xref: /freebsd/sys/contrib/openzfs/module/zfs/dsl_dir.c (revision 32a95656b51ebefcdf3e0b02c110825f59abd26f)
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  * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
24  * Copyright (c) 2013 Martin Matuska. All rights reserved.
25  * Copyright (c) 2014 Joyent, Inc. All rights reserved.
26  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
27  * Copyright (c) 2016 Actifio, Inc. All rights reserved.
28  * Copyright (c) 2018, loli10K <ezomori.nozomu@gmail.com>. All rights reserved.
29  */
30 
31 #include <sys/dmu.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dsl_dataset.h>
35 #include <sys/dsl_dir.h>
36 #include <sys/dsl_prop.h>
37 #include <sys/dsl_synctask.h>
38 #include <sys/dsl_deleg.h>
39 #include <sys/dmu_impl.h>
40 #include <sys/spa.h>
41 #include <sys/spa_impl.h>
42 #include <sys/metaslab.h>
43 #include <sys/zap.h>
44 #include <sys/zio.h>
45 #include <sys/arc.h>
46 #include <sys/sunddi.h>
47 #include <sys/zfeature.h>
48 #include <sys/policy.h>
49 #include <sys/zfs_vfsops.h>
50 #include <sys/zfs_znode.h>
51 #include <sys/zvol.h>
52 #include <sys/zthr.h>
53 #include "zfs_namecheck.h"
54 #include "zfs_prop.h"
55 
56 /*
57  * Filesystem and Snapshot Limits
58  * ------------------------------
59  *
60  * These limits are used to restrict the number of filesystems and/or snapshots
61  * that can be created at a given level in the tree or below. A typical
62  * use-case is with a delegated dataset where the administrator wants to ensure
63  * that a user within the zone is not creating too many additional filesystems
64  * or snapshots, even though they're not exceeding their space quota.
65  *
66  * The filesystem and snapshot counts are stored as extensible properties. This
67  * capability is controlled by a feature flag and must be enabled to be used.
68  * Once enabled, the feature is not active until the first limit is set. At
69  * that point, future operations to create/destroy filesystems or snapshots
70  * will validate and update the counts.
71  *
72  * Because the count properties will not exist before the feature is active,
73  * the counts are updated when a limit is first set on an uninitialized
74  * dsl_dir node in the tree (The filesystem/snapshot count on a node includes
75  * all of the nested filesystems/snapshots. Thus, a new leaf node has a
76  * filesystem count of 0 and a snapshot count of 0. Non-existent filesystem and
77  * snapshot count properties on a node indicate uninitialized counts on that
78  * node.) When first setting a limit on an uninitialized node, the code starts
79  * at the filesystem with the new limit and descends into all sub-filesystems
80  * to add the count properties.
81  *
82  * In practice this is lightweight since a limit is typically set when the
83  * filesystem is created and thus has no children. Once valid, changing the
84  * limit value won't require a re-traversal since the counts are already valid.
85  * When recursively fixing the counts, if a node with a limit is encountered
86  * during the descent, the counts are known to be valid and there is no need to
87  * descend into that filesystem's children. The counts on filesystems above the
88  * one with the new limit will still be uninitialized, unless a limit is
89  * eventually set on one of those filesystems. The counts are always recursively
90  * updated when a limit is set on a dataset, unless there is already a limit.
91  * When a new limit value is set on a filesystem with an existing limit, it is
92  * possible for the new limit to be less than the current count at that level
93  * since a user who can change the limit is also allowed to exceed the limit.
94  *
95  * Once the feature is active, then whenever a filesystem or snapshot is
96  * created, the code recurses up the tree, validating the new count against the
97  * limit at each initialized level. In practice, most levels will not have a
98  * limit set. If there is a limit at any initialized level up the tree, the
99  * check must pass or the creation will fail. Likewise, when a filesystem or
100  * snapshot is destroyed, the counts are recursively adjusted all the way up
101  * the initialized nodes in the tree. Renaming a filesystem into different point
102  * in the tree will first validate, then update the counts on each branch up to
103  * the common ancestor. A receive will also validate the counts and then update
104  * them.
105  *
106  * An exception to the above behavior is that the limit is not enforced if the
107  * user has permission to modify the limit. This is primarily so that
108  * recursive snapshots in the global zone always work. We want to prevent a
109  * denial-of-service in which a lower level delegated dataset could max out its
110  * limit and thus block recursive snapshots from being taken in the global zone.
111  * Because of this, it is possible for the snapshot count to be over the limit
112  * and snapshots taken in the global zone could cause a lower level dataset to
113  * hit or exceed its limit. The administrator taking the global zone recursive
114  * snapshot should be aware of this side-effect and behave accordingly.
115  * For consistency, the filesystem limit is also not enforced if the user can
116  * modify the limit.
117  *
118  * The filesystem and snapshot limits are validated by dsl_fs_ss_limit_check()
119  * and updated by dsl_fs_ss_count_adjust(). A new limit value is setup in
120  * dsl_dir_activate_fs_ss_limit() and the counts are adjusted, if necessary, by
121  * dsl_dir_init_fs_ss_count().
122  */
123 
124 extern inline dsl_dir_phys_t *dsl_dir_phys(dsl_dir_t *dd);
125 
126 static uint64_t dsl_dir_space_towrite(dsl_dir_t *dd);
127 
128 typedef struct ddulrt_arg {
129 	dsl_dir_t	*ddulrta_dd;
130 	uint64_t	ddlrta_txg;
131 } ddulrt_arg_t;
132 
133 static void
134 dsl_dir_evict_async(void *dbu)
135 {
136 	dsl_dir_t *dd = dbu;
137 	int t;
138 	dsl_pool_t *dp __maybe_unused = dd->dd_pool;
139 
140 	dd->dd_dbuf = NULL;
141 
142 	for (t = 0; t < TXG_SIZE; t++) {
143 		ASSERT(!txg_list_member(&dp->dp_dirty_dirs, dd, t));
144 		ASSERT(dd->dd_tempreserved[t] == 0);
145 		ASSERT(dd->dd_space_towrite[t] == 0);
146 	}
147 
148 	if (dd->dd_parent)
149 		dsl_dir_async_rele(dd->dd_parent, dd);
150 
151 	spa_async_close(dd->dd_pool->dp_spa, dd);
152 
153 	if (dsl_deadlist_is_open(&dd->dd_livelist))
154 		dsl_dir_livelist_close(dd);
155 
156 	dsl_prop_fini(dd);
157 	cv_destroy(&dd->dd_activity_cv);
158 	mutex_destroy(&dd->dd_activity_lock);
159 	mutex_destroy(&dd->dd_lock);
160 	kmem_free(dd, sizeof (dsl_dir_t));
161 }
162 
163 int
164 dsl_dir_hold_obj(dsl_pool_t *dp, uint64_t ddobj,
165     const char *tail, void *tag, dsl_dir_t **ddp)
166 {
167 	dmu_buf_t *dbuf;
168 	dsl_dir_t *dd;
169 	dmu_object_info_t doi;
170 	int err;
171 
172 	ASSERT(dsl_pool_config_held(dp));
173 
174 	err = dmu_bonus_hold(dp->dp_meta_objset, ddobj, tag, &dbuf);
175 	if (err != 0)
176 		return (err);
177 	dd = dmu_buf_get_user(dbuf);
178 
179 	dmu_object_info_from_db(dbuf, &doi);
180 	ASSERT3U(doi.doi_bonus_type, ==, DMU_OT_DSL_DIR);
181 	ASSERT3U(doi.doi_bonus_size, >=, sizeof (dsl_dir_phys_t));
182 
183 	if (dd == NULL) {
184 		dsl_dir_t *winner;
185 
186 		dd = kmem_zalloc(sizeof (dsl_dir_t), KM_SLEEP);
187 		dd->dd_object = ddobj;
188 		dd->dd_dbuf = dbuf;
189 		dd->dd_pool = dp;
190 
191 		mutex_init(&dd->dd_lock, NULL, MUTEX_DEFAULT, NULL);
192 		mutex_init(&dd->dd_activity_lock, NULL, MUTEX_DEFAULT, NULL);
193 		cv_init(&dd->dd_activity_cv, NULL, CV_DEFAULT, NULL);
194 		dsl_prop_init(dd);
195 
196 		if (dsl_dir_is_zapified(dd)) {
197 			err = zap_lookup(dp->dp_meta_objset,
198 			    ddobj, DD_FIELD_CRYPTO_KEY_OBJ,
199 			    sizeof (uint64_t), 1, &dd->dd_crypto_obj);
200 			if (err == 0) {
201 				/* check for on-disk format errata */
202 				if (dsl_dir_incompatible_encryption_version(
203 				    dd)) {
204 					dp->dp_spa->spa_errata =
205 					    ZPOOL_ERRATA_ZOL_6845_ENCRYPTION;
206 				}
207 			} else if (err != ENOENT) {
208 				goto errout;
209 			}
210 		}
211 
212 		dsl_dir_snap_cmtime_update(dd);
213 
214 		if (dsl_dir_phys(dd)->dd_parent_obj) {
215 			err = dsl_dir_hold_obj(dp,
216 			    dsl_dir_phys(dd)->dd_parent_obj, NULL, dd,
217 			    &dd->dd_parent);
218 			if (err != 0)
219 				goto errout;
220 			if (tail) {
221 #ifdef ZFS_DEBUG
222 				uint64_t foundobj;
223 
224 				err = zap_lookup(dp->dp_meta_objset,
225 				    dsl_dir_phys(dd->dd_parent)->
226 				    dd_child_dir_zapobj, tail,
227 				    sizeof (foundobj), 1, &foundobj);
228 				ASSERT(err || foundobj == ddobj);
229 #endif
230 				(void) strlcpy(dd->dd_myname, tail,
231 				    sizeof (dd->dd_myname));
232 			} else {
233 				err = zap_value_search(dp->dp_meta_objset,
234 				    dsl_dir_phys(dd->dd_parent)->
235 				    dd_child_dir_zapobj,
236 				    ddobj, 0, dd->dd_myname);
237 			}
238 			if (err != 0)
239 				goto errout;
240 		} else {
241 			(void) strlcpy(dd->dd_myname, spa_name(dp->dp_spa),
242 			    sizeof (dd->dd_myname));
243 		}
244 
245 		if (dsl_dir_is_clone(dd)) {
246 			dmu_buf_t *origin_bonus;
247 			dsl_dataset_phys_t *origin_phys;
248 
249 			/*
250 			 * We can't open the origin dataset, because
251 			 * that would require opening this dsl_dir.
252 			 * Just look at its phys directly instead.
253 			 */
254 			err = dmu_bonus_hold(dp->dp_meta_objset,
255 			    dsl_dir_phys(dd)->dd_origin_obj, FTAG,
256 			    &origin_bonus);
257 			if (err != 0)
258 				goto errout;
259 			origin_phys = origin_bonus->db_data;
260 			dd->dd_origin_txg =
261 			    origin_phys->ds_creation_txg;
262 			dmu_buf_rele(origin_bonus, FTAG);
263 			if (dsl_dir_is_zapified(dd)) {
264 				uint64_t obj;
265 				err = zap_lookup(dp->dp_meta_objset,
266 				    dd->dd_object, DD_FIELD_LIVELIST,
267 				    sizeof (uint64_t), 1, &obj);
268 				if (err == 0)
269 					dsl_dir_livelist_open(dd, obj);
270 				else if (err != ENOENT)
271 					goto errout;
272 			}
273 		}
274 
275 		dmu_buf_init_user(&dd->dd_dbu, NULL, dsl_dir_evict_async,
276 		    &dd->dd_dbuf);
277 		winner = dmu_buf_set_user_ie(dbuf, &dd->dd_dbu);
278 		if (winner != NULL) {
279 			if (dd->dd_parent)
280 				dsl_dir_rele(dd->dd_parent, dd);
281 			if (dsl_deadlist_is_open(&dd->dd_livelist))
282 				dsl_dir_livelist_close(dd);
283 			dsl_prop_fini(dd);
284 			cv_destroy(&dd->dd_activity_cv);
285 			mutex_destroy(&dd->dd_activity_lock);
286 			mutex_destroy(&dd->dd_lock);
287 			kmem_free(dd, sizeof (dsl_dir_t));
288 			dd = winner;
289 		} else {
290 			spa_open_ref(dp->dp_spa, dd);
291 		}
292 	}
293 
294 	/*
295 	 * The dsl_dir_t has both open-to-close and instantiate-to-evict
296 	 * holds on the spa.  We need the open-to-close holds because
297 	 * otherwise the spa_refcnt wouldn't change when we open a
298 	 * dir which the spa also has open, so we could incorrectly
299 	 * think it was OK to unload/export/destroy the pool.  We need
300 	 * the instantiate-to-evict hold because the dsl_dir_t has a
301 	 * pointer to the dd_pool, which has a pointer to the spa_t.
302 	 */
303 	spa_open_ref(dp->dp_spa, tag);
304 	ASSERT3P(dd->dd_pool, ==, dp);
305 	ASSERT3U(dd->dd_object, ==, ddobj);
306 	ASSERT3P(dd->dd_dbuf, ==, dbuf);
307 	*ddp = dd;
308 	return (0);
309 
310 errout:
311 	if (dd->dd_parent)
312 		dsl_dir_rele(dd->dd_parent, dd);
313 	if (dsl_deadlist_is_open(&dd->dd_livelist))
314 		dsl_dir_livelist_close(dd);
315 	dsl_prop_fini(dd);
316 	cv_destroy(&dd->dd_activity_cv);
317 	mutex_destroy(&dd->dd_activity_lock);
318 	mutex_destroy(&dd->dd_lock);
319 	kmem_free(dd, sizeof (dsl_dir_t));
320 	dmu_buf_rele(dbuf, tag);
321 	return (err);
322 }
323 
324 void
325 dsl_dir_rele(dsl_dir_t *dd, void *tag)
326 {
327 	dprintf_dd(dd, "%s\n", "");
328 	spa_close(dd->dd_pool->dp_spa, tag);
329 	dmu_buf_rele(dd->dd_dbuf, tag);
330 }
331 
332 /*
333  * Remove a reference to the given dsl dir that is being asynchronously
334  * released.  Async releases occur from a taskq performing eviction of
335  * dsl datasets and dirs.  This process is identical to a normal release
336  * with the exception of using the async API for releasing the reference on
337  * the spa.
338  */
339 void
340 dsl_dir_async_rele(dsl_dir_t *dd, void *tag)
341 {
342 	dprintf_dd(dd, "%s\n", "");
343 	spa_async_close(dd->dd_pool->dp_spa, tag);
344 	dmu_buf_rele(dd->dd_dbuf, tag);
345 }
346 
347 /* buf must be at least ZFS_MAX_DATASET_NAME_LEN bytes */
348 void
349 dsl_dir_name(dsl_dir_t *dd, char *buf)
350 {
351 	if (dd->dd_parent) {
352 		dsl_dir_name(dd->dd_parent, buf);
353 		VERIFY3U(strlcat(buf, "/", ZFS_MAX_DATASET_NAME_LEN), <,
354 		    ZFS_MAX_DATASET_NAME_LEN);
355 	} else {
356 		buf[0] = '\0';
357 	}
358 	if (!MUTEX_HELD(&dd->dd_lock)) {
359 		/*
360 		 * recursive mutex so that we can use
361 		 * dprintf_dd() with dd_lock held
362 		 */
363 		mutex_enter(&dd->dd_lock);
364 		VERIFY3U(strlcat(buf, dd->dd_myname, ZFS_MAX_DATASET_NAME_LEN),
365 		    <, ZFS_MAX_DATASET_NAME_LEN);
366 		mutex_exit(&dd->dd_lock);
367 	} else {
368 		VERIFY3U(strlcat(buf, dd->dd_myname, ZFS_MAX_DATASET_NAME_LEN),
369 		    <, ZFS_MAX_DATASET_NAME_LEN);
370 	}
371 }
372 
373 /* Calculate name length, avoiding all the strcat calls of dsl_dir_name */
374 int
375 dsl_dir_namelen(dsl_dir_t *dd)
376 {
377 	int result = 0;
378 
379 	if (dd->dd_parent) {
380 		/* parent's name + 1 for the "/" */
381 		result = dsl_dir_namelen(dd->dd_parent) + 1;
382 	}
383 
384 	if (!MUTEX_HELD(&dd->dd_lock)) {
385 		/* see dsl_dir_name */
386 		mutex_enter(&dd->dd_lock);
387 		result += strlen(dd->dd_myname);
388 		mutex_exit(&dd->dd_lock);
389 	} else {
390 		result += strlen(dd->dd_myname);
391 	}
392 
393 	return (result);
394 }
395 
396 static int
397 getcomponent(const char *path, char *component, const char **nextp)
398 {
399 	char *p;
400 
401 	if ((path == NULL) || (path[0] == '\0'))
402 		return (SET_ERROR(ENOENT));
403 	/* This would be a good place to reserve some namespace... */
404 	p = strpbrk(path, "/@");
405 	if (p && (p[1] == '/' || p[1] == '@')) {
406 		/* two separators in a row */
407 		return (SET_ERROR(EINVAL));
408 	}
409 	if (p == NULL || p == path) {
410 		/*
411 		 * if the first thing is an @ or /, it had better be an
412 		 * @ and it had better not have any more ats or slashes,
413 		 * and it had better have something after the @.
414 		 */
415 		if (p != NULL &&
416 		    (p[0] != '@' || strpbrk(path+1, "/@") || p[1] == '\0'))
417 			return (SET_ERROR(EINVAL));
418 		if (strlen(path) >= ZFS_MAX_DATASET_NAME_LEN)
419 			return (SET_ERROR(ENAMETOOLONG));
420 		(void) strlcpy(component, path, ZFS_MAX_DATASET_NAME_LEN);
421 		p = NULL;
422 	} else if (p[0] == '/') {
423 		if (p - path >= ZFS_MAX_DATASET_NAME_LEN)
424 			return (SET_ERROR(ENAMETOOLONG));
425 		(void) strncpy(component, path, p - path);
426 		component[p - path] = '\0';
427 		p++;
428 	} else if (p[0] == '@') {
429 		/*
430 		 * if the next separator is an @, there better not be
431 		 * any more slashes.
432 		 */
433 		if (strchr(path, '/'))
434 			return (SET_ERROR(EINVAL));
435 		if (p - path >= ZFS_MAX_DATASET_NAME_LEN)
436 			return (SET_ERROR(ENAMETOOLONG));
437 		(void) strncpy(component, path, p - path);
438 		component[p - path] = '\0';
439 	} else {
440 		panic("invalid p=%p", (void *)p);
441 	}
442 	*nextp = p;
443 	return (0);
444 }
445 
446 /*
447  * Return the dsl_dir_t, and possibly the last component which couldn't
448  * be found in *tail.  The name must be in the specified dsl_pool_t.  This
449  * thread must hold the dp_config_rwlock for the pool.  Returns NULL if the
450  * path is bogus, or if tail==NULL and we couldn't parse the whole name.
451  * (*tail)[0] == '@' means that the last component is a snapshot.
452  */
453 int
454 dsl_dir_hold(dsl_pool_t *dp, const char *name, void *tag,
455     dsl_dir_t **ddp, const char **tailp)
456 {
457 	char *buf;
458 	const char *spaname, *next, *nextnext = NULL;
459 	int err;
460 	dsl_dir_t *dd;
461 	uint64_t ddobj;
462 
463 	buf = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
464 	err = getcomponent(name, buf, &next);
465 	if (err != 0)
466 		goto error;
467 
468 	/* Make sure the name is in the specified pool. */
469 	spaname = spa_name(dp->dp_spa);
470 	if (strcmp(buf, spaname) != 0) {
471 		err = SET_ERROR(EXDEV);
472 		goto error;
473 	}
474 
475 	ASSERT(dsl_pool_config_held(dp));
476 
477 	err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj, NULL, tag, &dd);
478 	if (err != 0) {
479 		goto error;
480 	}
481 
482 	while (next != NULL) {
483 		dsl_dir_t *child_dd;
484 		err = getcomponent(next, buf, &nextnext);
485 		if (err != 0)
486 			break;
487 		ASSERT(next[0] != '\0');
488 		if (next[0] == '@')
489 			break;
490 		dprintf("looking up %s in obj%lld\n",
491 		    buf, (longlong_t)dsl_dir_phys(dd)->dd_child_dir_zapobj);
492 
493 		err = zap_lookup(dp->dp_meta_objset,
494 		    dsl_dir_phys(dd)->dd_child_dir_zapobj,
495 		    buf, sizeof (ddobj), 1, &ddobj);
496 		if (err != 0) {
497 			if (err == ENOENT)
498 				err = 0;
499 			break;
500 		}
501 
502 		err = dsl_dir_hold_obj(dp, ddobj, buf, tag, &child_dd);
503 		if (err != 0)
504 			break;
505 		dsl_dir_rele(dd, tag);
506 		dd = child_dd;
507 		next = nextnext;
508 	}
509 
510 	if (err != 0) {
511 		dsl_dir_rele(dd, tag);
512 		goto error;
513 	}
514 
515 	/*
516 	 * It's an error if there's more than one component left, or
517 	 * tailp==NULL and there's any component left.
518 	 */
519 	if (next != NULL &&
520 	    (tailp == NULL || (nextnext && nextnext[0] != '\0'))) {
521 		/* bad path name */
522 		dsl_dir_rele(dd, tag);
523 		dprintf("next=%p (%s) tail=%p\n", next, next?next:"", tailp);
524 		err = SET_ERROR(ENOENT);
525 	}
526 	if (tailp != NULL)
527 		*tailp = next;
528 	if (err == 0)
529 		*ddp = dd;
530 error:
531 	kmem_free(buf, ZFS_MAX_DATASET_NAME_LEN);
532 	return (err);
533 }
534 
535 /*
536  * If the counts are already initialized for this filesystem and its
537  * descendants then do nothing, otherwise initialize the counts.
538  *
539  * The counts on this filesystem, and those below, may be uninitialized due to
540  * either the use of a pre-existing pool which did not support the
541  * filesystem/snapshot limit feature, or one in which the feature had not yet
542  * been enabled.
543  *
544  * Recursively descend the filesystem tree and update the filesystem/snapshot
545  * counts on each filesystem below, then update the cumulative count on the
546  * current filesystem. If the filesystem already has a count set on it,
547  * then we know that its counts, and the counts on the filesystems below it,
548  * are already correct, so we don't have to update this filesystem.
549  */
550 static void
551 dsl_dir_init_fs_ss_count(dsl_dir_t *dd, dmu_tx_t *tx)
552 {
553 	uint64_t my_fs_cnt = 0;
554 	uint64_t my_ss_cnt = 0;
555 	dsl_pool_t *dp = dd->dd_pool;
556 	objset_t *os = dp->dp_meta_objset;
557 	zap_cursor_t *zc;
558 	zap_attribute_t *za;
559 	dsl_dataset_t *ds;
560 
561 	ASSERT(spa_feature_is_active(dp->dp_spa, SPA_FEATURE_FS_SS_LIMIT));
562 	ASSERT(dsl_pool_config_held(dp));
563 	ASSERT(dmu_tx_is_syncing(tx));
564 
565 	dsl_dir_zapify(dd, tx);
566 
567 	/*
568 	 * If the filesystem count has already been initialized then we
569 	 * don't need to recurse down any further.
570 	 */
571 	if (zap_contains(os, dd->dd_object, DD_FIELD_FILESYSTEM_COUNT) == 0)
572 		return;
573 
574 	zc = kmem_alloc(sizeof (zap_cursor_t), KM_SLEEP);
575 	za = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
576 
577 	/* Iterate my child dirs */
578 	for (zap_cursor_init(zc, os, dsl_dir_phys(dd)->dd_child_dir_zapobj);
579 	    zap_cursor_retrieve(zc, za) == 0; zap_cursor_advance(zc)) {
580 		dsl_dir_t *chld_dd;
581 		uint64_t count;
582 
583 		VERIFY0(dsl_dir_hold_obj(dp, za->za_first_integer, NULL, FTAG,
584 		    &chld_dd));
585 
586 		/*
587 		 * Ignore hidden ($FREE, $MOS & $ORIGIN) objsets.
588 		 */
589 		if (chld_dd->dd_myname[0] == '$') {
590 			dsl_dir_rele(chld_dd, FTAG);
591 			continue;
592 		}
593 
594 		my_fs_cnt++;	/* count this child */
595 
596 		dsl_dir_init_fs_ss_count(chld_dd, tx);
597 
598 		VERIFY0(zap_lookup(os, chld_dd->dd_object,
599 		    DD_FIELD_FILESYSTEM_COUNT, sizeof (count), 1, &count));
600 		my_fs_cnt += count;
601 		VERIFY0(zap_lookup(os, chld_dd->dd_object,
602 		    DD_FIELD_SNAPSHOT_COUNT, sizeof (count), 1, &count));
603 		my_ss_cnt += count;
604 
605 		dsl_dir_rele(chld_dd, FTAG);
606 	}
607 	zap_cursor_fini(zc);
608 	/* Count my snapshots (we counted children's snapshots above) */
609 	VERIFY0(dsl_dataset_hold_obj(dd->dd_pool,
610 	    dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds));
611 
612 	for (zap_cursor_init(zc, os, dsl_dataset_phys(ds)->ds_snapnames_zapobj);
613 	    zap_cursor_retrieve(zc, za) == 0;
614 	    zap_cursor_advance(zc)) {
615 		/* Don't count temporary snapshots */
616 		if (za->za_name[0] != '%')
617 			my_ss_cnt++;
618 	}
619 	zap_cursor_fini(zc);
620 
621 	dsl_dataset_rele(ds, FTAG);
622 
623 	kmem_free(zc, sizeof (zap_cursor_t));
624 	kmem_free(za, sizeof (zap_attribute_t));
625 
626 	/* we're in a sync task, update counts */
627 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
628 	VERIFY0(zap_add(os, dd->dd_object, DD_FIELD_FILESYSTEM_COUNT,
629 	    sizeof (my_fs_cnt), 1, &my_fs_cnt, tx));
630 	VERIFY0(zap_add(os, dd->dd_object, DD_FIELD_SNAPSHOT_COUNT,
631 	    sizeof (my_ss_cnt), 1, &my_ss_cnt, tx));
632 }
633 
634 static int
635 dsl_dir_actv_fs_ss_limit_check(void *arg, dmu_tx_t *tx)
636 {
637 	char *ddname = (char *)arg;
638 	dsl_pool_t *dp = dmu_tx_pool(tx);
639 	dsl_dataset_t *ds;
640 	dsl_dir_t *dd;
641 	int error;
642 
643 	error = dsl_dataset_hold(dp, ddname, FTAG, &ds);
644 	if (error != 0)
645 		return (error);
646 
647 	if (!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_FS_SS_LIMIT)) {
648 		dsl_dataset_rele(ds, FTAG);
649 		return (SET_ERROR(ENOTSUP));
650 	}
651 
652 	dd = ds->ds_dir;
653 	if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_FS_SS_LIMIT) &&
654 	    dsl_dir_is_zapified(dd) &&
655 	    zap_contains(dp->dp_meta_objset, dd->dd_object,
656 	    DD_FIELD_FILESYSTEM_COUNT) == 0) {
657 		dsl_dataset_rele(ds, FTAG);
658 		return (SET_ERROR(EALREADY));
659 	}
660 
661 	dsl_dataset_rele(ds, FTAG);
662 	return (0);
663 }
664 
665 static void
666 dsl_dir_actv_fs_ss_limit_sync(void *arg, dmu_tx_t *tx)
667 {
668 	char *ddname = (char *)arg;
669 	dsl_pool_t *dp = dmu_tx_pool(tx);
670 	dsl_dataset_t *ds;
671 	spa_t *spa;
672 
673 	VERIFY0(dsl_dataset_hold(dp, ddname, FTAG, &ds));
674 
675 	spa = dsl_dataset_get_spa(ds);
676 
677 	if (!spa_feature_is_active(spa, SPA_FEATURE_FS_SS_LIMIT)) {
678 		/*
679 		 * Since the feature was not active and we're now setting a
680 		 * limit, increment the feature-active counter so that the
681 		 * feature becomes active for the first time.
682 		 *
683 		 * We are already in a sync task so we can update the MOS.
684 		 */
685 		spa_feature_incr(spa, SPA_FEATURE_FS_SS_LIMIT, tx);
686 	}
687 
688 	/*
689 	 * Since we are now setting a non-UINT64_MAX limit on the filesystem,
690 	 * we need to ensure the counts are correct. Descend down the tree from
691 	 * this point and update all of the counts to be accurate.
692 	 */
693 	dsl_dir_init_fs_ss_count(ds->ds_dir, tx);
694 
695 	dsl_dataset_rele(ds, FTAG);
696 }
697 
698 /*
699  * Make sure the feature is enabled and activate it if necessary.
700  * Since we're setting a limit, ensure the on-disk counts are valid.
701  * This is only called by the ioctl path when setting a limit value.
702  *
703  * We do not need to validate the new limit, since users who can change the
704  * limit are also allowed to exceed the limit.
705  */
706 int
707 dsl_dir_activate_fs_ss_limit(const char *ddname)
708 {
709 	int error;
710 
711 	error = dsl_sync_task(ddname, dsl_dir_actv_fs_ss_limit_check,
712 	    dsl_dir_actv_fs_ss_limit_sync, (void *)ddname, 0,
713 	    ZFS_SPACE_CHECK_RESERVED);
714 
715 	if (error == EALREADY)
716 		error = 0;
717 
718 	return (error);
719 }
720 
721 /*
722  * Used to determine if the filesystem_limit or snapshot_limit should be
723  * enforced. We allow the limit to be exceeded if the user has permission to
724  * write the property value. We pass in the creds that we got in the open
725  * context since we will always be the GZ root in syncing context. We also have
726  * to handle the case where we are allowed to change the limit on the current
727  * dataset, but there may be another limit in the tree above.
728  *
729  * We can never modify these two properties within a non-global zone. In
730  * addition, the other checks are modeled on zfs_secpolicy_write_perms. We
731  * can't use that function since we are already holding the dp_config_rwlock.
732  * In addition, we already have the dd and dealing with snapshots is simplified
733  * in this code.
734  */
735 
736 typedef enum {
737 	ENFORCE_ALWAYS,
738 	ENFORCE_NEVER,
739 	ENFORCE_ABOVE
740 } enforce_res_t;
741 
742 static enforce_res_t
743 dsl_enforce_ds_ss_limits(dsl_dir_t *dd, zfs_prop_t prop,
744     cred_t *cr, proc_t *proc)
745 {
746 	enforce_res_t enforce = ENFORCE_ALWAYS;
747 	uint64_t obj;
748 	dsl_dataset_t *ds;
749 	uint64_t zoned;
750 	const char *zonedstr;
751 
752 	ASSERT(prop == ZFS_PROP_FILESYSTEM_LIMIT ||
753 	    prop == ZFS_PROP_SNAPSHOT_LIMIT);
754 
755 #ifdef _KERNEL
756 	if (crgetzoneid(cr) != GLOBAL_ZONEID)
757 		return (ENFORCE_ALWAYS);
758 
759 	/*
760 	 * We are checking the saved credentials of the user process, which is
761 	 * not the current process.  Note that we can't use secpolicy_zfs(),
762 	 * because it only works if the cred is that of the current process (on
763 	 * Linux).
764 	 */
765 	if (secpolicy_zfs_proc(cr, proc) == 0)
766 		return (ENFORCE_NEVER);
767 #endif
768 
769 	if ((obj = dsl_dir_phys(dd)->dd_head_dataset_obj) == 0)
770 		return (ENFORCE_ALWAYS);
771 
772 	ASSERT(dsl_pool_config_held(dd->dd_pool));
773 
774 	if (dsl_dataset_hold_obj(dd->dd_pool, obj, FTAG, &ds) != 0)
775 		return (ENFORCE_ALWAYS);
776 
777 	zonedstr = zfs_prop_to_name(ZFS_PROP_ZONED);
778 	if (dsl_prop_get_ds(ds, zonedstr, 8, 1, &zoned, NULL) || zoned) {
779 		/* Only root can access zoned fs's from the GZ */
780 		enforce = ENFORCE_ALWAYS;
781 	} else {
782 		if (dsl_deleg_access_impl(ds, zfs_prop_to_name(prop), cr) == 0)
783 			enforce = ENFORCE_ABOVE;
784 	}
785 
786 	dsl_dataset_rele(ds, FTAG);
787 	return (enforce);
788 }
789 
790 /*
791  * Check if adding additional child filesystem(s) would exceed any filesystem
792  * limits or adding additional snapshot(s) would exceed any snapshot limits.
793  * The prop argument indicates which limit to check.
794  *
795  * Note that all filesystem limits up to the root (or the highest
796  * initialized) filesystem or the given ancestor must be satisfied.
797  */
798 int
799 dsl_fs_ss_limit_check(dsl_dir_t *dd, uint64_t delta, zfs_prop_t prop,
800     dsl_dir_t *ancestor, cred_t *cr, proc_t *proc)
801 {
802 	objset_t *os = dd->dd_pool->dp_meta_objset;
803 	uint64_t limit, count;
804 	char *count_prop;
805 	enforce_res_t enforce;
806 	int err = 0;
807 
808 	ASSERT(dsl_pool_config_held(dd->dd_pool));
809 	ASSERT(prop == ZFS_PROP_FILESYSTEM_LIMIT ||
810 	    prop == ZFS_PROP_SNAPSHOT_LIMIT);
811 
812 	/*
813 	 * If we're allowed to change the limit, don't enforce the limit
814 	 * e.g. this can happen if a snapshot is taken by an administrative
815 	 * user in the global zone (i.e. a recursive snapshot by root).
816 	 * However, we must handle the case of delegated permissions where we
817 	 * are allowed to change the limit on the current dataset, but there
818 	 * is another limit in the tree above.
819 	 */
820 	enforce = dsl_enforce_ds_ss_limits(dd, prop, cr, proc);
821 	if (enforce == ENFORCE_NEVER)
822 		return (0);
823 
824 	/*
825 	 * e.g. if renaming a dataset with no snapshots, count adjustment
826 	 * is 0.
827 	 */
828 	if (delta == 0)
829 		return (0);
830 
831 	if (prop == ZFS_PROP_SNAPSHOT_LIMIT) {
832 		/*
833 		 * We don't enforce the limit for temporary snapshots. This is
834 		 * indicated by a NULL cred_t argument.
835 		 */
836 		if (cr == NULL)
837 			return (0);
838 
839 		count_prop = DD_FIELD_SNAPSHOT_COUNT;
840 	} else {
841 		count_prop = DD_FIELD_FILESYSTEM_COUNT;
842 	}
843 
844 	/*
845 	 * If an ancestor has been provided, stop checking the limit once we
846 	 * hit that dir. We need this during rename so that we don't overcount
847 	 * the check once we recurse up to the common ancestor.
848 	 */
849 	if (ancestor == dd)
850 		return (0);
851 
852 	/*
853 	 * If we hit an uninitialized node while recursing up the tree, we can
854 	 * stop since we know there is no limit here (or above). The counts are
855 	 * not valid on this node and we know we won't touch this node's counts.
856 	 */
857 	if (!dsl_dir_is_zapified(dd))
858 		return (0);
859 	err = zap_lookup(os, dd->dd_object,
860 	    count_prop, sizeof (count), 1, &count);
861 	if (err == ENOENT)
862 		return (0);
863 	if (err != 0)
864 		return (err);
865 
866 	err = dsl_prop_get_dd(dd, zfs_prop_to_name(prop), 8, 1, &limit, NULL,
867 	    B_FALSE);
868 	if (err != 0)
869 		return (err);
870 
871 	/* Is there a limit which we've hit? */
872 	if (enforce == ENFORCE_ALWAYS && (count + delta) > limit)
873 		return (SET_ERROR(EDQUOT));
874 
875 	if (dd->dd_parent != NULL)
876 		err = dsl_fs_ss_limit_check(dd->dd_parent, delta, prop,
877 		    ancestor, cr, proc);
878 
879 	return (err);
880 }
881 
882 /*
883  * Adjust the filesystem or snapshot count for the specified dsl_dir_t and all
884  * parents. When a new filesystem/snapshot is created, increment the count on
885  * all parents, and when a filesystem/snapshot is destroyed, decrement the
886  * count.
887  */
888 void
889 dsl_fs_ss_count_adjust(dsl_dir_t *dd, int64_t delta, const char *prop,
890     dmu_tx_t *tx)
891 {
892 	int err;
893 	objset_t *os = dd->dd_pool->dp_meta_objset;
894 	uint64_t count;
895 
896 	ASSERT(dsl_pool_config_held(dd->dd_pool));
897 	ASSERT(dmu_tx_is_syncing(tx));
898 	ASSERT(strcmp(prop, DD_FIELD_FILESYSTEM_COUNT) == 0 ||
899 	    strcmp(prop, DD_FIELD_SNAPSHOT_COUNT) == 0);
900 
901 	/*
902 	 * We don't do accounting for hidden ($FREE, $MOS & $ORIGIN) objsets.
903 	 */
904 	if (dd->dd_myname[0] == '$' && strcmp(prop,
905 	    DD_FIELD_FILESYSTEM_COUNT) == 0) {
906 		return;
907 	}
908 
909 	/*
910 	 * e.g. if renaming a dataset with no snapshots, count adjustment is 0
911 	 */
912 	if (delta == 0)
913 		return;
914 
915 	/*
916 	 * If we hit an uninitialized node while recursing up the tree, we can
917 	 * stop since we know the counts are not valid on this node and we
918 	 * know we shouldn't touch this node's counts. An uninitialized count
919 	 * on the node indicates that either the feature has not yet been
920 	 * activated or there are no limits on this part of the tree.
921 	 */
922 	if (!dsl_dir_is_zapified(dd) || (err = zap_lookup(os, dd->dd_object,
923 	    prop, sizeof (count), 1, &count)) == ENOENT)
924 		return;
925 	VERIFY0(err);
926 
927 	count += delta;
928 	/* Use a signed verify to make sure we're not neg. */
929 	VERIFY3S(count, >=, 0);
930 
931 	VERIFY0(zap_update(os, dd->dd_object, prop, sizeof (count), 1, &count,
932 	    tx));
933 
934 	/* Roll up this additional count into our ancestors */
935 	if (dd->dd_parent != NULL)
936 		dsl_fs_ss_count_adjust(dd->dd_parent, delta, prop, tx);
937 }
938 
939 uint64_t
940 dsl_dir_create_sync(dsl_pool_t *dp, dsl_dir_t *pds, const char *name,
941     dmu_tx_t *tx)
942 {
943 	objset_t *mos = dp->dp_meta_objset;
944 	uint64_t ddobj;
945 	dsl_dir_phys_t *ddphys;
946 	dmu_buf_t *dbuf;
947 
948 	ddobj = dmu_object_alloc(mos, DMU_OT_DSL_DIR, 0,
949 	    DMU_OT_DSL_DIR, sizeof (dsl_dir_phys_t), tx);
950 	if (pds) {
951 		VERIFY0(zap_add(mos, dsl_dir_phys(pds)->dd_child_dir_zapobj,
952 		    name, sizeof (uint64_t), 1, &ddobj, tx));
953 	} else {
954 		/* it's the root dir */
955 		VERIFY0(zap_add(mos, DMU_POOL_DIRECTORY_OBJECT,
956 		    DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1, &ddobj, tx));
957 	}
958 	VERIFY0(dmu_bonus_hold(mos, ddobj, FTAG, &dbuf));
959 	dmu_buf_will_dirty(dbuf, tx);
960 	ddphys = dbuf->db_data;
961 
962 	ddphys->dd_creation_time = gethrestime_sec();
963 	if (pds) {
964 		ddphys->dd_parent_obj = pds->dd_object;
965 
966 		/* update the filesystem counts */
967 		dsl_fs_ss_count_adjust(pds, 1, DD_FIELD_FILESYSTEM_COUNT, tx);
968 	}
969 	ddphys->dd_props_zapobj = zap_create(mos,
970 	    DMU_OT_DSL_PROPS, DMU_OT_NONE, 0, tx);
971 	ddphys->dd_child_dir_zapobj = zap_create(mos,
972 	    DMU_OT_DSL_DIR_CHILD_MAP, DMU_OT_NONE, 0, tx);
973 	if (spa_version(dp->dp_spa) >= SPA_VERSION_USED_BREAKDOWN)
974 		ddphys->dd_flags |= DD_FLAG_USED_BREAKDOWN;
975 
976 	dmu_buf_rele(dbuf, FTAG);
977 
978 	return (ddobj);
979 }
980 
981 boolean_t
982 dsl_dir_is_clone(dsl_dir_t *dd)
983 {
984 	return (dsl_dir_phys(dd)->dd_origin_obj &&
985 	    (dd->dd_pool->dp_origin_snap == NULL ||
986 	    dsl_dir_phys(dd)->dd_origin_obj !=
987 	    dd->dd_pool->dp_origin_snap->ds_object));
988 }
989 
990 uint64_t
991 dsl_dir_get_used(dsl_dir_t *dd)
992 {
993 	return (dsl_dir_phys(dd)->dd_used_bytes);
994 }
995 
996 uint64_t
997 dsl_dir_get_compressed(dsl_dir_t *dd)
998 {
999 	return (dsl_dir_phys(dd)->dd_compressed_bytes);
1000 }
1001 
1002 uint64_t
1003 dsl_dir_get_quota(dsl_dir_t *dd)
1004 {
1005 	return (dsl_dir_phys(dd)->dd_quota);
1006 }
1007 
1008 uint64_t
1009 dsl_dir_get_reservation(dsl_dir_t *dd)
1010 {
1011 	return (dsl_dir_phys(dd)->dd_reserved);
1012 }
1013 
1014 uint64_t
1015 dsl_dir_get_compressratio(dsl_dir_t *dd)
1016 {
1017 	/* a fixed point number, 100x the ratio */
1018 	return (dsl_dir_phys(dd)->dd_compressed_bytes == 0 ? 100 :
1019 	    (dsl_dir_phys(dd)->dd_uncompressed_bytes * 100 /
1020 	    dsl_dir_phys(dd)->dd_compressed_bytes));
1021 }
1022 
1023 uint64_t
1024 dsl_dir_get_logicalused(dsl_dir_t *dd)
1025 {
1026 	return (dsl_dir_phys(dd)->dd_uncompressed_bytes);
1027 }
1028 
1029 uint64_t
1030 dsl_dir_get_usedsnap(dsl_dir_t *dd)
1031 {
1032 	return (dsl_dir_phys(dd)->dd_used_breakdown[DD_USED_SNAP]);
1033 }
1034 
1035 uint64_t
1036 dsl_dir_get_usedds(dsl_dir_t *dd)
1037 {
1038 	return (dsl_dir_phys(dd)->dd_used_breakdown[DD_USED_HEAD]);
1039 }
1040 
1041 uint64_t
1042 dsl_dir_get_usedrefreserv(dsl_dir_t *dd)
1043 {
1044 	return (dsl_dir_phys(dd)->dd_used_breakdown[DD_USED_REFRSRV]);
1045 }
1046 
1047 uint64_t
1048 dsl_dir_get_usedchild(dsl_dir_t *dd)
1049 {
1050 	return (dsl_dir_phys(dd)->dd_used_breakdown[DD_USED_CHILD] +
1051 	    dsl_dir_phys(dd)->dd_used_breakdown[DD_USED_CHILD_RSRV]);
1052 }
1053 
1054 void
1055 dsl_dir_get_origin(dsl_dir_t *dd, char *buf)
1056 {
1057 	dsl_dataset_t *ds;
1058 	VERIFY0(dsl_dataset_hold_obj(dd->dd_pool,
1059 	    dsl_dir_phys(dd)->dd_origin_obj, FTAG, &ds));
1060 
1061 	dsl_dataset_name(ds, buf);
1062 
1063 	dsl_dataset_rele(ds, FTAG);
1064 }
1065 
1066 int
1067 dsl_dir_get_filesystem_count(dsl_dir_t *dd, uint64_t *count)
1068 {
1069 	if (dsl_dir_is_zapified(dd)) {
1070 		objset_t *os = dd->dd_pool->dp_meta_objset;
1071 		return (zap_lookup(os, dd->dd_object, DD_FIELD_FILESYSTEM_COUNT,
1072 		    sizeof (*count), 1, count));
1073 	} else {
1074 		return (SET_ERROR(ENOENT));
1075 	}
1076 }
1077 
1078 int
1079 dsl_dir_get_snapshot_count(dsl_dir_t *dd, uint64_t *count)
1080 {
1081 	if (dsl_dir_is_zapified(dd)) {
1082 		objset_t *os = dd->dd_pool->dp_meta_objset;
1083 		return (zap_lookup(os, dd->dd_object, DD_FIELD_SNAPSHOT_COUNT,
1084 		    sizeof (*count), 1, count));
1085 	} else {
1086 		return (SET_ERROR(ENOENT));
1087 	}
1088 }
1089 
1090 void
1091 dsl_dir_stats(dsl_dir_t *dd, nvlist_t *nv)
1092 {
1093 	mutex_enter(&dd->dd_lock);
1094 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_QUOTA,
1095 	    dsl_dir_get_quota(dd));
1096 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_RESERVATION,
1097 	    dsl_dir_get_reservation(dd));
1098 	dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_LOGICALUSED,
1099 	    dsl_dir_get_logicalused(dd));
1100 	if (dsl_dir_phys(dd)->dd_flags & DD_FLAG_USED_BREAKDOWN) {
1101 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USEDSNAP,
1102 		    dsl_dir_get_usedsnap(dd));
1103 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USEDDS,
1104 		    dsl_dir_get_usedds(dd));
1105 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USEDREFRESERV,
1106 		    dsl_dir_get_usedrefreserv(dd));
1107 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USEDCHILD,
1108 		    dsl_dir_get_usedchild(dd));
1109 	}
1110 	mutex_exit(&dd->dd_lock);
1111 
1112 	uint64_t count;
1113 	if (dsl_dir_get_filesystem_count(dd, &count) == 0) {
1114 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_FILESYSTEM_COUNT,
1115 		    count);
1116 	}
1117 	if (dsl_dir_get_snapshot_count(dd, &count) == 0) {
1118 		dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_SNAPSHOT_COUNT,
1119 		    count);
1120 	}
1121 
1122 	if (dsl_dir_is_clone(dd)) {
1123 		char buf[ZFS_MAX_DATASET_NAME_LEN];
1124 		dsl_dir_get_origin(dd, buf);
1125 		dsl_prop_nvlist_add_string(nv, ZFS_PROP_ORIGIN, buf);
1126 	}
1127 
1128 }
1129 
1130 void
1131 dsl_dir_dirty(dsl_dir_t *dd, dmu_tx_t *tx)
1132 {
1133 	dsl_pool_t *dp = dd->dd_pool;
1134 
1135 	ASSERT(dsl_dir_phys(dd));
1136 
1137 	if (txg_list_add(&dp->dp_dirty_dirs, dd, tx->tx_txg)) {
1138 		/* up the hold count until we can be written out */
1139 		dmu_buf_add_ref(dd->dd_dbuf, dd);
1140 	}
1141 }
1142 
1143 static int64_t
1144 parent_delta(dsl_dir_t *dd, uint64_t used, int64_t delta)
1145 {
1146 	uint64_t old_accounted = MAX(used, dsl_dir_phys(dd)->dd_reserved);
1147 	uint64_t new_accounted =
1148 	    MAX(used + delta, dsl_dir_phys(dd)->dd_reserved);
1149 	return (new_accounted - old_accounted);
1150 }
1151 
1152 void
1153 dsl_dir_sync(dsl_dir_t *dd, dmu_tx_t *tx)
1154 {
1155 	ASSERT(dmu_tx_is_syncing(tx));
1156 
1157 	mutex_enter(&dd->dd_lock);
1158 	ASSERT0(dd->dd_tempreserved[tx->tx_txg & TXG_MASK]);
1159 	dprintf_dd(dd, "txg=%llu towrite=%lluK\n", (u_longlong_t)tx->tx_txg,
1160 	    (u_longlong_t)dd->dd_space_towrite[tx->tx_txg & TXG_MASK] / 1024);
1161 	dd->dd_space_towrite[tx->tx_txg & TXG_MASK] = 0;
1162 	mutex_exit(&dd->dd_lock);
1163 
1164 	/* release the hold from dsl_dir_dirty */
1165 	dmu_buf_rele(dd->dd_dbuf, dd);
1166 }
1167 
1168 static uint64_t
1169 dsl_dir_space_towrite(dsl_dir_t *dd)
1170 {
1171 	uint64_t space = 0;
1172 
1173 	ASSERT(MUTEX_HELD(&dd->dd_lock));
1174 
1175 	for (int i = 0; i < TXG_SIZE; i++) {
1176 		space += dd->dd_space_towrite[i & TXG_MASK];
1177 		ASSERT3U(dd->dd_space_towrite[i & TXG_MASK], >=, 0);
1178 	}
1179 	return (space);
1180 }
1181 
1182 /*
1183  * How much space would dd have available if ancestor had delta applied
1184  * to it?  If ondiskonly is set, we're only interested in what's
1185  * on-disk, not estimated pending changes.
1186  */
1187 uint64_t
1188 dsl_dir_space_available(dsl_dir_t *dd,
1189     dsl_dir_t *ancestor, int64_t delta, int ondiskonly)
1190 {
1191 	uint64_t parentspace, myspace, quota, used;
1192 
1193 	/*
1194 	 * If there are no restrictions otherwise, assume we have
1195 	 * unlimited space available.
1196 	 */
1197 	quota = UINT64_MAX;
1198 	parentspace = UINT64_MAX;
1199 
1200 	if (dd->dd_parent != NULL) {
1201 		parentspace = dsl_dir_space_available(dd->dd_parent,
1202 		    ancestor, delta, ondiskonly);
1203 	}
1204 
1205 	mutex_enter(&dd->dd_lock);
1206 	if (dsl_dir_phys(dd)->dd_quota != 0)
1207 		quota = dsl_dir_phys(dd)->dd_quota;
1208 	used = dsl_dir_phys(dd)->dd_used_bytes;
1209 	if (!ondiskonly)
1210 		used += dsl_dir_space_towrite(dd);
1211 
1212 	if (dd->dd_parent == NULL) {
1213 		uint64_t poolsize = dsl_pool_adjustedsize(dd->dd_pool,
1214 		    ZFS_SPACE_CHECK_NORMAL);
1215 		quota = MIN(quota, poolsize);
1216 	}
1217 
1218 	if (dsl_dir_phys(dd)->dd_reserved > used && parentspace != UINT64_MAX) {
1219 		/*
1220 		 * We have some space reserved, in addition to what our
1221 		 * parent gave us.
1222 		 */
1223 		parentspace += dsl_dir_phys(dd)->dd_reserved - used;
1224 	}
1225 
1226 	if (dd == ancestor) {
1227 		ASSERT(delta <= 0);
1228 		ASSERT(used >= -delta);
1229 		used += delta;
1230 		if (parentspace != UINT64_MAX)
1231 			parentspace -= delta;
1232 	}
1233 
1234 	if (used > quota) {
1235 		/* over quota */
1236 		myspace = 0;
1237 	} else {
1238 		/*
1239 		 * the lesser of the space provided by our parent and
1240 		 * the space left in our quota
1241 		 */
1242 		myspace = MIN(parentspace, quota - used);
1243 	}
1244 
1245 	mutex_exit(&dd->dd_lock);
1246 
1247 	return (myspace);
1248 }
1249 
1250 struct tempreserve {
1251 	list_node_t tr_node;
1252 	dsl_dir_t *tr_ds;
1253 	uint64_t tr_size;
1254 };
1255 
1256 static int
1257 dsl_dir_tempreserve_impl(dsl_dir_t *dd, uint64_t asize, boolean_t netfree,
1258     boolean_t ignorequota, list_t *tr_list,
1259     dmu_tx_t *tx, boolean_t first)
1260 {
1261 	uint64_t txg;
1262 	uint64_t quota;
1263 	struct tempreserve *tr;
1264 	int retval;
1265 	uint64_t ref_rsrv;
1266 
1267 top_of_function:
1268 	txg = tx->tx_txg;
1269 	retval = EDQUOT;
1270 	ref_rsrv = 0;
1271 
1272 	ASSERT3U(txg, !=, 0);
1273 	ASSERT3S(asize, >, 0);
1274 
1275 	mutex_enter(&dd->dd_lock);
1276 
1277 	/*
1278 	 * Check against the dsl_dir's quota.  We don't add in the delta
1279 	 * when checking for over-quota because they get one free hit.
1280 	 */
1281 	uint64_t est_inflight = dsl_dir_space_towrite(dd);
1282 	for (int i = 0; i < TXG_SIZE; i++)
1283 		est_inflight += dd->dd_tempreserved[i];
1284 	uint64_t used_on_disk = dsl_dir_phys(dd)->dd_used_bytes;
1285 
1286 	/*
1287 	 * On the first iteration, fetch the dataset's used-on-disk and
1288 	 * refreservation values. Also, if checkrefquota is set, test if
1289 	 * allocating this space would exceed the dataset's refquota.
1290 	 */
1291 	if (first && tx->tx_objset) {
1292 		int error;
1293 		dsl_dataset_t *ds = tx->tx_objset->os_dsl_dataset;
1294 
1295 		error = dsl_dataset_check_quota(ds, !netfree,
1296 		    asize, est_inflight, &used_on_disk, &ref_rsrv);
1297 		if (error != 0) {
1298 			mutex_exit(&dd->dd_lock);
1299 			DMU_TX_STAT_BUMP(dmu_tx_quota);
1300 			return (error);
1301 		}
1302 	}
1303 
1304 	/*
1305 	 * If this transaction will result in a net free of space,
1306 	 * we want to let it through.
1307 	 */
1308 	if (ignorequota || netfree || dsl_dir_phys(dd)->dd_quota == 0)
1309 		quota = UINT64_MAX;
1310 	else
1311 		quota = dsl_dir_phys(dd)->dd_quota;
1312 
1313 	/*
1314 	 * Adjust the quota against the actual pool size at the root
1315 	 * minus any outstanding deferred frees.
1316 	 * To ensure that it's possible to remove files from a full
1317 	 * pool without inducing transient overcommits, we throttle
1318 	 * netfree transactions against a quota that is slightly larger,
1319 	 * but still within the pool's allocation slop.  In cases where
1320 	 * we're very close to full, this will allow a steady trickle of
1321 	 * removes to get through.
1322 	 */
1323 	uint64_t deferred = 0;
1324 	if (dd->dd_parent == NULL) {
1325 		uint64_t avail = dsl_pool_unreserved_space(dd->dd_pool,
1326 		    (netfree) ?
1327 		    ZFS_SPACE_CHECK_RESERVED : ZFS_SPACE_CHECK_NORMAL);
1328 
1329 		if (avail < quota) {
1330 			quota = avail;
1331 			retval = SET_ERROR(ENOSPC);
1332 		}
1333 	}
1334 
1335 	/*
1336 	 * If they are requesting more space, and our current estimate
1337 	 * is over quota, they get to try again unless the actual
1338 	 * on-disk is over quota and there are no pending changes (which
1339 	 * may free up space for us).
1340 	 */
1341 	if (used_on_disk + est_inflight >= quota) {
1342 		if (est_inflight > 0 || used_on_disk < quota ||
1343 		    (retval == ENOSPC && used_on_disk < quota + deferred))
1344 			retval = ERESTART;
1345 		dprintf_dd(dd, "failing: used=%lluK inflight = %lluK "
1346 		    "quota=%lluK tr=%lluK err=%d\n",
1347 		    (u_longlong_t)used_on_disk>>10,
1348 		    (u_longlong_t)est_inflight>>10,
1349 		    (u_longlong_t)quota>>10, (u_longlong_t)asize>>10, retval);
1350 		mutex_exit(&dd->dd_lock);
1351 		DMU_TX_STAT_BUMP(dmu_tx_quota);
1352 		return (SET_ERROR(retval));
1353 	}
1354 
1355 	/* We need to up our estimated delta before dropping dd_lock */
1356 	dd->dd_tempreserved[txg & TXG_MASK] += asize;
1357 
1358 	uint64_t parent_rsrv = parent_delta(dd, used_on_disk + est_inflight,
1359 	    asize - ref_rsrv);
1360 	mutex_exit(&dd->dd_lock);
1361 
1362 	tr = kmem_zalloc(sizeof (struct tempreserve), KM_SLEEP);
1363 	tr->tr_ds = dd;
1364 	tr->tr_size = asize;
1365 	list_insert_tail(tr_list, tr);
1366 
1367 	/* see if it's OK with our parent */
1368 	if (dd->dd_parent != NULL && parent_rsrv != 0) {
1369 		/*
1370 		 * Recurse on our parent without recursion. This has been
1371 		 * observed to be potentially large stack usage even within
1372 		 * the test suite. Largest seen stack was 7632 bytes on linux.
1373 		 */
1374 
1375 		dd = dd->dd_parent;
1376 		asize = parent_rsrv;
1377 		ignorequota = (dsl_dir_phys(dd)->dd_head_dataset_obj == 0);
1378 		first = B_FALSE;
1379 		goto top_of_function;
1380 
1381 	} else {
1382 		return (0);
1383 	}
1384 }
1385 
1386 /*
1387  * Reserve space in this dsl_dir, to be used in this tx's txg.
1388  * After the space has been dirtied (and dsl_dir_willuse_space()
1389  * has been called), the reservation should be canceled, using
1390  * dsl_dir_tempreserve_clear().
1391  */
1392 int
1393 dsl_dir_tempreserve_space(dsl_dir_t *dd, uint64_t lsize, uint64_t asize,
1394     boolean_t netfree, void **tr_cookiep, dmu_tx_t *tx)
1395 {
1396 	int err;
1397 	list_t *tr_list;
1398 
1399 	if (asize == 0) {
1400 		*tr_cookiep = NULL;
1401 		return (0);
1402 	}
1403 
1404 	tr_list = kmem_alloc(sizeof (list_t), KM_SLEEP);
1405 	list_create(tr_list, sizeof (struct tempreserve),
1406 	    offsetof(struct tempreserve, tr_node));
1407 	ASSERT3S(asize, >, 0);
1408 
1409 	err = arc_tempreserve_space(dd->dd_pool->dp_spa, lsize, tx->tx_txg);
1410 	if (err == 0) {
1411 		struct tempreserve *tr;
1412 
1413 		tr = kmem_zalloc(sizeof (struct tempreserve), KM_SLEEP);
1414 		tr->tr_size = lsize;
1415 		list_insert_tail(tr_list, tr);
1416 	} else {
1417 		if (err == EAGAIN) {
1418 			/*
1419 			 * If arc_memory_throttle() detected that pageout
1420 			 * is running and we are low on memory, we delay new
1421 			 * non-pageout transactions to give pageout an
1422 			 * advantage.
1423 			 *
1424 			 * It is unfortunate to be delaying while the caller's
1425 			 * locks are held.
1426 			 */
1427 			txg_delay(dd->dd_pool, tx->tx_txg,
1428 			    MSEC2NSEC(10), MSEC2NSEC(10));
1429 			err = SET_ERROR(ERESTART);
1430 		}
1431 	}
1432 
1433 	if (err == 0) {
1434 		err = dsl_dir_tempreserve_impl(dd, asize, netfree,
1435 		    B_FALSE, tr_list, tx, B_TRUE);
1436 	}
1437 
1438 	if (err != 0)
1439 		dsl_dir_tempreserve_clear(tr_list, tx);
1440 	else
1441 		*tr_cookiep = tr_list;
1442 
1443 	return (err);
1444 }
1445 
1446 /*
1447  * Clear a temporary reservation that we previously made with
1448  * dsl_dir_tempreserve_space().
1449  */
1450 void
1451 dsl_dir_tempreserve_clear(void *tr_cookie, dmu_tx_t *tx)
1452 {
1453 	int txgidx = tx->tx_txg & TXG_MASK;
1454 	list_t *tr_list = tr_cookie;
1455 	struct tempreserve *tr;
1456 
1457 	ASSERT3U(tx->tx_txg, !=, 0);
1458 
1459 	if (tr_cookie == NULL)
1460 		return;
1461 
1462 	while ((tr = list_head(tr_list)) != NULL) {
1463 		if (tr->tr_ds) {
1464 			mutex_enter(&tr->tr_ds->dd_lock);
1465 			ASSERT3U(tr->tr_ds->dd_tempreserved[txgidx], >=,
1466 			    tr->tr_size);
1467 			tr->tr_ds->dd_tempreserved[txgidx] -= tr->tr_size;
1468 			mutex_exit(&tr->tr_ds->dd_lock);
1469 		} else {
1470 			arc_tempreserve_clear(tr->tr_size);
1471 		}
1472 		list_remove(tr_list, tr);
1473 		kmem_free(tr, sizeof (struct tempreserve));
1474 	}
1475 
1476 	kmem_free(tr_list, sizeof (list_t));
1477 }
1478 
1479 /*
1480  * This should be called from open context when we think we're going to write
1481  * or free space, for example when dirtying data. Be conservative; it's okay
1482  * to write less space or free more, but we don't want to write more or free
1483  * less than the amount specified.
1484  *
1485  * NOTE: The behavior of this function is identical to the Illumos / FreeBSD
1486  * version however it has been adjusted to use an iterative rather than
1487  * recursive algorithm to minimize stack usage.
1488  */
1489 void
1490 dsl_dir_willuse_space(dsl_dir_t *dd, int64_t space, dmu_tx_t *tx)
1491 {
1492 	int64_t parent_space;
1493 	uint64_t est_used;
1494 
1495 	do {
1496 		mutex_enter(&dd->dd_lock);
1497 		if (space > 0)
1498 			dd->dd_space_towrite[tx->tx_txg & TXG_MASK] += space;
1499 
1500 		est_used = dsl_dir_space_towrite(dd) +
1501 		    dsl_dir_phys(dd)->dd_used_bytes;
1502 		parent_space = parent_delta(dd, est_used, space);
1503 		mutex_exit(&dd->dd_lock);
1504 
1505 		/* Make sure that we clean up dd_space_to* */
1506 		dsl_dir_dirty(dd, tx);
1507 
1508 		dd = dd->dd_parent;
1509 		space = parent_space;
1510 	} while (space && dd);
1511 }
1512 
1513 /* call from syncing context when we actually write/free space for this dd */
1514 void
1515 dsl_dir_diduse_space(dsl_dir_t *dd, dd_used_t type,
1516     int64_t used, int64_t compressed, int64_t uncompressed, dmu_tx_t *tx)
1517 {
1518 	int64_t accounted_delta;
1519 
1520 	ASSERT(dmu_tx_is_syncing(tx));
1521 	ASSERT(type < DD_USED_NUM);
1522 
1523 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
1524 
1525 	/*
1526 	 * dsl_dataset_set_refreservation_sync_impl() calls this with
1527 	 * dd_lock held, so that it can atomically update
1528 	 * ds->ds_reserved and the dsl_dir accounting, so that
1529 	 * dsl_dataset_check_quota() can see dataset and dir accounting
1530 	 * consistently.
1531 	 */
1532 	boolean_t needlock = !MUTEX_HELD(&dd->dd_lock);
1533 	if (needlock)
1534 		mutex_enter(&dd->dd_lock);
1535 	dsl_dir_phys_t *ddp = dsl_dir_phys(dd);
1536 	accounted_delta = parent_delta(dd, ddp->dd_used_bytes, used);
1537 	ASSERT(used >= 0 || ddp->dd_used_bytes >= -used);
1538 	ASSERT(compressed >= 0 || ddp->dd_compressed_bytes >= -compressed);
1539 	ASSERT(uncompressed >= 0 ||
1540 	    ddp->dd_uncompressed_bytes >= -uncompressed);
1541 	ddp->dd_used_bytes += used;
1542 	ddp->dd_uncompressed_bytes += uncompressed;
1543 	ddp->dd_compressed_bytes += compressed;
1544 
1545 	if (ddp->dd_flags & DD_FLAG_USED_BREAKDOWN) {
1546 		ASSERT(used >= 0 || ddp->dd_used_breakdown[type] >= -used);
1547 		ddp->dd_used_breakdown[type] += used;
1548 #ifdef ZFS_DEBUG
1549 		{
1550 			dd_used_t t;
1551 			uint64_t u = 0;
1552 			for (t = 0; t < DD_USED_NUM; t++)
1553 				u += ddp->dd_used_breakdown[t];
1554 			ASSERT3U(u, ==, ddp->dd_used_bytes);
1555 		}
1556 #endif
1557 	}
1558 	if (needlock)
1559 		mutex_exit(&dd->dd_lock);
1560 
1561 	if (dd->dd_parent != NULL) {
1562 		dsl_dir_diduse_transfer_space(dd->dd_parent,
1563 		    accounted_delta, compressed, uncompressed,
1564 		    used, DD_USED_CHILD_RSRV, DD_USED_CHILD, tx);
1565 	}
1566 }
1567 
1568 void
1569 dsl_dir_transfer_space(dsl_dir_t *dd, int64_t delta,
1570     dd_used_t oldtype, dd_used_t newtype, dmu_tx_t *tx)
1571 {
1572 	ASSERT(dmu_tx_is_syncing(tx));
1573 	ASSERT(oldtype < DD_USED_NUM);
1574 	ASSERT(newtype < DD_USED_NUM);
1575 
1576 	dsl_dir_phys_t *ddp = dsl_dir_phys(dd);
1577 	if (delta == 0 ||
1578 	    !(ddp->dd_flags & DD_FLAG_USED_BREAKDOWN))
1579 		return;
1580 
1581 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
1582 	mutex_enter(&dd->dd_lock);
1583 	ASSERT(delta > 0 ?
1584 	    ddp->dd_used_breakdown[oldtype] >= delta :
1585 	    ddp->dd_used_breakdown[newtype] >= -delta);
1586 	ASSERT(ddp->dd_used_bytes >= ABS(delta));
1587 	ddp->dd_used_breakdown[oldtype] -= delta;
1588 	ddp->dd_used_breakdown[newtype] += delta;
1589 	mutex_exit(&dd->dd_lock);
1590 }
1591 
1592 void
1593 dsl_dir_diduse_transfer_space(dsl_dir_t *dd, int64_t used,
1594     int64_t compressed, int64_t uncompressed, int64_t tonew,
1595     dd_used_t oldtype, dd_used_t newtype, dmu_tx_t *tx)
1596 {
1597 	int64_t accounted_delta;
1598 
1599 	ASSERT(dmu_tx_is_syncing(tx));
1600 	ASSERT(oldtype < DD_USED_NUM);
1601 	ASSERT(newtype < DD_USED_NUM);
1602 
1603 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
1604 
1605 	mutex_enter(&dd->dd_lock);
1606 	dsl_dir_phys_t *ddp = dsl_dir_phys(dd);
1607 	accounted_delta = parent_delta(dd, ddp->dd_used_bytes, used);
1608 	ASSERT(used >= 0 || ddp->dd_used_bytes >= -used);
1609 	ASSERT(compressed >= 0 || ddp->dd_compressed_bytes >= -compressed);
1610 	ASSERT(uncompressed >= 0 ||
1611 	    ddp->dd_uncompressed_bytes >= -uncompressed);
1612 	ddp->dd_used_bytes += used;
1613 	ddp->dd_uncompressed_bytes += uncompressed;
1614 	ddp->dd_compressed_bytes += compressed;
1615 
1616 	if (ddp->dd_flags & DD_FLAG_USED_BREAKDOWN) {
1617 		ASSERT(tonew - used <= 0 ||
1618 		    ddp->dd_used_breakdown[oldtype] >= tonew - used);
1619 		ASSERT(tonew >= 0 ||
1620 		    ddp->dd_used_breakdown[newtype] >= -tonew);
1621 		ddp->dd_used_breakdown[oldtype] -= tonew - used;
1622 		ddp->dd_used_breakdown[newtype] += tonew;
1623 #ifdef ZFS_DEBUG
1624 		{
1625 			dd_used_t t;
1626 			uint64_t u = 0;
1627 			for (t = 0; t < DD_USED_NUM; t++)
1628 				u += ddp->dd_used_breakdown[t];
1629 			ASSERT3U(u, ==, ddp->dd_used_bytes);
1630 		}
1631 #endif
1632 	}
1633 	mutex_exit(&dd->dd_lock);
1634 
1635 	if (dd->dd_parent != NULL) {
1636 		dsl_dir_diduse_transfer_space(dd->dd_parent,
1637 		    accounted_delta, compressed, uncompressed,
1638 		    used, DD_USED_CHILD_RSRV, DD_USED_CHILD, tx);
1639 	}
1640 }
1641 
1642 typedef struct dsl_dir_set_qr_arg {
1643 	const char *ddsqra_name;
1644 	zprop_source_t ddsqra_source;
1645 	uint64_t ddsqra_value;
1646 } dsl_dir_set_qr_arg_t;
1647 
1648 static int
1649 dsl_dir_set_quota_check(void *arg, dmu_tx_t *tx)
1650 {
1651 	dsl_dir_set_qr_arg_t *ddsqra = arg;
1652 	dsl_pool_t *dp = dmu_tx_pool(tx);
1653 	dsl_dataset_t *ds;
1654 	int error;
1655 	uint64_t towrite, newval;
1656 
1657 	error = dsl_dataset_hold(dp, ddsqra->ddsqra_name, FTAG, &ds);
1658 	if (error != 0)
1659 		return (error);
1660 
1661 	error = dsl_prop_predict(ds->ds_dir, "quota",
1662 	    ddsqra->ddsqra_source, ddsqra->ddsqra_value, &newval);
1663 	if (error != 0) {
1664 		dsl_dataset_rele(ds, FTAG);
1665 		return (error);
1666 	}
1667 
1668 	if (newval == 0) {
1669 		dsl_dataset_rele(ds, FTAG);
1670 		return (0);
1671 	}
1672 
1673 	mutex_enter(&ds->ds_dir->dd_lock);
1674 	/*
1675 	 * If we are doing the preliminary check in open context, and
1676 	 * there are pending changes, then don't fail it, since the
1677 	 * pending changes could under-estimate the amount of space to be
1678 	 * freed up.
1679 	 */
1680 	towrite = dsl_dir_space_towrite(ds->ds_dir);
1681 	if ((dmu_tx_is_syncing(tx) || towrite == 0) &&
1682 	    (newval < dsl_dir_phys(ds->ds_dir)->dd_reserved ||
1683 	    newval < dsl_dir_phys(ds->ds_dir)->dd_used_bytes + towrite)) {
1684 		error = SET_ERROR(ENOSPC);
1685 	}
1686 	mutex_exit(&ds->ds_dir->dd_lock);
1687 	dsl_dataset_rele(ds, FTAG);
1688 	return (error);
1689 }
1690 
1691 static void
1692 dsl_dir_set_quota_sync(void *arg, dmu_tx_t *tx)
1693 {
1694 	dsl_dir_set_qr_arg_t *ddsqra = arg;
1695 	dsl_pool_t *dp = dmu_tx_pool(tx);
1696 	dsl_dataset_t *ds;
1697 	uint64_t newval;
1698 
1699 	VERIFY0(dsl_dataset_hold(dp, ddsqra->ddsqra_name, FTAG, &ds));
1700 
1701 	if (spa_version(dp->dp_spa) >= SPA_VERSION_RECVD_PROPS) {
1702 		dsl_prop_set_sync_impl(ds, zfs_prop_to_name(ZFS_PROP_QUOTA),
1703 		    ddsqra->ddsqra_source, sizeof (ddsqra->ddsqra_value), 1,
1704 		    &ddsqra->ddsqra_value, tx);
1705 
1706 		VERIFY0(dsl_prop_get_int_ds(ds,
1707 		    zfs_prop_to_name(ZFS_PROP_QUOTA), &newval));
1708 	} else {
1709 		newval = ddsqra->ddsqra_value;
1710 		spa_history_log_internal_ds(ds, "set", tx, "%s=%lld",
1711 		    zfs_prop_to_name(ZFS_PROP_QUOTA), (longlong_t)newval);
1712 	}
1713 
1714 	dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
1715 	mutex_enter(&ds->ds_dir->dd_lock);
1716 	dsl_dir_phys(ds->ds_dir)->dd_quota = newval;
1717 	mutex_exit(&ds->ds_dir->dd_lock);
1718 	dsl_dataset_rele(ds, FTAG);
1719 }
1720 
1721 int
1722 dsl_dir_set_quota(const char *ddname, zprop_source_t source, uint64_t quota)
1723 {
1724 	dsl_dir_set_qr_arg_t ddsqra;
1725 
1726 	ddsqra.ddsqra_name = ddname;
1727 	ddsqra.ddsqra_source = source;
1728 	ddsqra.ddsqra_value = quota;
1729 
1730 	return (dsl_sync_task(ddname, dsl_dir_set_quota_check,
1731 	    dsl_dir_set_quota_sync, &ddsqra, 0,
1732 	    ZFS_SPACE_CHECK_EXTRA_RESERVED));
1733 }
1734 
1735 static int
1736 dsl_dir_set_reservation_check(void *arg, dmu_tx_t *tx)
1737 {
1738 	dsl_dir_set_qr_arg_t *ddsqra = arg;
1739 	dsl_pool_t *dp = dmu_tx_pool(tx);
1740 	dsl_dataset_t *ds;
1741 	dsl_dir_t *dd;
1742 	uint64_t newval, used, avail;
1743 	int error;
1744 
1745 	error = dsl_dataset_hold(dp, ddsqra->ddsqra_name, FTAG, &ds);
1746 	if (error != 0)
1747 		return (error);
1748 	dd = ds->ds_dir;
1749 
1750 	/*
1751 	 * If we are doing the preliminary check in open context, the
1752 	 * space estimates may be inaccurate.
1753 	 */
1754 	if (!dmu_tx_is_syncing(tx)) {
1755 		dsl_dataset_rele(ds, FTAG);
1756 		return (0);
1757 	}
1758 
1759 	error = dsl_prop_predict(ds->ds_dir,
1760 	    zfs_prop_to_name(ZFS_PROP_RESERVATION),
1761 	    ddsqra->ddsqra_source, ddsqra->ddsqra_value, &newval);
1762 	if (error != 0) {
1763 		dsl_dataset_rele(ds, FTAG);
1764 		return (error);
1765 	}
1766 
1767 	mutex_enter(&dd->dd_lock);
1768 	used = dsl_dir_phys(dd)->dd_used_bytes;
1769 	mutex_exit(&dd->dd_lock);
1770 
1771 	if (dd->dd_parent) {
1772 		avail = dsl_dir_space_available(dd->dd_parent,
1773 		    NULL, 0, FALSE);
1774 	} else {
1775 		avail = dsl_pool_adjustedsize(dd->dd_pool,
1776 		    ZFS_SPACE_CHECK_NORMAL) - used;
1777 	}
1778 
1779 	if (MAX(used, newval) > MAX(used, dsl_dir_phys(dd)->dd_reserved)) {
1780 		uint64_t delta = MAX(used, newval) -
1781 		    MAX(used, dsl_dir_phys(dd)->dd_reserved);
1782 
1783 		if (delta > avail ||
1784 		    (dsl_dir_phys(dd)->dd_quota > 0 &&
1785 		    newval > dsl_dir_phys(dd)->dd_quota))
1786 			error = SET_ERROR(ENOSPC);
1787 	}
1788 
1789 	dsl_dataset_rele(ds, FTAG);
1790 	return (error);
1791 }
1792 
1793 void
1794 dsl_dir_set_reservation_sync_impl(dsl_dir_t *dd, uint64_t value, dmu_tx_t *tx)
1795 {
1796 	uint64_t used;
1797 	int64_t delta;
1798 
1799 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
1800 
1801 	mutex_enter(&dd->dd_lock);
1802 	used = dsl_dir_phys(dd)->dd_used_bytes;
1803 	delta = MAX(used, value) - MAX(used, dsl_dir_phys(dd)->dd_reserved);
1804 	dsl_dir_phys(dd)->dd_reserved = value;
1805 
1806 	if (dd->dd_parent != NULL) {
1807 		/* Roll up this additional usage into our ancestors */
1808 		dsl_dir_diduse_space(dd->dd_parent, DD_USED_CHILD_RSRV,
1809 		    delta, 0, 0, tx);
1810 	}
1811 	mutex_exit(&dd->dd_lock);
1812 }
1813 
1814 static void
1815 dsl_dir_set_reservation_sync(void *arg, dmu_tx_t *tx)
1816 {
1817 	dsl_dir_set_qr_arg_t *ddsqra = arg;
1818 	dsl_pool_t *dp = dmu_tx_pool(tx);
1819 	dsl_dataset_t *ds;
1820 	uint64_t newval;
1821 
1822 	VERIFY0(dsl_dataset_hold(dp, ddsqra->ddsqra_name, FTAG, &ds));
1823 
1824 	if (spa_version(dp->dp_spa) >= SPA_VERSION_RECVD_PROPS) {
1825 		dsl_prop_set_sync_impl(ds,
1826 		    zfs_prop_to_name(ZFS_PROP_RESERVATION),
1827 		    ddsqra->ddsqra_source, sizeof (ddsqra->ddsqra_value), 1,
1828 		    &ddsqra->ddsqra_value, tx);
1829 
1830 		VERIFY0(dsl_prop_get_int_ds(ds,
1831 		    zfs_prop_to_name(ZFS_PROP_RESERVATION), &newval));
1832 	} else {
1833 		newval = ddsqra->ddsqra_value;
1834 		spa_history_log_internal_ds(ds, "set", tx, "%s=%lld",
1835 		    zfs_prop_to_name(ZFS_PROP_RESERVATION),
1836 		    (longlong_t)newval);
1837 	}
1838 
1839 	dsl_dir_set_reservation_sync_impl(ds->ds_dir, newval, tx);
1840 	dsl_dataset_rele(ds, FTAG);
1841 }
1842 
1843 int
1844 dsl_dir_set_reservation(const char *ddname, zprop_source_t source,
1845     uint64_t reservation)
1846 {
1847 	dsl_dir_set_qr_arg_t ddsqra;
1848 
1849 	ddsqra.ddsqra_name = ddname;
1850 	ddsqra.ddsqra_source = source;
1851 	ddsqra.ddsqra_value = reservation;
1852 
1853 	return (dsl_sync_task(ddname, dsl_dir_set_reservation_check,
1854 	    dsl_dir_set_reservation_sync, &ddsqra, 0,
1855 	    ZFS_SPACE_CHECK_EXTRA_RESERVED));
1856 }
1857 
1858 static dsl_dir_t *
1859 closest_common_ancestor(dsl_dir_t *ds1, dsl_dir_t *ds2)
1860 {
1861 	for (; ds1; ds1 = ds1->dd_parent) {
1862 		dsl_dir_t *dd;
1863 		for (dd = ds2; dd; dd = dd->dd_parent) {
1864 			if (ds1 == dd)
1865 				return (dd);
1866 		}
1867 	}
1868 	return (NULL);
1869 }
1870 
1871 /*
1872  * If delta is applied to dd, how much of that delta would be applied to
1873  * ancestor?  Syncing context only.
1874  */
1875 static int64_t
1876 would_change(dsl_dir_t *dd, int64_t delta, dsl_dir_t *ancestor)
1877 {
1878 	if (dd == ancestor)
1879 		return (delta);
1880 
1881 	mutex_enter(&dd->dd_lock);
1882 	delta = parent_delta(dd, dsl_dir_phys(dd)->dd_used_bytes, delta);
1883 	mutex_exit(&dd->dd_lock);
1884 	return (would_change(dd->dd_parent, delta, ancestor));
1885 }
1886 
1887 typedef struct dsl_dir_rename_arg {
1888 	const char *ddra_oldname;
1889 	const char *ddra_newname;
1890 	cred_t *ddra_cred;
1891 	proc_t *ddra_proc;
1892 } dsl_dir_rename_arg_t;
1893 
1894 typedef struct dsl_valid_rename_arg {
1895 	int char_delta;
1896 	int nest_delta;
1897 } dsl_valid_rename_arg_t;
1898 
1899 /* ARGSUSED */
1900 static int
1901 dsl_valid_rename(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
1902 {
1903 	dsl_valid_rename_arg_t *dvra = arg;
1904 	char namebuf[ZFS_MAX_DATASET_NAME_LEN];
1905 
1906 	dsl_dataset_name(ds, namebuf);
1907 
1908 	ASSERT3U(strnlen(namebuf, ZFS_MAX_DATASET_NAME_LEN),
1909 	    <, ZFS_MAX_DATASET_NAME_LEN);
1910 	int namelen = strlen(namebuf) + dvra->char_delta;
1911 	int depth = get_dataset_depth(namebuf) + dvra->nest_delta;
1912 
1913 	if (namelen >= ZFS_MAX_DATASET_NAME_LEN)
1914 		return (SET_ERROR(ENAMETOOLONG));
1915 	if (dvra->nest_delta > 0 && depth >= zfs_max_dataset_nesting)
1916 		return (SET_ERROR(ENAMETOOLONG));
1917 	return (0);
1918 }
1919 
1920 static int
1921 dsl_dir_rename_check(void *arg, dmu_tx_t *tx)
1922 {
1923 	dsl_dir_rename_arg_t *ddra = arg;
1924 	dsl_pool_t *dp = dmu_tx_pool(tx);
1925 	dsl_dir_t *dd, *newparent;
1926 	dsl_valid_rename_arg_t dvra;
1927 	dsl_dataset_t *parentds;
1928 	objset_t *parentos;
1929 	const char *mynewname;
1930 	int error;
1931 
1932 	/* target dir should exist */
1933 	error = dsl_dir_hold(dp, ddra->ddra_oldname, FTAG, &dd, NULL);
1934 	if (error != 0)
1935 		return (error);
1936 
1937 	/* new parent should exist */
1938 	error = dsl_dir_hold(dp, ddra->ddra_newname, FTAG,
1939 	    &newparent, &mynewname);
1940 	if (error != 0) {
1941 		dsl_dir_rele(dd, FTAG);
1942 		return (error);
1943 	}
1944 
1945 	/* can't rename to different pool */
1946 	if (dd->dd_pool != newparent->dd_pool) {
1947 		dsl_dir_rele(newparent, FTAG);
1948 		dsl_dir_rele(dd, FTAG);
1949 		return (SET_ERROR(EXDEV));
1950 	}
1951 
1952 	/* new name should not already exist */
1953 	if (mynewname == NULL) {
1954 		dsl_dir_rele(newparent, FTAG);
1955 		dsl_dir_rele(dd, FTAG);
1956 		return (SET_ERROR(EEXIST));
1957 	}
1958 
1959 	/* can't rename below anything but filesystems (eg. no ZVOLs) */
1960 	error = dsl_dataset_hold_obj(newparent->dd_pool,
1961 	    dsl_dir_phys(newparent)->dd_head_dataset_obj, FTAG, &parentds);
1962 	if (error != 0) {
1963 		dsl_dir_rele(newparent, FTAG);
1964 		dsl_dir_rele(dd, FTAG);
1965 		return (error);
1966 	}
1967 	error = dmu_objset_from_ds(parentds, &parentos);
1968 	if (error != 0) {
1969 		dsl_dataset_rele(parentds, FTAG);
1970 		dsl_dir_rele(newparent, FTAG);
1971 		dsl_dir_rele(dd, FTAG);
1972 		return (error);
1973 	}
1974 	if (dmu_objset_type(parentos) != DMU_OST_ZFS) {
1975 		dsl_dataset_rele(parentds, FTAG);
1976 		dsl_dir_rele(newparent, FTAG);
1977 		dsl_dir_rele(dd, FTAG);
1978 		return (SET_ERROR(ZFS_ERR_WRONG_PARENT));
1979 	}
1980 	dsl_dataset_rele(parentds, FTAG);
1981 
1982 	ASSERT3U(strnlen(ddra->ddra_newname, ZFS_MAX_DATASET_NAME_LEN),
1983 	    <, ZFS_MAX_DATASET_NAME_LEN);
1984 	ASSERT3U(strnlen(ddra->ddra_oldname, ZFS_MAX_DATASET_NAME_LEN),
1985 	    <, ZFS_MAX_DATASET_NAME_LEN);
1986 	dvra.char_delta = strlen(ddra->ddra_newname)
1987 	    - strlen(ddra->ddra_oldname);
1988 	dvra.nest_delta = get_dataset_depth(ddra->ddra_newname)
1989 	    - get_dataset_depth(ddra->ddra_oldname);
1990 
1991 	/* if the name length is growing, validate child name lengths */
1992 	if (dvra.char_delta > 0 || dvra.nest_delta > 0) {
1993 		error = dmu_objset_find_dp(dp, dd->dd_object, dsl_valid_rename,
1994 		    &dvra, DS_FIND_CHILDREN | DS_FIND_SNAPSHOTS);
1995 		if (error != 0) {
1996 			dsl_dir_rele(newparent, FTAG);
1997 			dsl_dir_rele(dd, FTAG);
1998 			return (error);
1999 		}
2000 	}
2001 
2002 	if (dmu_tx_is_syncing(tx)) {
2003 		if (spa_feature_is_active(dp->dp_spa,
2004 		    SPA_FEATURE_FS_SS_LIMIT)) {
2005 			/*
2006 			 * Although this is the check function and we don't
2007 			 * normally make on-disk changes in check functions,
2008 			 * we need to do that here.
2009 			 *
2010 			 * Ensure this portion of the tree's counts have been
2011 			 * initialized in case the new parent has limits set.
2012 			 */
2013 			dsl_dir_init_fs_ss_count(dd, tx);
2014 		}
2015 	}
2016 
2017 	if (newparent != dd->dd_parent) {
2018 		/* is there enough space? */
2019 		uint64_t myspace =
2020 		    MAX(dsl_dir_phys(dd)->dd_used_bytes,
2021 		    dsl_dir_phys(dd)->dd_reserved);
2022 		objset_t *os = dd->dd_pool->dp_meta_objset;
2023 		uint64_t fs_cnt = 0;
2024 		uint64_t ss_cnt = 0;
2025 
2026 		if (dsl_dir_is_zapified(dd)) {
2027 			int err;
2028 
2029 			err = zap_lookup(os, dd->dd_object,
2030 			    DD_FIELD_FILESYSTEM_COUNT, sizeof (fs_cnt), 1,
2031 			    &fs_cnt);
2032 			if (err != ENOENT && err != 0) {
2033 				dsl_dir_rele(newparent, FTAG);
2034 				dsl_dir_rele(dd, FTAG);
2035 				return (err);
2036 			}
2037 
2038 			/*
2039 			 * have to add 1 for the filesystem itself that we're
2040 			 * moving
2041 			 */
2042 			fs_cnt++;
2043 
2044 			err = zap_lookup(os, dd->dd_object,
2045 			    DD_FIELD_SNAPSHOT_COUNT, sizeof (ss_cnt), 1,
2046 			    &ss_cnt);
2047 			if (err != ENOENT && err != 0) {
2048 				dsl_dir_rele(newparent, FTAG);
2049 				dsl_dir_rele(dd, FTAG);
2050 				return (err);
2051 			}
2052 		}
2053 
2054 		/* check for encryption errors */
2055 		error = dsl_dir_rename_crypt_check(dd, newparent);
2056 		if (error != 0) {
2057 			dsl_dir_rele(newparent, FTAG);
2058 			dsl_dir_rele(dd, FTAG);
2059 			return (SET_ERROR(EACCES));
2060 		}
2061 
2062 		/* no rename into our descendant */
2063 		if (closest_common_ancestor(dd, newparent) == dd) {
2064 			dsl_dir_rele(newparent, FTAG);
2065 			dsl_dir_rele(dd, FTAG);
2066 			return (SET_ERROR(EINVAL));
2067 		}
2068 
2069 		error = dsl_dir_transfer_possible(dd->dd_parent,
2070 		    newparent, fs_cnt, ss_cnt, myspace,
2071 		    ddra->ddra_cred, ddra->ddra_proc);
2072 		if (error != 0) {
2073 			dsl_dir_rele(newparent, FTAG);
2074 			dsl_dir_rele(dd, FTAG);
2075 			return (error);
2076 		}
2077 	}
2078 
2079 	dsl_dir_rele(newparent, FTAG);
2080 	dsl_dir_rele(dd, FTAG);
2081 	return (0);
2082 }
2083 
2084 static void
2085 dsl_dir_rename_sync(void *arg, dmu_tx_t *tx)
2086 {
2087 	dsl_dir_rename_arg_t *ddra = arg;
2088 	dsl_pool_t *dp = dmu_tx_pool(tx);
2089 	dsl_dir_t *dd, *newparent;
2090 	const char *mynewname;
2091 	objset_t *mos = dp->dp_meta_objset;
2092 
2093 	VERIFY0(dsl_dir_hold(dp, ddra->ddra_oldname, FTAG, &dd, NULL));
2094 	VERIFY0(dsl_dir_hold(dp, ddra->ddra_newname, FTAG, &newparent,
2095 	    &mynewname));
2096 
2097 	/* Log this before we change the name. */
2098 	spa_history_log_internal_dd(dd, "rename", tx,
2099 	    "-> %s", ddra->ddra_newname);
2100 
2101 	if (newparent != dd->dd_parent) {
2102 		objset_t *os = dd->dd_pool->dp_meta_objset;
2103 		uint64_t fs_cnt = 0;
2104 		uint64_t ss_cnt = 0;
2105 
2106 		/*
2107 		 * We already made sure the dd counts were initialized in the
2108 		 * check function.
2109 		 */
2110 		if (spa_feature_is_active(dp->dp_spa,
2111 		    SPA_FEATURE_FS_SS_LIMIT)) {
2112 			VERIFY0(zap_lookup(os, dd->dd_object,
2113 			    DD_FIELD_FILESYSTEM_COUNT, sizeof (fs_cnt), 1,
2114 			    &fs_cnt));
2115 			/* add 1 for the filesystem itself that we're moving */
2116 			fs_cnt++;
2117 
2118 			VERIFY0(zap_lookup(os, dd->dd_object,
2119 			    DD_FIELD_SNAPSHOT_COUNT, sizeof (ss_cnt), 1,
2120 			    &ss_cnt));
2121 		}
2122 
2123 		dsl_fs_ss_count_adjust(dd->dd_parent, -fs_cnt,
2124 		    DD_FIELD_FILESYSTEM_COUNT, tx);
2125 		dsl_fs_ss_count_adjust(newparent, fs_cnt,
2126 		    DD_FIELD_FILESYSTEM_COUNT, tx);
2127 
2128 		dsl_fs_ss_count_adjust(dd->dd_parent, -ss_cnt,
2129 		    DD_FIELD_SNAPSHOT_COUNT, tx);
2130 		dsl_fs_ss_count_adjust(newparent, ss_cnt,
2131 		    DD_FIELD_SNAPSHOT_COUNT, tx);
2132 
2133 		dsl_dir_diduse_space(dd->dd_parent, DD_USED_CHILD,
2134 		    -dsl_dir_phys(dd)->dd_used_bytes,
2135 		    -dsl_dir_phys(dd)->dd_compressed_bytes,
2136 		    -dsl_dir_phys(dd)->dd_uncompressed_bytes, tx);
2137 		dsl_dir_diduse_space(newparent, DD_USED_CHILD,
2138 		    dsl_dir_phys(dd)->dd_used_bytes,
2139 		    dsl_dir_phys(dd)->dd_compressed_bytes,
2140 		    dsl_dir_phys(dd)->dd_uncompressed_bytes, tx);
2141 
2142 		if (dsl_dir_phys(dd)->dd_reserved >
2143 		    dsl_dir_phys(dd)->dd_used_bytes) {
2144 			uint64_t unused_rsrv = dsl_dir_phys(dd)->dd_reserved -
2145 			    dsl_dir_phys(dd)->dd_used_bytes;
2146 
2147 			dsl_dir_diduse_space(dd->dd_parent, DD_USED_CHILD_RSRV,
2148 			    -unused_rsrv, 0, 0, tx);
2149 			dsl_dir_diduse_space(newparent, DD_USED_CHILD_RSRV,
2150 			    unused_rsrv, 0, 0, tx);
2151 		}
2152 	}
2153 
2154 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
2155 
2156 	/* remove from old parent zapobj */
2157 	VERIFY0(zap_remove(mos,
2158 	    dsl_dir_phys(dd->dd_parent)->dd_child_dir_zapobj,
2159 	    dd->dd_myname, tx));
2160 
2161 	(void) strlcpy(dd->dd_myname, mynewname,
2162 	    sizeof (dd->dd_myname));
2163 	dsl_dir_rele(dd->dd_parent, dd);
2164 	dsl_dir_phys(dd)->dd_parent_obj = newparent->dd_object;
2165 	VERIFY0(dsl_dir_hold_obj(dp,
2166 	    newparent->dd_object, NULL, dd, &dd->dd_parent));
2167 
2168 	/* add to new parent zapobj */
2169 	VERIFY0(zap_add(mos, dsl_dir_phys(newparent)->dd_child_dir_zapobj,
2170 	    dd->dd_myname, 8, 1, &dd->dd_object, tx));
2171 
2172 	/* TODO: A rename callback to avoid these layering violations. */
2173 	zfsvfs_update_fromname(ddra->ddra_oldname, ddra->ddra_newname);
2174 	zvol_rename_minors(dp->dp_spa, ddra->ddra_oldname,
2175 	    ddra->ddra_newname, B_TRUE);
2176 
2177 	dsl_prop_notify_all(dd);
2178 
2179 	dsl_dir_rele(newparent, FTAG);
2180 	dsl_dir_rele(dd, FTAG);
2181 }
2182 
2183 int
2184 dsl_dir_rename(const char *oldname, const char *newname)
2185 {
2186 	dsl_dir_rename_arg_t ddra;
2187 
2188 	ddra.ddra_oldname = oldname;
2189 	ddra.ddra_newname = newname;
2190 	ddra.ddra_cred = CRED();
2191 	ddra.ddra_proc = curproc;
2192 
2193 	return (dsl_sync_task(oldname,
2194 	    dsl_dir_rename_check, dsl_dir_rename_sync, &ddra,
2195 	    3, ZFS_SPACE_CHECK_RESERVED));
2196 }
2197 
2198 int
2199 dsl_dir_transfer_possible(dsl_dir_t *sdd, dsl_dir_t *tdd,
2200     uint64_t fs_cnt, uint64_t ss_cnt, uint64_t space,
2201     cred_t *cr, proc_t *proc)
2202 {
2203 	dsl_dir_t *ancestor;
2204 	int64_t adelta;
2205 	uint64_t avail;
2206 	int err;
2207 
2208 	ancestor = closest_common_ancestor(sdd, tdd);
2209 	adelta = would_change(sdd, -space, ancestor);
2210 	avail = dsl_dir_space_available(tdd, ancestor, adelta, FALSE);
2211 	if (avail < space)
2212 		return (SET_ERROR(ENOSPC));
2213 
2214 	err = dsl_fs_ss_limit_check(tdd, fs_cnt, ZFS_PROP_FILESYSTEM_LIMIT,
2215 	    ancestor, cr, proc);
2216 	if (err != 0)
2217 		return (err);
2218 	err = dsl_fs_ss_limit_check(tdd, ss_cnt, ZFS_PROP_SNAPSHOT_LIMIT,
2219 	    ancestor, cr, proc);
2220 	if (err != 0)
2221 		return (err);
2222 
2223 	return (0);
2224 }
2225 
2226 inode_timespec_t
2227 dsl_dir_snap_cmtime(dsl_dir_t *dd)
2228 {
2229 	inode_timespec_t t;
2230 
2231 	mutex_enter(&dd->dd_lock);
2232 	t = dd->dd_snap_cmtime;
2233 	mutex_exit(&dd->dd_lock);
2234 
2235 	return (t);
2236 }
2237 
2238 void
2239 dsl_dir_snap_cmtime_update(dsl_dir_t *dd)
2240 {
2241 	inode_timespec_t t;
2242 
2243 	gethrestime(&t);
2244 	mutex_enter(&dd->dd_lock);
2245 	dd->dd_snap_cmtime = t;
2246 	mutex_exit(&dd->dd_lock);
2247 }
2248 
2249 void
2250 dsl_dir_zapify(dsl_dir_t *dd, dmu_tx_t *tx)
2251 {
2252 	objset_t *mos = dd->dd_pool->dp_meta_objset;
2253 	dmu_object_zapify(mos, dd->dd_object, DMU_OT_DSL_DIR, tx);
2254 }
2255 
2256 boolean_t
2257 dsl_dir_is_zapified(dsl_dir_t *dd)
2258 {
2259 	dmu_object_info_t doi;
2260 
2261 	dmu_object_info_from_db(dd->dd_dbuf, &doi);
2262 	return (doi.doi_type == DMU_OTN_ZAP_METADATA);
2263 }
2264 
2265 void
2266 dsl_dir_livelist_open(dsl_dir_t *dd, uint64_t obj)
2267 {
2268 	objset_t *mos = dd->dd_pool->dp_meta_objset;
2269 	ASSERT(spa_feature_is_active(dd->dd_pool->dp_spa,
2270 	    SPA_FEATURE_LIVELIST));
2271 	dsl_deadlist_open(&dd->dd_livelist, mos, obj);
2272 	bplist_create(&dd->dd_pending_allocs);
2273 	bplist_create(&dd->dd_pending_frees);
2274 }
2275 
2276 void
2277 dsl_dir_livelist_close(dsl_dir_t *dd)
2278 {
2279 	dsl_deadlist_close(&dd->dd_livelist);
2280 	bplist_destroy(&dd->dd_pending_allocs);
2281 	bplist_destroy(&dd->dd_pending_frees);
2282 }
2283 
2284 void
2285 dsl_dir_remove_livelist(dsl_dir_t *dd, dmu_tx_t *tx, boolean_t total)
2286 {
2287 	uint64_t obj;
2288 	dsl_pool_t *dp = dmu_tx_pool(tx);
2289 	spa_t *spa = dp->dp_spa;
2290 	livelist_condense_entry_t to_condense = spa->spa_to_condense;
2291 
2292 	if (!dsl_deadlist_is_open(&dd->dd_livelist))
2293 		return;
2294 
2295 	/*
2296 	 * If the livelist being removed is set to be condensed, stop the
2297 	 * condense zthr and indicate the cancellation in the spa_to_condense
2298 	 * struct in case the condense no-wait synctask has already started
2299 	 */
2300 	zthr_t *ll_condense_thread = spa->spa_livelist_condense_zthr;
2301 	if (ll_condense_thread != NULL &&
2302 	    (to_condense.ds != NULL) && (to_condense.ds->ds_dir == dd)) {
2303 		/*
2304 		 * We use zthr_wait_cycle_done instead of zthr_cancel
2305 		 * because we don't want to destroy the zthr, just have
2306 		 * it skip its current task.
2307 		 */
2308 		spa->spa_to_condense.cancelled = B_TRUE;
2309 		zthr_wait_cycle_done(ll_condense_thread);
2310 		/*
2311 		 * If we've returned from zthr_wait_cycle_done without
2312 		 * clearing the to_condense data structure it's either
2313 		 * because the no-wait synctask has started (which is
2314 		 * indicated by 'syncing' field of to_condense) and we
2315 		 * can expect it to clear to_condense on its own.
2316 		 * Otherwise, we returned before the zthr ran. The
2317 		 * checkfunc will now fail as cancelled == B_TRUE so we
2318 		 * can safely NULL out ds, allowing a different dir's
2319 		 * livelist to be condensed.
2320 		 *
2321 		 * We can be sure that the to_condense struct will not
2322 		 * be repopulated at this stage because both this
2323 		 * function and dsl_livelist_try_condense execute in
2324 		 * syncing context.
2325 		 */
2326 		if ((spa->spa_to_condense.ds != NULL) &&
2327 		    !spa->spa_to_condense.syncing) {
2328 			dmu_buf_rele(spa->spa_to_condense.ds->ds_dbuf,
2329 			    spa);
2330 			spa->spa_to_condense.ds = NULL;
2331 		}
2332 	}
2333 
2334 	dsl_dir_livelist_close(dd);
2335 	VERIFY0(zap_lookup(dp->dp_meta_objset, dd->dd_object,
2336 	    DD_FIELD_LIVELIST, sizeof (uint64_t), 1, &obj));
2337 	VERIFY0(zap_remove(dp->dp_meta_objset, dd->dd_object,
2338 	    DD_FIELD_LIVELIST, tx));
2339 	if (total) {
2340 		dsl_deadlist_free(dp->dp_meta_objset, obj, tx);
2341 		spa_feature_decr(spa, SPA_FEATURE_LIVELIST, tx);
2342 	}
2343 }
2344 
2345 static int
2346 dsl_dir_activity_in_progress(dsl_dir_t *dd, dsl_dataset_t *ds,
2347     zfs_wait_activity_t activity, boolean_t *in_progress)
2348 {
2349 	int error = 0;
2350 
2351 	ASSERT(MUTEX_HELD(&dd->dd_activity_lock));
2352 
2353 	switch (activity) {
2354 	case ZFS_WAIT_DELETEQ: {
2355 #ifdef _KERNEL
2356 		objset_t *os;
2357 		error = dmu_objset_from_ds(ds, &os);
2358 		if (error != 0)
2359 			break;
2360 
2361 		mutex_enter(&os->os_user_ptr_lock);
2362 		void *user = dmu_objset_get_user(os);
2363 		mutex_exit(&os->os_user_ptr_lock);
2364 		if (dmu_objset_type(os) != DMU_OST_ZFS ||
2365 		    user == NULL || zfs_get_vfs_flag_unmounted(os)) {
2366 			*in_progress = B_FALSE;
2367 			return (0);
2368 		}
2369 
2370 		uint64_t readonly = B_FALSE;
2371 		error = zfs_get_temporary_prop(ds, ZFS_PROP_READONLY, &readonly,
2372 		    NULL);
2373 
2374 		if (error != 0)
2375 			break;
2376 
2377 		if (readonly || !spa_writeable(dd->dd_pool->dp_spa)) {
2378 			*in_progress = B_FALSE;
2379 			return (0);
2380 		}
2381 
2382 		uint64_t count, unlinked_obj;
2383 		error = zap_lookup(os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET, 8, 1,
2384 		    &unlinked_obj);
2385 		if (error != 0) {
2386 			dsl_dataset_rele(ds, FTAG);
2387 			break;
2388 		}
2389 		error = zap_count(os, unlinked_obj, &count);
2390 
2391 		if (error == 0)
2392 			*in_progress = (count != 0);
2393 		break;
2394 #else
2395 		/*
2396 		 * The delete queue is ZPL specific, and libzpool doesn't have
2397 		 * it. It doesn't make sense to wait for it.
2398 		 */
2399 		*in_progress = B_FALSE;
2400 		break;
2401 #endif
2402 	}
2403 	default:
2404 		panic("unrecognized value for activity %d", activity);
2405 	}
2406 
2407 	return (error);
2408 }
2409 
2410 int
2411 dsl_dir_wait(dsl_dir_t *dd, dsl_dataset_t *ds, zfs_wait_activity_t activity,
2412     boolean_t *waited)
2413 {
2414 	int error = 0;
2415 	boolean_t in_progress;
2416 	dsl_pool_t *dp = dd->dd_pool;
2417 	for (;;) {
2418 		dsl_pool_config_enter(dp, FTAG);
2419 		error = dsl_dir_activity_in_progress(dd, ds, activity,
2420 		    &in_progress);
2421 		dsl_pool_config_exit(dp, FTAG);
2422 		if (error != 0 || !in_progress)
2423 			break;
2424 
2425 		*waited = B_TRUE;
2426 
2427 		if (cv_wait_sig(&dd->dd_activity_cv, &dd->dd_activity_lock) ==
2428 		    0 || dd->dd_activity_cancelled) {
2429 			error = SET_ERROR(EINTR);
2430 			break;
2431 		}
2432 	}
2433 	return (error);
2434 }
2435 
2436 void
2437 dsl_dir_cancel_waiters(dsl_dir_t *dd)
2438 {
2439 	mutex_enter(&dd->dd_activity_lock);
2440 	dd->dd_activity_cancelled = B_TRUE;
2441 	cv_broadcast(&dd->dd_activity_cv);
2442 	while (dd->dd_activity_waiters > 0)
2443 		cv_wait(&dd->dd_activity_cv, &dd->dd_activity_lock);
2444 	mutex_exit(&dd->dd_activity_lock);
2445 }
2446 
2447 #if defined(_KERNEL)
2448 EXPORT_SYMBOL(dsl_dir_set_quota);
2449 EXPORT_SYMBOL(dsl_dir_set_reservation);
2450 #endif
2451