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, 2017 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/dsl_deadlist.h>
46 #include <sys/bptree.h>
47 #include <sys/zfeature.h>
48 #include <sys/zil_impl.h>
49 #include <sys/dsl_userhold.h>
50 #include <sys/trace_txg.h>
57 * ZFS must limit the rate of incoming writes to the rate at which it is able
58 * to sync data modifications to the backend storage. Throttling by too much
59 * creates an artificial limit; throttling by too little can only be sustained
60 * for short periods and would lead to highly lumpy performance. On a per-pool
61 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
62 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
63 * of dirty data decreases. When the amount of dirty data exceeds a
64 * predetermined threshold further modifications are blocked until the amount
65 * of dirty data decreases (as data is synced out).
67 * The limit on dirty data is tunable, and should be adjusted according to
68 * both the IO capacity and available memory of the system. The larger the
69 * window, the more ZFS is able to aggregate and amortize metadata (and data)
70 * changes. However, memory is a limited resource, and allowing for more dirty
71 * data comes at the cost of keeping other useful data in memory (for example
72 * ZFS data cached by the ARC).
76 * As buffers are modified dsl_pool_willuse_space() increments both the per-
77 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
78 * dirty space used; dsl_pool_dirty_space() decrements those values as data
79 * is synced out from dsl_pool_sync(). While only the poolwide value is
80 * relevant, the per-txg value is useful for debugging. The tunable
81 * zfs_dirty_data_max determines the dirty space limit. Once that value is
82 * exceeded, new writes are halted until space frees up.
84 * The zfs_dirty_data_sync tunable dictates the threshold at which we
85 * ensure that there is a txg syncing (see the comment in txg.c for a full
86 * description of transaction group stages).
88 * The IO scheduler uses both the dirty space limit and current amount of
89 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
90 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
92 * The delay is also calculated based on the amount of dirty data. See the
93 * comment above dmu_tx_delay() for details.
97 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
98 * capped at zfs_dirty_data_max_max. It can also be overridden with a module
101 unsigned long zfs_dirty_data_max
= 0;
102 unsigned long zfs_dirty_data_max_max
= 0;
103 int zfs_dirty_data_max_percent
= 10;
104 int zfs_dirty_data_max_max_percent
= 25;
107 * If there is at least this much dirty data, push out a txg.
109 unsigned long zfs_dirty_data_sync
= 64 * 1024 * 1024;
112 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
113 * and delay each transaction.
114 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
116 int zfs_delay_min_dirty_percent
= 60;
119 * This controls how quickly the delay approaches infinity.
120 * Larger values cause it to delay more for a given amount of dirty data.
121 * Therefore larger values will cause there to be less dirty data for a
124 * For the smoothest delay, this value should be about 1 billion divided
125 * by the maximum number of operations per second. This will smoothly
126 * handle between 10x and 1/10th this number.
128 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
129 * multiply in dmu_tx_delay().
131 unsigned long zfs_delay_scale
= 1000 * 1000 * 1000 / 2000;
134 * This determines the number of threads used by the dp_sync_taskq.
136 int zfs_sync_taskq_batch_pct
= 75;
139 * These tunables determine the behavior of how zil_itxg_clean() is
140 * called via zil_clean() in the context of spa_sync(). When an itxg
141 * list needs to be cleaned, TQ_NOSLEEP will be used when dispatching.
142 * If the dispatch fails, the call to zil_itxg_clean() will occur
143 * synchronously in the context of spa_sync(), which can negatively
144 * impact the performance of spa_sync() (e.g. in the case of the itxg
145 * list having a large number of itxs that needs to be cleaned).
147 * Thus, these tunables can be used to manipulate the behavior of the
148 * taskq used by zil_clean(); they determine the number of taskq entries
149 * that are pre-populated when the taskq is first created (via the
150 * "zfs_zil_clean_taskq_minalloc" tunable) and the maximum number of
151 * taskq entries that are cached after an on-demand allocation (via the
152 * "zfs_zil_clean_taskq_maxalloc").
154 * The idea being, we want to try reasonably hard to ensure there will
155 * already be a taskq entry pre-allocated by the time that it is needed
156 * by zil_clean(). This way, we can avoid the possibility of an
157 * on-demand allocation of a new taskq entry from failing, which would
158 * result in zil_itxg_clean() being called synchronously from zil_clean()
159 * (which can adversely affect performance of spa_sync()).
161 * Additionally, the number of threads used by the taskq can be
162 * configured via the "zfs_zil_clean_taskq_nthr_pct" tunable.
164 int zfs_zil_clean_taskq_nthr_pct
= 100;
165 int zfs_zil_clean_taskq_minalloc
= 1024;
166 int zfs_zil_clean_taskq_maxalloc
= 1024 * 1024;
169 dsl_pool_open_special_dir(dsl_pool_t
*dp
, const char *name
, dsl_dir_t
**ddp
)
174 err
= zap_lookup(dp
->dp_meta_objset
,
175 dsl_dir_phys(dp
->dp_root_dir
)->dd_child_dir_zapobj
,
176 name
, sizeof (obj
), 1, &obj
);
180 return (dsl_dir_hold_obj(dp
, obj
, name
, dp
, ddp
));
184 dsl_pool_open_impl(spa_t
*spa
, uint64_t txg
)
187 blkptr_t
*bp
= spa_get_rootblkptr(spa
);
189 dp
= kmem_zalloc(sizeof (dsl_pool_t
), KM_SLEEP
);
191 dp
->dp_meta_rootbp
= *bp
;
192 rrw_init(&dp
->dp_config_rwlock
, B_TRUE
);
196 txg_list_create(&dp
->dp_dirty_datasets
, spa
,
197 offsetof(dsl_dataset_t
, ds_dirty_link
));
198 txg_list_create(&dp
->dp_dirty_zilogs
, spa
,
199 offsetof(zilog_t
, zl_dirty_link
));
200 txg_list_create(&dp
->dp_dirty_dirs
, spa
,
201 offsetof(dsl_dir_t
, dd_dirty_link
));
202 txg_list_create(&dp
->dp_sync_tasks
, spa
,
203 offsetof(dsl_sync_task_t
, dst_node
));
205 dp
->dp_sync_taskq
= taskq_create("dp_sync_taskq",
206 zfs_sync_taskq_batch_pct
, minclsyspri
, 1, INT_MAX
,
207 TASKQ_THREADS_CPU_PCT
);
209 dp
->dp_zil_clean_taskq
= taskq_create("dp_zil_clean_taskq",
210 zfs_zil_clean_taskq_nthr_pct
, minclsyspri
,
211 zfs_zil_clean_taskq_minalloc
,
212 zfs_zil_clean_taskq_maxalloc
,
213 TASKQ_PREPOPULATE
| TASKQ_THREADS_CPU_PCT
);
215 mutex_init(&dp
->dp_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
216 cv_init(&dp
->dp_spaceavail_cv
, NULL
, CV_DEFAULT
, NULL
);
218 dp
->dp_iput_taskq
= taskq_create("z_iput", max_ncpus
, defclsyspri
,
219 max_ncpus
* 8, INT_MAX
, TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
);
225 dsl_pool_init(spa_t
*spa
, uint64_t txg
, dsl_pool_t
**dpp
)
228 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
231 * Initialize the caller's dsl_pool_t structure before we actually open
232 * the meta objset. This is done because a self-healing write zio may
233 * be issued as part of dmu_objset_open_impl() and the spa needs its
234 * dsl_pool_t initialized in order to handle the write.
238 err
= dmu_objset_open_impl(spa
, NULL
, &dp
->dp_meta_rootbp
,
239 &dp
->dp_meta_objset
);
249 dsl_pool_open(dsl_pool_t
*dp
)
256 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
257 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
258 DMU_POOL_ROOT_DATASET
, sizeof (uint64_t), 1,
259 &dp
->dp_root_dir_obj
);
263 err
= dsl_dir_hold_obj(dp
, dp
->dp_root_dir_obj
,
264 NULL
, dp
, &dp
->dp_root_dir
);
268 err
= dsl_pool_open_special_dir(dp
, MOS_DIR_NAME
, &dp
->dp_mos_dir
);
272 if (spa_version(dp
->dp_spa
) >= SPA_VERSION_ORIGIN
) {
273 err
= dsl_pool_open_special_dir(dp
, ORIGIN_DIR_NAME
, &dd
);
276 err
= dsl_dataset_hold_obj(dp
,
277 dsl_dir_phys(dd
)->dd_head_dataset_obj
, FTAG
, &ds
);
279 err
= dsl_dataset_hold_obj(dp
,
280 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, dp
,
281 &dp
->dp_origin_snap
);
282 dsl_dataset_rele(ds
, FTAG
);
284 dsl_dir_rele(dd
, dp
);
289 if (spa_version(dp
->dp_spa
) >= SPA_VERSION_DEADLISTS
) {
290 err
= dsl_pool_open_special_dir(dp
, FREE_DIR_NAME
,
295 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
296 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
);
299 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
,
300 dp
->dp_meta_objset
, obj
));
304 * Note: errors ignored, because the leak dir will not exist if we
305 * have not encountered a leak yet.
307 (void) dsl_pool_open_special_dir(dp
, LEAK_DIR_NAME
,
310 if (spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_ASYNC_DESTROY
)) {
311 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
312 DMU_POOL_BPTREE_OBJ
, sizeof (uint64_t), 1,
318 if (spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_EMPTY_BPOBJ
)) {
319 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
320 DMU_POOL_EMPTY_BPOBJ
, sizeof (uint64_t), 1,
321 &dp
->dp_empty_bpobj
);
326 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
327 DMU_POOL_TMP_USERREFS
, sizeof (uint64_t), 1,
328 &dp
->dp_tmp_userrefs_obj
);
334 err
= dsl_scan_init(dp
, dp
->dp_tx
.tx_open_txg
);
337 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
342 dsl_pool_close(dsl_pool_t
*dp
)
345 * Drop our references from dsl_pool_open().
347 * Since we held the origin_snap from "syncing" context (which
348 * includes pool-opening context), it actually only got a "ref"
349 * and not a hold, so just drop that here.
351 if (dp
->dp_origin_snap
)
352 dsl_dataset_rele(dp
->dp_origin_snap
, dp
);
354 dsl_dir_rele(dp
->dp_mos_dir
, dp
);
356 dsl_dir_rele(dp
->dp_free_dir
, dp
);
358 dsl_dir_rele(dp
->dp_leak_dir
, dp
);
360 dsl_dir_rele(dp
->dp_root_dir
, dp
);
362 bpobj_close(&dp
->dp_free_bpobj
);
364 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
365 if (dp
->dp_meta_objset
)
366 dmu_objset_evict(dp
->dp_meta_objset
);
368 txg_list_destroy(&dp
->dp_dirty_datasets
);
369 txg_list_destroy(&dp
->dp_dirty_zilogs
);
370 txg_list_destroy(&dp
->dp_sync_tasks
);
371 txg_list_destroy(&dp
->dp_dirty_dirs
);
373 taskq_destroy(dp
->dp_zil_clean_taskq
);
374 taskq_destroy(dp
->dp_sync_taskq
);
377 * We can't set retry to TRUE since we're explicitly specifying
378 * a spa to flush. This is good enough; any missed buffers for
379 * this spa won't cause trouble, and they'll eventually fall
380 * out of the ARC just like any other unused buffer.
382 arc_flush(dp
->dp_spa
, FALSE
);
384 mmp_fini(dp
->dp_spa
);
387 dmu_buf_user_evict_wait();
389 rrw_destroy(&dp
->dp_config_rwlock
);
390 mutex_destroy(&dp
->dp_lock
);
391 cv_destroy(&dp
->dp_spaceavail_cv
);
392 taskq_destroy(dp
->dp_iput_taskq
);
394 vmem_free(dp
->dp_blkstats
, sizeof (zfs_all_blkstats_t
));
395 kmem_free(dp
, sizeof (dsl_pool_t
));
399 dsl_pool_create(spa_t
*spa
, nvlist_t
*zplprops
, dsl_crypto_params_t
*dcp
,
403 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
404 dmu_tx_t
*tx
= dmu_tx_create_assigned(dp
, txg
);
408 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
410 /* create and open the MOS (meta-objset) */
411 dp
->dp_meta_objset
= dmu_objset_create_impl(spa
,
412 NULL
, &dp
->dp_meta_rootbp
, DMU_OST_META
, tx
);
413 spa
->spa_meta_objset
= dp
->dp_meta_objset
;
415 /* create the pool directory */
416 err
= zap_create_claim(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
417 DMU_OT_OBJECT_DIRECTORY
, DMU_OT_NONE
, 0, tx
);
420 /* Initialize scan structures */
421 VERIFY0(dsl_scan_init(dp
, txg
));
423 /* create and open the root dir */
424 dp
->dp_root_dir_obj
= dsl_dir_create_sync(dp
, NULL
, NULL
, tx
);
425 VERIFY0(dsl_dir_hold_obj(dp
, dp
->dp_root_dir_obj
,
426 NULL
, dp
, &dp
->dp_root_dir
));
428 /* create and open the meta-objset dir */
429 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, MOS_DIR_NAME
, tx
);
430 VERIFY0(dsl_pool_open_special_dir(dp
,
431 MOS_DIR_NAME
, &dp
->dp_mos_dir
));
433 if (spa_version(spa
) >= SPA_VERSION_DEADLISTS
) {
434 /* create and open the free dir */
435 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
,
437 VERIFY0(dsl_pool_open_special_dir(dp
,
438 FREE_DIR_NAME
, &dp
->dp_free_dir
));
440 /* create and open the free_bplist */
441 obj
= bpobj_alloc(dp
->dp_meta_objset
, SPA_OLD_MAXBLOCKSIZE
, tx
);
442 VERIFY(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
443 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
) == 0);
444 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
,
445 dp
->dp_meta_objset
, obj
));
448 if (spa_version(spa
) >= SPA_VERSION_DSL_SCRUB
)
449 dsl_pool_create_origin(dp
, tx
);
452 * Some features may be needed when creating the root dataset, so we
453 * create the feature objects here.
455 if (spa_version(spa
) >= SPA_VERSION_FEATURES
)
456 spa_feature_create_zap_objects(spa
, tx
);
458 if (dcp
!= NULL
&& dcp
->cp_crypt
!= ZIO_CRYPT_OFF
&&
459 dcp
->cp_crypt
!= ZIO_CRYPT_INHERIT
)
460 spa_feature_enable(spa
, SPA_FEATURE_ENCRYPTION
, tx
);
462 /* create the root dataset */
463 obj
= dsl_dataset_create_sync_dd(dp
->dp_root_dir
, NULL
, dcp
, 0, tx
);
465 /* create the root objset */
466 VERIFY0(dsl_dataset_hold_obj(dp
, obj
, FTAG
, &ds
));
470 rrw_enter(&ds
->ds_bp_rwlock
, RW_READER
, FTAG
);
471 os
= dmu_objset_create_impl(dp
->dp_spa
, ds
,
472 dsl_dataset_get_blkptr(ds
), DMU_OST_ZFS
, tx
);
473 rrw_exit(&ds
->ds_bp_rwlock
, FTAG
);
474 zfs_create_fs(os
, kcred
, zplprops
, tx
);
477 dsl_dataset_rele(ds
, FTAG
);
481 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
487 * Account for the meta-objset space in its placeholder dsl_dir.
490 dsl_pool_mos_diduse_space(dsl_pool_t
*dp
,
491 int64_t used
, int64_t comp
, int64_t uncomp
)
493 ASSERT3U(comp
, ==, uncomp
); /* it's all metadata */
494 mutex_enter(&dp
->dp_lock
);
495 dp
->dp_mos_used_delta
+= used
;
496 dp
->dp_mos_compressed_delta
+= comp
;
497 dp
->dp_mos_uncompressed_delta
+= uncomp
;
498 mutex_exit(&dp
->dp_lock
);
502 dsl_pool_sync_mos(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
504 zio_t
*zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
505 dmu_objset_sync(dp
->dp_meta_objset
, zio
, tx
);
506 VERIFY0(zio_wait(zio
));
507 dprintf_bp(&dp
->dp_meta_rootbp
, "meta objset rootbp is %s", "");
508 spa_set_rootblkptr(dp
->dp_spa
, &dp
->dp_meta_rootbp
);
512 dsl_pool_dirty_delta(dsl_pool_t
*dp
, int64_t delta
)
514 ASSERT(MUTEX_HELD(&dp
->dp_lock
));
517 ASSERT3U(-delta
, <=, dp
->dp_dirty_total
);
519 dp
->dp_dirty_total
+= delta
;
522 * Note: we signal even when increasing dp_dirty_total.
523 * This ensures forward progress -- each thread wakes the next waiter.
525 if (dp
->dp_dirty_total
< zfs_dirty_data_max
)
526 cv_signal(&dp
->dp_spaceavail_cv
);
530 dsl_pool_sync(dsl_pool_t
*dp
, uint64_t txg
)
536 objset_t
*mos
= dp
->dp_meta_objset
;
537 list_t synced_datasets
;
539 list_create(&synced_datasets
, sizeof (dsl_dataset_t
),
540 offsetof(dsl_dataset_t
, ds_synced_link
));
542 tx
= dmu_tx_create_assigned(dp
, txg
);
545 * Write out all dirty blocks of dirty datasets.
547 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
548 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
550 * We must not sync any non-MOS datasets twice, because
551 * we may have taken a snapshot of them. However, we
552 * may sync newly-created datasets on pass 2.
554 ASSERT(!list_link_active(&ds
->ds_synced_link
));
555 list_insert_tail(&synced_datasets
, ds
);
556 dsl_dataset_sync(ds
, zio
, tx
);
558 VERIFY0(zio_wait(zio
));
561 * We have written all of the accounted dirty data, so our
562 * dp_space_towrite should now be zero. However, some seldom-used
563 * code paths do not adhere to this (e.g. dbuf_undirty(), also
564 * rounding error in dbuf_write_physdone).
565 * Shore up the accounting of any dirtied space now.
567 dsl_pool_undirty_space(dp
, dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], txg
);
570 * Update the long range free counter after
571 * we're done syncing user data
573 mutex_enter(&dp
->dp_lock
);
574 ASSERT(spa_sync_pass(dp
->dp_spa
) == 1 ||
575 dp
->dp_long_free_dirty_pertxg
[txg
& TXG_MASK
] == 0);
576 dp
->dp_long_free_dirty_pertxg
[txg
& TXG_MASK
] = 0;
577 mutex_exit(&dp
->dp_lock
);
580 * After the data blocks have been written (ensured by the zio_wait()
581 * above), update the user/group space accounting. This happens
582 * in tasks dispatched to dp_sync_taskq, so wait for them before
585 for (ds
= list_head(&synced_datasets
); ds
!= NULL
;
586 ds
= list_next(&synced_datasets
, ds
)) {
587 dmu_objset_do_userquota_updates(ds
->ds_objset
, tx
);
589 taskq_wait(dp
->dp_sync_taskq
);
592 * Sync the datasets again to push out the changes due to
593 * userspace updates. This must be done before we process the
594 * sync tasks, so that any snapshots will have the correct
595 * user accounting information (and we won't get confused
596 * about which blocks are part of the snapshot).
598 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
599 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
600 ASSERT(list_link_active(&ds
->ds_synced_link
));
601 dmu_buf_rele(ds
->ds_dbuf
, ds
);
602 dsl_dataset_sync(ds
, zio
, tx
);
604 VERIFY0(zio_wait(zio
));
607 * Now that the datasets have been completely synced, we can
608 * clean up our in-memory structures accumulated while syncing:
610 * - move dead blocks from the pending deadlist to the on-disk deadlist
611 * - release hold from dsl_dataset_dirty()
613 while ((ds
= list_remove_head(&synced_datasets
)) != NULL
) {
614 dsl_dataset_sync_done(ds
, tx
);
617 while ((dd
= txg_list_remove(&dp
->dp_dirty_dirs
, txg
)) != NULL
) {
618 dsl_dir_sync(dd
, tx
);
622 * The MOS's space is accounted for in the pool/$MOS
623 * (dp_mos_dir). We can't modify the mos while we're syncing
624 * it, so we remember the deltas and apply them here.
626 if (dp
->dp_mos_used_delta
!= 0 || dp
->dp_mos_compressed_delta
!= 0 ||
627 dp
->dp_mos_uncompressed_delta
!= 0) {
628 dsl_dir_diduse_space(dp
->dp_mos_dir
, DD_USED_HEAD
,
629 dp
->dp_mos_used_delta
,
630 dp
->dp_mos_compressed_delta
,
631 dp
->dp_mos_uncompressed_delta
, tx
);
632 dp
->dp_mos_used_delta
= 0;
633 dp
->dp_mos_compressed_delta
= 0;
634 dp
->dp_mos_uncompressed_delta
= 0;
637 if (!multilist_is_empty(mos
->os_dirty_dnodes
[txg
& TXG_MASK
])) {
638 dsl_pool_sync_mos(dp
, tx
);
642 * If we modify a dataset in the same txg that we want to destroy it,
643 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
644 * dsl_dir_destroy_check() will fail if there are unexpected holds.
645 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
646 * and clearing the hold on it) before we process the sync_tasks.
647 * The MOS data dirtied by the sync_tasks will be synced on the next
650 if (!txg_list_empty(&dp
->dp_sync_tasks
, txg
)) {
651 dsl_sync_task_t
*dst
;
653 * No more sync tasks should have been added while we
656 ASSERT3U(spa_sync_pass(dp
->dp_spa
), ==, 1);
657 while ((dst
= txg_list_remove(&dp
->dp_sync_tasks
, txg
)) != NULL
)
658 dsl_sync_task_sync(dst
, tx
);
663 DTRACE_PROBE2(dsl_pool_sync__done
, dsl_pool_t
*dp
, dp
, uint64_t, txg
);
667 dsl_pool_sync_done(dsl_pool_t
*dp
, uint64_t txg
)
671 while ((zilog
= txg_list_head(&dp
->dp_dirty_zilogs
, txg
))) {
672 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
674 * We don't remove the zilog from the dp_dirty_zilogs
675 * list until after we've cleaned it. This ensures that
676 * callers of zilog_is_dirty() receive an accurate
677 * answer when they are racing with the spa sync thread.
679 zil_clean(zilog
, txg
);
680 (void) txg_list_remove_this(&dp
->dp_dirty_zilogs
, zilog
, txg
);
681 ASSERT(!dmu_objset_is_dirty(zilog
->zl_os
, txg
));
682 dmu_buf_rele(ds
->ds_dbuf
, zilog
);
684 ASSERT(!dmu_objset_is_dirty(dp
->dp_meta_objset
, txg
));
688 * TRUE if the current thread is the tx_sync_thread or if we
689 * are being called from SPA context during pool initialization.
692 dsl_pool_sync_context(dsl_pool_t
*dp
)
694 return (curthread
== dp
->dp_tx
.tx_sync_thread
||
695 spa_is_initializing(dp
->dp_spa
) ||
696 taskq_member(dp
->dp_sync_taskq
, curthread
));
700 dsl_pool_adjustedsize(dsl_pool_t
*dp
, boolean_t netfree
)
702 uint64_t space
, resv
;
705 * If we're trying to assess whether it's OK to do a free,
706 * cut the reservation in half to allow forward progress
707 * (e.g. make it possible to rm(1) files from a full pool).
709 space
= spa_get_dspace(dp
->dp_spa
);
710 resv
= spa_get_slop_space(dp
->dp_spa
);
714 return (space
- resv
);
718 dsl_pool_need_dirty_delay(dsl_pool_t
*dp
)
720 uint64_t delay_min_bytes
=
721 zfs_dirty_data_max
* zfs_delay_min_dirty_percent
/ 100;
724 mutex_enter(&dp
->dp_lock
);
725 if (dp
->dp_dirty_total
> zfs_dirty_data_sync
)
727 rv
= (dp
->dp_dirty_total
> delay_min_bytes
);
728 mutex_exit(&dp
->dp_lock
);
733 dsl_pool_dirty_space(dsl_pool_t
*dp
, int64_t space
, dmu_tx_t
*tx
)
736 mutex_enter(&dp
->dp_lock
);
737 dp
->dp_dirty_pertxg
[tx
->tx_txg
& TXG_MASK
] += space
;
738 dsl_pool_dirty_delta(dp
, space
);
739 mutex_exit(&dp
->dp_lock
);
744 dsl_pool_undirty_space(dsl_pool_t
*dp
, int64_t space
, uint64_t txg
)
746 ASSERT3S(space
, >=, 0);
750 mutex_enter(&dp
->dp_lock
);
751 if (dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] < space
) {
752 /* XXX writing something we didn't dirty? */
753 space
= dp
->dp_dirty_pertxg
[txg
& TXG_MASK
];
755 ASSERT3U(dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], >=, space
);
756 dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] -= space
;
757 ASSERT3U(dp
->dp_dirty_total
, >=, space
);
758 dsl_pool_dirty_delta(dp
, -space
);
759 mutex_exit(&dp
->dp_lock
);
764 upgrade_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*hds
, void *arg
)
767 dsl_dataset_t
*ds
, *prev
= NULL
;
770 err
= dsl_dataset_hold_obj(dp
, hds
->ds_object
, FTAG
, &ds
);
774 while (dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= 0) {
775 err
= dsl_dataset_hold_obj(dp
,
776 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, FTAG
, &prev
);
778 dsl_dataset_rele(ds
, FTAG
);
782 if (dsl_dataset_phys(prev
)->ds_next_snap_obj
!= ds
->ds_object
)
784 dsl_dataset_rele(ds
, FTAG
);
790 prev
= dp
->dp_origin_snap
;
793 * The $ORIGIN can't have any data, or the accounting
796 rrw_enter(&ds
->ds_bp_rwlock
, RW_READER
, FTAG
);
797 ASSERT0(dsl_dataset_phys(prev
)->ds_bp
.blk_birth
);
798 rrw_exit(&ds
->ds_bp_rwlock
, FTAG
);
800 /* The origin doesn't get attached to itself */
801 if (ds
->ds_object
== prev
->ds_object
) {
802 dsl_dataset_rele(ds
, FTAG
);
806 dmu_buf_will_dirty(ds
->ds_dbuf
, tx
);
807 dsl_dataset_phys(ds
)->ds_prev_snap_obj
= prev
->ds_object
;
808 dsl_dataset_phys(ds
)->ds_prev_snap_txg
=
809 dsl_dataset_phys(prev
)->ds_creation_txg
;
811 dmu_buf_will_dirty(ds
->ds_dir
->dd_dbuf
, tx
);
812 dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
= prev
->ds_object
;
814 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
815 dsl_dataset_phys(prev
)->ds_num_children
++;
817 if (dsl_dataset_phys(ds
)->ds_next_snap_obj
== 0) {
818 ASSERT(ds
->ds_prev
== NULL
);
819 VERIFY0(dsl_dataset_hold_obj(dp
,
820 dsl_dataset_phys(ds
)->ds_prev_snap_obj
,
825 ASSERT3U(dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
, ==, prev
->ds_object
);
826 ASSERT3U(dsl_dataset_phys(ds
)->ds_prev_snap_obj
, ==, prev
->ds_object
);
828 if (dsl_dataset_phys(prev
)->ds_next_clones_obj
== 0) {
829 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
830 dsl_dataset_phys(prev
)->ds_next_clones_obj
=
831 zap_create(dp
->dp_meta_objset
,
832 DMU_OT_NEXT_CLONES
, DMU_OT_NONE
, 0, tx
);
834 VERIFY0(zap_add_int(dp
->dp_meta_objset
,
835 dsl_dataset_phys(prev
)->ds_next_clones_obj
, ds
->ds_object
, tx
));
837 dsl_dataset_rele(ds
, FTAG
);
838 if (prev
!= dp
->dp_origin_snap
)
839 dsl_dataset_rele(prev
, FTAG
);
844 dsl_pool_upgrade_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
846 ASSERT(dmu_tx_is_syncing(tx
));
847 ASSERT(dp
->dp_origin_snap
!= NULL
);
849 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
, upgrade_clones_cb
,
850 tx
, DS_FIND_CHILDREN
| DS_FIND_SERIALIZE
));
855 upgrade_dir_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *arg
)
858 objset_t
*mos
= dp
->dp_meta_objset
;
860 if (dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
!= 0) {
861 dsl_dataset_t
*origin
;
863 VERIFY0(dsl_dataset_hold_obj(dp
,
864 dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
, FTAG
, &origin
));
866 if (dsl_dir_phys(origin
->ds_dir
)->dd_clones
== 0) {
867 dmu_buf_will_dirty(origin
->ds_dir
->dd_dbuf
, tx
);
868 dsl_dir_phys(origin
->ds_dir
)->dd_clones
=
869 zap_create(mos
, DMU_OT_DSL_CLONES
, DMU_OT_NONE
,
873 VERIFY0(zap_add_int(dp
->dp_meta_objset
,
874 dsl_dir_phys(origin
->ds_dir
)->dd_clones
,
877 dsl_dataset_rele(origin
, FTAG
);
883 dsl_pool_upgrade_dir_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
887 ASSERT(dmu_tx_is_syncing(tx
));
889 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, FREE_DIR_NAME
, tx
);
890 VERIFY0(dsl_pool_open_special_dir(dp
,
891 FREE_DIR_NAME
, &dp
->dp_free_dir
));
894 * We can't use bpobj_alloc(), because spa_version() still
895 * returns the old version, and we need a new-version bpobj with
896 * subobj support. So call dmu_object_alloc() directly.
898 obj
= dmu_object_alloc(dp
->dp_meta_objset
, DMU_OT_BPOBJ
,
899 SPA_OLD_MAXBLOCKSIZE
, DMU_OT_BPOBJ_HDR
, sizeof (bpobj_phys_t
), tx
);
900 VERIFY0(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
901 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
));
902 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
, dp
->dp_meta_objset
, obj
));
904 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
,
905 upgrade_dir_clones_cb
, tx
, DS_FIND_CHILDREN
| DS_FIND_SERIALIZE
));
909 dsl_pool_create_origin(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
914 ASSERT(dmu_tx_is_syncing(tx
));
915 ASSERT(dp
->dp_origin_snap
== NULL
);
916 ASSERT(rrw_held(&dp
->dp_config_rwlock
, RW_WRITER
));
918 /* create the origin dir, ds, & snap-ds */
919 dsobj
= dsl_dataset_create_sync(dp
->dp_root_dir
, ORIGIN_DIR_NAME
,
920 NULL
, 0, kcred
, NULL
, tx
);
921 VERIFY0(dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
922 dsl_dataset_snapshot_sync_impl(ds
, ORIGIN_DIR_NAME
, tx
);
923 VERIFY0(dsl_dataset_hold_obj(dp
, dsl_dataset_phys(ds
)->ds_prev_snap_obj
,
924 dp
, &dp
->dp_origin_snap
));
925 dsl_dataset_rele(ds
, FTAG
);
929 dsl_pool_iput_taskq(dsl_pool_t
*dp
)
931 return (dp
->dp_iput_taskq
);
935 * Walk through the pool-wide zap object of temporary snapshot user holds
939 dsl_pool_clean_tmp_userrefs(dsl_pool_t
*dp
)
943 objset_t
*mos
= dp
->dp_meta_objset
;
944 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
949 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
951 holds
= fnvlist_alloc();
953 for (zap_cursor_init(&zc
, mos
, zapobj
);
954 zap_cursor_retrieve(&zc
, &za
) == 0;
955 zap_cursor_advance(&zc
)) {
959 htag
= strchr(za
.za_name
, '-');
962 if (nvlist_lookup_nvlist(holds
, za
.za_name
, &tags
) != 0) {
963 tags
= fnvlist_alloc();
964 fnvlist_add_boolean(tags
, htag
);
965 fnvlist_add_nvlist(holds
, za
.za_name
, tags
);
968 fnvlist_add_boolean(tags
, htag
);
971 dsl_dataset_user_release_tmp(dp
, holds
);
973 zap_cursor_fini(&zc
);
977 * Create the pool-wide zap object for storing temporary snapshot holds.
980 dsl_pool_user_hold_create_obj(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
982 objset_t
*mos
= dp
->dp_meta_objset
;
984 ASSERT(dp
->dp_tmp_userrefs_obj
== 0);
985 ASSERT(dmu_tx_is_syncing(tx
));
987 dp
->dp_tmp_userrefs_obj
= zap_create_link(mos
, DMU_OT_USERREFS
,
988 DMU_POOL_DIRECTORY_OBJECT
, DMU_POOL_TMP_USERREFS
, tx
);
992 dsl_pool_user_hold_rele_impl(dsl_pool_t
*dp
, uint64_t dsobj
,
993 const char *tag
, uint64_t now
, dmu_tx_t
*tx
, boolean_t holding
)
995 objset_t
*mos
= dp
->dp_meta_objset
;
996 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
1000 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
1001 ASSERT(dmu_tx_is_syncing(tx
));
1004 * If the pool was created prior to SPA_VERSION_USERREFS, the
1005 * zap object for temporary holds might not exist yet.
1009 dsl_pool_user_hold_create_obj(dp
, tx
);
1010 zapobj
= dp
->dp_tmp_userrefs_obj
;
1012 return (SET_ERROR(ENOENT
));
1016 name
= kmem_asprintf("%llx-%s", (u_longlong_t
)dsobj
, tag
);
1018 error
= zap_add(mos
, zapobj
, name
, 8, 1, &now
, tx
);
1020 error
= zap_remove(mos
, zapobj
, name
, tx
);
1027 * Add a temporary hold for the given dataset object and tag.
1030 dsl_pool_user_hold(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
1031 uint64_t now
, dmu_tx_t
*tx
)
1033 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, now
, tx
, B_TRUE
));
1037 * Release a temporary hold for the given dataset object and tag.
1040 dsl_pool_user_release(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
1043 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, 0,
1048 * DSL Pool Configuration Lock
1050 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
1051 * creation / destruction / rename / property setting). It must be held for
1052 * read to hold a dataset or dsl_dir. I.e. you must call
1053 * dsl_pool_config_enter() or dsl_pool_hold() before calling
1054 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
1055 * must be held continuously until all datasets and dsl_dirs are released.
1057 * The only exception to this rule is that if a "long hold" is placed on
1058 * a dataset, then the dp_config_rwlock may be dropped while the dataset
1059 * is still held. The long hold will prevent the dataset from being
1060 * destroyed -- the destroy will fail with EBUSY. A long hold can be
1061 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
1062 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
1064 * Legitimate long-holders (including owners) should be long-running, cancelable
1065 * tasks that should cause "zfs destroy" to fail. This includes DMU
1066 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
1067 * "zfs send", and "zfs diff". There are several other long-holders whose
1068 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
1070 * The usual formula for long-holding would be:
1072 * dsl_dataset_hold()
1073 * ... perform checks ...
1074 * dsl_dataset_long_hold()
1076 * ... perform long-running task ...
1077 * dsl_dataset_long_rele()
1078 * dsl_dataset_rele()
1080 * Note that when the long hold is released, the dataset is still held but
1081 * the pool is not held. The dataset may change arbitrarily during this time
1082 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1083 * dataset except release it.
1085 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1086 * or modifying operations.
1088 * Modifying operations should generally use dsl_sync_task(). The synctask
1089 * infrastructure enforces proper locking strategy with respect to the
1090 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1092 * Read-only operations will manually hold the pool, then the dataset, obtain
1093 * information from the dataset, then release the pool and dataset.
1094 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1099 dsl_pool_hold(const char *name
, void *tag
, dsl_pool_t
**dp
)
1104 error
= spa_open(name
, &spa
, tag
);
1106 *dp
= spa_get_dsl(spa
);
1107 dsl_pool_config_enter(*dp
, tag
);
1113 dsl_pool_rele(dsl_pool_t
*dp
, void *tag
)
1115 dsl_pool_config_exit(dp
, tag
);
1116 spa_close(dp
->dp_spa
, tag
);
1120 dsl_pool_config_enter(dsl_pool_t
*dp
, void *tag
)
1123 * We use a "reentrant" reader-writer lock, but not reentrantly.
1125 * The rrwlock can (with the track_all flag) track all reading threads,
1126 * which is very useful for debugging which code path failed to release
1127 * the lock, and for verifying that the *current* thread does hold
1130 * (Unlike a rwlock, which knows that N threads hold it for
1131 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1132 * if any thread holds it for read, even if this thread doesn't).
1134 ASSERT(!rrw_held(&dp
->dp_config_rwlock
, RW_READER
));
1135 rrw_enter(&dp
->dp_config_rwlock
, RW_READER
, tag
);
1139 dsl_pool_config_enter_prio(dsl_pool_t
*dp
, void *tag
)
1141 ASSERT(!rrw_held(&dp
->dp_config_rwlock
, RW_READER
));
1142 rrw_enter_read_prio(&dp
->dp_config_rwlock
, tag
);
1146 dsl_pool_config_exit(dsl_pool_t
*dp
, void *tag
)
1148 rrw_exit(&dp
->dp_config_rwlock
, tag
);
1152 dsl_pool_config_held(dsl_pool_t
*dp
)
1154 return (RRW_LOCK_HELD(&dp
->dp_config_rwlock
));
1158 dsl_pool_config_held_writer(dsl_pool_t
*dp
)
1160 return (RRW_WRITE_HELD(&dp
->dp_config_rwlock
));
1163 #if defined(_KERNEL) && defined(HAVE_SPL)
1164 EXPORT_SYMBOL(dsl_pool_config_enter
);
1165 EXPORT_SYMBOL(dsl_pool_config_exit
);
1168 /* zfs_dirty_data_max_percent only applied at module load in arc_init(). */
1169 module_param(zfs_dirty_data_max_percent
, int, 0444);
1170 MODULE_PARM_DESC(zfs_dirty_data_max_percent
, "percent of ram can be dirty");
1172 /* zfs_dirty_data_max_max_percent only applied at module load in arc_init(). */
1173 module_param(zfs_dirty_data_max_max_percent
, int, 0444);
1174 MODULE_PARM_DESC(zfs_dirty_data_max_max_percent
,
1175 "zfs_dirty_data_max upper bound as % of RAM");
1177 module_param(zfs_delay_min_dirty_percent
, int, 0644);
1178 MODULE_PARM_DESC(zfs_delay_min_dirty_percent
, "transaction delay threshold");
1180 module_param(zfs_dirty_data_max
, ulong
, 0644);
1181 MODULE_PARM_DESC(zfs_dirty_data_max
, "determines the dirty space limit");
1183 /* zfs_dirty_data_max_max only applied at module load in arc_init(). */
1184 module_param(zfs_dirty_data_max_max
, ulong
, 0444);
1185 MODULE_PARM_DESC(zfs_dirty_data_max_max
,
1186 "zfs_dirty_data_max upper bound in bytes");
1188 module_param(zfs_dirty_data_sync
, ulong
, 0644);
1189 MODULE_PARM_DESC(zfs_dirty_data_sync
, "sync txg when this much dirty data");
1191 module_param(zfs_delay_scale
, ulong
, 0644);
1192 MODULE_PARM_DESC(zfs_delay_scale
, "how quickly delay approaches infinity");
1194 module_param(zfs_sync_taskq_batch_pct
, int, 0644);
1195 MODULE_PARM_DESC(zfs_sync_taskq_batch_pct
,
1196 "max percent of CPUs that are used to sync dirty data");
1198 module_param(zfs_zil_clean_taskq_nthr_pct
, int, 0644);
1199 MODULE_PARM_DESC(zfs_zil_clean_taskq_nthr_pct
,
1200 "max percent of CPUs that are used per dp_sync_taskq");
1202 module_param(zfs_zil_clean_taskq_minalloc
, int, 0644);
1203 MODULE_PARM_DESC(zfs_zil_clean_taskq_minalloc
,
1204 "number of taskq entries that are pre-populated");
1206 module_param(zfs_zil_clean_taskq_maxalloc
, int, 0644);
1207 MODULE_PARM_DESC(zfs_zil_clean_taskq_maxalloc
,
1208 "max number of taskq entries that are cached");