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 (c) 2014 Integros [integros.com]
27 * Copyright 2016 Nexenta Systems, Inc. All rights reserved.
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>
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/bptree.h>
48 #include <sys/zfeature.h>
49 #include <sys/zil_impl.h>
50 #include <sys/dsl_userhold.h>
56 * ZFS must limit the rate of incoming writes to the rate at which it is able
57 * to sync data modifications to the backend storage. Throttling by too much
58 * creates an artificial limit; throttling by too little can only be sustained
59 * for short periods and would lead to highly lumpy performance. On a per-pool
60 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
61 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
62 * of dirty data decreases. When the amount of dirty data exceeds a
63 * predetermined threshold further modifications are blocked until the amount
64 * of dirty data decreases (as data is synced out).
66 * The limit on dirty data is tunable, and should be adjusted according to
67 * both the IO capacity and available memory of the system. The larger the
68 * window, the more ZFS is able to aggregate and amortize metadata (and data)
69 * changes. However, memory is a limited resource, and allowing for more dirty
70 * data comes at the cost of keeping other useful data in memory (for example
71 * ZFS data cached by the ARC).
75 * As buffers are modified dsl_pool_willuse_space() increments both the per-
76 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
77 * dirty space used; dsl_pool_dirty_space() decrements those values as data
78 * is synced out from dsl_pool_sync(). While only the poolwide value is
79 * relevant, the per-txg value is useful for debugging. The tunable
80 * zfs_dirty_data_max determines the dirty space limit. Once that value is
81 * exceeded, new writes are halted until space frees up.
83 * The zfs_dirty_data_sync tunable dictates the threshold at which we
84 * ensure that there is a txg syncing (see the comment in txg.c for a full
85 * description of transaction group stages).
87 * The IO scheduler uses both the dirty space limit and current amount of
88 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
89 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
91 * The delay is also calculated based on the amount of dirty data. See the
92 * comment above dmu_tx_delay() for details.
96 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
97 * capped at zfs_dirty_data_max_max. It can also be overridden in /etc/system.
99 uint64_t zfs_dirty_data_max
;
100 uint64_t zfs_dirty_data_max_max
= 4ULL * 1024 * 1024 * 1024;
101 int zfs_dirty_data_max_percent
= 10;
104 * If there is at least this much dirty data, push out a txg.
106 uint64_t zfs_dirty_data_sync
= 64 * 1024 * 1024;
109 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
110 * and delay each transaction.
111 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
113 int zfs_delay_min_dirty_percent
= 60;
116 * This controls how quickly the delay approaches infinity.
117 * Larger values cause it to delay more for a given amount of dirty data.
118 * Therefore larger values will cause there to be less dirty data for a
121 * For the smoothest delay, this value should be about 1 billion divided
122 * by the maximum number of operations per second. This will smoothly
123 * handle between 10x and 1/10th this number.
125 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
126 * multiply in dmu_tx_delay().
128 uint64_t zfs_delay_scale
= 1000 * 1000 * 1000 / 2000;
131 * This determines the number of threads used by the dp_sync_taskq.
133 int zfs_sync_taskq_batch_pct
= 75;
136 * These tunables determine the behavior of how zil_itxg_clean() is
137 * called via zil_clean() in the context of spa_sync(). When an itxg
138 * list needs to be cleaned, TQ_NOSLEEP will be used when dispatching.
139 * If the dispatch fails, the call to zil_itxg_clean() will occur
140 * synchronously in the context of spa_sync(), which can negatively
141 * impact the performance of spa_sync() (e.g. in the case of the itxg
142 * list having a large number of itxs that needs to be cleaned).
144 * Thus, these tunables can be used to manipulate the behavior of the
145 * taskq used by zil_clean(); they determine the number of taskq entries
146 * that are pre-populated when the taskq is first created (via the
147 * "zfs_zil_clean_taskq_minalloc" tunable) and the maximum number of
148 * taskq entries that are cached after an on-demand allocation (via the
149 * "zfs_zil_clean_taskq_maxalloc").
151 * The idea being, we want to try reasonably hard to ensure there will
152 * already be a taskq entry pre-allocated by the time that it is needed
153 * by zil_clean(). This way, we can avoid the possibility of an
154 * on-demand allocation of a new taskq entry from failing, which would
155 * result in zil_itxg_clean() being called synchronously from zil_clean()
156 * (which can adversely affect performance of spa_sync()).
158 * Additionally, the number of threads used by the taskq can be
159 * configured via the "zfs_zil_clean_taskq_nthr_pct" tunable.
161 int zfs_zil_clean_taskq_nthr_pct
= 100;
162 int zfs_zil_clean_taskq_minalloc
= 1024;
163 int zfs_zil_clean_taskq_maxalloc
= 1024 * 1024;
166 dsl_pool_open_special_dir(dsl_pool_t
*dp
, const char *name
, dsl_dir_t
**ddp
)
171 err
= zap_lookup(dp
->dp_meta_objset
,
172 dsl_dir_phys(dp
->dp_root_dir
)->dd_child_dir_zapobj
,
173 name
, sizeof (obj
), 1, &obj
);
177 return (dsl_dir_hold_obj(dp
, obj
, name
, dp
, ddp
));
181 dsl_pool_open_impl(spa_t
*spa
, uint64_t txg
)
184 blkptr_t
*bp
= spa_get_rootblkptr(spa
);
186 dp
= kmem_zalloc(sizeof (dsl_pool_t
), KM_SLEEP
);
188 dp
->dp_meta_rootbp
= *bp
;
189 rrw_init(&dp
->dp_config_rwlock
, B_TRUE
);
192 txg_list_create(&dp
->dp_dirty_datasets
, spa
,
193 offsetof(dsl_dataset_t
, ds_dirty_link
));
194 txg_list_create(&dp
->dp_dirty_zilogs
, spa
,
195 offsetof(zilog_t
, zl_dirty_link
));
196 txg_list_create(&dp
->dp_dirty_dirs
, spa
,
197 offsetof(dsl_dir_t
, dd_dirty_link
));
198 txg_list_create(&dp
->dp_sync_tasks
, spa
,
199 offsetof(dsl_sync_task_t
, dst_node
));
201 dp
->dp_sync_taskq
= taskq_create("dp_sync_taskq",
202 zfs_sync_taskq_batch_pct
, minclsyspri
, 1, INT_MAX
,
203 TASKQ_THREADS_CPU_PCT
);
205 dp
->dp_zil_clean_taskq
= taskq_create("dp_zil_clean_taskq",
206 zfs_zil_clean_taskq_nthr_pct
, minclsyspri
,
207 zfs_zil_clean_taskq_minalloc
,
208 zfs_zil_clean_taskq_maxalloc
,
209 TASKQ_PREPOPULATE
| TASKQ_THREADS_CPU_PCT
);
211 mutex_init(&dp
->dp_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
212 cv_init(&dp
->dp_spaceavail_cv
, NULL
, CV_DEFAULT
, NULL
);
214 dp
->dp_vnrele_taskq
= taskq_create("zfs_vn_rele_taskq", 1, minclsyspri
,
221 dsl_pool_init(spa_t
*spa
, uint64_t txg
, dsl_pool_t
**dpp
)
224 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
226 err
= dmu_objset_open_impl(spa
, NULL
, &dp
->dp_meta_rootbp
,
227 &dp
->dp_meta_objset
);
237 dsl_pool_open(dsl_pool_t
*dp
)
244 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
245 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
246 DMU_POOL_ROOT_DATASET
, sizeof (uint64_t), 1,
247 &dp
->dp_root_dir_obj
);
251 err
= dsl_dir_hold_obj(dp
, dp
->dp_root_dir_obj
,
252 NULL
, dp
, &dp
->dp_root_dir
);
256 err
= dsl_pool_open_special_dir(dp
, MOS_DIR_NAME
, &dp
->dp_mos_dir
);
260 if (spa_version(dp
->dp_spa
) >= SPA_VERSION_ORIGIN
) {
261 err
= dsl_pool_open_special_dir(dp
, ORIGIN_DIR_NAME
, &dd
);
264 err
= dsl_dataset_hold_obj(dp
,
265 dsl_dir_phys(dd
)->dd_head_dataset_obj
, FTAG
, &ds
);
267 err
= dsl_dataset_hold_obj(dp
,
268 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, dp
,
269 &dp
->dp_origin_snap
);
270 dsl_dataset_rele(ds
, FTAG
);
272 dsl_dir_rele(dd
, dp
);
277 if (spa_version(dp
->dp_spa
) >= SPA_VERSION_DEADLISTS
) {
278 err
= dsl_pool_open_special_dir(dp
, FREE_DIR_NAME
,
283 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
284 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
);
287 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
,
288 dp
->dp_meta_objset
, obj
));
292 * Note: errors ignored, because the leak dir will not exist if we
293 * have not encountered a leak yet.
295 (void) dsl_pool_open_special_dir(dp
, LEAK_DIR_NAME
,
298 if (spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_ASYNC_DESTROY
)) {
299 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
300 DMU_POOL_BPTREE_OBJ
, sizeof (uint64_t), 1,
306 if (spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_EMPTY_BPOBJ
)) {
307 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
308 DMU_POOL_EMPTY_BPOBJ
, sizeof (uint64_t), 1,
309 &dp
->dp_empty_bpobj
);
314 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
315 DMU_POOL_TMP_USERREFS
, sizeof (uint64_t), 1,
316 &dp
->dp_tmp_userrefs_obj
);
322 err
= dsl_scan_init(dp
, dp
->dp_tx
.tx_open_txg
);
325 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
330 dsl_pool_close(dsl_pool_t
*dp
)
333 * Drop our references from dsl_pool_open().
335 * Since we held the origin_snap from "syncing" context (which
336 * includes pool-opening context), it actually only got a "ref"
337 * and not a hold, so just drop that here.
339 if (dp
->dp_origin_snap
)
340 dsl_dataset_rele(dp
->dp_origin_snap
, dp
);
342 dsl_dir_rele(dp
->dp_mos_dir
, dp
);
344 dsl_dir_rele(dp
->dp_free_dir
, dp
);
346 dsl_dir_rele(dp
->dp_leak_dir
, dp
);
348 dsl_dir_rele(dp
->dp_root_dir
, dp
);
350 bpobj_close(&dp
->dp_free_bpobj
);
352 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
353 if (dp
->dp_meta_objset
)
354 dmu_objset_evict(dp
->dp_meta_objset
);
356 txg_list_destroy(&dp
->dp_dirty_datasets
);
357 txg_list_destroy(&dp
->dp_dirty_zilogs
);
358 txg_list_destroy(&dp
->dp_sync_tasks
);
359 txg_list_destroy(&dp
->dp_dirty_dirs
);
361 taskq_destroy(dp
->dp_zil_clean_taskq
);
362 taskq_destroy(dp
->dp_sync_taskq
);
365 * We can't set retry to TRUE since we're explicitly specifying
366 * a spa to flush. This is good enough; any missed buffers for
367 * this spa won't cause trouble, and they'll eventually fall
368 * out of the ARC just like any other unused buffer.
370 arc_flush(dp
->dp_spa
, FALSE
);
374 dmu_buf_user_evict_wait();
376 rrw_destroy(&dp
->dp_config_rwlock
);
377 mutex_destroy(&dp
->dp_lock
);
378 taskq_destroy(dp
->dp_vnrele_taskq
);
380 kmem_free(dp
->dp_blkstats
, sizeof (zfs_all_blkstats_t
));
381 kmem_free(dp
, sizeof (dsl_pool_t
));
385 dsl_pool_create(spa_t
*spa
, nvlist_t
*zplprops
, uint64_t txg
)
388 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
389 dmu_tx_t
*tx
= dmu_tx_create_assigned(dp
, txg
);
394 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
396 /* create and open the MOS (meta-objset) */
397 dp
->dp_meta_objset
= dmu_objset_create_impl(spa
,
398 NULL
, &dp
->dp_meta_rootbp
, DMU_OST_META
, tx
);
400 /* create the pool directory */
401 err
= zap_create_claim(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
402 DMU_OT_OBJECT_DIRECTORY
, DMU_OT_NONE
, 0, tx
);
405 /* Initialize scan structures */
406 VERIFY0(dsl_scan_init(dp
, txg
));
408 /* create and open the root dir */
409 dp
->dp_root_dir_obj
= dsl_dir_create_sync(dp
, NULL
, NULL
, tx
);
410 VERIFY0(dsl_dir_hold_obj(dp
, dp
->dp_root_dir_obj
,
411 NULL
, dp
, &dp
->dp_root_dir
));
413 /* create and open the meta-objset dir */
414 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, MOS_DIR_NAME
, tx
);
415 VERIFY0(dsl_pool_open_special_dir(dp
,
416 MOS_DIR_NAME
, &dp
->dp_mos_dir
));
418 if (spa_version(spa
) >= SPA_VERSION_DEADLISTS
) {
419 /* create and open the free dir */
420 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
,
422 VERIFY0(dsl_pool_open_special_dir(dp
,
423 FREE_DIR_NAME
, &dp
->dp_free_dir
));
425 /* create and open the free_bplist */
426 obj
= bpobj_alloc(dp
->dp_meta_objset
, SPA_OLD_MAXBLOCKSIZE
, tx
);
427 VERIFY(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
428 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
) == 0);
429 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
,
430 dp
->dp_meta_objset
, obj
));
433 if (spa_version(spa
) >= SPA_VERSION_DSL_SCRUB
)
434 dsl_pool_create_origin(dp
, tx
);
436 /* create the root dataset */
437 obj
= dsl_dataset_create_sync_dd(dp
->dp_root_dir
, NULL
, 0, tx
);
439 /* create the root objset */
440 VERIFY0(dsl_dataset_hold_obj(dp
, obj
, FTAG
, &ds
));
441 rrw_enter(&ds
->ds_bp_rwlock
, RW_READER
, FTAG
);
442 os
= dmu_objset_create_impl(dp
->dp_spa
, ds
,
443 dsl_dataset_get_blkptr(ds
), DMU_OST_ZFS
, tx
);
444 rrw_exit(&ds
->ds_bp_rwlock
, FTAG
);
446 zfs_create_fs(os
, kcred
, zplprops
, tx
);
448 dsl_dataset_rele(ds
, FTAG
);
452 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
458 * Account for the meta-objset space in its placeholder dsl_dir.
461 dsl_pool_mos_diduse_space(dsl_pool_t
*dp
,
462 int64_t used
, int64_t comp
, int64_t uncomp
)
464 ASSERT3U(comp
, ==, uncomp
); /* it's all metadata */
465 mutex_enter(&dp
->dp_lock
);
466 dp
->dp_mos_used_delta
+= used
;
467 dp
->dp_mos_compressed_delta
+= comp
;
468 dp
->dp_mos_uncompressed_delta
+= uncomp
;
469 mutex_exit(&dp
->dp_lock
);
473 dsl_pool_sync_mos(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
475 zio_t
*zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
476 dmu_objset_sync(dp
->dp_meta_objset
, zio
, tx
);
477 VERIFY0(zio_wait(zio
));
478 dprintf_bp(&dp
->dp_meta_rootbp
, "meta objset rootbp is %s", "");
479 spa_set_rootblkptr(dp
->dp_spa
, &dp
->dp_meta_rootbp
);
483 dsl_pool_dirty_delta(dsl_pool_t
*dp
, int64_t delta
)
485 ASSERT(MUTEX_HELD(&dp
->dp_lock
));
488 ASSERT3U(-delta
, <=, dp
->dp_dirty_total
);
490 dp
->dp_dirty_total
+= delta
;
493 * Note: we signal even when increasing dp_dirty_total.
494 * This ensures forward progress -- each thread wakes the next waiter.
496 if (dp
->dp_dirty_total
< zfs_dirty_data_max
)
497 cv_signal(&dp
->dp_spaceavail_cv
);
501 dsl_pool_sync(dsl_pool_t
*dp
, uint64_t txg
)
507 objset_t
*mos
= dp
->dp_meta_objset
;
508 list_t synced_datasets
;
510 list_create(&synced_datasets
, sizeof (dsl_dataset_t
),
511 offsetof(dsl_dataset_t
, ds_synced_link
));
513 tx
= dmu_tx_create_assigned(dp
, txg
);
516 * Write out all dirty blocks of dirty datasets.
518 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
519 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
521 * We must not sync any non-MOS datasets twice, because
522 * we may have taken a snapshot of them. However, we
523 * may sync newly-created datasets on pass 2.
525 ASSERT(!list_link_active(&ds
->ds_synced_link
));
526 list_insert_tail(&synced_datasets
, ds
);
527 dsl_dataset_sync(ds
, zio
, tx
);
529 VERIFY0(zio_wait(zio
));
532 * We have written all of the accounted dirty data, so our
533 * dp_space_towrite should now be zero. However, some seldom-used
534 * code paths do not adhere to this (e.g. dbuf_undirty(), also
535 * rounding error in dbuf_write_physdone).
536 * Shore up the accounting of any dirtied space now.
538 dsl_pool_undirty_space(dp
, dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], txg
);
541 * Update the long range free counter after
542 * we're done syncing user data
544 mutex_enter(&dp
->dp_lock
);
545 ASSERT(spa_sync_pass(dp
->dp_spa
) == 1 ||
546 dp
->dp_long_free_dirty_pertxg
[txg
& TXG_MASK
] == 0);
547 dp
->dp_long_free_dirty_pertxg
[txg
& TXG_MASK
] = 0;
548 mutex_exit(&dp
->dp_lock
);
551 * After the data blocks have been written (ensured by the zio_wait()
552 * above), update the user/group space accounting. This happens
553 * in tasks dispatched to dp_sync_taskq, so wait for them before
556 for (ds
= list_head(&synced_datasets
); ds
!= NULL
;
557 ds
= list_next(&synced_datasets
, ds
)) {
558 dmu_objset_do_userquota_updates(ds
->ds_objset
, tx
);
560 taskq_wait(dp
->dp_sync_taskq
);
563 * Sync the datasets again to push out the changes due to
564 * userspace updates. This must be done before we process the
565 * sync tasks, so that any snapshots will have the correct
566 * user accounting information (and we won't get confused
567 * about which blocks are part of the snapshot).
569 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
570 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
571 ASSERT(list_link_active(&ds
->ds_synced_link
));
572 dmu_buf_rele(ds
->ds_dbuf
, ds
);
573 dsl_dataset_sync(ds
, zio
, tx
);
575 VERIFY0(zio_wait(zio
));
578 * Now that the datasets have been completely synced, we can
579 * clean up our in-memory structures accumulated while syncing:
581 * - move dead blocks from the pending deadlist to the on-disk deadlist
582 * - release hold from dsl_dataset_dirty()
584 while ((ds
= list_remove_head(&synced_datasets
)) != NULL
) {
585 dsl_dataset_sync_done(ds
, tx
);
587 while ((dd
= txg_list_remove(&dp
->dp_dirty_dirs
, txg
)) != NULL
) {
588 dsl_dir_sync(dd
, tx
);
592 * The MOS's space is accounted for in the pool/$MOS
593 * (dp_mos_dir). We can't modify the mos while we're syncing
594 * it, so we remember the deltas and apply them here.
596 if (dp
->dp_mos_used_delta
!= 0 || dp
->dp_mos_compressed_delta
!= 0 ||
597 dp
->dp_mos_uncompressed_delta
!= 0) {
598 dsl_dir_diduse_space(dp
->dp_mos_dir
, DD_USED_HEAD
,
599 dp
->dp_mos_used_delta
,
600 dp
->dp_mos_compressed_delta
,
601 dp
->dp_mos_uncompressed_delta
, tx
);
602 dp
->dp_mos_used_delta
= 0;
603 dp
->dp_mos_compressed_delta
= 0;
604 dp
->dp_mos_uncompressed_delta
= 0;
607 if (!multilist_is_empty(mos
->os_dirty_dnodes
[txg
& TXG_MASK
])) {
608 dsl_pool_sync_mos(dp
, tx
);
612 * If we modify a dataset in the same txg that we want to destroy it,
613 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
614 * dsl_dir_destroy_check() will fail if there are unexpected holds.
615 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
616 * and clearing the hold on it) before we process the sync_tasks.
617 * The MOS data dirtied by the sync_tasks will be synced on the next
620 if (!txg_list_empty(&dp
->dp_sync_tasks
, txg
)) {
621 dsl_sync_task_t
*dst
;
623 * No more sync tasks should have been added while we
626 ASSERT3U(spa_sync_pass(dp
->dp_spa
), ==, 1);
627 while ((dst
= txg_list_remove(&dp
->dp_sync_tasks
, txg
)) != NULL
)
628 dsl_sync_task_sync(dst
, tx
);
633 DTRACE_PROBE2(dsl_pool_sync__done
, dsl_pool_t
*dp
, dp
, uint64_t, txg
);
637 dsl_pool_sync_done(dsl_pool_t
*dp
, uint64_t txg
)
641 while (zilog
= txg_list_head(&dp
->dp_dirty_zilogs
, txg
)) {
642 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
644 * We don't remove the zilog from the dp_dirty_zilogs
645 * list until after we've cleaned it. This ensures that
646 * callers of zilog_is_dirty() receive an accurate
647 * answer when they are racing with the spa sync thread.
649 zil_clean(zilog
, txg
);
650 (void) txg_list_remove_this(&dp
->dp_dirty_zilogs
, zilog
, txg
);
651 ASSERT(!dmu_objset_is_dirty(zilog
->zl_os
, txg
));
652 dmu_buf_rele(ds
->ds_dbuf
, zilog
);
654 ASSERT(!dmu_objset_is_dirty(dp
->dp_meta_objset
, txg
));
658 * TRUE if the current thread is the tx_sync_thread or if we
659 * are being called from SPA context during pool initialization.
662 dsl_pool_sync_context(dsl_pool_t
*dp
)
664 return (curthread
== dp
->dp_tx
.tx_sync_thread
||
665 spa_is_initializing(dp
->dp_spa
) ||
666 taskq_member(dp
->dp_sync_taskq
, curthread
));
670 dsl_pool_adjustedsize(dsl_pool_t
*dp
, boolean_t netfree
)
672 uint64_t space
, resv
;
675 * If we're trying to assess whether it's OK to do a free,
676 * cut the reservation in half to allow forward progress
677 * (e.g. make it possible to rm(1) files from a full pool).
679 space
= spa_get_dspace(dp
->dp_spa
);
680 resv
= spa_get_slop_space(dp
->dp_spa
);
684 return (space
- resv
);
688 dsl_pool_need_dirty_delay(dsl_pool_t
*dp
)
690 uint64_t delay_min_bytes
=
691 zfs_dirty_data_max
* zfs_delay_min_dirty_percent
/ 100;
694 mutex_enter(&dp
->dp_lock
);
695 if (dp
->dp_dirty_total
> zfs_dirty_data_sync
)
697 rv
= (dp
->dp_dirty_total
> delay_min_bytes
);
698 mutex_exit(&dp
->dp_lock
);
703 dsl_pool_dirty_space(dsl_pool_t
*dp
, int64_t space
, dmu_tx_t
*tx
)
706 mutex_enter(&dp
->dp_lock
);
707 dp
->dp_dirty_pertxg
[tx
->tx_txg
& TXG_MASK
] += space
;
708 dsl_pool_dirty_delta(dp
, space
);
709 mutex_exit(&dp
->dp_lock
);
714 dsl_pool_undirty_space(dsl_pool_t
*dp
, int64_t space
, uint64_t txg
)
716 ASSERT3S(space
, >=, 0);
719 mutex_enter(&dp
->dp_lock
);
720 if (dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] < space
) {
721 /* XXX writing something we didn't dirty? */
722 space
= dp
->dp_dirty_pertxg
[txg
& TXG_MASK
];
724 ASSERT3U(dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], >=, space
);
725 dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] -= space
;
726 ASSERT3U(dp
->dp_dirty_total
, >=, space
);
727 dsl_pool_dirty_delta(dp
, -space
);
728 mutex_exit(&dp
->dp_lock
);
733 upgrade_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*hds
, void *arg
)
736 dsl_dataset_t
*ds
, *prev
= NULL
;
739 err
= dsl_dataset_hold_obj(dp
, hds
->ds_object
, FTAG
, &ds
);
743 while (dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= 0) {
744 err
= dsl_dataset_hold_obj(dp
,
745 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, FTAG
, &prev
);
747 dsl_dataset_rele(ds
, FTAG
);
751 if (dsl_dataset_phys(prev
)->ds_next_snap_obj
!= ds
->ds_object
)
753 dsl_dataset_rele(ds
, FTAG
);
759 prev
= dp
->dp_origin_snap
;
762 * The $ORIGIN can't have any data, or the accounting
765 rrw_enter(&ds
->ds_bp_rwlock
, RW_READER
, FTAG
);
766 ASSERT0(dsl_dataset_phys(prev
)->ds_bp
.blk_birth
);
767 rrw_exit(&ds
->ds_bp_rwlock
, FTAG
);
769 /* The origin doesn't get attached to itself */
770 if (ds
->ds_object
== prev
->ds_object
) {
771 dsl_dataset_rele(ds
, FTAG
);
775 dmu_buf_will_dirty(ds
->ds_dbuf
, tx
);
776 dsl_dataset_phys(ds
)->ds_prev_snap_obj
= prev
->ds_object
;
777 dsl_dataset_phys(ds
)->ds_prev_snap_txg
=
778 dsl_dataset_phys(prev
)->ds_creation_txg
;
780 dmu_buf_will_dirty(ds
->ds_dir
->dd_dbuf
, tx
);
781 dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
= prev
->ds_object
;
783 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
784 dsl_dataset_phys(prev
)->ds_num_children
++;
786 if (dsl_dataset_phys(ds
)->ds_next_snap_obj
== 0) {
787 ASSERT(ds
->ds_prev
== NULL
);
788 VERIFY0(dsl_dataset_hold_obj(dp
,
789 dsl_dataset_phys(ds
)->ds_prev_snap_obj
,
794 ASSERT3U(dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
, ==, prev
->ds_object
);
795 ASSERT3U(dsl_dataset_phys(ds
)->ds_prev_snap_obj
, ==, prev
->ds_object
);
797 if (dsl_dataset_phys(prev
)->ds_next_clones_obj
== 0) {
798 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
799 dsl_dataset_phys(prev
)->ds_next_clones_obj
=
800 zap_create(dp
->dp_meta_objset
,
801 DMU_OT_NEXT_CLONES
, DMU_OT_NONE
, 0, tx
);
803 VERIFY0(zap_add_int(dp
->dp_meta_objset
,
804 dsl_dataset_phys(prev
)->ds_next_clones_obj
, ds
->ds_object
, tx
));
806 dsl_dataset_rele(ds
, FTAG
);
807 if (prev
!= dp
->dp_origin_snap
)
808 dsl_dataset_rele(prev
, FTAG
);
813 dsl_pool_upgrade_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
815 ASSERT(dmu_tx_is_syncing(tx
));
816 ASSERT(dp
->dp_origin_snap
!= NULL
);
818 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
, upgrade_clones_cb
,
819 tx
, DS_FIND_CHILDREN
| DS_FIND_SERIALIZE
));
824 upgrade_dir_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *arg
)
827 objset_t
*mos
= dp
->dp_meta_objset
;
829 if (dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
!= 0) {
830 dsl_dataset_t
*origin
;
832 VERIFY0(dsl_dataset_hold_obj(dp
,
833 dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
, FTAG
, &origin
));
835 if (dsl_dir_phys(origin
->ds_dir
)->dd_clones
== 0) {
836 dmu_buf_will_dirty(origin
->ds_dir
->dd_dbuf
, tx
);
837 dsl_dir_phys(origin
->ds_dir
)->dd_clones
=
838 zap_create(mos
, DMU_OT_DSL_CLONES
, DMU_OT_NONE
,
842 VERIFY0(zap_add_int(dp
->dp_meta_objset
,
843 dsl_dir_phys(origin
->ds_dir
)->dd_clones
,
846 dsl_dataset_rele(origin
, FTAG
);
852 dsl_pool_upgrade_dir_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
854 ASSERT(dmu_tx_is_syncing(tx
));
857 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, FREE_DIR_NAME
, tx
);
858 VERIFY0(dsl_pool_open_special_dir(dp
,
859 FREE_DIR_NAME
, &dp
->dp_free_dir
));
862 * We can't use bpobj_alloc(), because spa_version() still
863 * returns the old version, and we need a new-version bpobj with
864 * subobj support. So call dmu_object_alloc() directly.
866 obj
= dmu_object_alloc(dp
->dp_meta_objset
, DMU_OT_BPOBJ
,
867 SPA_OLD_MAXBLOCKSIZE
, DMU_OT_BPOBJ_HDR
, sizeof (bpobj_phys_t
), tx
);
868 VERIFY0(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
869 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
));
870 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
, dp
->dp_meta_objset
, obj
));
872 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
,
873 upgrade_dir_clones_cb
, tx
, DS_FIND_CHILDREN
| DS_FIND_SERIALIZE
));
877 dsl_pool_create_origin(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
882 ASSERT(dmu_tx_is_syncing(tx
));
883 ASSERT(dp
->dp_origin_snap
== NULL
);
884 ASSERT(rrw_held(&dp
->dp_config_rwlock
, RW_WRITER
));
886 /* create the origin dir, ds, & snap-ds */
887 dsobj
= dsl_dataset_create_sync(dp
->dp_root_dir
, ORIGIN_DIR_NAME
,
889 VERIFY0(dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
890 dsl_dataset_snapshot_sync_impl(ds
, ORIGIN_DIR_NAME
, tx
);
891 VERIFY0(dsl_dataset_hold_obj(dp
, dsl_dataset_phys(ds
)->ds_prev_snap_obj
,
892 dp
, &dp
->dp_origin_snap
));
893 dsl_dataset_rele(ds
, FTAG
);
897 dsl_pool_vnrele_taskq(dsl_pool_t
*dp
)
899 return (dp
->dp_vnrele_taskq
);
903 * Walk through the pool-wide zap object of temporary snapshot user holds
907 dsl_pool_clean_tmp_userrefs(dsl_pool_t
*dp
)
911 objset_t
*mos
= dp
->dp_meta_objset
;
912 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
917 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
919 holds
= fnvlist_alloc();
921 for (zap_cursor_init(&zc
, mos
, zapobj
);
922 zap_cursor_retrieve(&zc
, &za
) == 0;
923 zap_cursor_advance(&zc
)) {
927 htag
= strchr(za
.za_name
, '-');
930 if (nvlist_lookup_nvlist(holds
, za
.za_name
, &tags
) != 0) {
931 tags
= fnvlist_alloc();
932 fnvlist_add_boolean(tags
, htag
);
933 fnvlist_add_nvlist(holds
, za
.za_name
, tags
);
936 fnvlist_add_boolean(tags
, htag
);
939 dsl_dataset_user_release_tmp(dp
, holds
);
941 zap_cursor_fini(&zc
);
945 * Create the pool-wide zap object for storing temporary snapshot holds.
948 dsl_pool_user_hold_create_obj(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
950 objset_t
*mos
= dp
->dp_meta_objset
;
952 ASSERT(dp
->dp_tmp_userrefs_obj
== 0);
953 ASSERT(dmu_tx_is_syncing(tx
));
955 dp
->dp_tmp_userrefs_obj
= zap_create_link(mos
, DMU_OT_USERREFS
,
956 DMU_POOL_DIRECTORY_OBJECT
, DMU_POOL_TMP_USERREFS
, tx
);
960 dsl_pool_user_hold_rele_impl(dsl_pool_t
*dp
, uint64_t dsobj
,
961 const char *tag
, uint64_t now
, dmu_tx_t
*tx
, boolean_t holding
)
963 objset_t
*mos
= dp
->dp_meta_objset
;
964 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
968 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
969 ASSERT(dmu_tx_is_syncing(tx
));
972 * If the pool was created prior to SPA_VERSION_USERREFS, the
973 * zap object for temporary holds might not exist yet.
977 dsl_pool_user_hold_create_obj(dp
, tx
);
978 zapobj
= dp
->dp_tmp_userrefs_obj
;
980 return (SET_ERROR(ENOENT
));
984 name
= kmem_asprintf("%llx-%s", (u_longlong_t
)dsobj
, tag
);
986 error
= zap_add(mos
, zapobj
, name
, 8, 1, &now
, tx
);
988 error
= zap_remove(mos
, zapobj
, name
, tx
);
995 * Add a temporary hold for the given dataset object and tag.
998 dsl_pool_user_hold(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
999 uint64_t now
, dmu_tx_t
*tx
)
1001 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, now
, tx
, B_TRUE
));
1005 * Release a temporary hold for the given dataset object and tag.
1008 dsl_pool_user_release(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
1011 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, (uintptr_t)NULL
,
1016 * DSL Pool Configuration Lock
1018 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
1019 * creation / destruction / rename / property setting). It must be held for
1020 * read to hold a dataset or dsl_dir. I.e. you must call
1021 * dsl_pool_config_enter() or dsl_pool_hold() before calling
1022 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
1023 * must be held continuously until all datasets and dsl_dirs are released.
1025 * The only exception to this rule is that if a "long hold" is placed on
1026 * a dataset, then the dp_config_rwlock may be dropped while the dataset
1027 * is still held. The long hold will prevent the dataset from being
1028 * destroyed -- the destroy will fail with EBUSY. A long hold can be
1029 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
1030 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
1032 * Legitimate long-holders (including owners) should be long-running, cancelable
1033 * tasks that should cause "zfs destroy" to fail. This includes DMU
1034 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
1035 * "zfs send", and "zfs diff". There are several other long-holders whose
1036 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
1038 * The usual formula for long-holding would be:
1040 * dsl_dataset_hold()
1041 * ... perform checks ...
1042 * dsl_dataset_long_hold()
1044 * ... perform long-running task ...
1045 * dsl_dataset_long_rele()
1046 * dsl_dataset_rele()
1048 * Note that when the long hold is released, the dataset is still held but
1049 * the pool is not held. The dataset may change arbitrarily during this time
1050 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1051 * dataset except release it.
1053 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1054 * or modifying operations.
1056 * Modifying operations should generally use dsl_sync_task(). The synctask
1057 * infrastructure enforces proper locking strategy with respect to the
1058 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1060 * Read-only operations will manually hold the pool, then the dataset, obtain
1061 * information from the dataset, then release the pool and dataset.
1062 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1067 dsl_pool_hold(const char *name
, void *tag
, dsl_pool_t
**dp
)
1072 error
= spa_open(name
, &spa
, tag
);
1074 *dp
= spa_get_dsl(spa
);
1075 dsl_pool_config_enter(*dp
, tag
);
1081 dsl_pool_rele(dsl_pool_t
*dp
, void *tag
)
1083 dsl_pool_config_exit(dp
, tag
);
1084 spa_close(dp
->dp_spa
, tag
);
1088 dsl_pool_config_enter(dsl_pool_t
*dp
, void *tag
)
1091 * We use a "reentrant" reader-writer lock, but not reentrantly.
1093 * The rrwlock can (with the track_all flag) track all reading threads,
1094 * which is very useful for debugging which code path failed to release
1095 * the lock, and for verifying that the *current* thread does hold
1098 * (Unlike a rwlock, which knows that N threads hold it for
1099 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1100 * if any thread holds it for read, even if this thread doesn't).
1102 ASSERT(!rrw_held(&dp
->dp_config_rwlock
, RW_READER
));
1103 rrw_enter(&dp
->dp_config_rwlock
, RW_READER
, tag
);
1107 dsl_pool_config_enter_prio(dsl_pool_t
*dp
, void *tag
)
1109 ASSERT(!rrw_held(&dp
->dp_config_rwlock
, RW_READER
));
1110 rrw_enter_read_prio(&dp
->dp_config_rwlock
, tag
);
1114 dsl_pool_config_exit(dsl_pool_t
*dp
, void *tag
)
1116 rrw_exit(&dp
->dp_config_rwlock
, tag
);
1120 dsl_pool_config_held(dsl_pool_t
*dp
)
1122 return (RRW_LOCK_HELD(&dp
->dp_config_rwlock
));
1126 dsl_pool_config_held_writer(dsl_pool_t
*dp
)
1128 return (RRW_WRITE_HELD(&dp
->dp_config_rwlock
));