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) 2013 by Delphix. All rights reserved.
24 * Copyright (c) 2013 Steven Hartland. All rights reserved.
27 #include <sys/dsl_pool.h>
28 #include <sys/dsl_dataset.h>
29 #include <sys/dsl_prop.h>
30 #include <sys/dsl_dir.h>
31 #include <sys/dsl_synctask.h>
32 #include <sys/dsl_scan.h>
33 #include <sys/dnode.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/dmu_objset.h>
39 #include <sys/zfs_context.h>
40 #include <sys/fs/zfs.h>
41 #include <sys/zfs_znode.h>
42 #include <sys/spa_impl.h>
43 #include <sys/dsl_deadlist.h>
44 #include <sys/bptree.h>
45 #include <sys/zfeature.h>
46 #include <sys/zil_impl.h>
47 #include <sys/dsl_userhold.h>
53 * ZFS must limit the rate of incoming writes to the rate at which it is able
54 * to sync data modifications to the backend storage. Throttling by too much
55 * creates an artificial limit; throttling by too little can only be sustained
56 * for short periods and would lead to highly lumpy performance. On a per-pool
57 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
58 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
59 * of dirty data decreases. When the amount of dirty data exceeds a
60 * predetermined threshold further modifications are blocked until the amount
61 * of dirty data decreases (as data is synced out).
63 * The limit on dirty data is tunable, and should be adjusted according to
64 * both the IO capacity and available memory of the system. The larger the
65 * window, the more ZFS is able to aggregate and amortize metadata (and data)
66 * changes. However, memory is a limited resource, and allowing for more dirty
67 * data comes at the cost of keeping other useful data in memory (for example
68 * ZFS data cached by the ARC).
72 * As buffers are modified dsl_pool_willuse_space() increments both the per-
73 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
74 * dirty space used; dsl_pool_dirty_space() decrements those values as data
75 * is synced out from dsl_pool_sync(). While only the poolwide value is
76 * relevant, the per-txg value is useful for debugging. The tunable
77 * zfs_dirty_data_max determines the dirty space limit. Once that value is
78 * exceeded, new writes are halted until space frees up.
80 * The zfs_dirty_data_sync tunable dictates the threshold at which we
81 * ensure that there is a txg syncing (see the comment in txg.c for a full
82 * description of transaction group stages).
84 * The IO scheduler uses both the dirty space limit and current amount of
85 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
86 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
88 * The delay is also calculated based on the amount of dirty data. See the
89 * comment above dmu_tx_delay() for details.
93 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
94 * capped at zfs_dirty_data_max_max. It can also be overridden with a module
97 unsigned long zfs_dirty_data_max
= 0;
98 unsigned long zfs_dirty_data_max_max
= 0;
99 int zfs_dirty_data_max_percent
= 10;
100 int zfs_dirty_data_max_max_percent
= 25;
103 * If there is at least this much dirty data, push out a txg.
105 unsigned long zfs_dirty_data_sync
= 64 * 1024 * 1024;
108 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
109 * and delay each transaction.
110 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
112 int zfs_delay_min_dirty_percent
= 60;
115 * This controls how quickly the delay approaches infinity.
116 * Larger values cause it to delay more for a given amount of dirty data.
117 * Therefore larger values will cause there to be less dirty data for a
120 * For the smoothest delay, this value should be about 1 billion divided
121 * by the maximum number of operations per second. This will smoothly
122 * handle between 10x and 1/10th this number.
124 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
125 * multiply in dmu_tx_delay().
127 unsigned long zfs_delay_scale
= 1000 * 1000 * 1000 / 2000;
129 hrtime_t zfs_throttle_delay
= MSEC2NSEC(10);
130 hrtime_t zfs_throttle_resolution
= MSEC2NSEC(10);
133 dsl_pool_open_special_dir(dsl_pool_t
*dp
, const char *name
, dsl_dir_t
**ddp
)
138 err
= zap_lookup(dp
->dp_meta_objset
,
139 dp
->dp_root_dir
->dd_phys
->dd_child_dir_zapobj
,
140 name
, sizeof (obj
), 1, &obj
);
144 return (dsl_dir_hold_obj(dp
, obj
, name
, dp
, ddp
));
148 dsl_pool_open_impl(spa_t
*spa
, uint64_t txg
)
151 blkptr_t
*bp
= spa_get_rootblkptr(spa
);
153 dp
= kmem_zalloc(sizeof (dsl_pool_t
), KM_SLEEP
);
155 dp
->dp_meta_rootbp
= *bp
;
156 rrw_init(&dp
->dp_config_rwlock
, B_TRUE
);
159 txg_list_create(&dp
->dp_dirty_datasets
,
160 offsetof(dsl_dataset_t
, ds_dirty_link
));
161 txg_list_create(&dp
->dp_dirty_zilogs
,
162 offsetof(zilog_t
, zl_dirty_link
));
163 txg_list_create(&dp
->dp_dirty_dirs
,
164 offsetof(dsl_dir_t
, dd_dirty_link
));
165 txg_list_create(&dp
->dp_sync_tasks
,
166 offsetof(dsl_sync_task_t
, dst_node
));
168 mutex_init(&dp
->dp_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
169 cv_init(&dp
->dp_spaceavail_cv
, NULL
, CV_DEFAULT
, NULL
);
171 dp
->dp_iput_taskq
= taskq_create("zfs_iput_taskq", 1, minclsyspri
,
178 dsl_pool_init(spa_t
*spa
, uint64_t txg
, dsl_pool_t
**dpp
)
181 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
183 err
= dmu_objset_open_impl(spa
, NULL
, &dp
->dp_meta_rootbp
,
184 &dp
->dp_meta_objset
);
194 dsl_pool_open(dsl_pool_t
*dp
)
201 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
202 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
203 DMU_POOL_ROOT_DATASET
, sizeof (uint64_t), 1,
204 &dp
->dp_root_dir_obj
);
208 err
= dsl_dir_hold_obj(dp
, dp
->dp_root_dir_obj
,
209 NULL
, dp
, &dp
->dp_root_dir
);
213 err
= dsl_pool_open_special_dir(dp
, MOS_DIR_NAME
, &dp
->dp_mos_dir
);
217 if (spa_version(dp
->dp_spa
) >= SPA_VERSION_ORIGIN
) {
218 err
= dsl_pool_open_special_dir(dp
, ORIGIN_DIR_NAME
, &dd
);
221 err
= dsl_dataset_hold_obj(dp
, dd
->dd_phys
->dd_head_dataset_obj
,
224 err
= dsl_dataset_hold_obj(dp
,
225 ds
->ds_phys
->ds_prev_snap_obj
, dp
,
226 &dp
->dp_origin_snap
);
227 dsl_dataset_rele(ds
, FTAG
);
229 dsl_dir_rele(dd
, dp
);
234 if (spa_version(dp
->dp_spa
) >= SPA_VERSION_DEADLISTS
) {
235 err
= dsl_pool_open_special_dir(dp
, FREE_DIR_NAME
,
240 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
241 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
);
244 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
,
245 dp
->dp_meta_objset
, obj
));
249 * Note: errors ignored, because the leak dir will not exist if we
250 * have not encountered a leak yet.
252 (void) dsl_pool_open_special_dir(dp
, LEAK_DIR_NAME
,
255 if (spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_ASYNC_DESTROY
)) {
256 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
257 DMU_POOL_BPTREE_OBJ
, sizeof (uint64_t), 1,
263 if (spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_EMPTY_BPOBJ
)) {
264 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
265 DMU_POOL_EMPTY_BPOBJ
, sizeof (uint64_t), 1,
266 &dp
->dp_empty_bpobj
);
271 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
272 DMU_POOL_TMP_USERREFS
, sizeof (uint64_t), 1,
273 &dp
->dp_tmp_userrefs_obj
);
279 err
= dsl_scan_init(dp
, dp
->dp_tx
.tx_open_txg
);
282 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
287 dsl_pool_close(dsl_pool_t
*dp
)
290 * Drop our references from dsl_pool_open().
292 * Since we held the origin_snap from "syncing" context (which
293 * includes pool-opening context), it actually only got a "ref"
294 * and not a hold, so just drop that here.
296 if (dp
->dp_origin_snap
)
297 dsl_dataset_rele(dp
->dp_origin_snap
, dp
);
299 dsl_dir_rele(dp
->dp_mos_dir
, dp
);
301 dsl_dir_rele(dp
->dp_free_dir
, dp
);
303 dsl_dir_rele(dp
->dp_leak_dir
, dp
);
305 dsl_dir_rele(dp
->dp_root_dir
, dp
);
307 bpobj_close(&dp
->dp_free_bpobj
);
309 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
310 if (dp
->dp_meta_objset
)
311 dmu_objset_evict(dp
->dp_meta_objset
);
313 txg_list_destroy(&dp
->dp_dirty_datasets
);
314 txg_list_destroy(&dp
->dp_dirty_zilogs
);
315 txg_list_destroy(&dp
->dp_sync_tasks
);
316 txg_list_destroy(&dp
->dp_dirty_dirs
);
318 arc_flush(dp
->dp_spa
);
321 rrw_destroy(&dp
->dp_config_rwlock
);
322 mutex_destroy(&dp
->dp_lock
);
323 taskq_destroy(dp
->dp_iput_taskq
);
325 kmem_free(dp
->dp_blkstats
, sizeof (zfs_all_blkstats_t
));
326 kmem_free(dp
, sizeof (dsl_pool_t
));
330 dsl_pool_create(spa_t
*spa
, nvlist_t
*zplprops
, uint64_t txg
)
333 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
334 dmu_tx_t
*tx
= dmu_tx_create_assigned(dp
, txg
);
339 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
341 /* create and open the MOS (meta-objset) */
342 dp
->dp_meta_objset
= dmu_objset_create_impl(spa
,
343 NULL
, &dp
->dp_meta_rootbp
, DMU_OST_META
, tx
);
345 /* create the pool directory */
346 err
= zap_create_claim(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
347 DMU_OT_OBJECT_DIRECTORY
, DMU_OT_NONE
, 0, tx
);
350 /* Initialize scan structures */
351 VERIFY0(dsl_scan_init(dp
, txg
));
353 /* create and open the root dir */
354 dp
->dp_root_dir_obj
= dsl_dir_create_sync(dp
, NULL
, NULL
, tx
);
355 VERIFY0(dsl_dir_hold_obj(dp
, dp
->dp_root_dir_obj
,
356 NULL
, dp
, &dp
->dp_root_dir
));
358 /* create and open the meta-objset dir */
359 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, MOS_DIR_NAME
, tx
);
360 VERIFY0(dsl_pool_open_special_dir(dp
,
361 MOS_DIR_NAME
, &dp
->dp_mos_dir
));
363 if (spa_version(spa
) >= SPA_VERSION_DEADLISTS
) {
364 /* create and open the free dir */
365 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
,
367 VERIFY0(dsl_pool_open_special_dir(dp
,
368 FREE_DIR_NAME
, &dp
->dp_free_dir
));
370 /* create and open the free_bplist */
371 obj
= bpobj_alloc(dp
->dp_meta_objset
, SPA_MAXBLOCKSIZE
, tx
);
372 VERIFY(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
373 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
) == 0);
374 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
,
375 dp
->dp_meta_objset
, obj
));
378 if (spa_version(spa
) >= SPA_VERSION_DSL_SCRUB
)
379 dsl_pool_create_origin(dp
, tx
);
381 /* create the root dataset */
382 obj
= dsl_dataset_create_sync_dd(dp
->dp_root_dir
, NULL
, 0, tx
);
384 /* create the root objset */
385 VERIFY0(dsl_dataset_hold_obj(dp
, obj
, FTAG
, &ds
));
386 VERIFY(NULL
!= (os
= dmu_objset_create_impl(dp
->dp_spa
, ds
,
387 dsl_dataset_get_blkptr(ds
), DMU_OST_ZFS
, tx
)));
389 zfs_create_fs(os
, kcred
, zplprops
, tx
);
391 dsl_dataset_rele(ds
, FTAG
);
395 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
401 * Account for the meta-objset space in its placeholder dsl_dir.
404 dsl_pool_mos_diduse_space(dsl_pool_t
*dp
,
405 int64_t used
, int64_t comp
, int64_t uncomp
)
407 ASSERT3U(comp
, ==, uncomp
); /* it's all metadata */
408 mutex_enter(&dp
->dp_lock
);
409 dp
->dp_mos_used_delta
+= used
;
410 dp
->dp_mos_compressed_delta
+= comp
;
411 dp
->dp_mos_uncompressed_delta
+= uncomp
;
412 mutex_exit(&dp
->dp_lock
);
416 deadlist_enqueue_cb(void *arg
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
418 dsl_deadlist_t
*dl
= arg
;
419 dsl_deadlist_insert(dl
, bp
, tx
);
424 dsl_pool_sync_mos(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
426 zio_t
*zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
427 dmu_objset_sync(dp
->dp_meta_objset
, zio
, tx
);
428 VERIFY0(zio_wait(zio
));
429 dprintf_bp(&dp
->dp_meta_rootbp
, "meta objset rootbp is %s", "");
430 spa_set_rootblkptr(dp
->dp_spa
, &dp
->dp_meta_rootbp
);
434 dsl_pool_dirty_delta(dsl_pool_t
*dp
, int64_t delta
)
436 ASSERT(MUTEX_HELD(&dp
->dp_lock
));
439 ASSERT3U(-delta
, <=, dp
->dp_dirty_total
);
441 dp
->dp_dirty_total
+= delta
;
444 * Note: we signal even when increasing dp_dirty_total.
445 * This ensures forward progress -- each thread wakes the next waiter.
447 if (dp
->dp_dirty_total
<= zfs_dirty_data_max
)
448 cv_signal(&dp
->dp_spaceavail_cv
);
452 dsl_pool_sync(dsl_pool_t
*dp
, uint64_t txg
)
458 objset_t
*mos
= dp
->dp_meta_objset
;
459 list_t synced_datasets
;
461 list_create(&synced_datasets
, sizeof (dsl_dataset_t
),
462 offsetof(dsl_dataset_t
, ds_synced_link
));
464 tx
= dmu_tx_create_assigned(dp
, txg
);
467 * Write out all dirty blocks of dirty datasets.
469 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
470 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
472 * We must not sync any non-MOS datasets twice, because
473 * we may have taken a snapshot of them. However, we
474 * may sync newly-created datasets on pass 2.
476 ASSERT(!list_link_active(&ds
->ds_synced_link
));
477 list_insert_tail(&synced_datasets
, ds
);
478 dsl_dataset_sync(ds
, zio
, tx
);
480 VERIFY0(zio_wait(zio
));
483 * We have written all of the accounted dirty data, so our
484 * dp_space_towrite should now be zero. However, some seldom-used
485 * code paths do not adhere to this (e.g. dbuf_undirty(), also
486 * rounding error in dbuf_write_physdone).
487 * Shore up the accounting of any dirtied space now.
489 dsl_pool_undirty_space(dp
, dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], txg
);
492 * After the data blocks have been written (ensured by the zio_wait()
493 * above), update the user/group space accounting.
495 for (ds
= list_head(&synced_datasets
); ds
!= NULL
;
496 ds
= list_next(&synced_datasets
, ds
)) {
497 dmu_objset_do_userquota_updates(ds
->ds_objset
, tx
);
501 * Sync the datasets again to push out the changes due to
502 * userspace updates. This must be done before we process the
503 * sync tasks, so that any snapshots will have the correct
504 * user accounting information (and we won't get confused
505 * about which blocks are part of the snapshot).
507 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
508 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
509 ASSERT(list_link_active(&ds
->ds_synced_link
));
510 dmu_buf_rele(ds
->ds_dbuf
, ds
);
511 dsl_dataset_sync(ds
, zio
, tx
);
513 VERIFY0(zio_wait(zio
));
516 * Now that the datasets have been completely synced, we can
517 * clean up our in-memory structures accumulated while syncing:
519 * - move dead blocks from the pending deadlist to the on-disk deadlist
520 * - release hold from dsl_dataset_dirty()
522 while ((ds
= list_remove_head(&synced_datasets
)) != NULL
) {
523 ASSERTV(objset_t
*os
= ds
->ds_objset
);
524 bplist_iterate(&ds
->ds_pending_deadlist
,
525 deadlist_enqueue_cb
, &ds
->ds_deadlist
, tx
);
526 ASSERT(!dmu_objset_is_dirty(os
, txg
));
527 dmu_buf_rele(ds
->ds_dbuf
, ds
);
530 while ((dd
= txg_list_remove(&dp
->dp_dirty_dirs
, txg
)) != NULL
) {
531 dsl_dir_sync(dd
, tx
);
535 * The MOS's space is accounted for in the pool/$MOS
536 * (dp_mos_dir). We can't modify the mos while we're syncing
537 * it, so we remember the deltas and apply them here.
539 if (dp
->dp_mos_used_delta
!= 0 || dp
->dp_mos_compressed_delta
!= 0 ||
540 dp
->dp_mos_uncompressed_delta
!= 0) {
541 dsl_dir_diduse_space(dp
->dp_mos_dir
, DD_USED_HEAD
,
542 dp
->dp_mos_used_delta
,
543 dp
->dp_mos_compressed_delta
,
544 dp
->dp_mos_uncompressed_delta
, tx
);
545 dp
->dp_mos_used_delta
= 0;
546 dp
->dp_mos_compressed_delta
= 0;
547 dp
->dp_mos_uncompressed_delta
= 0;
550 if (list_head(&mos
->os_dirty_dnodes
[txg
& TXG_MASK
]) != NULL
||
551 list_head(&mos
->os_free_dnodes
[txg
& TXG_MASK
]) != NULL
) {
552 dsl_pool_sync_mos(dp
, tx
);
556 * If we modify a dataset in the same txg that we want to destroy it,
557 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
558 * dsl_dir_destroy_check() will fail if there are unexpected holds.
559 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
560 * and clearing the hold on it) before we process the sync_tasks.
561 * The MOS data dirtied by the sync_tasks will be synced on the next
564 if (!txg_list_empty(&dp
->dp_sync_tasks
, txg
)) {
565 dsl_sync_task_t
*dst
;
567 * No more sync tasks should have been added while we
570 ASSERT3U(spa_sync_pass(dp
->dp_spa
), ==, 1);
571 while ((dst
= txg_list_remove(&dp
->dp_sync_tasks
, txg
)) != NULL
)
572 dsl_sync_task_sync(dst
, tx
);
577 DTRACE_PROBE2(dsl_pool_sync__done
, dsl_pool_t
*dp
, dp
, uint64_t, txg
);
581 dsl_pool_sync_done(dsl_pool_t
*dp
, uint64_t txg
)
585 while ((zilog
= txg_list_remove(&dp
->dp_dirty_zilogs
, txg
))) {
586 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
587 zil_clean(zilog
, txg
);
588 ASSERT(!dmu_objset_is_dirty(zilog
->zl_os
, txg
));
589 dmu_buf_rele(ds
->ds_dbuf
, zilog
);
591 ASSERT(!dmu_objset_is_dirty(dp
->dp_meta_objset
, txg
));
595 * TRUE if the current thread is the tx_sync_thread or if we
596 * are being called from SPA context during pool initialization.
599 dsl_pool_sync_context(dsl_pool_t
*dp
)
601 return (curthread
== dp
->dp_tx
.tx_sync_thread
||
602 spa_is_initializing(dp
->dp_spa
));
606 dsl_pool_adjustedsize(dsl_pool_t
*dp
, boolean_t netfree
)
608 uint64_t space
, resv
;
611 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
613 * XXX The intent log is not accounted for, so it must fit
616 * If we're trying to assess whether it's OK to do a free,
617 * cut the reservation in half to allow forward progress
618 * (e.g. make it possible to rm(1) files from a full pool).
620 space
= spa_get_dspace(dp
->dp_spa
);
621 resv
= MAX(space
>> 6, SPA_MINDEVSIZE
>> 1);
625 return (space
- resv
);
629 dsl_pool_need_dirty_delay(dsl_pool_t
*dp
)
631 uint64_t delay_min_bytes
=
632 zfs_dirty_data_max
* zfs_delay_min_dirty_percent
/ 100;
635 mutex_enter(&dp
->dp_lock
);
636 if (dp
->dp_dirty_total
> zfs_dirty_data_sync
)
638 rv
= (dp
->dp_dirty_total
> delay_min_bytes
);
639 mutex_exit(&dp
->dp_lock
);
644 dsl_pool_dirty_space(dsl_pool_t
*dp
, int64_t space
, dmu_tx_t
*tx
)
647 mutex_enter(&dp
->dp_lock
);
648 dp
->dp_dirty_pertxg
[tx
->tx_txg
& TXG_MASK
] += space
;
649 dsl_pool_dirty_delta(dp
, space
);
650 mutex_exit(&dp
->dp_lock
);
655 dsl_pool_undirty_space(dsl_pool_t
*dp
, int64_t space
, uint64_t txg
)
657 ASSERT3S(space
, >=, 0);
661 mutex_enter(&dp
->dp_lock
);
662 if (dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] < space
) {
663 /* XXX writing something we didn't dirty? */
664 space
= dp
->dp_dirty_pertxg
[txg
& TXG_MASK
];
666 ASSERT3U(dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], >=, space
);
667 dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] -= space
;
668 ASSERT3U(dp
->dp_dirty_total
, >=, space
);
669 dsl_pool_dirty_delta(dp
, -space
);
670 mutex_exit(&dp
->dp_lock
);
675 upgrade_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*hds
, void *arg
)
678 dsl_dataset_t
*ds
, *prev
= NULL
;
681 err
= dsl_dataset_hold_obj(dp
, hds
->ds_object
, FTAG
, &ds
);
685 while (ds
->ds_phys
->ds_prev_snap_obj
!= 0) {
686 err
= dsl_dataset_hold_obj(dp
, ds
->ds_phys
->ds_prev_snap_obj
,
689 dsl_dataset_rele(ds
, FTAG
);
693 if (prev
->ds_phys
->ds_next_snap_obj
!= ds
->ds_object
)
695 dsl_dataset_rele(ds
, FTAG
);
701 prev
= dp
->dp_origin_snap
;
704 * The $ORIGIN can't have any data, or the accounting
707 ASSERT0(prev
->ds_phys
->ds_bp
.blk_birth
);
709 /* The origin doesn't get attached to itself */
710 if (ds
->ds_object
== prev
->ds_object
) {
711 dsl_dataset_rele(ds
, FTAG
);
715 dmu_buf_will_dirty(ds
->ds_dbuf
, tx
);
716 ds
->ds_phys
->ds_prev_snap_obj
= prev
->ds_object
;
717 ds
->ds_phys
->ds_prev_snap_txg
= prev
->ds_phys
->ds_creation_txg
;
719 dmu_buf_will_dirty(ds
->ds_dir
->dd_dbuf
, tx
);
720 ds
->ds_dir
->dd_phys
->dd_origin_obj
= prev
->ds_object
;
722 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
723 prev
->ds_phys
->ds_num_children
++;
725 if (ds
->ds_phys
->ds_next_snap_obj
== 0) {
726 ASSERT(ds
->ds_prev
== NULL
);
727 VERIFY0(dsl_dataset_hold_obj(dp
,
728 ds
->ds_phys
->ds_prev_snap_obj
, ds
, &ds
->ds_prev
));
732 ASSERT3U(ds
->ds_dir
->dd_phys
->dd_origin_obj
, ==, prev
->ds_object
);
733 ASSERT3U(ds
->ds_phys
->ds_prev_snap_obj
, ==, prev
->ds_object
);
735 if (prev
->ds_phys
->ds_next_clones_obj
== 0) {
736 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
737 prev
->ds_phys
->ds_next_clones_obj
=
738 zap_create(dp
->dp_meta_objset
,
739 DMU_OT_NEXT_CLONES
, DMU_OT_NONE
, 0, tx
);
741 VERIFY0(zap_add_int(dp
->dp_meta_objset
,
742 prev
->ds_phys
->ds_next_clones_obj
, ds
->ds_object
, tx
));
744 dsl_dataset_rele(ds
, FTAG
);
745 if (prev
!= dp
->dp_origin_snap
)
746 dsl_dataset_rele(prev
, FTAG
);
751 dsl_pool_upgrade_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
753 ASSERT(dmu_tx_is_syncing(tx
));
754 ASSERT(dp
->dp_origin_snap
!= NULL
);
756 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
, upgrade_clones_cb
,
757 tx
, DS_FIND_CHILDREN
));
762 upgrade_dir_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *arg
)
765 objset_t
*mos
= dp
->dp_meta_objset
;
767 if (ds
->ds_dir
->dd_phys
->dd_origin_obj
!= 0) {
768 dsl_dataset_t
*origin
;
770 VERIFY0(dsl_dataset_hold_obj(dp
,
771 ds
->ds_dir
->dd_phys
->dd_origin_obj
, FTAG
, &origin
));
773 if (origin
->ds_dir
->dd_phys
->dd_clones
== 0) {
774 dmu_buf_will_dirty(origin
->ds_dir
->dd_dbuf
, tx
);
775 origin
->ds_dir
->dd_phys
->dd_clones
= zap_create(mos
,
776 DMU_OT_DSL_CLONES
, DMU_OT_NONE
, 0, tx
);
779 VERIFY0(zap_add_int(dp
->dp_meta_objset
,
780 origin
->ds_dir
->dd_phys
->dd_clones
, ds
->ds_object
, tx
));
782 dsl_dataset_rele(origin
, FTAG
);
788 dsl_pool_upgrade_dir_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
792 ASSERT(dmu_tx_is_syncing(tx
));
794 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, FREE_DIR_NAME
, tx
);
795 VERIFY0(dsl_pool_open_special_dir(dp
,
796 FREE_DIR_NAME
, &dp
->dp_free_dir
));
799 * We can't use bpobj_alloc(), because spa_version() still
800 * returns the old version, and we need a new-version bpobj with
801 * subobj support. So call dmu_object_alloc() directly.
803 obj
= dmu_object_alloc(dp
->dp_meta_objset
, DMU_OT_BPOBJ
,
804 SPA_MAXBLOCKSIZE
, DMU_OT_BPOBJ_HDR
, sizeof (bpobj_phys_t
), tx
);
805 VERIFY0(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
806 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
));
807 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
, dp
->dp_meta_objset
, obj
));
809 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
,
810 upgrade_dir_clones_cb
, tx
, DS_FIND_CHILDREN
));
814 dsl_pool_create_origin(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
819 ASSERT(dmu_tx_is_syncing(tx
));
820 ASSERT(dp
->dp_origin_snap
== NULL
);
821 ASSERT(rrw_held(&dp
->dp_config_rwlock
, RW_WRITER
));
823 /* create the origin dir, ds, & snap-ds */
824 dsobj
= dsl_dataset_create_sync(dp
->dp_root_dir
, ORIGIN_DIR_NAME
,
826 VERIFY0(dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
827 dsl_dataset_snapshot_sync_impl(ds
, ORIGIN_DIR_NAME
, tx
);
828 VERIFY0(dsl_dataset_hold_obj(dp
, ds
->ds_phys
->ds_prev_snap_obj
,
829 dp
, &dp
->dp_origin_snap
));
830 dsl_dataset_rele(ds
, FTAG
);
834 dsl_pool_iput_taskq(dsl_pool_t
*dp
)
836 return (dp
->dp_iput_taskq
);
840 * Walk through the pool-wide zap object of temporary snapshot user holds
844 dsl_pool_clean_tmp_userrefs(dsl_pool_t
*dp
)
848 objset_t
*mos
= dp
->dp_meta_objset
;
849 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
854 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
856 holds
= fnvlist_alloc();
858 for (zap_cursor_init(&zc
, mos
, zapobj
);
859 zap_cursor_retrieve(&zc
, &za
) == 0;
860 zap_cursor_advance(&zc
)) {
864 htag
= strchr(za
.za_name
, '-');
867 if (nvlist_lookup_nvlist(holds
, za
.za_name
, &tags
) != 0) {
868 tags
= fnvlist_alloc();
869 fnvlist_add_boolean(tags
, htag
);
870 fnvlist_add_nvlist(holds
, za
.za_name
, tags
);
873 fnvlist_add_boolean(tags
, htag
);
876 dsl_dataset_user_release_tmp(dp
, holds
);
878 zap_cursor_fini(&zc
);
882 * Create the pool-wide zap object for storing temporary snapshot holds.
885 dsl_pool_user_hold_create_obj(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
887 objset_t
*mos
= dp
->dp_meta_objset
;
889 ASSERT(dp
->dp_tmp_userrefs_obj
== 0);
890 ASSERT(dmu_tx_is_syncing(tx
));
892 dp
->dp_tmp_userrefs_obj
= zap_create_link(mos
, DMU_OT_USERREFS
,
893 DMU_POOL_DIRECTORY_OBJECT
, DMU_POOL_TMP_USERREFS
, tx
);
897 dsl_pool_user_hold_rele_impl(dsl_pool_t
*dp
, uint64_t dsobj
,
898 const char *tag
, uint64_t now
, dmu_tx_t
*tx
, boolean_t holding
)
900 objset_t
*mos
= dp
->dp_meta_objset
;
901 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
905 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
906 ASSERT(dmu_tx_is_syncing(tx
));
909 * If the pool was created prior to SPA_VERSION_USERREFS, the
910 * zap object for temporary holds might not exist yet.
914 dsl_pool_user_hold_create_obj(dp
, tx
);
915 zapobj
= dp
->dp_tmp_userrefs_obj
;
917 return (SET_ERROR(ENOENT
));
921 name
= kmem_asprintf("%llx-%s", (u_longlong_t
)dsobj
, tag
);
923 error
= zap_add(mos
, zapobj
, name
, 8, 1, &now
, tx
);
925 error
= zap_remove(mos
, zapobj
, name
, tx
);
932 * Add a temporary hold for the given dataset object and tag.
935 dsl_pool_user_hold(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
936 uint64_t now
, dmu_tx_t
*tx
)
938 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, now
, tx
, B_TRUE
));
942 * Release a temporary hold for the given dataset object and tag.
945 dsl_pool_user_release(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
948 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, 0,
953 * DSL Pool Configuration Lock
955 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
956 * creation / destruction / rename / property setting). It must be held for
957 * read to hold a dataset or dsl_dir. I.e. you must call
958 * dsl_pool_config_enter() or dsl_pool_hold() before calling
959 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
960 * must be held continuously until all datasets and dsl_dirs are released.
962 * The only exception to this rule is that if a "long hold" is placed on
963 * a dataset, then the dp_config_rwlock may be dropped while the dataset
964 * is still held. The long hold will prevent the dataset from being
965 * destroyed -- the destroy will fail with EBUSY. A long hold can be
966 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
967 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
969 * Legitimate long-holders (including owners) should be long-running, cancelable
970 * tasks that should cause "zfs destroy" to fail. This includes DMU
971 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
972 * "zfs send", and "zfs diff". There are several other long-holders whose
973 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
975 * The usual formula for long-holding would be:
978 * ... perform checks ...
979 * dsl_dataset_long_hold()
981 * ... perform long-running task ...
982 * dsl_dataset_long_rele()
985 * Note that when the long hold is released, the dataset is still held but
986 * the pool is not held. The dataset may change arbitrarily during this time
987 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
988 * dataset except release it.
990 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
991 * or modifying operations.
993 * Modifying operations should generally use dsl_sync_task(). The synctask
994 * infrastructure enforces proper locking strategy with respect to the
995 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
997 * Read-only operations will manually hold the pool, then the dataset, obtain
998 * information from the dataset, then release the pool and dataset.
999 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1004 dsl_pool_hold(const char *name
, void *tag
, dsl_pool_t
**dp
)
1009 error
= spa_open(name
, &spa
, tag
);
1011 *dp
= spa_get_dsl(spa
);
1012 dsl_pool_config_enter(*dp
, tag
);
1018 dsl_pool_rele(dsl_pool_t
*dp
, void *tag
)
1020 dsl_pool_config_exit(dp
, tag
);
1021 spa_close(dp
->dp_spa
, tag
);
1025 dsl_pool_config_enter(dsl_pool_t
*dp
, void *tag
)
1028 * We use a "reentrant" reader-writer lock, but not reentrantly.
1030 * The rrwlock can (with the track_all flag) track all reading threads,
1031 * which is very useful for debugging which code path failed to release
1032 * the lock, and for verifying that the *current* thread does hold
1035 * (Unlike a rwlock, which knows that N threads hold it for
1036 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1037 * if any thread holds it for read, even if this thread doesn't).
1039 ASSERT(!rrw_held(&dp
->dp_config_rwlock
, RW_READER
));
1040 rrw_enter(&dp
->dp_config_rwlock
, RW_READER
, tag
);
1044 dsl_pool_config_exit(dsl_pool_t
*dp
, void *tag
)
1046 rrw_exit(&dp
->dp_config_rwlock
, tag
);
1050 dsl_pool_config_held(dsl_pool_t
*dp
)
1052 return (RRW_LOCK_HELD(&dp
->dp_config_rwlock
));
1055 #if defined(_KERNEL) && defined(HAVE_SPL)
1056 EXPORT_SYMBOL(dsl_pool_config_enter
);
1057 EXPORT_SYMBOL(dsl_pool_config_exit
);
1059 /* zfs_dirty_data_max_percent only applied at module load in arc_init(). */
1060 module_param(zfs_dirty_data_max_percent
, int, 0444);
1061 MODULE_PARM_DESC(zfs_dirty_data_max_percent
, "percent of ram can be dirty");
1063 /* zfs_dirty_data_max_max_percent only applied at module load in arc_init(). */
1064 module_param(zfs_dirty_data_max_max_percent
, int, 0444);
1065 MODULE_PARM_DESC(zfs_dirty_data_max_max_percent
,
1066 "zfs_dirty_data_max upper bound as % of RAM");
1068 module_param(zfs_delay_min_dirty_percent
, int, 0644);
1069 MODULE_PARM_DESC(zfs_delay_min_dirty_percent
, "transaction delay threshold");
1071 module_param(zfs_dirty_data_max
, ulong
, 0644);
1072 MODULE_PARM_DESC(zfs_dirty_data_max
, "determines the dirty space limit");
1074 /* zfs_dirty_data_max_max only applied at module load in arc_init(). */
1075 module_param(zfs_dirty_data_max_max
, ulong
, 0444);
1076 MODULE_PARM_DESC(zfs_dirty_data_max_max
,
1077 "zfs_dirty_data_max upper bound in bytes");
1079 module_param(zfs_dirty_data_sync
, ulong
, 0644);
1080 MODULE_PARM_DESC(zfs_dirty_data_sync
, "sync txg when this much dirty data");
1082 module_param(zfs_delay_scale
, ulong
, 0644);
1083 MODULE_PARM_DESC(zfs_delay_scale
, "how quickly delay approaches infinity");