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.
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.
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]
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 2016 Nexenta Systems, Inc. All rights reserved.
29 #include <sys/dsl_pool.h>
30 #include <sys/dsl_dataset.h>
31 #include <sys/dsl_prop.h>
32 #include <sys/dsl_dir.h>
33 #include <sys/dsl_synctask.h>
34 #include <sys/dsl_scan.h>
35 #include <sys/dnode.h>
36 #include <sys/dmu_tx.h>
37 #include <sys/dmu_objset.h>
41 #include <sys/zfs_context.h>
42 #include <sys/fs/zfs.h>
43 #include <sys/zfs_znode.h>
44 #include <sys/spa_impl.h>
45 #include <sys/vdev_impl.h>
46 #include <sys/metaslab_impl.h>
47 #include <sys/bptree.h>
48 #include <sys/zfeature.h>
49 #include <sys/zil_impl.h>
50 #include <sys/dsl_userhold.h>
51 #include <sys/trace_zfs.h>
58 * ZFS must limit the rate of incoming writes to the rate at which it is able
59 * to sync data modifications to the backend storage. Throttling by too much
60 * creates an artificial limit; throttling by too little can only be sustained
61 * for short periods and would lead to highly lumpy performance. On a per-pool
62 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
63 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
64 * of dirty data decreases. When the amount of dirty data exceeds a
65 * predetermined threshold further modifications are blocked until the amount
66 * of dirty data decreases (as data is synced out).
68 * The limit on dirty data is tunable, and should be adjusted according to
69 * both the IO capacity and available memory of the system. The larger the
70 * window, the more ZFS is able to aggregate and amortize metadata (and data)
71 * changes. However, memory is a limited resource, and allowing for more dirty
72 * data comes at the cost of keeping other useful data in memory (for example
73 * ZFS data cached by the ARC).
77 * As buffers are modified dsl_pool_willuse_space() increments both the per-
78 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
79 * dirty space used; dsl_pool_dirty_space() decrements those values as data
80 * is synced out from dsl_pool_sync(). While only the poolwide value is
81 * relevant, the per-txg value is useful for debugging. The tunable
82 * zfs_dirty_data_max determines the dirty space limit. Once that value is
83 * exceeded, new writes are halted until space frees up.
85 * The zfs_dirty_data_sync_percent tunable dictates the threshold at which we
86 * ensure that there is a txg syncing (see the comment in txg.c for a full
87 * description of transaction group stages).
89 * The IO scheduler uses both the dirty space limit and current amount of
90 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
91 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
93 * The delay is also calculated based on the amount of dirty data. See the
94 * comment above dmu_tx_delay() for details.
98 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
99 * capped at zfs_dirty_data_max_max. It can also be overridden with a module
102 unsigned long zfs_dirty_data_max
= 0;
103 unsigned long zfs_dirty_data_max_max
= 0;
104 int zfs_dirty_data_max_percent
= 10;
105 int zfs_dirty_data_max_max_percent
= 25;
108 * If there's at least this much dirty data (as a percentage of
109 * zfs_dirty_data_max), push out a txg. This should be less than
110 * zfs_vdev_async_write_active_min_dirty_percent.
112 int zfs_dirty_data_sync_percent
= 20;
115 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
116 * and delay each transaction.
117 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
119 int zfs_delay_min_dirty_percent
= 60;
122 * This controls how quickly the delay approaches infinity.
123 * Larger values cause it to delay more for a given amount of dirty data.
124 * Therefore larger values will cause there to be less dirty data for a
127 * For the smoothest delay, this value should be about 1 billion divided
128 * by the maximum number of operations per second. This will smoothly
129 * handle between 10x and 1/10th this number.
131 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
132 * multiply in dmu_tx_delay().
134 unsigned long zfs_delay_scale
= 1000 * 1000 * 1000 / 2000;
137 * This determines the number of threads used by the dp_sync_taskq.
139 int zfs_sync_taskq_batch_pct
= 75;
142 * These tunables determine the behavior of how zil_itxg_clean() is
143 * called via zil_clean() in the context of spa_sync(). When an itxg
144 * list needs to be cleaned, TQ_NOSLEEP will be used when dispatching.
145 * If the dispatch fails, the call to zil_itxg_clean() will occur
146 * synchronously in the context of spa_sync(), which can negatively
147 * impact the performance of spa_sync() (e.g. in the case of the itxg
148 * list having a large number of itxs that needs to be cleaned).
150 * Thus, these tunables can be used to manipulate the behavior of the
151 * taskq used by zil_clean(); they determine the number of taskq entries
152 * that are pre-populated when the taskq is first created (via the
153 * "zfs_zil_clean_taskq_minalloc" tunable) and the maximum number of
154 * taskq entries that are cached after an on-demand allocation (via the
155 * "zfs_zil_clean_taskq_maxalloc").
157 * The idea being, we want to try reasonably hard to ensure there will
158 * already be a taskq entry pre-allocated by the time that it is needed
159 * by zil_clean(). This way, we can avoid the possibility of an
160 * on-demand allocation of a new taskq entry from failing, which would
161 * result in zil_itxg_clean() being called synchronously from zil_clean()
162 * (which can adversely affect performance of spa_sync()).
164 * Additionally, the number of threads used by the taskq can be
165 * configured via the "zfs_zil_clean_taskq_nthr_pct" tunable.
167 int zfs_zil_clean_taskq_nthr_pct
= 100;
168 int zfs_zil_clean_taskq_minalloc
= 1024;
169 int zfs_zil_clean_taskq_maxalloc
= 1024 * 1024;
172 dsl_pool_open_special_dir(dsl_pool_t
*dp
, const char *name
, dsl_dir_t
**ddp
)
177 err
= zap_lookup(dp
->dp_meta_objset
,
178 dsl_dir_phys(dp
->dp_root_dir
)->dd_child_dir_zapobj
,
179 name
, sizeof (obj
), 1, &obj
);
183 return (dsl_dir_hold_obj(dp
, obj
, name
, dp
, ddp
));
187 dsl_pool_open_impl(spa_t
*spa
, uint64_t txg
)
190 blkptr_t
*bp
= spa_get_rootblkptr(spa
);
192 dp
= kmem_zalloc(sizeof (dsl_pool_t
), KM_SLEEP
);
194 dp
->dp_meta_rootbp
= *bp
;
195 rrw_init(&dp
->dp_config_rwlock
, B_TRUE
);
199 txg_list_create(&dp
->dp_dirty_datasets
, spa
,
200 offsetof(dsl_dataset_t
, ds_dirty_link
));
201 txg_list_create(&dp
->dp_dirty_zilogs
, spa
,
202 offsetof(zilog_t
, zl_dirty_link
));
203 txg_list_create(&dp
->dp_dirty_dirs
, spa
,
204 offsetof(dsl_dir_t
, dd_dirty_link
));
205 txg_list_create(&dp
->dp_sync_tasks
, spa
,
206 offsetof(dsl_sync_task_t
, dst_node
));
207 txg_list_create(&dp
->dp_early_sync_tasks
, spa
,
208 offsetof(dsl_sync_task_t
, dst_node
));
210 dp
->dp_sync_taskq
= taskq_create("dp_sync_taskq",
211 zfs_sync_taskq_batch_pct
, minclsyspri
, 1, INT_MAX
,
212 TASKQ_THREADS_CPU_PCT
);
214 dp
->dp_zil_clean_taskq
= taskq_create("dp_zil_clean_taskq",
215 zfs_zil_clean_taskq_nthr_pct
, minclsyspri
,
216 zfs_zil_clean_taskq_minalloc
,
217 zfs_zil_clean_taskq_maxalloc
,
218 TASKQ_PREPOPULATE
| TASKQ_THREADS_CPU_PCT
);
220 mutex_init(&dp
->dp_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
221 cv_init(&dp
->dp_spaceavail_cv
, NULL
, CV_DEFAULT
, NULL
);
223 dp
->dp_zrele_taskq
= taskq_create("z_zrele", 100, defclsyspri
,
224 boot_ncpus
* 8, INT_MAX
, TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
|
225 TASKQ_THREADS_CPU_PCT
);
226 dp
->dp_unlinked_drain_taskq
= taskq_create("z_unlinked_drain",
227 100, defclsyspri
, boot_ncpus
, INT_MAX
,
228 TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
| TASKQ_THREADS_CPU_PCT
);
234 dsl_pool_init(spa_t
*spa
, uint64_t txg
, dsl_pool_t
**dpp
)
237 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
240 * Initialize the caller's dsl_pool_t structure before we actually open
241 * the meta objset. This is done because a self-healing write zio may
242 * be issued as part of dmu_objset_open_impl() and the spa needs its
243 * dsl_pool_t initialized in order to handle the write.
247 err
= dmu_objset_open_impl(spa
, NULL
, &dp
->dp_meta_rootbp
,
248 &dp
->dp_meta_objset
);
258 dsl_pool_open(dsl_pool_t
*dp
)
265 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
266 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
267 DMU_POOL_ROOT_DATASET
, sizeof (uint64_t), 1,
268 &dp
->dp_root_dir_obj
);
272 err
= dsl_dir_hold_obj(dp
, dp
->dp_root_dir_obj
,
273 NULL
, dp
, &dp
->dp_root_dir
);
277 err
= dsl_pool_open_special_dir(dp
, MOS_DIR_NAME
, &dp
->dp_mos_dir
);
281 if (spa_version(dp
->dp_spa
) >= SPA_VERSION_ORIGIN
) {
282 err
= dsl_pool_open_special_dir(dp
, ORIGIN_DIR_NAME
, &dd
);
285 err
= dsl_dataset_hold_obj(dp
,
286 dsl_dir_phys(dd
)->dd_head_dataset_obj
, FTAG
, &ds
);
288 err
= dsl_dataset_hold_obj(dp
,
289 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, dp
,
290 &dp
->dp_origin_snap
);
291 dsl_dataset_rele(ds
, FTAG
);
293 dsl_dir_rele(dd
, dp
);
298 if (spa_version(dp
->dp_spa
) >= SPA_VERSION_DEADLISTS
) {
299 err
= dsl_pool_open_special_dir(dp
, FREE_DIR_NAME
,
304 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
305 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
);
308 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
,
309 dp
->dp_meta_objset
, obj
));
312 if (spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_OBSOLETE_COUNTS
)) {
313 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
314 DMU_POOL_OBSOLETE_BPOBJ
, sizeof (uint64_t), 1, &obj
);
316 VERIFY0(bpobj_open(&dp
->dp_obsolete_bpobj
,
317 dp
->dp_meta_objset
, obj
));
318 } else if (err
== ENOENT
) {
320 * We might not have created the remap bpobj yet.
329 * Note: errors ignored, because the these special dirs, used for
330 * space accounting, are only created on demand.
332 (void) dsl_pool_open_special_dir(dp
, LEAK_DIR_NAME
,
335 if (spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_ASYNC_DESTROY
)) {
336 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
337 DMU_POOL_BPTREE_OBJ
, sizeof (uint64_t), 1,
343 if (spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_EMPTY_BPOBJ
)) {
344 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
345 DMU_POOL_EMPTY_BPOBJ
, sizeof (uint64_t), 1,
346 &dp
->dp_empty_bpobj
);
351 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
352 DMU_POOL_TMP_USERREFS
, sizeof (uint64_t), 1,
353 &dp
->dp_tmp_userrefs_obj
);
359 err
= dsl_scan_init(dp
, dp
->dp_tx
.tx_open_txg
);
362 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
367 dsl_pool_close(dsl_pool_t
*dp
)
370 * Drop our references from dsl_pool_open().
372 * Since we held the origin_snap from "syncing" context (which
373 * includes pool-opening context), it actually only got a "ref"
374 * and not a hold, so just drop that here.
376 if (dp
->dp_origin_snap
!= NULL
)
377 dsl_dataset_rele(dp
->dp_origin_snap
, dp
);
378 if (dp
->dp_mos_dir
!= NULL
)
379 dsl_dir_rele(dp
->dp_mos_dir
, dp
);
380 if (dp
->dp_free_dir
!= NULL
)
381 dsl_dir_rele(dp
->dp_free_dir
, dp
);
382 if (dp
->dp_leak_dir
!= NULL
)
383 dsl_dir_rele(dp
->dp_leak_dir
, dp
);
384 if (dp
->dp_root_dir
!= NULL
)
385 dsl_dir_rele(dp
->dp_root_dir
, dp
);
387 bpobj_close(&dp
->dp_free_bpobj
);
388 bpobj_close(&dp
->dp_obsolete_bpobj
);
390 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
391 if (dp
->dp_meta_objset
!= NULL
)
392 dmu_objset_evict(dp
->dp_meta_objset
);
394 txg_list_destroy(&dp
->dp_dirty_datasets
);
395 txg_list_destroy(&dp
->dp_dirty_zilogs
);
396 txg_list_destroy(&dp
->dp_sync_tasks
);
397 txg_list_destroy(&dp
->dp_early_sync_tasks
);
398 txg_list_destroy(&dp
->dp_dirty_dirs
);
400 taskq_destroy(dp
->dp_zil_clean_taskq
);
401 taskq_destroy(dp
->dp_sync_taskq
);
404 * We can't set retry to TRUE since we're explicitly specifying
405 * a spa to flush. This is good enough; any missed buffers for
406 * this spa won't cause trouble, and they'll eventually fall
407 * out of the ARC just like any other unused buffer.
409 arc_flush(dp
->dp_spa
, FALSE
);
411 mmp_fini(dp
->dp_spa
);
414 dmu_buf_user_evict_wait();
416 rrw_destroy(&dp
->dp_config_rwlock
);
417 mutex_destroy(&dp
->dp_lock
);
418 cv_destroy(&dp
->dp_spaceavail_cv
);
419 taskq_destroy(dp
->dp_unlinked_drain_taskq
);
420 taskq_destroy(dp
->dp_zrele_taskq
);
421 if (dp
->dp_blkstats
!= NULL
) {
422 mutex_destroy(&dp
->dp_blkstats
->zab_lock
);
423 vmem_free(dp
->dp_blkstats
, sizeof (zfs_all_blkstats_t
));
425 kmem_free(dp
, sizeof (dsl_pool_t
));
429 dsl_pool_create_obsolete_bpobj(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
433 * Currently, we only create the obsolete_bpobj where there are
434 * indirect vdevs with referenced mappings.
436 ASSERT(spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_DEVICE_REMOVAL
));
437 /* create and open the obsolete_bpobj */
438 obj
= bpobj_alloc(dp
->dp_meta_objset
, SPA_OLD_MAXBLOCKSIZE
, tx
);
439 VERIFY0(bpobj_open(&dp
->dp_obsolete_bpobj
, dp
->dp_meta_objset
, obj
));
440 VERIFY0(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
441 DMU_POOL_OBSOLETE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
));
442 spa_feature_incr(dp
->dp_spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
446 dsl_pool_destroy_obsolete_bpobj(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
448 spa_feature_decr(dp
->dp_spa
, SPA_FEATURE_OBSOLETE_COUNTS
, tx
);
449 VERIFY0(zap_remove(dp
->dp_meta_objset
,
450 DMU_POOL_DIRECTORY_OBJECT
,
451 DMU_POOL_OBSOLETE_BPOBJ
, tx
));
452 bpobj_free(dp
->dp_meta_objset
,
453 dp
->dp_obsolete_bpobj
.bpo_object
, tx
);
454 bpobj_close(&dp
->dp_obsolete_bpobj
);
458 dsl_pool_create(spa_t
*spa
, nvlist_t
*zplprops
, dsl_crypto_params_t
*dcp
,
462 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
463 dmu_tx_t
*tx
= dmu_tx_create_assigned(dp
, txg
);
467 objset_t
*os
__attribute__((unused
));
472 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
474 /* create and open the MOS (meta-objset) */
475 dp
->dp_meta_objset
= dmu_objset_create_impl(spa
,
476 NULL
, &dp
->dp_meta_rootbp
, DMU_OST_META
, tx
);
477 spa
->spa_meta_objset
= dp
->dp_meta_objset
;
479 /* create the pool directory */
480 err
= zap_create_claim(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
481 DMU_OT_OBJECT_DIRECTORY
, DMU_OT_NONE
, 0, tx
);
484 /* Initialize scan structures */
485 VERIFY0(dsl_scan_init(dp
, txg
));
487 /* create and open the root dir */
488 dp
->dp_root_dir_obj
= dsl_dir_create_sync(dp
, NULL
, NULL
, tx
);
489 VERIFY0(dsl_dir_hold_obj(dp
, dp
->dp_root_dir_obj
,
490 NULL
, dp
, &dp
->dp_root_dir
));
492 /* create and open the meta-objset dir */
493 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, MOS_DIR_NAME
, tx
);
494 VERIFY0(dsl_pool_open_special_dir(dp
,
495 MOS_DIR_NAME
, &dp
->dp_mos_dir
));
497 if (spa_version(spa
) >= SPA_VERSION_DEADLISTS
) {
498 /* create and open the free dir */
499 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
,
501 VERIFY0(dsl_pool_open_special_dir(dp
,
502 FREE_DIR_NAME
, &dp
->dp_free_dir
));
504 /* create and open the free_bplist */
505 obj
= bpobj_alloc(dp
->dp_meta_objset
, SPA_OLD_MAXBLOCKSIZE
, tx
);
506 VERIFY(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
507 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
) == 0);
508 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
,
509 dp
->dp_meta_objset
, obj
));
512 if (spa_version(spa
) >= SPA_VERSION_DSL_SCRUB
)
513 dsl_pool_create_origin(dp
, tx
);
516 * Some features may be needed when creating the root dataset, so we
517 * create the feature objects here.
519 if (spa_version(spa
) >= SPA_VERSION_FEATURES
)
520 spa_feature_create_zap_objects(spa
, tx
);
522 if (dcp
!= NULL
&& dcp
->cp_crypt
!= ZIO_CRYPT_OFF
&&
523 dcp
->cp_crypt
!= ZIO_CRYPT_INHERIT
)
524 spa_feature_enable(spa
, SPA_FEATURE_ENCRYPTION
, tx
);
526 /* create the root dataset */
527 obj
= dsl_dataset_create_sync_dd(dp
->dp_root_dir
, NULL
, dcp
, 0, tx
);
529 /* create the root objset */
530 VERIFY0(dsl_dataset_hold_obj_flags(dp
, obj
,
531 DS_HOLD_FLAG_DECRYPT
, FTAG
, &ds
));
532 rrw_enter(&ds
->ds_bp_rwlock
, RW_READER
, FTAG
);
533 os
= dmu_objset_create_impl(dp
->dp_spa
, ds
,
534 dsl_dataset_get_blkptr(ds
), DMU_OST_ZFS
, tx
);
535 rrw_exit(&ds
->ds_bp_rwlock
, FTAG
);
537 zfs_create_fs(os
, kcred
, zplprops
, tx
);
539 dsl_dataset_rele_flags(ds
, DS_HOLD_FLAG_DECRYPT
, FTAG
);
543 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
549 * Account for the meta-objset space in its placeholder dsl_dir.
552 dsl_pool_mos_diduse_space(dsl_pool_t
*dp
,
553 int64_t used
, int64_t comp
, int64_t uncomp
)
555 ASSERT3U(comp
, ==, uncomp
); /* it's all metadata */
556 mutex_enter(&dp
->dp_lock
);
557 dp
->dp_mos_used_delta
+= used
;
558 dp
->dp_mos_compressed_delta
+= comp
;
559 dp
->dp_mos_uncompressed_delta
+= uncomp
;
560 mutex_exit(&dp
->dp_lock
);
564 dsl_pool_sync_mos(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
566 zio_t
*zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
567 dmu_objset_sync(dp
->dp_meta_objset
, zio
, tx
);
568 VERIFY0(zio_wait(zio
));
569 dprintf_bp(&dp
->dp_meta_rootbp
, "meta objset rootbp is %s", "");
570 spa_set_rootblkptr(dp
->dp_spa
, &dp
->dp_meta_rootbp
);
574 dsl_pool_dirty_delta(dsl_pool_t
*dp
, int64_t delta
)
576 ASSERT(MUTEX_HELD(&dp
->dp_lock
));
579 ASSERT3U(-delta
, <=, dp
->dp_dirty_total
);
581 dp
->dp_dirty_total
+= delta
;
584 * Note: we signal even when increasing dp_dirty_total.
585 * This ensures forward progress -- each thread wakes the next waiter.
587 if (dp
->dp_dirty_total
< zfs_dirty_data_max
)
588 cv_signal(&dp
->dp_spaceavail_cv
);
593 dsl_early_sync_task_verify(dsl_pool_t
*dp
, uint64_t txg
)
595 spa_t
*spa
= dp
->dp_spa
;
596 vdev_t
*rvd
= spa
->spa_root_vdev
;
598 for (uint64_t c
= 0; c
< rvd
->vdev_children
; c
++) {
599 vdev_t
*vd
= rvd
->vdev_child
[c
];
600 txg_list_t
*tl
= &vd
->vdev_ms_list
;
603 for (ms
= txg_list_head(tl
, TXG_CLEAN(txg
)); ms
;
604 ms
= txg_list_next(tl
, ms
, TXG_CLEAN(txg
))) {
605 VERIFY(range_tree_is_empty(ms
->ms_freeing
));
606 VERIFY(range_tree_is_empty(ms
->ms_checkpointing
));
615 dsl_pool_sync(dsl_pool_t
*dp
, uint64_t txg
)
621 objset_t
*mos
= dp
->dp_meta_objset
;
622 list_t synced_datasets
;
624 list_create(&synced_datasets
, sizeof (dsl_dataset_t
),
625 offsetof(dsl_dataset_t
, ds_synced_link
));
627 tx
= dmu_tx_create_assigned(dp
, txg
);
630 * Run all early sync tasks before writing out any dirty blocks.
631 * For more info on early sync tasks see block comment in
632 * dsl_early_sync_task().
634 if (!txg_list_empty(&dp
->dp_early_sync_tasks
, txg
)) {
635 dsl_sync_task_t
*dst
;
637 ASSERT3U(spa_sync_pass(dp
->dp_spa
), ==, 1);
639 txg_list_remove(&dp
->dp_early_sync_tasks
, txg
)) != NULL
) {
640 ASSERT(dsl_early_sync_task_verify(dp
, txg
));
641 dsl_sync_task_sync(dst
, tx
);
643 ASSERT(dsl_early_sync_task_verify(dp
, txg
));
647 * Write out all dirty blocks of dirty datasets.
649 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
650 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
652 * We must not sync any non-MOS datasets twice, because
653 * we may have taken a snapshot of them. However, we
654 * may sync newly-created datasets on pass 2.
656 ASSERT(!list_link_active(&ds
->ds_synced_link
));
657 list_insert_tail(&synced_datasets
, ds
);
658 dsl_dataset_sync(ds
, zio
, tx
);
660 VERIFY0(zio_wait(zio
));
663 * Update the long range free counter after
664 * we're done syncing user data
666 mutex_enter(&dp
->dp_lock
);
667 ASSERT(spa_sync_pass(dp
->dp_spa
) == 1 ||
668 dp
->dp_long_free_dirty_pertxg
[txg
& TXG_MASK
] == 0);
669 dp
->dp_long_free_dirty_pertxg
[txg
& TXG_MASK
] = 0;
670 mutex_exit(&dp
->dp_lock
);
673 * After the data blocks have been written (ensured by the zio_wait()
674 * above), update the user/group/project space accounting. This happens
675 * in tasks dispatched to dp_sync_taskq, so wait for them before
678 for (ds
= list_head(&synced_datasets
); ds
!= NULL
;
679 ds
= list_next(&synced_datasets
, ds
)) {
680 dmu_objset_do_userquota_updates(ds
->ds_objset
, tx
);
682 taskq_wait(dp
->dp_sync_taskq
);
685 * Sync the datasets again to push out the changes due to
686 * userspace updates. This must be done before we process the
687 * sync tasks, so that any snapshots will have the correct
688 * user accounting information (and we won't get confused
689 * about which blocks are part of the snapshot).
691 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
692 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
693 objset_t
*os
= ds
->ds_objset
;
695 ASSERT(list_link_active(&ds
->ds_synced_link
));
696 dmu_buf_rele(ds
->ds_dbuf
, ds
);
697 dsl_dataset_sync(ds
, zio
, tx
);
700 * Release any key mappings created by calls to
701 * dsl_dataset_dirty() from the userquota accounting
704 if (os
->os_encrypted
&& !os
->os_raw_receive
&&
705 !os
->os_next_write_raw
[txg
& TXG_MASK
]) {
706 ASSERT3P(ds
->ds_key_mapping
, !=, NULL
);
707 key_mapping_rele(dp
->dp_spa
, ds
->ds_key_mapping
, ds
);
710 VERIFY0(zio_wait(zio
));
713 * Now that the datasets have been completely synced, we can
714 * clean up our in-memory structures accumulated while syncing:
716 * - move dead blocks from the pending deadlist and livelists
717 * to the on-disk versions
718 * - release hold from dsl_dataset_dirty()
719 * - release key mapping hold from dsl_dataset_dirty()
721 while ((ds
= list_remove_head(&synced_datasets
)) != NULL
) {
722 objset_t
*os
= ds
->ds_objset
;
724 if (os
->os_encrypted
&& !os
->os_raw_receive
&&
725 !os
->os_next_write_raw
[txg
& TXG_MASK
]) {
726 ASSERT3P(ds
->ds_key_mapping
, !=, NULL
);
727 key_mapping_rele(dp
->dp_spa
, ds
->ds_key_mapping
, ds
);
730 dsl_dataset_sync_done(ds
, tx
);
733 while ((dd
= txg_list_remove(&dp
->dp_dirty_dirs
, txg
)) != NULL
) {
734 dsl_dir_sync(dd
, tx
);
738 * The MOS's space is accounted for in the pool/$MOS
739 * (dp_mos_dir). We can't modify the mos while we're syncing
740 * it, so we remember the deltas and apply them here.
742 if (dp
->dp_mos_used_delta
!= 0 || dp
->dp_mos_compressed_delta
!= 0 ||
743 dp
->dp_mos_uncompressed_delta
!= 0) {
744 dsl_dir_diduse_space(dp
->dp_mos_dir
, DD_USED_HEAD
,
745 dp
->dp_mos_used_delta
,
746 dp
->dp_mos_compressed_delta
,
747 dp
->dp_mos_uncompressed_delta
, tx
);
748 dp
->dp_mos_used_delta
= 0;
749 dp
->dp_mos_compressed_delta
= 0;
750 dp
->dp_mos_uncompressed_delta
= 0;
753 if (dmu_objset_is_dirty(mos
, txg
)) {
754 dsl_pool_sync_mos(dp
, tx
);
758 * We have written all of the accounted dirty data, so our
759 * dp_space_towrite should now be zero. However, some seldom-used
760 * code paths do not adhere to this (e.g. dbuf_undirty()). Shore up
761 * the accounting of any dirtied space now.
763 * Note that, besides any dirty data from datasets, the amount of
764 * dirty data in the MOS is also accounted by the pool. Therefore,
765 * we want to do this cleanup after dsl_pool_sync_mos() so we don't
766 * attempt to update the accounting for the same dirty data twice.
767 * (i.e. at this point we only update the accounting for the space
768 * that we know that we "leaked").
770 dsl_pool_undirty_space(dp
, dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], txg
);
773 * If we modify a dataset in the same txg that we want to destroy it,
774 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
775 * dsl_dir_destroy_check() will fail if there are unexpected holds.
776 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
777 * and clearing the hold on it) before we process the sync_tasks.
778 * The MOS data dirtied by the sync_tasks will be synced on the next
781 if (!txg_list_empty(&dp
->dp_sync_tasks
, txg
)) {
782 dsl_sync_task_t
*dst
;
784 * No more sync tasks should have been added while we
787 ASSERT3U(spa_sync_pass(dp
->dp_spa
), ==, 1);
788 while ((dst
= txg_list_remove(&dp
->dp_sync_tasks
, txg
)) != NULL
)
789 dsl_sync_task_sync(dst
, tx
);
794 DTRACE_PROBE2(dsl_pool_sync__done
, dsl_pool_t
*dp
, dp
, uint64_t, txg
);
798 dsl_pool_sync_done(dsl_pool_t
*dp
, uint64_t txg
)
802 while ((zilog
= txg_list_head(&dp
->dp_dirty_zilogs
, txg
))) {
803 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
805 * We don't remove the zilog from the dp_dirty_zilogs
806 * list until after we've cleaned it. This ensures that
807 * callers of zilog_is_dirty() receive an accurate
808 * answer when they are racing with the spa sync thread.
810 zil_clean(zilog
, txg
);
811 (void) txg_list_remove_this(&dp
->dp_dirty_zilogs
, zilog
, txg
);
812 ASSERT(!dmu_objset_is_dirty(zilog
->zl_os
, txg
));
813 dmu_buf_rele(ds
->ds_dbuf
, zilog
);
815 ASSERT(!dmu_objset_is_dirty(dp
->dp_meta_objset
, txg
));
819 * TRUE if the current thread is the tx_sync_thread or if we
820 * are being called from SPA context during pool initialization.
823 dsl_pool_sync_context(dsl_pool_t
*dp
)
825 return (curthread
== dp
->dp_tx
.tx_sync_thread
||
826 spa_is_initializing(dp
->dp_spa
) ||
827 taskq_member(dp
->dp_sync_taskq
, curthread
));
831 * This function returns the amount of allocatable space in the pool
832 * minus whatever space is currently reserved by ZFS for specific
833 * purposes. Specifically:
835 * 1] Any reserved SLOP space
836 * 2] Any space used by the checkpoint
837 * 3] Any space used for deferred frees
839 * The latter 2 are especially important because they are needed to
840 * rectify the SPA's and DMU's different understanding of how much space
841 * is used. Now the DMU is aware of that extra space tracked by the SPA
842 * without having to maintain a separate special dir (e.g similar to
843 * $MOS, $FREEING, and $LEAKED).
845 * Note: By deferred frees here, we mean the frees that were deferred
846 * in spa_sync() after sync pass 1 (spa_deferred_bpobj), and not the
847 * segments placed in ms_defer trees during metaslab_sync_done().
850 dsl_pool_adjustedsize(dsl_pool_t
*dp
, zfs_space_check_t slop_policy
)
852 spa_t
*spa
= dp
->dp_spa
;
853 uint64_t space
, resv
, adjustedsize
;
854 uint64_t spa_deferred_frees
=
855 spa
->spa_deferred_bpobj
.bpo_phys
->bpo_bytes
;
857 space
= spa_get_dspace(spa
)
858 - spa_get_checkpoint_space(spa
) - spa_deferred_frees
;
859 resv
= spa_get_slop_space(spa
);
861 switch (slop_policy
) {
862 case ZFS_SPACE_CHECK_NORMAL
:
864 case ZFS_SPACE_CHECK_RESERVED
:
867 case ZFS_SPACE_CHECK_EXTRA_RESERVED
:
870 case ZFS_SPACE_CHECK_NONE
:
874 panic("invalid slop policy value: %d", slop_policy
);
877 adjustedsize
= (space
>= resv
) ? (space
- resv
) : 0;
879 return (adjustedsize
);
883 dsl_pool_unreserved_space(dsl_pool_t
*dp
, zfs_space_check_t slop_policy
)
885 uint64_t poolsize
= dsl_pool_adjustedsize(dp
, slop_policy
);
887 metaslab_class_get_deferred(spa_normal_class(dp
->dp_spa
));
888 uint64_t quota
= (poolsize
>= deferred
) ? (poolsize
- deferred
) : 0;
893 dsl_pool_need_dirty_delay(dsl_pool_t
*dp
)
895 uint64_t delay_min_bytes
=
896 zfs_dirty_data_max
* zfs_delay_min_dirty_percent
/ 100;
897 uint64_t dirty_min_bytes
=
898 zfs_dirty_data_max
* zfs_dirty_data_sync_percent
/ 100;
901 mutex_enter(&dp
->dp_lock
);
902 dirty
= dp
->dp_dirty_total
;
903 mutex_exit(&dp
->dp_lock
);
904 if (dirty
> dirty_min_bytes
)
906 return (dirty
> delay_min_bytes
);
910 dsl_pool_dirty_space(dsl_pool_t
*dp
, int64_t space
, dmu_tx_t
*tx
)
913 mutex_enter(&dp
->dp_lock
);
914 dp
->dp_dirty_pertxg
[tx
->tx_txg
& TXG_MASK
] += space
;
915 dsl_pool_dirty_delta(dp
, space
);
916 mutex_exit(&dp
->dp_lock
);
921 dsl_pool_undirty_space(dsl_pool_t
*dp
, int64_t space
, uint64_t txg
)
923 ASSERT3S(space
, >=, 0);
927 mutex_enter(&dp
->dp_lock
);
928 if (dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] < space
) {
929 /* XXX writing something we didn't dirty? */
930 space
= dp
->dp_dirty_pertxg
[txg
& TXG_MASK
];
932 ASSERT3U(dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], >=, space
);
933 dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] -= space
;
934 ASSERT3U(dp
->dp_dirty_total
, >=, space
);
935 dsl_pool_dirty_delta(dp
, -space
);
936 mutex_exit(&dp
->dp_lock
);
941 upgrade_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*hds
, void *arg
)
944 dsl_dataset_t
*ds
, *prev
= NULL
;
947 err
= dsl_dataset_hold_obj(dp
, hds
->ds_object
, FTAG
, &ds
);
951 while (dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= 0) {
952 err
= dsl_dataset_hold_obj(dp
,
953 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, FTAG
, &prev
);
955 dsl_dataset_rele(ds
, FTAG
);
959 if (dsl_dataset_phys(prev
)->ds_next_snap_obj
!= ds
->ds_object
)
961 dsl_dataset_rele(ds
, FTAG
);
967 prev
= dp
->dp_origin_snap
;
970 * The $ORIGIN can't have any data, or the accounting
973 rrw_enter(&ds
->ds_bp_rwlock
, RW_READER
, FTAG
);
974 ASSERT0(dsl_dataset_phys(prev
)->ds_bp
.blk_birth
);
975 rrw_exit(&ds
->ds_bp_rwlock
, FTAG
);
977 /* The origin doesn't get attached to itself */
978 if (ds
->ds_object
== prev
->ds_object
) {
979 dsl_dataset_rele(ds
, FTAG
);
983 dmu_buf_will_dirty(ds
->ds_dbuf
, tx
);
984 dsl_dataset_phys(ds
)->ds_prev_snap_obj
= prev
->ds_object
;
985 dsl_dataset_phys(ds
)->ds_prev_snap_txg
=
986 dsl_dataset_phys(prev
)->ds_creation_txg
;
988 dmu_buf_will_dirty(ds
->ds_dir
->dd_dbuf
, tx
);
989 dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
= prev
->ds_object
;
991 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
992 dsl_dataset_phys(prev
)->ds_num_children
++;
994 if (dsl_dataset_phys(ds
)->ds_next_snap_obj
== 0) {
995 ASSERT(ds
->ds_prev
== NULL
);
996 VERIFY0(dsl_dataset_hold_obj(dp
,
997 dsl_dataset_phys(ds
)->ds_prev_snap_obj
,
1002 ASSERT3U(dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
, ==, prev
->ds_object
);
1003 ASSERT3U(dsl_dataset_phys(ds
)->ds_prev_snap_obj
, ==, prev
->ds_object
);
1005 if (dsl_dataset_phys(prev
)->ds_next_clones_obj
== 0) {
1006 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
1007 dsl_dataset_phys(prev
)->ds_next_clones_obj
=
1008 zap_create(dp
->dp_meta_objset
,
1009 DMU_OT_NEXT_CLONES
, DMU_OT_NONE
, 0, tx
);
1011 VERIFY0(zap_add_int(dp
->dp_meta_objset
,
1012 dsl_dataset_phys(prev
)->ds_next_clones_obj
, ds
->ds_object
, tx
));
1014 dsl_dataset_rele(ds
, FTAG
);
1015 if (prev
!= dp
->dp_origin_snap
)
1016 dsl_dataset_rele(prev
, FTAG
);
1021 dsl_pool_upgrade_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
1023 ASSERT(dmu_tx_is_syncing(tx
));
1024 ASSERT(dp
->dp_origin_snap
!= NULL
);
1026 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
, upgrade_clones_cb
,
1027 tx
, DS_FIND_CHILDREN
| DS_FIND_SERIALIZE
));
1032 upgrade_dir_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *arg
)
1035 objset_t
*mos
= dp
->dp_meta_objset
;
1037 if (dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
!= 0) {
1038 dsl_dataset_t
*origin
;
1040 VERIFY0(dsl_dataset_hold_obj(dp
,
1041 dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
, FTAG
, &origin
));
1043 if (dsl_dir_phys(origin
->ds_dir
)->dd_clones
== 0) {
1044 dmu_buf_will_dirty(origin
->ds_dir
->dd_dbuf
, tx
);
1045 dsl_dir_phys(origin
->ds_dir
)->dd_clones
=
1046 zap_create(mos
, DMU_OT_DSL_CLONES
, DMU_OT_NONE
,
1050 VERIFY0(zap_add_int(dp
->dp_meta_objset
,
1051 dsl_dir_phys(origin
->ds_dir
)->dd_clones
,
1052 ds
->ds_object
, tx
));
1054 dsl_dataset_rele(origin
, FTAG
);
1060 dsl_pool_upgrade_dir_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
1064 ASSERT(dmu_tx_is_syncing(tx
));
1066 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, FREE_DIR_NAME
, tx
);
1067 VERIFY0(dsl_pool_open_special_dir(dp
,
1068 FREE_DIR_NAME
, &dp
->dp_free_dir
));
1071 * We can't use bpobj_alloc(), because spa_version() still
1072 * returns the old version, and we need a new-version bpobj with
1073 * subobj support. So call dmu_object_alloc() directly.
1075 obj
= dmu_object_alloc(dp
->dp_meta_objset
, DMU_OT_BPOBJ
,
1076 SPA_OLD_MAXBLOCKSIZE
, DMU_OT_BPOBJ_HDR
, sizeof (bpobj_phys_t
), tx
);
1077 VERIFY0(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
1078 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
));
1079 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
, dp
->dp_meta_objset
, obj
));
1081 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
,
1082 upgrade_dir_clones_cb
, tx
, DS_FIND_CHILDREN
| DS_FIND_SERIALIZE
));
1086 dsl_pool_create_origin(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
1091 ASSERT(dmu_tx_is_syncing(tx
));
1092 ASSERT(dp
->dp_origin_snap
== NULL
);
1093 ASSERT(rrw_held(&dp
->dp_config_rwlock
, RW_WRITER
));
1095 /* create the origin dir, ds, & snap-ds */
1096 dsobj
= dsl_dataset_create_sync(dp
->dp_root_dir
, ORIGIN_DIR_NAME
,
1097 NULL
, 0, kcred
, NULL
, tx
);
1098 VERIFY0(dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
1099 dsl_dataset_snapshot_sync_impl(ds
, ORIGIN_DIR_NAME
, tx
);
1100 VERIFY0(dsl_dataset_hold_obj(dp
, dsl_dataset_phys(ds
)->ds_prev_snap_obj
,
1101 dp
, &dp
->dp_origin_snap
));
1102 dsl_dataset_rele(ds
, FTAG
);
1106 dsl_pool_zrele_taskq(dsl_pool_t
*dp
)
1108 return (dp
->dp_zrele_taskq
);
1112 dsl_pool_unlinked_drain_taskq(dsl_pool_t
*dp
)
1114 return (dp
->dp_unlinked_drain_taskq
);
1118 * Walk through the pool-wide zap object of temporary snapshot user holds
1122 dsl_pool_clean_tmp_userrefs(dsl_pool_t
*dp
)
1126 objset_t
*mos
= dp
->dp_meta_objset
;
1127 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
1132 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
1134 holds
= fnvlist_alloc();
1136 for (zap_cursor_init(&zc
, mos
, zapobj
);
1137 zap_cursor_retrieve(&zc
, &za
) == 0;
1138 zap_cursor_advance(&zc
)) {
1142 htag
= strchr(za
.za_name
, '-');
1145 if (nvlist_lookup_nvlist(holds
, za
.za_name
, &tags
) != 0) {
1146 tags
= fnvlist_alloc();
1147 fnvlist_add_boolean(tags
, htag
);
1148 fnvlist_add_nvlist(holds
, za
.za_name
, tags
);
1151 fnvlist_add_boolean(tags
, htag
);
1154 dsl_dataset_user_release_tmp(dp
, holds
);
1155 fnvlist_free(holds
);
1156 zap_cursor_fini(&zc
);
1160 * Create the pool-wide zap object for storing temporary snapshot holds.
1163 dsl_pool_user_hold_create_obj(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
1165 objset_t
*mos
= dp
->dp_meta_objset
;
1167 ASSERT(dp
->dp_tmp_userrefs_obj
== 0);
1168 ASSERT(dmu_tx_is_syncing(tx
));
1170 dp
->dp_tmp_userrefs_obj
= zap_create_link(mos
, DMU_OT_USERREFS
,
1171 DMU_POOL_DIRECTORY_OBJECT
, DMU_POOL_TMP_USERREFS
, tx
);
1175 dsl_pool_user_hold_rele_impl(dsl_pool_t
*dp
, uint64_t dsobj
,
1176 const char *tag
, uint64_t now
, dmu_tx_t
*tx
, boolean_t holding
)
1178 objset_t
*mos
= dp
->dp_meta_objset
;
1179 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
1183 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
1184 ASSERT(dmu_tx_is_syncing(tx
));
1187 * If the pool was created prior to SPA_VERSION_USERREFS, the
1188 * zap object for temporary holds might not exist yet.
1192 dsl_pool_user_hold_create_obj(dp
, tx
);
1193 zapobj
= dp
->dp_tmp_userrefs_obj
;
1195 return (SET_ERROR(ENOENT
));
1199 name
= kmem_asprintf("%llx-%s", (u_longlong_t
)dsobj
, tag
);
1201 error
= zap_add(mos
, zapobj
, name
, 8, 1, &now
, tx
);
1203 error
= zap_remove(mos
, zapobj
, name
, tx
);
1210 * Add a temporary hold for the given dataset object and tag.
1213 dsl_pool_user_hold(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
1214 uint64_t now
, dmu_tx_t
*tx
)
1216 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, now
, tx
, B_TRUE
));
1220 * Release a temporary hold for the given dataset object and tag.
1223 dsl_pool_user_release(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
1226 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, 0,
1231 * DSL Pool Configuration Lock
1233 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
1234 * creation / destruction / rename / property setting). It must be held for
1235 * read to hold a dataset or dsl_dir. I.e. you must call
1236 * dsl_pool_config_enter() or dsl_pool_hold() before calling
1237 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
1238 * must be held continuously until all datasets and dsl_dirs are released.
1240 * The only exception to this rule is that if a "long hold" is placed on
1241 * a dataset, then the dp_config_rwlock may be dropped while the dataset
1242 * is still held. The long hold will prevent the dataset from being
1243 * destroyed -- the destroy will fail with EBUSY. A long hold can be
1244 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
1245 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
1247 * Legitimate long-holders (including owners) should be long-running, cancelable
1248 * tasks that should cause "zfs destroy" to fail. This includes DMU
1249 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
1250 * "zfs send", and "zfs diff". There are several other long-holders whose
1251 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
1253 * The usual formula for long-holding would be:
1255 * dsl_dataset_hold()
1256 * ... perform checks ...
1257 * dsl_dataset_long_hold()
1259 * ... perform long-running task ...
1260 * dsl_dataset_long_rele()
1261 * dsl_dataset_rele()
1263 * Note that when the long hold is released, the dataset is still held but
1264 * the pool is not held. The dataset may change arbitrarily during this time
1265 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1266 * dataset except release it.
1268 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1269 * or modifying operations.
1271 * Modifying operations should generally use dsl_sync_task(). The synctask
1272 * infrastructure enforces proper locking strategy with respect to the
1273 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1275 * Read-only operations will manually hold the pool, then the dataset, obtain
1276 * information from the dataset, then release the pool and dataset.
1277 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1282 dsl_pool_hold(const char *name
, void *tag
, dsl_pool_t
**dp
)
1287 error
= spa_open(name
, &spa
, tag
);
1289 *dp
= spa_get_dsl(spa
);
1290 dsl_pool_config_enter(*dp
, tag
);
1296 dsl_pool_rele(dsl_pool_t
*dp
, void *tag
)
1298 dsl_pool_config_exit(dp
, tag
);
1299 spa_close(dp
->dp_spa
, tag
);
1303 dsl_pool_config_enter(dsl_pool_t
*dp
, void *tag
)
1306 * We use a "reentrant" reader-writer lock, but not reentrantly.
1308 * The rrwlock can (with the track_all flag) track all reading threads,
1309 * which is very useful for debugging which code path failed to release
1310 * the lock, and for verifying that the *current* thread does hold
1313 * (Unlike a rwlock, which knows that N threads hold it for
1314 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1315 * if any thread holds it for read, even if this thread doesn't).
1317 ASSERT(!rrw_held(&dp
->dp_config_rwlock
, RW_READER
));
1318 rrw_enter(&dp
->dp_config_rwlock
, RW_READER
, tag
);
1322 dsl_pool_config_enter_prio(dsl_pool_t
*dp
, void *tag
)
1324 ASSERT(!rrw_held(&dp
->dp_config_rwlock
, RW_READER
));
1325 rrw_enter_read_prio(&dp
->dp_config_rwlock
, tag
);
1329 dsl_pool_config_exit(dsl_pool_t
*dp
, void *tag
)
1331 rrw_exit(&dp
->dp_config_rwlock
, tag
);
1335 dsl_pool_config_held(dsl_pool_t
*dp
)
1337 return (RRW_LOCK_HELD(&dp
->dp_config_rwlock
));
1341 dsl_pool_config_held_writer(dsl_pool_t
*dp
)
1343 return (RRW_WRITE_HELD(&dp
->dp_config_rwlock
));
1346 EXPORT_SYMBOL(dsl_pool_config_enter
);
1347 EXPORT_SYMBOL(dsl_pool_config_exit
);
1350 /* zfs_dirty_data_max_percent only applied at module load in arc_init(). */
1351 ZFS_MODULE_PARAM(zfs
, zfs_
, dirty_data_max_percent
, INT
, ZMOD_RD
,
1352 "Max percent of RAM allowed to be dirty");
1354 /* zfs_dirty_data_max_max_percent only applied at module load in arc_init(). */
1355 ZFS_MODULE_PARAM(zfs
, zfs_
, dirty_data_max_max_percent
, INT
, ZMOD_RD
,
1356 "zfs_dirty_data_max upper bound as % of RAM");
1358 ZFS_MODULE_PARAM(zfs
, zfs_
, delay_min_dirty_percent
, INT
, ZMOD_RW
,
1359 "Transaction delay threshold");
1361 ZFS_MODULE_PARAM(zfs
, zfs_
, dirty_data_max
, ULONG
, ZMOD_RW
,
1362 "Determines the dirty space limit");
1364 /* zfs_dirty_data_max_max only applied at module load in arc_init(). */
1365 ZFS_MODULE_PARAM(zfs
, zfs_
, dirty_data_max_max
, ULONG
, ZMOD_RD
,
1366 "zfs_dirty_data_max upper bound in bytes");
1368 ZFS_MODULE_PARAM(zfs
, zfs_
, dirty_data_sync_percent
, INT
, ZMOD_RW
,
1369 "Dirty data txg sync threshold as a percentage of zfs_dirty_data_max");
1371 ZFS_MODULE_PARAM(zfs
, zfs_
, delay_scale
, ULONG
, ZMOD_RW
,
1372 "How quickly delay approaches infinity");
1374 ZFS_MODULE_PARAM(zfs
, zfs_
, sync_taskq_batch_pct
, INT
, ZMOD_RW
,
1375 "Max percent of CPUs that are used to sync dirty data");
1377 ZFS_MODULE_PARAM(zfs_zil
, zfs_zil_
, clean_taskq_nthr_pct
, INT
, ZMOD_RW
,
1378 "Max percent of CPUs that are used per dp_sync_taskq");
1380 ZFS_MODULE_PARAM(zfs_zil
, zfs_zil_
, clean_taskq_minalloc
, INT
, ZMOD_RW
,
1381 "Number of taskq entries that are pre-populated");
1383 ZFS_MODULE_PARAM(zfs_zil
, zfs_zil_
, clean_taskq_maxalloc
, INT
, ZMOD_RW
,
1384 "Max number of taskq entries that are cached");