xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_objset.c (revision adb52d92)
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 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
25  * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26  * Copyright (c) 2013, Joyent, Inc. All rights reserved.
27  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
28  * Copyright (c) 2015, STRATO AG, Inc. All rights reserved.
29  * Copyright (c) 2014 Integros [integros.com]
30  * Copyright 2017 Nexenta Systems, Inc.
31  */
32 
33 /* Portions Copyright 2010 Robert Milkowski */
34 
35 #include <sys/cred.h>
36 #include <sys/zfs_context.h>
37 #include <sys/dmu_objset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dsl_dataset.h>
40 #include <sys/dsl_prop.h>
41 #include <sys/dsl_pool.h>
42 #include <sys/dsl_synctask.h>
43 #include <sys/dsl_deleg.h>
44 #include <sys/dnode.h>
45 #include <sys/dbuf.h>
46 #include <sys/zvol.h>
47 #include <sys/dmu_tx.h>
48 #include <sys/zap.h>
49 #include <sys/zil.h>
50 #include <sys/dmu_impl.h>
51 #include <sys/zfs_ioctl.h>
52 #include <sys/sa.h>
53 #include <sys/zfs_onexit.h>
54 #include <sys/dsl_destroy.h>
55 #include <sys/vdev.h>
56 #include <sys/zfeature.h>
57 
58 /*
59  * Needed to close a window in dnode_move() that allows the objset to be freed
60  * before it can be safely accessed.
61  */
62 krwlock_t os_lock;
63 
64 /*
65  * Tunable to overwrite the maximum number of threads for the parallization
66  * of dmu_objset_find_dp, needed to speed up the import of pools with many
67  * datasets.
68  * Default is 4 times the number of leaf vdevs.
69  */
70 int dmu_find_threads = 0;
71 
72 /*
73  * Backfill lower metadnode objects after this many have been freed.
74  * Backfilling negatively impacts object creation rates, so only do it
75  * if there are enough holes to fill.
76  */
77 int dmu_rescan_dnode_threshold = 131072;
78 
79 static void dmu_objset_find_dp_cb(void *arg);
80 
81 void
82 dmu_objset_init(void)
83 {
84 	rw_init(&os_lock, NULL, RW_DEFAULT, NULL);
85 }
86 
87 void
88 dmu_objset_fini(void)
89 {
90 	rw_destroy(&os_lock);
91 }
92 
93 spa_t *
94 dmu_objset_spa(objset_t *os)
95 {
96 	return (os->os_spa);
97 }
98 
99 zilog_t *
100 dmu_objset_zil(objset_t *os)
101 {
102 	return (os->os_zil);
103 }
104 
105 dsl_pool_t *
106 dmu_objset_pool(objset_t *os)
107 {
108 	dsl_dataset_t *ds;
109 
110 	if ((ds = os->os_dsl_dataset) != NULL && ds->ds_dir)
111 		return (ds->ds_dir->dd_pool);
112 	else
113 		return (spa_get_dsl(os->os_spa));
114 }
115 
116 dsl_dataset_t *
117 dmu_objset_ds(objset_t *os)
118 {
119 	return (os->os_dsl_dataset);
120 }
121 
122 dmu_objset_type_t
123 dmu_objset_type(objset_t *os)
124 {
125 	return (os->os_phys->os_type);
126 }
127 
128 void
129 dmu_objset_name(objset_t *os, char *buf)
130 {
131 	dsl_dataset_name(os->os_dsl_dataset, buf);
132 }
133 
134 uint64_t
135 dmu_objset_id(objset_t *os)
136 {
137 	dsl_dataset_t *ds = os->os_dsl_dataset;
138 
139 	return (ds ? ds->ds_object : 0);
140 }
141 
142 zfs_sync_type_t
143 dmu_objset_syncprop(objset_t *os)
144 {
145 	return (os->os_sync);
146 }
147 
148 zfs_logbias_op_t
149 dmu_objset_logbias(objset_t *os)
150 {
151 	return (os->os_logbias);
152 }
153 
154 static void
155 checksum_changed_cb(void *arg, uint64_t newval)
156 {
157 	objset_t *os = arg;
158 
159 	/*
160 	 * Inheritance should have been done by now.
161 	 */
162 	ASSERT(newval != ZIO_CHECKSUM_INHERIT);
163 
164 	os->os_checksum = zio_checksum_select(newval, ZIO_CHECKSUM_ON_VALUE);
165 }
166 
167 static void
168 compression_changed_cb(void *arg, uint64_t newval)
169 {
170 	objset_t *os = arg;
171 
172 	/*
173 	 * Inheritance and range checking should have been done by now.
174 	 */
175 	ASSERT(newval != ZIO_COMPRESS_INHERIT);
176 
177 	os->os_compress = zio_compress_select(os->os_spa, newval,
178 	    ZIO_COMPRESS_ON);
179 }
180 
181 static void
182 copies_changed_cb(void *arg, uint64_t newval)
183 {
184 	objset_t *os = arg;
185 
186 	/*
187 	 * Inheritance and range checking should have been done by now.
188 	 */
189 	ASSERT(newval > 0);
190 	ASSERT(newval <= spa_max_replication(os->os_spa));
191 
192 	os->os_copies = newval;
193 }
194 
195 static void
196 dedup_changed_cb(void *arg, uint64_t newval)
197 {
198 	objset_t *os = arg;
199 	spa_t *spa = os->os_spa;
200 	enum zio_checksum checksum;
201 
202 	/*
203 	 * Inheritance should have been done by now.
204 	 */
205 	ASSERT(newval != ZIO_CHECKSUM_INHERIT);
206 
207 	checksum = zio_checksum_dedup_select(spa, newval, ZIO_CHECKSUM_OFF);
208 
209 	os->os_dedup_checksum = checksum & ZIO_CHECKSUM_MASK;
210 	os->os_dedup_verify = !!(checksum & ZIO_CHECKSUM_VERIFY);
211 }
212 
213 static void
214 primary_cache_changed_cb(void *arg, uint64_t newval)
215 {
216 	objset_t *os = arg;
217 
218 	/*
219 	 * Inheritance and range checking should have been done by now.
220 	 */
221 	ASSERT(newval == ZFS_CACHE_ALL || newval == ZFS_CACHE_NONE ||
222 	    newval == ZFS_CACHE_METADATA);
223 
224 	os->os_primary_cache = newval;
225 }
226 
227 static void
228 secondary_cache_changed_cb(void *arg, uint64_t newval)
229 {
230 	objset_t *os = arg;
231 
232 	/*
233 	 * Inheritance and range checking should have been done by now.
234 	 */
235 	ASSERT(newval == ZFS_CACHE_ALL || newval == ZFS_CACHE_NONE ||
236 	    newval == ZFS_CACHE_METADATA);
237 
238 	os->os_secondary_cache = newval;
239 }
240 
241 static void
242 sync_changed_cb(void *arg, uint64_t newval)
243 {
244 	objset_t *os = arg;
245 
246 	/*
247 	 * Inheritance and range checking should have been done by now.
248 	 */
249 	ASSERT(newval == ZFS_SYNC_STANDARD || newval == ZFS_SYNC_ALWAYS ||
250 	    newval == ZFS_SYNC_DISABLED);
251 
252 	os->os_sync = newval;
253 	if (os->os_zil)
254 		zil_set_sync(os->os_zil, newval);
255 }
256 
257 static void
258 redundant_metadata_changed_cb(void *arg, uint64_t newval)
259 {
260 	objset_t *os = arg;
261 
262 	/*
263 	 * Inheritance and range checking should have been done by now.
264 	 */
265 	ASSERT(newval == ZFS_REDUNDANT_METADATA_ALL ||
266 	    newval == ZFS_REDUNDANT_METADATA_MOST);
267 
268 	os->os_redundant_metadata = newval;
269 }
270 
271 static void
272 logbias_changed_cb(void *arg, uint64_t newval)
273 {
274 	objset_t *os = arg;
275 
276 	ASSERT(newval == ZFS_LOGBIAS_LATENCY ||
277 	    newval == ZFS_LOGBIAS_THROUGHPUT);
278 	os->os_logbias = newval;
279 	if (os->os_zil)
280 		zil_set_logbias(os->os_zil, newval);
281 }
282 
283 static void
284 recordsize_changed_cb(void *arg, uint64_t newval)
285 {
286 	objset_t *os = arg;
287 
288 	os->os_recordsize = newval;
289 }
290 
291 void
292 dmu_objset_byteswap(void *buf, size_t size)
293 {
294 	objset_phys_t *osp = buf;
295 
296 	ASSERT(size == OBJSET_OLD_PHYS_SIZE || size == sizeof (objset_phys_t));
297 	dnode_byteswap(&osp->os_meta_dnode);
298 	byteswap_uint64_array(&osp->os_zil_header, sizeof (zil_header_t));
299 	osp->os_type = BSWAP_64(osp->os_type);
300 	osp->os_flags = BSWAP_64(osp->os_flags);
301 	if (size == sizeof (objset_phys_t)) {
302 		dnode_byteswap(&osp->os_userused_dnode);
303 		dnode_byteswap(&osp->os_groupused_dnode);
304 	}
305 }
306 
307 /*
308  * The hash is a CRC-based hash of the objset_t pointer and the object number.
309  */
310 static uint64_t
311 dnode_hash(const objset_t *os, uint64_t obj)
312 {
313 	uintptr_t osv = (uintptr_t)os;
314 	uint64_t crc = -1ULL;
315 
316 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
317 	/*
318 	 * The low 6 bits of the pointer don't have much entropy, because
319 	 * the objset_t is larger than 2^6 bytes long.
320 	 */
321 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
322 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
323 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
324 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 16)) & 0xFF];
325 
326 	crc ^= (osv>>14) ^ (obj>>24);
327 
328 	return (crc);
329 }
330 
331 unsigned int
332 dnode_multilist_index_func(multilist_t *ml, void *obj)
333 {
334 	dnode_t *dn = obj;
335 	return (dnode_hash(dn->dn_objset, dn->dn_object) %
336 	    multilist_get_num_sublists(ml));
337 }
338 
339 /*
340  * Instantiates the objset_t in-memory structure corresponding to the
341  * objset_phys_t that's pointed to by the specified blkptr_t.
342  */
343 int
344 dmu_objset_open_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
345     objset_t **osp)
346 {
347 	objset_t *os;
348 	int i, err;
349 
350 	ASSERT(ds == NULL || MUTEX_HELD(&ds->ds_opening_lock));
351 
352 	/*
353 	 * The $ORIGIN dataset (if it exists) doesn't have an associated
354 	 * objset, so there's no reason to open it. The $ORIGIN dataset
355 	 * will not exist on pools older than SPA_VERSION_ORIGIN.
356 	 */
357 	if (ds != NULL && spa_get_dsl(spa) != NULL &&
358 	    spa_get_dsl(spa)->dp_origin_snap != NULL) {
359 		ASSERT3P(ds->ds_dir, !=,
360 		    spa_get_dsl(spa)->dp_origin_snap->ds_dir);
361 	}
362 
363 	os = kmem_zalloc(sizeof (objset_t), KM_SLEEP);
364 	os->os_dsl_dataset = ds;
365 	os->os_spa = spa;
366 	os->os_rootbp = bp;
367 	if (!BP_IS_HOLE(os->os_rootbp)) {
368 		arc_flags_t aflags = ARC_FLAG_WAIT;
369 		zbookmark_phys_t zb;
370 		SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
371 		    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
372 
373 		if (DMU_OS_IS_L2CACHEABLE(os))
374 			aflags |= ARC_FLAG_L2CACHE;
375 
376 		dprintf_bp(os->os_rootbp, "reading %s", "");
377 		err = arc_read(NULL, spa, os->os_rootbp,
378 		    arc_getbuf_func, &os->os_phys_buf,
379 		    ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL, &aflags, &zb);
380 		if (err != 0) {
381 			kmem_free(os, sizeof (objset_t));
382 			/* convert checksum errors into IO errors */
383 			if (err == ECKSUM)
384 				err = SET_ERROR(EIO);
385 			return (err);
386 		}
387 
388 		/* Increase the blocksize if we are permitted. */
389 		if (spa_version(spa) >= SPA_VERSION_USERSPACE &&
390 		    arc_buf_size(os->os_phys_buf) < sizeof (objset_phys_t)) {
391 			arc_buf_t *buf = arc_alloc_buf(spa, &os->os_phys_buf,
392 			    ARC_BUFC_METADATA, sizeof (objset_phys_t));
393 			bzero(buf->b_data, sizeof (objset_phys_t));
394 			bcopy(os->os_phys_buf->b_data, buf->b_data,
395 			    arc_buf_size(os->os_phys_buf));
396 			arc_buf_destroy(os->os_phys_buf, &os->os_phys_buf);
397 			os->os_phys_buf = buf;
398 		}
399 
400 		os->os_phys = os->os_phys_buf->b_data;
401 		os->os_flags = os->os_phys->os_flags;
402 	} else {
403 		int size = spa_version(spa) >= SPA_VERSION_USERSPACE ?
404 		    sizeof (objset_phys_t) : OBJSET_OLD_PHYS_SIZE;
405 		os->os_phys_buf = arc_alloc_buf(spa, &os->os_phys_buf,
406 		    ARC_BUFC_METADATA, size);
407 		os->os_phys = os->os_phys_buf->b_data;
408 		bzero(os->os_phys, size);
409 	}
410 
411 	/*
412 	 * Note: the changed_cb will be called once before the register
413 	 * func returns, thus changing the checksum/compression from the
414 	 * default (fletcher2/off).  Snapshots don't need to know about
415 	 * checksum/compression/copies.
416 	 */
417 	if (ds != NULL) {
418 		boolean_t needlock = B_FALSE;
419 
420 		/*
421 		 * Note: it's valid to open the objset if the dataset is
422 		 * long-held, in which case the pool_config lock will not
423 		 * be held.
424 		 */
425 		if (!dsl_pool_config_held(dmu_objset_pool(os))) {
426 			needlock = B_TRUE;
427 			dsl_pool_config_enter(dmu_objset_pool(os), FTAG);
428 		}
429 		err = dsl_prop_register(ds,
430 		    zfs_prop_to_name(ZFS_PROP_PRIMARYCACHE),
431 		    primary_cache_changed_cb, os);
432 		if (err == 0) {
433 			err = dsl_prop_register(ds,
434 			    zfs_prop_to_name(ZFS_PROP_SECONDARYCACHE),
435 			    secondary_cache_changed_cb, os);
436 		}
437 		if (!ds->ds_is_snapshot) {
438 			if (err == 0) {
439 				err = dsl_prop_register(ds,
440 				    zfs_prop_to_name(ZFS_PROP_CHECKSUM),
441 				    checksum_changed_cb, os);
442 			}
443 			if (err == 0) {
444 				err = dsl_prop_register(ds,
445 				    zfs_prop_to_name(ZFS_PROP_COMPRESSION),
446 				    compression_changed_cb, os);
447 			}
448 			if (err == 0) {
449 				err = dsl_prop_register(ds,
450 				    zfs_prop_to_name(ZFS_PROP_COPIES),
451 				    copies_changed_cb, os);
452 			}
453 			if (err == 0) {
454 				err = dsl_prop_register(ds,
455 				    zfs_prop_to_name(ZFS_PROP_DEDUP),
456 				    dedup_changed_cb, os);
457 			}
458 			if (err == 0) {
459 				err = dsl_prop_register(ds,
460 				    zfs_prop_to_name(ZFS_PROP_LOGBIAS),
461 				    logbias_changed_cb, os);
462 			}
463 			if (err == 0) {
464 				err = dsl_prop_register(ds,
465 				    zfs_prop_to_name(ZFS_PROP_SYNC),
466 				    sync_changed_cb, os);
467 			}
468 			if (err == 0) {
469 				err = dsl_prop_register(ds,
470 				    zfs_prop_to_name(
471 				    ZFS_PROP_REDUNDANT_METADATA),
472 				    redundant_metadata_changed_cb, os);
473 			}
474 			if (err == 0) {
475 				err = dsl_prop_register(ds,
476 				    zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
477 				    recordsize_changed_cb, os);
478 			}
479 		}
480 		if (needlock)
481 			dsl_pool_config_exit(dmu_objset_pool(os), FTAG);
482 		if (err != 0) {
483 			arc_buf_destroy(os->os_phys_buf, &os->os_phys_buf);
484 			kmem_free(os, sizeof (objset_t));
485 			return (err);
486 		}
487 	} else {
488 		/* It's the meta-objset. */
489 		os->os_checksum = ZIO_CHECKSUM_FLETCHER_4;
490 		os->os_compress = ZIO_COMPRESS_ON;
491 		os->os_copies = spa_max_replication(spa);
492 		os->os_dedup_checksum = ZIO_CHECKSUM_OFF;
493 		os->os_dedup_verify = B_FALSE;
494 		os->os_logbias = ZFS_LOGBIAS_LATENCY;
495 		os->os_sync = ZFS_SYNC_STANDARD;
496 		os->os_primary_cache = ZFS_CACHE_ALL;
497 		os->os_secondary_cache = ZFS_CACHE_ALL;
498 	}
499 	/*
500 	 * These properties will be filled in by the logic in zfs_get_zplprop()
501 	 * when they are queried for the first time.
502 	 */
503 	os->os_version = OBJSET_PROP_UNINITIALIZED;
504 	os->os_normalization = OBJSET_PROP_UNINITIALIZED;
505 	os->os_utf8only = OBJSET_PROP_UNINITIALIZED;
506 	os->os_casesensitivity = OBJSET_PROP_UNINITIALIZED;
507 
508 	if (ds == NULL || !ds->ds_is_snapshot)
509 		os->os_zil_header = os->os_phys->os_zil_header;
510 	os->os_zil = zil_alloc(os, &os->os_zil_header);
511 
512 	for (i = 0; i < TXG_SIZE; i++) {
513 		os->os_dirty_dnodes[i] = multilist_create(sizeof (dnode_t),
514 		    offsetof(dnode_t, dn_dirty_link[i]),
515 		    dnode_multilist_index_func);
516 	}
517 	list_create(&os->os_dnodes, sizeof (dnode_t),
518 	    offsetof(dnode_t, dn_link));
519 	list_create(&os->os_downgraded_dbufs, sizeof (dmu_buf_impl_t),
520 	    offsetof(dmu_buf_impl_t, db_link));
521 
522 	mutex_init(&os->os_lock, NULL, MUTEX_DEFAULT, NULL);
523 	mutex_init(&os->os_userused_lock, NULL, MUTEX_DEFAULT, NULL);
524 	mutex_init(&os->os_obj_lock, NULL, MUTEX_DEFAULT, NULL);
525 	mutex_init(&os->os_user_ptr_lock, NULL, MUTEX_DEFAULT, NULL);
526 
527 	dnode_special_open(os, &os->os_phys->os_meta_dnode,
528 	    DMU_META_DNODE_OBJECT, &os->os_meta_dnode);
529 	if (arc_buf_size(os->os_phys_buf) >= sizeof (objset_phys_t)) {
530 		dnode_special_open(os, &os->os_phys->os_userused_dnode,
531 		    DMU_USERUSED_OBJECT, &os->os_userused_dnode);
532 		dnode_special_open(os, &os->os_phys->os_groupused_dnode,
533 		    DMU_GROUPUSED_OBJECT, &os->os_groupused_dnode);
534 	}
535 
536 	*osp = os;
537 	return (0);
538 }
539 
540 int
541 dmu_objset_from_ds(dsl_dataset_t *ds, objset_t **osp)
542 {
543 	int err = 0;
544 
545 	/*
546 	 * We shouldn't be doing anything with dsl_dataset_t's unless the
547 	 * pool_config lock is held, or the dataset is long-held.
548 	 */
549 	ASSERT(dsl_pool_config_held(ds->ds_dir->dd_pool) ||
550 	    dsl_dataset_long_held(ds));
551 
552 	mutex_enter(&ds->ds_opening_lock);
553 	if (ds->ds_objset == NULL) {
554 		objset_t *os;
555 		rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
556 		err = dmu_objset_open_impl(dsl_dataset_get_spa(ds),
557 		    ds, dsl_dataset_get_blkptr(ds), &os);
558 		rrw_exit(&ds->ds_bp_rwlock, FTAG);
559 
560 		if (err == 0) {
561 			mutex_enter(&ds->ds_lock);
562 			ASSERT(ds->ds_objset == NULL);
563 			ds->ds_objset = os;
564 			mutex_exit(&ds->ds_lock);
565 		}
566 	}
567 	*osp = ds->ds_objset;
568 	mutex_exit(&ds->ds_opening_lock);
569 	return (err);
570 }
571 
572 /*
573  * Holds the pool while the objset is held.  Therefore only one objset
574  * can be held at a time.
575  */
576 int
577 dmu_objset_hold(const char *name, void *tag, objset_t **osp)
578 {
579 	dsl_pool_t *dp;
580 	dsl_dataset_t *ds;
581 	int err;
582 
583 	err = dsl_pool_hold(name, tag, &dp);
584 	if (err != 0)
585 		return (err);
586 	err = dsl_dataset_hold(dp, name, tag, &ds);
587 	if (err != 0) {
588 		dsl_pool_rele(dp, tag);
589 		return (err);
590 	}
591 
592 	err = dmu_objset_from_ds(ds, osp);
593 	if (err != 0) {
594 		dsl_dataset_rele(ds, tag);
595 		dsl_pool_rele(dp, tag);
596 	}
597 
598 	return (err);
599 }
600 
601 static int
602 dmu_objset_own_impl(dsl_dataset_t *ds, dmu_objset_type_t type,
603     boolean_t readonly, void *tag, objset_t **osp)
604 {
605 	int err;
606 
607 	err = dmu_objset_from_ds(ds, osp);
608 	if (err != 0) {
609 		dsl_dataset_disown(ds, tag);
610 	} else if (type != DMU_OST_ANY && type != (*osp)->os_phys->os_type) {
611 		dsl_dataset_disown(ds, tag);
612 		return (SET_ERROR(EINVAL));
613 	} else if (!readonly && dsl_dataset_is_snapshot(ds)) {
614 		dsl_dataset_disown(ds, tag);
615 		return (SET_ERROR(EROFS));
616 	}
617 	return (err);
618 }
619 
620 /*
621  * dsl_pool must not be held when this is called.
622  * Upon successful return, there will be a longhold on the dataset,
623  * and the dsl_pool will not be held.
624  */
625 int
626 dmu_objset_own(const char *name, dmu_objset_type_t type,
627     boolean_t readonly, void *tag, objset_t **osp)
628 {
629 	dsl_pool_t *dp;
630 	dsl_dataset_t *ds;
631 	int err;
632 
633 	err = dsl_pool_hold(name, FTAG, &dp);
634 	if (err != 0)
635 		return (err);
636 	err = dsl_dataset_own(dp, name, tag, &ds);
637 	if (err != 0) {
638 		dsl_pool_rele(dp, FTAG);
639 		return (err);
640 	}
641 	err = dmu_objset_own_impl(ds, type, readonly, tag, osp);
642 	dsl_pool_rele(dp, FTAG);
643 
644 	return (err);
645 }
646 
647 int
648 dmu_objset_own_obj(dsl_pool_t *dp, uint64_t obj, dmu_objset_type_t type,
649     boolean_t readonly, void *tag, objset_t **osp)
650 {
651 	dsl_dataset_t *ds;
652 	int err;
653 
654 	err = dsl_dataset_own_obj(dp, obj, tag, &ds);
655 	if (err != 0)
656 		return (err);
657 
658 	return (dmu_objset_own_impl(ds, type, readonly, tag, osp));
659 }
660 
661 void
662 dmu_objset_rele(objset_t *os, void *tag)
663 {
664 	dsl_pool_t *dp = dmu_objset_pool(os);
665 	dsl_dataset_rele(os->os_dsl_dataset, tag);
666 	dsl_pool_rele(dp, tag);
667 }
668 
669 /*
670  * When we are called, os MUST refer to an objset associated with a dataset
671  * that is owned by 'tag'; that is, is held and long held by 'tag' and ds_owner
672  * == tag.  We will then release and reacquire ownership of the dataset while
673  * holding the pool config_rwlock to avoid intervening namespace or ownership
674  * changes may occur.
675  *
676  * This exists solely to accommodate zfs_ioc_userspace_upgrade()'s desire to
677  * release the hold on its dataset and acquire a new one on the dataset of the
678  * same name so that it can be partially torn down and reconstructed.
679  */
680 void
681 dmu_objset_refresh_ownership(dsl_dataset_t *ds, dsl_dataset_t **newds,
682     void *tag)
683 {
684 	dsl_pool_t *dp;
685 	char name[ZFS_MAX_DATASET_NAME_LEN];
686 
687 	VERIFY3P(ds, !=, NULL);
688 	VERIFY3P(ds->ds_owner, ==, tag);
689 	VERIFY(dsl_dataset_long_held(ds));
690 
691 	dsl_dataset_name(ds, name);
692 	dp = ds->ds_dir->dd_pool;
693 	dsl_pool_config_enter(dp, FTAG);
694 	dsl_dataset_disown(ds, tag);
695 	VERIFY0(dsl_dataset_own(dp, name, tag, newds));
696 	dsl_pool_config_exit(dp, FTAG);
697 }
698 
699 void
700 dmu_objset_disown(objset_t *os, void *tag)
701 {
702 	dsl_dataset_disown(os->os_dsl_dataset, tag);
703 }
704 
705 void
706 dmu_objset_evict_dbufs(objset_t *os)
707 {
708 	dnode_t dn_marker;
709 	dnode_t *dn;
710 
711 	mutex_enter(&os->os_lock);
712 	dn = list_head(&os->os_dnodes);
713 	while (dn != NULL) {
714 		/*
715 		 * Skip dnodes without holds.  We have to do this dance
716 		 * because dnode_add_ref() only works if there is already a
717 		 * hold.  If the dnode has no holds, then it has no dbufs.
718 		 */
719 		if (dnode_add_ref(dn, FTAG)) {
720 			list_insert_after(&os->os_dnodes, dn, &dn_marker);
721 			mutex_exit(&os->os_lock);
722 
723 			dnode_evict_dbufs(dn);
724 			dnode_rele(dn, FTAG);
725 
726 			mutex_enter(&os->os_lock);
727 			dn = list_next(&os->os_dnodes, &dn_marker);
728 			list_remove(&os->os_dnodes, &dn_marker);
729 		} else {
730 			dn = list_next(&os->os_dnodes, dn);
731 		}
732 	}
733 	mutex_exit(&os->os_lock);
734 
735 	if (DMU_USERUSED_DNODE(os) != NULL) {
736 		dnode_evict_dbufs(DMU_GROUPUSED_DNODE(os));
737 		dnode_evict_dbufs(DMU_USERUSED_DNODE(os));
738 	}
739 	dnode_evict_dbufs(DMU_META_DNODE(os));
740 }
741 
742 /*
743  * Objset eviction processing is split into into two pieces.
744  * The first marks the objset as evicting, evicts any dbufs that
745  * have a refcount of zero, and then queues up the objset for the
746  * second phase of eviction.  Once os->os_dnodes has been cleared by
747  * dnode_buf_pageout()->dnode_destroy(), the second phase is executed.
748  * The second phase closes the special dnodes, dequeues the objset from
749  * the list of those undergoing eviction, and finally frees the objset.
750  *
751  * NOTE: Due to asynchronous eviction processing (invocation of
752  *       dnode_buf_pageout()), it is possible for the meta dnode for the
753  *       objset to have no holds even though os->os_dnodes is not empty.
754  */
755 void
756 dmu_objset_evict(objset_t *os)
757 {
758 	dsl_dataset_t *ds = os->os_dsl_dataset;
759 
760 	for (int t = 0; t < TXG_SIZE; t++)
761 		ASSERT(!dmu_objset_is_dirty(os, t));
762 
763 	if (ds)
764 		dsl_prop_unregister_all(ds, os);
765 
766 	if (os->os_sa)
767 		sa_tear_down(os);
768 
769 	dmu_objset_evict_dbufs(os);
770 
771 	mutex_enter(&os->os_lock);
772 	spa_evicting_os_register(os->os_spa, os);
773 	if (list_is_empty(&os->os_dnodes)) {
774 		mutex_exit(&os->os_lock);
775 		dmu_objset_evict_done(os);
776 	} else {
777 		mutex_exit(&os->os_lock);
778 	}
779 }
780 
781 void
782 dmu_objset_evict_done(objset_t *os)
783 {
784 	ASSERT3P(list_head(&os->os_dnodes), ==, NULL);
785 
786 	dnode_special_close(&os->os_meta_dnode);
787 	if (DMU_USERUSED_DNODE(os)) {
788 		dnode_special_close(&os->os_userused_dnode);
789 		dnode_special_close(&os->os_groupused_dnode);
790 	}
791 	zil_free(os->os_zil);
792 
793 	arc_buf_destroy(os->os_phys_buf, &os->os_phys_buf);
794 
795 	/*
796 	 * This is a barrier to prevent the objset from going away in
797 	 * dnode_move() until we can safely ensure that the objset is still in
798 	 * use. We consider the objset valid before the barrier and invalid
799 	 * after the barrier.
800 	 */
801 	rw_enter(&os_lock, RW_READER);
802 	rw_exit(&os_lock);
803 
804 	mutex_destroy(&os->os_lock);
805 	mutex_destroy(&os->os_userused_lock);
806 	mutex_destroy(&os->os_obj_lock);
807 	mutex_destroy(&os->os_user_ptr_lock);
808 	for (int i = 0; i < TXG_SIZE; i++) {
809 		multilist_destroy(os->os_dirty_dnodes[i]);
810 	}
811 	spa_evicting_os_deregister(os->os_spa, os);
812 	kmem_free(os, sizeof (objset_t));
813 }
814 
815 timestruc_t
816 dmu_objset_snap_cmtime(objset_t *os)
817 {
818 	return (dsl_dir_snap_cmtime(os->os_dsl_dataset->ds_dir));
819 }
820 
821 /* called from dsl for meta-objset */
822 objset_t *
823 dmu_objset_create_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
824     dmu_objset_type_t type, dmu_tx_t *tx)
825 {
826 	objset_t *os;
827 	dnode_t *mdn;
828 
829 	ASSERT(dmu_tx_is_syncing(tx));
830 
831 	if (ds != NULL)
832 		VERIFY0(dmu_objset_from_ds(ds, &os));
833 	else
834 		VERIFY0(dmu_objset_open_impl(spa, NULL, bp, &os));
835 
836 	mdn = DMU_META_DNODE(os);
837 
838 	dnode_allocate(mdn, DMU_OT_DNODE, 1 << DNODE_BLOCK_SHIFT,
839 	    DN_MAX_INDBLKSHIFT, DMU_OT_NONE, 0, tx);
840 
841 	/*
842 	 * We don't want to have to increase the meta-dnode's nlevels
843 	 * later, because then we could do it in quescing context while
844 	 * we are also accessing it in open context.
845 	 *
846 	 * This precaution is not necessary for the MOS (ds == NULL),
847 	 * because the MOS is only updated in syncing context.
848 	 * This is most fortunate: the MOS is the only objset that
849 	 * needs to be synced multiple times as spa_sync() iterates
850 	 * to convergence, so minimizing its dn_nlevels matters.
851 	 */
852 	if (ds != NULL) {
853 		int levels = 1;
854 
855 		/*
856 		 * Determine the number of levels necessary for the meta-dnode
857 		 * to contain DN_MAX_OBJECT dnodes.  Note that in order to
858 		 * ensure that we do not overflow 64 bits, there has to be
859 		 * a nlevels that gives us a number of blocks > DN_MAX_OBJECT
860 		 * but < 2^64.  Therefore,
861 		 * (mdn->dn_indblkshift - SPA_BLKPTRSHIFT) (10) must be
862 		 * less than (64 - log2(DN_MAX_OBJECT)) (16).
863 		 */
864 		while ((uint64_t)mdn->dn_nblkptr <<
865 		    (mdn->dn_datablkshift - DNODE_SHIFT +
866 		    (levels - 1) * (mdn->dn_indblkshift - SPA_BLKPTRSHIFT)) <
867 		    DN_MAX_OBJECT)
868 			levels++;
869 
870 		mdn->dn_next_nlevels[tx->tx_txg & TXG_MASK] =
871 		    mdn->dn_nlevels = levels;
872 	}
873 
874 	ASSERT(type != DMU_OST_NONE);
875 	ASSERT(type != DMU_OST_ANY);
876 	ASSERT(type < DMU_OST_NUMTYPES);
877 	os->os_phys->os_type = type;
878 	if (dmu_objset_userused_enabled(os)) {
879 		os->os_phys->os_flags |= OBJSET_FLAG_USERACCOUNTING_COMPLETE;
880 		os->os_flags = os->os_phys->os_flags;
881 	}
882 
883 	dsl_dataset_dirty(ds, tx);
884 
885 	return (os);
886 }
887 
888 typedef struct dmu_objset_create_arg {
889 	const char *doca_name;
890 	cred_t *doca_cred;
891 	void (*doca_userfunc)(objset_t *os, void *arg,
892 	    cred_t *cr, dmu_tx_t *tx);
893 	void *doca_userarg;
894 	dmu_objset_type_t doca_type;
895 	uint64_t doca_flags;
896 } dmu_objset_create_arg_t;
897 
898 /*ARGSUSED*/
899 static int
900 dmu_objset_create_check(void *arg, dmu_tx_t *tx)
901 {
902 	dmu_objset_create_arg_t *doca = arg;
903 	dsl_pool_t *dp = dmu_tx_pool(tx);
904 	dsl_dir_t *pdd;
905 	const char *tail;
906 	int error;
907 
908 	if (strchr(doca->doca_name, '@') != NULL)
909 		return (SET_ERROR(EINVAL));
910 
911 	if (strlen(doca->doca_name) >= ZFS_MAX_DATASET_NAME_LEN)
912 		return (SET_ERROR(ENAMETOOLONG));
913 
914 	error = dsl_dir_hold(dp, doca->doca_name, FTAG, &pdd, &tail);
915 	if (error != 0)
916 		return (error);
917 	if (tail == NULL) {
918 		dsl_dir_rele(pdd, FTAG);
919 		return (SET_ERROR(EEXIST));
920 	}
921 	error = dsl_fs_ss_limit_check(pdd, 1, ZFS_PROP_FILESYSTEM_LIMIT, NULL,
922 	    doca->doca_cred);
923 	dsl_dir_rele(pdd, FTAG);
924 
925 	return (error);
926 }
927 
928 static void
929 dmu_objset_create_sync(void *arg, dmu_tx_t *tx)
930 {
931 	dmu_objset_create_arg_t *doca = arg;
932 	dsl_pool_t *dp = dmu_tx_pool(tx);
933 	dsl_dir_t *pdd;
934 	const char *tail;
935 	dsl_dataset_t *ds;
936 	uint64_t obj;
937 	blkptr_t *bp;
938 	objset_t *os;
939 
940 	VERIFY0(dsl_dir_hold(dp, doca->doca_name, FTAG, &pdd, &tail));
941 
942 	obj = dsl_dataset_create_sync(pdd, tail, NULL, doca->doca_flags,
943 	    doca->doca_cred, tx);
944 
945 	VERIFY0(dsl_dataset_hold_obj(pdd->dd_pool, obj, FTAG, &ds));
946 	rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
947 	bp = dsl_dataset_get_blkptr(ds);
948 	os = dmu_objset_create_impl(pdd->dd_pool->dp_spa,
949 	    ds, bp, doca->doca_type, tx);
950 	rrw_exit(&ds->ds_bp_rwlock, FTAG);
951 
952 	if (doca->doca_userfunc != NULL) {
953 		doca->doca_userfunc(os, doca->doca_userarg,
954 		    doca->doca_cred, tx);
955 	}
956 
957 	spa_history_log_internal_ds(ds, "create", tx, "");
958 	dsl_dataset_rele(ds, FTAG);
959 	dsl_dir_rele(pdd, FTAG);
960 }
961 
962 int
963 dmu_objset_create(const char *name, dmu_objset_type_t type, uint64_t flags,
964     void (*func)(objset_t *os, void *arg, cred_t *cr, dmu_tx_t *tx), void *arg)
965 {
966 	dmu_objset_create_arg_t doca;
967 
968 	doca.doca_name = name;
969 	doca.doca_cred = CRED();
970 	doca.doca_flags = flags;
971 	doca.doca_userfunc = func;
972 	doca.doca_userarg = arg;
973 	doca.doca_type = type;
974 
975 	return (dsl_sync_task(name,
976 	    dmu_objset_create_check, dmu_objset_create_sync, &doca,
977 	    5, ZFS_SPACE_CHECK_NORMAL));
978 }
979 
980 typedef struct dmu_objset_clone_arg {
981 	const char *doca_clone;
982 	const char *doca_origin;
983 	cred_t *doca_cred;
984 } dmu_objset_clone_arg_t;
985 
986 /*ARGSUSED*/
987 static int
988 dmu_objset_clone_check(void *arg, dmu_tx_t *tx)
989 {
990 	dmu_objset_clone_arg_t *doca = arg;
991 	dsl_dir_t *pdd;
992 	const char *tail;
993 	int error;
994 	dsl_dataset_t *origin;
995 	dsl_pool_t *dp = dmu_tx_pool(tx);
996 
997 	if (strchr(doca->doca_clone, '@') != NULL)
998 		return (SET_ERROR(EINVAL));
999 
1000 	if (strlen(doca->doca_clone) >= ZFS_MAX_DATASET_NAME_LEN)
1001 		return (SET_ERROR(ENAMETOOLONG));
1002 
1003 	error = dsl_dir_hold(dp, doca->doca_clone, FTAG, &pdd, &tail);
1004 	if (error != 0)
1005 		return (error);
1006 	if (tail == NULL) {
1007 		dsl_dir_rele(pdd, FTAG);
1008 		return (SET_ERROR(EEXIST));
1009 	}
1010 
1011 	error = dsl_fs_ss_limit_check(pdd, 1, ZFS_PROP_FILESYSTEM_LIMIT, NULL,
1012 	    doca->doca_cred);
1013 	if (error != 0) {
1014 		dsl_dir_rele(pdd, FTAG);
1015 		return (SET_ERROR(EDQUOT));
1016 	}
1017 	dsl_dir_rele(pdd, FTAG);
1018 
1019 	error = dsl_dataset_hold(dp, doca->doca_origin, FTAG, &origin);
1020 	if (error != 0)
1021 		return (error);
1022 
1023 	/* You can only clone snapshots, not the head datasets. */
1024 	if (!origin->ds_is_snapshot) {
1025 		dsl_dataset_rele(origin, FTAG);
1026 		return (SET_ERROR(EINVAL));
1027 	}
1028 	dsl_dataset_rele(origin, FTAG);
1029 
1030 	return (0);
1031 }
1032 
1033 static void
1034 dmu_objset_clone_sync(void *arg, dmu_tx_t *tx)
1035 {
1036 	dmu_objset_clone_arg_t *doca = arg;
1037 	dsl_pool_t *dp = dmu_tx_pool(tx);
1038 	dsl_dir_t *pdd;
1039 	const char *tail;
1040 	dsl_dataset_t *origin, *ds;
1041 	uint64_t obj;
1042 	char namebuf[ZFS_MAX_DATASET_NAME_LEN];
1043 
1044 	VERIFY0(dsl_dir_hold(dp, doca->doca_clone, FTAG, &pdd, &tail));
1045 	VERIFY0(dsl_dataset_hold(dp, doca->doca_origin, FTAG, &origin));
1046 
1047 	obj = dsl_dataset_create_sync(pdd, tail, origin, 0,
1048 	    doca->doca_cred, tx);
1049 
1050 	VERIFY0(dsl_dataset_hold_obj(pdd->dd_pool, obj, FTAG, &ds));
1051 	dsl_dataset_name(origin, namebuf);
1052 	spa_history_log_internal_ds(ds, "clone", tx,
1053 	    "origin=%s (%llu)", namebuf, origin->ds_object);
1054 	dsl_dataset_rele(ds, FTAG);
1055 	dsl_dataset_rele(origin, FTAG);
1056 	dsl_dir_rele(pdd, FTAG);
1057 }
1058 
1059 int
1060 dmu_objset_clone(const char *clone, const char *origin)
1061 {
1062 	dmu_objset_clone_arg_t doca;
1063 
1064 	doca.doca_clone = clone;
1065 	doca.doca_origin = origin;
1066 	doca.doca_cred = CRED();
1067 
1068 	return (dsl_sync_task(clone,
1069 	    dmu_objset_clone_check, dmu_objset_clone_sync, &doca,
1070 	    5, ZFS_SPACE_CHECK_NORMAL));
1071 }
1072 
1073 static int
1074 dmu_objset_remap_indirects_impl(objset_t *os, uint64_t last_removed_txg)
1075 {
1076 	int error = 0;
1077 	uint64_t object = 0;
1078 	while ((error = dmu_object_next(os, &object, B_FALSE, 0)) == 0) {
1079 		error = dmu_object_remap_indirects(os, object,
1080 		    last_removed_txg);
1081 		/*
1082 		 * If the ZPL removed the object before we managed to dnode_hold
1083 		 * it, we would get an ENOENT. If the ZPL declares its intent
1084 		 * to remove the object (dnode_free) before we manage to
1085 		 * dnode_hold it, we would get an EEXIST. In either case, we
1086 		 * want to continue remapping the other objects in the objset;
1087 		 * in all other cases, we want to break early.
1088 		 */
1089 		if (error != 0 && error != ENOENT && error != EEXIST) {
1090 			break;
1091 		}
1092 	}
1093 	if (error == ESRCH) {
1094 		error = 0;
1095 	}
1096 	return (error);
1097 }
1098 
1099 int
1100 dmu_objset_remap_indirects(const char *fsname)
1101 {
1102 	int error = 0;
1103 	objset_t *os = NULL;
1104 	uint64_t last_removed_txg;
1105 	uint64_t remap_start_txg;
1106 	dsl_dir_t *dd;
1107 
1108 	error = dmu_objset_hold(fsname, FTAG, &os);
1109 	if (error != 0) {
1110 		return (error);
1111 	}
1112 	dd = dmu_objset_ds(os)->ds_dir;
1113 
1114 	if (!spa_feature_is_enabled(dmu_objset_spa(os),
1115 	    SPA_FEATURE_OBSOLETE_COUNTS)) {
1116 		dmu_objset_rele(os, FTAG);
1117 		return (SET_ERROR(ENOTSUP));
1118 	}
1119 
1120 	if (dsl_dataset_is_snapshot(dmu_objset_ds(os))) {
1121 		dmu_objset_rele(os, FTAG);
1122 		return (SET_ERROR(EINVAL));
1123 	}
1124 
1125 	/*
1126 	 * If there has not been a removal, we're done.
1127 	 */
1128 	last_removed_txg = spa_get_last_removal_txg(dmu_objset_spa(os));
1129 	if (last_removed_txg == -1ULL) {
1130 		dmu_objset_rele(os, FTAG);
1131 		return (0);
1132 	}
1133 
1134 	/*
1135 	 * If we have remapped since the last removal, we're done.
1136 	 */
1137 	if (dsl_dir_is_zapified(dd)) {
1138 		uint64_t last_remap_txg;
1139 		if (zap_lookup(spa_meta_objset(dmu_objset_spa(os)),
1140 		    dd->dd_object, DD_FIELD_LAST_REMAP_TXG,
1141 		    sizeof (last_remap_txg), 1, &last_remap_txg) == 0 &&
1142 		    last_remap_txg > last_removed_txg) {
1143 			dmu_objset_rele(os, FTAG);
1144 			return (0);
1145 		}
1146 	}
1147 
1148 	dsl_dataset_long_hold(dmu_objset_ds(os), FTAG);
1149 	dsl_pool_rele(dmu_objset_pool(os), FTAG);
1150 
1151 	remap_start_txg = spa_last_synced_txg(dmu_objset_spa(os));
1152 	error = dmu_objset_remap_indirects_impl(os, last_removed_txg);
1153 	if (error == 0) {
1154 		/*
1155 		 * We update the last_remap_txg to be the start txg so that
1156 		 * we can guarantee that every block older than last_remap_txg
1157 		 * that can be remapped has been remapped.
1158 		 */
1159 		error = dsl_dir_update_last_remap_txg(dd, remap_start_txg);
1160 	}
1161 
1162 	dsl_dataset_long_rele(dmu_objset_ds(os), FTAG);
1163 	dsl_dataset_rele(dmu_objset_ds(os), FTAG);
1164 
1165 	return (error);
1166 }
1167 
1168 int
1169 dmu_objset_snapshot_one(const char *fsname, const char *snapname)
1170 {
1171 	int err;
1172 	char *longsnap = kmem_asprintf("%s@%s", fsname, snapname);
1173 	nvlist_t *snaps = fnvlist_alloc();
1174 
1175 	fnvlist_add_boolean(snaps, longsnap);
1176 	strfree(longsnap);
1177 	err = dsl_dataset_snapshot(snaps, NULL, NULL);
1178 	fnvlist_free(snaps);
1179 	return (err);
1180 }
1181 
1182 static void
1183 dmu_objset_sync_dnodes(multilist_sublist_t *list, dmu_tx_t *tx)
1184 {
1185 	dnode_t *dn;
1186 
1187 	while ((dn = multilist_sublist_head(list)) != NULL) {
1188 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
1189 		ASSERT(dn->dn_dbuf->db_data_pending);
1190 		/*
1191 		 * Initialize dn_zio outside dnode_sync() because the
1192 		 * meta-dnode needs to set it ouside dnode_sync().
1193 		 */
1194 		dn->dn_zio = dn->dn_dbuf->db_data_pending->dr_zio;
1195 		ASSERT(dn->dn_zio);
1196 
1197 		ASSERT3U(dn->dn_nlevels, <=, DN_MAX_LEVELS);
1198 		multilist_sublist_remove(list, dn);
1199 
1200 		multilist_t *newlist = dn->dn_objset->os_synced_dnodes;
1201 		if (newlist != NULL) {
1202 			(void) dnode_add_ref(dn, newlist);
1203 			multilist_insert(newlist, dn);
1204 		}
1205 
1206 		dnode_sync(dn, tx);
1207 	}
1208 }
1209 
1210 /* ARGSUSED */
1211 static void
1212 dmu_objset_write_ready(zio_t *zio, arc_buf_t *abuf, void *arg)
1213 {
1214 	blkptr_t *bp = zio->io_bp;
1215 	objset_t *os = arg;
1216 	dnode_phys_t *dnp = &os->os_phys->os_meta_dnode;
1217 
1218 	ASSERT(!BP_IS_EMBEDDED(bp));
1219 	ASSERT3U(BP_GET_TYPE(bp), ==, DMU_OT_OBJSET);
1220 	ASSERT0(BP_GET_LEVEL(bp));
1221 
1222 	/*
1223 	 * Update rootbp fill count: it should be the number of objects
1224 	 * allocated in the object set (not counting the "special"
1225 	 * objects that are stored in the objset_phys_t -- the meta
1226 	 * dnode and user/group accounting objects).
1227 	 */
1228 	bp->blk_fill = 0;
1229 	for (int i = 0; i < dnp->dn_nblkptr; i++)
1230 		bp->blk_fill += BP_GET_FILL(&dnp->dn_blkptr[i]);
1231 	if (os->os_dsl_dataset != NULL)
1232 		rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_WRITER, FTAG);
1233 	*os->os_rootbp = *bp;
1234 	if (os->os_dsl_dataset != NULL)
1235 		rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1236 }
1237 
1238 /* ARGSUSED */
1239 static void
1240 dmu_objset_write_done(zio_t *zio, arc_buf_t *abuf, void *arg)
1241 {
1242 	blkptr_t *bp = zio->io_bp;
1243 	blkptr_t *bp_orig = &zio->io_bp_orig;
1244 	objset_t *os = arg;
1245 
1246 	if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
1247 		ASSERT(BP_EQUAL(bp, bp_orig));
1248 	} else {
1249 		dsl_dataset_t *ds = os->os_dsl_dataset;
1250 		dmu_tx_t *tx = os->os_synctx;
1251 
1252 		(void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
1253 		dsl_dataset_block_born(ds, bp, tx);
1254 	}
1255 	kmem_free(bp, sizeof (*bp));
1256 }
1257 
1258 typedef struct sync_dnodes_arg {
1259 	multilist_t *sda_list;
1260 	int sda_sublist_idx;
1261 	multilist_t *sda_newlist;
1262 	dmu_tx_t *sda_tx;
1263 } sync_dnodes_arg_t;
1264 
1265 static void
1266 sync_dnodes_task(void *arg)
1267 {
1268 	sync_dnodes_arg_t *sda = arg;
1269 
1270 	multilist_sublist_t *ms =
1271 	    multilist_sublist_lock(sda->sda_list, sda->sda_sublist_idx);
1272 
1273 	dmu_objset_sync_dnodes(ms, sda->sda_tx);
1274 
1275 	multilist_sublist_unlock(ms);
1276 
1277 	kmem_free(sda, sizeof (*sda));
1278 }
1279 
1280 
1281 /* called from dsl */
1282 void
1283 dmu_objset_sync(objset_t *os, zio_t *pio, dmu_tx_t *tx)
1284 {
1285 	int txgoff;
1286 	zbookmark_phys_t zb;
1287 	zio_prop_t zp;
1288 	zio_t *zio;
1289 	list_t *list;
1290 	dbuf_dirty_record_t *dr;
1291 	blkptr_t *blkptr_copy = kmem_alloc(sizeof (*os->os_rootbp), KM_SLEEP);
1292 	*blkptr_copy = *os->os_rootbp;
1293 
1294 	dprintf_ds(os->os_dsl_dataset, "txg=%llu\n", tx->tx_txg);
1295 
1296 	ASSERT(dmu_tx_is_syncing(tx));
1297 	/* XXX the write_done callback should really give us the tx... */
1298 	os->os_synctx = tx;
1299 
1300 	if (os->os_dsl_dataset == NULL) {
1301 		/*
1302 		 * This is the MOS.  If we have upgraded,
1303 		 * spa_max_replication() could change, so reset
1304 		 * os_copies here.
1305 		 */
1306 		os->os_copies = spa_max_replication(os->os_spa);
1307 	}
1308 
1309 	/*
1310 	 * Create the root block IO
1311 	 */
1312 	SET_BOOKMARK(&zb, os->os_dsl_dataset ?
1313 	    os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1314 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
1315 	arc_release(os->os_phys_buf, &os->os_phys_buf);
1316 
1317 	dmu_write_policy(os, NULL, 0, 0, &zp);
1318 
1319 	zio = arc_write(pio, os->os_spa, tx->tx_txg,
1320 	    blkptr_copy, os->os_phys_buf, DMU_OS_IS_L2CACHEABLE(os),
1321 	    &zp, dmu_objset_write_ready, NULL, NULL, dmu_objset_write_done,
1322 	    os, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
1323 
1324 	/*
1325 	 * Sync special dnodes - the parent IO for the sync is the root block
1326 	 */
1327 	DMU_META_DNODE(os)->dn_zio = zio;
1328 	dnode_sync(DMU_META_DNODE(os), tx);
1329 
1330 	os->os_phys->os_flags = os->os_flags;
1331 
1332 	if (DMU_USERUSED_DNODE(os) &&
1333 	    DMU_USERUSED_DNODE(os)->dn_type != DMU_OT_NONE) {
1334 		DMU_USERUSED_DNODE(os)->dn_zio = zio;
1335 		dnode_sync(DMU_USERUSED_DNODE(os), tx);
1336 		DMU_GROUPUSED_DNODE(os)->dn_zio = zio;
1337 		dnode_sync(DMU_GROUPUSED_DNODE(os), tx);
1338 	}
1339 
1340 	txgoff = tx->tx_txg & TXG_MASK;
1341 
1342 	if (dmu_objset_userused_enabled(os)) {
1343 		/*
1344 		 * We must create the list here because it uses the
1345 		 * dn_dirty_link[] of this txg.  But it may already
1346 		 * exist because we call dsl_dataset_sync() twice per txg.
1347 		 */
1348 		if (os->os_synced_dnodes == NULL) {
1349 			os->os_synced_dnodes =
1350 			    multilist_create(sizeof (dnode_t),
1351 			    offsetof(dnode_t, dn_dirty_link[txgoff]),
1352 			    dnode_multilist_index_func);
1353 		} else {
1354 			ASSERT3U(os->os_synced_dnodes->ml_offset, ==,
1355 			    offsetof(dnode_t, dn_dirty_link[txgoff]));
1356 		}
1357 	}
1358 
1359 	for (int i = 0;
1360 	    i < multilist_get_num_sublists(os->os_dirty_dnodes[txgoff]); i++) {
1361 		sync_dnodes_arg_t *sda = kmem_alloc(sizeof (*sda), KM_SLEEP);
1362 		sda->sda_list = os->os_dirty_dnodes[txgoff];
1363 		sda->sda_sublist_idx = i;
1364 		sda->sda_tx = tx;
1365 		(void) taskq_dispatch(dmu_objset_pool(os)->dp_sync_taskq,
1366 		    sync_dnodes_task, sda, 0);
1367 		/* callback frees sda */
1368 	}
1369 	taskq_wait(dmu_objset_pool(os)->dp_sync_taskq);
1370 
1371 	list = &DMU_META_DNODE(os)->dn_dirty_records[txgoff];
1372 	while ((dr = list_head(list)) != NULL) {
1373 		ASSERT0(dr->dr_dbuf->db_level);
1374 		list_remove(list, dr);
1375 		if (dr->dr_zio)
1376 			zio_nowait(dr->dr_zio);
1377 	}
1378 
1379 	/* Enable dnode backfill if enough objects have been freed. */
1380 	if (os->os_freed_dnodes >= dmu_rescan_dnode_threshold) {
1381 		os->os_rescan_dnodes = B_TRUE;
1382 		os->os_freed_dnodes = 0;
1383 	}
1384 
1385 	/*
1386 	 * Free intent log blocks up to this tx.
1387 	 */
1388 	zil_sync(os->os_zil, tx);
1389 	os->os_phys->os_zil_header = os->os_zil_header;
1390 	zio_nowait(zio);
1391 }
1392 
1393 boolean_t
1394 dmu_objset_is_dirty(objset_t *os, uint64_t txg)
1395 {
1396 	return (!multilist_is_empty(os->os_dirty_dnodes[txg & TXG_MASK]));
1397 }
1398 
1399 static objset_used_cb_t *used_cbs[DMU_OST_NUMTYPES];
1400 
1401 void
1402 dmu_objset_register_type(dmu_objset_type_t ost, objset_used_cb_t *cb)
1403 {
1404 	used_cbs[ost] = cb;
1405 }
1406 
1407 boolean_t
1408 dmu_objset_userused_enabled(objset_t *os)
1409 {
1410 	return (spa_version(os->os_spa) >= SPA_VERSION_USERSPACE &&
1411 	    used_cbs[os->os_phys->os_type] != NULL &&
1412 	    DMU_USERUSED_DNODE(os) != NULL);
1413 }
1414 
1415 typedef struct userquota_node {
1416 	uint64_t uqn_id;
1417 	int64_t uqn_delta;
1418 	avl_node_t uqn_node;
1419 } userquota_node_t;
1420 
1421 typedef struct userquota_cache {
1422 	avl_tree_t uqc_user_deltas;
1423 	avl_tree_t uqc_group_deltas;
1424 } userquota_cache_t;
1425 
1426 static int
1427 userquota_compare(const void *l, const void *r)
1428 {
1429 	const userquota_node_t *luqn = l;
1430 	const userquota_node_t *ruqn = r;
1431 
1432 	if (luqn->uqn_id < ruqn->uqn_id)
1433 		return (-1);
1434 	if (luqn->uqn_id > ruqn->uqn_id)
1435 		return (1);
1436 	return (0);
1437 }
1438 
1439 static void
1440 do_userquota_cacheflush(objset_t *os, userquota_cache_t *cache, dmu_tx_t *tx)
1441 {
1442 	void *cookie;
1443 	userquota_node_t *uqn;
1444 
1445 	ASSERT(dmu_tx_is_syncing(tx));
1446 
1447 	cookie = NULL;
1448 	while ((uqn = avl_destroy_nodes(&cache->uqc_user_deltas,
1449 	    &cookie)) != NULL) {
1450 		/*
1451 		 * os_userused_lock protects against concurrent calls to
1452 		 * zap_increment_int().  It's needed because zap_increment_int()
1453 		 * is not thread-safe (i.e. not atomic).
1454 		 */
1455 		mutex_enter(&os->os_userused_lock);
1456 		VERIFY0(zap_increment_int(os, DMU_USERUSED_OBJECT,
1457 		    uqn->uqn_id, uqn->uqn_delta, tx));
1458 		mutex_exit(&os->os_userused_lock);
1459 		kmem_free(uqn, sizeof (*uqn));
1460 	}
1461 	avl_destroy(&cache->uqc_user_deltas);
1462 
1463 	cookie = NULL;
1464 	while ((uqn = avl_destroy_nodes(&cache->uqc_group_deltas,
1465 	    &cookie)) != NULL) {
1466 		mutex_enter(&os->os_userused_lock);
1467 		VERIFY0(zap_increment_int(os, DMU_GROUPUSED_OBJECT,
1468 		    uqn->uqn_id, uqn->uqn_delta, tx));
1469 		mutex_exit(&os->os_userused_lock);
1470 		kmem_free(uqn, sizeof (*uqn));
1471 	}
1472 	avl_destroy(&cache->uqc_group_deltas);
1473 }
1474 
1475 static void
1476 userquota_update_cache(avl_tree_t *avl, uint64_t id, int64_t delta)
1477 {
1478 	userquota_node_t search = { .uqn_id = id };
1479 	avl_index_t idx;
1480 
1481 	userquota_node_t *uqn = avl_find(avl, &search, &idx);
1482 	if (uqn == NULL) {
1483 		uqn = kmem_zalloc(sizeof (*uqn), KM_SLEEP);
1484 		uqn->uqn_id = id;
1485 		avl_insert(avl, uqn, idx);
1486 	}
1487 	uqn->uqn_delta += delta;
1488 }
1489 
1490 static void
1491 do_userquota_update(userquota_cache_t *cache, uint64_t used, uint64_t flags,
1492     uint64_t user, uint64_t group, boolean_t subtract)
1493 {
1494 	if ((flags & DNODE_FLAG_USERUSED_ACCOUNTED)) {
1495 		int64_t delta = DNODE_SIZE + used;
1496 		if (subtract)
1497 			delta = -delta;
1498 
1499 		userquota_update_cache(&cache->uqc_user_deltas, user, delta);
1500 		userquota_update_cache(&cache->uqc_group_deltas, group, delta);
1501 	}
1502 }
1503 
1504 typedef struct userquota_updates_arg {
1505 	objset_t *uua_os;
1506 	int uua_sublist_idx;
1507 	dmu_tx_t *uua_tx;
1508 } userquota_updates_arg_t;
1509 
1510 static void
1511 userquota_updates_task(void *arg)
1512 {
1513 	userquota_updates_arg_t *uua = arg;
1514 	objset_t *os = uua->uua_os;
1515 	dmu_tx_t *tx = uua->uua_tx;
1516 	dnode_t *dn;
1517 	userquota_cache_t cache = { 0 };
1518 
1519 	multilist_sublist_t *list =
1520 	    multilist_sublist_lock(os->os_synced_dnodes, uua->uua_sublist_idx);
1521 
1522 	ASSERT(multilist_sublist_head(list) == NULL ||
1523 	    dmu_objset_userused_enabled(os));
1524 	avl_create(&cache.uqc_user_deltas, userquota_compare,
1525 	    sizeof (userquota_node_t), offsetof(userquota_node_t, uqn_node));
1526 	avl_create(&cache.uqc_group_deltas, userquota_compare,
1527 	    sizeof (userquota_node_t), offsetof(userquota_node_t, uqn_node));
1528 
1529 	while ((dn = multilist_sublist_head(list)) != NULL) {
1530 		int flags;
1531 		ASSERT(!DMU_OBJECT_IS_SPECIAL(dn->dn_object));
1532 		ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE ||
1533 		    dn->dn_phys->dn_flags &
1534 		    DNODE_FLAG_USERUSED_ACCOUNTED);
1535 
1536 		flags = dn->dn_id_flags;
1537 		ASSERT(flags);
1538 		if (flags & DN_ID_OLD_EXIST)  {
1539 			do_userquota_update(&cache,
1540 			    dn->dn_oldused, dn->dn_oldflags,
1541 			    dn->dn_olduid, dn->dn_oldgid, B_TRUE);
1542 		}
1543 		if (flags & DN_ID_NEW_EXIST) {
1544 			do_userquota_update(&cache,
1545 			    DN_USED_BYTES(dn->dn_phys),
1546 			    dn->dn_phys->dn_flags,  dn->dn_newuid,
1547 			    dn->dn_newgid, B_FALSE);
1548 		}
1549 
1550 		mutex_enter(&dn->dn_mtx);
1551 		dn->dn_oldused = 0;
1552 		dn->dn_oldflags = 0;
1553 		if (dn->dn_id_flags & DN_ID_NEW_EXIST) {
1554 			dn->dn_olduid = dn->dn_newuid;
1555 			dn->dn_oldgid = dn->dn_newgid;
1556 			dn->dn_id_flags |= DN_ID_OLD_EXIST;
1557 			if (dn->dn_bonuslen == 0)
1558 				dn->dn_id_flags |= DN_ID_CHKED_SPILL;
1559 			else
1560 				dn->dn_id_flags |= DN_ID_CHKED_BONUS;
1561 		}
1562 		dn->dn_id_flags &= ~(DN_ID_NEW_EXIST);
1563 		mutex_exit(&dn->dn_mtx);
1564 
1565 		multilist_sublist_remove(list, dn);
1566 		dnode_rele(dn, os->os_synced_dnodes);
1567 	}
1568 	do_userquota_cacheflush(os, &cache, tx);
1569 	multilist_sublist_unlock(list);
1570 	kmem_free(uua, sizeof (*uua));
1571 }
1572 
1573 void
1574 dmu_objset_do_userquota_updates(objset_t *os, dmu_tx_t *tx)
1575 {
1576 	if (!dmu_objset_userused_enabled(os))
1577 		return;
1578 
1579 	/* Allocate the user/groupused objects if necessary. */
1580 	if (DMU_USERUSED_DNODE(os)->dn_type == DMU_OT_NONE) {
1581 		VERIFY0(zap_create_claim(os,
1582 		    DMU_USERUSED_OBJECT,
1583 		    DMU_OT_USERGROUP_USED, DMU_OT_NONE, 0, tx));
1584 		VERIFY0(zap_create_claim(os,
1585 		    DMU_GROUPUSED_OBJECT,
1586 		    DMU_OT_USERGROUP_USED, DMU_OT_NONE, 0, tx));
1587 	}
1588 
1589 	for (int i = 0;
1590 	    i < multilist_get_num_sublists(os->os_synced_dnodes); i++) {
1591 		userquota_updates_arg_t *uua =
1592 		    kmem_alloc(sizeof (*uua), KM_SLEEP);
1593 		uua->uua_os = os;
1594 		uua->uua_sublist_idx = i;
1595 		uua->uua_tx = tx;
1596 		/* note: caller does taskq_wait() */
1597 		(void) taskq_dispatch(dmu_objset_pool(os)->dp_sync_taskq,
1598 		    userquota_updates_task, uua, 0);
1599 		/* callback frees uua */
1600 	}
1601 }
1602 
1603 /*
1604  * Returns a pointer to data to find uid/gid from
1605  *
1606  * If a dirty record for transaction group that is syncing can't
1607  * be found then NULL is returned.  In the NULL case it is assumed
1608  * the uid/gid aren't changing.
1609  */
1610 static void *
1611 dmu_objset_userquota_find_data(dmu_buf_impl_t *db, dmu_tx_t *tx)
1612 {
1613 	dbuf_dirty_record_t *dr, **drp;
1614 	void *data;
1615 
1616 	if (db->db_dirtycnt == 0)
1617 		return (db->db.db_data);  /* Nothing is changing */
1618 
1619 	for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1620 		if (dr->dr_txg == tx->tx_txg)
1621 			break;
1622 
1623 	if (dr == NULL) {
1624 		data = NULL;
1625 	} else {
1626 		dnode_t *dn;
1627 
1628 		DB_DNODE_ENTER(dr->dr_dbuf);
1629 		dn = DB_DNODE(dr->dr_dbuf);
1630 
1631 		if (dn->dn_bonuslen == 0 &&
1632 		    dr->dr_dbuf->db_blkid == DMU_SPILL_BLKID)
1633 			data = dr->dt.dl.dr_data->b_data;
1634 		else
1635 			data = dr->dt.dl.dr_data;
1636 
1637 		DB_DNODE_EXIT(dr->dr_dbuf);
1638 	}
1639 
1640 	return (data);
1641 }
1642 
1643 void
1644 dmu_objset_userquota_get_ids(dnode_t *dn, boolean_t before, dmu_tx_t *tx)
1645 {
1646 	objset_t *os = dn->dn_objset;
1647 	void *data = NULL;
1648 	dmu_buf_impl_t *db = NULL;
1649 	uint64_t *user = NULL;
1650 	uint64_t *group = NULL;
1651 	int flags = dn->dn_id_flags;
1652 	int error;
1653 	boolean_t have_spill = B_FALSE;
1654 
1655 	if (!dmu_objset_userused_enabled(dn->dn_objset))
1656 		return;
1657 
1658 	if (before && (flags & (DN_ID_CHKED_BONUS|DN_ID_OLD_EXIST|
1659 	    DN_ID_CHKED_SPILL)))
1660 		return;
1661 
1662 	if (before && dn->dn_bonuslen != 0)
1663 		data = DN_BONUS(dn->dn_phys);
1664 	else if (!before && dn->dn_bonuslen != 0) {
1665 		if (dn->dn_bonus) {
1666 			db = dn->dn_bonus;
1667 			mutex_enter(&db->db_mtx);
1668 			data = dmu_objset_userquota_find_data(db, tx);
1669 		} else {
1670 			data = DN_BONUS(dn->dn_phys);
1671 		}
1672 	} else if (dn->dn_bonuslen == 0 && dn->dn_bonustype == DMU_OT_SA) {
1673 			int rf = 0;
1674 
1675 			if (RW_WRITE_HELD(&dn->dn_struct_rwlock))
1676 				rf |= DB_RF_HAVESTRUCT;
1677 			error = dmu_spill_hold_by_dnode(dn,
1678 			    rf | DB_RF_MUST_SUCCEED,
1679 			    FTAG, (dmu_buf_t **)&db);
1680 			ASSERT(error == 0);
1681 			mutex_enter(&db->db_mtx);
1682 			data = (before) ? db->db.db_data :
1683 			    dmu_objset_userquota_find_data(db, tx);
1684 			have_spill = B_TRUE;
1685 	} else {
1686 		mutex_enter(&dn->dn_mtx);
1687 		dn->dn_id_flags |= DN_ID_CHKED_BONUS;
1688 		mutex_exit(&dn->dn_mtx);
1689 		return;
1690 	}
1691 
1692 	if (before) {
1693 		ASSERT(data);
1694 		user = &dn->dn_olduid;
1695 		group = &dn->dn_oldgid;
1696 	} else if (data) {
1697 		user = &dn->dn_newuid;
1698 		group = &dn->dn_newgid;
1699 	}
1700 
1701 	/*
1702 	 * Must always call the callback in case the object
1703 	 * type has changed and that type isn't an object type to track
1704 	 */
1705 	error = used_cbs[os->os_phys->os_type](dn->dn_bonustype, data,
1706 	    user, group);
1707 
1708 	/*
1709 	 * Preserve existing uid/gid when the callback can't determine
1710 	 * what the new uid/gid are and the callback returned EEXIST.
1711 	 * The EEXIST error tells us to just use the existing uid/gid.
1712 	 * If we don't know what the old values are then just assign
1713 	 * them to 0, since that is a new file  being created.
1714 	 */
1715 	if (!before && data == NULL && error == EEXIST) {
1716 		if (flags & DN_ID_OLD_EXIST) {
1717 			dn->dn_newuid = dn->dn_olduid;
1718 			dn->dn_newgid = dn->dn_oldgid;
1719 		} else {
1720 			dn->dn_newuid = 0;
1721 			dn->dn_newgid = 0;
1722 		}
1723 		error = 0;
1724 	}
1725 
1726 	if (db)
1727 		mutex_exit(&db->db_mtx);
1728 
1729 	mutex_enter(&dn->dn_mtx);
1730 	if (error == 0 && before)
1731 		dn->dn_id_flags |= DN_ID_OLD_EXIST;
1732 	if (error == 0 && !before)
1733 		dn->dn_id_flags |= DN_ID_NEW_EXIST;
1734 
1735 	if (have_spill) {
1736 		dn->dn_id_flags |= DN_ID_CHKED_SPILL;
1737 	} else {
1738 		dn->dn_id_flags |= DN_ID_CHKED_BONUS;
1739 	}
1740 	mutex_exit(&dn->dn_mtx);
1741 	if (have_spill)
1742 		dmu_buf_rele((dmu_buf_t *)db, FTAG);
1743 }
1744 
1745 boolean_t
1746 dmu_objset_userspace_present(objset_t *os)
1747 {
1748 	return (os->os_phys->os_flags &
1749 	    OBJSET_FLAG_USERACCOUNTING_COMPLETE);
1750 }
1751 
1752 int
1753 dmu_objset_userspace_upgrade(objset_t *os)
1754 {
1755 	uint64_t obj;
1756 	int err = 0;
1757 
1758 	if (dmu_objset_userspace_present(os))
1759 		return (0);
1760 	if (!dmu_objset_userused_enabled(os))
1761 		return (SET_ERROR(ENOTSUP));
1762 	if (dmu_objset_is_snapshot(os))
1763 		return (SET_ERROR(EINVAL));
1764 
1765 	/*
1766 	 * We simply need to mark every object dirty, so that it will be
1767 	 * synced out and now accounted.  If this is called
1768 	 * concurrently, or if we already did some work before crashing,
1769 	 * that's fine, since we track each object's accounted state
1770 	 * independently.
1771 	 */
1772 
1773 	for (obj = 0; err == 0; err = dmu_object_next(os, &obj, FALSE, 0)) {
1774 		dmu_tx_t *tx;
1775 		dmu_buf_t *db;
1776 		int objerr;
1777 
1778 		if (issig(JUSTLOOKING) && issig(FORREAL))
1779 			return (SET_ERROR(EINTR));
1780 
1781 		objerr = dmu_bonus_hold(os, obj, FTAG, &db);
1782 		if (objerr != 0)
1783 			continue;
1784 		tx = dmu_tx_create(os);
1785 		dmu_tx_hold_bonus(tx, obj);
1786 		objerr = dmu_tx_assign(tx, TXG_WAIT);
1787 		if (objerr != 0) {
1788 			dmu_tx_abort(tx);
1789 			continue;
1790 		}
1791 		dmu_buf_will_dirty(db, tx);
1792 		dmu_buf_rele(db, FTAG);
1793 		dmu_tx_commit(tx);
1794 	}
1795 
1796 	os->os_flags |= OBJSET_FLAG_USERACCOUNTING_COMPLETE;
1797 	txg_wait_synced(dmu_objset_pool(os), 0);
1798 	return (0);
1799 }
1800 
1801 void
1802 dmu_objset_space(objset_t *os, uint64_t *refdbytesp, uint64_t *availbytesp,
1803     uint64_t *usedobjsp, uint64_t *availobjsp)
1804 {
1805 	dsl_dataset_space(os->os_dsl_dataset, refdbytesp, availbytesp,
1806 	    usedobjsp, availobjsp);
1807 }
1808 
1809 uint64_t
1810 dmu_objset_fsid_guid(objset_t *os)
1811 {
1812 	return (dsl_dataset_fsid_guid(os->os_dsl_dataset));
1813 }
1814 
1815 void
1816 dmu_objset_fast_stat(objset_t *os, dmu_objset_stats_t *stat)
1817 {
1818 	stat->dds_type = os->os_phys->os_type;
1819 	if (os->os_dsl_dataset)
1820 		dsl_dataset_fast_stat(os->os_dsl_dataset, stat);
1821 }
1822 
1823 void
1824 dmu_objset_stats(objset_t *os, nvlist_t *nv)
1825 {
1826 	ASSERT(os->os_dsl_dataset ||
1827 	    os->os_phys->os_type == DMU_OST_META);
1828 
1829 	if (os->os_dsl_dataset != NULL)
1830 		dsl_dataset_stats(os->os_dsl_dataset, nv);
1831 
1832 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_TYPE,
1833 	    os->os_phys->os_type);
1834 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USERACCOUNTING,
1835 	    dmu_objset_userspace_present(os));
1836 }
1837 
1838 int
1839 dmu_objset_is_snapshot(objset_t *os)
1840 {
1841 	if (os->os_dsl_dataset != NULL)
1842 		return (os->os_dsl_dataset->ds_is_snapshot);
1843 	else
1844 		return (B_FALSE);
1845 }
1846 
1847 int
1848 dmu_snapshot_realname(objset_t *os, char *name, char *real, int maxlen,
1849     boolean_t *conflict)
1850 {
1851 	dsl_dataset_t *ds = os->os_dsl_dataset;
1852 	uint64_t ignored;
1853 
1854 	if (dsl_dataset_phys(ds)->ds_snapnames_zapobj == 0)
1855 		return (SET_ERROR(ENOENT));
1856 
1857 	return (zap_lookup_norm(ds->ds_dir->dd_pool->dp_meta_objset,
1858 	    dsl_dataset_phys(ds)->ds_snapnames_zapobj, name, 8, 1, &ignored,
1859 	    MT_NORMALIZE, real, maxlen, conflict));
1860 }
1861 
1862 int
1863 dmu_snapshot_list_next(objset_t *os, int namelen, char *name,
1864     uint64_t *idp, uint64_t *offp, boolean_t *case_conflict)
1865 {
1866 	dsl_dataset_t *ds = os->os_dsl_dataset;
1867 	zap_cursor_t cursor;
1868 	zap_attribute_t attr;
1869 
1870 	ASSERT(dsl_pool_config_held(dmu_objset_pool(os)));
1871 
1872 	if (dsl_dataset_phys(ds)->ds_snapnames_zapobj == 0)
1873 		return (SET_ERROR(ENOENT));
1874 
1875 	zap_cursor_init_serialized(&cursor,
1876 	    ds->ds_dir->dd_pool->dp_meta_objset,
1877 	    dsl_dataset_phys(ds)->ds_snapnames_zapobj, *offp);
1878 
1879 	if (zap_cursor_retrieve(&cursor, &attr) != 0) {
1880 		zap_cursor_fini(&cursor);
1881 		return (SET_ERROR(ENOENT));
1882 	}
1883 
1884 	if (strlen(attr.za_name) + 1 > namelen) {
1885 		zap_cursor_fini(&cursor);
1886 		return (SET_ERROR(ENAMETOOLONG));
1887 	}
1888 
1889 	(void) strcpy(name, attr.za_name);
1890 	if (idp)
1891 		*idp = attr.za_first_integer;
1892 	if (case_conflict)
1893 		*case_conflict = attr.za_normalization_conflict;
1894 	zap_cursor_advance(&cursor);
1895 	*offp = zap_cursor_serialize(&cursor);
1896 	zap_cursor_fini(&cursor);
1897 
1898 	return (0);
1899 }
1900 
1901 int
1902 dmu_dir_list_next(objset_t *os, int namelen, char *name,
1903     uint64_t *idp, uint64_t *offp)
1904 {
1905 	dsl_dir_t *dd = os->os_dsl_dataset->ds_dir;
1906 	zap_cursor_t cursor;
1907 	zap_attribute_t attr;
1908 
1909 	/* there is no next dir on a snapshot! */
1910 	if (os->os_dsl_dataset->ds_object !=
1911 	    dsl_dir_phys(dd)->dd_head_dataset_obj)
1912 		return (SET_ERROR(ENOENT));
1913 
1914 	zap_cursor_init_serialized(&cursor,
1915 	    dd->dd_pool->dp_meta_objset,
1916 	    dsl_dir_phys(dd)->dd_child_dir_zapobj, *offp);
1917 
1918 	if (zap_cursor_retrieve(&cursor, &attr) != 0) {
1919 		zap_cursor_fini(&cursor);
1920 		return (SET_ERROR(ENOENT));
1921 	}
1922 
1923 	if (strlen(attr.za_name) + 1 > namelen) {
1924 		zap_cursor_fini(&cursor);
1925 		return (SET_ERROR(ENAMETOOLONG));
1926 	}
1927 
1928 	(void) strcpy(name, attr.za_name);
1929 	if (idp)
1930 		*idp = attr.za_first_integer;
1931 	zap_cursor_advance(&cursor);
1932 	*offp = zap_cursor_serialize(&cursor);
1933 	zap_cursor_fini(&cursor);
1934 
1935 	return (0);
1936 }
1937 
1938 typedef struct dmu_objset_find_ctx {
1939 	taskq_t		*dc_tq;
1940 	dsl_pool_t	*dc_dp;
1941 	uint64_t	dc_ddobj;
1942 	char		*dc_ddname; /* last component of ddobj's name */
1943 	int		(*dc_func)(dsl_pool_t *, dsl_dataset_t *, void *);
1944 	void		*dc_arg;
1945 	int		dc_flags;
1946 	kmutex_t	*dc_error_lock;
1947 	int		*dc_error;
1948 } dmu_objset_find_ctx_t;
1949 
1950 static void
1951 dmu_objset_find_dp_impl(dmu_objset_find_ctx_t *dcp)
1952 {
1953 	dsl_pool_t *dp = dcp->dc_dp;
1954 	dsl_dir_t *dd;
1955 	dsl_dataset_t *ds;
1956 	zap_cursor_t zc;
1957 	zap_attribute_t *attr;
1958 	uint64_t thisobj;
1959 	int err = 0;
1960 
1961 	/* don't process if there already was an error */
1962 	if (*dcp->dc_error != 0)
1963 		goto out;
1964 
1965 	/*
1966 	 * Note: passing the name (dc_ddname) here is optional, but it
1967 	 * improves performance because we don't need to call
1968 	 * zap_value_search() to determine the name.
1969 	 */
1970 	err = dsl_dir_hold_obj(dp, dcp->dc_ddobj, dcp->dc_ddname, FTAG, &dd);
1971 	if (err != 0)
1972 		goto out;
1973 
1974 	/* Don't visit hidden ($MOS & $ORIGIN) objsets. */
1975 	if (dd->dd_myname[0] == '$') {
1976 		dsl_dir_rele(dd, FTAG);
1977 		goto out;
1978 	}
1979 
1980 	thisobj = dsl_dir_phys(dd)->dd_head_dataset_obj;
1981 	attr = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
1982 
1983 	/*
1984 	 * Iterate over all children.
1985 	 */
1986 	if (dcp->dc_flags & DS_FIND_CHILDREN) {
1987 		for (zap_cursor_init(&zc, dp->dp_meta_objset,
1988 		    dsl_dir_phys(dd)->dd_child_dir_zapobj);
1989 		    zap_cursor_retrieve(&zc, attr) == 0;
1990 		    (void) zap_cursor_advance(&zc)) {
1991 			ASSERT3U(attr->za_integer_length, ==,
1992 			    sizeof (uint64_t));
1993 			ASSERT3U(attr->za_num_integers, ==, 1);
1994 
1995 			dmu_objset_find_ctx_t *child_dcp =
1996 			    kmem_alloc(sizeof (*child_dcp), KM_SLEEP);
1997 			*child_dcp = *dcp;
1998 			child_dcp->dc_ddobj = attr->za_first_integer;
1999 			child_dcp->dc_ddname = spa_strdup(attr->za_name);
2000 			if (dcp->dc_tq != NULL)
2001 				(void) taskq_dispatch(dcp->dc_tq,
2002 				    dmu_objset_find_dp_cb, child_dcp, TQ_SLEEP);
2003 			else
2004 				dmu_objset_find_dp_impl(child_dcp);
2005 		}
2006 		zap_cursor_fini(&zc);
2007 	}
2008 
2009 	/*
2010 	 * Iterate over all snapshots.
2011 	 */
2012 	if (dcp->dc_flags & DS_FIND_SNAPSHOTS) {
2013 		dsl_dataset_t *ds;
2014 		err = dsl_dataset_hold_obj(dp, thisobj, FTAG, &ds);
2015 
2016 		if (err == 0) {
2017 			uint64_t snapobj;
2018 
2019 			snapobj = dsl_dataset_phys(ds)->ds_snapnames_zapobj;
2020 			dsl_dataset_rele(ds, FTAG);
2021 
2022 			for (zap_cursor_init(&zc, dp->dp_meta_objset, snapobj);
2023 			    zap_cursor_retrieve(&zc, attr) == 0;
2024 			    (void) zap_cursor_advance(&zc)) {
2025 				ASSERT3U(attr->za_integer_length, ==,
2026 				    sizeof (uint64_t));
2027 				ASSERT3U(attr->za_num_integers, ==, 1);
2028 
2029 				err = dsl_dataset_hold_obj(dp,
2030 				    attr->za_first_integer, FTAG, &ds);
2031 				if (err != 0)
2032 					break;
2033 				err = dcp->dc_func(dp, ds, dcp->dc_arg);
2034 				dsl_dataset_rele(ds, FTAG);
2035 				if (err != 0)
2036 					break;
2037 			}
2038 			zap_cursor_fini(&zc);
2039 		}
2040 	}
2041 
2042 	kmem_free(attr, sizeof (zap_attribute_t));
2043 
2044 	if (err != 0) {
2045 		dsl_dir_rele(dd, FTAG);
2046 		goto out;
2047 	}
2048 
2049 	/*
2050 	 * Apply to self.
2051 	 */
2052 	err = dsl_dataset_hold_obj(dp, thisobj, FTAG, &ds);
2053 
2054 	/*
2055 	 * Note: we hold the dir while calling dsl_dataset_hold_obj() so
2056 	 * that the dir will remain cached, and we won't have to re-instantiate
2057 	 * it (which could be expensive due to finding its name via
2058 	 * zap_value_search()).
2059 	 */
2060 	dsl_dir_rele(dd, FTAG);
2061 	if (err != 0)
2062 		goto out;
2063 	err = dcp->dc_func(dp, ds, dcp->dc_arg);
2064 	dsl_dataset_rele(ds, FTAG);
2065 
2066 out:
2067 	if (err != 0) {
2068 		mutex_enter(dcp->dc_error_lock);
2069 		/* only keep first error */
2070 		if (*dcp->dc_error == 0)
2071 			*dcp->dc_error = err;
2072 		mutex_exit(dcp->dc_error_lock);
2073 	}
2074 
2075 	if (dcp->dc_ddname != NULL)
2076 		spa_strfree(dcp->dc_ddname);
2077 	kmem_free(dcp, sizeof (*dcp));
2078 }
2079 
2080 static void
2081 dmu_objset_find_dp_cb(void *arg)
2082 {
2083 	dmu_objset_find_ctx_t *dcp = arg;
2084 	dsl_pool_t *dp = dcp->dc_dp;
2085 
2086 	/*
2087 	 * We need to get a pool_config_lock here, as there are several
2088 	 * asssert(pool_config_held) down the stack. Getting a lock via
2089 	 * dsl_pool_config_enter is risky, as it might be stalled by a
2090 	 * pending writer. This would deadlock, as the write lock can
2091 	 * only be granted when our parent thread gives up the lock.
2092 	 * The _prio interface gives us priority over a pending writer.
2093 	 */
2094 	dsl_pool_config_enter_prio(dp, FTAG);
2095 
2096 	dmu_objset_find_dp_impl(dcp);
2097 
2098 	dsl_pool_config_exit(dp, FTAG);
2099 }
2100 
2101 /*
2102  * Find objsets under and including ddobj, call func(ds) on each.
2103  * The order for the enumeration is completely undefined.
2104  * func is called with dsl_pool_config held.
2105  */
2106 int
2107 dmu_objset_find_dp(dsl_pool_t *dp, uint64_t ddobj,
2108     int func(dsl_pool_t *, dsl_dataset_t *, void *), void *arg, int flags)
2109 {
2110 	int error = 0;
2111 	taskq_t *tq = NULL;
2112 	int ntasks;
2113 	dmu_objset_find_ctx_t *dcp;
2114 	kmutex_t err_lock;
2115 
2116 	mutex_init(&err_lock, NULL, MUTEX_DEFAULT, NULL);
2117 	dcp = kmem_alloc(sizeof (*dcp), KM_SLEEP);
2118 	dcp->dc_tq = NULL;
2119 	dcp->dc_dp = dp;
2120 	dcp->dc_ddobj = ddobj;
2121 	dcp->dc_ddname = NULL;
2122 	dcp->dc_func = func;
2123 	dcp->dc_arg = arg;
2124 	dcp->dc_flags = flags;
2125 	dcp->dc_error_lock = &err_lock;
2126 	dcp->dc_error = &error;
2127 
2128 	if ((flags & DS_FIND_SERIALIZE) || dsl_pool_config_held_writer(dp)) {
2129 		/*
2130 		 * In case a write lock is held we can't make use of
2131 		 * parallelism, as down the stack of the worker threads
2132 		 * the lock is asserted via dsl_pool_config_held.
2133 		 * In case of a read lock this is solved by getting a read
2134 		 * lock in each worker thread, which isn't possible in case
2135 		 * of a writer lock. So we fall back to the synchronous path
2136 		 * here.
2137 		 * In the future it might be possible to get some magic into
2138 		 * dsl_pool_config_held in a way that it returns true for
2139 		 * the worker threads so that a single lock held from this
2140 		 * thread suffices. For now, stay single threaded.
2141 		 */
2142 		dmu_objset_find_dp_impl(dcp);
2143 		mutex_destroy(&err_lock);
2144 
2145 		return (error);
2146 	}
2147 
2148 	ntasks = dmu_find_threads;
2149 	if (ntasks == 0)
2150 		ntasks = vdev_count_leaves(dp->dp_spa) * 4;
2151 	tq = taskq_create("dmu_objset_find", ntasks, minclsyspri, ntasks,
2152 	    INT_MAX, 0);
2153 	if (tq == NULL) {
2154 		kmem_free(dcp, sizeof (*dcp));
2155 		mutex_destroy(&err_lock);
2156 
2157 		return (SET_ERROR(ENOMEM));
2158 	}
2159 	dcp->dc_tq = tq;
2160 
2161 	/* dcp will be freed by task */
2162 	(void) taskq_dispatch(tq, dmu_objset_find_dp_cb, dcp, TQ_SLEEP);
2163 
2164 	/*
2165 	 * PORTING: this code relies on the property of taskq_wait to wait
2166 	 * until no more tasks are queued and no more tasks are active. As
2167 	 * we always queue new tasks from within other tasks, task_wait
2168 	 * reliably waits for the full recursion to finish, even though we
2169 	 * enqueue new tasks after taskq_wait has been called.
2170 	 * On platforms other than illumos, taskq_wait may not have this
2171 	 * property.
2172 	 */
2173 	taskq_wait(tq);
2174 	taskq_destroy(tq);
2175 	mutex_destroy(&err_lock);
2176 
2177 	return (error);
2178 }
2179 
2180 /*
2181  * Find all objsets under name, and for each, call 'func(child_name, arg)'.
2182  * The dp_config_rwlock must not be held when this is called, and it
2183  * will not be held when the callback is called.
2184  * Therefore this function should only be used when the pool is not changing
2185  * (e.g. in syncing context), or the callback can deal with the possible races.
2186  */
2187 static int
2188 dmu_objset_find_impl(spa_t *spa, const char *name,
2189     int func(const char *, void *), void *arg, int flags)
2190 {
2191 	dsl_dir_t *dd;
2192 	dsl_pool_t *dp = spa_get_dsl(spa);
2193 	dsl_dataset_t *ds;
2194 	zap_cursor_t zc;
2195 	zap_attribute_t *attr;
2196 	char *child;
2197 	uint64_t thisobj;
2198 	int err;
2199 
2200 	dsl_pool_config_enter(dp, FTAG);
2201 
2202 	err = dsl_dir_hold(dp, name, FTAG, &dd, NULL);
2203 	if (err != 0) {
2204 		dsl_pool_config_exit(dp, FTAG);
2205 		return (err);
2206 	}
2207 
2208 	/* Don't visit hidden ($MOS & $ORIGIN) objsets. */
2209 	if (dd->dd_myname[0] == '$') {
2210 		dsl_dir_rele(dd, FTAG);
2211 		dsl_pool_config_exit(dp, FTAG);
2212 		return (0);
2213 	}
2214 
2215 	thisobj = dsl_dir_phys(dd)->dd_head_dataset_obj;
2216 	attr = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
2217 
2218 	/*
2219 	 * Iterate over all children.
2220 	 */
2221 	if (flags & DS_FIND_CHILDREN) {
2222 		for (zap_cursor_init(&zc, dp->dp_meta_objset,
2223 		    dsl_dir_phys(dd)->dd_child_dir_zapobj);
2224 		    zap_cursor_retrieve(&zc, attr) == 0;
2225 		    (void) zap_cursor_advance(&zc)) {
2226 			ASSERT3U(attr->za_integer_length, ==,
2227 			    sizeof (uint64_t));
2228 			ASSERT3U(attr->za_num_integers, ==, 1);
2229 
2230 			child = kmem_asprintf("%s/%s", name, attr->za_name);
2231 			dsl_pool_config_exit(dp, FTAG);
2232 			err = dmu_objset_find_impl(spa, child,
2233 			    func, arg, flags);
2234 			dsl_pool_config_enter(dp, FTAG);
2235 			strfree(child);
2236 			if (err != 0)
2237 				break;
2238 		}
2239 		zap_cursor_fini(&zc);
2240 
2241 		if (err != 0) {
2242 			dsl_dir_rele(dd, FTAG);
2243 			dsl_pool_config_exit(dp, FTAG);
2244 			kmem_free(attr, sizeof (zap_attribute_t));
2245 			return (err);
2246 		}
2247 	}
2248 
2249 	/*
2250 	 * Iterate over all snapshots.
2251 	 */
2252 	if (flags & DS_FIND_SNAPSHOTS) {
2253 		err = dsl_dataset_hold_obj(dp, thisobj, FTAG, &ds);
2254 
2255 		if (err == 0) {
2256 			uint64_t snapobj;
2257 
2258 			snapobj = dsl_dataset_phys(ds)->ds_snapnames_zapobj;
2259 			dsl_dataset_rele(ds, FTAG);
2260 
2261 			for (zap_cursor_init(&zc, dp->dp_meta_objset, snapobj);
2262 			    zap_cursor_retrieve(&zc, attr) == 0;
2263 			    (void) zap_cursor_advance(&zc)) {
2264 				ASSERT3U(attr->za_integer_length, ==,
2265 				    sizeof (uint64_t));
2266 				ASSERT3U(attr->za_num_integers, ==, 1);
2267 
2268 				child = kmem_asprintf("%s@%s",
2269 				    name, attr->za_name);
2270 				dsl_pool_config_exit(dp, FTAG);
2271 				err = func(child, arg);
2272 				dsl_pool_config_enter(dp, FTAG);
2273 				strfree(child);
2274 				if (err != 0)
2275 					break;
2276 			}
2277 			zap_cursor_fini(&zc);
2278 		}
2279 	}
2280 
2281 	dsl_dir_rele(dd, FTAG);
2282 	kmem_free(attr, sizeof (zap_attribute_t));
2283 	dsl_pool_config_exit(dp, FTAG);
2284 
2285 	if (err != 0)
2286 		return (err);
2287 
2288 	/* Apply to self. */
2289 	return (func(name, arg));
2290 }
2291 
2292 /*
2293  * See comment above dmu_objset_find_impl().
2294  */
2295 int
2296 dmu_objset_find(char *name, int func(const char *, void *), void *arg,
2297     int flags)
2298 {
2299 	spa_t *spa;
2300 	int error;
2301 
2302 	error = spa_open(name, &spa, FTAG);
2303 	if (error != 0)
2304 		return (error);
2305 	error = dmu_objset_find_impl(spa, name, func, arg, flags);
2306 	spa_close(spa, FTAG);
2307 	return (error);
2308 }
2309 
2310 void
2311 dmu_objset_set_user(objset_t *os, void *user_ptr)
2312 {
2313 	ASSERT(MUTEX_HELD(&os->os_user_ptr_lock));
2314 	os->os_user_ptr = user_ptr;
2315 }
2316 
2317 void *
2318 dmu_objset_get_user(objset_t *os)
2319 {
2320 	ASSERT(MUTEX_HELD(&os->os_user_ptr_lock));
2321 	return (os->os_user_ptr);
2322 }
2323 
2324 /*
2325  * Determine name of filesystem, given name of snapshot.
2326  * buf must be at least ZFS_MAX_DATASET_NAME_LEN bytes
2327  */
2328 int
2329 dmu_fsname(const char *snapname, char *buf)
2330 {
2331 	char *atp = strchr(snapname, '@');
2332 	if (atp == NULL)
2333 		return (SET_ERROR(EINVAL));
2334 	if (atp - snapname >= ZFS_MAX_DATASET_NAME_LEN)
2335 		return (SET_ERROR(ENAMETOOLONG));
2336 	(void) strlcpy(buf, snapname, atp - snapname + 1);
2337 	return (0);
2338 }
2339 
2340 /*
2341  * Call when we think we're going to write/free space in open context to track
2342  * the amount of dirty data in the open txg, which is also the amount
2343  * of memory that can not be evicted until this txg syncs.
2344  */
2345 void
2346 dmu_objset_willuse_space(objset_t *os, int64_t space, dmu_tx_t *tx)
2347 {
2348 	dsl_dataset_t *ds = os->os_dsl_dataset;
2349 	int64_t aspace = spa_get_worst_case_asize(os->os_spa, space);
2350 
2351 	if (ds != NULL) {
2352 		dsl_dir_willuse_space(ds->ds_dir, aspace, tx);
2353 		dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx);
2354 	}
2355 }
2356