xref: /titanic_41/usr/src/uts/common/fs/zfs/dmu_objset.c (revision 1dd08564e4a3aafe66b00aee6f222b0885346fe8)
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 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/cred.h>
29 #include <sys/zfs_context.h>
30 #include <sys/dmu_objset.h>
31 #include <sys/dsl_dir.h>
32 #include <sys/dsl_dataset.h>
33 #include <sys/dsl_prop.h>
34 #include <sys/dsl_pool.h>
35 #include <sys/dsl_synctask.h>
36 #include <sys/dsl_deleg.h>
37 #include <sys/dnode.h>
38 #include <sys/dbuf.h>
39 #include <sys/zvol.h>
40 #include <sys/dmu_tx.h>
41 #include <sys/zio_checksum.h>
42 #include <sys/zap.h>
43 #include <sys/zil.h>
44 #include <sys/dmu_impl.h>
45 #include <sys/zfs_ioctl.h>
46 
47 spa_t *
48 dmu_objset_spa(objset_t *os)
49 {
50 	return (os->os->os_spa);
51 }
52 
53 zilog_t *
54 dmu_objset_zil(objset_t *os)
55 {
56 	return (os->os->os_zil);
57 }
58 
59 dsl_pool_t *
60 dmu_objset_pool(objset_t *os)
61 {
62 	dsl_dataset_t *ds;
63 
64 	if ((ds = os->os->os_dsl_dataset) != NULL && ds->ds_dir)
65 		return (ds->ds_dir->dd_pool);
66 	else
67 		return (spa_get_dsl(os->os->os_spa));
68 }
69 
70 dsl_dataset_t *
71 dmu_objset_ds(objset_t *os)
72 {
73 	return (os->os->os_dsl_dataset);
74 }
75 
76 dmu_objset_type_t
77 dmu_objset_type(objset_t *os)
78 {
79 	return (os->os->os_phys->os_type);
80 }
81 
82 void
83 dmu_objset_name(objset_t *os, char *buf)
84 {
85 	dsl_dataset_name(os->os->os_dsl_dataset, buf);
86 }
87 
88 uint64_t
89 dmu_objset_id(objset_t *os)
90 {
91 	dsl_dataset_t *ds = os->os->os_dsl_dataset;
92 
93 	return (ds ? ds->ds_object : 0);
94 }
95 
96 static void
97 checksum_changed_cb(void *arg, uint64_t newval)
98 {
99 	objset_impl_t *osi = arg;
100 
101 	/*
102 	 * Inheritance should have been done by now.
103 	 */
104 	ASSERT(newval != ZIO_CHECKSUM_INHERIT);
105 
106 	osi->os_checksum = zio_checksum_select(newval, ZIO_CHECKSUM_ON_VALUE);
107 }
108 
109 static void
110 compression_changed_cb(void *arg, uint64_t newval)
111 {
112 	objset_impl_t *osi = arg;
113 
114 	/*
115 	 * Inheritance and range checking should have been done by now.
116 	 */
117 	ASSERT(newval != ZIO_COMPRESS_INHERIT);
118 
119 	osi->os_compress = zio_compress_select(newval, ZIO_COMPRESS_ON_VALUE);
120 }
121 
122 static void
123 copies_changed_cb(void *arg, uint64_t newval)
124 {
125 	objset_impl_t *osi = arg;
126 
127 	/*
128 	 * Inheritance and range checking should have been done by now.
129 	 */
130 	ASSERT(newval > 0);
131 	ASSERT(newval <= spa_max_replication(osi->os_spa));
132 
133 	osi->os_copies = newval;
134 }
135 
136 void
137 dmu_objset_byteswap(void *buf, size_t size)
138 {
139 	objset_phys_t *osp = buf;
140 
141 	ASSERT(size == sizeof (objset_phys_t));
142 	dnode_byteswap(&osp->os_meta_dnode);
143 	byteswap_uint64_array(&osp->os_zil_header, sizeof (zil_header_t));
144 	osp->os_type = BSWAP_64(osp->os_type);
145 }
146 
147 int
148 dmu_objset_open_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
149     objset_impl_t **osip)
150 {
151 	objset_impl_t *osi;
152 	int i, err, checksum;
153 
154 	ASSERT(ds == NULL || MUTEX_HELD(&ds->ds_opening_lock));
155 
156 	osi = kmem_zalloc(sizeof (objset_impl_t), KM_SLEEP);
157 	osi->os.os = osi;
158 	osi->os_dsl_dataset = ds;
159 	osi->os_spa = spa;
160 	osi->os_rootbp = bp;
161 	if (!BP_IS_HOLE(osi->os_rootbp)) {
162 		uint32_t aflags = ARC_WAIT;
163 		zbookmark_t zb;
164 		zb.zb_objset = ds ? ds->ds_object : 0;
165 		zb.zb_object = 0;
166 		zb.zb_level = -1;
167 		zb.zb_blkid = 0;
168 
169 		dprintf_bp(osi->os_rootbp, "reading %s", "");
170 		err = arc_read(NULL, spa, osi->os_rootbp,
171 		    dmu_ot[DMU_OT_OBJSET].ot_byteswap,
172 		    arc_getbuf_func, &osi->os_phys_buf,
173 		    ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL, &aflags, &zb);
174 		if (err) {
175 			kmem_free(osi, sizeof (objset_impl_t));
176 			return (err);
177 		}
178 		osi->os_phys = osi->os_phys_buf->b_data;
179 		if (ds == NULL || dsl_dataset_is_snapshot(ds) == 0)
180 			arc_release(osi->os_phys_buf, &osi->os_phys_buf);
181 	} else {
182 		osi->os_phys_buf = arc_buf_alloc(spa, sizeof (objset_phys_t),
183 		    &osi->os_phys_buf, ARC_BUFC_METADATA);
184 		osi->os_phys = osi->os_phys_buf->b_data;
185 		bzero(osi->os_phys, sizeof (objset_phys_t));
186 	}
187 
188 	/*
189 	 * Note: the changed_cb will be called once before the register
190 	 * func returns, thus changing the checksum/compression from the
191 	 * default (fletcher2/off).  Snapshots don't need to know, and
192 	 * registering would complicate clone promotion.
193 	 */
194 	if (ds && ds->ds_phys->ds_num_children == 0) {
195 		err = dsl_prop_register(ds, "checksum",
196 		    checksum_changed_cb, osi);
197 		if (err == 0)
198 			err = dsl_prop_register(ds, "compression",
199 			    compression_changed_cb, osi);
200 		if (err == 0)
201 			err = dsl_prop_register(ds, "copies",
202 			    copies_changed_cb, osi);
203 		if (err) {
204 			VERIFY(arc_buf_remove_ref(osi->os_phys_buf,
205 			    &osi->os_phys_buf) == 1);
206 			kmem_free(osi, sizeof (objset_impl_t));
207 			return (err);
208 		}
209 	} else if (ds == NULL) {
210 		/* It's the meta-objset. */
211 		osi->os_checksum = ZIO_CHECKSUM_FLETCHER_4;
212 		osi->os_compress = ZIO_COMPRESS_LZJB;
213 		osi->os_copies = spa_max_replication(spa);
214 	}
215 
216 	osi->os_zil = zil_alloc(&osi->os, &osi->os_phys->os_zil_header);
217 
218 	/*
219 	 * Metadata always gets compressed and checksummed.
220 	 * If the data checksum is multi-bit correctable, and it's not
221 	 * a ZBT-style checksum, then it's suitable for metadata as well.
222 	 * Otherwise, the metadata checksum defaults to fletcher4.
223 	 */
224 	checksum = osi->os_checksum;
225 
226 	if (zio_checksum_table[checksum].ci_correctable &&
227 	    !zio_checksum_table[checksum].ci_zbt)
228 		osi->os_md_checksum = checksum;
229 	else
230 		osi->os_md_checksum = ZIO_CHECKSUM_FLETCHER_4;
231 	osi->os_md_compress = ZIO_COMPRESS_LZJB;
232 
233 	for (i = 0; i < TXG_SIZE; i++) {
234 		list_create(&osi->os_dirty_dnodes[i], sizeof (dnode_t),
235 		    offsetof(dnode_t, dn_dirty_link[i]));
236 		list_create(&osi->os_free_dnodes[i], sizeof (dnode_t),
237 		    offsetof(dnode_t, dn_dirty_link[i]));
238 	}
239 	list_create(&osi->os_dnodes, sizeof (dnode_t),
240 	    offsetof(dnode_t, dn_link));
241 	list_create(&osi->os_downgraded_dbufs, sizeof (dmu_buf_impl_t),
242 	    offsetof(dmu_buf_impl_t, db_link));
243 
244 	mutex_init(&osi->os_lock, NULL, MUTEX_DEFAULT, NULL);
245 	mutex_init(&osi->os_obj_lock, NULL, MUTEX_DEFAULT, NULL);
246 	mutex_init(&osi->os_user_ptr_lock, NULL, MUTEX_DEFAULT, NULL);
247 
248 	osi->os_meta_dnode = dnode_special_open(osi,
249 	    &osi->os_phys->os_meta_dnode, DMU_META_DNODE_OBJECT);
250 
251 	/*
252 	 * We should be the only thread trying to do this because we
253 	 * have ds_opening_lock
254 	 */
255 	if (ds) {
256 		VERIFY(NULL == dsl_dataset_set_user_ptr(ds, osi,
257 		    dmu_objset_evict));
258 	}
259 
260 	*osip = osi;
261 	return (0);
262 }
263 
264 static int
265 dmu_objset_open_ds_os(dsl_dataset_t *ds, objset_t *os, dmu_objset_type_t type)
266 {
267 	objset_impl_t *osi;
268 
269 	mutex_enter(&ds->ds_opening_lock);
270 	osi = dsl_dataset_get_user_ptr(ds);
271 	if (osi == NULL) {
272 		int err;
273 
274 		err = dmu_objset_open_impl(dsl_dataset_get_spa(ds),
275 		    ds, &ds->ds_phys->ds_bp, &osi);
276 		if (err) {
277 			mutex_exit(&ds->ds_opening_lock);
278 			return (err);
279 		}
280 	}
281 	mutex_exit(&ds->ds_opening_lock);
282 
283 	os->os = osi;
284 	os->os_mode = DS_MODE_NOHOLD;
285 
286 	if (type != DMU_OST_ANY && type != os->os->os_phys->os_type)
287 		return (EINVAL);
288 	return (0);
289 }
290 
291 int
292 dmu_objset_open_ds(dsl_dataset_t *ds, dmu_objset_type_t type, objset_t **osp)
293 {
294 	objset_t *os;
295 	int err;
296 
297 	os = kmem_alloc(sizeof (objset_t), KM_SLEEP);
298 	err = dmu_objset_open_ds_os(ds, os, type);
299 	if (err)
300 		kmem_free(os, sizeof (objset_t));
301 	else
302 		*osp = os;
303 	return (err);
304 }
305 
306 /* called from zpl */
307 int
308 dmu_objset_open(const char *name, dmu_objset_type_t type, int mode,
309     objset_t **osp)
310 {
311 	objset_t *os;
312 	dsl_dataset_t *ds;
313 	int err;
314 
315 	ASSERT(DS_MODE_TYPE(mode) == DS_MODE_USER ||
316 	    DS_MODE_TYPE(mode) == DS_MODE_OWNER);
317 
318 	os = kmem_alloc(sizeof (objset_t), KM_SLEEP);
319 	if (DS_MODE_TYPE(mode) == DS_MODE_USER)
320 		err = dsl_dataset_hold(name, os, &ds);
321 	else
322 		err = dsl_dataset_own(name, mode, os, &ds);
323 	if (err) {
324 		kmem_free(os, sizeof (objset_t));
325 		return (err);
326 	}
327 
328 	err = dmu_objset_open_ds_os(ds, os, type);
329 	if (err) {
330 		if (DS_MODE_TYPE(mode) == DS_MODE_USER)
331 			dsl_dataset_rele(ds, os);
332 		else
333 			dsl_dataset_disown(ds, os);
334 		kmem_free(os, sizeof (objset_t));
335 	} else {
336 		os->os_mode = mode;
337 		*osp = os;
338 	}
339 	return (err);
340 }
341 
342 void
343 dmu_objset_close(objset_t *os)
344 {
345 	ASSERT(DS_MODE_TYPE(os->os_mode) == DS_MODE_USER ||
346 	    DS_MODE_TYPE(os->os_mode) == DS_MODE_OWNER ||
347 	    DS_MODE_TYPE(os->os_mode) == DS_MODE_NOHOLD);
348 
349 	if (DS_MODE_TYPE(os->os_mode) == DS_MODE_USER)
350 		dsl_dataset_rele(os->os->os_dsl_dataset, os);
351 	else if (DS_MODE_TYPE(os->os_mode) == DS_MODE_OWNER)
352 		dsl_dataset_disown(os->os->os_dsl_dataset, os);
353 	kmem_free(os, sizeof (objset_t));
354 }
355 
356 int
357 dmu_objset_evict_dbufs(objset_t *os)
358 {
359 	objset_impl_t *osi = os->os;
360 	dnode_t *dn;
361 
362 	mutex_enter(&osi->os_lock);
363 
364 	/* process the mdn last, since the other dnodes have holds on it */
365 	list_remove(&osi->os_dnodes, osi->os_meta_dnode);
366 	list_insert_tail(&osi->os_dnodes, osi->os_meta_dnode);
367 
368 	/*
369 	 * Find the first dnode with holds.  We have to do this dance
370 	 * because dnode_add_ref() only works if you already have a
371 	 * hold.  If there are no holds then it has no dbufs so OK to
372 	 * skip.
373 	 */
374 	for (dn = list_head(&osi->os_dnodes);
375 	    dn && !dnode_add_ref(dn, FTAG);
376 	    dn = list_next(&osi->os_dnodes, dn))
377 		continue;
378 
379 	while (dn) {
380 		dnode_t *next_dn = dn;
381 
382 		do {
383 			next_dn = list_next(&osi->os_dnodes, next_dn);
384 		} while (next_dn && !dnode_add_ref(next_dn, FTAG));
385 
386 		mutex_exit(&osi->os_lock);
387 		dnode_evict_dbufs(dn);
388 		dnode_rele(dn, FTAG);
389 		mutex_enter(&osi->os_lock);
390 		dn = next_dn;
391 	}
392 	mutex_exit(&osi->os_lock);
393 	return (list_head(&osi->os_dnodes) != osi->os_meta_dnode);
394 }
395 
396 void
397 dmu_objset_evict(dsl_dataset_t *ds, void *arg)
398 {
399 	objset_impl_t *osi = arg;
400 	objset_t os;
401 	int i;
402 
403 	for (i = 0; i < TXG_SIZE; i++) {
404 		ASSERT(list_head(&osi->os_dirty_dnodes[i]) == NULL);
405 		ASSERT(list_head(&osi->os_free_dnodes[i]) == NULL);
406 	}
407 
408 	if (ds && ds->ds_phys && ds->ds_phys->ds_num_children == 0) {
409 		VERIFY(0 == dsl_prop_unregister(ds, "checksum",
410 		    checksum_changed_cb, osi));
411 		VERIFY(0 == dsl_prop_unregister(ds, "compression",
412 		    compression_changed_cb, osi));
413 		VERIFY(0 == dsl_prop_unregister(ds, "copies",
414 		    copies_changed_cb, osi));
415 	}
416 
417 	/*
418 	 * We should need only a single pass over the dnode list, since
419 	 * nothing can be added to the list at this point.
420 	 */
421 	os.os = osi;
422 	(void) dmu_objset_evict_dbufs(&os);
423 
424 	ASSERT3P(list_head(&osi->os_dnodes), ==, osi->os_meta_dnode);
425 	ASSERT3P(list_tail(&osi->os_dnodes), ==, osi->os_meta_dnode);
426 	ASSERT3P(list_head(&osi->os_meta_dnode->dn_dbufs), ==, NULL);
427 
428 	dnode_special_close(osi->os_meta_dnode);
429 	zil_free(osi->os_zil);
430 
431 	VERIFY(arc_buf_remove_ref(osi->os_phys_buf, &osi->os_phys_buf) == 1);
432 	mutex_destroy(&osi->os_lock);
433 	mutex_destroy(&osi->os_obj_lock);
434 	mutex_destroy(&osi->os_user_ptr_lock);
435 	kmem_free(osi, sizeof (objset_impl_t));
436 }
437 
438 /* called from dsl for meta-objset */
439 objset_impl_t *
440 dmu_objset_create_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
441     dmu_objset_type_t type, dmu_tx_t *tx)
442 {
443 	objset_impl_t *osi;
444 	dnode_t *mdn;
445 
446 	ASSERT(dmu_tx_is_syncing(tx));
447 	if (ds)
448 		mutex_enter(&ds->ds_opening_lock);
449 	VERIFY(0 == dmu_objset_open_impl(spa, ds, bp, &osi));
450 	if (ds)
451 		mutex_exit(&ds->ds_opening_lock);
452 	mdn = osi->os_meta_dnode;
453 
454 	dnode_allocate(mdn, DMU_OT_DNODE, 1 << DNODE_BLOCK_SHIFT,
455 	    DN_MAX_INDBLKSHIFT, DMU_OT_NONE, 0, tx);
456 
457 	/*
458 	 * We don't want to have to increase the meta-dnode's nlevels
459 	 * later, because then we could do it in quescing context while
460 	 * we are also accessing it in open context.
461 	 *
462 	 * This precaution is not necessary for the MOS (ds == NULL),
463 	 * because the MOS is only updated in syncing context.
464 	 * This is most fortunate: the MOS is the only objset that
465 	 * needs to be synced multiple times as spa_sync() iterates
466 	 * to convergence, so minimizing its dn_nlevels matters.
467 	 */
468 	if (ds != NULL) {
469 		int levels = 1;
470 
471 		/*
472 		 * Determine the number of levels necessary for the meta-dnode
473 		 * to contain DN_MAX_OBJECT dnodes.
474 		 */
475 		while ((uint64_t)mdn->dn_nblkptr << (mdn->dn_datablkshift +
476 		    (levels - 1) * (mdn->dn_indblkshift - SPA_BLKPTRSHIFT)) <
477 		    DN_MAX_OBJECT * sizeof (dnode_phys_t))
478 			levels++;
479 
480 		mdn->dn_next_nlevels[tx->tx_txg & TXG_MASK] =
481 		    mdn->dn_nlevels = levels;
482 	}
483 
484 	ASSERT(type != DMU_OST_NONE);
485 	ASSERT(type != DMU_OST_ANY);
486 	ASSERT(type < DMU_OST_NUMTYPES);
487 	osi->os_phys->os_type = type;
488 
489 	dsl_dataset_dirty(ds, tx);
490 
491 	return (osi);
492 }
493 
494 struct oscarg {
495 	void (*userfunc)(objset_t *os, void *arg, cred_t *cr, dmu_tx_t *tx);
496 	void *userarg;
497 	dsl_dataset_t *clone_parent;
498 	const char *lastname;
499 	dmu_objset_type_t type;
500 	uint64_t flags;
501 };
502 
503 /*ARGSUSED*/
504 static int
505 dmu_objset_create_check(void *arg1, void *arg2, dmu_tx_t *tx)
506 {
507 	dsl_dir_t *dd = arg1;
508 	struct oscarg *oa = arg2;
509 	objset_t *mos = dd->dd_pool->dp_meta_objset;
510 	int err;
511 	uint64_t ddobj;
512 
513 	err = zap_lookup(mos, dd->dd_phys->dd_child_dir_zapobj,
514 	    oa->lastname, sizeof (uint64_t), 1, &ddobj);
515 	if (err != ENOENT)
516 		return (err ? err : EEXIST);
517 
518 	if (oa->clone_parent != NULL) {
519 		/*
520 		 * You can't clone across pools.
521 		 */
522 		if (oa->clone_parent->ds_dir->dd_pool != dd->dd_pool)
523 			return (EXDEV);
524 
525 		/*
526 		 * You can only clone snapshots, not the head datasets.
527 		 */
528 		if (oa->clone_parent->ds_phys->ds_num_children == 0)
529 			return (EINVAL);
530 	}
531 
532 	return (0);
533 }
534 
535 static void
536 dmu_objset_create_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
537 {
538 	dsl_dir_t *dd = arg1;
539 	struct oscarg *oa = arg2;
540 	dsl_dataset_t *ds;
541 	blkptr_t *bp;
542 	uint64_t dsobj;
543 
544 	ASSERT(dmu_tx_is_syncing(tx));
545 
546 	dsobj = dsl_dataset_create_sync(dd, oa->lastname,
547 	    oa->clone_parent, oa->flags, cr, tx);
548 
549 	VERIFY(0 == dsl_dataset_hold_obj(dd->dd_pool, dsobj, FTAG, &ds));
550 	bp = dsl_dataset_get_blkptr(ds);
551 	if (BP_IS_HOLE(bp)) {
552 		objset_impl_t *osi;
553 
554 		/* This is an empty dmu_objset; not a clone. */
555 		osi = dmu_objset_create_impl(dsl_dataset_get_spa(ds),
556 		    ds, bp, oa->type, tx);
557 
558 		if (oa->userfunc)
559 			oa->userfunc(&osi->os, oa->userarg, cr, tx);
560 	}
561 
562 	spa_history_internal_log(LOG_DS_CREATE, dd->dd_pool->dp_spa,
563 	    tx, cr, "dataset = %llu", dsobj);
564 
565 	dsl_dataset_rele(ds, FTAG);
566 }
567 
568 int
569 dmu_objset_create(const char *name, dmu_objset_type_t type,
570     objset_t *clone_parent, uint64_t flags,
571     void (*func)(objset_t *os, void *arg, cred_t *cr, dmu_tx_t *tx), void *arg)
572 {
573 	dsl_dir_t *pdd;
574 	const char *tail;
575 	int err = 0;
576 	struct oscarg oa = { 0 };
577 
578 	ASSERT(strchr(name, '@') == NULL);
579 	err = dsl_dir_open(name, FTAG, &pdd, &tail);
580 	if (err)
581 		return (err);
582 	if (tail == NULL) {
583 		dsl_dir_close(pdd, FTAG);
584 		return (EEXIST);
585 	}
586 
587 	dprintf("name=%s\n", name);
588 
589 	oa.userfunc = func;
590 	oa.userarg = arg;
591 	oa.lastname = tail;
592 	oa.type = type;
593 	oa.flags = flags;
594 
595 	if (clone_parent != NULL) {
596 		/*
597 		 * You can't clone to a different type.
598 		 */
599 		if (clone_parent->os->os_phys->os_type != type) {
600 			dsl_dir_close(pdd, FTAG);
601 			return (EINVAL);
602 		}
603 		oa.clone_parent = clone_parent->os->os_dsl_dataset;
604 	}
605 	err = dsl_sync_task_do(pdd->dd_pool, dmu_objset_create_check,
606 	    dmu_objset_create_sync, pdd, &oa, 5);
607 	dsl_dir_close(pdd, FTAG);
608 	return (err);
609 }
610 
611 int
612 dmu_objset_destroy(const char *name)
613 {
614 	objset_t *os;
615 	int error;
616 
617 	/*
618 	 * If it looks like we'll be able to destroy it, and there's
619 	 * an unplayed replay log sitting around, destroy the log.
620 	 * It would be nicer to do this in dsl_dataset_destroy_sync(),
621 	 * but the replay log objset is modified in open context.
622 	 */
623 	error = dmu_objset_open(name, DMU_OST_ANY,
624 	    DS_MODE_OWNER|DS_MODE_READONLY|DS_MODE_INCONSISTENT, &os);
625 	if (error == 0) {
626 		dsl_dataset_t *ds = os->os->os_dsl_dataset;
627 		zil_destroy(dmu_objset_zil(os), B_FALSE);
628 
629 		error = dsl_dataset_destroy(ds, os);
630 		/*
631 		 * dsl_dataset_destroy() closes the ds.
632 		 */
633 		kmem_free(os, sizeof (objset_t));
634 	}
635 
636 	return (error);
637 }
638 
639 /*
640  * This will close the objset.
641  */
642 int
643 dmu_objset_rollback(objset_t *os)
644 {
645 	int err;
646 	dsl_dataset_t *ds;
647 
648 	ds = os->os->os_dsl_dataset;
649 
650 	if (!dsl_dataset_tryown(ds, TRUE, os)) {
651 		dmu_objset_close(os);
652 		return (EBUSY);
653 	}
654 
655 	err = dsl_dataset_rollback(ds, os->os->os_phys->os_type);
656 
657 	/*
658 	 * NB: we close the objset manually because the rollback
659 	 * actually implicitly called dmu_objset_evict(), thus freeing
660 	 * the objset_impl_t.
661 	 */
662 	dsl_dataset_disown(ds, os);
663 	kmem_free(os, sizeof (objset_t));
664 	return (err);
665 }
666 
667 struct snaparg {
668 	dsl_sync_task_group_t *dstg;
669 	char *snapname;
670 	char failed[MAXPATHLEN];
671 	boolean_t checkperms;
672 	list_t objsets;
673 };
674 
675 struct osnode {
676 	list_node_t node;
677 	objset_t *os;
678 };
679 
680 static int
681 dmu_objset_snapshot_one(char *name, void *arg)
682 {
683 	struct snaparg *sn = arg;
684 	objset_t *os;
685 	int err;
686 
687 	(void) strcpy(sn->failed, name);
688 
689 	/*
690 	 * Check permissions only when requested.  This only applies when
691 	 * doing a recursive snapshot.  The permission checks for the starting
692 	 * dataset have already been performed in zfs_secpolicy_snapshot()
693 	 */
694 	if (sn->checkperms == B_TRUE &&
695 	    (err = zfs_secpolicy_snapshot_perms(name, CRED())))
696 		return (err);
697 
698 	err = dmu_objset_open(name, DMU_OST_ANY, DS_MODE_USER, &os);
699 	if (err != 0)
700 		return (err);
701 
702 	/* If the objset is in an inconsistent state, return busy */
703 	if (os->os->os_dsl_dataset->ds_phys->ds_flags & DS_FLAG_INCONSISTENT) {
704 		dmu_objset_close(os);
705 		return (EBUSY);
706 	}
707 
708 	/*
709 	 * NB: we need to wait for all in-flight changes to get to disk,
710 	 * so that we snapshot those changes.  zil_suspend does this as
711 	 * a side effect.
712 	 */
713 	err = zil_suspend(dmu_objset_zil(os));
714 	if (err == 0) {
715 		struct osnode *osn;
716 		dsl_sync_task_create(sn->dstg, dsl_dataset_snapshot_check,
717 		    dsl_dataset_snapshot_sync, os->os->os_dsl_dataset,
718 		    sn->snapname, 3);
719 		osn = kmem_alloc(sizeof (struct osnode), KM_SLEEP);
720 		osn->os = os;
721 		list_insert_tail(&sn->objsets, osn);
722 	} else {
723 		dmu_objset_close(os);
724 	}
725 
726 	return (err);
727 }
728 
729 int
730 dmu_objset_snapshot(char *fsname, char *snapname, boolean_t recursive)
731 {
732 	dsl_sync_task_t *dst;
733 	struct osnode *osn;
734 	struct snaparg sn = { 0 };
735 	spa_t *spa;
736 	int err;
737 
738 	(void) strcpy(sn.failed, fsname);
739 
740 	err = spa_open(fsname, &spa, FTAG);
741 	if (err)
742 		return (err);
743 
744 	sn.dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
745 	sn.snapname = snapname;
746 	list_create(&sn.objsets, sizeof (struct osnode),
747 	    offsetof(struct osnode, node));
748 
749 	if (recursive) {
750 		sn.checkperms = B_TRUE;
751 		err = dmu_objset_find(fsname,
752 		    dmu_objset_snapshot_one, &sn, DS_FIND_CHILDREN);
753 	} else {
754 		sn.checkperms = B_FALSE;
755 		err = dmu_objset_snapshot_one(fsname, &sn);
756 	}
757 
758 	if (err)
759 		goto out;
760 
761 	err = dsl_sync_task_group_wait(sn.dstg);
762 
763 	for (dst = list_head(&sn.dstg->dstg_tasks); dst;
764 	    dst = list_next(&sn.dstg->dstg_tasks, dst)) {
765 		dsl_dataset_t *ds = dst->dst_arg1;
766 		if (dst->dst_err)
767 			dsl_dataset_name(ds, sn.failed);
768 	}
769 
770 out:
771 	while (osn = list_head(&sn.objsets)) {
772 		list_remove(&sn.objsets, osn);
773 		zil_resume(dmu_objset_zil(osn->os));
774 		dmu_objset_close(osn->os);
775 		kmem_free(osn, sizeof (struct osnode));
776 	}
777 	list_destroy(&sn.objsets);
778 
779 	if (err)
780 		(void) strcpy(fsname, sn.failed);
781 	dsl_sync_task_group_destroy(sn.dstg);
782 	spa_close(spa, FTAG);
783 	return (err);
784 }
785 
786 static void
787 dmu_objset_sync_dnodes(list_t *list, dmu_tx_t *tx)
788 {
789 	dnode_t *dn;
790 
791 	while (dn = list_head(list)) {
792 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
793 		ASSERT(dn->dn_dbuf->db_data_pending);
794 		/*
795 		 * Initialize dn_zio outside dnode_sync()
796 		 * to accomodate meta-dnode
797 		 */
798 		dn->dn_zio = dn->dn_dbuf->db_data_pending->dr_zio;
799 		ASSERT(dn->dn_zio);
800 
801 		ASSERT3U(dn->dn_nlevels, <=, DN_MAX_LEVELS);
802 		list_remove(list, dn);
803 		dnode_sync(dn, tx);
804 	}
805 }
806 
807 /* ARGSUSED */
808 static void
809 ready(zio_t *zio, arc_buf_t *abuf, void *arg)
810 {
811 	objset_impl_t *os = arg;
812 	blkptr_t *bp = os->os_rootbp;
813 	dnode_phys_t *dnp = &os->os_phys->os_meta_dnode;
814 	int i;
815 
816 	ASSERT(bp == zio->io_bp);
817 
818 	/*
819 	 * Update rootbp fill count.
820 	 */
821 	bp->blk_fill = 1;	/* count the meta-dnode */
822 	for (i = 0; i < dnp->dn_nblkptr; i++)
823 		bp->blk_fill += dnp->dn_blkptr[i].blk_fill;
824 
825 	BP_SET_TYPE(bp, DMU_OT_OBJSET);
826 	BP_SET_LEVEL(bp, 0);
827 
828 	/* We must do this after we've set the bp's type and level */
829 	if (!DVA_EQUAL(BP_IDENTITY(bp),
830 	    BP_IDENTITY(&zio->io_bp_orig))) {
831 		if (zio->io_bp_orig.blk_birth == os->os_synctx->tx_txg)
832 			dsl_dataset_block_kill(os->os_dsl_dataset,
833 			    &zio->io_bp_orig, NULL, os->os_synctx);
834 		dsl_dataset_block_born(os->os_dsl_dataset, bp, os->os_synctx);
835 	}
836 }
837 
838 /* ARGSUSED */
839 static void
840 killer(zio_t *zio, arc_buf_t *abuf, void *arg)
841 {
842 	objset_impl_t *os = arg;
843 
844 	ASSERT3U(zio->io_error, ==, 0);
845 	arc_release(os->os_phys_buf, &os->os_phys_buf);
846 }
847 
848 /* called from dsl */
849 void
850 dmu_objset_sync(objset_impl_t *os, zio_t *pio, dmu_tx_t *tx)
851 {
852 	int txgoff;
853 	zbookmark_t zb;
854 	zio_t *zio;
855 	list_t *list;
856 	dbuf_dirty_record_t *dr;
857 
858 	dprintf_ds(os->os_dsl_dataset, "txg=%llu\n", tx->tx_txg);
859 
860 	ASSERT(dmu_tx_is_syncing(tx));
861 	/* XXX the write_done callback should really give us the tx... */
862 	os->os_synctx = tx;
863 
864 	if (os->os_dsl_dataset == NULL) {
865 		/*
866 		 * This is the MOS.  If we have upgraded,
867 		 * spa_max_replication() could change, so reset
868 		 * os_copies here.
869 		 */
870 		os->os_copies = spa_max_replication(os->os_spa);
871 	}
872 
873 	/*
874 	 * Create the root block IO
875 	 */
876 	zb.zb_objset = os->os_dsl_dataset ? os->os_dsl_dataset->ds_object : 0;
877 	zb.zb_object = 0;
878 	zb.zb_level = -1;
879 	zb.zb_blkid = 0;
880 	if (BP_IS_OLDER(os->os_rootbp, tx->tx_txg)) {
881 		dsl_dataset_block_kill(os->os_dsl_dataset,
882 		    os->os_rootbp, pio, tx);
883 	}
884 	zio = arc_write(pio, os->os_spa, os->os_md_checksum,
885 	    os->os_md_compress,
886 	    dmu_get_replication_level(os, &zb, DMU_OT_OBJSET),
887 	    tx->tx_txg, os->os_rootbp, os->os_phys_buf, ready, killer, os,
888 	    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_METADATA,
889 	    &zb);
890 
891 	/*
892 	 * Sync meta-dnode - the parent IO for the sync is the root block
893 	 */
894 	os->os_meta_dnode->dn_zio = zio;
895 	dnode_sync(os->os_meta_dnode, tx);
896 
897 	txgoff = tx->tx_txg & TXG_MASK;
898 
899 	dmu_objset_sync_dnodes(&os->os_free_dnodes[txgoff], tx);
900 	dmu_objset_sync_dnodes(&os->os_dirty_dnodes[txgoff], tx);
901 
902 	list = &os->os_meta_dnode->dn_dirty_records[txgoff];
903 	while (dr = list_head(list)) {
904 		ASSERT(dr->dr_dbuf->db_level == 0);
905 		list_remove(list, dr);
906 		if (dr->dr_zio)
907 			zio_nowait(dr->dr_zio);
908 	}
909 	/*
910 	 * Free intent log blocks up to this tx.
911 	 */
912 	zil_sync(os->os_zil, tx);
913 	zio_nowait(zio);
914 }
915 
916 void
917 dmu_objset_space(objset_t *os, uint64_t *refdbytesp, uint64_t *availbytesp,
918     uint64_t *usedobjsp, uint64_t *availobjsp)
919 {
920 	dsl_dataset_space(os->os->os_dsl_dataset, refdbytesp, availbytesp,
921 	    usedobjsp, availobjsp);
922 }
923 
924 uint64_t
925 dmu_objset_fsid_guid(objset_t *os)
926 {
927 	return (dsl_dataset_fsid_guid(os->os->os_dsl_dataset));
928 }
929 
930 void
931 dmu_objset_fast_stat(objset_t *os, dmu_objset_stats_t *stat)
932 {
933 	stat->dds_type = os->os->os_phys->os_type;
934 	if (os->os->os_dsl_dataset)
935 		dsl_dataset_fast_stat(os->os->os_dsl_dataset, stat);
936 }
937 
938 void
939 dmu_objset_stats(objset_t *os, nvlist_t *nv)
940 {
941 	ASSERT(os->os->os_dsl_dataset ||
942 	    os->os->os_phys->os_type == DMU_OST_META);
943 
944 	if (os->os->os_dsl_dataset != NULL)
945 		dsl_dataset_stats(os->os->os_dsl_dataset, nv);
946 
947 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_TYPE,
948 	    os->os->os_phys->os_type);
949 }
950 
951 int
952 dmu_objset_is_snapshot(objset_t *os)
953 {
954 	if (os->os->os_dsl_dataset != NULL)
955 		return (dsl_dataset_is_snapshot(os->os->os_dsl_dataset));
956 	else
957 		return (B_FALSE);
958 }
959 
960 int
961 dmu_snapshot_realname(objset_t *os, char *name, char *real, int maxlen,
962     boolean_t *conflict)
963 {
964 	dsl_dataset_t *ds = os->os->os_dsl_dataset;
965 	uint64_t ignored;
966 
967 	if (ds->ds_phys->ds_snapnames_zapobj == 0)
968 		return (ENOENT);
969 
970 	return (zap_lookup_norm(ds->ds_dir->dd_pool->dp_meta_objset,
971 	    ds->ds_phys->ds_snapnames_zapobj, name, 8, 1, &ignored, MT_FIRST,
972 	    real, maxlen, conflict));
973 }
974 
975 int
976 dmu_snapshot_list_next(objset_t *os, int namelen, char *name,
977     uint64_t *idp, uint64_t *offp, boolean_t *case_conflict)
978 {
979 	dsl_dataset_t *ds = os->os->os_dsl_dataset;
980 	zap_cursor_t cursor;
981 	zap_attribute_t attr;
982 
983 	if (ds->ds_phys->ds_snapnames_zapobj == 0)
984 		return (ENOENT);
985 
986 	zap_cursor_init_serialized(&cursor,
987 	    ds->ds_dir->dd_pool->dp_meta_objset,
988 	    ds->ds_phys->ds_snapnames_zapobj, *offp);
989 
990 	if (zap_cursor_retrieve(&cursor, &attr) != 0) {
991 		zap_cursor_fini(&cursor);
992 		return (ENOENT);
993 	}
994 
995 	if (strlen(attr.za_name) + 1 > namelen) {
996 		zap_cursor_fini(&cursor);
997 		return (ENAMETOOLONG);
998 	}
999 
1000 	(void) strcpy(name, attr.za_name);
1001 	if (idp)
1002 		*idp = attr.za_first_integer;
1003 	if (case_conflict)
1004 		*case_conflict = attr.za_normalization_conflict;
1005 	zap_cursor_advance(&cursor);
1006 	*offp = zap_cursor_serialize(&cursor);
1007 	zap_cursor_fini(&cursor);
1008 
1009 	return (0);
1010 }
1011 
1012 int
1013 dmu_dir_list_next(objset_t *os, int namelen, char *name,
1014     uint64_t *idp, uint64_t *offp)
1015 {
1016 	dsl_dir_t *dd = os->os->os_dsl_dataset->ds_dir;
1017 	zap_cursor_t cursor;
1018 	zap_attribute_t attr;
1019 
1020 	/* there is no next dir on a snapshot! */
1021 	if (os->os->os_dsl_dataset->ds_object !=
1022 	    dd->dd_phys->dd_head_dataset_obj)
1023 		return (ENOENT);
1024 
1025 	zap_cursor_init_serialized(&cursor,
1026 	    dd->dd_pool->dp_meta_objset,
1027 	    dd->dd_phys->dd_child_dir_zapobj, *offp);
1028 
1029 	if (zap_cursor_retrieve(&cursor, &attr) != 0) {
1030 		zap_cursor_fini(&cursor);
1031 		return (ENOENT);
1032 	}
1033 
1034 	if (strlen(attr.za_name) + 1 > namelen) {
1035 		zap_cursor_fini(&cursor);
1036 		return (ENAMETOOLONG);
1037 	}
1038 
1039 	(void) strcpy(name, attr.za_name);
1040 	if (idp)
1041 		*idp = attr.za_first_integer;
1042 	zap_cursor_advance(&cursor);
1043 	*offp = zap_cursor_serialize(&cursor);
1044 	zap_cursor_fini(&cursor);
1045 
1046 	return (0);
1047 }
1048 
1049 /*
1050  * Find all objsets under name, and for each, call 'func(child_name, arg)'.
1051  */
1052 int
1053 dmu_objset_find(char *name, int func(char *, void *), void *arg, int flags)
1054 {
1055 	dsl_dir_t *dd;
1056 	objset_t *os;
1057 	uint64_t snapobj;
1058 	zap_cursor_t zc;
1059 	zap_attribute_t *attr;
1060 	char *child;
1061 	int do_self, err;
1062 
1063 	err = dsl_dir_open(name, FTAG, &dd, NULL);
1064 	if (err)
1065 		return (err);
1066 
1067 	/* NB: the $MOS dir doesn't have a head dataset */
1068 	do_self = (dd->dd_phys->dd_head_dataset_obj != 0);
1069 	attr = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
1070 
1071 	/*
1072 	 * Iterate over all children.
1073 	 */
1074 	if (flags & DS_FIND_CHILDREN) {
1075 		for (zap_cursor_init(&zc, dd->dd_pool->dp_meta_objset,
1076 		    dd->dd_phys->dd_child_dir_zapobj);
1077 		    zap_cursor_retrieve(&zc, attr) == 0;
1078 		    (void) zap_cursor_advance(&zc)) {
1079 			ASSERT(attr->za_integer_length == sizeof (uint64_t));
1080 			ASSERT(attr->za_num_integers == 1);
1081 
1082 			/*
1083 			 * No separating '/' because parent's name ends in /.
1084 			 */
1085 			child = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1086 			/* XXX could probably just use name here */
1087 			dsl_dir_name(dd, child);
1088 			(void) strcat(child, "/");
1089 			(void) strcat(child, attr->za_name);
1090 			err = dmu_objset_find(child, func, arg, flags);
1091 			kmem_free(child, MAXPATHLEN);
1092 			if (err)
1093 				break;
1094 		}
1095 		zap_cursor_fini(&zc);
1096 
1097 		if (err) {
1098 			dsl_dir_close(dd, FTAG);
1099 			kmem_free(attr, sizeof (zap_attribute_t));
1100 			return (err);
1101 		}
1102 	}
1103 
1104 	/*
1105 	 * Iterate over all snapshots.
1106 	 */
1107 	if ((flags & DS_FIND_SNAPSHOTS) &&
1108 	    dmu_objset_open(name, DMU_OST_ANY,
1109 	    DS_MODE_USER | DS_MODE_READONLY, &os) == 0) {
1110 
1111 		snapobj = os->os->os_dsl_dataset->ds_phys->ds_snapnames_zapobj;
1112 		dmu_objset_close(os);
1113 
1114 		for (zap_cursor_init(&zc, dd->dd_pool->dp_meta_objset, snapobj);
1115 		    zap_cursor_retrieve(&zc, attr) == 0;
1116 		    (void) zap_cursor_advance(&zc)) {
1117 			ASSERT(attr->za_integer_length == sizeof (uint64_t));
1118 			ASSERT(attr->za_num_integers == 1);
1119 
1120 			child = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1121 			/* XXX could probably just use name here */
1122 			dsl_dir_name(dd, child);
1123 			(void) strcat(child, "@");
1124 			(void) strcat(child, attr->za_name);
1125 			err = func(child, arg);
1126 			kmem_free(child, MAXPATHLEN);
1127 			if (err)
1128 				break;
1129 		}
1130 		zap_cursor_fini(&zc);
1131 	}
1132 
1133 	dsl_dir_close(dd, FTAG);
1134 	kmem_free(attr, sizeof (zap_attribute_t));
1135 
1136 	if (err)
1137 		return (err);
1138 
1139 	/*
1140 	 * Apply to self if appropriate.
1141 	 */
1142 	if (do_self)
1143 		err = func(name, arg);
1144 	return (err);
1145 }
1146 
1147 void
1148 dmu_objset_set_user(objset_t *os, void *user_ptr)
1149 {
1150 	ASSERT(MUTEX_HELD(&os->os->os_user_ptr_lock));
1151 	os->os->os_user_ptr = user_ptr;
1152 }
1153 
1154 void *
1155 dmu_objset_get_user(objset_t *os)
1156 {
1157 	ASSERT(MUTEX_HELD(&os->os->os_user_ptr_lock));
1158 	return (os->os->os_user_ptr);
1159 }
1160