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) 2011, 2019 by Delphix. All rights reserved.
24 * Copyright (c) 2013 Steven Hartland. All rights reserved.
25 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
26 * Copyright (c) 2014 Integros [integros.com]
27 * Copyright 2016 Nexenta Systems, Inc.  All rights reserved.
28 */
29
30#include <sys/dsl_pool.h>
31#include <sys/dsl_dataset.h>
32#include <sys/dsl_prop.h>
33#include <sys/dsl_dir.h>
34#include <sys/dsl_synctask.h>
35#include <sys/dsl_scan.h>
36#include <sys/dnode.h>
37#include <sys/dmu_tx.h>
38#include <sys/dmu_objset.h>
39#include <sys/arc.h>
40#include <sys/zap.h>
41#include <sys/zio.h>
42#include <sys/zfs_context.h>
43#include <sys/fs/zfs.h>
44#include <sys/zfs_znode.h>
45#include <sys/spa_impl.h>
46#include <sys/dsl_deadlist.h>
47#include <sys/vdev_impl.h>
48#include <sys/metaslab_impl.h>
49#include <sys/bptree.h>
50#include <sys/zfeature.h>
51#include <sys/zil_impl.h>
52#include <sys/dsl_userhold.h>
53#include <sys/mmp.h>
54
55/*
56 * ZFS Write Throttle
57 * ------------------
58 *
59 * ZFS must limit the rate of incoming writes to the rate at which it is able
60 * to sync data modifications to the backend storage. Throttling by too much
61 * creates an artificial limit; throttling by too little can only be sustained
62 * for short periods and would lead to highly lumpy performance. On a per-pool
63 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
64 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
65 * of dirty data decreases. When the amount of dirty data exceeds a
66 * predetermined threshold further modifications are blocked until the amount
67 * of dirty data decreases (as data is synced out).
68 *
69 * The limit on dirty data is tunable, and should be adjusted according to
70 * both the IO capacity and available memory of the system. The larger the
71 * window, the more ZFS is able to aggregate and amortize metadata (and data)
72 * changes. However, memory is a limited resource, and allowing for more dirty
73 * data comes at the cost of keeping other useful data in memory (for example
74 * ZFS data cached by the ARC).
75 *
76 * Implementation
77 *
78 * As buffers are modified dsl_pool_willuse_space() increments both the per-
79 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
80 * dirty space used; dsl_pool_dirty_space() decrements those values as data
81 * is synced out from dsl_pool_sync(). While only the poolwide value is
82 * relevant, the per-txg value is useful for debugging. The tunable
83 * zfs_dirty_data_max determines the dirty space limit. Once that value is
84 * exceeded, new writes are halted until space frees up.
85 *
86 * The zfs_dirty_data_sync tunable dictates the threshold at which we
87 * ensure that there is a txg syncing (see the comment in txg.c for a full
88 * description of transaction group stages).
89 *
90 * The IO scheduler uses both the dirty space limit and current amount of
91 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
92 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
93 *
94 * The delay is also calculated based on the amount of dirty data.  See the
95 * comment above dmu_tx_delay() for details.
96 */
97
98/*
99 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
100 * capped at zfs_dirty_data_max_max.  It can also be overridden in /etc/system.
101 */
102uint64_t zfs_dirty_data_max;
103uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024;
104int zfs_dirty_data_max_percent = 10;
105
106/*
107 * If there's at least this much dirty data (as a percentage of
108 * zfs_dirty_data_max), push out a txg.  This should be less than
109 * zfs_vdev_async_write_active_min_dirty_percent.
110 */
111uint64_t zfs_dirty_data_sync_pct = 20;
112
113/*
114 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
115 * and delay each transaction.
116 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
117 */
118int zfs_delay_min_dirty_percent = 60;
119
120/*
121 * This controls how quickly the delay approaches infinity.
122 * Larger values cause it to delay more for a given amount of dirty data.
123 * Therefore larger values will cause there to be less dirty data for a
124 * given throughput.
125 *
126 * For the smoothest delay, this value should be about 1 billion divided
127 * by the maximum number of operations per second.  This will smoothly
128 * handle between 10x and 1/10th this number.
129 *
130 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
131 * multiply in dmu_tx_delay().
132 */
133uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
134
135/*
136 * This determines the number of threads used by the dp_sync_taskq.
137 */
138int zfs_sync_taskq_batch_pct = 75;
139
140/*
141 * These tunables determine the behavior of how zil_itxg_clean() is
142 * called via zil_clean() in the context of spa_sync(). When an itxg
143 * list needs to be cleaned, TQ_NOSLEEP will be used when dispatching.
144 * If the dispatch fails, the call to zil_itxg_clean() will occur
145 * synchronously in the context of spa_sync(), which can negatively
146 * impact the performance of spa_sync() (e.g. in the case of the itxg
147 * list having a large number of itxs that needs to be cleaned).
148 *
149 * Thus, these tunables can be used to manipulate the behavior of the
150 * taskq used by zil_clean(); they determine the number of taskq entries
151 * that are pre-populated when the taskq is first created (via the
152 * "zfs_zil_clean_taskq_minalloc" tunable) and the maximum number of
153 * taskq entries that are cached after an on-demand allocation (via the
154 * "zfs_zil_clean_taskq_maxalloc").
155 *
156 * The idea being, we want to try reasonably hard to ensure there will
157 * already be a taskq entry pre-allocated by the time that it is needed
158 * by zil_clean(). This way, we can avoid the possibility of an
159 * on-demand allocation of a new taskq entry from failing, which would
160 * result in zil_itxg_clean() being called synchronously from zil_clean()
161 * (which can adversely affect performance of spa_sync()).
162 *
163 * Additionally, the number of threads used by the taskq can be
164 * configured via the "zfs_zil_clean_taskq_nthr_pct" tunable.
165 */
166int zfs_zil_clean_taskq_nthr_pct = 100;
167int zfs_zil_clean_taskq_minalloc = 1024;
168int zfs_zil_clean_taskq_maxalloc = 1024 * 1024;
169
170int
171dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
172{
173	uint64_t obj;
174	int err;
175
176	err = zap_lookup(dp->dp_meta_objset,
177	    dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
178	    name, sizeof (obj), 1, &obj);
179	if (err)
180		return (err);
181
182	return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
183}
184
185static dsl_pool_t *
186dsl_pool_open_impl(spa_t *spa, uint64_t txg)
187{
188	dsl_pool_t *dp;
189	blkptr_t *bp = spa_get_rootblkptr(spa);
190
191	dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
192	dp->dp_spa = spa;
193	dp->dp_meta_rootbp = *bp;
194	rrw_init(&dp->dp_config_rwlock, B_TRUE);
195	txg_init(dp, txg);
196	mmp_init(spa);
197
198	txg_list_create(&dp->dp_dirty_datasets, spa,
199	    offsetof(dsl_dataset_t, ds_dirty_link));
200	txg_list_create(&dp->dp_dirty_zilogs, spa,
201	    offsetof(zilog_t, zl_dirty_link));
202	txg_list_create(&dp->dp_dirty_dirs, spa,
203	    offsetof(dsl_dir_t, dd_dirty_link));
204	txg_list_create(&dp->dp_sync_tasks, spa,
205	    offsetof(dsl_sync_task_t, dst_node));
206	txg_list_create(&dp->dp_early_sync_tasks, spa,
207	    offsetof(dsl_sync_task_t, dst_node));
208
209	dp->dp_sync_taskq = taskq_create("dp_sync_taskq",
210	    zfs_sync_taskq_batch_pct, minclsyspri, 1, INT_MAX,
211	    TASKQ_THREADS_CPU_PCT);
212
213	dp->dp_zil_clean_taskq = taskq_create("dp_zil_clean_taskq",
214	    zfs_zil_clean_taskq_nthr_pct, minclsyspri,
215	    zfs_zil_clean_taskq_minalloc,
216	    zfs_zil_clean_taskq_maxalloc,
217	    TASKQ_PREPOPULATE | TASKQ_THREADS_CPU_PCT);
218
219	mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
220	cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
221
222	dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
223	    1, 4, 0);
224	dp->dp_unlinked_drain_taskq = taskq_create("z_unlinked_drain",
225	    max_ncpus, minclsyspri, max_ncpus, INT_MAX,
226	    TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
227
228	return (dp);
229}
230
231int
232dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
233{
234	int err;
235	dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
236
237	err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
238	    &dp->dp_meta_objset);
239	if (err != 0)
240		dsl_pool_close(dp);
241	else
242		*dpp = dp;
243
244	return (err);
245}
246
247int
248dsl_pool_open(dsl_pool_t *dp)
249{
250	int err;
251	dsl_dir_t *dd;
252	dsl_dataset_t *ds;
253	uint64_t obj;
254
255	rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
256	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
257	    DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
258	    &dp->dp_root_dir_obj);
259	if (err)
260		goto out;
261
262	err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
263	    NULL, dp, &dp->dp_root_dir);
264	if (err)
265		goto out;
266
267	err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
268	if (err)
269		goto out;
270
271	if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
272		err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
273		if (err)
274			goto out;
275		err = dsl_dataset_hold_obj(dp,
276		    dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds);
277		if (err == 0) {
278			err = dsl_dataset_hold_obj(dp,
279			    dsl_dataset_phys(ds)->ds_prev_snap_obj, dp,
280			    &dp->dp_origin_snap);
281			dsl_dataset_rele(ds, FTAG);
282		}
283		dsl_dir_rele(dd, dp);
284		if (err)
285			goto out;
286	}
287
288	if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
289		err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
290		    &dp->dp_free_dir);
291		if (err)
292			goto out;
293
294		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
295		    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
296		if (err)
297			goto out;
298		VERIFY0(bpobj_open(&dp->dp_free_bpobj,
299		    dp->dp_meta_objset, obj));
300	}
301
302	if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS)) {
303		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
304		    DMU_POOL_OBSOLETE_BPOBJ, sizeof (uint64_t), 1, &obj);
305		if (err == 0) {
306			VERIFY0(bpobj_open(&dp->dp_obsolete_bpobj,
307			    dp->dp_meta_objset, obj));
308		} else if (err == ENOENT) {
309			/*
310			 * We might not have created the remap bpobj yet.
311			 */
312			err = 0;
313		} else {
314			goto out;
315		}
316	}
317
318	/*
319	 * Note: errors ignored, because the these special dirs, used for
320	 * space accounting, are only created on demand.
321	 */
322	(void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
323	    &dp->dp_leak_dir);
324
325	if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
326		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
327		    DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
328		    &dp->dp_bptree_obj);
329		if (err != 0)
330			goto out;
331	}
332
333	if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
334		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
335		    DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
336		    &dp->dp_empty_bpobj);
337		if (err != 0)
338			goto out;
339	}
340
341	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
342	    DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
343	    &dp->dp_tmp_userrefs_obj);
344	if (err == ENOENT)
345		err = 0;
346	if (err)
347		goto out;
348
349	err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
350
351out:
352	rrw_exit(&dp->dp_config_rwlock, FTAG);
353	return (err);
354}
355
356void
357dsl_pool_close(dsl_pool_t *dp)
358{
359	/*
360	 * Drop our references from dsl_pool_open().
361	 *
362	 * Since we held the origin_snap from "syncing" context (which
363	 * includes pool-opening context), it actually only got a "ref"
364	 * and not a hold, so just drop that here.
365	 */
366	if (dp->dp_origin_snap != NULL)
367		dsl_dataset_rele(dp->dp_origin_snap, dp);
368	if (dp->dp_mos_dir != NULL)
369		dsl_dir_rele(dp->dp_mos_dir, dp);
370	if (dp->dp_free_dir != NULL)
371		dsl_dir_rele(dp->dp_free_dir, dp);
372	if (dp->dp_leak_dir != NULL)
373		dsl_dir_rele(dp->dp_leak_dir, dp);
374	if (dp->dp_root_dir != NULL)
375		dsl_dir_rele(dp->dp_root_dir, dp);
376
377	bpobj_close(&dp->dp_free_bpobj);
378	bpobj_close(&dp->dp_obsolete_bpobj);
379
380	/* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
381	if (dp->dp_meta_objset != NULL)
382		dmu_objset_evict(dp->dp_meta_objset);
383
384	txg_list_destroy(&dp->dp_dirty_datasets);
385	txg_list_destroy(&dp->dp_dirty_zilogs);
386	txg_list_destroy(&dp->dp_sync_tasks);
387	txg_list_destroy(&dp->dp_early_sync_tasks);
388	txg_list_destroy(&dp->dp_dirty_dirs);
389
390	taskq_destroy(dp->dp_zil_clean_taskq);
391	taskq_destroy(dp->dp_sync_taskq);
392
393	/*
394	 * We can't set retry to TRUE since we're explicitly specifying
395	 * a spa to flush. This is good enough; any missed buffers for
396	 * this spa won't cause trouble, and they'll eventually fall
397	 * out of the ARC just like any other unused buffer.
398	 */
399	arc_flush(dp->dp_spa, FALSE);
400
401	mmp_fini(dp->dp_spa);
402	txg_fini(dp);
403	dsl_scan_fini(dp);
404	dmu_buf_user_evict_wait();
405
406	rrw_destroy(&dp->dp_config_rwlock);
407	mutex_destroy(&dp->dp_lock);
408	taskq_destroy(dp->dp_unlinked_drain_taskq);
409	taskq_destroy(dp->dp_vnrele_taskq);
410	if (dp->dp_blkstats != NULL)
411		kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
412	kmem_free(dp, sizeof (dsl_pool_t));
413}
414
415void
416dsl_pool_create_obsolete_bpobj(dsl_pool_t *dp, dmu_tx_t *tx)
417{
418	uint64_t obj;
419	/*
420	 * Currently, we only create the obsolete_bpobj where there are
421	 * indirect vdevs with referenced mappings.
422	 */
423	ASSERT(spa_feature_is_active(dp->dp_spa, SPA_FEATURE_DEVICE_REMOVAL));
424	/* create and open the obsolete_bpobj */
425	obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
426	VERIFY0(bpobj_open(&dp->dp_obsolete_bpobj, dp->dp_meta_objset, obj));
427	VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
428	    DMU_POOL_OBSOLETE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
429	spa_feature_incr(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
430}
431
432void
433dsl_pool_destroy_obsolete_bpobj(dsl_pool_t *dp, dmu_tx_t *tx)
434{
435	spa_feature_decr(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
436	VERIFY0(zap_remove(dp->dp_meta_objset,
437	    DMU_POOL_DIRECTORY_OBJECT,
438	    DMU_POOL_OBSOLETE_BPOBJ, tx));
439	bpobj_free(dp->dp_meta_objset,
440	    dp->dp_obsolete_bpobj.bpo_object, tx);
441	bpobj_close(&dp->dp_obsolete_bpobj);
442}
443
444dsl_pool_t *
445dsl_pool_create(spa_t *spa, nvlist_t *zplprops, dsl_crypto_params_t *dcp,
446    uint64_t txg)
447{
448	int err;
449	dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
450	dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
451	dsl_dataset_t *ds;
452	uint64_t obj;
453
454	rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
455
456	/* create and open the MOS (meta-objset) */
457	dp->dp_meta_objset = dmu_objset_create_impl(spa,
458	    NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
459	spa->spa_meta_objset = dp->dp_meta_objset;
460
461	/* create the pool directory */
462	err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
463	    DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
464	ASSERT0(err);
465
466	/* Initialize scan structures */
467	VERIFY0(dsl_scan_init(dp, txg));
468
469	/* create and open the root dir */
470	dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
471	VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
472	    NULL, dp, &dp->dp_root_dir));
473
474	/* create and open the meta-objset dir */
475	(void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
476	VERIFY0(dsl_pool_open_special_dir(dp,
477	    MOS_DIR_NAME, &dp->dp_mos_dir));
478
479	if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
480		/* create and open the free dir */
481		(void) dsl_dir_create_sync(dp, dp->dp_root_dir,
482		    FREE_DIR_NAME, tx);
483		VERIFY0(dsl_pool_open_special_dir(dp,
484		    FREE_DIR_NAME, &dp->dp_free_dir));
485
486		/* create and open the free_bplist */
487		obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
488		VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
489		    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
490		VERIFY0(bpobj_open(&dp->dp_free_bpobj,
491		    dp->dp_meta_objset, obj));
492	}
493
494	if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
495		dsl_pool_create_origin(dp, tx);
496
497	/*
498	 * Some features may be needed when creating the root dataset, so we
499	 * create the feature objects here.
500	 */
501	if (spa_version(spa) >= SPA_VERSION_FEATURES)
502		spa_feature_create_zap_objects(spa, tx);
503
504	if (dcp != NULL && dcp->cp_crypt != ZIO_CRYPT_OFF &&
505	    dcp->cp_crypt != ZIO_CRYPT_INHERIT)
506		spa_feature_enable(spa, SPA_FEATURE_ENCRYPTION, tx);
507
508	/* create the root dataset */
509	obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, dcp, 0, tx);
510
511	/* create the root objset */
512	VERIFY0(dsl_dataset_hold_obj_flags(dp, obj,
513	    DS_HOLD_FLAG_DECRYPT, FTAG, &ds));
514#ifdef _KERNEL
515	{
516		objset_t *os;
517		rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
518		os = dmu_objset_create_impl(dp->dp_spa, ds,
519		    dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
520		rrw_exit(&ds->ds_bp_rwlock, FTAG);
521		zfs_create_fs(os, kcred, zplprops, tx);
522	}
523#endif
524	dsl_dataset_rele_flags(ds, DS_HOLD_FLAG_DECRYPT, FTAG);
525
526	dmu_tx_commit(tx);
527
528	rrw_exit(&dp->dp_config_rwlock, FTAG);
529
530	return (dp);
531}
532
533/*
534 * Account for the meta-objset space in its placeholder dsl_dir.
535 */
536void
537dsl_pool_mos_diduse_space(dsl_pool_t *dp,
538    int64_t used, int64_t comp, int64_t uncomp)
539{
540	ASSERT3U(comp, ==, uncomp); /* it's all metadata */
541	mutex_enter(&dp->dp_lock);
542	dp->dp_mos_used_delta += used;
543	dp->dp_mos_compressed_delta += comp;
544	dp->dp_mos_uncompressed_delta += uncomp;
545	mutex_exit(&dp->dp_lock);
546}
547
548static void
549dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
550{
551	zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
552	dmu_objset_sync(dp->dp_meta_objset, zio, tx);
553	VERIFY0(zio_wait(zio));
554	dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
555	spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
556}
557
558static void
559dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
560{
561	ASSERT(MUTEX_HELD(&dp->dp_lock));
562
563	if (delta < 0)
564		ASSERT3U(-delta, <=, dp->dp_dirty_total);
565
566	dp->dp_dirty_total += delta;
567
568	/*
569	 * Note: we signal even when increasing dp_dirty_total.
570	 * This ensures forward progress -- each thread wakes the next waiter.
571	 */
572	if (dp->dp_dirty_total < zfs_dirty_data_max)
573		cv_signal(&dp->dp_spaceavail_cv);
574}
575
576static boolean_t
577dsl_early_sync_task_verify(dsl_pool_t *dp, uint64_t txg)
578{
579	spa_t *spa = dp->dp_spa;
580	vdev_t *rvd = spa->spa_root_vdev;
581
582	for (uint64_t c = 0; c < rvd->vdev_children; c++) {
583		vdev_t *vd = rvd->vdev_child[c];
584		txg_list_t *tl = &vd->vdev_ms_list;
585		metaslab_t *ms;
586
587		for (ms = txg_list_head(tl, TXG_CLEAN(txg)); ms;
588		    ms = txg_list_next(tl, ms, TXG_CLEAN(txg))) {
589			VERIFY(range_tree_is_empty(ms->ms_freeing));
590			VERIFY(range_tree_is_empty(ms->ms_checkpointing));
591		}
592	}
593
594	return (B_TRUE);
595}
596
597void
598dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
599{
600	zio_t *zio;
601	dmu_tx_t *tx;
602	dsl_dir_t *dd;
603	dsl_dataset_t *ds;
604	objset_t *mos = dp->dp_meta_objset;
605	list_t synced_datasets;
606
607	list_create(&synced_datasets, sizeof (dsl_dataset_t),
608	    offsetof(dsl_dataset_t, ds_synced_link));
609
610	tx = dmu_tx_create_assigned(dp, txg);
611
612	/*
613	 * Run all early sync tasks before writing out any dirty blocks.
614	 * For more info on early sync tasks see block comment in
615	 * dsl_early_sync_task().
616	 */
617	if (!txg_list_empty(&dp->dp_early_sync_tasks, txg)) {
618		dsl_sync_task_t *dst;
619
620		ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
621		while ((dst =
622		    txg_list_remove(&dp->dp_early_sync_tasks, txg)) != NULL) {
623			ASSERT(dsl_early_sync_task_verify(dp, txg));
624			dsl_sync_task_sync(dst, tx);
625		}
626		ASSERT(dsl_early_sync_task_verify(dp, txg));
627	}
628
629	/*
630	 * Write out all dirty blocks of dirty datasets.
631	 */
632	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
633	while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
634		/*
635		 * We must not sync any non-MOS datasets twice, because
636		 * we may have taken a snapshot of them.  However, we
637		 * may sync newly-created datasets on pass 2.
638		 */
639		ASSERT(!list_link_active(&ds->ds_synced_link));
640		list_insert_tail(&synced_datasets, ds);
641		dsl_dataset_sync(ds, zio, tx);
642	}
643	VERIFY0(zio_wait(zio));
644
645	/*
646	 * We have written all of the accounted dirty data, so our
647	 * dp_space_towrite should now be zero.  However, some seldom-used
648	 * code paths do not adhere to this (e.g. dbuf_undirty(), also
649	 * rounding error in dbuf_write_physdone).
650	 * Shore up the accounting of any dirtied space now.
651	 */
652	dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
653
654	/*
655	 * Update the long range free counter after
656	 * we're done syncing user data
657	 */
658	mutex_enter(&dp->dp_lock);
659	ASSERT(spa_sync_pass(dp->dp_spa) == 1 ||
660	    dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] == 0);
661	dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] = 0;
662	mutex_exit(&dp->dp_lock);
663
664	/*
665	 * After the data blocks have been written (ensured by the zio_wait()
666	 * above), update the user/group/project space accounting.  This happens
667	 * in tasks dispatched to dp_sync_taskq, so wait for them before
668	 * continuing.
669	 */
670	for (ds = list_head(&synced_datasets); ds != NULL;
671	    ds = list_next(&synced_datasets, ds)) {
672		dmu_objset_do_userquota_updates(ds->ds_objset, tx);
673	}
674	taskq_wait(dp->dp_sync_taskq);
675
676	/*
677	 * Sync the datasets again to push out the changes due to
678	 * userspace updates.  This must be done before we process the
679	 * sync tasks, so that any snapshots will have the correct
680	 * user accounting information (and we won't get confused
681	 * about which blocks are part of the snapshot).
682	 */
683	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
684	while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
685		objset_t *os = ds->ds_objset;
686
687		ASSERT(list_link_active(&ds->ds_synced_link));
688		dmu_buf_rele(ds->ds_dbuf, ds);
689		dsl_dataset_sync(ds, zio, tx);
690
691		/*
692		 * Release any key mappings created by calls to
693		 * dsl_dataset_dirty() from the userquota accounting
694		 * code paths.
695		 */
696		if (os->os_encrypted && !os->os_raw_receive &&
697		    !os->os_next_write_raw[txg & TXG_MASK]) {
698			ASSERT3P(ds->ds_key_mapping, !=, NULL);
699			key_mapping_rele(dp->dp_spa, ds->ds_key_mapping, ds);
700		}
701	}
702	VERIFY0(zio_wait(zio));
703
704	/*
705	 * Now that the datasets have been completely synced, we can
706	 * clean up our in-memory structures accumulated while syncing:
707	 *
708	 *  - move dead blocks from the pending deadlist to the on-disk deadlist
709	 *  - release hold from dsl_dataset_dirty()
710	 *  - release key mapping hold from dsl_dataset_dirty()
711	 */
712	while ((ds = list_remove_head(&synced_datasets)) != NULL) {
713		objset_t *os = ds->ds_objset;
714
715		if (os->os_encrypted && !os->os_raw_receive &&
716		    !os->os_next_write_raw[txg & TXG_MASK]) {
717			ASSERT3P(ds->ds_key_mapping, !=, NULL);
718			key_mapping_rele(dp->dp_spa, ds->ds_key_mapping, ds);
719		}
720
721		dsl_dataset_sync_done(ds, tx);
722	}
723	while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
724		dsl_dir_sync(dd, tx);
725	}
726
727	/*
728	 * The MOS's space is accounted for in the pool/$MOS
729	 * (dp_mos_dir).  We can't modify the mos while we're syncing
730	 * it, so we remember the deltas and apply them here.
731	 */
732	if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
733	    dp->dp_mos_uncompressed_delta != 0) {
734		dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
735		    dp->dp_mos_used_delta,
736		    dp->dp_mos_compressed_delta,
737		    dp->dp_mos_uncompressed_delta, tx);
738		dp->dp_mos_used_delta = 0;
739		dp->dp_mos_compressed_delta = 0;
740		dp->dp_mos_uncompressed_delta = 0;
741	}
742
743	if (dmu_objset_is_dirty(mos, txg)) {
744		dsl_pool_sync_mos(dp, tx);
745	}
746
747	/*
748	 * If we modify a dataset in the same txg that we want to destroy it,
749	 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
750	 * dsl_dir_destroy_check() will fail if there are unexpected holds.
751	 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
752	 * and clearing the hold on it) before we process the sync_tasks.
753	 * The MOS data dirtied by the sync_tasks will be synced on the next
754	 * pass.
755	 */
756	if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
757		dsl_sync_task_t *dst;
758		/*
759		 * No more sync tasks should have been added while we
760		 * were syncing.
761		 */
762		ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
763		while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
764			dsl_sync_task_sync(dst, tx);
765	}
766
767	dmu_tx_commit(tx);
768
769	DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
770}
771
772void
773dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
774{
775	zilog_t *zilog;
776
777	while (zilog = txg_list_head(&dp->dp_dirty_zilogs, txg)) {
778		dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
779		/*
780		 * We don't remove the zilog from the dp_dirty_zilogs
781		 * list until after we've cleaned it. This ensures that
782		 * callers of zilog_is_dirty() receive an accurate
783		 * answer when they are racing with the spa sync thread.
784		 */
785		zil_clean(zilog, txg);
786		(void) txg_list_remove_this(&dp->dp_dirty_zilogs, zilog, txg);
787		ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
788		dmu_buf_rele(ds->ds_dbuf, zilog);
789	}
790	ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
791}
792
793/*
794 * TRUE if the current thread is the tx_sync_thread or if we
795 * are being called from SPA context during pool initialization.
796 */
797int
798dsl_pool_sync_context(dsl_pool_t *dp)
799{
800	return (curthread == dp->dp_tx.tx_sync_thread ||
801	    spa_is_initializing(dp->dp_spa) ||
802	    taskq_member(dp->dp_sync_taskq, curthread));
803}
804
805/*
806 * This function returns the amount of allocatable space in the pool
807 * minus whatever space is currently reserved by ZFS for specific
808 * purposes. Specifically:
809 *
810 * 1] Any reserved SLOP space
811 * 2] Any space used by the checkpoint
812 * 3] Any space used for deferred frees
813 *
814 * The latter 2 are especially important because they are needed to
815 * rectify the SPA's and DMU's different understanding of how much space
816 * is used. Now the DMU is aware of that extra space tracked by the SPA
817 * without having to maintain a separate special dir (e.g similar to
818 * $MOS, $FREEING, and $LEAKED).
819 *
820 * Note: By deferred frees here, we mean the frees that were deferred
821 * in spa_sync() after sync pass 1 (spa_deferred_bpobj), and not the
822 * segments placed in ms_defer trees during metaslab_sync_done().
823 */
824uint64_t
825dsl_pool_adjustedsize(dsl_pool_t *dp, zfs_space_check_t slop_policy)
826{
827	spa_t *spa = dp->dp_spa;
828	uint64_t space, resv, adjustedsize;
829	uint64_t spa_deferred_frees =
830	    spa->spa_deferred_bpobj.bpo_phys->bpo_bytes;
831
832	space = spa_get_dspace(spa)
833	    - spa_get_checkpoint_space(spa) - spa_deferred_frees;
834	resv = spa_get_slop_space(spa);
835
836	switch (slop_policy) {
837	case ZFS_SPACE_CHECK_NORMAL:
838		break;
839	case ZFS_SPACE_CHECK_RESERVED:
840		resv >>= 1;
841		break;
842	case ZFS_SPACE_CHECK_EXTRA_RESERVED:
843		resv >>= 2;
844		break;
845	case ZFS_SPACE_CHECK_NONE:
846		resv = 0;
847		break;
848	default:
849		panic("invalid slop policy value: %d", slop_policy);
850		break;
851	}
852	adjustedsize = (space >= resv) ? (space - resv) : 0;
853
854	return (adjustedsize);
855}
856
857uint64_t
858dsl_pool_unreserved_space(dsl_pool_t *dp, zfs_space_check_t slop_policy)
859{
860	uint64_t poolsize = dsl_pool_adjustedsize(dp, slop_policy);
861	uint64_t deferred =
862	    metaslab_class_get_deferred(spa_normal_class(dp->dp_spa));
863	uint64_t quota = (poolsize >= deferred) ? (poolsize - deferred) : 0;
864	return (quota);
865}
866
867boolean_t
868dsl_pool_need_dirty_delay(dsl_pool_t *dp)
869{
870	uint64_t delay_min_bytes =
871	    zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
872	uint64_t dirty_min_bytes =
873	    zfs_dirty_data_max * zfs_dirty_data_sync_pct / 100;
874	boolean_t rv;
875
876	mutex_enter(&dp->dp_lock);
877	if (dp->dp_dirty_total > dirty_min_bytes)
878		txg_kick(dp);
879	rv = (dp->dp_dirty_total > delay_min_bytes);
880	mutex_exit(&dp->dp_lock);
881	return (rv);
882}
883
884void
885dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
886{
887	if (space > 0) {
888		mutex_enter(&dp->dp_lock);
889		dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
890		dsl_pool_dirty_delta(dp, space);
891		mutex_exit(&dp->dp_lock);
892	}
893}
894
895void
896dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
897{
898	ASSERT3S(space, >=, 0);
899	if (space == 0)
900		return;
901	mutex_enter(&dp->dp_lock);
902	if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
903		/* XXX writing something we didn't dirty? */
904		space = dp->dp_dirty_pertxg[txg & TXG_MASK];
905	}
906	ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
907	dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
908	ASSERT3U(dp->dp_dirty_total, >=, space);
909	dsl_pool_dirty_delta(dp, -space);
910	mutex_exit(&dp->dp_lock);
911}
912
913/* ARGSUSED */
914static int
915upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
916{
917	dmu_tx_t *tx = arg;
918	dsl_dataset_t *ds, *prev = NULL;
919	int err;
920
921	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
922	if (err)
923		return (err);
924
925	while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
926		err = dsl_dataset_hold_obj(dp,
927		    dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
928		if (err) {
929			dsl_dataset_rele(ds, FTAG);
930			return (err);
931		}
932
933		if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
934			break;
935		dsl_dataset_rele(ds, FTAG);
936		ds = prev;
937		prev = NULL;
938	}
939
940	if (prev == NULL) {
941		prev = dp->dp_origin_snap;
942
943		/*
944		 * The $ORIGIN can't have any data, or the accounting
945		 * will be wrong.
946		 */
947		rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
948		ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
949		rrw_exit(&ds->ds_bp_rwlock, FTAG);
950
951		/* The origin doesn't get attached to itself */
952		if (ds->ds_object == prev->ds_object) {
953			dsl_dataset_rele(ds, FTAG);
954			return (0);
955		}
956
957		dmu_buf_will_dirty(ds->ds_dbuf, tx);
958		dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
959		dsl_dataset_phys(ds)->ds_prev_snap_txg =
960		    dsl_dataset_phys(prev)->ds_creation_txg;
961
962		dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
963		dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
964
965		dmu_buf_will_dirty(prev->ds_dbuf, tx);
966		dsl_dataset_phys(prev)->ds_num_children++;
967
968		if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
969			ASSERT(ds->ds_prev == NULL);
970			VERIFY0(dsl_dataset_hold_obj(dp,
971			    dsl_dataset_phys(ds)->ds_prev_snap_obj,
972			    ds, &ds->ds_prev));
973		}
974	}
975
976	ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
977	ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
978
979	if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
980		dmu_buf_will_dirty(prev->ds_dbuf, tx);
981		dsl_dataset_phys(prev)->ds_next_clones_obj =
982		    zap_create(dp->dp_meta_objset,
983		    DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
984	}
985	VERIFY0(zap_add_int(dp->dp_meta_objset,
986	    dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
987
988	dsl_dataset_rele(ds, FTAG);
989	if (prev != dp->dp_origin_snap)
990		dsl_dataset_rele(prev, FTAG);
991	return (0);
992}
993
994void
995dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
996{
997	ASSERT(dmu_tx_is_syncing(tx));
998	ASSERT(dp->dp_origin_snap != NULL);
999
1000	VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
1001	    tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
1002}
1003
1004/* ARGSUSED */
1005static int
1006upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
1007{
1008	dmu_tx_t *tx = arg;
1009	objset_t *mos = dp->dp_meta_objset;
1010
1011	if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
1012		dsl_dataset_t *origin;
1013
1014		VERIFY0(dsl_dataset_hold_obj(dp,
1015		    dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
1016
1017		if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
1018			dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
1019			dsl_dir_phys(origin->ds_dir)->dd_clones =
1020			    zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
1021			    0, tx);
1022		}
1023
1024		VERIFY0(zap_add_int(dp->dp_meta_objset,
1025		    dsl_dir_phys(origin->ds_dir)->dd_clones,
1026		    ds->ds_object, tx));
1027
1028		dsl_dataset_rele(origin, FTAG);
1029	}
1030	return (0);
1031}
1032
1033void
1034dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
1035{
1036	ASSERT(dmu_tx_is_syncing(tx));
1037	uint64_t obj;
1038
1039	(void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
1040	VERIFY0(dsl_pool_open_special_dir(dp,
1041	    FREE_DIR_NAME, &dp->dp_free_dir));
1042
1043	/*
1044	 * We can't use bpobj_alloc(), because spa_version() still
1045	 * returns the old version, and we need a new-version bpobj with
1046	 * subobj support.  So call dmu_object_alloc() directly.
1047	 */
1048	obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
1049	    SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
1050	VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
1051	    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
1052	VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
1053
1054	VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1055	    upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
1056}
1057
1058void
1059dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
1060{
1061	uint64_t dsobj;
1062	dsl_dataset_t *ds;
1063
1064	ASSERT(dmu_tx_is_syncing(tx));
1065	ASSERT(dp->dp_origin_snap == NULL);
1066	ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
1067
1068	/* create the origin dir, ds, & snap-ds */
1069	dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
1070	    NULL, 0, kcred, NULL, tx);
1071	VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1072	dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
1073	VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
1074	    dp, &dp->dp_origin_snap));
1075	dsl_dataset_rele(ds, FTAG);
1076}
1077
1078taskq_t *
1079dsl_pool_vnrele_taskq(dsl_pool_t *dp)
1080{
1081	return (dp->dp_vnrele_taskq);
1082}
1083
1084taskq_t *
1085dsl_pool_unlinked_drain_taskq(dsl_pool_t *dp)
1086{
1087	return (dp->dp_unlinked_drain_taskq);
1088}
1089
1090/*
1091 * Walk through the pool-wide zap object of temporary snapshot user holds
1092 * and release them.
1093 */
1094void
1095dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
1096{
1097	zap_attribute_t za;
1098	zap_cursor_t zc;
1099	objset_t *mos = dp->dp_meta_objset;
1100	uint64_t zapobj = dp->dp_tmp_userrefs_obj;
1101	nvlist_t *holds;
1102
1103	if (zapobj == 0)
1104		return;
1105	ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1106
1107	holds = fnvlist_alloc();
1108
1109	for (zap_cursor_init(&zc, mos, zapobj);
1110	    zap_cursor_retrieve(&zc, &za) == 0;
1111	    zap_cursor_advance(&zc)) {
1112		char *htag;
1113		nvlist_t *tags;
1114
1115		htag = strchr(za.za_name, '-');
1116		*htag = '\0';
1117		++htag;
1118		if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
1119			tags = fnvlist_alloc();
1120			fnvlist_add_boolean(tags, htag);
1121			fnvlist_add_nvlist(holds, za.za_name, tags);
1122			fnvlist_free(tags);
1123		} else {
1124			fnvlist_add_boolean(tags, htag);
1125		}
1126	}
1127	dsl_dataset_user_release_tmp(dp, holds);
1128	fnvlist_free(holds);
1129	zap_cursor_fini(&zc);
1130}
1131
1132/*
1133 * Create the pool-wide zap object for storing temporary snapshot holds.
1134 */
1135void
1136dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
1137{
1138	objset_t *mos = dp->dp_meta_objset;
1139
1140	ASSERT(dp->dp_tmp_userrefs_obj == 0);
1141	ASSERT(dmu_tx_is_syncing(tx));
1142
1143	dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
1144	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
1145}
1146
1147static int
1148dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
1149    const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
1150{
1151	objset_t *mos = dp->dp_meta_objset;
1152	uint64_t zapobj = dp->dp_tmp_userrefs_obj;
1153	char *name;
1154	int error;
1155
1156	ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1157	ASSERT(dmu_tx_is_syncing(tx));
1158
1159	/*
1160	 * If the pool was created prior to SPA_VERSION_USERREFS, the
1161	 * zap object for temporary holds might not exist yet.
1162	 */
1163	if (zapobj == 0) {
1164		if (holding) {
1165			dsl_pool_user_hold_create_obj(dp, tx);
1166			zapobj = dp->dp_tmp_userrefs_obj;
1167		} else {
1168			return (SET_ERROR(ENOENT));
1169		}
1170	}
1171
1172	name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
1173	if (holding)
1174		error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
1175	else
1176		error = zap_remove(mos, zapobj, name, tx);
1177	strfree(name);
1178
1179	return (error);
1180}
1181
1182/*
1183 * Add a temporary hold for the given dataset object and tag.
1184 */
1185int
1186dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1187    uint64_t now, dmu_tx_t *tx)
1188{
1189	return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
1190}
1191
1192/*
1193 * Release a temporary hold for the given dataset object and tag.
1194 */
1195int
1196dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1197    dmu_tx_t *tx)
1198{
1199	return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0, tx, B_FALSE));
1200}
1201
1202/*
1203 * DSL Pool Configuration Lock
1204 *
1205 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
1206 * creation / destruction / rename / property setting).  It must be held for
1207 * read to hold a dataset or dsl_dir.  I.e. you must call
1208 * dsl_pool_config_enter() or dsl_pool_hold() before calling
1209 * dsl_{dataset,dir}_hold{_obj}.  In most circumstances, the dp_config_rwlock
1210 * must be held continuously until all datasets and dsl_dirs are released.
1211 *
1212 * The only exception to this rule is that if a "long hold" is placed on
1213 * a dataset, then the dp_config_rwlock may be dropped while the dataset
1214 * is still held.  The long hold will prevent the dataset from being
1215 * destroyed -- the destroy will fail with EBUSY.  A long hold can be
1216 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
1217 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
1218 *
1219 * Legitimate long-holders (including owners) should be long-running, cancelable
1220 * tasks that should cause "zfs destroy" to fail.  This includes DMU
1221 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
1222 * "zfs send", and "zfs diff".  There are several other long-holders whose
1223 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
1224 *
1225 * The usual formula for long-holding would be:
1226 * dsl_pool_hold()
1227 * dsl_dataset_hold()
1228 * ... perform checks ...
1229 * dsl_dataset_long_hold()
1230 * dsl_pool_rele()
1231 * ... perform long-running task ...
1232 * dsl_dataset_long_rele()
1233 * dsl_dataset_rele()
1234 *
1235 * Note that when the long hold is released, the dataset is still held but
1236 * the pool is not held.  The dataset may change arbitrarily during this time
1237 * (e.g. it could be destroyed).  Therefore you shouldn't do anything to the
1238 * dataset except release it.
1239 *
1240 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1241 * or modifying operations.
1242 *
1243 * Modifying operations should generally use dsl_sync_task().  The synctask
1244 * infrastructure enforces proper locking strategy with respect to the
1245 * dp_config_rwlock.  See the comment above dsl_sync_task() for details.
1246 *
1247 * Read-only operations will manually hold the pool, then the dataset, obtain
1248 * information from the dataset, then release the pool and dataset.
1249 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1250 * hold/rele.
1251 */
1252
1253int
1254dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1255{
1256	spa_t *spa;
1257	int error;
1258
1259	error = spa_open(name, &spa, tag);
1260	if (error == 0) {
1261		*dp = spa_get_dsl(spa);
1262		dsl_pool_config_enter(*dp, tag);
1263	}
1264	return (error);
1265}
1266
1267void
1268dsl_pool_rele(dsl_pool_t *dp, void *tag)
1269{
1270	dsl_pool_config_exit(dp, tag);
1271	spa_close(dp->dp_spa, tag);
1272}
1273
1274void
1275dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1276{
1277	/*
1278	 * We use a "reentrant" reader-writer lock, but not reentrantly.
1279	 *
1280	 * The rrwlock can (with the track_all flag) track all reading threads,
1281	 * which is very useful for debugging which code path failed to release
1282	 * the lock, and for verifying that the *current* thread does hold
1283	 * the lock.
1284	 *
1285	 * (Unlike a rwlock, which knows that N threads hold it for
1286	 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1287	 * if any thread holds it for read, even if this thread doesn't).
1288	 */
1289	ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1290	rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1291}
1292
1293void
1294dsl_pool_config_enter_prio(dsl_pool_t *dp, void *tag)
1295{
1296	ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1297	rrw_enter_read_prio(&dp->dp_config_rwlock, tag);
1298}
1299
1300void
1301dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1302{
1303	rrw_exit(&dp->dp_config_rwlock, tag);
1304}
1305
1306boolean_t
1307dsl_pool_config_held(dsl_pool_t *dp)
1308{
1309	return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1310}
1311
1312boolean_t
1313dsl_pool_config_held_writer(dsl_pool_t *dp)
1314{
1315	return (RRW_WRITE_HELD(&dp->dp_config_rwlock));
1316}
1317