xref: /illumos-gate/usr/src/uts/common/fs/zfs/dsl_pool.c (revision 69962b5647e4a8b9b14998733b765925381b727e)
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) 2013 by Delphix. All rights reserved.
24  * Copyright (c) 2013 Steven Hartland. All rights reserved.
25  */
26 
27 #include <sys/dsl_pool.h>
28 #include <sys/dsl_dataset.h>
29 #include <sys/dsl_prop.h>
30 #include <sys/dsl_dir.h>
31 #include <sys/dsl_synctask.h>
32 #include <sys/dsl_scan.h>
33 #include <sys/dnode.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/arc.h>
37 #include <sys/zap.h>
38 #include <sys/zio.h>
39 #include <sys/zfs_context.h>
40 #include <sys/fs/zfs.h>
41 #include <sys/zfs_znode.h>
42 #include <sys/spa_impl.h>
43 #include <sys/dsl_deadlist.h>
44 #include <sys/bptree.h>
45 #include <sys/zfeature.h>
46 #include <sys/zil_impl.h>
47 #include <sys/dsl_userhold.h>
48 
49 /*
50  * ZFS Write Throttle
51  * ------------------
52  *
53  * ZFS must limit the rate of incoming writes to the rate at which it is able
54  * to sync data modifications to the backend storage. Throttling by too much
55  * creates an artificial limit; throttling by too little can only be sustained
56  * for short periods and would lead to highly lumpy performance. On a per-pool
57  * basis, ZFS tracks the amount of modified (dirty) data. As operations change
58  * data, the amount of dirty data increases; as ZFS syncs out data, the amount
59  * of dirty data decreases. When the amount of dirty data exceeds a
60  * predetermined threshold further modifications are blocked until the amount
61  * of dirty data decreases (as data is synced out).
62  *
63  * The limit on dirty data is tunable, and should be adjusted according to
64  * both the IO capacity and available memory of the system. The larger the
65  * window, the more ZFS is able to aggregate and amortize metadata (and data)
66  * changes. However, memory is a limited resource, and allowing for more dirty
67  * data comes at the cost of keeping other useful data in memory (for example
68  * ZFS data cached by the ARC).
69  *
70  * Implementation
71  *
72  * As buffers are modified dsl_pool_willuse_space() increments both the per-
73  * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
74  * dirty space used; dsl_pool_dirty_space() decrements those values as data
75  * is synced out from dsl_pool_sync(). While only the poolwide value is
76  * relevant, the per-txg value is useful for debugging. The tunable
77  * zfs_dirty_data_max determines the dirty space limit. Once that value is
78  * exceeded, new writes are halted until space frees up.
79  *
80  * The zfs_dirty_data_sync tunable dictates the threshold at which we
81  * ensure that there is a txg syncing (see the comment in txg.c for a full
82  * description of transaction group stages).
83  *
84  * The IO scheduler uses both the dirty space limit and current amount of
85  * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
86  * issues. See the comment in vdev_queue.c for details of the IO scheduler.
87  *
88  * The delay is also calculated based on the amount of dirty data.  See the
89  * comment above dmu_tx_delay() for details.
90  */
91 
92 /*
93  * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
94  * capped at zfs_dirty_data_max_max.  It can also be overridden in /etc/system.
95  */
96 uint64_t zfs_dirty_data_max;
97 uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024;
98 int zfs_dirty_data_max_percent = 10;
99 
100 /*
101  * If there is at least this much dirty data, push out a txg.
102  */
103 uint64_t zfs_dirty_data_sync = 64 * 1024 * 1024;
104 
105 /*
106  * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
107  * and delay each transaction.
108  * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
109  */
110 int zfs_delay_min_dirty_percent = 60;
111 
112 /*
113  * This controls how quickly the delay approaches infinity.
114  * Larger values cause it to delay less for a given amount of dirty data.
115  * Therefore larger values will cause there to be more dirty data for a
116  * given throughput.
117  *
118  * For the smoothest delay, this value should be about 1 billion divided
119  * by the maximum number of operations per second.  This will smoothly
120  * handle between 10x and 1/10th this number.
121  *
122  * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
123  * multiply in dmu_tx_delay().
124  */
125 uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
126 
127 
128 /*
129  * XXX someday maybe turn these into #defines, and you have to tune it on a
130  * per-pool basis using zfs.conf.
131  */
132 
133 
134 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
135 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
136 
137 int
138 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
139 {
140 	uint64_t obj;
141 	int err;
142 
143 	err = zap_lookup(dp->dp_meta_objset,
144 	    dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
145 	    name, sizeof (obj), 1, &obj);
146 	if (err)
147 		return (err);
148 
149 	return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
150 }
151 
152 static dsl_pool_t *
153 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
154 {
155 	dsl_pool_t *dp;
156 	blkptr_t *bp = spa_get_rootblkptr(spa);
157 
158 	dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
159 	dp->dp_spa = spa;
160 	dp->dp_meta_rootbp = *bp;
161 	rrw_init(&dp->dp_config_rwlock, B_TRUE);
162 	txg_init(dp, txg);
163 
164 	txg_list_create(&dp->dp_dirty_datasets,
165 	    offsetof(dsl_dataset_t, ds_dirty_link));
166 	txg_list_create(&dp->dp_dirty_zilogs,
167 	    offsetof(zilog_t, zl_dirty_link));
168 	txg_list_create(&dp->dp_dirty_dirs,
169 	    offsetof(dsl_dir_t, dd_dirty_link));
170 	txg_list_create(&dp->dp_sync_tasks,
171 	    offsetof(dsl_sync_task_t, dst_node));
172 
173 	mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
174 	cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
175 
176 	dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
177 	    1, 4, 0);
178 
179 	return (dp);
180 }
181 
182 int
183 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
184 {
185 	int err;
186 	dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
187 
188 	err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
189 	    &dp->dp_meta_objset);
190 	if (err != 0)
191 		dsl_pool_close(dp);
192 	else
193 		*dpp = dp;
194 
195 	return (err);
196 }
197 
198 int
199 dsl_pool_open(dsl_pool_t *dp)
200 {
201 	int err;
202 	dsl_dir_t *dd;
203 	dsl_dataset_t *ds;
204 	uint64_t obj;
205 
206 	rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
207 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
208 	    DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
209 	    &dp->dp_root_dir_obj);
210 	if (err)
211 		goto out;
212 
213 	err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
214 	    NULL, dp, &dp->dp_root_dir);
215 	if (err)
216 		goto out;
217 
218 	err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
219 	if (err)
220 		goto out;
221 
222 	if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
223 		err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
224 		if (err)
225 			goto out;
226 		err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
227 		    FTAG, &ds);
228 		if (err == 0) {
229 			err = dsl_dataset_hold_obj(dp,
230 			    ds->ds_phys->ds_prev_snap_obj, dp,
231 			    &dp->dp_origin_snap);
232 			dsl_dataset_rele(ds, FTAG);
233 		}
234 		dsl_dir_rele(dd, dp);
235 		if (err)
236 			goto out;
237 	}
238 
239 	if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
240 		err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
241 		    &dp->dp_free_dir);
242 		if (err)
243 			goto out;
244 
245 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
246 		    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
247 		if (err)
248 			goto out;
249 		VERIFY0(bpobj_open(&dp->dp_free_bpobj,
250 		    dp->dp_meta_objset, obj));
251 	}
252 
253 	if (spa_feature_is_active(dp->dp_spa,
254 	    &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
255 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
256 		    DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
257 		    &dp->dp_bptree_obj);
258 		if (err != 0)
259 			goto out;
260 	}
261 
262 	if (spa_feature_is_active(dp->dp_spa,
263 	    &spa_feature_table[SPA_FEATURE_EMPTY_BPOBJ])) {
264 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
265 		    DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
266 		    &dp->dp_empty_bpobj);
267 		if (err != 0)
268 			goto out;
269 	}
270 
271 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
272 	    DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
273 	    &dp->dp_tmp_userrefs_obj);
274 	if (err == ENOENT)
275 		err = 0;
276 	if (err)
277 		goto out;
278 
279 	err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
280 
281 out:
282 	rrw_exit(&dp->dp_config_rwlock, FTAG);
283 	return (err);
284 }
285 
286 void
287 dsl_pool_close(dsl_pool_t *dp)
288 {
289 	/*
290 	 * Drop our references from dsl_pool_open().
291 	 *
292 	 * Since we held the origin_snap from "syncing" context (which
293 	 * includes pool-opening context), it actually only got a "ref"
294 	 * and not a hold, so just drop that here.
295 	 */
296 	if (dp->dp_origin_snap)
297 		dsl_dataset_rele(dp->dp_origin_snap, dp);
298 	if (dp->dp_mos_dir)
299 		dsl_dir_rele(dp->dp_mos_dir, dp);
300 	if (dp->dp_free_dir)
301 		dsl_dir_rele(dp->dp_free_dir, dp);
302 	if (dp->dp_root_dir)
303 		dsl_dir_rele(dp->dp_root_dir, dp);
304 
305 	bpobj_close(&dp->dp_free_bpobj);
306 
307 	/* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
308 	if (dp->dp_meta_objset)
309 		dmu_objset_evict(dp->dp_meta_objset);
310 
311 	txg_list_destroy(&dp->dp_dirty_datasets);
312 	txg_list_destroy(&dp->dp_dirty_zilogs);
313 	txg_list_destroy(&dp->dp_sync_tasks);
314 	txg_list_destroy(&dp->dp_dirty_dirs);
315 
316 	arc_flush(dp->dp_spa);
317 	txg_fini(dp);
318 	dsl_scan_fini(dp);
319 	rrw_destroy(&dp->dp_config_rwlock);
320 	mutex_destroy(&dp->dp_lock);
321 	taskq_destroy(dp->dp_vnrele_taskq);
322 	if (dp->dp_blkstats)
323 		kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
324 	kmem_free(dp, sizeof (dsl_pool_t));
325 }
326 
327 dsl_pool_t *
328 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
329 {
330 	int err;
331 	dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
332 	dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
333 	objset_t *os;
334 	dsl_dataset_t *ds;
335 	uint64_t obj;
336 
337 	rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
338 
339 	/* create and open the MOS (meta-objset) */
340 	dp->dp_meta_objset = dmu_objset_create_impl(spa,
341 	    NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
342 
343 	/* create the pool directory */
344 	err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
345 	    DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
346 	ASSERT0(err);
347 
348 	/* Initialize scan structures */
349 	VERIFY0(dsl_scan_init(dp, txg));
350 
351 	/* create and open the root dir */
352 	dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
353 	VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
354 	    NULL, dp, &dp->dp_root_dir));
355 
356 	/* create and open the meta-objset dir */
357 	(void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
358 	VERIFY0(dsl_pool_open_special_dir(dp,
359 	    MOS_DIR_NAME, &dp->dp_mos_dir));
360 
361 	if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
362 		/* create and open the free dir */
363 		(void) dsl_dir_create_sync(dp, dp->dp_root_dir,
364 		    FREE_DIR_NAME, tx);
365 		VERIFY0(dsl_pool_open_special_dir(dp,
366 		    FREE_DIR_NAME, &dp->dp_free_dir));
367 
368 		/* create and open the free_bplist */
369 		obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
370 		VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
371 		    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
372 		VERIFY0(bpobj_open(&dp->dp_free_bpobj,
373 		    dp->dp_meta_objset, obj));
374 	}
375 
376 	if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
377 		dsl_pool_create_origin(dp, tx);
378 
379 	/* create the root dataset */
380 	obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
381 
382 	/* create the root objset */
383 	VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
384 	os = dmu_objset_create_impl(dp->dp_spa, ds,
385 	    dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
386 #ifdef _KERNEL
387 	zfs_create_fs(os, kcred, zplprops, tx);
388 #endif
389 	dsl_dataset_rele(ds, FTAG);
390 
391 	dmu_tx_commit(tx);
392 
393 	rrw_exit(&dp->dp_config_rwlock, FTAG);
394 
395 	return (dp);
396 }
397 
398 /*
399  * Account for the meta-objset space in its placeholder dsl_dir.
400  */
401 void
402 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
403     int64_t used, int64_t comp, int64_t uncomp)
404 {
405 	ASSERT3U(comp, ==, uncomp); /* it's all metadata */
406 	mutex_enter(&dp->dp_lock);
407 	dp->dp_mos_used_delta += used;
408 	dp->dp_mos_compressed_delta += comp;
409 	dp->dp_mos_uncompressed_delta += uncomp;
410 	mutex_exit(&dp->dp_lock);
411 }
412 
413 static int
414 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
415 {
416 	dsl_deadlist_t *dl = arg;
417 	dsl_deadlist_insert(dl, bp, tx);
418 	return (0);
419 }
420 
421 static void
422 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
423 {
424 	zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
425 	dmu_objset_sync(dp->dp_meta_objset, zio, tx);
426 	VERIFY0(zio_wait(zio));
427 	dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
428 	spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
429 }
430 
431 static void
432 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
433 {
434 	ASSERT(MUTEX_HELD(&dp->dp_lock));
435 
436 	if (delta < 0)
437 		ASSERT3U(-delta, <=, dp->dp_dirty_total);
438 
439 	dp->dp_dirty_total += delta;
440 
441 	/*
442 	 * Note: we signal even when increasing dp_dirty_total.
443 	 * This ensures forward progress -- each thread wakes the next waiter.
444 	 */
445 	if (dp->dp_dirty_total <= zfs_dirty_data_max)
446 		cv_signal(&dp->dp_spaceavail_cv);
447 }
448 
449 void
450 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
451 {
452 	zio_t *zio;
453 	dmu_tx_t *tx;
454 	dsl_dir_t *dd;
455 	dsl_dataset_t *ds;
456 	objset_t *mos = dp->dp_meta_objset;
457 	list_t synced_datasets;
458 
459 	list_create(&synced_datasets, sizeof (dsl_dataset_t),
460 	    offsetof(dsl_dataset_t, ds_synced_link));
461 
462 	tx = dmu_tx_create_assigned(dp, txg);
463 
464 	/*
465 	 * Write out all dirty blocks of dirty datasets.
466 	 */
467 	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
468 	while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
469 		/*
470 		 * We must not sync any non-MOS datasets twice, because
471 		 * we may have taken a snapshot of them.  However, we
472 		 * may sync newly-created datasets on pass 2.
473 		 */
474 		ASSERT(!list_link_active(&ds->ds_synced_link));
475 		list_insert_tail(&synced_datasets, ds);
476 		dsl_dataset_sync(ds, zio, tx);
477 	}
478 	VERIFY0(zio_wait(zio));
479 
480 	/*
481 	 * We have written all of the accounted dirty data, so our
482 	 * dp_space_towrite should now be zero.  However, some seldom-used
483 	 * code paths do not adhere to this (e.g. dbuf_undirty(), also
484 	 * rounding error in dbuf_write_physdone).
485 	 * Shore up the accounting of any dirtied space now.
486 	 */
487 	dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
488 
489 	/*
490 	 * After the data blocks have been written (ensured by the zio_wait()
491 	 * above), update the user/group space accounting.
492 	 */
493 	for (ds = list_head(&synced_datasets); ds != NULL;
494 	    ds = list_next(&synced_datasets, ds)) {
495 		dmu_objset_do_userquota_updates(ds->ds_objset, tx);
496 	}
497 
498 	/*
499 	 * Sync the datasets again to push out the changes due to
500 	 * userspace updates.  This must be done before we process the
501 	 * sync tasks, so that any snapshots will have the correct
502 	 * user accounting information (and we won't get confused
503 	 * about which blocks are part of the snapshot).
504 	 */
505 	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
506 	while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
507 		ASSERT(list_link_active(&ds->ds_synced_link));
508 		dmu_buf_rele(ds->ds_dbuf, ds);
509 		dsl_dataset_sync(ds, zio, tx);
510 	}
511 	VERIFY0(zio_wait(zio));
512 
513 	/*
514 	 * Now that the datasets have been completely synced, we can
515 	 * clean up our in-memory structures accumulated while syncing:
516 	 *
517 	 *  - move dead blocks from the pending deadlist to the on-disk deadlist
518 	 *  - release hold from dsl_dataset_dirty()
519 	 */
520 	while ((ds = list_remove_head(&synced_datasets)) != NULL) {
521 		objset_t *os = ds->ds_objset;
522 		bplist_iterate(&ds->ds_pending_deadlist,
523 		    deadlist_enqueue_cb, &ds->ds_deadlist, tx);
524 		ASSERT(!dmu_objset_is_dirty(os, txg));
525 		dmu_buf_rele(ds->ds_dbuf, ds);
526 	}
527 	while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
528 		dsl_dir_sync(dd, tx);
529 	}
530 
531 	/*
532 	 * The MOS's space is accounted for in the pool/$MOS
533 	 * (dp_mos_dir).  We can't modify the mos while we're syncing
534 	 * it, so we remember the deltas and apply them here.
535 	 */
536 	if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
537 	    dp->dp_mos_uncompressed_delta != 0) {
538 		dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
539 		    dp->dp_mos_used_delta,
540 		    dp->dp_mos_compressed_delta,
541 		    dp->dp_mos_uncompressed_delta, tx);
542 		dp->dp_mos_used_delta = 0;
543 		dp->dp_mos_compressed_delta = 0;
544 		dp->dp_mos_uncompressed_delta = 0;
545 	}
546 
547 	if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
548 	    list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
549 		dsl_pool_sync_mos(dp, tx);
550 	}
551 
552 	/*
553 	 * If we modify a dataset in the same txg that we want to destroy it,
554 	 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
555 	 * dsl_dir_destroy_check() will fail if there are unexpected holds.
556 	 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
557 	 * and clearing the hold on it) before we process the sync_tasks.
558 	 * The MOS data dirtied by the sync_tasks will be synced on the next
559 	 * pass.
560 	 */
561 	if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
562 		dsl_sync_task_t *dst;
563 		/*
564 		 * No more sync tasks should have been added while we
565 		 * were syncing.
566 		 */
567 		ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
568 		while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
569 			dsl_sync_task_sync(dst, tx);
570 	}
571 
572 	dmu_tx_commit(tx);
573 
574 	DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
575 }
576 
577 void
578 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
579 {
580 	zilog_t *zilog;
581 
582 	while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
583 		dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
584 		zil_clean(zilog, txg);
585 		ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
586 		dmu_buf_rele(ds->ds_dbuf, zilog);
587 	}
588 	ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
589 }
590 
591 /*
592  * TRUE if the current thread is the tx_sync_thread or if we
593  * are being called from SPA context during pool initialization.
594  */
595 int
596 dsl_pool_sync_context(dsl_pool_t *dp)
597 {
598 	return (curthread == dp->dp_tx.tx_sync_thread ||
599 	    spa_is_initializing(dp->dp_spa));
600 }
601 
602 uint64_t
603 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
604 {
605 	uint64_t space, resv;
606 
607 	/*
608 	 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
609 	 * efficiency.
610 	 * XXX The intent log is not accounted for, so it must fit
611 	 * within this slop.
612 	 *
613 	 * If we're trying to assess whether it's OK to do a free,
614 	 * cut the reservation in half to allow forward progress
615 	 * (e.g. make it possible to rm(1) files from a full pool).
616 	 */
617 	space = spa_get_dspace(dp->dp_spa);
618 	resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
619 	if (netfree)
620 		resv >>= 1;
621 
622 	return (space - resv);
623 }
624 
625 boolean_t
626 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
627 {
628 	uint64_t delay_min_bytes =
629 	    zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
630 	boolean_t rv;
631 
632 	mutex_enter(&dp->dp_lock);
633 	if (dp->dp_dirty_total > zfs_dirty_data_sync)
634 		txg_kick(dp);
635 	rv = (dp->dp_dirty_total > delay_min_bytes);
636 	mutex_exit(&dp->dp_lock);
637 	return (rv);
638 }
639 
640 void
641 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
642 {
643 	if (space > 0) {
644 		mutex_enter(&dp->dp_lock);
645 		dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
646 		dsl_pool_dirty_delta(dp, space);
647 		mutex_exit(&dp->dp_lock);
648 	}
649 }
650 
651 void
652 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
653 {
654 	ASSERT3S(space, >=, 0);
655 	if (space == 0)
656 		return;
657 	mutex_enter(&dp->dp_lock);
658 	if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
659 		/* XXX writing something we didn't dirty? */
660 		space = dp->dp_dirty_pertxg[txg & TXG_MASK];
661 	}
662 	ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
663 	dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
664 	ASSERT3U(dp->dp_dirty_total, >=, space);
665 	dsl_pool_dirty_delta(dp, -space);
666 	mutex_exit(&dp->dp_lock);
667 }
668 
669 /* ARGSUSED */
670 static int
671 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
672 {
673 	dmu_tx_t *tx = arg;
674 	dsl_dataset_t *ds, *prev = NULL;
675 	int err;
676 
677 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
678 	if (err)
679 		return (err);
680 
681 	while (ds->ds_phys->ds_prev_snap_obj != 0) {
682 		err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
683 		    FTAG, &prev);
684 		if (err) {
685 			dsl_dataset_rele(ds, FTAG);
686 			return (err);
687 		}
688 
689 		if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
690 			break;
691 		dsl_dataset_rele(ds, FTAG);
692 		ds = prev;
693 		prev = NULL;
694 	}
695 
696 	if (prev == NULL) {
697 		prev = dp->dp_origin_snap;
698 
699 		/*
700 		 * The $ORIGIN can't have any data, or the accounting
701 		 * will be wrong.
702 		 */
703 		ASSERT0(prev->ds_phys->ds_bp.blk_birth);
704 
705 		/* The origin doesn't get attached to itself */
706 		if (ds->ds_object == prev->ds_object) {
707 			dsl_dataset_rele(ds, FTAG);
708 			return (0);
709 		}
710 
711 		dmu_buf_will_dirty(ds->ds_dbuf, tx);
712 		ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
713 		ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
714 
715 		dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
716 		ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
717 
718 		dmu_buf_will_dirty(prev->ds_dbuf, tx);
719 		prev->ds_phys->ds_num_children++;
720 
721 		if (ds->ds_phys->ds_next_snap_obj == 0) {
722 			ASSERT(ds->ds_prev == NULL);
723 			VERIFY0(dsl_dataset_hold_obj(dp,
724 			    ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
725 		}
726 	}
727 
728 	ASSERT3U(ds->ds_dir->dd_phys->dd_origin_obj, ==, prev->ds_object);
729 	ASSERT3U(ds->ds_phys->ds_prev_snap_obj, ==, prev->ds_object);
730 
731 	if (prev->ds_phys->ds_next_clones_obj == 0) {
732 		dmu_buf_will_dirty(prev->ds_dbuf, tx);
733 		prev->ds_phys->ds_next_clones_obj =
734 		    zap_create(dp->dp_meta_objset,
735 		    DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
736 	}
737 	VERIFY0(zap_add_int(dp->dp_meta_objset,
738 	    prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
739 
740 	dsl_dataset_rele(ds, FTAG);
741 	if (prev != dp->dp_origin_snap)
742 		dsl_dataset_rele(prev, FTAG);
743 	return (0);
744 }
745 
746 void
747 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
748 {
749 	ASSERT(dmu_tx_is_syncing(tx));
750 	ASSERT(dp->dp_origin_snap != NULL);
751 
752 	VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
753 	    tx, DS_FIND_CHILDREN));
754 }
755 
756 /* ARGSUSED */
757 static int
758 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
759 {
760 	dmu_tx_t *tx = arg;
761 	objset_t *mos = dp->dp_meta_objset;
762 
763 	if (ds->ds_dir->dd_phys->dd_origin_obj != 0) {
764 		dsl_dataset_t *origin;
765 
766 		VERIFY0(dsl_dataset_hold_obj(dp,
767 		    ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
768 
769 		if (origin->ds_dir->dd_phys->dd_clones == 0) {
770 			dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
771 			origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
772 			    DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
773 		}
774 
775 		VERIFY0(zap_add_int(dp->dp_meta_objset,
776 		    origin->ds_dir->dd_phys->dd_clones, ds->ds_object, tx));
777 
778 		dsl_dataset_rele(origin, FTAG);
779 	}
780 	return (0);
781 }
782 
783 void
784 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
785 {
786 	ASSERT(dmu_tx_is_syncing(tx));
787 	uint64_t obj;
788 
789 	(void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
790 	VERIFY0(dsl_pool_open_special_dir(dp,
791 	    FREE_DIR_NAME, &dp->dp_free_dir));
792 
793 	/*
794 	 * We can't use bpobj_alloc(), because spa_version() still
795 	 * returns the old version, and we need a new-version bpobj with
796 	 * subobj support.  So call dmu_object_alloc() directly.
797 	 */
798 	obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
799 	    SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
800 	VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
801 	    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
802 	VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
803 
804 	VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
805 	    upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
806 }
807 
808 void
809 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
810 {
811 	uint64_t dsobj;
812 	dsl_dataset_t *ds;
813 
814 	ASSERT(dmu_tx_is_syncing(tx));
815 	ASSERT(dp->dp_origin_snap == NULL);
816 	ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
817 
818 	/* create the origin dir, ds, & snap-ds */
819 	dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
820 	    NULL, 0, kcred, tx);
821 	VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
822 	dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
823 	VERIFY0(dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
824 	    dp, &dp->dp_origin_snap));
825 	dsl_dataset_rele(ds, FTAG);
826 }
827 
828 taskq_t *
829 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
830 {
831 	return (dp->dp_vnrele_taskq);
832 }
833 
834 /*
835  * Walk through the pool-wide zap object of temporary snapshot user holds
836  * and release them.
837  */
838 void
839 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
840 {
841 	zap_attribute_t za;
842 	zap_cursor_t zc;
843 	objset_t *mos = dp->dp_meta_objset;
844 	uint64_t zapobj = dp->dp_tmp_userrefs_obj;
845 	nvlist_t *holds;
846 
847 	if (zapobj == 0)
848 		return;
849 	ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
850 
851 	holds = fnvlist_alloc();
852 
853 	for (zap_cursor_init(&zc, mos, zapobj);
854 	    zap_cursor_retrieve(&zc, &za) == 0;
855 	    zap_cursor_advance(&zc)) {
856 		char *htag;
857 		nvlist_t *tags;
858 
859 		htag = strchr(za.za_name, '-');
860 		*htag = '\0';
861 		++htag;
862 		if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
863 			tags = fnvlist_alloc();
864 			fnvlist_add_boolean(tags, htag);
865 			fnvlist_add_nvlist(holds, za.za_name, tags);
866 			fnvlist_free(tags);
867 		} else {
868 			fnvlist_add_boolean(tags, htag);
869 		}
870 	}
871 	dsl_dataset_user_release_tmp(dp, holds);
872 	fnvlist_free(holds);
873 	zap_cursor_fini(&zc);
874 }
875 
876 /*
877  * Create the pool-wide zap object for storing temporary snapshot holds.
878  */
879 void
880 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
881 {
882 	objset_t *mos = dp->dp_meta_objset;
883 
884 	ASSERT(dp->dp_tmp_userrefs_obj == 0);
885 	ASSERT(dmu_tx_is_syncing(tx));
886 
887 	dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
888 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
889 }
890 
891 static int
892 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
893     const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
894 {
895 	objset_t *mos = dp->dp_meta_objset;
896 	uint64_t zapobj = dp->dp_tmp_userrefs_obj;
897 	char *name;
898 	int error;
899 
900 	ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
901 	ASSERT(dmu_tx_is_syncing(tx));
902 
903 	/*
904 	 * If the pool was created prior to SPA_VERSION_USERREFS, the
905 	 * zap object for temporary holds might not exist yet.
906 	 */
907 	if (zapobj == 0) {
908 		if (holding) {
909 			dsl_pool_user_hold_create_obj(dp, tx);
910 			zapobj = dp->dp_tmp_userrefs_obj;
911 		} else {
912 			return (SET_ERROR(ENOENT));
913 		}
914 	}
915 
916 	name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
917 	if (holding)
918 		error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
919 	else
920 		error = zap_remove(mos, zapobj, name, tx);
921 	strfree(name);
922 
923 	return (error);
924 }
925 
926 /*
927  * Add a temporary hold for the given dataset object and tag.
928  */
929 int
930 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
931     uint64_t now, dmu_tx_t *tx)
932 {
933 	return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
934 }
935 
936 /*
937  * Release a temporary hold for the given dataset object and tag.
938  */
939 int
940 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
941     dmu_tx_t *tx)
942 {
943 	return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
944 	    tx, B_FALSE));
945 }
946 
947 /*
948  * DSL Pool Configuration Lock
949  *
950  * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
951  * creation / destruction / rename / property setting).  It must be held for
952  * read to hold a dataset or dsl_dir.  I.e. you must call
953  * dsl_pool_config_enter() or dsl_pool_hold() before calling
954  * dsl_{dataset,dir}_hold{_obj}.  In most circumstances, the dp_config_rwlock
955  * must be held continuously until all datasets and dsl_dirs are released.
956  *
957  * The only exception to this rule is that if a "long hold" is placed on
958  * a dataset, then the dp_config_rwlock may be dropped while the dataset
959  * is still held.  The long hold will prevent the dataset from being
960  * destroyed -- the destroy will fail with EBUSY.  A long hold can be
961  * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
962  * (by calling dsl_{dataset,objset}_{try}own{_obj}).
963  *
964  * Legitimate long-holders (including owners) should be long-running, cancelable
965  * tasks that should cause "zfs destroy" to fail.  This includes DMU
966  * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
967  * "zfs send", and "zfs diff".  There are several other long-holders whose
968  * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
969  *
970  * The usual formula for long-holding would be:
971  * dsl_pool_hold()
972  * dsl_dataset_hold()
973  * ... perform checks ...
974  * dsl_dataset_long_hold()
975  * dsl_pool_rele()
976  * ... perform long-running task ...
977  * dsl_dataset_long_rele()
978  * dsl_dataset_rele()
979  *
980  * Note that when the long hold is released, the dataset is still held but
981  * the pool is not held.  The dataset may change arbitrarily during this time
982  * (e.g. it could be destroyed).  Therefore you shouldn't do anything to the
983  * dataset except release it.
984  *
985  * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
986  * or modifying operations.
987  *
988  * Modifying operations should generally use dsl_sync_task().  The synctask
989  * infrastructure enforces proper locking strategy with respect to the
990  * dp_config_rwlock.  See the comment above dsl_sync_task() for details.
991  *
992  * Read-only operations will manually hold the pool, then the dataset, obtain
993  * information from the dataset, then release the pool and dataset.
994  * dmu_objset_{hold,rele}() are convenience routines that also do the pool
995  * hold/rele.
996  */
997 
998 int
999 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1000 {
1001 	spa_t *spa;
1002 	int error;
1003 
1004 	error = spa_open(name, &spa, tag);
1005 	if (error == 0) {
1006 		*dp = spa_get_dsl(spa);
1007 		dsl_pool_config_enter(*dp, tag);
1008 	}
1009 	return (error);
1010 }
1011 
1012 void
1013 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1014 {
1015 	dsl_pool_config_exit(dp, tag);
1016 	spa_close(dp->dp_spa, tag);
1017 }
1018 
1019 void
1020 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1021 {
1022 	/*
1023 	 * We use a "reentrant" reader-writer lock, but not reentrantly.
1024 	 *
1025 	 * The rrwlock can (with the track_all flag) track all reading threads,
1026 	 * which is very useful for debugging which code path failed to release
1027 	 * the lock, and for verifying that the *current* thread does hold
1028 	 * the lock.
1029 	 *
1030 	 * (Unlike a rwlock, which knows that N threads hold it for
1031 	 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1032 	 * if any thread holds it for read, even if this thread doesn't).
1033 	 */
1034 	ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1035 	rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1036 }
1037 
1038 void
1039 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1040 {
1041 	rrw_exit(&dp->dp_config_rwlock, tag);
1042 }
1043 
1044 boolean_t
1045 dsl_pool_config_held(dsl_pool_t *dp)
1046 {
1047 	return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1048 }
1049