xref: /titanic_51/usr/src/uts/common/fs/zfs/dmu_objset.c (revision 1117ab727b886504c594e34371b4408daacba77e)
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 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/cred.h>
27 #include <sys/zfs_context.h>
28 #include <sys/dmu_objset.h>
29 #include <sys/dsl_dir.h>
30 #include <sys/dsl_dataset.h>
31 #include <sys/dsl_prop.h>
32 #include <sys/dsl_pool.h>
33 #include <sys/dsl_synctask.h>
34 #include <sys/dsl_deleg.h>
35 #include <sys/dnode.h>
36 #include <sys/dbuf.h>
37 #include <sys/zvol.h>
38 #include <sys/dmu_tx.h>
39 #include <sys/zio_checksum.h>
40 #include <sys/zap.h>
41 #include <sys/zil.h>
42 #include <sys/dmu_impl.h>
43 #include <sys/zfs_ioctl.h>
44 
45 spa_t *
46 dmu_objset_spa(objset_t *os)
47 {
48 	return (os->os->os_spa);
49 }
50 
51 zilog_t *
52 dmu_objset_zil(objset_t *os)
53 {
54 	return (os->os->os_zil);
55 }
56 
57 dsl_pool_t *
58 dmu_objset_pool(objset_t *os)
59 {
60 	dsl_dataset_t *ds;
61 
62 	if ((ds = os->os->os_dsl_dataset) != NULL && ds->ds_dir)
63 		return (ds->ds_dir->dd_pool);
64 	else
65 		return (spa_get_dsl(os->os->os_spa));
66 }
67 
68 dsl_dataset_t *
69 dmu_objset_ds(objset_t *os)
70 {
71 	return (os->os->os_dsl_dataset);
72 }
73 
74 dmu_objset_type_t
75 dmu_objset_type(objset_t *os)
76 {
77 	return (os->os->os_phys->os_type);
78 }
79 
80 void
81 dmu_objset_name(objset_t *os, char *buf)
82 {
83 	dsl_dataset_name(os->os->os_dsl_dataset, buf);
84 }
85 
86 uint64_t
87 dmu_objset_id(objset_t *os)
88 {
89 	dsl_dataset_t *ds = os->os->os_dsl_dataset;
90 
91 	return (ds ? ds->ds_object : 0);
92 }
93 
94 static void
95 checksum_changed_cb(void *arg, uint64_t newval)
96 {
97 	objset_impl_t *osi = arg;
98 
99 	/*
100 	 * Inheritance should have been done by now.
101 	 */
102 	ASSERT(newval != ZIO_CHECKSUM_INHERIT);
103 
104 	osi->os_checksum = zio_checksum_select(newval, ZIO_CHECKSUM_ON_VALUE);
105 }
106 
107 static void
108 compression_changed_cb(void *arg, uint64_t newval)
109 {
110 	objset_impl_t *osi = arg;
111 
112 	/*
113 	 * Inheritance and range checking should have been done by now.
114 	 */
115 	ASSERT(newval != ZIO_COMPRESS_INHERIT);
116 
117 	osi->os_compress = zio_compress_select(newval, ZIO_COMPRESS_ON_VALUE);
118 }
119 
120 static void
121 copies_changed_cb(void *arg, uint64_t newval)
122 {
123 	objset_impl_t *osi = arg;
124 
125 	/*
126 	 * Inheritance and range checking should have been done by now.
127 	 */
128 	ASSERT(newval > 0);
129 	ASSERT(newval <= spa_max_replication(osi->os_spa));
130 
131 	osi->os_copies = newval;
132 }
133 
134 static void
135 primary_cache_changed_cb(void *arg, uint64_t newval)
136 {
137 	objset_impl_t *osi = arg;
138 
139 	/*
140 	 * Inheritance and range checking should have been done by now.
141 	 */
142 	ASSERT(newval == ZFS_CACHE_ALL || newval == ZFS_CACHE_NONE ||
143 	    newval == ZFS_CACHE_METADATA);
144 
145 	osi->os_primary_cache = newval;
146 }
147 
148 static void
149 secondary_cache_changed_cb(void *arg, uint64_t newval)
150 {
151 	objset_impl_t *osi = arg;
152 
153 	/*
154 	 * Inheritance and range checking should have been done by now.
155 	 */
156 	ASSERT(newval == ZFS_CACHE_ALL || newval == ZFS_CACHE_NONE ||
157 	    newval == ZFS_CACHE_METADATA);
158 
159 	osi->os_secondary_cache = newval;
160 }
161 
162 void
163 dmu_objset_byteswap(void *buf, size_t size)
164 {
165 	objset_phys_t *osp = buf;
166 
167 	ASSERT(size == sizeof (objset_phys_t));
168 	dnode_byteswap(&osp->os_meta_dnode);
169 	byteswap_uint64_array(&osp->os_zil_header, sizeof (zil_header_t));
170 	osp->os_type = BSWAP_64(osp->os_type);
171 }
172 
173 int
174 dmu_objset_open_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
175     objset_impl_t **osip)
176 {
177 	objset_impl_t *osi;
178 	int i, err;
179 
180 	ASSERT(ds == NULL || MUTEX_HELD(&ds->ds_opening_lock));
181 
182 	osi = kmem_zalloc(sizeof (objset_impl_t), KM_SLEEP);
183 	osi->os.os = osi;
184 	osi->os_dsl_dataset = ds;
185 	osi->os_spa = spa;
186 	osi->os_rootbp = bp;
187 	if (!BP_IS_HOLE(osi->os_rootbp)) {
188 		uint32_t aflags = ARC_WAIT;
189 		zbookmark_t zb;
190 		zb.zb_objset = ds ? ds->ds_object : 0;
191 		zb.zb_object = 0;
192 		zb.zb_level = -1;
193 		zb.zb_blkid = 0;
194 		if (DMU_OS_IS_L2CACHEABLE(osi))
195 			aflags |= ARC_L2CACHE;
196 
197 		dprintf_bp(osi->os_rootbp, "reading %s", "");
198 		/*
199 		 * NB: when bprewrite scrub can change the bp,
200 		 * and this is called from dmu_objset_open_ds_os, the bp
201 		 * could change, and we'll need a lock.
202 		 */
203 		err = arc_read_nolock(NULL, spa, osi->os_rootbp,
204 		    arc_getbuf_func, &osi->os_phys_buf,
205 		    ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL, &aflags, &zb);
206 		if (err) {
207 			kmem_free(osi, sizeof (objset_impl_t));
208 			/* convert checksum errors into IO errors */
209 			if (err == ECKSUM)
210 				err = EIO;
211 			return (err);
212 		}
213 		osi->os_phys = osi->os_phys_buf->b_data;
214 	} else {
215 		osi->os_phys_buf = arc_buf_alloc(spa, sizeof (objset_phys_t),
216 		    &osi->os_phys_buf, ARC_BUFC_METADATA);
217 		osi->os_phys = osi->os_phys_buf->b_data;
218 		bzero(osi->os_phys, sizeof (objset_phys_t));
219 	}
220 
221 	/*
222 	 * Note: the changed_cb will be called once before the register
223 	 * func returns, thus changing the checksum/compression from the
224 	 * default (fletcher2/off).  Snapshots don't need to know about
225 	 * checksum/compression/copies.
226 	 */
227 	if (ds) {
228 		err = dsl_prop_register(ds, "primarycache",
229 		    primary_cache_changed_cb, osi);
230 		if (err == 0)
231 			err = dsl_prop_register(ds, "secondarycache",
232 			    secondary_cache_changed_cb, osi);
233 		if (!dsl_dataset_is_snapshot(ds)) {
234 			if (err == 0)
235 				err = dsl_prop_register(ds, "checksum",
236 				    checksum_changed_cb, osi);
237 			if (err == 0)
238 				err = dsl_prop_register(ds, "compression",
239 				    compression_changed_cb, osi);
240 			if (err == 0)
241 				err = dsl_prop_register(ds, "copies",
242 				    copies_changed_cb, osi);
243 		}
244 		if (err) {
245 			VERIFY(arc_buf_remove_ref(osi->os_phys_buf,
246 			    &osi->os_phys_buf) == 1);
247 			kmem_free(osi, sizeof (objset_impl_t));
248 			return (err);
249 		}
250 	} else if (ds == NULL) {
251 		/* It's the meta-objset. */
252 		osi->os_checksum = ZIO_CHECKSUM_FLETCHER_4;
253 		osi->os_compress = ZIO_COMPRESS_LZJB;
254 		osi->os_copies = spa_max_replication(spa);
255 		osi->os_primary_cache = ZFS_CACHE_ALL;
256 		osi->os_secondary_cache = ZFS_CACHE_ALL;
257 	}
258 
259 	osi->os_zil_header = osi->os_phys->os_zil_header;
260 	osi->os_zil = zil_alloc(&osi->os, &osi->os_zil_header);
261 
262 	for (i = 0; i < TXG_SIZE; i++) {
263 		list_create(&osi->os_dirty_dnodes[i], sizeof (dnode_t),
264 		    offsetof(dnode_t, dn_dirty_link[i]));
265 		list_create(&osi->os_free_dnodes[i], sizeof (dnode_t),
266 		    offsetof(dnode_t, dn_dirty_link[i]));
267 	}
268 	list_create(&osi->os_dnodes, sizeof (dnode_t),
269 	    offsetof(dnode_t, dn_link));
270 	list_create(&osi->os_downgraded_dbufs, sizeof (dmu_buf_impl_t),
271 	    offsetof(dmu_buf_impl_t, db_link));
272 
273 	mutex_init(&osi->os_lock, NULL, MUTEX_DEFAULT, NULL);
274 	mutex_init(&osi->os_obj_lock, NULL, MUTEX_DEFAULT, NULL);
275 	mutex_init(&osi->os_user_ptr_lock, NULL, MUTEX_DEFAULT, NULL);
276 
277 	osi->os_meta_dnode = dnode_special_open(osi,
278 	    &osi->os_phys->os_meta_dnode, DMU_META_DNODE_OBJECT);
279 
280 	/*
281 	 * We should be the only thread trying to do this because we
282 	 * have ds_opening_lock
283 	 */
284 	if (ds) {
285 		VERIFY(NULL == dsl_dataset_set_user_ptr(ds, osi,
286 		    dmu_objset_evict));
287 	}
288 
289 	*osip = osi;
290 	return (0);
291 }
292 
293 static int
294 dmu_objset_open_ds_os(dsl_dataset_t *ds, objset_t *os, dmu_objset_type_t type)
295 {
296 	objset_impl_t *osi;
297 
298 	mutex_enter(&ds->ds_opening_lock);
299 	osi = dsl_dataset_get_user_ptr(ds);
300 	if (osi == NULL) {
301 		int err;
302 
303 		err = dmu_objset_open_impl(dsl_dataset_get_spa(ds),
304 		    ds, &ds->ds_phys->ds_bp, &osi);
305 		if (err) {
306 			mutex_exit(&ds->ds_opening_lock);
307 			return (err);
308 		}
309 	}
310 	mutex_exit(&ds->ds_opening_lock);
311 
312 	os->os = osi;
313 	os->os_mode = DS_MODE_NOHOLD;
314 
315 	if (type != DMU_OST_ANY && type != os->os->os_phys->os_type)
316 		return (EINVAL);
317 	return (0);
318 }
319 
320 int
321 dmu_objset_open_ds(dsl_dataset_t *ds, dmu_objset_type_t type, objset_t **osp)
322 {
323 	objset_t *os;
324 	int err;
325 
326 	os = kmem_alloc(sizeof (objset_t), KM_SLEEP);
327 	err = dmu_objset_open_ds_os(ds, os, type);
328 	if (err)
329 		kmem_free(os, sizeof (objset_t));
330 	else
331 		*osp = os;
332 	return (err);
333 }
334 
335 /* called from zpl */
336 int
337 dmu_objset_open(const char *name, dmu_objset_type_t type, int mode,
338     objset_t **osp)
339 {
340 	objset_t *os;
341 	dsl_dataset_t *ds;
342 	int err;
343 
344 	ASSERT(DS_MODE_TYPE(mode) == DS_MODE_USER ||
345 	    DS_MODE_TYPE(mode) == DS_MODE_OWNER);
346 
347 	os = kmem_alloc(sizeof (objset_t), KM_SLEEP);
348 	if (DS_MODE_TYPE(mode) == DS_MODE_USER)
349 		err = dsl_dataset_hold(name, os, &ds);
350 	else
351 		err = dsl_dataset_own(name, mode, os, &ds);
352 	if (err) {
353 		kmem_free(os, sizeof (objset_t));
354 		return (err);
355 	}
356 
357 	err = dmu_objset_open_ds_os(ds, os, type);
358 	if (err) {
359 		if (DS_MODE_TYPE(mode) == DS_MODE_USER)
360 			dsl_dataset_rele(ds, os);
361 		else
362 			dsl_dataset_disown(ds, os);
363 		kmem_free(os, sizeof (objset_t));
364 	} else {
365 		os->os_mode = mode;
366 		*osp = os;
367 	}
368 	return (err);
369 }
370 
371 void
372 dmu_objset_close(objset_t *os)
373 {
374 	ASSERT(DS_MODE_TYPE(os->os_mode) == DS_MODE_USER ||
375 	    DS_MODE_TYPE(os->os_mode) == DS_MODE_OWNER ||
376 	    DS_MODE_TYPE(os->os_mode) == DS_MODE_NOHOLD);
377 
378 	if (DS_MODE_TYPE(os->os_mode) == DS_MODE_USER)
379 		dsl_dataset_rele(os->os->os_dsl_dataset, os);
380 	else if (DS_MODE_TYPE(os->os_mode) == DS_MODE_OWNER)
381 		dsl_dataset_disown(os->os->os_dsl_dataset, os);
382 	kmem_free(os, sizeof (objset_t));
383 }
384 
385 int
386 dmu_objset_evict_dbufs(objset_t *os)
387 {
388 	objset_impl_t *osi = os->os;
389 	dnode_t *dn;
390 
391 	mutex_enter(&osi->os_lock);
392 
393 	/* process the mdn last, since the other dnodes have holds on it */
394 	list_remove(&osi->os_dnodes, osi->os_meta_dnode);
395 	list_insert_tail(&osi->os_dnodes, osi->os_meta_dnode);
396 
397 	/*
398 	 * Find the first dnode with holds.  We have to do this dance
399 	 * because dnode_add_ref() only works if you already have a
400 	 * hold.  If there are no holds then it has no dbufs so OK to
401 	 * skip.
402 	 */
403 	for (dn = list_head(&osi->os_dnodes);
404 	    dn && !dnode_add_ref(dn, FTAG);
405 	    dn = list_next(&osi->os_dnodes, dn))
406 		continue;
407 
408 	while (dn) {
409 		dnode_t *next_dn = dn;
410 
411 		do {
412 			next_dn = list_next(&osi->os_dnodes, next_dn);
413 		} while (next_dn && !dnode_add_ref(next_dn, FTAG));
414 
415 		mutex_exit(&osi->os_lock);
416 		dnode_evict_dbufs(dn);
417 		dnode_rele(dn, FTAG);
418 		mutex_enter(&osi->os_lock);
419 		dn = next_dn;
420 	}
421 	mutex_exit(&osi->os_lock);
422 	return (list_head(&osi->os_dnodes) != osi->os_meta_dnode);
423 }
424 
425 void
426 dmu_objset_evict(dsl_dataset_t *ds, void *arg)
427 {
428 	objset_impl_t *osi = arg;
429 	objset_t os;
430 	int i;
431 
432 	for (i = 0; i < TXG_SIZE; i++) {
433 		ASSERT(list_head(&osi->os_dirty_dnodes[i]) == NULL);
434 		ASSERT(list_head(&osi->os_free_dnodes[i]) == NULL);
435 	}
436 
437 	if (ds) {
438 		if (!dsl_dataset_is_snapshot(ds)) {
439 			VERIFY(0 == dsl_prop_unregister(ds, "checksum",
440 			    checksum_changed_cb, osi));
441 			VERIFY(0 == dsl_prop_unregister(ds, "compression",
442 			    compression_changed_cb, osi));
443 			VERIFY(0 == dsl_prop_unregister(ds, "copies",
444 			    copies_changed_cb, osi));
445 		}
446 		VERIFY(0 == dsl_prop_unregister(ds, "primarycache",
447 		    primary_cache_changed_cb, osi));
448 		VERIFY(0 == dsl_prop_unregister(ds, "secondarycache",
449 		    secondary_cache_changed_cb, osi));
450 	}
451 
452 	/*
453 	 * We should need only a single pass over the dnode list, since
454 	 * nothing can be added to the list at this point.
455 	 */
456 	os.os = osi;
457 	(void) dmu_objset_evict_dbufs(&os);
458 
459 	ASSERT3P(list_head(&osi->os_dnodes), ==, osi->os_meta_dnode);
460 	ASSERT3P(list_tail(&osi->os_dnodes), ==, osi->os_meta_dnode);
461 	ASSERT3P(list_head(&osi->os_meta_dnode->dn_dbufs), ==, NULL);
462 
463 	dnode_special_close(osi->os_meta_dnode);
464 	zil_free(osi->os_zil);
465 
466 	VERIFY(arc_buf_remove_ref(osi->os_phys_buf, &osi->os_phys_buf) == 1);
467 	mutex_destroy(&osi->os_lock);
468 	mutex_destroy(&osi->os_obj_lock);
469 	mutex_destroy(&osi->os_user_ptr_lock);
470 	kmem_free(osi, sizeof (objset_impl_t));
471 }
472 
473 /* called from dsl for meta-objset */
474 objset_impl_t *
475 dmu_objset_create_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
476     dmu_objset_type_t type, dmu_tx_t *tx)
477 {
478 	objset_impl_t *osi;
479 	dnode_t *mdn;
480 
481 	ASSERT(dmu_tx_is_syncing(tx));
482 	if (ds)
483 		mutex_enter(&ds->ds_opening_lock);
484 	VERIFY(0 == dmu_objset_open_impl(spa, ds, bp, &osi));
485 	if (ds)
486 		mutex_exit(&ds->ds_opening_lock);
487 	mdn = osi->os_meta_dnode;
488 
489 	dnode_allocate(mdn, DMU_OT_DNODE, 1 << DNODE_BLOCK_SHIFT,
490 	    DN_MAX_INDBLKSHIFT, DMU_OT_NONE, 0, tx);
491 
492 	/*
493 	 * We don't want to have to increase the meta-dnode's nlevels
494 	 * later, because then we could do it in quescing context while
495 	 * we are also accessing it in open context.
496 	 *
497 	 * This precaution is not necessary for the MOS (ds == NULL),
498 	 * because the MOS is only updated in syncing context.
499 	 * This is most fortunate: the MOS is the only objset that
500 	 * needs to be synced multiple times as spa_sync() iterates
501 	 * to convergence, so minimizing its dn_nlevels matters.
502 	 */
503 	if (ds != NULL) {
504 		int levels = 1;
505 
506 		/*
507 		 * Determine the number of levels necessary for the meta-dnode
508 		 * to contain DN_MAX_OBJECT dnodes.
509 		 */
510 		while ((uint64_t)mdn->dn_nblkptr << (mdn->dn_datablkshift +
511 		    (levels - 1) * (mdn->dn_indblkshift - SPA_BLKPTRSHIFT)) <
512 		    DN_MAX_OBJECT * sizeof (dnode_phys_t))
513 			levels++;
514 
515 		mdn->dn_next_nlevels[tx->tx_txg & TXG_MASK] =
516 		    mdn->dn_nlevels = levels;
517 	}
518 
519 	ASSERT(type != DMU_OST_NONE);
520 	ASSERT(type != DMU_OST_ANY);
521 	ASSERT(type < DMU_OST_NUMTYPES);
522 	osi->os_phys->os_type = type;
523 
524 	dsl_dataset_dirty(ds, tx);
525 
526 	return (osi);
527 }
528 
529 struct oscarg {
530 	void (*userfunc)(objset_t *os, void *arg, cred_t *cr, dmu_tx_t *tx);
531 	void *userarg;
532 	dsl_dataset_t *clone_parent;
533 	const char *lastname;
534 	dmu_objset_type_t type;
535 	uint64_t flags;
536 };
537 
538 /*ARGSUSED*/
539 static int
540 dmu_objset_create_check(void *arg1, void *arg2, dmu_tx_t *tx)
541 {
542 	dsl_dir_t *dd = arg1;
543 	struct oscarg *oa = arg2;
544 	objset_t *mos = dd->dd_pool->dp_meta_objset;
545 	int err;
546 	uint64_t ddobj;
547 
548 	err = zap_lookup(mos, dd->dd_phys->dd_child_dir_zapobj,
549 	    oa->lastname, sizeof (uint64_t), 1, &ddobj);
550 	if (err != ENOENT)
551 		return (err ? err : EEXIST);
552 
553 	if (oa->clone_parent != NULL) {
554 		/*
555 		 * You can't clone across pools.
556 		 */
557 		if (oa->clone_parent->ds_dir->dd_pool != dd->dd_pool)
558 			return (EXDEV);
559 
560 		/*
561 		 * You can only clone snapshots, not the head datasets.
562 		 */
563 		if (oa->clone_parent->ds_phys->ds_num_children == 0)
564 			return (EINVAL);
565 	}
566 
567 	return (0);
568 }
569 
570 static void
571 dmu_objset_create_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
572 {
573 	dsl_dir_t *dd = arg1;
574 	struct oscarg *oa = arg2;
575 	dsl_dataset_t *ds;
576 	blkptr_t *bp;
577 	uint64_t dsobj;
578 
579 	ASSERT(dmu_tx_is_syncing(tx));
580 
581 	dsobj = dsl_dataset_create_sync(dd, oa->lastname,
582 	    oa->clone_parent, oa->flags, cr, tx);
583 
584 	VERIFY(0 == dsl_dataset_hold_obj(dd->dd_pool, dsobj, FTAG, &ds));
585 	bp = dsl_dataset_get_blkptr(ds);
586 	if (BP_IS_HOLE(bp)) {
587 		objset_impl_t *osi;
588 
589 		/* This is an empty dmu_objset; not a clone. */
590 		osi = dmu_objset_create_impl(dsl_dataset_get_spa(ds),
591 		    ds, bp, oa->type, tx);
592 
593 		if (oa->userfunc)
594 			oa->userfunc(&osi->os, oa->userarg, cr, tx);
595 	}
596 
597 	spa_history_internal_log(LOG_DS_CREATE, dd->dd_pool->dp_spa,
598 	    tx, cr, "dataset = %llu", dsobj);
599 
600 	dsl_dataset_rele(ds, FTAG);
601 }
602 
603 int
604 dmu_objset_create(const char *name, dmu_objset_type_t type,
605     objset_t *clone_parent, uint64_t flags,
606     void (*func)(objset_t *os, void *arg, cred_t *cr, dmu_tx_t *tx), void *arg)
607 {
608 	dsl_dir_t *pdd;
609 	const char *tail;
610 	int err = 0;
611 	struct oscarg oa = { 0 };
612 
613 	ASSERT(strchr(name, '@') == NULL);
614 	err = dsl_dir_open(name, FTAG, &pdd, &tail);
615 	if (err)
616 		return (err);
617 	if (tail == NULL) {
618 		dsl_dir_close(pdd, FTAG);
619 		return (EEXIST);
620 	}
621 
622 	dprintf("name=%s\n", name);
623 
624 	oa.userfunc = func;
625 	oa.userarg = arg;
626 	oa.lastname = tail;
627 	oa.type = type;
628 	oa.flags = flags;
629 
630 	if (clone_parent != NULL) {
631 		/*
632 		 * You can't clone to a different type.
633 		 */
634 		if (clone_parent->os->os_phys->os_type != type) {
635 			dsl_dir_close(pdd, FTAG);
636 			return (EINVAL);
637 		}
638 		oa.clone_parent = clone_parent->os->os_dsl_dataset;
639 	}
640 	err = dsl_sync_task_do(pdd->dd_pool, dmu_objset_create_check,
641 	    dmu_objset_create_sync, pdd, &oa, 5);
642 	dsl_dir_close(pdd, FTAG);
643 	return (err);
644 }
645 
646 int
647 dmu_objset_destroy(const char *name)
648 {
649 	objset_t *os;
650 	int error;
651 
652 	/*
653 	 * If it looks like we'll be able to destroy it, and there's
654 	 * an unplayed replay log sitting around, destroy the log.
655 	 * It would be nicer to do this in dsl_dataset_destroy_sync(),
656 	 * but the replay log objset is modified in open context.
657 	 */
658 	error = dmu_objset_open(name, DMU_OST_ANY,
659 	    DS_MODE_OWNER|DS_MODE_READONLY|DS_MODE_INCONSISTENT, &os);
660 	if (error == 0) {
661 		dsl_dataset_t *ds = os->os->os_dsl_dataset;
662 		zil_destroy(dmu_objset_zil(os), B_FALSE);
663 
664 		error = dsl_dataset_destroy(ds, os);
665 		/*
666 		 * dsl_dataset_destroy() closes the ds.
667 		 */
668 		kmem_free(os, sizeof (objset_t));
669 	}
670 
671 	return (error);
672 }
673 
674 /*
675  * This will close the objset.
676  */
677 int
678 dmu_objset_rollback(objset_t *os)
679 {
680 	int err;
681 	dsl_dataset_t *ds;
682 
683 	ds = os->os->os_dsl_dataset;
684 
685 	if (!dsl_dataset_tryown(ds, TRUE, os)) {
686 		dmu_objset_close(os);
687 		return (EBUSY);
688 	}
689 
690 	err = dsl_dataset_rollback(ds, os->os->os_phys->os_type);
691 
692 	/*
693 	 * NB: we close the objset manually because the rollback
694 	 * actually implicitly called dmu_objset_evict(), thus freeing
695 	 * the objset_impl_t.
696 	 */
697 	dsl_dataset_disown(ds, os);
698 	kmem_free(os, sizeof (objset_t));
699 	return (err);
700 }
701 
702 struct snaparg {
703 	dsl_sync_task_group_t *dstg;
704 	char *snapname;
705 	char failed[MAXPATHLEN];
706 	boolean_t checkperms;
707 	list_t objsets;
708 };
709 
710 struct osnode {
711 	list_node_t node;
712 	objset_t *os;
713 };
714 
715 static int
716 dmu_objset_snapshot_one(char *name, void *arg)
717 {
718 	struct snaparg *sn = arg;
719 	objset_t *os;
720 	int err;
721 
722 	(void) strcpy(sn->failed, name);
723 
724 	/*
725 	 * Check permissions only when requested.  This only applies when
726 	 * doing a recursive snapshot.  The permission checks for the starting
727 	 * dataset have already been performed in zfs_secpolicy_snapshot()
728 	 */
729 	if (sn->checkperms == B_TRUE &&
730 	    (err = zfs_secpolicy_snapshot_perms(name, CRED())))
731 		return (err);
732 
733 	err = dmu_objset_open(name, DMU_OST_ANY, DS_MODE_USER, &os);
734 	if (err != 0)
735 		return (err);
736 
737 	/* If the objset is in an inconsistent state, return busy */
738 	if (os->os->os_dsl_dataset->ds_phys->ds_flags & DS_FLAG_INCONSISTENT) {
739 		dmu_objset_close(os);
740 		return (EBUSY);
741 	}
742 
743 	/*
744 	 * NB: we need to wait for all in-flight changes to get to disk,
745 	 * so that we snapshot those changes.  zil_suspend does this as
746 	 * a side effect.
747 	 */
748 	err = zil_suspend(dmu_objset_zil(os));
749 	if (err == 0) {
750 		struct osnode *osn;
751 		dsl_sync_task_create(sn->dstg, dsl_dataset_snapshot_check,
752 		    dsl_dataset_snapshot_sync, os->os->os_dsl_dataset,
753 		    sn->snapname, 3);
754 		osn = kmem_alloc(sizeof (struct osnode), KM_SLEEP);
755 		osn->os = os;
756 		list_insert_tail(&sn->objsets, osn);
757 	} else {
758 		dmu_objset_close(os);
759 	}
760 
761 	return (err);
762 }
763 
764 int
765 dmu_objset_snapshot(char *fsname, char *snapname, boolean_t recursive)
766 {
767 	dsl_sync_task_t *dst;
768 	struct osnode *osn;
769 	struct snaparg sn = { 0 };
770 	spa_t *spa;
771 	int err;
772 
773 	(void) strcpy(sn.failed, fsname);
774 
775 	err = spa_open(fsname, &spa, FTAG);
776 	if (err)
777 		return (err);
778 
779 	sn.dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
780 	sn.snapname = snapname;
781 	list_create(&sn.objsets, sizeof (struct osnode),
782 	    offsetof(struct osnode, node));
783 
784 	if (recursive) {
785 		sn.checkperms = B_TRUE;
786 		err = dmu_objset_find(fsname,
787 		    dmu_objset_snapshot_one, &sn, DS_FIND_CHILDREN);
788 	} else {
789 		sn.checkperms = B_FALSE;
790 		err = dmu_objset_snapshot_one(fsname, &sn);
791 	}
792 
793 	if (err)
794 		goto out;
795 
796 	err = dsl_sync_task_group_wait(sn.dstg);
797 
798 	for (dst = list_head(&sn.dstg->dstg_tasks); dst;
799 	    dst = list_next(&sn.dstg->dstg_tasks, dst)) {
800 		dsl_dataset_t *ds = dst->dst_arg1;
801 		if (dst->dst_err)
802 			dsl_dataset_name(ds, sn.failed);
803 	}
804 
805 out:
806 	while (osn = list_head(&sn.objsets)) {
807 		list_remove(&sn.objsets, osn);
808 		zil_resume(dmu_objset_zil(osn->os));
809 		dmu_objset_close(osn->os);
810 		kmem_free(osn, sizeof (struct osnode));
811 	}
812 	list_destroy(&sn.objsets);
813 
814 	if (err)
815 		(void) strcpy(fsname, sn.failed);
816 	dsl_sync_task_group_destroy(sn.dstg);
817 	spa_close(spa, FTAG);
818 	return (err);
819 }
820 
821 static void
822 dmu_objset_sync_dnodes(list_t *list, dmu_tx_t *tx)
823 {
824 	dnode_t *dn;
825 
826 	while (dn = list_head(list)) {
827 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
828 		ASSERT(dn->dn_dbuf->db_data_pending);
829 		/*
830 		 * Initialize dn_zio outside dnode_sync()
831 		 * to accomodate meta-dnode
832 		 */
833 		dn->dn_zio = dn->dn_dbuf->db_data_pending->dr_zio;
834 		ASSERT(dn->dn_zio);
835 
836 		ASSERT3U(dn->dn_nlevels, <=, DN_MAX_LEVELS);
837 		list_remove(list, dn);
838 		dnode_sync(dn, tx);
839 	}
840 }
841 
842 /* ARGSUSED */
843 static void
844 ready(zio_t *zio, arc_buf_t *abuf, void *arg)
845 {
846 	blkptr_t *bp = zio->io_bp;
847 	blkptr_t *bp_orig = &zio->io_bp_orig;
848 	objset_impl_t *os = arg;
849 	dnode_phys_t *dnp = &os->os_phys->os_meta_dnode;
850 
851 	ASSERT(bp == os->os_rootbp);
852 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_OBJSET);
853 	ASSERT(BP_GET_LEVEL(bp) == 0);
854 
855 	/*
856 	 * Update rootbp fill count.
857 	 */
858 	bp->blk_fill = 1;	/* count the meta-dnode */
859 	for (int i = 0; i < dnp->dn_nblkptr; i++)
860 		bp->blk_fill += dnp->dn_blkptr[i].blk_fill;
861 
862 	if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
863 		ASSERT(DVA_EQUAL(BP_IDENTITY(bp), BP_IDENTITY(bp_orig)));
864 	} else {
865 		if (zio->io_bp_orig.blk_birth == os->os_synctx->tx_txg)
866 			(void) dsl_dataset_block_kill(os->os_dsl_dataset,
867 			    &zio->io_bp_orig, zio, os->os_synctx);
868 		dsl_dataset_block_born(os->os_dsl_dataset, bp, os->os_synctx);
869 	}
870 }
871 
872 /* called from dsl */
873 void
874 dmu_objset_sync(objset_impl_t *os, zio_t *pio, dmu_tx_t *tx)
875 {
876 	int txgoff;
877 	zbookmark_t zb;
878 	writeprops_t wp = { 0 };
879 	zio_t *zio;
880 	list_t *list;
881 	dbuf_dirty_record_t *dr;
882 
883 	dprintf_ds(os->os_dsl_dataset, "txg=%llu\n", tx->tx_txg);
884 
885 	ASSERT(dmu_tx_is_syncing(tx));
886 	/* XXX the write_done callback should really give us the tx... */
887 	os->os_synctx = tx;
888 
889 	if (os->os_dsl_dataset == NULL) {
890 		/*
891 		 * This is the MOS.  If we have upgraded,
892 		 * spa_max_replication() could change, so reset
893 		 * os_copies here.
894 		 */
895 		os->os_copies = spa_max_replication(os->os_spa);
896 	}
897 
898 	/*
899 	 * Create the root block IO
900 	 */
901 	zb.zb_objset = os->os_dsl_dataset ? os->os_dsl_dataset->ds_object : 0;
902 	zb.zb_object = 0;
903 	zb.zb_level = -1;	/* for block ordering; it's level 0 on disk */
904 	zb.zb_blkid = 0;
905 
906 	wp.wp_type = DMU_OT_OBJSET;
907 	wp.wp_level = 0;	/* on-disk BP level; see above */
908 	wp.wp_copies = os->os_copies;
909 	wp.wp_oschecksum = os->os_checksum;
910 	wp.wp_oscompress = os->os_compress;
911 
912 	if (BP_IS_OLDER(os->os_rootbp, tx->tx_txg)) {
913 		(void) dsl_dataset_block_kill(os->os_dsl_dataset,
914 		    os->os_rootbp, pio, tx);
915 	}
916 
917 	arc_release(os->os_phys_buf, &os->os_phys_buf);
918 	zio = arc_write(pio, os->os_spa, &wp, DMU_OS_IS_L2CACHEABLE(os),
919 	    tx->tx_txg, os->os_rootbp, os->os_phys_buf, ready, NULL, os,
920 	    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
921 
922 	/*
923 	 * Sync meta-dnode - the parent IO for the sync is the root block
924 	 */
925 	os->os_meta_dnode->dn_zio = zio;
926 	dnode_sync(os->os_meta_dnode, tx);
927 
928 	txgoff = tx->tx_txg & TXG_MASK;
929 
930 	dmu_objset_sync_dnodes(&os->os_free_dnodes[txgoff], tx);
931 	dmu_objset_sync_dnodes(&os->os_dirty_dnodes[txgoff], tx);
932 
933 	list = &os->os_meta_dnode->dn_dirty_records[txgoff];
934 	while (dr = list_head(list)) {
935 		ASSERT(dr->dr_dbuf->db_level == 0);
936 		list_remove(list, dr);
937 		if (dr->dr_zio)
938 			zio_nowait(dr->dr_zio);
939 	}
940 	/*
941 	 * Free intent log blocks up to this tx.
942 	 */
943 	zil_sync(os->os_zil, tx);
944 	os->os_phys->os_zil_header = os->os_zil_header;
945 	zio_nowait(zio);
946 }
947 
948 void
949 dmu_objset_space(objset_t *os, uint64_t *refdbytesp, uint64_t *availbytesp,
950     uint64_t *usedobjsp, uint64_t *availobjsp)
951 {
952 	dsl_dataset_space(os->os->os_dsl_dataset, refdbytesp, availbytesp,
953 	    usedobjsp, availobjsp);
954 }
955 
956 uint64_t
957 dmu_objset_fsid_guid(objset_t *os)
958 {
959 	return (dsl_dataset_fsid_guid(os->os->os_dsl_dataset));
960 }
961 
962 void
963 dmu_objset_fast_stat(objset_t *os, dmu_objset_stats_t *stat)
964 {
965 	stat->dds_type = os->os->os_phys->os_type;
966 	if (os->os->os_dsl_dataset)
967 		dsl_dataset_fast_stat(os->os->os_dsl_dataset, stat);
968 }
969 
970 void
971 dmu_objset_stats(objset_t *os, nvlist_t *nv)
972 {
973 	ASSERT(os->os->os_dsl_dataset ||
974 	    os->os->os_phys->os_type == DMU_OST_META);
975 
976 	if (os->os->os_dsl_dataset != NULL)
977 		dsl_dataset_stats(os->os->os_dsl_dataset, nv);
978 
979 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_TYPE,
980 	    os->os->os_phys->os_type);
981 }
982 
983 int
984 dmu_objset_is_snapshot(objset_t *os)
985 {
986 	if (os->os->os_dsl_dataset != NULL)
987 		return (dsl_dataset_is_snapshot(os->os->os_dsl_dataset));
988 	else
989 		return (B_FALSE);
990 }
991 
992 int
993 dmu_snapshot_realname(objset_t *os, char *name, char *real, int maxlen,
994     boolean_t *conflict)
995 {
996 	dsl_dataset_t *ds = os->os->os_dsl_dataset;
997 	uint64_t ignored;
998 
999 	if (ds->ds_phys->ds_snapnames_zapobj == 0)
1000 		return (ENOENT);
1001 
1002 	return (zap_lookup_norm(ds->ds_dir->dd_pool->dp_meta_objset,
1003 	    ds->ds_phys->ds_snapnames_zapobj, name, 8, 1, &ignored, MT_FIRST,
1004 	    real, maxlen, conflict));
1005 }
1006 
1007 int
1008 dmu_snapshot_list_next(objset_t *os, int namelen, char *name,
1009     uint64_t *idp, uint64_t *offp, boolean_t *case_conflict)
1010 {
1011 	dsl_dataset_t *ds = os->os->os_dsl_dataset;
1012 	zap_cursor_t cursor;
1013 	zap_attribute_t attr;
1014 
1015 	if (ds->ds_phys->ds_snapnames_zapobj == 0)
1016 		return (ENOENT);
1017 
1018 	zap_cursor_init_serialized(&cursor,
1019 	    ds->ds_dir->dd_pool->dp_meta_objset,
1020 	    ds->ds_phys->ds_snapnames_zapobj, *offp);
1021 
1022 	if (zap_cursor_retrieve(&cursor, &attr) != 0) {
1023 		zap_cursor_fini(&cursor);
1024 		return (ENOENT);
1025 	}
1026 
1027 	if (strlen(attr.za_name) + 1 > namelen) {
1028 		zap_cursor_fini(&cursor);
1029 		return (ENAMETOOLONG);
1030 	}
1031 
1032 	(void) strcpy(name, attr.za_name);
1033 	if (idp)
1034 		*idp = attr.za_first_integer;
1035 	if (case_conflict)
1036 		*case_conflict = attr.za_normalization_conflict;
1037 	zap_cursor_advance(&cursor);
1038 	*offp = zap_cursor_serialize(&cursor);
1039 	zap_cursor_fini(&cursor);
1040 
1041 	return (0);
1042 }
1043 
1044 int
1045 dmu_dir_list_next(objset_t *os, int namelen, char *name,
1046     uint64_t *idp, uint64_t *offp)
1047 {
1048 	dsl_dir_t *dd = os->os->os_dsl_dataset->ds_dir;
1049 	zap_cursor_t cursor;
1050 	zap_attribute_t attr;
1051 
1052 	/* there is no next dir on a snapshot! */
1053 	if (os->os->os_dsl_dataset->ds_object !=
1054 	    dd->dd_phys->dd_head_dataset_obj)
1055 		return (ENOENT);
1056 
1057 	zap_cursor_init_serialized(&cursor,
1058 	    dd->dd_pool->dp_meta_objset,
1059 	    dd->dd_phys->dd_child_dir_zapobj, *offp);
1060 
1061 	if (zap_cursor_retrieve(&cursor, &attr) != 0) {
1062 		zap_cursor_fini(&cursor);
1063 		return (ENOENT);
1064 	}
1065 
1066 	if (strlen(attr.za_name) + 1 > namelen) {
1067 		zap_cursor_fini(&cursor);
1068 		return (ENAMETOOLONG);
1069 	}
1070 
1071 	(void) strcpy(name, attr.za_name);
1072 	if (idp)
1073 		*idp = attr.za_first_integer;
1074 	zap_cursor_advance(&cursor);
1075 	*offp = zap_cursor_serialize(&cursor);
1076 	zap_cursor_fini(&cursor);
1077 
1078 	return (0);
1079 }
1080 
1081 struct findarg {
1082 	int (*func)(char *, void *);
1083 	void *arg;
1084 };
1085 
1086 /* ARGSUSED */
1087 static int
1088 findfunc(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1089 {
1090 	struct findarg *fa = arg;
1091 	return (fa->func((char *)dsname, fa->arg));
1092 }
1093 
1094 /*
1095  * Find all objsets under name, and for each, call 'func(child_name, arg)'.
1096  * Perhaps change all callers to use dmu_objset_find_spa()?
1097  */
1098 int
1099 dmu_objset_find(char *name, int func(char *, void *), void *arg, int flags)
1100 {
1101 	struct findarg fa;
1102 	fa.func = func;
1103 	fa.arg = arg;
1104 	return (dmu_objset_find_spa(NULL, name, findfunc, &fa, flags));
1105 }
1106 
1107 /*
1108  * Find all objsets under name, call func on each
1109  */
1110 int
1111 dmu_objset_find_spa(spa_t *spa, const char *name,
1112     int func(spa_t *, uint64_t, const char *, void *), void *arg, int flags)
1113 {
1114 	dsl_dir_t *dd;
1115 	dsl_pool_t *dp;
1116 	dsl_dataset_t *ds;
1117 	zap_cursor_t zc;
1118 	zap_attribute_t *attr;
1119 	char *child;
1120 	uint64_t thisobj;
1121 	int err;
1122 
1123 	if (name == NULL)
1124 		name = spa_name(spa);
1125 	err = dsl_dir_open_spa(spa, name, FTAG, &dd, NULL);
1126 	if (err)
1127 		return (err);
1128 
1129 	/* Don't visit hidden ($MOS & $ORIGIN) objsets. */
1130 	if (dd->dd_myname[0] == '$') {
1131 		dsl_dir_close(dd, FTAG);
1132 		return (0);
1133 	}
1134 
1135 	thisobj = dd->dd_phys->dd_head_dataset_obj;
1136 	attr = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
1137 	dp = dd->dd_pool;
1138 
1139 	/*
1140 	 * Iterate over all children.
1141 	 */
1142 	if (flags & DS_FIND_CHILDREN) {
1143 		for (zap_cursor_init(&zc, dp->dp_meta_objset,
1144 		    dd->dd_phys->dd_child_dir_zapobj);
1145 		    zap_cursor_retrieve(&zc, attr) == 0;
1146 		    (void) zap_cursor_advance(&zc)) {
1147 			ASSERT(attr->za_integer_length == sizeof (uint64_t));
1148 			ASSERT(attr->za_num_integers == 1);
1149 
1150 			child = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1151 			(void) strcpy(child, name);
1152 			(void) strcat(child, "/");
1153 			(void) strcat(child, attr->za_name);
1154 			err = dmu_objset_find_spa(spa, child, func, arg, flags);
1155 			kmem_free(child, MAXPATHLEN);
1156 			if (err)
1157 				break;
1158 		}
1159 		zap_cursor_fini(&zc);
1160 
1161 		if (err) {
1162 			dsl_dir_close(dd, FTAG);
1163 			kmem_free(attr, sizeof (zap_attribute_t));
1164 			return (err);
1165 		}
1166 	}
1167 
1168 	/*
1169 	 * Iterate over all snapshots.
1170 	 */
1171 	if (flags & DS_FIND_SNAPSHOTS) {
1172 		if (!dsl_pool_sync_context(dp))
1173 			rw_enter(&dp->dp_config_rwlock, RW_READER);
1174 		err = dsl_dataset_hold_obj(dp, thisobj, FTAG, &ds);
1175 		if (!dsl_pool_sync_context(dp))
1176 			rw_exit(&dp->dp_config_rwlock);
1177 
1178 		if (err == 0) {
1179 			uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
1180 			dsl_dataset_rele(ds, FTAG);
1181 
1182 			for (zap_cursor_init(&zc, dp->dp_meta_objset, snapobj);
1183 			    zap_cursor_retrieve(&zc, attr) == 0;
1184 			    (void) zap_cursor_advance(&zc)) {
1185 				ASSERT(attr->za_integer_length ==
1186 				    sizeof (uint64_t));
1187 				ASSERT(attr->za_num_integers == 1);
1188 
1189 				child = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1190 				(void) strcpy(child, name);
1191 				(void) strcat(child, "@");
1192 				(void) strcat(child, attr->za_name);
1193 				err = func(spa, attr->za_first_integer,
1194 				    child, arg);
1195 				kmem_free(child, MAXPATHLEN);
1196 				if (err)
1197 					break;
1198 			}
1199 			zap_cursor_fini(&zc);
1200 		}
1201 	}
1202 
1203 	dsl_dir_close(dd, FTAG);
1204 	kmem_free(attr, sizeof (zap_attribute_t));
1205 
1206 	if (err)
1207 		return (err);
1208 
1209 	/*
1210 	 * Apply to self if appropriate.
1211 	 */
1212 	err = func(spa, thisobj, name, arg);
1213 	return (err);
1214 }
1215 
1216 void
1217 dmu_objset_set_user(objset_t *os, void *user_ptr)
1218 {
1219 	ASSERT(MUTEX_HELD(&os->os->os_user_ptr_lock));
1220 	os->os->os_user_ptr = user_ptr;
1221 }
1222 
1223 void *
1224 dmu_objset_get_user(objset_t *os)
1225 {
1226 	ASSERT(MUTEX_HELD(&os->os->os_user_ptr_lock));
1227 	return (os->os->os_user_ptr);
1228 }
1229