xref: /titanic_44/usr/src/uts/common/fs/zfs/dmu_objset.c (revision ee5416c9d7e449233197d5d20bc6b81e4ff091b2)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/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 
48 spa_t *
49 dmu_objset_spa(objset_t *os)
50 {
51 	return (os->os->os_spa);
52 }
53 
54 zilog_t *
55 dmu_objset_zil(objset_t *os)
56 {
57 	return (os->os->os_zil);
58 }
59 
60 dsl_pool_t *
61 dmu_objset_pool(objset_t *os)
62 {
63 	dsl_dataset_t *ds;
64 
65 	if ((ds = os->os->os_dsl_dataset) != NULL && ds->ds_dir)
66 		return (ds->ds_dir->dd_pool);
67 	else
68 		return (spa_get_dsl(os->os->os_spa));
69 }
70 
71 dsl_dataset_t *
72 dmu_objset_ds(objset_t *os)
73 {
74 	return (os->os->os_dsl_dataset);
75 }
76 
77 dmu_objset_type_t
78 dmu_objset_type(objset_t *os)
79 {
80 	return (os->os->os_phys->os_type);
81 }
82 
83 void
84 dmu_objset_name(objset_t *os, char *buf)
85 {
86 	dsl_dataset_name(os->os->os_dsl_dataset, buf);
87 }
88 
89 uint64_t
90 dmu_objset_id(objset_t *os)
91 {
92 	dsl_dataset_t *ds = os->os->os_dsl_dataset;
93 
94 	return (ds ? ds->ds_object : 0);
95 }
96 
97 static void
98 checksum_changed_cb(void *arg, uint64_t newval)
99 {
100 	objset_impl_t *osi = arg;
101 
102 	/*
103 	 * Inheritance should have been done by now.
104 	 */
105 	ASSERT(newval != ZIO_CHECKSUM_INHERIT);
106 
107 	osi->os_checksum = zio_checksum_select(newval, ZIO_CHECKSUM_ON_VALUE);
108 }
109 
110 static void
111 compression_changed_cb(void *arg, uint64_t newval)
112 {
113 	objset_impl_t *osi = arg;
114 
115 	/*
116 	 * Inheritance and range checking should have been done by now.
117 	 */
118 	ASSERT(newval != ZIO_COMPRESS_INHERIT);
119 
120 	osi->os_compress = zio_compress_select(newval, ZIO_COMPRESS_ON_VALUE);
121 }
122 
123 static void
124 copies_changed_cb(void *arg, uint64_t newval)
125 {
126 	objset_impl_t *osi = arg;
127 
128 	/*
129 	 * Inheritance and range checking should have been done by now.
130 	 */
131 	ASSERT(newval > 0);
132 	ASSERT(newval <= spa_max_replication(osi->os_spa));
133 
134 	osi->os_copies = newval;
135 }
136 
137 void
138 dmu_objset_byteswap(void *buf, size_t size)
139 {
140 	objset_phys_t *osp = buf;
141 
142 	ASSERT(size == sizeof (objset_phys_t));
143 	dnode_byteswap(&osp->os_meta_dnode);
144 	byteswap_uint64_array(&osp->os_zil_header, sizeof (zil_header_t));
145 	osp->os_type = BSWAP_64(osp->os_type);
146 }
147 
148 int
149 dmu_objset_open_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
150     objset_impl_t **osip)
151 {
152 	objset_impl_t *osi;
153 	int i, err, checksum;
154 
155 	ASSERT(ds == NULL || MUTEX_HELD(&ds->ds_opening_lock));
156 
157 	osi = kmem_zalloc(sizeof (objset_impl_t), KM_SLEEP);
158 	osi->os.os = osi;
159 	osi->os_dsl_dataset = ds;
160 	osi->os_spa = spa;
161 	osi->os_rootbp = bp;
162 	if (!BP_IS_HOLE(osi->os_rootbp)) {
163 		uint32_t aflags = ARC_WAIT;
164 		zbookmark_t zb;
165 		zb.zb_objset = ds ? ds->ds_object : 0;
166 		zb.zb_object = 0;
167 		zb.zb_level = -1;
168 		zb.zb_blkid = 0;
169 
170 		dprintf_bp(osi->os_rootbp, "reading %s", "");
171 		err = arc_read(NULL, spa, osi->os_rootbp,
172 		    dmu_ot[DMU_OT_OBJSET].ot_byteswap,
173 		    arc_getbuf_func, &osi->os_phys_buf,
174 		    ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL, &aflags, &zb);
175 		if (err) {
176 			kmem_free(osi, sizeof (objset_impl_t));
177 			return (err);
178 		}
179 		osi->os_phys = osi->os_phys_buf->b_data;
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 
247 	osi->os_meta_dnode = dnode_special_open(osi,
248 	    &osi->os_phys->os_meta_dnode, DMU_META_DNODE_OBJECT);
249 
250 	/*
251 	 * We should be the only thread trying to do this because we
252 	 * have ds_opening_lock
253 	 */
254 	if (ds) {
255 		VERIFY(NULL == dsl_dataset_set_user_ptr(ds, osi,
256 		    dmu_objset_evict));
257 	}
258 
259 	*osip = osi;
260 	return (0);
261 }
262 
263 /* called from zpl */
264 int
265 dmu_objset_open(const char *name, dmu_objset_type_t type, int mode,
266     objset_t **osp)
267 {
268 	dsl_dataset_t *ds;
269 	int err;
270 	objset_t *os;
271 	objset_impl_t *osi;
272 
273 	os = kmem_alloc(sizeof (objset_t), KM_SLEEP);
274 	err = dsl_dataset_open(name, mode, os, &ds);
275 	if (err) {
276 		kmem_free(os, sizeof (objset_t));
277 		return (err);
278 	}
279 
280 	mutex_enter(&ds->ds_opening_lock);
281 	osi = dsl_dataset_get_user_ptr(ds);
282 	if (osi == NULL) {
283 		err = dmu_objset_open_impl(dsl_dataset_get_spa(ds),
284 		    ds, &ds->ds_phys->ds_bp, &osi);
285 		if (err) {
286 			dsl_dataset_close(ds, mode, os);
287 			kmem_free(os, sizeof (objset_t));
288 			return (err);
289 		}
290 	}
291 	mutex_exit(&ds->ds_opening_lock);
292 
293 	os->os = osi;
294 	os->os_mode = mode;
295 
296 	if (type != DMU_OST_ANY && type != os->os->os_phys->os_type) {
297 		dmu_objset_close(os);
298 		return (EINVAL);
299 	}
300 	*osp = os;
301 	return (0);
302 }
303 
304 void
305 dmu_objset_close(objset_t *os)
306 {
307 	dsl_dataset_close(os->os->os_dsl_dataset, os->os_mode, os);
308 	kmem_free(os, sizeof (objset_t));
309 }
310 
311 int
312 dmu_objset_evict_dbufs(objset_t *os)
313 {
314 	objset_impl_t *osi = os->os;
315 	dnode_t *dn;
316 
317 	mutex_enter(&osi->os_lock);
318 
319 	/* process the mdn last, since the other dnodes have holds on it */
320 	list_remove(&osi->os_dnodes, osi->os_meta_dnode);
321 	list_insert_tail(&osi->os_dnodes, osi->os_meta_dnode);
322 
323 	/*
324 	 * Find the first dnode with holds.  We have to do this dance
325 	 * because dnode_add_ref() only works if you already have a
326 	 * hold.  If there are no holds then it has no dbufs so OK to
327 	 * skip.
328 	 */
329 	for (dn = list_head(&osi->os_dnodes);
330 	    dn && !dnode_add_ref(dn, FTAG);
331 	    dn = list_next(&osi->os_dnodes, dn))
332 		continue;
333 
334 	while (dn) {
335 		dnode_t *next_dn = dn;
336 
337 		do {
338 			next_dn = list_next(&osi->os_dnodes, next_dn);
339 		} while (next_dn && !dnode_add_ref(next_dn, FTAG));
340 
341 		mutex_exit(&osi->os_lock);
342 		dnode_evict_dbufs(dn);
343 		dnode_rele(dn, FTAG);
344 		mutex_enter(&osi->os_lock);
345 		dn = next_dn;
346 	}
347 	mutex_exit(&osi->os_lock);
348 	return (list_head(&osi->os_dnodes) != osi->os_meta_dnode);
349 }
350 
351 void
352 dmu_objset_evict(dsl_dataset_t *ds, void *arg)
353 {
354 	objset_impl_t *osi = arg;
355 	objset_t os;
356 	int i;
357 
358 	for (i = 0; i < TXG_SIZE; i++) {
359 		ASSERT(list_head(&osi->os_dirty_dnodes[i]) == NULL);
360 		ASSERT(list_head(&osi->os_free_dnodes[i]) == NULL);
361 	}
362 
363 	if (ds && ds->ds_phys->ds_num_children == 0) {
364 		VERIFY(0 == dsl_prop_unregister(ds, "checksum",
365 		    checksum_changed_cb, osi));
366 		VERIFY(0 == dsl_prop_unregister(ds, "compression",
367 		    compression_changed_cb, osi));
368 		VERIFY(0 == dsl_prop_unregister(ds, "copies",
369 		    copies_changed_cb, osi));
370 	}
371 
372 	/*
373 	 * We should need only a single pass over the dnode list, since
374 	 * nothing can be added to the list at this point.
375 	 */
376 	os.os = osi;
377 	(void) dmu_objset_evict_dbufs(&os);
378 
379 	ASSERT3P(list_head(&osi->os_dnodes), ==, osi->os_meta_dnode);
380 	ASSERT3P(list_tail(&osi->os_dnodes), ==, osi->os_meta_dnode);
381 	ASSERT3P(list_head(&osi->os_meta_dnode->dn_dbufs), ==, NULL);
382 
383 	dnode_special_close(osi->os_meta_dnode);
384 	zil_free(osi->os_zil);
385 
386 	VERIFY(arc_buf_remove_ref(osi->os_phys_buf, &osi->os_phys_buf) == 1);
387 	mutex_destroy(&osi->os_lock);
388 	mutex_destroy(&osi->os_obj_lock);
389 	kmem_free(osi, sizeof (objset_impl_t));
390 }
391 
392 /* called from dsl for meta-objset */
393 objset_impl_t *
394 dmu_objset_create_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
395     dmu_objset_type_t type, dmu_tx_t *tx)
396 {
397 	objset_impl_t *osi;
398 	dnode_t *mdn;
399 
400 	ASSERT(dmu_tx_is_syncing(tx));
401 	if (ds)
402 		mutex_enter(&ds->ds_opening_lock);
403 	VERIFY(0 == dmu_objset_open_impl(spa, ds, bp, &osi));
404 	if (ds)
405 		mutex_exit(&ds->ds_opening_lock);
406 	mdn = osi->os_meta_dnode;
407 
408 	dnode_allocate(mdn, DMU_OT_DNODE, 1 << DNODE_BLOCK_SHIFT,
409 	    DN_MAX_INDBLKSHIFT, DMU_OT_NONE, 0, tx);
410 
411 	/*
412 	 * We don't want to have to increase the meta-dnode's nlevels
413 	 * later, because then we could do it in quescing context while
414 	 * we are also accessing it in open context.
415 	 *
416 	 * This precaution is not necessary for the MOS (ds == NULL),
417 	 * because the MOS is only updated in syncing context.
418 	 * This is most fortunate: the MOS is the only objset that
419 	 * needs to be synced multiple times as spa_sync() iterates
420 	 * to convergence, so minimizing its dn_nlevels matters.
421 	 */
422 	if (ds != NULL) {
423 		int levels = 1;
424 
425 		/*
426 		 * Determine the number of levels necessary for the meta-dnode
427 		 * to contain DN_MAX_OBJECT dnodes.
428 		 */
429 		while ((uint64_t)mdn->dn_nblkptr << (mdn->dn_datablkshift +
430 		    (levels - 1) * (mdn->dn_indblkshift - SPA_BLKPTRSHIFT)) <
431 		    DN_MAX_OBJECT * sizeof (dnode_phys_t))
432 			levels++;
433 
434 		mdn->dn_next_nlevels[tx->tx_txg & TXG_MASK] =
435 		    mdn->dn_nlevels = levels;
436 	}
437 
438 	ASSERT(type != DMU_OST_NONE);
439 	ASSERT(type != DMU_OST_ANY);
440 	ASSERT(type < DMU_OST_NUMTYPES);
441 	osi->os_phys->os_type = type;
442 
443 	dsl_dataset_dirty(ds, tx);
444 
445 	return (osi);
446 }
447 
448 struct oscarg {
449 	void (*userfunc)(objset_t *os, void *arg, cred_t *cr, dmu_tx_t *tx);
450 	void *userarg;
451 	dsl_dataset_t *clone_parent;
452 	const char *lastname;
453 	dmu_objset_type_t type;
454 };
455 
456 /*ARGSUSED*/
457 static int
458 dmu_objset_create_check(void *arg1, void *arg2, dmu_tx_t *tx)
459 {
460 	dsl_dir_t *dd = arg1;
461 	struct oscarg *oa = arg2;
462 	objset_t *mos = dd->dd_pool->dp_meta_objset;
463 	int err;
464 	uint64_t ddobj;
465 
466 	err = zap_lookup(mos, dd->dd_phys->dd_child_dir_zapobj,
467 	    oa->lastname, sizeof (uint64_t), 1, &ddobj);
468 	if (err != ENOENT)
469 		return (err ? err : EEXIST);
470 
471 	if (oa->clone_parent != NULL) {
472 		/*
473 		 * You can't clone across pools.
474 		 */
475 		if (oa->clone_parent->ds_dir->dd_pool != dd->dd_pool)
476 			return (EXDEV);
477 
478 		/*
479 		 * You can only clone snapshots, not the head datasets.
480 		 */
481 		if (oa->clone_parent->ds_phys->ds_num_children == 0)
482 			return (EINVAL);
483 	}
484 
485 	return (0);
486 }
487 
488 static void
489 dmu_objset_create_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
490 {
491 	dsl_dir_t *dd = arg1;
492 	struct oscarg *oa = arg2;
493 	dsl_dataset_t *ds;
494 	blkptr_t *bp;
495 	uint64_t dsobj;
496 
497 	ASSERT(dmu_tx_is_syncing(tx));
498 
499 	dsobj = dsl_dataset_create_sync(dd, oa->lastname,
500 	    oa->clone_parent, tx);
501 
502 	VERIFY(0 == dsl_dataset_open_obj(dd->dd_pool, dsobj, NULL,
503 	    DS_MODE_STANDARD | DS_MODE_READONLY, FTAG, &ds));
504 	bp = dsl_dataset_get_blkptr(ds);
505 	if (BP_IS_HOLE(bp)) {
506 		objset_impl_t *osi;
507 
508 		/* This is an empty dmu_objset; not a clone. */
509 		osi = dmu_objset_create_impl(dsl_dataset_get_spa(ds),
510 		    ds, bp, oa->type, tx);
511 
512 		if (oa->userfunc)
513 			oa->userfunc(&osi->os, oa->userarg, cr, tx);
514 	}
515 
516 	/*
517 	 * Create create time permission if any?
518 	 */
519 	dsl_deleg_set_create_perms(ds->ds_dir, tx, cr);
520 
521 	spa_history_internal_log(LOG_DS_CREATE, dd->dd_pool->dp_spa,
522 	    tx, cr, "dataset = %llu", dsobj);
523 
524 	dsl_dataset_close(ds, DS_MODE_STANDARD | DS_MODE_READONLY, FTAG);
525 }
526 
527 int
528 dmu_objset_create(const char *name, dmu_objset_type_t type,
529     objset_t *clone_parent,
530     void (*func)(objset_t *os, void *arg, cred_t *cr, dmu_tx_t *tx), void *arg)
531 {
532 	dsl_dir_t *pdd;
533 	const char *tail;
534 	int err = 0;
535 	struct oscarg oa = { 0 };
536 
537 	ASSERT(strchr(name, '@') == NULL);
538 	err = dsl_dir_open(name, FTAG, &pdd, &tail);
539 	if (err)
540 		return (err);
541 	if (tail == NULL) {
542 		dsl_dir_close(pdd, FTAG);
543 		return (EEXIST);
544 	}
545 
546 	dprintf("name=%s\n", name);
547 
548 	oa.userfunc = func;
549 	oa.userarg = arg;
550 	oa.lastname = tail;
551 	oa.type = type;
552 
553 	if (clone_parent != NULL) {
554 		/*
555 		 * You can't clone to a different type.
556 		 */
557 		if (clone_parent->os->os_phys->os_type != type) {
558 			dsl_dir_close(pdd, FTAG);
559 			return (EINVAL);
560 		}
561 		oa.clone_parent = clone_parent->os->os_dsl_dataset;
562 	}
563 	err = dsl_sync_task_do(pdd->dd_pool, dmu_objset_create_check,
564 	    dmu_objset_create_sync, pdd, &oa, 5);
565 	dsl_dir_close(pdd, FTAG);
566 	return (err);
567 }
568 
569 int
570 dmu_objset_destroy(const char *name)
571 {
572 	objset_t *os;
573 	int error;
574 
575 	/*
576 	 * If it looks like we'll be able to destroy it, and there's
577 	 * an unplayed replay log sitting around, destroy the log.
578 	 * It would be nicer to do this in dsl_dataset_destroy_sync(),
579 	 * but the replay log objset is modified in open context.
580 	 */
581 	error = dmu_objset_open(name, DMU_OST_ANY, DS_MODE_EXCLUSIVE, &os);
582 	if (error == 0) {
583 		zil_destroy(dmu_objset_zil(os), B_FALSE);
584 		dmu_objset_close(os);
585 	}
586 
587 	return (dsl_dataset_destroy(name));
588 }
589 
590 int
591 dmu_objset_rollback(const char *name)
592 {
593 	int err;
594 	objset_t *os;
595 
596 	err = dmu_objset_open(name, DMU_OST_ANY,
597 	    DS_MODE_EXCLUSIVE | DS_MODE_INCONSISTENT, &os);
598 	if (err)
599 		return (err);
600 
601 	/* XXX uncache everything? */
602 	err = dsl_dataset_rollback(os->os->os_dsl_dataset);
603 
604 	dmu_objset_close(os);
605 	return (err);
606 }
607 
608 struct snaparg {
609 	dsl_sync_task_group_t *dstg;
610 	char *snapname;
611 	char failed[MAXPATHLEN];
612 	boolean_t checkperms;
613 };
614 
615 static int
616 dmu_objset_snapshot_one(char *name, void *arg)
617 {
618 	struct snaparg *sn = arg;
619 	objset_t *os;
620 	dmu_objset_stats_t stat;
621 	int err;
622 
623 	(void) strcpy(sn->failed, name);
624 
625 	/*
626 	 * Check permissions only when requested.  This only applies when
627 	 * doing a recursive snapshot.  The permission checks for the starting
628 	 * dataset have already been performed in zfs_secpolicy_snapshot()
629 	 */
630 	if (sn->checkperms == B_TRUE &&
631 	    (err = zfs_secpolicy_snapshot_perms(name, CRED())))
632 		return (err);
633 
634 	err = dmu_objset_open(name, DMU_OST_ANY, DS_MODE_STANDARD, &os);
635 	if (err != 0)
636 		return (err);
637 
638 	/*
639 	 * If the objset is in an inconsistent state, return busy.
640 	 */
641 	dmu_objset_fast_stat(os, &stat);
642 	if (stat.dds_inconsistent) {
643 		dmu_objset_close(os);
644 		return (EBUSY);
645 	}
646 
647 	/*
648 	 * NB: we need to wait for all in-flight changes to get to disk,
649 	 * so that we snapshot those changes.  zil_suspend does this as
650 	 * a side effect.
651 	 */
652 	err = zil_suspend(dmu_objset_zil(os));
653 	if (err == 0) {
654 		dsl_sync_task_create(sn->dstg, dsl_dataset_snapshot_check,
655 		    dsl_dataset_snapshot_sync, os, sn->snapname, 3);
656 	} else {
657 		dmu_objset_close(os);
658 	}
659 
660 	return (err);
661 }
662 
663 int
664 dmu_objset_snapshot(char *fsname, char *snapname, boolean_t recursive)
665 {
666 	dsl_sync_task_t *dst;
667 	struct snaparg sn = { 0 };
668 	spa_t *spa;
669 	int err;
670 
671 	(void) strcpy(sn.failed, fsname);
672 
673 	err = spa_open(fsname, &spa, FTAG);
674 	if (err)
675 		return (err);
676 
677 	sn.dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
678 	sn.snapname = snapname;
679 
680 	if (recursive) {
681 		sn.checkperms = B_TRUE;
682 		err = dmu_objset_find(fsname,
683 		    dmu_objset_snapshot_one, &sn, DS_FIND_CHILDREN);
684 	} else {
685 		sn.checkperms = B_FALSE;
686 		err = dmu_objset_snapshot_one(fsname, &sn);
687 	}
688 
689 	if (err)
690 		goto out;
691 
692 	err = dsl_sync_task_group_wait(sn.dstg);
693 
694 	for (dst = list_head(&sn.dstg->dstg_tasks); dst;
695 	    dst = list_next(&sn.dstg->dstg_tasks, dst)) {
696 		objset_t *os = dst->dst_arg1;
697 		if (dst->dst_err)
698 			dmu_objset_name(os, sn.failed);
699 		zil_resume(dmu_objset_zil(os));
700 		dmu_objset_close(os);
701 	}
702 out:
703 	if (err)
704 		(void) strcpy(fsname, sn.failed);
705 	dsl_sync_task_group_destroy(sn.dstg);
706 	spa_close(spa, FTAG);
707 	return (err);
708 }
709 
710 static void
711 dmu_objset_sync_dnodes(list_t *list, dmu_tx_t *tx)
712 {
713 	dnode_t *dn;
714 
715 	while (dn = list_head(list)) {
716 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
717 		ASSERT(dn->dn_dbuf->db_data_pending);
718 		/*
719 		 * Initialize dn_zio outside dnode_sync()
720 		 * to accomodate meta-dnode
721 		 */
722 		dn->dn_zio = dn->dn_dbuf->db_data_pending->dr_zio;
723 		ASSERT(dn->dn_zio);
724 
725 		ASSERT3U(dn->dn_nlevels, <=, DN_MAX_LEVELS);
726 		list_remove(list, dn);
727 		dnode_sync(dn, tx);
728 	}
729 }
730 
731 /* ARGSUSED */
732 static void
733 ready(zio_t *zio, arc_buf_t *abuf, void *arg)
734 {
735 	objset_impl_t *os = arg;
736 	blkptr_t *bp = os->os_rootbp;
737 	dnode_phys_t *dnp = &os->os_phys->os_meta_dnode;
738 	int i;
739 
740 	/*
741 	 * Update rootbp fill count.
742 	 */
743 	bp->blk_fill = 1;	/* count the meta-dnode */
744 	for (i = 0; i < dnp->dn_nblkptr; i++)
745 		bp->blk_fill += dnp->dn_blkptr[i].blk_fill;
746 }
747 
748 /* ARGSUSED */
749 static void
750 killer(zio_t *zio, arc_buf_t *abuf, void *arg)
751 {
752 	objset_impl_t *os = arg;
753 
754 	ASSERT3U(zio->io_error, ==, 0);
755 
756 	BP_SET_TYPE(zio->io_bp, DMU_OT_OBJSET);
757 	BP_SET_LEVEL(zio->io_bp, 0);
758 
759 	if (!DVA_EQUAL(BP_IDENTITY(zio->io_bp),
760 	    BP_IDENTITY(&zio->io_bp_orig))) {
761 		if (zio->io_bp_orig.blk_birth == os->os_synctx->tx_txg)
762 			dsl_dataset_block_kill(os->os_dsl_dataset,
763 			    &zio->io_bp_orig, NULL, os->os_synctx);
764 		dsl_dataset_block_born(os->os_dsl_dataset, zio->io_bp,
765 		    os->os_synctx);
766 	}
767 	arc_release(os->os_phys_buf, &os->os_phys_buf);
768 }
769 
770 /* called from dsl */
771 void
772 dmu_objset_sync(objset_impl_t *os, zio_t *pio, dmu_tx_t *tx)
773 {
774 	int txgoff;
775 	zbookmark_t zb;
776 	zio_t *zio;
777 	list_t *list;
778 	dbuf_dirty_record_t *dr;
779 
780 	dprintf_ds(os->os_dsl_dataset, "txg=%llu\n", tx->tx_txg);
781 
782 	ASSERT(dmu_tx_is_syncing(tx));
783 	/* XXX the write_done callback should really give us the tx... */
784 	os->os_synctx = tx;
785 
786 	if (os->os_dsl_dataset == NULL) {
787 		/*
788 		 * This is the MOS.  If we have upgraded,
789 		 * spa_max_replication() could change, so reset
790 		 * os_copies here.
791 		 */
792 		os->os_copies = spa_max_replication(os->os_spa);
793 	}
794 
795 	/*
796 	 * Create the root block IO
797 	 */
798 	zb.zb_objset = os->os_dsl_dataset ? os->os_dsl_dataset->ds_object : 0;
799 	zb.zb_object = 0;
800 	zb.zb_level = -1;
801 	zb.zb_blkid = 0;
802 	if (BP_IS_OLDER(os->os_rootbp, tx->tx_txg)) {
803 		dsl_dataset_block_kill(os->os_dsl_dataset,
804 		    os->os_rootbp, pio, tx);
805 	}
806 	zio = arc_write(pio, os->os_spa, os->os_md_checksum,
807 	    os->os_md_compress,
808 	    dmu_get_replication_level(os, &zb, DMU_OT_OBJSET),
809 	    tx->tx_txg, os->os_rootbp, os->os_phys_buf, ready, killer, os,
810 	    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_METADATA,
811 	    &zb);
812 
813 	/*
814 	 * Sync meta-dnode - the parent IO for the sync is the root block
815 	 */
816 	os->os_meta_dnode->dn_zio = zio;
817 	dnode_sync(os->os_meta_dnode, tx);
818 
819 	txgoff = tx->tx_txg & TXG_MASK;
820 
821 	dmu_objset_sync_dnodes(&os->os_free_dnodes[txgoff], tx);
822 	dmu_objset_sync_dnodes(&os->os_dirty_dnodes[txgoff], tx);
823 
824 	list = &os->os_meta_dnode->dn_dirty_records[txgoff];
825 	while (dr = list_head(list)) {
826 		ASSERT(dr->dr_dbuf->db_level == 0);
827 		list_remove(list, dr);
828 		if (dr->dr_zio)
829 			zio_nowait(dr->dr_zio);
830 	}
831 	/*
832 	 * Free intent log blocks up to this tx.
833 	 */
834 	zil_sync(os->os_zil, tx);
835 	zio_nowait(zio);
836 }
837 
838 void
839 dmu_objset_space(objset_t *os, uint64_t *refdbytesp, uint64_t *availbytesp,
840     uint64_t *usedobjsp, uint64_t *availobjsp)
841 {
842 	dsl_dataset_space(os->os->os_dsl_dataset, refdbytesp, availbytesp,
843 	    usedobjsp, availobjsp);
844 }
845 
846 uint64_t
847 dmu_objset_fsid_guid(objset_t *os)
848 {
849 	return (dsl_dataset_fsid_guid(os->os->os_dsl_dataset));
850 }
851 
852 void
853 dmu_objset_fast_stat(objset_t *os, dmu_objset_stats_t *stat)
854 {
855 	stat->dds_type = os->os->os_phys->os_type;
856 	if (os->os->os_dsl_dataset)
857 		dsl_dataset_fast_stat(os->os->os_dsl_dataset, stat);
858 }
859 
860 void
861 dmu_objset_stats(objset_t *os, nvlist_t *nv)
862 {
863 	ASSERT(os->os->os_dsl_dataset ||
864 	    os->os->os_phys->os_type == DMU_OST_META);
865 
866 	if (os->os->os_dsl_dataset != NULL)
867 		dsl_dataset_stats(os->os->os_dsl_dataset, nv);
868 
869 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_TYPE,
870 	    os->os->os_phys->os_type);
871 }
872 
873 int
874 dmu_objset_is_snapshot(objset_t *os)
875 {
876 	if (os->os->os_dsl_dataset != NULL)
877 		return (dsl_dataset_is_snapshot(os->os->os_dsl_dataset));
878 	else
879 		return (B_FALSE);
880 }
881 
882 int
883 dmu_snapshot_list_next(objset_t *os, int namelen, char *name,
884     uint64_t *idp, uint64_t *offp)
885 {
886 	dsl_dataset_t *ds = os->os->os_dsl_dataset;
887 	zap_cursor_t cursor;
888 	zap_attribute_t attr;
889 
890 	if (ds->ds_phys->ds_snapnames_zapobj == 0)
891 		return (ENOENT);
892 
893 	zap_cursor_init_serialized(&cursor,
894 	    ds->ds_dir->dd_pool->dp_meta_objset,
895 	    ds->ds_phys->ds_snapnames_zapobj, *offp);
896 
897 	if (zap_cursor_retrieve(&cursor, &attr) != 0) {
898 		zap_cursor_fini(&cursor);
899 		return (ENOENT);
900 	}
901 
902 	if (strlen(attr.za_name) + 1 > namelen) {
903 		zap_cursor_fini(&cursor);
904 		return (ENAMETOOLONG);
905 	}
906 
907 	(void) strcpy(name, attr.za_name);
908 	if (idp)
909 		*idp = attr.za_first_integer;
910 	zap_cursor_advance(&cursor);
911 	*offp = zap_cursor_serialize(&cursor);
912 	zap_cursor_fini(&cursor);
913 
914 	return (0);
915 }
916 
917 int
918 dmu_dir_list_next(objset_t *os, int namelen, char *name,
919     uint64_t *idp, uint64_t *offp)
920 {
921 	dsl_dir_t *dd = os->os->os_dsl_dataset->ds_dir;
922 	zap_cursor_t cursor;
923 	zap_attribute_t attr;
924 
925 	/* there is no next dir on a snapshot! */
926 	if (os->os->os_dsl_dataset->ds_object !=
927 	    dd->dd_phys->dd_head_dataset_obj)
928 		return (ENOENT);
929 
930 	zap_cursor_init_serialized(&cursor,
931 	    dd->dd_pool->dp_meta_objset,
932 	    dd->dd_phys->dd_child_dir_zapobj, *offp);
933 
934 	if (zap_cursor_retrieve(&cursor, &attr) != 0) {
935 		zap_cursor_fini(&cursor);
936 		return (ENOENT);
937 	}
938 
939 	if (strlen(attr.za_name) + 1 > namelen) {
940 		zap_cursor_fini(&cursor);
941 		return (ENAMETOOLONG);
942 	}
943 
944 	(void) strcpy(name, attr.za_name);
945 	if (idp)
946 		*idp = attr.za_first_integer;
947 	zap_cursor_advance(&cursor);
948 	*offp = zap_cursor_serialize(&cursor);
949 	zap_cursor_fini(&cursor);
950 
951 	return (0);
952 }
953 
954 /*
955  * Find all objsets under name, and for each, call 'func(child_name, arg)'.
956  */
957 int
958 dmu_objset_find(char *name, int func(char *, void *), void *arg, int flags)
959 {
960 	dsl_dir_t *dd;
961 	objset_t *os;
962 	uint64_t snapobj;
963 	zap_cursor_t zc;
964 	zap_attribute_t *attr;
965 	char *child;
966 	int do_self, err;
967 
968 	err = dsl_dir_open(name, FTAG, &dd, NULL);
969 	if (err)
970 		return (err);
971 
972 	/* NB: the $MOS dir doesn't have a head dataset */
973 	do_self = (dd->dd_phys->dd_head_dataset_obj != 0);
974 	attr = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
975 
976 	/*
977 	 * Iterate over all children.
978 	 */
979 	if (flags & DS_FIND_CHILDREN) {
980 		for (zap_cursor_init(&zc, dd->dd_pool->dp_meta_objset,
981 		    dd->dd_phys->dd_child_dir_zapobj);
982 		    zap_cursor_retrieve(&zc, attr) == 0;
983 		    (void) zap_cursor_advance(&zc)) {
984 			ASSERT(attr->za_integer_length == sizeof (uint64_t));
985 			ASSERT(attr->za_num_integers == 1);
986 
987 			/*
988 			 * No separating '/' because parent's name ends in /.
989 			 */
990 			child = kmem_alloc(MAXPATHLEN, KM_SLEEP);
991 			/* XXX could probably just use name here */
992 			dsl_dir_name(dd, child);
993 			(void) strcat(child, "/");
994 			(void) strcat(child, attr->za_name);
995 			err = dmu_objset_find(child, func, arg, flags);
996 			kmem_free(child, MAXPATHLEN);
997 			if (err)
998 				break;
999 		}
1000 		zap_cursor_fini(&zc);
1001 
1002 		if (err) {
1003 			dsl_dir_close(dd, FTAG);
1004 			kmem_free(attr, sizeof (zap_attribute_t));
1005 			return (err);
1006 		}
1007 	}
1008 
1009 	/*
1010 	 * Iterate over all snapshots.
1011 	 */
1012 	if ((flags & DS_FIND_SNAPSHOTS) &&
1013 	    dmu_objset_open(name, DMU_OST_ANY,
1014 	    DS_MODE_STANDARD | DS_MODE_READONLY, &os) == 0) {
1015 
1016 		snapobj = os->os->os_dsl_dataset->ds_phys->ds_snapnames_zapobj;
1017 		dmu_objset_close(os);
1018 
1019 		for (zap_cursor_init(&zc, dd->dd_pool->dp_meta_objset, snapobj);
1020 		    zap_cursor_retrieve(&zc, attr) == 0;
1021 		    (void) zap_cursor_advance(&zc)) {
1022 			ASSERT(attr->za_integer_length == sizeof (uint64_t));
1023 			ASSERT(attr->za_num_integers == 1);
1024 
1025 			child = kmem_alloc(MAXPATHLEN, KM_SLEEP);
1026 			/* XXX could probably just use name here */
1027 			dsl_dir_name(dd, child);
1028 			(void) strcat(child, "@");
1029 			(void) strcat(child, attr->za_name);
1030 			err = func(child, arg);
1031 			kmem_free(child, MAXPATHLEN);
1032 			if (err)
1033 				break;
1034 		}
1035 		zap_cursor_fini(&zc);
1036 	}
1037 
1038 	dsl_dir_close(dd, FTAG);
1039 	kmem_free(attr, sizeof (zap_attribute_t));
1040 
1041 	if (err)
1042 		return (err);
1043 
1044 	/*
1045 	 * Apply to self if appropriate.
1046 	 */
1047 	if (do_self)
1048 		err = func(name, arg);
1049 	return (err);
1050 }
1051