xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_objset.c (revision e77c795bcbe51aebd7579fe13cbf2a6d56eca47f)
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 (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /* Portions Copyright 2010 Robert Milkowski */
26 
27 #include <sys/cred.h>
28 #include <sys/zfs_context.h>
29 #include <sys/dmu_objset.h>
30 #include <sys/dsl_dir.h>
31 #include <sys/dsl_dataset.h>
32 #include <sys/dsl_prop.h>
33 #include <sys/dsl_pool.h>
34 #include <sys/dsl_synctask.h>
35 #include <sys/dsl_deleg.h>
36 #include <sys/dnode.h>
37 #include <sys/dbuf.h>
38 #include <sys/zvol.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/zap.h>
41 #include <sys/zil.h>
42 #include <sys/dmu_impl.h>
43 #include <sys/zfs_ioctl.h>
44 #include <sys/sa.h>
45 #include <sys/zfs_onexit.h>
46 
47 /*
48  * Needed to close a window in dnode_move() that allows the objset to be freed
49  * before it can be safely accessed.
50  */
51 krwlock_t os_lock;
52 
53 void
54 dmu_objset_init(void)
55 {
56 	rw_init(&os_lock, NULL, RW_DEFAULT, NULL);
57 }
58 
59 void
60 dmu_objset_fini(void)
61 {
62 	rw_destroy(&os_lock);
63 }
64 
65 spa_t *
66 dmu_objset_spa(objset_t *os)
67 {
68 	return (os->os_spa);
69 }
70 
71 zilog_t *
72 dmu_objset_zil(objset_t *os)
73 {
74 	return (os->os_zil);
75 }
76 
77 dsl_pool_t *
78 dmu_objset_pool(objset_t *os)
79 {
80 	dsl_dataset_t *ds;
81 
82 	if ((ds = os->os_dsl_dataset) != NULL && ds->ds_dir)
83 		return (ds->ds_dir->dd_pool);
84 	else
85 		return (spa_get_dsl(os->os_spa));
86 }
87 
88 dsl_dataset_t *
89 dmu_objset_ds(objset_t *os)
90 {
91 	return (os->os_dsl_dataset);
92 }
93 
94 dmu_objset_type_t
95 dmu_objset_type(objset_t *os)
96 {
97 	return (os->os_phys->os_type);
98 }
99 
100 void
101 dmu_objset_name(objset_t *os, char *buf)
102 {
103 	dsl_dataset_name(os->os_dsl_dataset, buf);
104 }
105 
106 uint64_t
107 dmu_objset_id(objset_t *os)
108 {
109 	dsl_dataset_t *ds = os->os_dsl_dataset;
110 
111 	return (ds ? ds->ds_object : 0);
112 }
113 
114 uint64_t
115 dmu_objset_syncprop(objset_t *os)
116 {
117 	return (os->os_sync);
118 }
119 
120 uint64_t
121 dmu_objset_logbias(objset_t *os)
122 {
123 	return (os->os_logbias);
124 }
125 
126 static void
127 checksum_changed_cb(void *arg, uint64_t newval)
128 {
129 	objset_t *os = arg;
130 
131 	/*
132 	 * Inheritance should have been done by now.
133 	 */
134 	ASSERT(newval != ZIO_CHECKSUM_INHERIT);
135 
136 	os->os_checksum = zio_checksum_select(newval, ZIO_CHECKSUM_ON_VALUE);
137 }
138 
139 static void
140 compression_changed_cb(void *arg, uint64_t newval)
141 {
142 	objset_t *os = arg;
143 
144 	/*
145 	 * Inheritance and range checking should have been done by now.
146 	 */
147 	ASSERT(newval != ZIO_COMPRESS_INHERIT);
148 
149 	os->os_compress = zio_compress_select(newval, ZIO_COMPRESS_ON_VALUE);
150 }
151 
152 static void
153 copies_changed_cb(void *arg, uint64_t newval)
154 {
155 	objset_t *os = arg;
156 
157 	/*
158 	 * Inheritance and range checking should have been done by now.
159 	 */
160 	ASSERT(newval > 0);
161 	ASSERT(newval <= spa_max_replication(os->os_spa));
162 
163 	os->os_copies = newval;
164 }
165 
166 static void
167 dedup_changed_cb(void *arg, uint64_t newval)
168 {
169 	objset_t *os = arg;
170 	spa_t *spa = os->os_spa;
171 	enum zio_checksum checksum;
172 
173 	/*
174 	 * Inheritance should have been done by now.
175 	 */
176 	ASSERT(newval != ZIO_CHECKSUM_INHERIT);
177 
178 	checksum = zio_checksum_dedup_select(spa, newval, ZIO_CHECKSUM_OFF);
179 
180 	os->os_dedup_checksum = checksum & ZIO_CHECKSUM_MASK;
181 	os->os_dedup_verify = !!(checksum & ZIO_CHECKSUM_VERIFY);
182 }
183 
184 static void
185 primary_cache_changed_cb(void *arg, uint64_t newval)
186 {
187 	objset_t *os = arg;
188 
189 	/*
190 	 * Inheritance and range checking should have been done by now.
191 	 */
192 	ASSERT(newval == ZFS_CACHE_ALL || newval == ZFS_CACHE_NONE ||
193 	    newval == ZFS_CACHE_METADATA);
194 
195 	os->os_primary_cache = newval;
196 }
197 
198 static void
199 secondary_cache_changed_cb(void *arg, uint64_t newval)
200 {
201 	objset_t *os = arg;
202 
203 	/*
204 	 * Inheritance and range checking should have been done by now.
205 	 */
206 	ASSERT(newval == ZFS_CACHE_ALL || newval == ZFS_CACHE_NONE ||
207 	    newval == ZFS_CACHE_METADATA);
208 
209 	os->os_secondary_cache = newval;
210 }
211 
212 static void
213 sync_changed_cb(void *arg, uint64_t newval)
214 {
215 	objset_t *os = arg;
216 
217 	/*
218 	 * Inheritance and range checking should have been done by now.
219 	 */
220 	ASSERT(newval == ZFS_SYNC_STANDARD || newval == ZFS_SYNC_ALWAYS ||
221 	    newval == ZFS_SYNC_DISABLED);
222 
223 	os->os_sync = newval;
224 	if (os->os_zil)
225 		zil_set_sync(os->os_zil, newval);
226 }
227 
228 static void
229 logbias_changed_cb(void *arg, uint64_t newval)
230 {
231 	objset_t *os = arg;
232 
233 	ASSERT(newval == ZFS_LOGBIAS_LATENCY ||
234 	    newval == ZFS_LOGBIAS_THROUGHPUT);
235 	os->os_logbias = newval;
236 	if (os->os_zil)
237 		zil_set_logbias(os->os_zil, newval);
238 }
239 
240 void
241 dmu_objset_byteswap(void *buf, size_t size)
242 {
243 	objset_phys_t *osp = buf;
244 
245 	ASSERT(size == OBJSET_OLD_PHYS_SIZE || size == sizeof (objset_phys_t));
246 	dnode_byteswap(&osp->os_meta_dnode);
247 	byteswap_uint64_array(&osp->os_zil_header, sizeof (zil_header_t));
248 	osp->os_type = BSWAP_64(osp->os_type);
249 	osp->os_flags = BSWAP_64(osp->os_flags);
250 	if (size == sizeof (objset_phys_t)) {
251 		dnode_byteswap(&osp->os_userused_dnode);
252 		dnode_byteswap(&osp->os_groupused_dnode);
253 	}
254 }
255 
256 int
257 dmu_objset_open_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
258     objset_t **osp)
259 {
260 	objset_t *os;
261 	int i, err;
262 
263 	ASSERT(ds == NULL || MUTEX_HELD(&ds->ds_opening_lock));
264 
265 	os = kmem_zalloc(sizeof (objset_t), KM_SLEEP);
266 	os->os_dsl_dataset = ds;
267 	os->os_spa = spa;
268 	os->os_rootbp = bp;
269 	if (!BP_IS_HOLE(os->os_rootbp)) {
270 		uint32_t aflags = ARC_WAIT;
271 		zbookmark_t zb;
272 		SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
273 		    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
274 
275 		if (DMU_OS_IS_L2CACHEABLE(os))
276 			aflags |= ARC_L2CACHE;
277 
278 		dprintf_bp(os->os_rootbp, "reading %s", "");
279 		/*
280 		 * XXX when bprewrite scrub can change the bp,
281 		 * and this is called from dmu_objset_open_ds_os, the bp
282 		 * could change, and we'll need a lock.
283 		 */
284 		err = dsl_read_nolock(NULL, spa, os->os_rootbp,
285 		    arc_getbuf_func, &os->os_phys_buf,
286 		    ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL, &aflags, &zb);
287 		if (err) {
288 			kmem_free(os, sizeof (objset_t));
289 			/* convert checksum errors into IO errors */
290 			if (err == ECKSUM)
291 				err = EIO;
292 			return (err);
293 		}
294 
295 		/* Increase the blocksize if we are permitted. */
296 		if (spa_version(spa) >= SPA_VERSION_USERSPACE &&
297 		    arc_buf_size(os->os_phys_buf) < sizeof (objset_phys_t)) {
298 			arc_buf_t *buf = arc_buf_alloc(spa,
299 			    sizeof (objset_phys_t), &os->os_phys_buf,
300 			    ARC_BUFC_METADATA);
301 			bzero(buf->b_data, sizeof (objset_phys_t));
302 			bcopy(os->os_phys_buf->b_data, buf->b_data,
303 			    arc_buf_size(os->os_phys_buf));
304 			(void) arc_buf_remove_ref(os->os_phys_buf,
305 			    &os->os_phys_buf);
306 			os->os_phys_buf = buf;
307 		}
308 
309 		os->os_phys = os->os_phys_buf->b_data;
310 		os->os_flags = os->os_phys->os_flags;
311 	} else {
312 		int size = spa_version(spa) >= SPA_VERSION_USERSPACE ?
313 		    sizeof (objset_phys_t) : OBJSET_OLD_PHYS_SIZE;
314 		os->os_phys_buf = arc_buf_alloc(spa, size,
315 		    &os->os_phys_buf, ARC_BUFC_METADATA);
316 		os->os_phys = os->os_phys_buf->b_data;
317 		bzero(os->os_phys, size);
318 	}
319 
320 	/*
321 	 * Note: the changed_cb will be called once before the register
322 	 * func returns, thus changing the checksum/compression from the
323 	 * default (fletcher2/off).  Snapshots don't need to know about
324 	 * checksum/compression/copies.
325 	 */
326 	if (ds) {
327 		err = dsl_prop_register(ds, "primarycache",
328 		    primary_cache_changed_cb, os);
329 		if (err == 0)
330 			err = dsl_prop_register(ds, "secondarycache",
331 			    secondary_cache_changed_cb, os);
332 		if (!dsl_dataset_is_snapshot(ds)) {
333 			if (err == 0)
334 				err = dsl_prop_register(ds, "checksum",
335 				    checksum_changed_cb, os);
336 			if (err == 0)
337 				err = dsl_prop_register(ds, "compression",
338 				    compression_changed_cb, os);
339 			if (err == 0)
340 				err = dsl_prop_register(ds, "copies",
341 				    copies_changed_cb, os);
342 			if (err == 0)
343 				err = dsl_prop_register(ds, "dedup",
344 				    dedup_changed_cb, os);
345 			if (err == 0)
346 				err = dsl_prop_register(ds, "logbias",
347 				    logbias_changed_cb, os);
348 			if (err == 0)
349 				err = dsl_prop_register(ds, "sync",
350 				    sync_changed_cb, os);
351 		}
352 		if (err) {
353 			VERIFY(arc_buf_remove_ref(os->os_phys_buf,
354 			    &os->os_phys_buf) == 1);
355 			kmem_free(os, sizeof (objset_t));
356 			return (err);
357 		}
358 	} else if (ds == NULL) {
359 		/* It's the meta-objset. */
360 		os->os_checksum = ZIO_CHECKSUM_FLETCHER_4;
361 		os->os_compress = ZIO_COMPRESS_LZJB;
362 		os->os_copies = spa_max_replication(spa);
363 		os->os_dedup_checksum = ZIO_CHECKSUM_OFF;
364 		os->os_dedup_verify = 0;
365 		os->os_logbias = 0;
366 		os->os_sync = 0;
367 		os->os_primary_cache = ZFS_CACHE_ALL;
368 		os->os_secondary_cache = ZFS_CACHE_ALL;
369 	}
370 
371 	if (ds == NULL || !dsl_dataset_is_snapshot(ds))
372 		os->os_zil_header = os->os_phys->os_zil_header;
373 	os->os_zil = zil_alloc(os, &os->os_zil_header);
374 
375 	for (i = 0; i < TXG_SIZE; i++) {
376 		list_create(&os->os_dirty_dnodes[i], sizeof (dnode_t),
377 		    offsetof(dnode_t, dn_dirty_link[i]));
378 		list_create(&os->os_free_dnodes[i], sizeof (dnode_t),
379 		    offsetof(dnode_t, dn_dirty_link[i]));
380 	}
381 	list_create(&os->os_dnodes, sizeof (dnode_t),
382 	    offsetof(dnode_t, dn_link));
383 	list_create(&os->os_downgraded_dbufs, sizeof (dmu_buf_impl_t),
384 	    offsetof(dmu_buf_impl_t, db_link));
385 
386 	mutex_init(&os->os_lock, NULL, MUTEX_DEFAULT, NULL);
387 	mutex_init(&os->os_obj_lock, NULL, MUTEX_DEFAULT, NULL);
388 	mutex_init(&os->os_user_ptr_lock, NULL, MUTEX_DEFAULT, NULL);
389 
390 	DMU_META_DNODE(os) = dnode_special_open(os,
391 	    &os->os_phys->os_meta_dnode, DMU_META_DNODE_OBJECT,
392 	    &os->os_meta_dnode);
393 	if (arc_buf_size(os->os_phys_buf) >= sizeof (objset_phys_t)) {
394 		DMU_USERUSED_DNODE(os) = dnode_special_open(os,
395 		    &os->os_phys->os_userused_dnode, DMU_USERUSED_OBJECT,
396 		    &os->os_userused_dnode);
397 		DMU_GROUPUSED_DNODE(os) = dnode_special_open(os,
398 		    &os->os_phys->os_groupused_dnode, DMU_GROUPUSED_OBJECT,
399 		    &os->os_groupused_dnode);
400 	}
401 
402 	/*
403 	 * We should be the only thread trying to do this because we
404 	 * have ds_opening_lock
405 	 */
406 	if (ds) {
407 		mutex_enter(&ds->ds_lock);
408 		ASSERT(ds->ds_objset == NULL);
409 		ds->ds_objset = os;
410 		mutex_exit(&ds->ds_lock);
411 	}
412 
413 	*osp = os;
414 	return (0);
415 }
416 
417 int
418 dmu_objset_from_ds(dsl_dataset_t *ds, objset_t **osp)
419 {
420 	int err = 0;
421 
422 	mutex_enter(&ds->ds_opening_lock);
423 	*osp = ds->ds_objset;
424 	if (*osp == NULL) {
425 		err = dmu_objset_open_impl(dsl_dataset_get_spa(ds),
426 		    ds, dsl_dataset_get_blkptr(ds), osp);
427 	}
428 	mutex_exit(&ds->ds_opening_lock);
429 	return (err);
430 }
431 
432 /* called from zpl */
433 int
434 dmu_objset_hold(const char *name, void *tag, objset_t **osp)
435 {
436 	dsl_dataset_t *ds;
437 	int err;
438 
439 	err = dsl_dataset_hold(name, tag, &ds);
440 	if (err)
441 		return (err);
442 
443 	err = dmu_objset_from_ds(ds, osp);
444 	if (err)
445 		dsl_dataset_rele(ds, tag);
446 
447 	return (err);
448 }
449 
450 /* called from zpl */
451 int
452 dmu_objset_own(const char *name, dmu_objset_type_t type,
453     boolean_t readonly, void *tag, objset_t **osp)
454 {
455 	dsl_dataset_t *ds;
456 	int err;
457 
458 	err = dsl_dataset_own(name, B_FALSE, tag, &ds);
459 	if (err)
460 		return (err);
461 
462 	err = dmu_objset_from_ds(ds, osp);
463 	if (err) {
464 		dsl_dataset_disown(ds, tag);
465 	} else if (type != DMU_OST_ANY && type != (*osp)->os_phys->os_type) {
466 		dmu_objset_disown(*osp, tag);
467 		return (EINVAL);
468 	} else if (!readonly && dsl_dataset_is_snapshot(ds)) {
469 		dmu_objset_disown(*osp, tag);
470 		return (EROFS);
471 	}
472 	return (err);
473 }
474 
475 void
476 dmu_objset_rele(objset_t *os, void *tag)
477 {
478 	dsl_dataset_rele(os->os_dsl_dataset, tag);
479 }
480 
481 void
482 dmu_objset_disown(objset_t *os, void *tag)
483 {
484 	dsl_dataset_disown(os->os_dsl_dataset, tag);
485 }
486 
487 int
488 dmu_objset_evict_dbufs(objset_t *os)
489 {
490 	dnode_t *dn;
491 
492 	mutex_enter(&os->os_lock);
493 
494 	/* process the mdn last, since the other dnodes have holds on it */
495 	list_remove(&os->os_dnodes, DMU_META_DNODE(os));
496 	list_insert_tail(&os->os_dnodes, DMU_META_DNODE(os));
497 
498 	/*
499 	 * Find the first dnode with holds.  We have to do this dance
500 	 * because dnode_add_ref() only works if you already have a
501 	 * hold.  If there are no holds then it has no dbufs so OK to
502 	 * skip.
503 	 */
504 	for (dn = list_head(&os->os_dnodes);
505 	    dn && !dnode_add_ref(dn, FTAG);
506 	    dn = list_next(&os->os_dnodes, dn))
507 		continue;
508 
509 	while (dn) {
510 		dnode_t *next_dn = dn;
511 
512 		do {
513 			next_dn = list_next(&os->os_dnodes, next_dn);
514 		} while (next_dn && !dnode_add_ref(next_dn, FTAG));
515 
516 		mutex_exit(&os->os_lock);
517 		dnode_evict_dbufs(dn);
518 		dnode_rele(dn, FTAG);
519 		mutex_enter(&os->os_lock);
520 		dn = next_dn;
521 	}
522 	dn = list_head(&os->os_dnodes);
523 	mutex_exit(&os->os_lock);
524 	return (dn != DMU_META_DNODE(os));
525 }
526 
527 void
528 dmu_objset_evict(objset_t *os)
529 {
530 	dsl_dataset_t *ds = os->os_dsl_dataset;
531 
532 	for (int t = 0; t < TXG_SIZE; t++)
533 		ASSERT(!dmu_objset_is_dirty(os, t));
534 
535 	if (ds) {
536 		if (!dsl_dataset_is_snapshot(ds)) {
537 			VERIFY(0 == dsl_prop_unregister(ds, "checksum",
538 			    checksum_changed_cb, os));
539 			VERIFY(0 == dsl_prop_unregister(ds, "compression",
540 			    compression_changed_cb, os));
541 			VERIFY(0 == dsl_prop_unregister(ds, "copies",
542 			    copies_changed_cb, os));
543 			VERIFY(0 == dsl_prop_unregister(ds, "dedup",
544 			    dedup_changed_cb, os));
545 			VERIFY(0 == dsl_prop_unregister(ds, "logbias",
546 			    logbias_changed_cb, os));
547 			VERIFY(0 == dsl_prop_unregister(ds, "sync",
548 			    sync_changed_cb, os));
549 		}
550 		VERIFY(0 == dsl_prop_unregister(ds, "primarycache",
551 		    primary_cache_changed_cb, os));
552 		VERIFY(0 == dsl_prop_unregister(ds, "secondarycache",
553 		    secondary_cache_changed_cb, os));
554 	}
555 
556 	if (os->os_sa)
557 		sa_tear_down(os);
558 
559 	/*
560 	 * We should need only a single pass over the dnode list, since
561 	 * nothing can be added to the list at this point.
562 	 */
563 	(void) dmu_objset_evict_dbufs(os);
564 
565 	dnode_special_close(&os->os_meta_dnode);
566 	if (DMU_USERUSED_DNODE(os)) {
567 		dnode_special_close(&os->os_userused_dnode);
568 		dnode_special_close(&os->os_groupused_dnode);
569 	}
570 	zil_free(os->os_zil);
571 
572 	ASSERT3P(list_head(&os->os_dnodes), ==, NULL);
573 
574 	VERIFY(arc_buf_remove_ref(os->os_phys_buf, &os->os_phys_buf) == 1);
575 
576 	/*
577 	 * This is a barrier to prevent the objset from going away in
578 	 * dnode_move() until we can safely ensure that the objset is still in
579 	 * use. We consider the objset valid before the barrier and invalid
580 	 * after the barrier.
581 	 */
582 	rw_enter(&os_lock, RW_READER);
583 	rw_exit(&os_lock);
584 
585 	mutex_destroy(&os->os_lock);
586 	mutex_destroy(&os->os_obj_lock);
587 	mutex_destroy(&os->os_user_ptr_lock);
588 	kmem_free(os, sizeof (objset_t));
589 }
590 
591 timestruc_t
592 dmu_objset_snap_cmtime(objset_t *os)
593 {
594 	return (dsl_dir_snap_cmtime(os->os_dsl_dataset->ds_dir));
595 }
596 
597 /* called from dsl for meta-objset */
598 objset_t *
599 dmu_objset_create_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
600     dmu_objset_type_t type, dmu_tx_t *tx)
601 {
602 	objset_t *os;
603 	dnode_t *mdn;
604 
605 	ASSERT(dmu_tx_is_syncing(tx));
606 	if (ds != NULL)
607 		VERIFY(0 == dmu_objset_from_ds(ds, &os));
608 	else
609 		VERIFY(0 == dmu_objset_open_impl(spa, NULL, bp, &os));
610 
611 	mdn = DMU_META_DNODE(os);
612 
613 	dnode_allocate(mdn, DMU_OT_DNODE, 1 << DNODE_BLOCK_SHIFT,
614 	    DN_MAX_INDBLKSHIFT, DMU_OT_NONE, 0, tx);
615 
616 	/*
617 	 * We don't want to have to increase the meta-dnode's nlevels
618 	 * later, because then we could do it in quescing context while
619 	 * we are also accessing it in open context.
620 	 *
621 	 * This precaution is not necessary for the MOS (ds == NULL),
622 	 * because the MOS is only updated in syncing context.
623 	 * This is most fortunate: the MOS is the only objset that
624 	 * needs to be synced multiple times as spa_sync() iterates
625 	 * to convergence, so minimizing its dn_nlevels matters.
626 	 */
627 	if (ds != NULL) {
628 		int levels = 1;
629 
630 		/*
631 		 * Determine the number of levels necessary for the meta-dnode
632 		 * to contain DN_MAX_OBJECT dnodes.
633 		 */
634 		while ((uint64_t)mdn->dn_nblkptr << (mdn->dn_datablkshift +
635 		    (levels - 1) * (mdn->dn_indblkshift - SPA_BLKPTRSHIFT)) <
636 		    DN_MAX_OBJECT * sizeof (dnode_phys_t))
637 			levels++;
638 
639 		mdn->dn_next_nlevels[tx->tx_txg & TXG_MASK] =
640 		    mdn->dn_nlevels = levels;
641 	}
642 
643 	ASSERT(type != DMU_OST_NONE);
644 	ASSERT(type != DMU_OST_ANY);
645 	ASSERT(type < DMU_OST_NUMTYPES);
646 	os->os_phys->os_type = type;
647 	if (dmu_objset_userused_enabled(os)) {
648 		os->os_phys->os_flags |= OBJSET_FLAG_USERACCOUNTING_COMPLETE;
649 		os->os_flags = os->os_phys->os_flags;
650 	}
651 
652 	dsl_dataset_dirty(ds, tx);
653 
654 	return (os);
655 }
656 
657 struct oscarg {
658 	void (*userfunc)(objset_t *os, void *arg, cred_t *cr, dmu_tx_t *tx);
659 	void *userarg;
660 	dsl_dataset_t *clone_origin;
661 	const char *lastname;
662 	dmu_objset_type_t type;
663 	uint64_t flags;
664 	cred_t *cr;
665 };
666 
667 /*ARGSUSED*/
668 static int
669 dmu_objset_create_check(void *arg1, void *arg2, dmu_tx_t *tx)
670 {
671 	dsl_dir_t *dd = arg1;
672 	struct oscarg *oa = arg2;
673 	objset_t *mos = dd->dd_pool->dp_meta_objset;
674 	int err;
675 	uint64_t ddobj;
676 
677 	err = zap_lookup(mos, dd->dd_phys->dd_child_dir_zapobj,
678 	    oa->lastname, sizeof (uint64_t), 1, &ddobj);
679 	if (err != ENOENT)
680 		return (err ? err : EEXIST);
681 
682 	if (oa->clone_origin != NULL) {
683 		/* You can't clone across pools. */
684 		if (oa->clone_origin->ds_dir->dd_pool != dd->dd_pool)
685 			return (EXDEV);
686 
687 		/* You can only clone snapshots, not the head datasets. */
688 		if (!dsl_dataset_is_snapshot(oa->clone_origin))
689 			return (EINVAL);
690 	}
691 
692 	return (0);
693 }
694 
695 static void
696 dmu_objset_create_sync(void *arg1, void *arg2, dmu_tx_t *tx)
697 {
698 	dsl_dir_t *dd = arg1;
699 	spa_t *spa = dd->dd_pool->dp_spa;
700 	struct oscarg *oa = arg2;
701 	uint64_t obj;
702 
703 	ASSERT(dmu_tx_is_syncing(tx));
704 
705 	obj = dsl_dataset_create_sync(dd, oa->lastname,
706 	    oa->clone_origin, oa->flags, oa->cr, tx);
707 
708 	if (oa->clone_origin == NULL) {
709 		dsl_pool_t *dp = dd->dd_pool;
710 		dsl_dataset_t *ds;
711 		blkptr_t *bp;
712 		objset_t *os;
713 
714 		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
715 		bp = dsl_dataset_get_blkptr(ds);
716 		ASSERT(BP_IS_HOLE(bp));
717 
718 		os = dmu_objset_create_impl(spa, ds, bp, oa->type, tx);
719 
720 		if (oa->userfunc)
721 			oa->userfunc(os, oa->userarg, oa->cr, tx);
722 		dsl_dataset_rele(ds, FTAG);
723 	}
724 
725 	spa_history_log_internal(LOG_DS_CREATE, spa, tx, "dataset = %llu", obj);
726 }
727 
728 int
729 dmu_objset_create(const char *name, dmu_objset_type_t type, uint64_t flags,
730     void (*func)(objset_t *os, void *arg, cred_t *cr, dmu_tx_t *tx), void *arg)
731 {
732 	dsl_dir_t *pdd;
733 	const char *tail;
734 	int err = 0;
735 	struct oscarg oa = { 0 };
736 
737 	ASSERT(strchr(name, '@') == NULL);
738 	err = dsl_dir_open(name, FTAG, &pdd, &tail);
739 	if (err)
740 		return (err);
741 	if (tail == NULL) {
742 		dsl_dir_close(pdd, FTAG);
743 		return (EEXIST);
744 	}
745 
746 	oa.userfunc = func;
747 	oa.userarg = arg;
748 	oa.lastname = tail;
749 	oa.type = type;
750 	oa.flags = flags;
751 	oa.cr = CRED();
752 
753 	err = dsl_sync_task_do(pdd->dd_pool, dmu_objset_create_check,
754 	    dmu_objset_create_sync, pdd, &oa, 5);
755 	dsl_dir_close(pdd, FTAG);
756 	return (err);
757 }
758 
759 int
760 dmu_objset_clone(const char *name, dsl_dataset_t *clone_origin, uint64_t flags)
761 {
762 	dsl_dir_t *pdd;
763 	const char *tail;
764 	int err = 0;
765 	struct oscarg oa = { 0 };
766 
767 	ASSERT(strchr(name, '@') == NULL);
768 	err = dsl_dir_open(name, FTAG, &pdd, &tail);
769 	if (err)
770 		return (err);
771 	if (tail == NULL) {
772 		dsl_dir_close(pdd, FTAG);
773 		return (EEXIST);
774 	}
775 
776 	oa.lastname = tail;
777 	oa.clone_origin = clone_origin;
778 	oa.flags = flags;
779 	oa.cr = CRED();
780 
781 	err = dsl_sync_task_do(pdd->dd_pool, dmu_objset_create_check,
782 	    dmu_objset_create_sync, pdd, &oa, 5);
783 	dsl_dir_close(pdd, FTAG);
784 	return (err);
785 }
786 
787 int
788 dmu_objset_destroy(const char *name, boolean_t defer)
789 {
790 	dsl_dataset_t *ds;
791 	int error;
792 
793 	error = dsl_dataset_own(name, B_TRUE, FTAG, &ds);
794 	if (error == 0) {
795 		error = dsl_dataset_destroy(ds, FTAG, defer);
796 		/* dsl_dataset_destroy() closes the ds. */
797 	}
798 
799 	return (error);
800 }
801 
802 struct snaparg {
803 	dsl_sync_task_group_t *dstg;
804 	char *snapname;
805 	char *htag;
806 	char failed[MAXPATHLEN];
807 	boolean_t recursive;
808 	boolean_t needsuspend;
809 	boolean_t temporary;
810 	nvlist_t *props;
811 	struct dsl_ds_holdarg *ha;	/* only needed in the temporary case */
812 	dsl_dataset_t *newds;
813 };
814 
815 static int
816 snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
817 {
818 	objset_t *os = arg1;
819 	struct snaparg *sn = arg2;
820 	int error;
821 
822 	/* The props have already been checked by zfs_check_userprops(). */
823 
824 	error = dsl_dataset_snapshot_check(os->os_dsl_dataset,
825 	    sn->snapname, tx);
826 	if (error)
827 		return (error);
828 
829 	if (sn->temporary) {
830 		/*
831 		 * Ideally we would just call
832 		 * dsl_dataset_user_hold_check() and
833 		 * dsl_dataset_destroy_check() here.  However the
834 		 * dataset we want to hold and destroy is the snapshot
835 		 * that we just confirmed we can create, but it won't
836 		 * exist until after these checks are run.  Do any
837 		 * checks we can here and if more checks are added to
838 		 * those routines in the future, similar checks may be
839 		 * necessary here.
840 		 */
841 		if (spa_version(os->os_spa) < SPA_VERSION_USERREFS)
842 			return (ENOTSUP);
843 		/*
844 		 * Not checking number of tags because the tag will be
845 		 * unique, as it will be the only tag.
846 		 */
847 		if (strlen(sn->htag) + MAX_TAG_PREFIX_LEN >= MAXNAMELEN)
848 			return (E2BIG);
849 
850 		sn->ha = kmem_alloc(sizeof (struct dsl_ds_holdarg), KM_SLEEP);
851 		sn->ha->temphold = B_TRUE;
852 		sn->ha->htag = sn->htag;
853 	}
854 	return (error);
855 }
856 
857 static void
858 snapshot_sync(void *arg1, void *arg2, dmu_tx_t *tx)
859 {
860 	objset_t *os = arg1;
861 	dsl_dataset_t *ds = os->os_dsl_dataset;
862 	struct snaparg *sn = arg2;
863 
864 	dsl_dataset_snapshot_sync(ds, sn->snapname, tx);
865 
866 	if (sn->props) {
867 		dsl_props_arg_t pa;
868 		pa.pa_props = sn->props;
869 		pa.pa_source = ZPROP_SRC_LOCAL;
870 		dsl_props_set_sync(ds->ds_prev, &pa, tx);
871 	}
872 
873 	if (sn->temporary) {
874 		struct dsl_ds_destroyarg da;
875 
876 		dsl_dataset_user_hold_sync(ds->ds_prev, sn->ha, tx);
877 		kmem_free(sn->ha, sizeof (struct dsl_ds_holdarg));
878 		sn->ha = NULL;
879 		sn->newds = ds->ds_prev;
880 
881 		da.ds = ds->ds_prev;
882 		da.defer = B_TRUE;
883 		dsl_dataset_destroy_sync(&da, FTAG, tx);
884 	}
885 }
886 
887 static int
888 dmu_objset_snapshot_one(const char *name, void *arg)
889 {
890 	struct snaparg *sn = arg;
891 	objset_t *os;
892 	int err;
893 	char *cp;
894 
895 	/*
896 	 * If the objset starts with a '%', then ignore it unless it was
897 	 * explicitly named (ie, not recursive).  These hidden datasets
898 	 * are always inconsistent, and by not opening them here, we can
899 	 * avoid a race with dsl_dir_destroy_check().
900 	 */
901 	cp = strrchr(name, '/');
902 	if (cp && cp[1] == '%' && sn->recursive)
903 		return (0);
904 
905 	(void) strcpy(sn->failed, name);
906 
907 	/*
908 	 * Check permissions if we are doing a recursive snapshot.  The
909 	 * permission checks for the starting dataset have already been
910 	 * performed in zfs_secpolicy_snapshot()
911 	 */
912 	if (sn->recursive && (err = zfs_secpolicy_snapshot_perms(name, CRED())))
913 		return (err);
914 
915 	err = dmu_objset_hold(name, sn, &os);
916 	if (err != 0)
917 		return (err);
918 
919 	/*
920 	 * If the objset is in an inconsistent state (eg, in the process
921 	 * of being destroyed), don't snapshot it.  As with %hidden
922 	 * datasets, we return EBUSY if this name was explicitly
923 	 * requested (ie, not recursive), and otherwise ignore it.
924 	 */
925 	if (os->os_dsl_dataset->ds_phys->ds_flags & DS_FLAG_INCONSISTENT) {
926 		dmu_objset_rele(os, sn);
927 		return (sn->recursive ? 0 : EBUSY);
928 	}
929 
930 	if (sn->needsuspend) {
931 		err = zil_suspend(dmu_objset_zil(os));
932 		if (err) {
933 			dmu_objset_rele(os, sn);
934 			return (err);
935 		}
936 	}
937 	dsl_sync_task_create(sn->dstg, snapshot_check, snapshot_sync,
938 	    os, sn, 3);
939 
940 	return (0);
941 }
942 
943 int
944 dmu_objset_snapshot(char *fsname, char *snapname, char *tag,
945     nvlist_t *props, boolean_t recursive, boolean_t temporary, int cleanup_fd)
946 {
947 	dsl_sync_task_t *dst;
948 	struct snaparg sn;
949 	spa_t *spa;
950 	minor_t minor;
951 	int err;
952 
953 	(void) strcpy(sn.failed, fsname);
954 
955 	err = spa_open(fsname, &spa, FTAG);
956 	if (err)
957 		return (err);
958 
959 	if (temporary) {
960 		if (cleanup_fd < 0) {
961 			spa_close(spa, FTAG);
962 			return (EINVAL);
963 		}
964 		if ((err = zfs_onexit_fd_hold(cleanup_fd, &minor)) != 0) {
965 			spa_close(spa, FTAG);
966 			return (err);
967 		}
968 	}
969 
970 	sn.dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
971 	sn.snapname = snapname;
972 	sn.htag = tag;
973 	sn.props = props;
974 	sn.recursive = recursive;
975 	sn.needsuspend = (spa_version(spa) < SPA_VERSION_FAST_SNAP);
976 	sn.temporary = temporary;
977 	sn.ha = NULL;
978 	sn.newds = NULL;
979 
980 	if (recursive) {
981 		err = dmu_objset_find(fsname,
982 		    dmu_objset_snapshot_one, &sn, DS_FIND_CHILDREN);
983 	} else {
984 		err = dmu_objset_snapshot_one(fsname, &sn);
985 	}
986 
987 	if (err == 0)
988 		err = dsl_sync_task_group_wait(sn.dstg);
989 
990 	for (dst = list_head(&sn.dstg->dstg_tasks); dst;
991 	    dst = list_next(&sn.dstg->dstg_tasks, dst)) {
992 		objset_t *os = dst->dst_arg1;
993 		dsl_dataset_t *ds = os->os_dsl_dataset;
994 		if (dst->dst_err) {
995 			dsl_dataset_name(ds, sn.failed);
996 		} else if (temporary) {
997 			dsl_register_onexit_hold_cleanup(sn.newds, tag, minor);
998 		}
999 		if (sn.needsuspend)
1000 			zil_resume(dmu_objset_zil(os));
1001 		dmu_objset_rele(os, &sn);
1002 	}
1003 
1004 	if (err)
1005 		(void) strcpy(fsname, sn.failed);
1006 	if (temporary)
1007 		zfs_onexit_fd_rele(cleanup_fd);
1008 	dsl_sync_task_group_destroy(sn.dstg);
1009 	spa_close(spa, FTAG);
1010 	return (err);
1011 }
1012 
1013 static void
1014 dmu_objset_sync_dnodes(list_t *list, list_t *newlist, dmu_tx_t *tx)
1015 {
1016 	dnode_t *dn;
1017 
1018 	while (dn = list_head(list)) {
1019 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
1020 		ASSERT(dn->dn_dbuf->db_data_pending);
1021 		/*
1022 		 * Initialize dn_zio outside dnode_sync() because the
1023 		 * meta-dnode needs to set it ouside dnode_sync().
1024 		 */
1025 		dn->dn_zio = dn->dn_dbuf->db_data_pending->dr_zio;
1026 		ASSERT(dn->dn_zio);
1027 
1028 		ASSERT3U(dn->dn_nlevels, <=, DN_MAX_LEVELS);
1029 		list_remove(list, dn);
1030 
1031 		if (newlist) {
1032 			(void) dnode_add_ref(dn, newlist);
1033 			list_insert_tail(newlist, dn);
1034 		}
1035 
1036 		dnode_sync(dn, tx);
1037 	}
1038 }
1039 
1040 /* ARGSUSED */
1041 static void
1042 dmu_objset_write_ready(zio_t *zio, arc_buf_t *abuf, void *arg)
1043 {
1044 	blkptr_t *bp = zio->io_bp;
1045 	objset_t *os = arg;
1046 	dnode_phys_t *dnp = &os->os_phys->os_meta_dnode;
1047 
1048 	ASSERT(bp == os->os_rootbp);
1049 	ASSERT(BP_GET_TYPE(bp) == DMU_OT_OBJSET);
1050 	ASSERT(BP_GET_LEVEL(bp) == 0);
1051 
1052 	/*
1053 	 * Update rootbp fill count: it should be the number of objects
1054 	 * allocated in the object set (not counting the "special"
1055 	 * objects that are stored in the objset_phys_t -- the meta
1056 	 * dnode and user/group accounting objects).
1057 	 */
1058 	bp->blk_fill = 0;
1059 	for (int i = 0; i < dnp->dn_nblkptr; i++)
1060 		bp->blk_fill += dnp->dn_blkptr[i].blk_fill;
1061 }
1062 
1063 /* ARGSUSED */
1064 static void
1065 dmu_objset_write_done(zio_t *zio, arc_buf_t *abuf, void *arg)
1066 {
1067 	blkptr_t *bp = zio->io_bp;
1068 	blkptr_t *bp_orig = &zio->io_bp_orig;
1069 	objset_t *os = arg;
1070 
1071 	if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
1072 		ASSERT(BP_EQUAL(bp, bp_orig));
1073 	} else {
1074 		dsl_dataset_t *ds = os->os_dsl_dataset;
1075 		dmu_tx_t *tx = os->os_synctx;
1076 
1077 		(void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
1078 		dsl_dataset_block_born(ds, bp, tx);
1079 	}
1080 }
1081 
1082 /* called from dsl */
1083 void
1084 dmu_objset_sync(objset_t *os, zio_t *pio, dmu_tx_t *tx)
1085 {
1086 	int txgoff;
1087 	zbookmark_t zb;
1088 	zio_prop_t zp;
1089 	zio_t *zio;
1090 	list_t *list;
1091 	list_t *newlist = NULL;
1092 	dbuf_dirty_record_t *dr;
1093 
1094 	dprintf_ds(os->os_dsl_dataset, "txg=%llu\n", tx->tx_txg);
1095 
1096 	ASSERT(dmu_tx_is_syncing(tx));
1097 	/* XXX the write_done callback should really give us the tx... */
1098 	os->os_synctx = tx;
1099 
1100 	if (os->os_dsl_dataset == NULL) {
1101 		/*
1102 		 * This is the MOS.  If we have upgraded,
1103 		 * spa_max_replication() could change, so reset
1104 		 * os_copies here.
1105 		 */
1106 		os->os_copies = spa_max_replication(os->os_spa);
1107 	}
1108 
1109 	/*
1110 	 * Create the root block IO
1111 	 */
1112 	SET_BOOKMARK(&zb, os->os_dsl_dataset ?
1113 	    os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1114 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
1115 	VERIFY3U(0, ==, arc_release_bp(os->os_phys_buf, &os->os_phys_buf,
1116 	    os->os_rootbp, os->os_spa, &zb));
1117 
1118 	dmu_write_policy(os, NULL, 0, 0, &zp);
1119 
1120 	zio = arc_write(pio, os->os_spa, tx->tx_txg,
1121 	    os->os_rootbp, os->os_phys_buf, DMU_OS_IS_L2CACHEABLE(os), &zp,
1122 	    dmu_objset_write_ready, dmu_objset_write_done, os,
1123 	    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
1124 
1125 	/*
1126 	 * Sync special dnodes - the parent IO for the sync is the root block
1127 	 */
1128 	DMU_META_DNODE(os)->dn_zio = zio;
1129 	dnode_sync(DMU_META_DNODE(os), tx);
1130 
1131 	os->os_phys->os_flags = os->os_flags;
1132 
1133 	if (DMU_USERUSED_DNODE(os) &&
1134 	    DMU_USERUSED_DNODE(os)->dn_type != DMU_OT_NONE) {
1135 		DMU_USERUSED_DNODE(os)->dn_zio = zio;
1136 		dnode_sync(DMU_USERUSED_DNODE(os), tx);
1137 		DMU_GROUPUSED_DNODE(os)->dn_zio = zio;
1138 		dnode_sync(DMU_GROUPUSED_DNODE(os), tx);
1139 	}
1140 
1141 	txgoff = tx->tx_txg & TXG_MASK;
1142 
1143 	if (dmu_objset_userused_enabled(os)) {
1144 		newlist = &os->os_synced_dnodes;
1145 		/*
1146 		 * We must create the list here because it uses the
1147 		 * dn_dirty_link[] of this txg.
1148 		 */
1149 		list_create(newlist, sizeof (dnode_t),
1150 		    offsetof(dnode_t, dn_dirty_link[txgoff]));
1151 	}
1152 
1153 	dmu_objset_sync_dnodes(&os->os_free_dnodes[txgoff], newlist, tx);
1154 	dmu_objset_sync_dnodes(&os->os_dirty_dnodes[txgoff], newlist, tx);
1155 
1156 	list = &DMU_META_DNODE(os)->dn_dirty_records[txgoff];
1157 	while (dr = list_head(list)) {
1158 		ASSERT(dr->dr_dbuf->db_level == 0);
1159 		list_remove(list, dr);
1160 		if (dr->dr_zio)
1161 			zio_nowait(dr->dr_zio);
1162 	}
1163 	/*
1164 	 * Free intent log blocks up to this tx.
1165 	 */
1166 	zil_sync(os->os_zil, tx);
1167 	os->os_phys->os_zil_header = os->os_zil_header;
1168 	zio_nowait(zio);
1169 }
1170 
1171 boolean_t
1172 dmu_objset_is_dirty(objset_t *os, uint64_t txg)
1173 {
1174 	return (!list_is_empty(&os->os_dirty_dnodes[txg & TXG_MASK]) ||
1175 	    !list_is_empty(&os->os_free_dnodes[txg & TXG_MASK]));
1176 }
1177 
1178 boolean_t
1179 dmu_objset_is_dirty_anywhere(objset_t *os)
1180 {
1181 	for (int t = 0; t < TXG_SIZE; t++)
1182 		if (dmu_objset_is_dirty(os, t))
1183 			return (B_TRUE);
1184 	return (B_FALSE);
1185 }
1186 
1187 static objset_used_cb_t *used_cbs[DMU_OST_NUMTYPES];
1188 
1189 void
1190 dmu_objset_register_type(dmu_objset_type_t ost, objset_used_cb_t *cb)
1191 {
1192 	used_cbs[ost] = cb;
1193 }
1194 
1195 boolean_t
1196 dmu_objset_userused_enabled(objset_t *os)
1197 {
1198 	return (spa_version(os->os_spa) >= SPA_VERSION_USERSPACE &&
1199 	    used_cbs[os->os_phys->os_type] != NULL &&
1200 	    DMU_USERUSED_DNODE(os) != NULL);
1201 }
1202 
1203 static void
1204 do_userquota_update(objset_t *os, uint64_t used, uint64_t flags,
1205     uint64_t user, uint64_t group, boolean_t subtract, dmu_tx_t *tx)
1206 {
1207 	if ((flags & DNODE_FLAG_USERUSED_ACCOUNTED)) {
1208 		int64_t delta = DNODE_SIZE + used;
1209 		if (subtract)
1210 			delta = -delta;
1211 		VERIFY3U(0, ==, zap_increment_int(os, DMU_USERUSED_OBJECT,
1212 		    user, delta, tx));
1213 		VERIFY3U(0, ==, zap_increment_int(os, DMU_GROUPUSED_OBJECT,
1214 		    group, delta, tx));
1215 	}
1216 }
1217 
1218 void
1219 dmu_objset_do_userquota_updates(objset_t *os, dmu_tx_t *tx)
1220 {
1221 	dnode_t *dn;
1222 	list_t *list = &os->os_synced_dnodes;
1223 
1224 	ASSERT(list_head(list) == NULL || dmu_objset_userused_enabled(os));
1225 
1226 	while (dn = list_head(list)) {
1227 		int flags;
1228 		ASSERT(!DMU_OBJECT_IS_SPECIAL(dn->dn_object));
1229 		ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE ||
1230 		    dn->dn_phys->dn_flags &
1231 		    DNODE_FLAG_USERUSED_ACCOUNTED);
1232 
1233 		/* Allocate the user/groupused objects if necessary. */
1234 		if (DMU_USERUSED_DNODE(os)->dn_type == DMU_OT_NONE) {
1235 			VERIFY(0 == zap_create_claim(os,
1236 			    DMU_USERUSED_OBJECT,
1237 			    DMU_OT_USERGROUP_USED, DMU_OT_NONE, 0, tx));
1238 			VERIFY(0 == zap_create_claim(os,
1239 			    DMU_GROUPUSED_OBJECT,
1240 			    DMU_OT_USERGROUP_USED, DMU_OT_NONE, 0, tx));
1241 		}
1242 
1243 		/*
1244 		 * We intentionally modify the zap object even if the
1245 		 * net delta is zero.  Otherwise
1246 		 * the block of the zap obj could be shared between
1247 		 * datasets but need to be different between them after
1248 		 * a bprewrite.
1249 		 */
1250 
1251 		flags = dn->dn_id_flags;
1252 		ASSERT(flags);
1253 		if (flags & DN_ID_OLD_EXIST)  {
1254 			do_userquota_update(os, dn->dn_oldused, dn->dn_oldflags,
1255 			    dn->dn_olduid, dn->dn_oldgid, B_TRUE, tx);
1256 		}
1257 		if (flags & DN_ID_NEW_EXIST) {
1258 			do_userquota_update(os, DN_USED_BYTES(dn->dn_phys),
1259 			    dn->dn_phys->dn_flags,  dn->dn_newuid,
1260 			    dn->dn_newgid, B_FALSE, tx);
1261 		}
1262 
1263 		mutex_enter(&dn->dn_mtx);
1264 		dn->dn_oldused = 0;
1265 		dn->dn_oldflags = 0;
1266 		if (dn->dn_id_flags & DN_ID_NEW_EXIST) {
1267 			dn->dn_olduid = dn->dn_newuid;
1268 			dn->dn_oldgid = dn->dn_newgid;
1269 			dn->dn_id_flags |= DN_ID_OLD_EXIST;
1270 			if (dn->dn_bonuslen == 0)
1271 				dn->dn_id_flags |= DN_ID_CHKED_SPILL;
1272 			else
1273 				dn->dn_id_flags |= DN_ID_CHKED_BONUS;
1274 		}
1275 		dn->dn_id_flags &= ~(DN_ID_NEW_EXIST);
1276 		mutex_exit(&dn->dn_mtx);
1277 
1278 		list_remove(list, dn);
1279 		dnode_rele(dn, list);
1280 	}
1281 }
1282 
1283 /*
1284  * Returns a pointer to data to find uid/gid from
1285  *
1286  * If a dirty record for transaction group that is syncing can't
1287  * be found then NULL is returned.  In the NULL case it is assumed
1288  * the uid/gid aren't changing.
1289  */
1290 static void *
1291 dmu_objset_userquota_find_data(dmu_buf_impl_t *db, dmu_tx_t *tx)
1292 {
1293 	dbuf_dirty_record_t *dr, **drp;
1294 	void *data;
1295 
1296 	if (db->db_dirtycnt == 0)
1297 		return (db->db.db_data);  /* Nothing is changing */
1298 
1299 	for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1300 		if (dr->dr_txg == tx->tx_txg)
1301 			break;
1302 
1303 	if (dr == NULL) {
1304 		data = NULL;
1305 	} else {
1306 		dnode_t *dn;
1307 
1308 		DB_DNODE_ENTER(dr->dr_dbuf);
1309 		dn = DB_DNODE(dr->dr_dbuf);
1310 
1311 		if (dn->dn_bonuslen == 0 &&
1312 		    dr->dr_dbuf->db_blkid == DMU_SPILL_BLKID)
1313 			data = dr->dt.dl.dr_data->b_data;
1314 		else
1315 			data = dr->dt.dl.dr_data;
1316 
1317 		DB_DNODE_EXIT(dr->dr_dbuf);
1318 	}
1319 
1320 	return (data);
1321 }
1322 
1323 void
1324 dmu_objset_userquota_get_ids(dnode_t *dn, boolean_t before, dmu_tx_t *tx)
1325 {
1326 	objset_t *os = dn->dn_objset;
1327 	void *data = NULL;
1328 	dmu_buf_impl_t *db = NULL;
1329 	uint64_t *user, *group;
1330 	int flags = dn->dn_id_flags;
1331 	int error;
1332 	boolean_t have_spill = B_FALSE;
1333 
1334 	if (!dmu_objset_userused_enabled(dn->dn_objset))
1335 		return;
1336 
1337 	if (before && (flags & (DN_ID_CHKED_BONUS|DN_ID_OLD_EXIST|
1338 	    DN_ID_CHKED_SPILL)))
1339 		return;
1340 
1341 	if (before && dn->dn_bonuslen != 0)
1342 		data = DN_BONUS(dn->dn_phys);
1343 	else if (!before && dn->dn_bonuslen != 0) {
1344 		if (dn->dn_bonus) {
1345 			db = dn->dn_bonus;
1346 			mutex_enter(&db->db_mtx);
1347 			data = dmu_objset_userquota_find_data(db, tx);
1348 		} else {
1349 			data = DN_BONUS(dn->dn_phys);
1350 		}
1351 	} else if (dn->dn_bonuslen == 0 && dn->dn_bonustype == DMU_OT_SA) {
1352 			int rf = 0;
1353 
1354 			if (RW_WRITE_HELD(&dn->dn_struct_rwlock))
1355 				rf |= DB_RF_HAVESTRUCT;
1356 			error = dmu_spill_hold_by_dnode(dn,
1357 			    rf | DB_RF_MUST_SUCCEED,
1358 			    FTAG, (dmu_buf_t **)&db);
1359 			ASSERT(error == 0);
1360 			mutex_enter(&db->db_mtx);
1361 			data = (before) ? db->db.db_data :
1362 			    dmu_objset_userquota_find_data(db, tx);
1363 			have_spill = B_TRUE;
1364 	} else {
1365 		mutex_enter(&dn->dn_mtx);
1366 		dn->dn_id_flags |= DN_ID_CHKED_BONUS;
1367 		mutex_exit(&dn->dn_mtx);
1368 		return;
1369 	}
1370 
1371 	if (before) {
1372 		ASSERT(data);
1373 		user = &dn->dn_olduid;
1374 		group = &dn->dn_oldgid;
1375 	} else if (data) {
1376 		user = &dn->dn_newuid;
1377 		group = &dn->dn_newgid;
1378 	}
1379 
1380 	/*
1381 	 * Must always call the callback in case the object
1382 	 * type has changed and that type isn't an object type to track
1383 	 */
1384 	error = used_cbs[os->os_phys->os_type](dn->dn_bonustype, data,
1385 	    user, group);
1386 
1387 	/*
1388 	 * Preserve existing uid/gid when the callback can't determine
1389 	 * what the new uid/gid are and the callback returned EEXIST.
1390 	 * The EEXIST error tells us to just use the existing uid/gid.
1391 	 * If we don't know what the old values are then just assign
1392 	 * them to 0, since that is a new file  being created.
1393 	 */
1394 	if (!before && data == NULL && error == EEXIST) {
1395 		if (flags & DN_ID_OLD_EXIST) {
1396 			dn->dn_newuid = dn->dn_olduid;
1397 			dn->dn_newgid = dn->dn_oldgid;
1398 		} else {
1399 			dn->dn_newuid = 0;
1400 			dn->dn_newgid = 0;
1401 		}
1402 		error = 0;
1403 	}
1404 
1405 	if (db)
1406 		mutex_exit(&db->db_mtx);
1407 
1408 	mutex_enter(&dn->dn_mtx);
1409 	if (error == 0 && before)
1410 		dn->dn_id_flags |= DN_ID_OLD_EXIST;
1411 	if (error == 0 && !before)
1412 		dn->dn_id_flags |= DN_ID_NEW_EXIST;
1413 
1414 	if (have_spill) {
1415 		dn->dn_id_flags |= DN_ID_CHKED_SPILL;
1416 	} else {
1417 		dn->dn_id_flags |= DN_ID_CHKED_BONUS;
1418 	}
1419 	mutex_exit(&dn->dn_mtx);
1420 	if (have_spill)
1421 		dmu_buf_rele((dmu_buf_t *)db, FTAG);
1422 }
1423 
1424 boolean_t
1425 dmu_objset_userspace_present(objset_t *os)
1426 {
1427 	return (os->os_phys->os_flags &
1428 	    OBJSET_FLAG_USERACCOUNTING_COMPLETE);
1429 }
1430 
1431 int
1432 dmu_objset_userspace_upgrade(objset_t *os)
1433 {
1434 	uint64_t obj;
1435 	int err = 0;
1436 
1437 	if (dmu_objset_userspace_present(os))
1438 		return (0);
1439 	if (!dmu_objset_userused_enabled(os))
1440 		return (ENOTSUP);
1441 	if (dmu_objset_is_snapshot(os))
1442 		return (EINVAL);
1443 
1444 	/*
1445 	 * We simply need to mark every object dirty, so that it will be
1446 	 * synced out and now accounted.  If this is called
1447 	 * concurrently, or if we already did some work before crashing,
1448 	 * that's fine, since we track each object's accounted state
1449 	 * independently.
1450 	 */
1451 
1452 	for (obj = 0; err == 0; err = dmu_object_next(os, &obj, FALSE, 0)) {
1453 		dmu_tx_t *tx;
1454 		dmu_buf_t *db;
1455 		int objerr;
1456 
1457 		if (issig(JUSTLOOKING) && issig(FORREAL))
1458 			return (EINTR);
1459 
1460 		objerr = dmu_bonus_hold(os, obj, FTAG, &db);
1461 		if (objerr)
1462 			continue;
1463 		tx = dmu_tx_create(os);
1464 		dmu_tx_hold_bonus(tx, obj);
1465 		objerr = dmu_tx_assign(tx, TXG_WAIT);
1466 		if (objerr) {
1467 			dmu_tx_abort(tx);
1468 			continue;
1469 		}
1470 		dmu_buf_will_dirty(db, tx);
1471 		dmu_buf_rele(db, FTAG);
1472 		dmu_tx_commit(tx);
1473 	}
1474 
1475 	os->os_flags |= OBJSET_FLAG_USERACCOUNTING_COMPLETE;
1476 	txg_wait_synced(dmu_objset_pool(os), 0);
1477 	return (0);
1478 }
1479 
1480 void
1481 dmu_objset_space(objset_t *os, uint64_t *refdbytesp, uint64_t *availbytesp,
1482     uint64_t *usedobjsp, uint64_t *availobjsp)
1483 {
1484 	dsl_dataset_space(os->os_dsl_dataset, refdbytesp, availbytesp,
1485 	    usedobjsp, availobjsp);
1486 }
1487 
1488 uint64_t
1489 dmu_objset_fsid_guid(objset_t *os)
1490 {
1491 	return (dsl_dataset_fsid_guid(os->os_dsl_dataset));
1492 }
1493 
1494 void
1495 dmu_objset_fast_stat(objset_t *os, dmu_objset_stats_t *stat)
1496 {
1497 	stat->dds_type = os->os_phys->os_type;
1498 	if (os->os_dsl_dataset)
1499 		dsl_dataset_fast_stat(os->os_dsl_dataset, stat);
1500 }
1501 
1502 void
1503 dmu_objset_stats(objset_t *os, nvlist_t *nv)
1504 {
1505 	ASSERT(os->os_dsl_dataset ||
1506 	    os->os_phys->os_type == DMU_OST_META);
1507 
1508 	if (os->os_dsl_dataset != NULL)
1509 		dsl_dataset_stats(os->os_dsl_dataset, nv);
1510 
1511 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_TYPE,
1512 	    os->os_phys->os_type);
1513 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USERACCOUNTING,
1514 	    dmu_objset_userspace_present(os));
1515 }
1516 
1517 int
1518 dmu_objset_is_snapshot(objset_t *os)
1519 {
1520 	if (os->os_dsl_dataset != NULL)
1521 		return (dsl_dataset_is_snapshot(os->os_dsl_dataset));
1522 	else
1523 		return (B_FALSE);
1524 }
1525 
1526 int
1527 dmu_snapshot_realname(objset_t *os, char *name, char *real, int maxlen,
1528     boolean_t *conflict)
1529 {
1530 	dsl_dataset_t *ds = os->os_dsl_dataset;
1531 	uint64_t ignored;
1532 
1533 	if (ds->ds_phys->ds_snapnames_zapobj == 0)
1534 		return (ENOENT);
1535 
1536 	return (zap_lookup_norm(ds->ds_dir->dd_pool->dp_meta_objset,
1537 	    ds->ds_phys->ds_snapnames_zapobj, name, 8, 1, &ignored, MT_FIRST,
1538 	    real, maxlen, conflict));
1539 }
1540 
1541 int
1542 dmu_snapshot_list_next(objset_t *os, int namelen, char *name,
1543     uint64_t *idp, uint64_t *offp, boolean_t *case_conflict)
1544 {
1545 	dsl_dataset_t *ds = os->os_dsl_dataset;
1546 	zap_cursor_t cursor;
1547 	zap_attribute_t attr;
1548 
1549 	if (ds->ds_phys->ds_snapnames_zapobj == 0)
1550 		return (ENOENT);
1551 
1552 	zap_cursor_init_serialized(&cursor,
1553 	    ds->ds_dir->dd_pool->dp_meta_objset,
1554 	    ds->ds_phys->ds_snapnames_zapobj, *offp);
1555 
1556 	if (zap_cursor_retrieve(&cursor, &attr) != 0) {
1557 		zap_cursor_fini(&cursor);
1558 		return (ENOENT);
1559 	}
1560 
1561 	if (strlen(attr.za_name) + 1 > namelen) {
1562 		zap_cursor_fini(&cursor);
1563 		return (ENAMETOOLONG);
1564 	}
1565 
1566 	(void) strcpy(name, attr.za_name);
1567 	if (idp)
1568 		*idp = attr.za_first_integer;
1569 	if (case_conflict)
1570 		*case_conflict = attr.za_normalization_conflict;
1571 	zap_cursor_advance(&cursor);
1572 	*offp = zap_cursor_serialize(&cursor);
1573 	zap_cursor_fini(&cursor);
1574 
1575 	return (0);
1576 }
1577 
1578 int
1579 dmu_dir_list_next(objset_t *os, int namelen, char *name,
1580     uint64_t *idp, uint64_t *offp)
1581 {
1582 	dsl_dir_t *dd = os->os_dsl_dataset->ds_dir;
1583 	zap_cursor_t cursor;
1584 	zap_attribute_t attr;
1585 
1586 	/* there is no next dir on a snapshot! */
1587 	if (os->os_dsl_dataset->ds_object !=
1588 	    dd->dd_phys->dd_head_dataset_obj)
1589 		return (ENOENT);
1590 
1591 	zap_cursor_init_serialized(&cursor,
1592 	    dd->dd_pool->dp_meta_objset,
1593 	    dd->dd_phys->dd_child_dir_zapobj, *offp);
1594 
1595 	if (zap_cursor_retrieve(&cursor, &attr) != 0) {
1596 		zap_cursor_fini(&cursor);
1597 		return (ENOENT);
1598 	}
1599 
1600 	if (strlen(attr.za_name) + 1 > namelen) {
1601 		zap_cursor_fini(&cursor);
1602 		return (ENAMETOOLONG);
1603 	}
1604 
1605 	(void) strcpy(name, attr.za_name);
1606 	if (idp)
1607 		*idp = attr.za_first_integer;
1608 	zap_cursor_advance(&cursor);
1609 	*offp = zap_cursor_serialize(&cursor);
1610 	zap_cursor_fini(&cursor);
1611 
1612 	return (0);
1613 }
1614 
1615 struct findarg {
1616 	int (*func)(const char *, void *);
1617 	void *arg;
1618 };
1619 
1620 /* ARGSUSED */
1621 static int
1622 findfunc(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1623 {
1624 	struct findarg *fa = arg;
1625 	return (fa->func(dsname, fa->arg));
1626 }
1627 
1628 /*
1629  * Find all objsets under name, and for each, call 'func(child_name, arg)'.
1630  * Perhaps change all callers to use dmu_objset_find_spa()?
1631  */
1632 int
1633 dmu_objset_find(char *name, int func(const char *, void *), void *arg,
1634     int flags)
1635 {
1636 	struct findarg fa;
1637 	fa.func = func;
1638 	fa.arg = arg;
1639 	return (dmu_objset_find_spa(NULL, name, findfunc, &fa, flags));
1640 }
1641 
1642 /*
1643  * Find all objsets under name, call func on each
1644  */
1645 int
1646 dmu_objset_find_spa(spa_t *spa, const char *name,
1647     int func(spa_t *, uint64_t, const char *, void *), void *arg, int flags)
1648 {
1649 	dsl_dir_t *dd;
1650 	dsl_pool_t *dp;
1651 	dsl_dataset_t *ds;
1652 	zap_cursor_t zc;
1653 	zap_attribute_t *attr;
1654 	char *child;
1655 	uint64_t thisobj;
1656 	int err;
1657 
1658 	if (name == NULL)
1659 		name = spa_name(spa);
1660 	err = dsl_dir_open_spa(spa, name, FTAG, &dd, NULL);
1661 	if (err)
1662 		return (err);
1663 
1664 	/* Don't visit hidden ($MOS & $ORIGIN) objsets. */
1665 	if (dd->dd_myname[0] == '$') {
1666 		dsl_dir_close(dd, FTAG);
1667 		return (0);
1668 	}
1669 
1670 	thisobj = dd->dd_phys->dd_head_dataset_obj;
1671 	attr = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
1672 	dp = dd->dd_pool;
1673 
1674 	/*
1675 	 * Iterate over all children.
1676 	 */
1677 	if (flags & DS_FIND_CHILDREN) {
1678 		for (zap_cursor_init(&zc, dp->dp_meta_objset,
1679 		    dd->dd_phys->dd_child_dir_zapobj);
1680 		    zap_cursor_retrieve(&zc, attr) == 0;
1681 		    (void) zap_cursor_advance(&zc)) {
1682 			ASSERT(attr->za_integer_length == sizeof (uint64_t));
1683 			ASSERT(attr->za_num_integers == 1);
1684 
1685 			child = kmem_asprintf("%s/%s", name, attr->za_name);
1686 			err = dmu_objset_find_spa(spa, child, func, arg, flags);
1687 			strfree(child);
1688 			if (err)
1689 				break;
1690 		}
1691 		zap_cursor_fini(&zc);
1692 
1693 		if (err) {
1694 			dsl_dir_close(dd, FTAG);
1695 			kmem_free(attr, sizeof (zap_attribute_t));
1696 			return (err);
1697 		}
1698 	}
1699 
1700 	/*
1701 	 * Iterate over all snapshots.
1702 	 */
1703 	if (flags & DS_FIND_SNAPSHOTS) {
1704 		if (!dsl_pool_sync_context(dp))
1705 			rw_enter(&dp->dp_config_rwlock, RW_READER);
1706 		err = dsl_dataset_hold_obj(dp, thisobj, FTAG, &ds);
1707 		if (!dsl_pool_sync_context(dp))
1708 			rw_exit(&dp->dp_config_rwlock);
1709 
1710 		if (err == 0) {
1711 			uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
1712 			dsl_dataset_rele(ds, FTAG);
1713 
1714 			for (zap_cursor_init(&zc, dp->dp_meta_objset, snapobj);
1715 			    zap_cursor_retrieve(&zc, attr) == 0;
1716 			    (void) zap_cursor_advance(&zc)) {
1717 				ASSERT(attr->za_integer_length ==
1718 				    sizeof (uint64_t));
1719 				ASSERT(attr->za_num_integers == 1);
1720 
1721 				child = kmem_asprintf("%s@%s",
1722 				    name, attr->za_name);
1723 				err = func(spa, attr->za_first_integer,
1724 				    child, arg);
1725 				strfree(child);
1726 				if (err)
1727 					break;
1728 			}
1729 			zap_cursor_fini(&zc);
1730 		}
1731 	}
1732 
1733 	dsl_dir_close(dd, FTAG);
1734 	kmem_free(attr, sizeof (zap_attribute_t));
1735 
1736 	if (err)
1737 		return (err);
1738 
1739 	/*
1740 	 * Apply to self if appropriate.
1741 	 */
1742 	err = func(spa, thisobj, name, arg);
1743 	return (err);
1744 }
1745 
1746 /* ARGSUSED */
1747 int
1748 dmu_objset_prefetch(const char *name, void *arg)
1749 {
1750 	dsl_dataset_t *ds;
1751 
1752 	if (dsl_dataset_hold(name, FTAG, &ds))
1753 		return (0);
1754 
1755 	if (!BP_IS_HOLE(&ds->ds_phys->ds_bp)) {
1756 		mutex_enter(&ds->ds_opening_lock);
1757 		if (ds->ds_objset == NULL) {
1758 			uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1759 			zbookmark_t zb;
1760 
1761 			SET_BOOKMARK(&zb, ds->ds_object, ZB_ROOT_OBJECT,
1762 			    ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
1763 
1764 			(void) dsl_read_nolock(NULL, dsl_dataset_get_spa(ds),
1765 			    &ds->ds_phys->ds_bp, NULL, NULL,
1766 			    ZIO_PRIORITY_ASYNC_READ,
1767 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1768 			    &aflags, &zb);
1769 		}
1770 		mutex_exit(&ds->ds_opening_lock);
1771 	}
1772 
1773 	dsl_dataset_rele(ds, FTAG);
1774 	return (0);
1775 }
1776 
1777 void
1778 dmu_objset_set_user(objset_t *os, void *user_ptr)
1779 {
1780 	ASSERT(MUTEX_HELD(&os->os_user_ptr_lock));
1781 	os->os_user_ptr = user_ptr;
1782 }
1783 
1784 void *
1785 dmu_objset_get_user(objset_t *os)
1786 {
1787 	ASSERT(MUTEX_HELD(&os->os_user_ptr_lock));
1788 	return (os->os_user_ptr);
1789 }
1790