >= tx->tx_txg)
1028 		return (EAGAIN);
1029 
1030 	/* XXX we should do some i/o error checking... */
1031 	return (0);
1032 }
1033 
1034 static void
1035 dsl_dataset_destroy_sync(void *arg1, void *tag, dmu_tx_t *tx)
1036 {
1037 	dsl_dataset_t *ds = arg1;
1038 	uint64_t used = 0, compressed = 0, uncompressed = 0;
1039 	zio_t *zio;
1040 	int err;
1041 	int after_branch_point = FALSE;
1042 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1043 	objset_t *mos = dp->dp_meta_objset;
1044 	dsl_dataset_t *ds_prev = NULL;
1045 	uint64_t obj;
1046 
1047 	ASSERT3U(ds->ds_open_refcount, ==, DOS_REF_MAX);
1048 	ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
1049 	ASSERT(ds->ds_prev == NULL ||
1050 	    ds->ds_prev->ds_phys->ds_next_snap_obj != ds->ds_object);
1051 	ASSERT3U(ds->ds_phys->ds_bp.blk_birth, <=, tx->tx_txg);
1052 
1053 	ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1054 
1055 	obj = ds->ds_object;
1056 
1057 	if (ds->ds_phys->ds_prev_snap_obj != 0) {
1058 		if (ds->ds_prev) {
1059 			ds_prev = ds->ds_prev;
1060 		} else {
1061 			VERIFY(0 == dsl_dataset_open_obj(dp,
1062 			    ds->ds_phys->ds_prev_snap_obj, NULL,
1063 			    DS_MODE_NONE, FTAG, &ds_prev));
1064 		}
1065 		after_branch_point =
1066 		    (ds_prev->ds_phys->ds_next_snap_obj != obj);
1067 
1068 		dmu_buf_will_dirty(ds_prev->ds_dbuf, tx);
1069 		if (after_branch_point &&
1070 		    ds->ds_phys->ds_next_snap_obj == 0) {
1071 			/* This clone is toast. */
1072 			ASSERT(ds_prev->ds_phys->ds_num_children > 1);
1073 			ds_prev->ds_phys->ds_num_children--;
1074 		} else if (!after_branch_point) {
1075 			ds_prev->ds_phys->ds_next_snap_obj =
1076 			    ds->ds_phys->ds_next_snap_obj;
1077 		}
1078 	}
1079 
1080 	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
1081 
1082 	if (ds->ds_phys->ds_next_snap_obj != 0) {
1083 		blkptr_t bp;
1084 		dsl_dataset_t *ds_next;
1085 		uint64_t itor = 0;
1086 
1087 		spa_scrub_restart(dp->dp_spa, tx->tx_txg);
1088 
1089 		VERIFY(0 == dsl_dataset_open_obj(dp,
1090 		    ds->ds_phys->ds_next_snap_obj, NULL,
1091 		    DS_MODE_NONE, FTAG, &ds_next));
1092 		ASSERT3U(ds_next->ds_phys->ds_prev_snap_obj, ==, obj);
1093 
1094 		dmu_buf_will_dirty(ds_next->ds_dbuf, tx);
1095 		ds_next->ds_phys->ds_prev_snap_obj =
1096 		    ds->ds_phys->ds_prev_snap_obj;
1097 		ds_next->ds_phys->ds_prev_snap_txg =
1098 		    ds->ds_phys->ds_prev_snap_txg;
1099 		ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1100 		    ds_prev ? ds_prev->ds_phys->ds_creation_txg : 0);
1101 
1102 		/*
1103 		 * Transfer to our deadlist (which will become next's
1104 		 * new deadlist) any entries from next's current
1105 		 * deadlist which were born before prev, and free the
1106 		 * other entries.
1107 		 *
1108 		 * XXX we're doing this long task with the config lock held
1109 		 */
1110 		while (bplist_iterate(&ds_next->ds_deadlist, &itor,
1111 		    &bp) == 0) {
1112 			if (bp.blk_birth <= ds->ds_phys->ds_prev_snap_txg) {
1113 				VERIFY(0 == bplist_enqueue(&ds->ds_deadlist,
1114 				    &bp, tx));
1115 				if (ds_prev && !after_branch_point &&
1116 				    bp.blk_birth >
1117 				    ds_prev->ds_phys->ds_prev_snap_txg) {
1118 					ds_prev->ds_phys->ds_unique_bytes +=
1119 					    bp_get_dasize(dp->dp_spa, &bp);
1120 				}
1121 			} else {
1122 				used += bp_get_dasize(dp->dp_spa, &bp);
1123 				compressed += BP_GET_PSIZE(&bp);
1124 				uncompressed += BP_GET_UCSIZE(&bp);
1125 				/* XXX check return value? */
1126 				(void) arc_free(zio, dp->dp_spa, tx->tx_txg,
1127 				    &bp, NULL, NULL, ARC_NOWAIT);
1128 			}
1129 		}
1130 
1131 		/* free next's deadlist */
1132 		bplist_close(&ds_next->ds_deadlist);
1133 		bplist_destroy(mos, ds_next->ds_phys->ds_deadlist_obj, tx);
1134 
1135 		/* set next's deadlist to our deadlist */
1136 		ds_next->ds_phys->ds_deadlist_obj =
1137 		    ds->ds_phys->ds_deadlist_obj;
1138 		VERIFY(0 == bplist_open(&ds_next->ds_deadlist, mos,
1139 		    ds_next->ds_phys->ds_deadlist_obj));
1140 		ds->ds_phys->ds_deadlist_obj = 0;
1141 
1142 		if (ds_next->ds_phys->ds_next_snap_obj != 0) {
1143 			/*
1144 			 * Update next's unique to include blocks which
1145 			 * were previously shared by only this snapshot
1146 			 * and it.  Those blocks will be born after the
1147 			 * prev snap and before this snap, and will have
1148 			 * died after the next snap and before the one
1149 			 * after that (ie. be on the snap after next's
1150 			 * deadlist).
1151 			 *
1152 			 * XXX we're doing this long task with the
1153 			 * config lock held
1154 			 */
1155 			dsl_dataset_t *ds_after_next;
1156 
1157 			VERIFY(0 == dsl_dataset_open_obj(dp,
1158 			    ds_next->ds_phys->ds_next_snap_obj, NULL,
1159 			    DS_MODE_NONE, FTAG, &ds_after_next));
1160 			itor = 0;
1161 			while (bplist_iterate(&ds_after_next->ds_deadlist,
1162 			    &itor, &bp) == 0) {
1163 				if (bp.blk_birth >
1164 				    ds->ds_phys->ds_prev_snap_txg &&
1165 				    bp.blk_birth <=
1166 				    ds->ds_phys->ds_creation_txg) {
1167 					ds_next->ds_phys->ds_unique_bytes +=
1168 					    bp_get_dasize(dp->dp_spa, &bp);
1169 				}
1170 			}
1171 
1172 			dsl_dataset_close(ds_after_next, DS_MODE_NONE, FTAG);
1173 			ASSERT3P(ds_next->ds_prev, ==, NULL);
1174 		} else {
1175 			/*
1176 			 * It would be nice to update the head dataset's
1177 			 * unique.  To do so we would have to traverse
1178 			 * it for blocks born after ds_prev, which is
1179 			 * pretty expensive just to maintain something
1180 			 * for debugging purposes.
1181 			 */
1182 			ASSERT3P(ds_next->ds_prev, ==, ds);
1183 			dsl_dataset_close(ds_next->ds_prev, DS_MODE_NONE,
1184 			    ds_next);
1185 			if (ds_prev) {
1186 				VERIFY(0 == dsl_dataset_open_obj(dp,
1187 				    ds->ds_phys->ds_prev_snap_obj, NULL,
1188 				    DS_MODE_NONE, ds_next, &ds_next->ds_prev));
1189 			} else {
1190 				ds_next->ds_prev = NULL;
1191 			}
1192 		}
1193 		dsl_dataset_close(ds_next, DS_MODE_NONE, FTAG);
1194 
1195 		/*
1196 		 * NB: unique_bytes is not accurate for head objsets
1197 		 * because we don't update it when we delete the most
1198 		 * recent snapshot -- see above comment.
1199 		 */
1200 		ASSERT3U(used, ==, ds->ds_phys->ds_unique_bytes);
1201 	} else {
1202 		/*
1203 		 * There's no next snapshot, so this is a head dataset.
1204 		 * Destroy the deadlist.  Unless it's a clone, the
1205 		 * deadlist should be empty.  (If it's a clone, it's
1206 		 * safe to ignore the deadlist contents.)
1207 		 */
1208 		struct killarg ka;
1209 
1210 		ASSERT(after_branch_point || bplist_empty(&ds->ds_deadlist));
1211 		bplist_close(&ds->ds_deadlist);
1212 		bplist_destroy(mos, ds->ds_phys->ds_deadlist_obj, tx);
1213 		ds->ds_phys->ds_deadlist_obj = 0;
1214 
1215 		/*
1216 		 * Free everything that we point to (that's born after
1217 		 * the previous snapshot, if we are a clone)
1218 		 *
1219 		 * XXX we're doing this long task with the config lock held
1220 		 */
1221 		ka.usedp = &used;
1222 		ka.compressedp = &compressed;
1223 		ka.uncompressedp = &uncompressed;
1224 		ka.zio = zio;
1225 		ka.tx = tx;
1226 		err = traverse_dsl_dataset(ds, ds->ds_phys->ds_prev_snap_txg,
1227 		    ADVANCE_POST, kill_blkptr, &ka);
1228 		ASSERT3U(err, ==, 0);
1229 	}
1230 
1231 	err = zio_wait(zio);
1232 	ASSERT3U(err, ==, 0);
1233 
1234 	dsl_dir_diduse_space(ds->ds_dir, -used, -compressed, -uncompressed, tx);
1235 
1236 	if (ds->ds_phys->ds_snapnames_zapobj) {
1237 		err = zap_destroy(mos, ds->ds_phys->ds_snapnames_zapobj, tx);
1238 		ASSERT(err == 0);
1239 	}
1240 
1241 	if (ds->ds_dir->dd_phys->dd_head_dataset_obj == ds->ds_object) {
1242 		/* Erase the link in the dataset */
1243 		dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
1244 		ds->ds_dir->dd_phys->dd_head_dataset_obj = 0;
1245 		/*
1246 		 * dsl_dir_sync_destroy() called us, they'll destroy
1247 		 * the dataset.
1248 		 */
1249 	} else {
1250 		/* remove from snapshot namespace */
1251 		dsl_dataset_t *ds_head;
1252 		VERIFY(0 == dsl_dataset_open_obj(dp,
1253 		    ds->ds_dir->dd_phys->dd_head_dataset_obj, NULL,
1254 		    DS_MODE_NONE, FTAG, &ds_head));
1255 		VERIFY(0 == dsl_dataset_get_snapname(ds));
1256 #ifdef ZFS_DEBUG
1257 		{
1258 			uint64_t val;
1259 			err = zap_lookup(mos,
1260 			    ds_head->ds_phys->ds_snapnames_zapobj,
1261 			    ds->ds_snapname, 8, 1, &val);
1262 			ASSERT3U(err, ==, 0);
1263 			ASSERT3U(val, ==, obj);
1264 		}
1265 #endif
1266 		err = zap_remove(mos, ds_head->ds_phys->ds_snapnames_zapobj,
1267 		    ds->ds_snapname, tx);
1268 		ASSERT(err == 0);
1269 		dsl_dataset_close(ds_head, DS_MODE_NONE, FTAG);
1270 	}
1271 
1272 	if (ds_prev && ds->ds_prev != ds_prev)
1273 		dsl_dataset_close(ds_prev, DS_MODE_NONE, FTAG);
1274 
1275 	dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, tag);
1276 	VERIFY(0 == dmu_object_free(mos, obj, tx));
1277 }
1278 
1279 /* ARGSUSED */
1280 int
1281 dsl_dataset_snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
1282 {
1283 	objset_t *os = arg1;
1284 	dsl_dataset_t *ds = os->os->os_dsl_dataset;
1285 	const char *snapname = arg2;
1286 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
1287 	int err;
1288 	uint64_t value;
1289 
1290 	/*
1291 	 * We don't allow multiple snapshots of the same txg.  If there
1292 	 * is already one, try again.
1293 	 */
1294 	if (ds->ds_phys->ds_prev_snap_txg >= tx->tx_txg)
1295 		return (EAGAIN);
1296 
1297 	/*
1298 	 * Check for conflicting name snapshot name.
1299 	 */
1300 	err = zap_lookup(mos, ds->ds_phys->ds_snapnames_zapobj,
1301 	    snapname, 8, 1, &value);
1302 	if (err == 0)
1303 		return (EEXIST);
1304 	if (err != ENOENT)
1305 		return (err);
1306 
1307 	ds->ds_trysnap_txg = tx->tx_txg;
1308 	return (0);
1309 }
1310 
1311 void
1312 dsl_dataset_snapshot_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1313 {
1314 	objset_t *os = arg1;
1315 	dsl_dataset_t *ds = os->os->os_dsl_dataset;
1316 	const char *snapname = arg2;
1317 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
1318 	dmu_buf_t *dbuf;
1319 	dsl_dataset_phys_t *dsphys;
1320 	uint64_t dsobj;
1321 	objset_t *mos = dp->dp_meta_objset;
1322 	int err;
1323 
1324 	spa_scrub_restart(dp->dp_spa, tx->tx_txg);
1325 	ASSERT(RW_WRITE_HELD(&dp->dp_config_rwlock));
1326 
1327 	dsobj = dmu_object_alloc(mos, DMU_OT_DSL_DATASET, 0,
1328 	    DMU_OT_DSL_DATASET, sizeof (dsl_dataset_phys_t), tx);
1329 	VERIFY(0 == dmu_bonus_hold(mos, dsobj, FTAG, &dbuf));
1330 	dmu_buf_will_dirty(dbuf, tx);
1331 	dsphys = dbuf->db_data;
1332 	dsphys->ds_dir_obj = ds->ds_dir->dd_object;
1333 	dsphys->ds_fsid_guid = unique_create();
1334 	unique_remove(dsphys->ds_fsid_guid); /* it isn't open yet */
1335 	(void) random_get_pseudo_bytes((void*)&dsphys->ds_guid,
1336 	    sizeof (dsphys->ds_guid));
1337 	dsphys->ds_prev_snap_obj = ds->ds_phys->ds_prev_snap_obj;
1338 	dsphys->ds_prev_snap_txg = ds->ds_phys->ds_prev_snap_txg;
1339 	dsphys->ds_next_snap_obj = ds->ds_object;
1340 	dsphys->ds_num_children = 1;
1341 	dsphys->ds_creation_time = gethrestime_sec();
1342 	dsphys->ds_creation_txg = tx->tx_txg;
1343 	dsphys->ds_deadlist_obj = ds->ds_phys->ds_deadlist_obj;
1344 	dsphys->ds_used_bytes = ds->ds_phys->ds_used_bytes;
1345 	dsphys->ds_compressed_bytes = ds->ds_phys->ds_compressed_bytes;
1346 	dsphys->ds_uncompressed_bytes = ds->ds_phys->ds_uncompressed_bytes;
1347 	dsphys->ds_flags = ds->ds_phys->ds_flags;
1348 	dsphys->ds_bp = ds->ds_phys->ds_bp;
1349 	dmu_buf_rele(dbuf, FTAG);
1350 
1351 	ASSERT3U(ds->ds_prev != 0, ==, ds->ds_phys->ds_prev_snap_obj != 0);
1352 	if (ds->ds_prev) {
1353 		ASSERT(ds->ds_prev->ds_phys->ds_next_snap_obj ==
1354 		    ds->ds_object ||
1355 		    ds->ds_prev->ds_phys->ds_num_children > 1);
1356 		if (ds->ds_prev->ds_phys->ds_next_snap_obj == ds->ds_object) {
1357 			dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
1358 			ASSERT3U(ds->ds_phys->ds_prev_snap_txg, ==,
1359 			    ds->ds_prev->ds_phys->ds_creation_txg);
1360 			ds->ds_prev->ds_phys->ds_next_snap_obj = dsobj;
1361 		}
1362 	}
1363 
1364 	bplist_close(&ds->ds_deadlist);
1365 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1366 	ASSERT3U(ds->ds_phys->ds_prev_snap_txg, <, dsphys->ds_creation_txg);
1367 	ds->ds_phys->ds_prev_snap_obj = dsobj;
1368 	ds->ds_phys->ds_prev_snap_txg = dsphys->ds_creation_txg;
1369 	ds->ds_phys->ds_unique_bytes = 0;
1370 	ds->ds_phys->ds_deadlist_obj =
1371 	    bplist_create(mos, DSL_DEADLIST_BLOCKSIZE, tx);
1372 	VERIFY(0 == bplist_open(&ds->ds_deadlist, mos,
1373 	    ds->ds_phys->ds_deadlist_obj));
1374 
1375 	dprintf("snap '%s' -> obj %llu\n", snapname, dsobj);
1376 	err = zap_add(mos, ds->ds_phys->ds_snapnames_zapobj,
1377 	    snapname, 8, 1, &dsobj, tx);
1378 	ASSERT(err == 0);
1379 
1380 	if (ds->ds_prev)
1381 		dsl_dataset_close(ds->ds_prev, DS_MODE_NONE, ds);
1382 	VERIFY(0 == dsl_dataset_open_obj(dp,
1383 	    ds->ds_phys->ds_prev_snap_obj, snapname,
1384 	    DS_MODE_NONE, ds, &ds->ds_prev));
1385 }
1386 
1387 void
1388 dsl_dataset_sync(dsl_dataset_t *ds, dmu_tx_t *tx)
1389 {
1390 	ASSERT(dmu_tx_is_syncing(tx));
1391 	ASSERT(ds->ds_user_ptr != NULL);
1392 	ASSERT(ds->ds_phys->ds_next_snap_obj == 0);
1393 
1394 	dmu_objset_sync(ds->ds_user_ptr, tx);
1395 	dsl_dir_dirty(ds->ds_dir, tx);
1396 	bplist_close(&ds->ds_deadlist);
1397 
1398 	dmu_buf_rele(ds->ds_dbuf, ds);
1399 }
1400 
1401 void
1402 dsl_dataset_stats(dsl_dataset_t *ds, dmu_objset_stats_t *dds)
1403 {
1404 	/* fill in properties crap */
1405 	dsl_dir_stats(ds->ds_dir, dds);
1406 
1407 	if (ds->ds_phys->ds_num_children != 0) {
1408 		dds->dds_is_snapshot = TRUE;
1409 		dds->dds_num_clones = ds->ds_phys->ds_num_children - 1;
1410 	}
1411 
1412 	dds->dds_inconsistent = ds->ds_phys->ds_flags & DS_FLAG_INCONSISTENT;
1413 	dds->dds_last_txg = ds->ds_phys->ds_bp.blk_birth;
1414 
1415 	dds->dds_objects_used = ds->ds_phys->ds_bp.blk_fill;
1416 	dds->dds_objects_avail = DN_MAX_OBJECT - dds->dds_objects_used;
1417 
1418 	/* We override the dataset's creation time... they should be the same */
1419 	dds->dds_creation_time = ds->ds_phys->ds_creation_time;
1420 	dds->dds_creation_txg = ds->ds_phys->ds_creation_txg;
1421 	dds->dds_space_refd = ds->ds_phys->ds_used_bytes;
1422 	dds->dds_fsid_guid = ds->ds_phys->ds_fsid_guid;
1423 
1424 	if (ds->ds_phys->ds_next_snap_obj) {
1425 		/*
1426 		 * This is a snapshot; override the dd's space used with
1427 		 * our unique space
1428 		 */
1429 		dds->dds_space_used = ds->ds_phys->ds_unique_bytes;
1430 		dds->dds_compressed_bytes =
1431 		    ds->ds_phys->ds_compressed_bytes;
1432 		dds->dds_uncompressed_bytes =
1433 		    ds->ds_phys->ds_uncompressed_bytes;
1434 	}
1435 }
1436 
1437 dsl_pool_t *
1438 dsl_dataset_pool(dsl_dataset_t *ds)
1439 {
1440 	return (ds->ds_dir->dd_pool);
1441 }
1442 
1443 /* ARGSUSED */
1444 static int
1445 dsl_dataset_snapshot_rename_check(void *arg1, void *arg2, dmu_tx_t *tx)
1446 {
1447 	dsl_dataset_t *ds = arg1;
1448 	char *newsnapname = arg2;
1449 	dsl_dir_t *dd = ds->ds_dir;
1450 	objset_t *mos = dd->dd_pool->dp_meta_objset;
1451 	dsl_dataset_t *hds;
1452 	uint64_t val;
1453 	int err;
1454 
1455 	err = dsl_dataset_open_obj(dd->dd_pool,
1456 	    dd->dd_phys->dd_head_dataset_obj, NULL, DS_MODE_NONE, FTAG, &hds);
1457 	if (err)
1458 		return (err);
1459 
1460 	/* new name better not be in use */
1461 	err = zap_lookup(mos, hds->ds_phys->ds_snapnames_zapobj,
1462 	    newsnapname, 8, 1, &val);
1463 	dsl_dataset_close(hds, DS_MODE_NONE, FTAG);
1464 
1465 	if (err == 0)
1466 		err = EEXIST;
1467 	else if (err == ENOENT)
1468 		err = 0;
1469 	return (err);
1470 }
1471 
1472 static void
1473 dsl_dataset_snapshot_rename_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1474 {
1475 	dsl_dataset_t *ds = arg1;
1476 	char *newsnapname = arg2;
1477 	dsl_dir_t *dd = ds->ds_dir;
1478 	objset_t *mos = dd->dd_pool->dp_meta_objset;
1479 	dsl_dataset_t *hds;
1480 	int err;
1481 
1482 	ASSERT(ds->ds_phys->ds_next_snap_obj != 0);
1483 
1484 	VERIFY(0 == dsl_dataset_open_obj(dd->dd_pool,
1485 	    dd->dd_phys->dd_head_dataset_obj, NULL, DS_MODE_NONE, FTAG, &hds));
1486 
1487 	VERIFY(0 == dsl_dataset_get_snapname(ds));
1488 	err = zap_remove(mos, hds->ds_phys->ds_snapnames_zapobj,
1489 	    ds->ds_snapname, tx);
1490 	ASSERT3U(err, ==, 0);
1491 	mutex_enter(&ds->ds_lock);
1492 	(void) strcpy(ds->ds_snapname, newsnapname);
1493 	mutex_exit(&ds->ds_lock);
1494 	err = zap_add(mos, hds->ds_phys->ds_snapnames_zapobj,
1495 	    ds->ds_snapname, 8, 1, &ds->ds_object, tx);
1496 	ASSERT3U(err, ==, 0);
1497 
1498 	dsl_dataset_close(hds, DS_MODE_NONE, FTAG);
1499 }
1500 
1501 #pragma weak dmu_objset_rename = dsl_dataset_rename
1502 int
1503 dsl_dataset_rename(const char *oldname, const char *newname)
1504 {
1505 	dsl_dir_t *dd;
1506 	dsl_dataset_t *ds;
1507 	const char *tail;
1508 	int err;
1509 
1510 	err = dsl_dir_open(oldname, FTAG, &dd, &tail);
1511 	if (err)
1512 		return (err);
1513 	if (tail == NULL) {
1514 		err = dsl_dir_rename(dd, newname);
1515 		dsl_dir_close(dd, FTAG);
1516 		return (err);
1517 	}
1518 	if (tail[0] != '@') {
1519 		/* the name ended in a nonexistant component */
1520 		dsl_dir_close(dd, FTAG);
1521 		return (ENOENT);
1522 	}
1523 
1524 	dsl_dir_close(dd, FTAG);
1525 
1526 	/* new name must be snapshot in same filesystem */
1527 	tail = strchr(newname, '@');
1528 	if (tail == NULL)
1529 		return (EINVAL);
1530 	tail++;
1531 	if (strncmp(oldname, newname, tail - newname) != 0)
1532 		return (EXDEV);
1533 
1534 	err = dsl_dataset_open(oldname,
1535 	    DS_MODE_READONLY | DS_MODE_STANDARD, FTAG, &ds);
1536 	if (err)
1537 		return (err);
1538 
1539 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1540 	    dsl_dataset_snapshot_rename_check,
1541 	    dsl_dataset_snapshot_rename_sync, ds, (char *)tail, 1);
1542 
1543 	dsl_dataset_close(ds, DS_MODE_STANDARD, FTAG);
1544 
1545 	return (err);
1546 }
1547 
1548 struct promotearg {
1549 	uint64_t used, comp, uncomp, unique;
1550 	uint64_t newnext_obj, snapnames_obj;
1551 };
1552 
1553 static int
1554 dsl_dataset_promote_check(void *arg1, void *arg2, dmu_tx_t *tx)
1555 {
1556 	dsl_dataset_t *hds = arg1;
1557 	struct promotearg *pa = arg2;
1558 	dsl_dir_t *dd = hds->ds_dir;
1559 	dsl_pool_t *dp = hds->ds_dir->dd_pool;
1560 	dsl_dir_t *pdd = NULL;
1561 	dsl_dataset_t *ds = NULL;
1562 	dsl_dataset_t *pivot_ds = NULL;
1563 	dsl_dataset_t *newnext_ds = NULL;
1564 	int err;
1565 	char *name = NULL;
1566 	uint64_t itor = 0;
1567 	blkptr_t bp;
1568 
1569 	bzero(pa, sizeof (*pa));
1570 
1571 	/* Check that it is a clone */
1572 	if (dd->dd_phys->dd_clone_parent_obj == 0)
1573 		return (EINVAL);
1574 
1575 	/* Since this is so expensive, don't do the preliminary check */
1576 	if (!dmu_tx_is_syncing(tx))
1577 		return (0);
1578 
1579 	if (err = dsl_dataset_open_obj(dp,
1580 	    dd->dd_phys->dd_clone_parent_obj,
1581 	    NULL, DS_MODE_EXCLUSIVE, FTAG, &pivot_ds))
1582 		goto out;
1583 	pdd = pivot_ds->ds_dir;
1584 
1585 	{
1586 		dsl_dataset_t *phds;
1587 		if (err = dsl_dataset_open_obj(dd->dd_pool,
1588 		    pdd->dd_phys->dd_head_dataset_obj,
1589 		    NULL, DS_MODE_NONE, FTAG, &phds))
1590 			goto out;
1591 		pa->snapnames_obj = phds->ds_phys->ds_snapnames_zapobj;
1592 		dsl_dataset_close(phds, DS_MODE_NONE, FTAG);
1593 	}
1594 
1595 	if (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE) {
1596 		err = EXDEV;
1597 		goto out;
1598 	}
1599 
1600 	/* find pivot point's new next ds */
1601 	VERIFY(0 == dsl_dataset_open_obj(dd->dd_pool, hds->ds_object,
1602 	    NULL, DS_MODE_NONE, FTAG, &newnext_ds));
1603 	while (newnext_ds->ds_phys->ds_prev_snap_obj != pivot_ds->ds_object) {
1604 		dsl_dataset_t *prev;
1605 
1606 		if (err = dsl_dataset_open_obj(dd->dd_pool,
1607 		    newnext_ds->ds_phys->ds_prev_snap_obj,
1608 		    NULL, DS_MODE_NONE, FTAG, &prev))
1609 			goto out;
1610 		dsl_dataset_close(newnext_ds, DS_MODE_NONE, FTAG);
1611 		newnext_ds = prev;
1612 	}
1613 	pa->newnext_obj = newnext_ds->ds_object;
1614 
1615 	/* compute pivot point's new unique space */
1616 	while ((err = bplist_iterate(&newnext_ds->ds_deadlist,
1617 	    &itor, &bp)) == 0) {
1618 		if (bp.blk_birth > pivot_ds->ds_phys->ds_prev_snap_txg)
1619 			pa->unique += bp_get_dasize(dd->dd_pool->dp_spa, &bp);
1620 	}
1621 	if (err != ENOENT)
1622 		goto out;
1623 
1624 	/* Walk the snapshots that we are moving */
1625 	name = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1626 	ds = pivot_ds;
1627 	/* CONSTCOND */
1628 	while (TRUE) {
1629 		uint64_t val, dlused, dlcomp, dluncomp;
1630 		dsl_dataset_t *prev;
1631 
1632 		/* Check that the snapshot name does not conflict */
1633 		dsl_dataset_name(ds, name);
1634 		err = zap_lookup(dd->dd_pool->dp_meta_objset,
1635 		    hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
1636 		    8, 1, &val);
1637 		if (err != ENOENT) {
1638 			if (err == 0)
1639 				err = EEXIST;
1640 			goto out;
1641 		}
1642 
1643 		/*
1644 		 * compute space to transfer.  Each snapshot gave birth to:
1645 		 * (my used) - (prev's used) + (deadlist's used)
1646 		 */
1647 		pa->used += ds->ds_phys->ds_used_bytes;
1648 		pa->comp += ds->ds_phys->ds_compressed_bytes;
1649 		pa->uncomp += ds->ds_phys->ds_uncompressed_bytes;
1650 
1651 		/* If we reach the first snapshot, we're done. */
1652 		if (ds->ds_phys->ds_prev_snap_obj == 0)
1653 			break;
1654 
1655 		if (err = bplist_space(&ds->ds_deadlist,
1656 		    &dlused, &dlcomp, &dluncomp))
1657 			goto out;
1658 		if (err = dsl_dataset_open_obj(dd->dd_pool,
1659 		    ds->ds_phys->ds_prev_snap_obj, NULL, DS_MODE_EXCLUSIVE,
1660 		    FTAG, &prev))
1661 			goto out;
1662 		pa->used += dlused - prev->ds_phys->ds_used_bytes;
1663 		pa->comp += dlcomp - prev->ds_phys->ds_compressed_bytes;
1664 		pa->uncomp += dluncomp - prev->ds_phys->ds_uncompressed_bytes;
1665 
1666 		/*
1667 		 * We could be a clone of a clone.  If we reach our
1668 		 * parent's branch point, we're done.
1669 		 */
1670 		if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1671 			dsl_dataset_close(prev, DS_MODE_EXCLUSIVE, FTAG);
1672 			break;
1673 		}
1674 		if (ds != pivot_ds)
1675 			dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
1676 		ds = prev;
1677 	}
1678 
1679 	/* Check that there is enough space here */
1680 	err = dsl_dir_transfer_possible(pdd, dd, pa->used);
1681 
1682 out:
1683 	if (ds && ds != pivot_ds)
1684 		dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
1685 	if (pivot_ds)
1686 		dsl_dataset_close(pivot_ds, DS_MODE_EXCLUSIVE, FTAG);
1687 	if (newnext_ds)
1688 		dsl_dataset_close(newnext_ds, DS_MODE_NONE, FTAG);
1689 	if (name)
1690 		kmem_free(name, MAXPATHLEN);
1691 	return (err);
1692 }
1693 
1694 static void
1695 dsl_dataset_promote_sync(void *arg1, void *arg2, dmu_tx_t *tx)
1696 {
1697 	dsl_dataset_t *hds = arg1;
1698 	struct promotearg *pa = arg2;
1699 	dsl_dir_t *dd = hds->ds_dir;
1700 	dsl_pool_t *dp = hds->ds_dir->dd_pool;
1701 	dsl_dir_t *pdd = NULL;
1702 	dsl_dataset_t *ds, *pivot_ds;
1703 	char *name;
1704 
1705 	ASSERT(dd->dd_phys->dd_clone_parent_obj != 0);
1706 	ASSERT(0 == (hds->ds_phys->ds_flags & DS_FLAG_NOPROMOTE));
1707 
1708 	VERIFY(0 == dsl_dataset_open_obj(dp,
1709 	    dd->dd_phys->dd_clone_parent_obj,
1710 	    NULL, DS_MODE_EXCLUSIVE, FTAG, &pivot_ds));
1711 	/*
1712 	 * We need to explicitly open pdd, since pivot_ds's pdd will be
1713 	 * changing.
1714 	 */
1715 	VERIFY(0 == dsl_dir_open_obj(dp, pivot_ds->ds_dir->dd_object,
1716 	    NULL, FTAG, &pdd));
1717 
1718 	/* move snapshots to this dir */
1719 	name = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1720 	ds = pivot_ds;
1721 	/* CONSTCOND */
1722 	while (TRUE) {
1723 		dsl_dataset_t *prev;
1724 
1725 		/* move snap name entry */
1726 		dsl_dataset_name(ds, name);
1727 		VERIFY(0 == zap_remove(dp->dp_meta_objset,
1728 		    pa->snapnames_obj, ds->ds_snapname, tx));
1729 		VERIFY(0 == zap_add(dp->dp_meta_objset,
1730 		    hds->ds_phys->ds_snapnames_zapobj, ds->ds_snapname,
1731 		    8, 1, &ds->ds_object, tx));
1732 
1733 		/* change containing dsl_dir */
1734 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
1735 		ASSERT3U(ds->ds_phys->ds_dir_obj, ==, pdd->dd_object);
1736 		ds->ds_phys->ds_dir_obj = dd->dd_object;
1737 		ASSERT3P(ds->ds_dir, ==, pdd);
1738 		dsl_dir_close(ds->ds_dir, ds);
1739 		VERIFY(0 == dsl_dir_open_obj(dp, dd->dd_object,
1740 		    NULL, ds, &ds->ds_dir));
1741 
1742 		ASSERT3U(dsl_prop_numcb(ds), ==, 0);
1743 
1744 		if (ds->ds_phys->ds_prev_snap_obj == 0)
1745 			break;
1746 
1747 		VERIFY(0 == dsl_dataset_open_obj(dp,
1748 		    ds->ds_phys->ds_prev_snap_obj, NULL, DS_MODE_EXCLUSIVE,
1749 		    FTAG, &prev));
1750 
1751 		if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1752 			dsl_dataset_close(prev, DS_MODE_EXCLUSIVE, FTAG);
1753 			break;
1754 		}
1755 		if (ds != pivot_ds)
1756 			dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
1757 		ds = prev;
1758 	}
1759 	if (ds != pivot_ds)
1760 		dsl_dataset_close(ds, DS_MODE_EXCLUSIVE, FTAG);
1761 
1762 	/* change pivot point's next snap */
1763 	dmu_buf_will_dirty(pivot_ds->ds_dbuf, tx);
1764 	pivot_ds->ds_phys->ds_next_snap_obj = pa->newnext_obj;
1765 
1766 	/* change clone_parent-age */
1767 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
1768 	ASSERT3U(dd->dd_phys->dd_clone_parent_obj, ==, pivot_ds->ds_object);
1769 	dd->dd_phys->dd_clone_parent_obj = pdd->dd_phys->dd_clone_parent_obj;
1770 	dmu_buf_will_dirty(pdd->dd_dbuf, tx);
1771 	pdd->dd_phys->dd_clone_parent_obj = pivot_ds->ds_object;
1772 
1773 	/* change space accounting */
1774 	dsl_dir_diduse_space(pdd, -pa->used, -pa->comp, -pa->uncomp, tx);
1775 	dsl_dir_diduse_space(dd, pa->used, pa->comp, pa->uncomp, tx);
1776 	pivot_ds->ds_phys->ds_unique_bytes = pa->unique;
1777 
1778 	dsl_dir_close(pdd, FTAG);
1779 	dsl_dataset_close(pivot_ds, DS_MODE_EXCLUSIVE, FTAG);
1780 	kmem_free(name, MAXPATHLEN);
1781 }
1782 
1783 int
1784 dsl_dataset_promote(const char *name)
1785 {
1786 	dsl_dataset_t *ds;
1787 	int err;
1788 	dmu_object_info_t doi;
1789 	struct promotearg pa;
1790 
1791 	err = dsl_dataset_open(name, DS_MODE_NONE, FTAG, &ds);
1792 	if (err)
1793 		return (err);
1794 
1795 	err = dmu_object_info(ds->ds_dir->dd_pool->dp_meta_objset,
1796 	    ds->ds_phys->ds_snapnames_zapobj, &doi);
1797 	if (err) {
1798 		dsl_dataset_close(ds, DS_MODE_NONE, FTAG);
1799 		return (err);
1800 	}
1801 
1802 	/*
1803 	 * Add in 128x the snapnames zapobj size, since we will be moving
1804 	 * a bunch of snapnames to the promoted ds, and dirtying their
1805 	 * bonus buffers.
1806 	 */
1807 	err = dsl_sync_task_do(ds->ds_dir->dd_pool,
1808 	    dsl_dataset_promote_check,
1809 	    dsl_dataset_promote_sync, ds, &pa, 2 + 2 * doi.doi_physical_blks);
1810 	dsl_dataset_close(ds, DS_MODE_NONE, FTAG);
1811 	return (err);
1812 }
1813