xref: /illumos-gate/usr/src/uts/common/fs/zfs/dsl_dir.c (revision 04580fdfa39e6e9d80ef4c60bfcf6e8461687692)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/dmu.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dsl_dataset.h>
32 #include <sys/dsl_dir.h>
33 #include <sys/dsl_prop.h>
34 #include <sys/spa.h>
35 #include <sys/zap.h>
36 #include <sys/zio.h>
37 #include <sys/arc.h>
38 #include "zfs_namecheck.h"
39 
40 static uint64_t dsl_dir_space_accounted(dsl_dir_t *dd);
41 static uint64_t dsl_dir_estimated_space(dsl_dir_t *dd);
42 static int dsl_dir_set_reservation_sync(dsl_dir_t *dd,
43     void *arg, dmu_tx_t *tx);
44 static uint64_t dsl_dir_space_available(dsl_dir_t *dd,
45     dsl_dir_t *ancestor, int64_t delta, int ondiskonly);
46 
47 
48 /* ARGSUSED */
49 static void
50 dsl_dir_evict(dmu_buf_t *db, void *arg)
51 {
52 	dsl_dir_t *dd = arg;
53 	dsl_pool_t *dp = dd->dd_pool;
54 	int t;
55 
56 	for (t = 0; t < TXG_SIZE; t++) {
57 		ASSERT(!txg_list_member(&dp->dp_dirty_dirs, dd, t));
58 		ASSERT(dd->dd_tempreserved[t] == 0);
59 		ASSERT(dd->dd_space_towrite[t] == 0);
60 	}
61 
62 	ASSERT3U(dd->dd_used_bytes, ==, dd->dd_phys->dd_used_bytes);
63 
64 	ASSERT(dd->dd_sync_txg == 0);
65 
66 	if (dd->dd_parent)
67 		dsl_dir_close(dd->dd_parent, dd);
68 
69 	spa_close(dd->dd_pool->dp_spa, dd);
70 
71 	/*
72 	 * The props callback list should be empty since they hold the
73 	 * dir open.
74 	 */
75 	list_destroy(&dd->dd_prop_cbs);
76 	kmem_free(dd, sizeof (dsl_dir_t));
77 }
78 
79 dsl_dir_t *
80 dsl_dir_open_obj(dsl_pool_t *dp, uint64_t ddobj,
81     const char *tail, void *tag)
82 {
83 	dmu_buf_t *dbuf;
84 	dsl_dir_t *dd;
85 
86 	ASSERT(RW_LOCK_HELD(&dp->dp_config_rwlock) ||
87 	    dsl_pool_sync_context(dp));
88 
89 	dbuf = dmu_bonus_hold_tag(dp->dp_meta_objset, ddobj, tag);
90 	dmu_buf_read(dbuf);
91 	dd = dmu_buf_get_user(dbuf);
92 #ifdef ZFS_DEBUG
93 	{
94 		dmu_object_info_t doi;
95 		dmu_object_info_from_db(dbuf, &doi);
96 		ASSERT3U(doi.doi_type, ==, DMU_OT_DSL_DIR);
97 	}
98 #endif
99 	/* XXX assert bonus buffer size is correct */
100 	if (dd == NULL) {
101 		dsl_dir_t *winner;
102 		int err;
103 
104 		dd = kmem_zalloc(sizeof (dsl_dir_t), KM_SLEEP);
105 		dd->dd_object = ddobj;
106 		dd->dd_dbuf = dbuf;
107 		dd->dd_pool = dp;
108 		dd->dd_phys = dbuf->db_data;
109 		dd->dd_used_bytes = dd->dd_phys->dd_used_bytes;
110 
111 		list_create(&dd->dd_prop_cbs, sizeof (dsl_prop_cb_record_t),
112 		    offsetof(dsl_prop_cb_record_t, cbr_node));
113 
114 		if (dd->dd_phys->dd_parent_obj) {
115 			dd->dd_parent = dsl_dir_open_obj(dp,
116 			    dd->dd_phys->dd_parent_obj, NULL, dd);
117 			if (tail) {
118 #ifdef ZFS_DEBUG
119 				uint64_t foundobj;
120 
121 				err = zap_lookup(dp->dp_meta_objset,
122 				    dd->dd_parent->dd_phys->
123 				    dd_child_dir_zapobj,
124 				    tail, sizeof (foundobj), 1, &foundobj);
125 				ASSERT3U(err, ==, 0);
126 				ASSERT3U(foundobj, ==, ddobj);
127 #endif
128 				(void) strcpy(dd->dd_myname, tail);
129 			} else {
130 				err = zap_value_search(dp->dp_meta_objset,
131 				    dd->dd_parent->dd_phys->
132 				    dd_child_dir_zapobj,
133 				    ddobj, dd->dd_myname);
134 				/*
135 				 * The caller should be protecting this ddobj
136 				 * from being deleted concurrently
137 				 */
138 				ASSERT(err == 0);
139 			}
140 		} else {
141 			(void) strcpy(dd->dd_myname, spa_name(dp->dp_spa));
142 		}
143 
144 		winner = dmu_buf_set_user_ie(dbuf, dd, &dd->dd_phys,
145 		    dsl_dir_evict);
146 		if (winner) {
147 			if (dd->dd_parent)
148 				dsl_dir_close(dd->dd_parent, dd);
149 			kmem_free(dd, sizeof (dsl_dir_t));
150 			dd = winner;
151 		} else {
152 			spa_open_ref(dp->dp_spa, dd);
153 		}
154 	}
155 
156 	/*
157 	 * The dsl_dir_t has both open-to-close and instantiate-to-evict
158 	 * holds on the spa.  We need the open-to-close holds because
159 	 * otherwise the spa_refcnt wouldn't change when we open a
160 	 * dir which the spa also has open, so we could incorrectly
161 	 * think it was OK to unload/export/destroy the pool.  We need
162 	 * the instantiate-to-evict hold because the dsl_dir_t has a
163 	 * pointer to the dd_pool, which has a pointer to the spa_t.
164 	 */
165 	spa_open_ref(dp->dp_spa, tag);
166 	ASSERT3P(dd->dd_pool, ==, dp);
167 	ASSERT3U(dd->dd_object, ==, ddobj);
168 	ASSERT3P(dd->dd_dbuf, ==, dbuf);
169 	return (dd);
170 }
171 
172 void
173 dsl_dir_close(dsl_dir_t *dd, void *tag)
174 {
175 	dprintf_dd(dd, "%s\n", "");
176 	spa_close(dd->dd_pool->dp_spa, tag);
177 	dmu_buf_rele_tag(dd->dd_dbuf, tag);
178 }
179 
180 /* buf must be long enough (MAXNAMELEN should do) */
181 void
182 dsl_dir_name(dsl_dir_t *dd, char *buf)
183 {
184 	if (dd->dd_parent) {
185 		dsl_dir_name(dd->dd_parent, buf);
186 		(void) strcat(buf, "/");
187 	} else {
188 		buf[0] = '\0';
189 	}
190 	if (!MUTEX_HELD(&dd->dd_lock)) {
191 		/*
192 		 * recursive mutex so that we can use
193 		 * dprintf_dd() with dd_lock held
194 		 */
195 		mutex_enter(&dd->dd_lock);
196 		(void) strcat(buf, dd->dd_myname);
197 		mutex_exit(&dd->dd_lock);
198 	} else {
199 		(void) strcat(buf, dd->dd_myname);
200 	}
201 }
202 
203 int
204 dsl_dir_is_private(dsl_dir_t *dd)
205 {
206 	int rv = FALSE;
207 
208 	if (dd->dd_parent && dsl_dir_is_private(dd->dd_parent))
209 		rv = TRUE;
210 	if (dataset_name_hidden(dd->dd_myname))
211 		rv = TRUE;
212 	return (rv);
213 }
214 
215 
216 static int
217 getcomponent(const char *path, char *component, const char **nextp)
218 {
219 	char *p;
220 	if (path == NULL)
221 		return (NULL);
222 	/* This would be a good place to reserve some namespace... */
223 	p = strpbrk(path, "/@");
224 	if (p && (p[1] == '/' || p[1] == '@')) {
225 		/* two separators in a row */
226 		return (EINVAL);
227 	}
228 	if (p == NULL || p == path) {
229 		/*
230 		 * if the first thing is an @ or /, it had better be an
231 		 * @ and it had better not have any more ats or slashes,
232 		 * and it had better have something after the @.
233 		 */
234 		if (p != NULL &&
235 		    (p[0] != '@' || strpbrk(path+1, "/@") || p[1] == '\0'))
236 			return (EINVAL);
237 		if (strlen(path) >= MAXNAMELEN)
238 			return (ENAMETOOLONG);
239 		(void) strcpy(component, path);
240 		p = NULL;
241 	} else if (p[0] == '/') {
242 		if (p-path >= MAXNAMELEN)
243 			return (ENAMETOOLONG);
244 		(void) strncpy(component, path, p - path);
245 		component[p-path] = '\0';
246 		p++;
247 	} else if (p[0] == '@') {
248 		/*
249 		 * if the next separator is an @, there better not be
250 		 * any more slashes.
251 		 */
252 		if (strchr(path, '/'))
253 			return (EINVAL);
254 		if (p-path >= MAXNAMELEN)
255 			return (ENAMETOOLONG);
256 		(void) strncpy(component, path, p - path);
257 		component[p-path] = '\0';
258 	} else {
259 		ASSERT(!"invalid p");
260 	}
261 	*nextp = p;
262 	return (0);
263 }
264 
265 /*
266  * same as dsl_open_dir, ignore the first component of name and use the
267  * spa instead
268  */
269 dsl_dir_t *
270 dsl_dir_open_spa(spa_t *spa, const char *name, void *tag, const char **tailp)
271 {
272 	char buf[MAXNAMELEN];
273 	const char *next, *nextnext = NULL;
274 	int err;
275 	dsl_dir_t *dd;
276 	dsl_pool_t *dp;
277 	uint64_t ddobj;
278 	int openedspa = FALSE;
279 
280 	dprintf("%s\n", name);
281 
282 	if (name == NULL)
283 		return (NULL);
284 	err = getcomponent(name, buf, &next);
285 	if (err)
286 		return (NULL);
287 	if (spa == NULL) {
288 		err = spa_open(buf, &spa, FTAG);
289 		if (err) {
290 			dprintf("spa_open(%s) failed\n", buf);
291 			return (NULL);
292 		}
293 		openedspa = TRUE;
294 
295 		/* XXX this assertion belongs in spa_open */
296 		ASSERT(!dsl_pool_sync_context(spa_get_dsl(spa)));
297 	}
298 
299 	dp = spa_get_dsl(spa);
300 
301 	rw_enter(&dp->dp_config_rwlock, RW_READER);
302 	dd = dsl_dir_open_obj(dp, dp->dp_root_dir_obj, NULL, tag);
303 	while (next != NULL) {
304 		dsl_dir_t *child_ds;
305 		err = getcomponent(next, buf, &nextnext);
306 		if (err) {
307 			dsl_dir_close(dd, tag);
308 			rw_exit(&dp->dp_config_rwlock);
309 			if (openedspa)
310 				spa_close(spa, FTAG);
311 			return (NULL);
312 		}
313 		ASSERT(next[0] != '\0');
314 		if (next[0] == '@')
315 			break;
316 		if (dd->dd_phys->dd_child_dir_zapobj == 0)
317 			break;
318 		dprintf("looking up %s in obj%lld\n",
319 		    buf, dd->dd_phys->dd_child_dir_zapobj);
320 
321 		err = zap_lookup(dp->dp_meta_objset,
322 		    dd->dd_phys->dd_child_dir_zapobj,
323 		    buf, sizeof (ddobj), 1, &ddobj);
324 		if (err == ENOENT) {
325 			break;
326 		}
327 		ASSERT(err == 0);
328 
329 		child_ds = dsl_dir_open_obj(dp, ddobj, buf, tag);
330 		dsl_dir_close(dd, tag);
331 		dd = child_ds;
332 		next = nextnext;
333 	}
334 	rw_exit(&dp->dp_config_rwlock);
335 
336 	/*
337 	 * It's an error if there's more than one component left, or
338 	 * tailp==NULL and there's any component left.
339 	 */
340 	if (next != NULL &&
341 	    (tailp == NULL || (nextnext && nextnext[0] != '\0'))) {
342 		/* bad path name */
343 		dsl_dir_close(dd, tag);
344 		dprintf("next=%p (%s) tail=%p\n", next, next?next:"", tailp);
345 		next = NULL;
346 		dd = NULL;
347 	}
348 	if (tailp)
349 		*tailp = next;
350 	if (openedspa)
351 		spa_close(spa, FTAG);
352 	return (dd);
353 }
354 
355 /*
356  * Return the dsl_dir_t, and possibly the last component which couldn't
357  * be found in *tail.  Return NULL if the path is bogus, or if
358  * tail==NULL and we couldn't parse the whole name.  (*tail)[0] == '@'
359  * means that the last component is a snapshot.
360  */
361 dsl_dir_t *
362 dsl_dir_open(const char *name, void *tag, const char **tailp)
363 {
364 	return (dsl_dir_open_spa(NULL, name, tag, tailp));
365 }
366 
367 int
368 dsl_dir_create_sync(dsl_dir_t *pds, const char *name, dmu_tx_t *tx)
369 {
370 	objset_t *mos = pds->dd_pool->dp_meta_objset;
371 	uint64_t ddobj;
372 	dsl_dir_phys_t *dsphys;
373 	dmu_buf_t *dbuf;
374 	int err;
375 
376 	ASSERT(dmu_tx_is_syncing(tx));
377 
378 	if (pds->dd_phys->dd_child_dir_zapobj == 0) {
379 		dmu_buf_will_dirty(pds->dd_dbuf, tx);
380 		pds->dd_phys->dd_child_dir_zapobj = zap_create(mos,
381 		    DMU_OT_DSL_DIR_CHILD_MAP, DMU_OT_NONE, 0, tx);
382 	}
383 
384 	rw_enter(&pds->dd_pool->dp_config_rwlock, RW_WRITER);
385 	err = zap_lookup(mos, pds->dd_phys->dd_child_dir_zapobj,
386 	    name, sizeof (uint64_t), 1, &ddobj);
387 	if (err != ENOENT) {
388 		rw_exit(&pds->dd_pool->dp_config_rwlock);
389 		return (err ? err : EEXIST);
390 	}
391 
392 	ddobj = dmu_object_alloc(mos, DMU_OT_DSL_DIR, 0,
393 	    DMU_OT_DSL_DIR, sizeof (dsl_dir_phys_t), tx);
394 	err = zap_add(mos, pds->dd_phys->dd_child_dir_zapobj,
395 	    name, sizeof (uint64_t), 1, &ddobj, tx);
396 	ASSERT3U(err, ==, 0);
397 	dprintf("dataset_create: zap_add %s->%lld to %lld returned %d\n",
398 	    name, ddobj, pds->dd_phys->dd_child_dir_zapobj, err);
399 
400 	dbuf = dmu_bonus_hold(mos, ddobj);
401 	dmu_buf_will_dirty(dbuf, tx);
402 	dsphys = dbuf->db_data;
403 
404 	dsphys->dd_creation_time = gethrestime_sec();
405 	dsphys->dd_parent_obj = pds->dd_object;
406 	dsphys->dd_props_zapobj = zap_create(mos,
407 	    DMU_OT_DSL_PROPS, DMU_OT_NONE, 0, tx);
408 	dsphys->dd_child_dir_zapobj = zap_create(mos,
409 	    DMU_OT_DSL_DIR_CHILD_MAP, DMU_OT_NONE, 0, tx);
410 	dmu_buf_rele(dbuf);
411 
412 	rw_exit(&pds->dd_pool->dp_config_rwlock);
413 
414 	return (0);
415 }
416 
417 int
418 dsl_dir_destroy_sync(dsl_dir_t *pds, void *arg, dmu_tx_t *tx)
419 {
420 	const char *name = arg;
421 	dsl_dir_t *dd = NULL;
422 	dsl_pool_t *dp = pds->dd_pool;
423 	objset_t *mos = dp->dp_meta_objset;
424 	uint64_t val, obj, child_zapobj, props_zapobj;
425 	int t, err;
426 
427 	rw_enter(&dp->dp_config_rwlock, RW_WRITER);
428 
429 	err = zap_lookup(mos, pds->dd_phys->dd_child_dir_zapobj, name,
430 	    8, 1, &obj);
431 	if (err)
432 		goto out;
433 
434 	dd = dsl_dir_open_obj(dp, obj, name, FTAG);
435 	ASSERT3U(dd->dd_phys->dd_parent_obj, ==, pds->dd_object);
436 
437 	if (dmu_buf_refcount(dd->dd_dbuf) > 1) {
438 		err = EBUSY;
439 		goto out;
440 	}
441 
442 	for (t = 0; t < TXG_SIZE; t++) {
443 		/*
444 		 * if they were dirty, they'd also be open.
445 		 * dp_config_rwlock ensures that it stays that way.
446 		 */
447 		ASSERT(!txg_list_member(&dp->dp_dirty_dirs, dd, t));
448 	}
449 
450 	child_zapobj = dd->dd_phys->dd_child_dir_zapobj;
451 	props_zapobj = dd->dd_phys->dd_props_zapobj;
452 
453 	if (child_zapobj != 0) {
454 		uint64_t count;
455 		err = EEXIST;
456 		(void) zap_count(mos, child_zapobj, &count);
457 		if (count != 0)
458 			goto out;
459 	}
460 
461 	if (dd->dd_phys->dd_head_dataset_obj != 0) {
462 		err = dsl_dataset_destroy_sync(dd, NULL, tx);
463 		if (err)
464 			goto out;
465 	}
466 	ASSERT(dd->dd_phys->dd_head_dataset_obj == 0);
467 
468 	/* The point of no (unsuccessful) return */
469 
470 	/* Make sure parent's used gets updated */
471 	val = 0;
472 	err = dsl_dir_set_reservation_sync(dd, &val, tx);
473 	ASSERT(err == 0);
474 	ASSERT3U(dd->dd_used_bytes, ==, 0);
475 	ASSERT3U(dd->dd_phys->dd_reserved, ==, 0);
476 	dsl_dir_close(dd, FTAG);
477 	dd = NULL;
478 
479 	err = dmu_object_free(mos, obj, tx);
480 	ASSERT(err == 0);
481 
482 	if (child_zapobj)
483 		err = zap_destroy(mos, child_zapobj, tx);
484 	ASSERT(err == 0);
485 
486 	if (props_zapobj)
487 		err = zap_destroy(mos, props_zapobj, tx);
488 	ASSERT(err == 0);
489 
490 	err = zap_remove(mos, pds->dd_phys->dd_child_dir_zapobj, name, tx);
491 	ASSERT(err == 0);
492 
493 out:
494 	rw_exit(&dp->dp_config_rwlock);
495 	if (dd)
496 		dsl_dir_close(dd, FTAG);
497 
498 	return (err);
499 }
500 
501 void
502 dsl_dir_create_root(objset_t *mos, uint64_t *ddobjp, dmu_tx_t *tx)
503 {
504 	dsl_dir_phys_t *dsp;
505 	dmu_buf_t *dbuf;
506 	int error;
507 
508 	*ddobjp = dmu_object_alloc(mos, DMU_OT_DSL_DIR, 0,
509 	    DMU_OT_DSL_DIR, sizeof (dsl_dir_phys_t), tx);
510 
511 	error = zap_add(mos, DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_ROOT_DATASET,
512 	    sizeof (uint64_t), 1, ddobjp, tx);
513 	ASSERT3U(error, ==, 0);
514 
515 	dbuf = dmu_bonus_hold(mos, *ddobjp);
516 	dmu_buf_will_dirty(dbuf, tx);
517 	dsp = dbuf->db_data;
518 
519 	dsp->dd_creation_time = gethrestime_sec();
520 	dsp->dd_props_zapobj = zap_create(mos,
521 	    DMU_OT_DSL_PROPS, DMU_OT_NONE, 0, tx);
522 	dsp->dd_child_dir_zapobj = zap_create(mos,
523 	    DMU_OT_DSL_DIR_CHILD_MAP, DMU_OT_NONE, 0, tx);
524 
525 	dmu_buf_rele(dbuf);
526 }
527 
528 void
529 dsl_dir_stats(dsl_dir_t *dd, dmu_objset_stats_t *dds)
530 {
531 	bzero(dds, sizeof (dmu_objset_stats_t));
532 
533 	dds->dds_dir_obj = dd->dd_object;
534 	dds->dds_available = dsl_dir_space_available(dd, NULL, 0, TRUE);
535 
536 	mutex_enter(&dd->dd_lock);
537 	dds->dds_space_used = dd->dd_used_bytes;
538 	dds->dds_compressed_bytes = dd->dd_phys->dd_compressed_bytes;
539 	dds->dds_uncompressed_bytes = dd->dd_phys->dd_uncompressed_bytes;
540 	dds->dds_quota = dd->dd_phys->dd_quota;
541 	dds->dds_reserved = dd->dd_phys->dd_reserved;
542 	mutex_exit(&dd->dd_lock);
543 
544 	dds->dds_creation_time = dd->dd_phys->dd_creation_time;
545 
546 	dds->dds_is_placeholder = (dd->dd_phys->dd_head_dataset_obj == 0);
547 
548 	if (dd->dd_phys->dd_clone_parent_obj) {
549 		dsl_dataset_t *ds;
550 
551 		rw_enter(&dd->dd_pool->dp_config_rwlock, RW_READER);
552 		ds = dsl_dataset_open_obj(dd->dd_pool,
553 		    dd->dd_phys->dd_clone_parent_obj, NULL, DS_MODE_NONE, FTAG);
554 		dsl_dataset_name(ds, dds->dds_clone_of);
555 		dds->dds_clone_of_obj = dd->dd_phys->dd_clone_parent_obj;
556 		dsl_dataset_close(ds, DS_MODE_NONE, FTAG);
557 		rw_exit(&dd->dd_pool->dp_config_rwlock);
558 	}
559 
560 	VERIFY(dsl_prop_get_ds_integer(dd, "checksum",
561 	    &dds->dds_checksum, dds->dds_checksum_setpoint) == 0);
562 
563 	VERIFY(dsl_prop_get_ds_integer(dd, "compression",
564 	    &dds->dds_compression, dds->dds_compression_setpoint) == 0);
565 
566 	VERIFY(dsl_prop_get_ds_integer(dd, "zoned",
567 	    &dds->dds_zoned, dds->dds_zoned_setpoint) == 0);
568 
569 	spa_altroot(dd->dd_pool->dp_spa, dds->dds_altroot,
570 	    sizeof (dds->dds_altroot));
571 }
572 
573 int
574 dsl_dir_sync_task(dsl_dir_t *dd,
575     int (*func)(dsl_dir_t *, void*, dmu_tx_t *), void *arg, uint64_t space)
576 {
577 	dmu_tx_t *tx;
578 	dsl_pool_t *dp = dd->dd_pool;
579 	int err = 0;
580 	uint64_t txg;
581 
582 	dprintf_dd(dd, "func=%p space=%llu\n", func, space);
583 
584 again:
585 	tx = dmu_tx_create_ds(dd);
586 	dmu_tx_hold_space(tx, space);
587 	err = dmu_tx_assign(tx, TXG_WAIT);
588 	if (err == ENOSPC || err == EDQUOT) {
589 		dsl_dir_t *rds;
590 		/*
591 		 * They can get their space from either this dd, or the
592 		 * root dd.
593 		 */
594 		for (rds = dd; rds->dd_parent; rds = rds->dd_parent)
595 			continue;
596 		dmu_tx_abort(tx);
597 		tx = dmu_tx_create_ds(rds);
598 		dmu_tx_hold_space(tx, space);
599 		err = dmu_tx_assign(tx, TXG_WAIT);
600 	}
601 	if (err) {
602 		dmu_tx_abort(tx);
603 		return (err);
604 	}
605 
606 	txg = dmu_tx_get_txg(tx);
607 	mutex_enter(&dd->dd_lock);
608 	if (dd->dd_sync_txg != 0) {
609 		mutex_exit(&dd->dd_lock);
610 		dmu_tx_commit(tx);
611 		txg_wait_synced(dp, 0);
612 		goto again;
613 	}
614 
615 	/* We're good to go */
616 
617 	dd->dd_sync_txg = txg;
618 	dd->dd_sync_func = func;
619 	dd->dd_sync_arg = arg;
620 
621 	mutex_exit(&dd->dd_lock);
622 
623 	dsl_dir_dirty(dd, tx);
624 	dmu_tx_commit(tx);
625 
626 	txg_wait_synced(dp, txg);
627 
628 	mutex_enter(&dd->dd_lock);
629 	ASSERT(dd->dd_sync_txg == txg);
630 	ASSERT(dd->dd_sync_func == NULL);
631 	err = dd->dd_sync_err;
632 	dd->dd_sync_txg = 0;
633 	mutex_exit(&dd->dd_lock);
634 
635 	return (err);
636 }
637 
638 void
639 dsl_dir_dirty(dsl_dir_t *dd, dmu_tx_t *tx)
640 {
641 	dsl_pool_t *dp = dd->dd_pool;
642 
643 	ASSERT(dd->dd_phys);
644 
645 	if (txg_list_add(&dp->dp_dirty_dirs, dd, tx->tx_txg) == 0) {
646 		/* up the hold count until we can be written out */
647 		dmu_buf_add_ref(dd->dd_dbuf, dd);
648 	}
649 }
650 
651 static int64_t
652 parent_delta(dsl_dir_t *dd, uint64_t used, int64_t delta)
653 {
654 	uint64_t old_accounted = MAX(used, dd->dd_phys->dd_reserved);
655 	uint64_t new_accounted = MAX(used + delta, dd->dd_phys->dd_reserved);
656 	return (new_accounted - old_accounted);
657 }
658 
659 void
660 dsl_dir_sync(dsl_dir_t *dd, dmu_tx_t *tx)
661 {
662 	if (dd->dd_sync_txg == tx->tx_txg && dd->dd_sync_func) {
663 		dd->dd_sync_err = dd->dd_sync_func(dd, dd->dd_sync_arg, tx);
664 		dd->dd_sync_func = NULL;
665 	}
666 
667 	ASSERT(dmu_tx_is_syncing(tx));
668 
669 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
670 
671 	mutex_enter(&dd->dd_lock);
672 	ASSERT3U(dd->dd_tempreserved[tx->tx_txg&TXG_MASK], ==, 0);
673 	dprintf_dd(dd, "txg=%llu towrite=%lluK\n", tx->tx_txg,
674 	    dd->dd_space_towrite[tx->tx_txg&TXG_MASK] / 1024);
675 	dd->dd_space_towrite[tx->tx_txg&TXG_MASK] = 0;
676 	dd->dd_phys->dd_used_bytes = dd->dd_used_bytes;
677 	mutex_exit(&dd->dd_lock);
678 
679 	/* release the hold from dsl_dir_dirty */
680 	dmu_buf_remove_ref(dd->dd_dbuf, dd);
681 }
682 
683 static uint64_t
684 dsl_dir_estimated_space(dsl_dir_t *dd)
685 {
686 	int64_t space;
687 	int i;
688 
689 	ASSERT(MUTEX_HELD(&dd->dd_lock));
690 
691 	space = dd->dd_used_bytes;
692 	ASSERT(space >= 0);
693 	for (i = 0; i < TXG_SIZE; i++) {
694 		space += dd->dd_space_towrite[i&TXG_MASK];
695 		ASSERT3U(dd->dd_space_towrite[i&TXG_MASK], >=, 0);
696 	}
697 	return (space);
698 }
699 
700 /*
701  * How much space would dd have available if ancestor had delta applied
702  * to it?  If ondiskonly is set, we're only interested in what's
703  * on-disk, not estimated pending changes.
704  */
705 static uint64_t
706 dsl_dir_space_available(dsl_dir_t *dd,
707     dsl_dir_t *ancestor, int64_t delta, int ondiskonly)
708 {
709 	uint64_t parentspace, myspace, quota, used;
710 
711 	/*
712 	 * If there are no restrictions otherwise, assume we have
713 	 * unlimited space available.
714 	 */
715 	quota = UINT64_MAX;
716 	parentspace = UINT64_MAX;
717 
718 	if (dd->dd_parent != NULL) {
719 		parentspace = dsl_dir_space_available(dd->dd_parent,
720 		    ancestor, delta, ondiskonly);
721 	}
722 
723 	mutex_enter(&dd->dd_lock);
724 	if (dd->dd_phys->dd_quota != 0)
725 		quota = dd->dd_phys->dd_quota;
726 	if (ondiskonly) {
727 		used = dd->dd_used_bytes;
728 	} else {
729 		used = dsl_dir_estimated_space(dd);
730 	}
731 	if (dd == ancestor)
732 		used += delta;
733 
734 	if (dd->dd_parent == NULL) {
735 		uint64_t poolsize = dsl_pool_adjustedsize(dd->dd_pool, B_FALSE);
736 		quota = MIN(quota, poolsize);
737 	}
738 
739 	if (dd->dd_phys->dd_reserved > used && parentspace != UINT64_MAX) {
740 		/*
741 		 * We have some space reserved, in addition to what our
742 		 * parent gave us.
743 		 */
744 		parentspace += dd->dd_phys->dd_reserved - used;
745 	}
746 
747 	if (used > quota) {
748 		/* over quota */
749 		myspace = 0;
750 #ifdef ZFS_DEBUG
751 		{
752 			/*
753 			 * While it's OK to be a little over quota, if
754 			 * we think we are using more space than there
755 			 * is in the pool (which is already 6% more than
756 			 * dsl_pool_adjustedsize()), something is very
757 			 * wrong.
758 			 */
759 			uint64_t space = spa_get_space(dd->dd_pool->dp_spa);
760 			ASSERT3U(used, <=, space);
761 		}
762 #endif
763 	} else {
764 		/*
765 		 * the lesser of parent's space and the space
766 		 * left in our quota
767 		 */
768 		myspace = MIN(parentspace, quota - used);
769 	}
770 
771 	mutex_exit(&dd->dd_lock);
772 
773 	return (myspace);
774 }
775 
776 struct tempreserve {
777 	list_node_t tr_node;
778 	dsl_dir_t *tr_ds;
779 	uint64_t tr_size;
780 };
781 
782 /*
783  * Reserve space in this dsl_dir, to be used in this tx's txg.
784  * After the space has been dirtied (and thus
785  * dsl_dir_willuse_space() has been called), the reservation should
786  * be canceled, using dsl_dir_tempreserve_clear().
787  */
788 static int
789 dsl_dir_tempreserve_impl(dsl_dir_t *dd,
790     uint64_t asize, boolean_t netfree, list_t *tr_list, dmu_tx_t *tx)
791 {
792 	uint64_t txg = tx->tx_txg;
793 	uint64_t est_used, quota, parent_rsrv;
794 	int edquot = EDQUOT;
795 	int txgidx = txg & TXG_MASK;
796 	int i;
797 	struct tempreserve *tr;
798 
799 	ASSERT3U(txg, !=, 0);
800 
801 	mutex_enter(&dd->dd_lock);
802 	/*
803 	 * Check against the dsl_dir's quota.  We don't add in the delta
804 	 * when checking for over-quota because they get one free hit.
805 	 */
806 	est_used = dsl_dir_estimated_space(dd);
807 	for (i = 0; i < TXG_SIZE; i++)
808 		est_used += dd->dd_tempreserved[i];
809 
810 	quota = UINT64_MAX;
811 
812 	if (dd->dd_phys->dd_quota)
813 		quota = dd->dd_phys->dd_quota;
814 
815 	/*
816 	 * If this transaction will result in a net free of space, we want
817 	 * to let it through, but we have to be careful: the space that it
818 	 * frees won't become available until *after* this txg syncs.
819 	 * Therefore, to ensure that it's possible to remove files from
820 	 * a full pool without inducing transient overcommits, we throttle
821 	 * netfree transactions against a quota that is slightly larger,
822 	 * but still within the pool's allocation slop.  In cases where
823 	 * we're very close to full, this will allow a steady trickle of
824 	 * removes to get through.
825 	 */
826 	if (dd->dd_parent == NULL) {
827 		uint64_t poolsize = dsl_pool_adjustedsize(dd->dd_pool, netfree);
828 		if (poolsize < quota) {
829 			quota = poolsize;
830 			edquot = ENOSPC;
831 		}
832 	} else if (netfree) {
833 		quota = UINT64_MAX;
834 	}
835 
836 	/*
837 	 * If they are requesting more space, and our current estimate
838 	 * is over quota.  They get to try again unless the actual
839 	 * on-disk is over quota.
840 	 */
841 	if (asize > 0 && est_used > quota) {
842 		if (dd->dd_used_bytes < quota)
843 			edquot = ERESTART;
844 		dprintf_dd(dd, "failing: used=%lluK est_used = %lluK "
845 		    "quota=%lluK tr=%lluK err=%d\n",
846 		    dd->dd_used_bytes>>10, est_used>>10,
847 		    quota>>10, asize>>10, edquot);
848 		mutex_exit(&dd->dd_lock);
849 		return (edquot);
850 	}
851 
852 	/* We need to up our estimated delta before dropping dd_lock */
853 	dd->dd_tempreserved[txgidx] += asize;
854 
855 	parent_rsrv = parent_delta(dd, est_used, asize);
856 	mutex_exit(&dd->dd_lock);
857 
858 	tr = kmem_alloc(sizeof (struct tempreserve), KM_SLEEP);
859 	tr->tr_ds = dd;
860 	tr->tr_size = asize;
861 	list_insert_tail(tr_list, tr);
862 
863 	/* see if it's OK with our parent */
864 	if (dd->dd_parent && parent_rsrv) {
865 		return (dsl_dir_tempreserve_impl(dd->dd_parent,
866 		    parent_rsrv, netfree, tr_list, tx));
867 	} else {
868 		return (0);
869 	}
870 }
871 
872 /*
873  * Reserve space in this dsl_dir, to be used in this tx's txg.
874  * After the space has been dirtied (and thus
875  * dsl_dir_willuse_space() has been called), the reservation should
876  * be canceled, using dsl_dir_tempreserve_clear().
877  */
878 int
879 dsl_dir_tempreserve_space(dsl_dir_t *dd, uint64_t lsize,
880     uint64_t asize, uint64_t fsize, void **tr_cookiep, dmu_tx_t *tx)
881 {
882 	int err = 0;
883 	list_t *tr_list;
884 
885 	tr_list = kmem_alloc(sizeof (list_t), KM_SLEEP);
886 	list_create(tr_list, sizeof (struct tempreserve),
887 	    offsetof(struct tempreserve, tr_node));
888 
889 	err = dsl_dir_tempreserve_impl(dd, asize, fsize >= asize,
890 	    tr_list, tx);
891 
892 	if (err == 0) {
893 		struct tempreserve *tr;
894 
895 		err = arc_tempreserve_space(lsize);
896 		if (err == 0) {
897 			tr = kmem_alloc(sizeof (struct tempreserve), KM_SLEEP);
898 			tr->tr_ds = NULL;
899 			tr->tr_size = lsize;
900 			list_insert_tail(tr_list, tr);
901 		}
902 	}
903 
904 	if (err)
905 		dsl_dir_tempreserve_clear(tr_list, tx);
906 	else
907 		*tr_cookiep = tr_list;
908 	return (err);
909 }
910 
911 /*
912  * Clear a temporary reservation that we previously made with
913  * dsl_dir_tempreserve_space().
914  */
915 void
916 dsl_dir_tempreserve_clear(void *tr_cookie, dmu_tx_t *tx)
917 {
918 	int txgidx = tx->tx_txg & TXG_MASK;
919 	list_t *tr_list = tr_cookie;
920 	struct tempreserve *tr;
921 
922 	ASSERT3U(tx->tx_txg, !=, 0);
923 
924 	while (tr = list_head(tr_list)) {
925 		if (tr->tr_ds == NULL) {
926 			arc_tempreserve_clear(tr->tr_size);
927 		} else {
928 			mutex_enter(&tr->tr_ds->dd_lock);
929 			ASSERT3U(tr->tr_ds->dd_tempreserved[txgidx], >=,
930 			    tr->tr_size);
931 			tr->tr_ds->dd_tempreserved[txgidx] -= tr->tr_size;
932 			mutex_exit(&tr->tr_ds->dd_lock);
933 		}
934 		list_remove(tr_list, tr);
935 		kmem_free(tr, sizeof (struct tempreserve));
936 	}
937 
938 	kmem_free(tr_list, sizeof (list_t));
939 }
940 
941 /*
942  * Call in open context when we think we're going to write/free space,
943  * eg. when dirtying data.  Be conservative (ie. OK to write less than
944  * this or free more than this, but don't write more or free less).
945  */
946 void
947 dsl_dir_willuse_space(dsl_dir_t *dd, int64_t space, dmu_tx_t *tx)
948 {
949 	int64_t parent_space;
950 	uint64_t est_used;
951 
952 	mutex_enter(&dd->dd_lock);
953 	if (space > 0)
954 		dd->dd_space_towrite[tx->tx_txg & TXG_MASK] += space;
955 
956 	est_used = dsl_dir_estimated_space(dd);
957 	parent_space = parent_delta(dd, est_used, space);
958 	mutex_exit(&dd->dd_lock);
959 
960 	/* Make sure that we clean up dd_space_to* */
961 	dsl_dir_dirty(dd, tx);
962 
963 	/* XXX this is potentially expensive and unnecessary... */
964 	if (parent_space && dd->dd_parent)
965 		dsl_dir_willuse_space(dd->dd_parent, parent_space, tx);
966 }
967 
968 /* call from syncing context when we actually write/free space for this dd */
969 void
970 dsl_dir_diduse_space(dsl_dir_t *dd,
971     int64_t used, int64_t compressed, int64_t uncompressed, dmu_tx_t *tx)
972 {
973 	int64_t accounted_delta;
974 
975 	ASSERT(dmu_tx_is_syncing(tx));
976 
977 	dsl_dir_dirty(dd, tx);
978 
979 	mutex_enter(&dd->dd_lock);
980 	accounted_delta = parent_delta(dd, dd->dd_used_bytes, used);
981 	ASSERT(used >= 0 || dd->dd_used_bytes >= -used);
982 	ASSERT(compressed >= 0 ||
983 	    dd->dd_phys->dd_compressed_bytes >= -compressed);
984 	ASSERT(uncompressed >= 0 ||
985 	    dd->dd_phys->dd_uncompressed_bytes >= -uncompressed);
986 	dd->dd_used_bytes += used;
987 	if (used > 0)
988 		dd->dd_space_towrite[tx->tx_txg & TXG_MASK] -= used;
989 	dd->dd_phys->dd_uncompressed_bytes += uncompressed;
990 	dd->dd_phys->dd_compressed_bytes += compressed;
991 	mutex_exit(&dd->dd_lock);
992 
993 	if (dd->dd_parent != NULL) {
994 		dsl_dir_diduse_space(dd->dd_parent,
995 		    accounted_delta, compressed, uncompressed, tx);
996 	}
997 }
998 
999 static int
1000 dsl_dir_set_quota_sync(dsl_dir_t *dd, void *arg, dmu_tx_t *tx)
1001 {
1002 	uint64_t *quotap = arg;
1003 	uint64_t new_quota = *quotap;
1004 	int err = 0;
1005 
1006 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
1007 
1008 	mutex_enter(&dd->dd_lock);
1009 	if (new_quota != 0 && (new_quota < dd->dd_phys->dd_reserved ||
1010 	    new_quota < dsl_dir_estimated_space(dd))) {
1011 		err = ENOSPC;
1012 	} else {
1013 		dd->dd_phys->dd_quota = new_quota;
1014 	}
1015 	mutex_exit(&dd->dd_lock);
1016 	return (err);
1017 }
1018 
1019 int
1020 dsl_dir_set_quota(const char *ddname, uint64_t quota)
1021 {
1022 	dsl_dir_t *dd;
1023 	int err;
1024 
1025 	dd = dsl_dir_open(ddname, FTAG, NULL);
1026 	if (dd == NULL)
1027 		return (ENOENT);
1028 	/*
1029 	 * If someone removes a file, then tries to set the quota, we
1030 	 * want to make sure the file freeing takes effect.
1031 	 */
1032 	txg_wait_open(dd->dd_pool, 0);
1033 
1034 	err = dsl_dir_sync_task(dd, dsl_dir_set_quota_sync, &quota, 0);
1035 	dsl_dir_close(dd, FTAG);
1036 	return (err);
1037 }
1038 
1039 static int
1040 dsl_dir_set_reservation_sync(dsl_dir_t *dd, void *arg, dmu_tx_t *tx)
1041 {
1042 	uint64_t *reservationp = arg;
1043 	uint64_t new_reservation = *reservationp;
1044 	uint64_t used, avail;
1045 	int64_t delta;
1046 
1047 	if (new_reservation > INT64_MAX)
1048 		return (EOVERFLOW);
1049 
1050 	mutex_enter(&dd->dd_lock);
1051 	used = dd->dd_used_bytes;
1052 	delta = MAX(used, new_reservation) -
1053 	    MAX(used, dd->dd_phys->dd_reserved);
1054 	mutex_exit(&dd->dd_lock);
1055 
1056 	if (dd->dd_parent) {
1057 		avail = dsl_dir_space_available(dd->dd_parent,
1058 		    NULL, 0, FALSE);
1059 	} else {
1060 		avail = dsl_pool_adjustedsize(dd->dd_pool, B_FALSE) - used;
1061 	}
1062 
1063 	if (delta > 0 && delta > avail)
1064 		return (ENOSPC);
1065 	if (delta > 0 && dd->dd_phys->dd_quota > 0 &&
1066 	    new_reservation > dd->dd_phys->dd_quota)
1067 		return (ENOSPC);
1068 
1069 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
1070 	dd->dd_phys->dd_reserved = new_reservation;
1071 
1072 	if (dd->dd_parent != NULL) {
1073 		/* Roll up this additional usage into our ancestors */
1074 		dsl_dir_diduse_space(dd->dd_parent, delta, 0, 0, tx);
1075 	}
1076 	return (0);
1077 }
1078 
1079 int
1080 dsl_dir_set_reservation(const char *ddname, uint64_t reservation)
1081 {
1082 	dsl_dir_t *dd;
1083 	int err;
1084 
1085 	dd = dsl_dir_open(ddname, FTAG, NULL);
1086 	if (dd == NULL)
1087 		return (ENOENT);
1088 	err = dsl_dir_sync_task(dd,
1089 	    dsl_dir_set_reservation_sync, &reservation, 0);
1090 	dsl_dir_close(dd, FTAG);
1091 	return (err);
1092 }
1093 
1094 static dsl_dir_t *
1095 closest_common_ancestor(dsl_dir_t *ds1, dsl_dir_t *ds2)
1096 {
1097 	for (; ds1; ds1 = ds1->dd_parent) {
1098 		dsl_dir_t *dd;
1099 		for (dd = ds2; dd; dd = dd->dd_parent) {
1100 			if (ds1 == dd)
1101 				return (dd);
1102 		}
1103 	}
1104 	return (NULL);
1105 }
1106 
1107 /*
1108  * If delta is applied to dd, how much of that delta would be applied to
1109  * ancestor?  Syncing context only.
1110  */
1111 static int64_t
1112 would_change(dsl_dir_t *dd, int64_t delta, dsl_dir_t *ancestor)
1113 {
1114 	if (dd == ancestor)
1115 		return (delta);
1116 
1117 	mutex_enter(&dd->dd_lock);
1118 	delta = parent_delta(dd, dd->dd_used_bytes, delta);
1119 	mutex_exit(&dd->dd_lock);
1120 	return (would_change(dd->dd_parent, delta, ancestor));
1121 }
1122 
1123 int
1124 dsl_dir_rename_sync(dsl_dir_t *dd, void *arg, dmu_tx_t *tx)
1125 {
1126 	const char *newname = arg;
1127 	dsl_pool_t *dp = dd->dd_pool;
1128 	objset_t *mos = dp->dp_meta_objset;
1129 	dsl_dir_t *newpds;
1130 	const char *tail;
1131 	int err, len;
1132 
1133 	/* can't rename to different pool */
1134 	len = strlen(dp->dp_root_dir->dd_myname);
1135 	if (strncmp(dp->dp_root_dir->dd_myname, newname, len != 0) ||
1136 	    newname[len] != '/') {
1137 		return (ENXIO);
1138 	}
1139 
1140 	newpds = dsl_dir_open_spa(dp->dp_spa, newname, FTAG, &tail);
1141 
1142 	/* new parent should exist */
1143 	if (newpds == NULL)
1144 		return (ENOENT);
1145 
1146 	/* new name should not already exist */
1147 	if (tail == NULL) {
1148 		dsl_dir_close(newpds, FTAG);
1149 		return (EEXIST);
1150 	}
1151 
1152 	rw_enter(&dp->dp_config_rwlock, RW_WRITER);
1153 
1154 	/* There should be 2 references: the open and the dirty */
1155 	if (dmu_buf_refcount(dd->dd_dbuf) > 2) {
1156 		rw_exit(&dp->dp_config_rwlock);
1157 		dsl_dir_close(newpds, FTAG);
1158 		return (EBUSY);
1159 	}
1160 
1161 	if (newpds != dd->dd_parent) {
1162 		dsl_dir_t *ancestor;
1163 		int64_t adelta;
1164 		uint64_t myspace, avail;
1165 
1166 		ancestor = closest_common_ancestor(dd, newpds);
1167 
1168 		/* no rename into our descendent */
1169 		if (ancestor == dd) {
1170 			dsl_dir_close(newpds, FTAG);
1171 			rw_exit(&dp->dp_config_rwlock);
1172 			return (EINVAL);
1173 		}
1174 
1175 		myspace = MAX(dd->dd_used_bytes, dd->dd_phys->dd_reserved);
1176 		adelta = would_change(dd->dd_parent, -myspace, ancestor);
1177 		avail = dsl_dir_space_available(newpds,
1178 		    ancestor, adelta, FALSE);
1179 		if (avail < myspace) {
1180 			dsl_dir_close(newpds, FTAG);
1181 			rw_exit(&dp->dp_config_rwlock);
1182 			return (ENOSPC);
1183 		}
1184 
1185 		/* The point of no (unsuccessful) return */
1186 
1187 		dsl_dir_diduse_space(dd->dd_parent, -myspace,
1188 		    -dd->dd_phys->dd_compressed_bytes,
1189 		    -dd->dd_phys->dd_uncompressed_bytes, tx);
1190 		dsl_dir_diduse_space(newpds, myspace,
1191 		    dd->dd_phys->dd_compressed_bytes,
1192 		    dd->dd_phys->dd_uncompressed_bytes, tx);
1193 	}
1194 
1195 	/* The point of no (unsuccessful) return */
1196 
1197 	dmu_buf_will_dirty(dd->dd_dbuf, tx);
1198 
1199 	/* remove from old parent zapobj */
1200 	err = zap_remove(mos, dd->dd_parent->dd_phys->dd_child_dir_zapobj,
1201 	    dd->dd_myname, tx);
1202 	ASSERT3U(err, ==, 0);
1203 
1204 	(void) strcpy(dd->dd_myname, tail);
1205 	dsl_dir_close(dd->dd_parent, dd);
1206 	dd->dd_phys->dd_parent_obj = newpds->dd_object;
1207 	dd->dd_parent = dsl_dir_open_obj(dd->dd_pool,
1208 	    newpds->dd_object, NULL, dd);
1209 
1210 	/* add to new parent zapobj */
1211 	err = zap_add(mos, newpds->dd_phys->dd_child_dir_zapobj,
1212 	    dd->dd_myname, 8, 1, &dd->dd_object, tx);
1213 	ASSERT3U(err, ==, 0);
1214 
1215 	dsl_dir_close(newpds, FTAG);
1216 	rw_exit(&dp->dp_config_rwlock);
1217 	return (0);
1218 }
1219