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, 2014 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>
48 #include <sys/trace_txg.h>
54 * ZFS must limit the rate of incoming writes to the rate at which it is able
55 * to sync data modifications to the backend storage. Throttling by too much
56 * creates an artificial limit; throttling by too little can only be sustained
57 * for short periods and would lead to highly lumpy performance. On a per-pool
58 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
59 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
60 * of dirty data decreases. When the amount of dirty data exceeds a
61 * predetermined threshold further modifications are blocked until the amount
62 * of dirty data decreases (as data is synced out).
64 * The limit on dirty data is tunable, and should be adjusted according to
65 * both the IO capacity and available memory of the system. The larger the
66 * window, the more ZFS is able to aggregate and amortize metadata (and data)
67 * changes. However, memory is a limited resource, and allowing for more dirty
68 * data comes at the cost of keeping other useful data in memory (for example
69 * ZFS data cached by the ARC).
73 * As buffers are modified dsl_pool_willuse_space() increments both the per-
74 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
75 * dirty space used; dsl_pool_dirty_space() decrements those values as data
76 * is synced out from dsl_pool_sync(). While only the poolwide value is
77 * relevant, the per-txg value is useful for debugging. The tunable
78 * zfs_dirty_data_max determines the dirty space limit. Once that value is
79 * exceeded, new writes are halted until space frees up.
81 * The zfs_dirty_data_sync tunable dictates the threshold at which we
82 * ensure that there is a txg syncing (see the comment in txg.c for a full
83 * description of transaction group stages).
85 * The IO scheduler uses both the dirty space limit and current amount of
86 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
87 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
89 * The delay is also calculated based on the amount of dirty data. See the
90 * comment above dmu_tx_delay() for details.
94 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
95 * capped at zfs_dirty_data_max_max. It can also be overridden with a module
98 unsigned long zfs_dirty_data_max
= 0;
99 unsigned long zfs_dirty_data_max_max
= 0;
100 int zfs_dirty_data_max_percent
= 10;
101 int zfs_dirty_data_max_max_percent
= 25;
104 * If there is at least this much dirty data, push out a txg.
106 unsigned long 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 unsigned long zfs_delay_scale
= 1000 * 1000 * 1000 / 2000;
130 hrtime_t zfs_throttle_delay
= MSEC2NSEC(10);
131 hrtime_t zfs_throttle_resolution
= MSEC2NSEC(10);
134 dsl_pool_open_special_dir(dsl_pool_t
*dp
, const char *name
, dsl_dir_t
**ddp
)
139 err
= zap_lookup(dp
->dp_meta_objset
,
140 dsl_dir_phys(dp
->dp_root_dir
)->dd_child_dir_zapobj
,
141 name
, sizeof (obj
), 1, &obj
);
145 return (dsl_dir_hold_obj(dp
, obj
, name
, dp
, ddp
));
149 dsl_pool_open_impl(spa_t
*spa
, uint64_t txg
)
152 blkptr_t
*bp
= spa_get_rootblkptr(spa
);
154 dp
= kmem_zalloc(sizeof (dsl_pool_t
), KM_SLEEP
);
156 dp
->dp_meta_rootbp
= *bp
;
157 rrw_init(&dp
->dp_config_rwlock
, B_TRUE
);
160 txg_list_create(&dp
->dp_dirty_datasets
,
161 offsetof(dsl_dataset_t
, ds_dirty_link
));
162 txg_list_create(&dp
->dp_dirty_zilogs
,
163 offsetof(zilog_t
, zl_dirty_link
));
164 txg_list_create(&dp
->dp_dirty_dirs
,
165 offsetof(dsl_dir_t
, dd_dirty_link
));
166 txg_list_create(&dp
->dp_sync_tasks
,
167 offsetof(dsl_sync_task_t
, dst_node
));
169 mutex_init(&dp
->dp_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
170 cv_init(&dp
->dp_spaceavail_cv
, NULL
, CV_DEFAULT
, NULL
);
172 dp
->dp_iput_taskq
= taskq_create("zfs_iput_taskq", 1, minclsyspri
,
179 dsl_pool_init(spa_t
*spa
, uint64_t txg
, dsl_pool_t
**dpp
)
182 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
184 err
= dmu_objset_open_impl(spa
, NULL
, &dp
->dp_meta_rootbp
,
185 &dp
->dp_meta_objset
);
195 dsl_pool_open(dsl_pool_t
*dp
)
202 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
203 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
204 DMU_POOL_ROOT_DATASET
, sizeof (uint64_t), 1,
205 &dp
->dp_root_dir_obj
);
209 err
= dsl_dir_hold_obj(dp
, dp
->dp_root_dir_obj
,
210 NULL
, dp
, &dp
->dp_root_dir
);
214 err
= dsl_pool_open_special_dir(dp
, MOS_DIR_NAME
, &dp
->dp_mos_dir
);
218 if (spa_version(dp
->dp_spa
) >= SPA_VERSION_ORIGIN
) {
219 err
= dsl_pool_open_special_dir(dp
, ORIGIN_DIR_NAME
, &dd
);
222 err
= dsl_dataset_hold_obj(dp
,
223 dsl_dir_phys(dd
)->dd_head_dataset_obj
, FTAG
, &ds
);
225 err
= dsl_dataset_hold_obj(dp
,
226 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, dp
,
227 &dp
->dp_origin_snap
);
228 dsl_dataset_rele(ds
, FTAG
);
230 dsl_dir_rele(dd
, dp
);
235 if (spa_version(dp
->dp_spa
) >= SPA_VERSION_DEADLISTS
) {
236 err
= dsl_pool_open_special_dir(dp
, FREE_DIR_NAME
,
241 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
242 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
);
245 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
,
246 dp
->dp_meta_objset
, obj
));
250 * Note: errors ignored, because the leak dir will not exist if we
251 * have not encountered a leak yet.
253 (void) dsl_pool_open_special_dir(dp
, LEAK_DIR_NAME
,
256 if (spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_ASYNC_DESTROY
)) {
257 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
258 DMU_POOL_BPTREE_OBJ
, sizeof (uint64_t), 1,
264 if (spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_EMPTY_BPOBJ
)) {
265 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
266 DMU_POOL_EMPTY_BPOBJ
, sizeof (uint64_t), 1,
267 &dp
->dp_empty_bpobj
);
272 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
273 DMU_POOL_TMP_USERREFS
, sizeof (uint64_t), 1,
274 &dp
->dp_tmp_userrefs_obj
);
280 err
= dsl_scan_init(dp
, dp
->dp_tx
.tx_open_txg
);
283 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
288 dsl_pool_close(dsl_pool_t
*dp
)
291 * Drop our references from dsl_pool_open().
293 * Since we held the origin_snap from "syncing" context (which
294 * includes pool-opening context), it actually only got a "ref"
295 * and not a hold, so just drop that here.
297 if (dp
->dp_origin_snap
)
298 dsl_dataset_rele(dp
->dp_origin_snap
, dp
);
300 dsl_dir_rele(dp
->dp_mos_dir
, dp
);
302 dsl_dir_rele(dp
->dp_free_dir
, dp
);
304 dsl_dir_rele(dp
->dp_leak_dir
, dp
);
306 dsl_dir_rele(dp
->dp_root_dir
, dp
);
308 bpobj_close(&dp
->dp_free_bpobj
);
310 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
311 if (dp
->dp_meta_objset
)
312 dmu_objset_evict(dp
->dp_meta_objset
);
314 txg_list_destroy(&dp
->dp_dirty_datasets
);
315 txg_list_destroy(&dp
->dp_dirty_zilogs
);
316 txg_list_destroy(&dp
->dp_sync_tasks
);
317 txg_list_destroy(&dp
->dp_dirty_dirs
);
319 arc_flush(dp
->dp_spa
);
322 rrw_destroy(&dp
->dp_config_rwlock
);
323 mutex_destroy(&dp
->dp_lock
);
324 taskq_destroy(dp
->dp_iput_taskq
);
326 vmem_free(dp
->dp_blkstats
, sizeof (zfs_all_blkstats_t
));
327 kmem_free(dp
, sizeof (dsl_pool_t
));
331 dsl_pool_create(spa_t
*spa
, nvlist_t
*zplprops
, uint64_t txg
)
334 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
335 dmu_tx_t
*tx
= dmu_tx_create_assigned(dp
, txg
);
340 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
342 /* create and open the MOS (meta-objset) */
343 dp
->dp_meta_objset
= dmu_objset_create_impl(spa
,
344 NULL
, &dp
->dp_meta_rootbp
, DMU_OST_META
, tx
);
346 /* create the pool directory */
347 err
= zap_create_claim(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
348 DMU_OT_OBJECT_DIRECTORY
, DMU_OT_NONE
, 0, tx
);
351 /* Initialize scan structures */
352 VERIFY0(dsl_scan_init(dp
, txg
));
354 /* create and open the root dir */
355 dp
->dp_root_dir_obj
= dsl_dir_create_sync(dp
, NULL
, NULL
, tx
);
356 VERIFY0(dsl_dir_hold_obj(dp
, dp
->dp_root_dir_obj
,
357 NULL
, dp
, &dp
->dp_root_dir
));
359 /* create and open the meta-objset dir */
360 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, MOS_DIR_NAME
, tx
);
361 VERIFY0(dsl_pool_open_special_dir(dp
,
362 MOS_DIR_NAME
, &dp
->dp_mos_dir
));
364 if (spa_version(spa
) >= SPA_VERSION_DEADLISTS
) {
365 /* create and open the free dir */
366 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
,
368 VERIFY0(dsl_pool_open_special_dir(dp
,
369 FREE_DIR_NAME
, &dp
->dp_free_dir
));
371 /* create and open the free_bplist */
372 obj
= bpobj_alloc(dp
->dp_meta_objset
, SPA_MAXBLOCKSIZE
, tx
);
373 VERIFY(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
374 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
) == 0);
375 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
,
376 dp
->dp_meta_objset
, obj
));
379 if (spa_version(spa
) >= SPA_VERSION_DSL_SCRUB
)
380 dsl_pool_create_origin(dp
, tx
);
382 /* create the root dataset */
383 obj
= dsl_dataset_create_sync_dd(dp
->dp_root_dir
, NULL
, 0, tx
);
385 /* create the root objset */
386 VERIFY0(dsl_dataset_hold_obj(dp
, obj
, FTAG
, &ds
));
387 VERIFY(NULL
!= (os
= dmu_objset_create_impl(dp
->dp_spa
, ds
,
388 dsl_dataset_get_blkptr(ds
), DMU_OST_ZFS
, tx
)));
390 zfs_create_fs(os
, kcred
, zplprops
, tx
);
392 dsl_dataset_rele(ds
, FTAG
);
396 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
402 * Account for the meta-objset space in its placeholder dsl_dir.
405 dsl_pool_mos_diduse_space(dsl_pool_t
*dp
,
406 int64_t used
, int64_t comp
, int64_t uncomp
)
408 ASSERT3U(comp
, ==, uncomp
); /* it's all metadata */
409 mutex_enter(&dp
->dp_lock
);
410 dp
->dp_mos_used_delta
+= used
;
411 dp
->dp_mos_compressed_delta
+= comp
;
412 dp
->dp_mos_uncompressed_delta
+= uncomp
;
413 mutex_exit(&dp
->dp_lock
);
417 deadlist_enqueue_cb(void *arg
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
419 dsl_deadlist_t
*dl
= arg
;
420 dsl_deadlist_insert(dl
, bp
, tx
);
425 dsl_pool_sync_mos(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
427 zio_t
*zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
428 dmu_objset_sync(dp
->dp_meta_objset
, zio
, tx
);
429 VERIFY0(zio_wait(zio
));
430 dprintf_bp(&dp
->dp_meta_rootbp
, "meta objset rootbp is %s", "");
431 spa_set_rootblkptr(dp
->dp_spa
, &dp
->dp_meta_rootbp
);
435 dsl_pool_dirty_delta(dsl_pool_t
*dp
, int64_t delta
)
437 ASSERT(MUTEX_HELD(&dp
->dp_lock
));
440 ASSERT3U(-delta
, <=, dp
->dp_dirty_total
);
442 dp
->dp_dirty_total
+= delta
;
445 * Note: we signal even when increasing dp_dirty_total.
446 * This ensures forward progress -- each thread wakes the next waiter.
448 if (dp
->dp_dirty_total
<= zfs_dirty_data_max
)
449 cv_signal(&dp
->dp_spaceavail_cv
);
453 dsl_pool_sync(dsl_pool_t
*dp
, uint64_t txg
)
459 objset_t
*mos
= dp
->dp_meta_objset
;
460 list_t synced_datasets
;
462 list_create(&synced_datasets
, sizeof (dsl_dataset_t
),
463 offsetof(dsl_dataset_t
, ds_synced_link
));
465 tx
= dmu_tx_create_assigned(dp
, txg
);
468 * Write out all dirty blocks of dirty datasets.
470 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
471 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
473 * We must not sync any non-MOS datasets twice, because
474 * we may have taken a snapshot of them. However, we
475 * may sync newly-created datasets on pass 2.
477 ASSERT(!list_link_active(&ds
->ds_synced_link
));
478 list_insert_tail(&synced_datasets
, ds
);
479 dsl_dataset_sync(ds
, zio
, tx
);
481 VERIFY0(zio_wait(zio
));
484 * We have written all of the accounted dirty data, so our
485 * dp_space_towrite should now be zero. However, some seldom-used
486 * code paths do not adhere to this (e.g. dbuf_undirty(), also
487 * rounding error in dbuf_write_physdone).
488 * Shore up the accounting of any dirtied space now.
490 dsl_pool_undirty_space(dp
, dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], txg
);
493 * After the data blocks have been written (ensured by the zio_wait()
494 * above), update the user/group space accounting.
496 for (ds
= list_head(&synced_datasets
); ds
!= NULL
;
497 ds
= list_next(&synced_datasets
, ds
)) {
498 dmu_objset_do_userquota_updates(ds
->ds_objset
, tx
);
502 * Sync the datasets again to push out the changes due to
503 * userspace updates. This must be done before we process the
504 * sync tasks, so that any snapshots will have the correct
505 * user accounting information (and we won't get confused
506 * about which blocks are part of the snapshot).
508 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
509 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
510 ASSERT(list_link_active(&ds
->ds_synced_link
));
511 dmu_buf_rele(ds
->ds_dbuf
, ds
);
512 dsl_dataset_sync(ds
, zio
, tx
);
514 VERIFY0(zio_wait(zio
));
517 * Now that the datasets have been completely synced, we can
518 * clean up our in-memory structures accumulated while syncing:
520 * - move dead blocks from the pending deadlist to the on-disk deadlist
521 * - release hold from dsl_dataset_dirty()
523 while ((ds
= list_remove_head(&synced_datasets
)) != NULL
) {
524 ASSERTV(objset_t
*os
= ds
->ds_objset
);
525 bplist_iterate(&ds
->ds_pending_deadlist
,
526 deadlist_enqueue_cb
, &ds
->ds_deadlist
, tx
);
527 ASSERT(!dmu_objset_is_dirty(os
, txg
));
528 dmu_buf_rele(ds
->ds_dbuf
, ds
);
531 while ((dd
= txg_list_remove(&dp
->dp_dirty_dirs
, txg
)) != NULL
) {
532 dsl_dir_sync(dd
, tx
);
536 * The MOS's space is accounted for in the pool/$MOS
537 * (dp_mos_dir). We can't modify the mos while we're syncing
538 * it, so we remember the deltas and apply them here.
540 if (dp
->dp_mos_used_delta
!= 0 || dp
->dp_mos_compressed_delta
!= 0 ||
541 dp
->dp_mos_uncompressed_delta
!= 0) {
542 dsl_dir_diduse_space(dp
->dp_mos_dir
, DD_USED_HEAD
,
543 dp
->dp_mos_used_delta
,
544 dp
->dp_mos_compressed_delta
,
545 dp
->dp_mos_uncompressed_delta
, tx
);
546 dp
->dp_mos_used_delta
= 0;
547 dp
->dp_mos_compressed_delta
= 0;
548 dp
->dp_mos_uncompressed_delta
= 0;
551 if (list_head(&mos
->os_dirty_dnodes
[txg
& TXG_MASK
]) != NULL
||
552 list_head(&mos
->os_free_dnodes
[txg
& TXG_MASK
]) != NULL
) {
553 dsl_pool_sync_mos(dp
, tx
);
557 * If we modify a dataset in the same txg that we want to destroy it,
558 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
559 * dsl_dir_destroy_check() will fail if there are unexpected holds.
560 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
561 * and clearing the hold on it) before we process the sync_tasks.
562 * The MOS data dirtied by the sync_tasks will be synced on the next
565 if (!txg_list_empty(&dp
->dp_sync_tasks
, txg
)) {
566 dsl_sync_task_t
*dst
;
568 * No more sync tasks should have been added while we
571 ASSERT3U(spa_sync_pass(dp
->dp_spa
), ==, 1);
572 while ((dst
= txg_list_remove(&dp
->dp_sync_tasks
, txg
)) != NULL
)
573 dsl_sync_task_sync(dst
, tx
);
578 DTRACE_PROBE2(dsl_pool_sync__done
, dsl_pool_t
*dp
, dp
, uint64_t, txg
);
582 dsl_pool_sync_done(dsl_pool_t
*dp
, uint64_t txg
)
586 while ((zilog
= txg_list_remove(&dp
->dp_dirty_zilogs
, txg
))) {
587 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
588 zil_clean(zilog
, txg
);
589 ASSERT(!dmu_objset_is_dirty(zilog
->zl_os
, txg
));
590 dmu_buf_rele(ds
->ds_dbuf
, zilog
);
592 ASSERT(!dmu_objset_is_dirty(dp
->dp_meta_objset
, txg
));
596 * TRUE if the current thread is the tx_sync_thread or if we
597 * are being called from SPA context during pool initialization.
600 dsl_pool_sync_context(dsl_pool_t
*dp
)
602 return (curthread
== dp
->dp_tx
.tx_sync_thread
||
603 spa_is_initializing(dp
->dp_spa
));
607 dsl_pool_adjustedsize(dsl_pool_t
*dp
, boolean_t netfree
)
609 uint64_t space
, resv
;
612 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
614 * XXX The intent log is not accounted for, so it must fit
617 * If we're trying to assess whether it's OK to do a free,
618 * cut the reservation in half to allow forward progress
619 * (e.g. make it possible to rm(1) files from a full pool).
621 space
= spa_get_dspace(dp
->dp_spa
);
622 resv
= MAX(space
>> 6, SPA_MINDEVSIZE
>> 1);
626 return (space
- resv
);
630 dsl_pool_need_dirty_delay(dsl_pool_t
*dp
)
632 uint64_t delay_min_bytes
=
633 zfs_dirty_data_max
* zfs_delay_min_dirty_percent
/ 100;
636 mutex_enter(&dp
->dp_lock
);
637 if (dp
->dp_dirty_total
> zfs_dirty_data_sync
)
639 rv
= (dp
->dp_dirty_total
> delay_min_bytes
);
640 mutex_exit(&dp
->dp_lock
);
645 dsl_pool_dirty_space(dsl_pool_t
*dp
, int64_t space
, dmu_tx_t
*tx
)
648 mutex_enter(&dp
->dp_lock
);
649 dp
->dp_dirty_pertxg
[tx
->tx_txg
& TXG_MASK
] += space
;
650 dsl_pool_dirty_delta(dp
, space
);
651 mutex_exit(&dp
->dp_lock
);
656 dsl_pool_undirty_space(dsl_pool_t
*dp
, int64_t space
, uint64_t txg
)
658 ASSERT3S(space
, >=, 0);
662 mutex_enter(&dp
->dp_lock
);
663 if (dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] < space
) {
664 /* XXX writing something we didn't dirty? */
665 space
= dp
->dp_dirty_pertxg
[txg
& TXG_MASK
];
667 ASSERT3U(dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], >=, space
);
668 dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] -= space
;
669 ASSERT3U(dp
->dp_dirty_total
, >=, space
);
670 dsl_pool_dirty_delta(dp
, -space
);
671 mutex_exit(&dp
->dp_lock
);
676 upgrade_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*hds
, void *arg
)
679 dsl_dataset_t
*ds
, *prev
= NULL
;
682 err
= dsl_dataset_hold_obj(dp
, hds
->ds_object
, FTAG
, &ds
);
686 while (dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= 0) {
687 err
= dsl_dataset_hold_obj(dp
,
688 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, FTAG
, &prev
);
690 dsl_dataset_rele(ds
, FTAG
);
694 if (dsl_dataset_phys(prev
)->ds_next_snap_obj
!= ds
->ds_object
)
696 dsl_dataset_rele(ds
, FTAG
);
702 prev
= dp
->dp_origin_snap
;
705 * The $ORIGIN can't have any data, or the accounting
708 ASSERT0(dsl_dataset_phys(prev
)->ds_bp
.blk_birth
);
710 /* The origin doesn't get attached to itself */
711 if (ds
->ds_object
== prev
->ds_object
) {
712 dsl_dataset_rele(ds
, FTAG
);
716 dmu_buf_will_dirty(ds
->ds_dbuf
, tx
);
717 dsl_dataset_phys(ds
)->ds_prev_snap_obj
= prev
->ds_object
;
718 dsl_dataset_phys(ds
)->ds_prev_snap_txg
=
719 dsl_dataset_phys(prev
)->ds_creation_txg
;
721 dmu_buf_will_dirty(ds
->ds_dir
->dd_dbuf
, tx
);
722 dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
= prev
->ds_object
;
724 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
725 dsl_dataset_phys(prev
)->ds_num_children
++;
727 if (dsl_dataset_phys(ds
)->ds_next_snap_obj
== 0) {
728 ASSERT(ds
->ds_prev
== NULL
);
729 VERIFY0(dsl_dataset_hold_obj(dp
,
730 dsl_dataset_phys(ds
)->ds_prev_snap_obj
,
735 ASSERT3U(dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
, ==, prev
->ds_object
);
736 ASSERT3U(dsl_dataset_phys(ds
)->ds_prev_snap_obj
, ==, prev
->ds_object
);
738 if (dsl_dataset_phys(prev
)->ds_next_clones_obj
== 0) {
739 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
740 dsl_dataset_phys(prev
)->ds_next_clones_obj
=
741 zap_create(dp
->dp_meta_objset
,
742 DMU_OT_NEXT_CLONES
, DMU_OT_NONE
, 0, tx
);
744 VERIFY0(zap_add_int(dp
->dp_meta_objset
,
745 dsl_dataset_phys(prev
)->ds_next_clones_obj
, ds
->ds_object
, tx
));
747 dsl_dataset_rele(ds
, FTAG
);
748 if (prev
!= dp
->dp_origin_snap
)
749 dsl_dataset_rele(prev
, FTAG
);
754 dsl_pool_upgrade_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
756 ASSERT(dmu_tx_is_syncing(tx
));
757 ASSERT(dp
->dp_origin_snap
!= NULL
);
759 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
, upgrade_clones_cb
,
760 tx
, DS_FIND_CHILDREN
));
765 upgrade_dir_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *arg
)
768 objset_t
*mos
= dp
->dp_meta_objset
;
770 if (dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
!= 0) {
771 dsl_dataset_t
*origin
;
773 VERIFY0(dsl_dataset_hold_obj(dp
,
774 dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
, FTAG
, &origin
));
776 if (dsl_dir_phys(origin
->ds_dir
)->dd_clones
== 0) {
777 dmu_buf_will_dirty(origin
->ds_dir
->dd_dbuf
, tx
);
778 dsl_dir_phys(origin
->ds_dir
)->dd_clones
=
779 zap_create(mos
, DMU_OT_DSL_CLONES
, DMU_OT_NONE
,
783 VERIFY0(zap_add_int(dp
->dp_meta_objset
,
784 dsl_dir_phys(origin
->ds_dir
)->dd_clones
,
787 dsl_dataset_rele(origin
, FTAG
);
793 dsl_pool_upgrade_dir_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
797 ASSERT(dmu_tx_is_syncing(tx
));
799 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, FREE_DIR_NAME
, tx
);
800 VERIFY0(dsl_pool_open_special_dir(dp
,
801 FREE_DIR_NAME
, &dp
->dp_free_dir
));
804 * We can't use bpobj_alloc(), because spa_version() still
805 * returns the old version, and we need a new-version bpobj with
806 * subobj support. So call dmu_object_alloc() directly.
808 obj
= dmu_object_alloc(dp
->dp_meta_objset
, DMU_OT_BPOBJ
,
809 SPA_MAXBLOCKSIZE
, DMU_OT_BPOBJ_HDR
, sizeof (bpobj_phys_t
), tx
);
810 VERIFY0(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
811 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
));
812 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
, dp
->dp_meta_objset
, obj
));
814 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
,
815 upgrade_dir_clones_cb
, tx
, DS_FIND_CHILDREN
));
819 dsl_pool_create_origin(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
824 ASSERT(dmu_tx_is_syncing(tx
));
825 ASSERT(dp
->dp_origin_snap
== NULL
);
826 ASSERT(rrw_held(&dp
->dp_config_rwlock
, RW_WRITER
));
828 /* create the origin dir, ds, & snap-ds */
829 dsobj
= dsl_dataset_create_sync(dp
->dp_root_dir
, ORIGIN_DIR_NAME
,
831 VERIFY0(dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
832 dsl_dataset_snapshot_sync_impl(ds
, ORIGIN_DIR_NAME
, tx
);
833 VERIFY0(dsl_dataset_hold_obj(dp
, dsl_dataset_phys(ds
)->ds_prev_snap_obj
,
834 dp
, &dp
->dp_origin_snap
));
835 dsl_dataset_rele(ds
, FTAG
);
839 dsl_pool_iput_taskq(dsl_pool_t
*dp
)
841 return (dp
->dp_iput_taskq
);
845 * Walk through the pool-wide zap object of temporary snapshot user holds
849 dsl_pool_clean_tmp_userrefs(dsl_pool_t
*dp
)
853 objset_t
*mos
= dp
->dp_meta_objset
;
854 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
859 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
861 holds
= fnvlist_alloc();
863 for (zap_cursor_init(&zc
, mos
, zapobj
);
864 zap_cursor_retrieve(&zc
, &za
) == 0;
865 zap_cursor_advance(&zc
)) {
869 htag
= strchr(za
.za_name
, '-');
872 if (nvlist_lookup_nvlist(holds
, za
.za_name
, &tags
) != 0) {
873 tags
= fnvlist_alloc();
874 fnvlist_add_boolean(tags
, htag
);
875 fnvlist_add_nvlist(holds
, za
.za_name
, tags
);
878 fnvlist_add_boolean(tags
, htag
);
881 dsl_dataset_user_release_tmp(dp
, holds
);
883 zap_cursor_fini(&zc
);
887 * Create the pool-wide zap object for storing temporary snapshot holds.
890 dsl_pool_user_hold_create_obj(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
892 objset_t
*mos
= dp
->dp_meta_objset
;
894 ASSERT(dp
->dp_tmp_userrefs_obj
== 0);
895 ASSERT(dmu_tx_is_syncing(tx
));
897 dp
->dp_tmp_userrefs_obj
= zap_create_link(mos
, DMU_OT_USERREFS
,
898 DMU_POOL_DIRECTORY_OBJECT
, DMU_POOL_TMP_USERREFS
, tx
);
902 dsl_pool_user_hold_rele_impl(dsl_pool_t
*dp
, uint64_t dsobj
,
903 const char *tag
, uint64_t now
, dmu_tx_t
*tx
, boolean_t holding
)
905 objset_t
*mos
= dp
->dp_meta_objset
;
906 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
910 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
911 ASSERT(dmu_tx_is_syncing(tx
));
914 * If the pool was created prior to SPA_VERSION_USERREFS, the
915 * zap object for temporary holds might not exist yet.
919 dsl_pool_user_hold_create_obj(dp
, tx
);
920 zapobj
= dp
->dp_tmp_userrefs_obj
;
922 return (SET_ERROR(ENOENT
));
926 name
= kmem_asprintf("%llx-%s", (u_longlong_t
)dsobj
, tag
);
928 error
= zap_add(mos
, zapobj
, name
, 8, 1, &now
, tx
);
930 error
= zap_remove(mos
, zapobj
, name
, tx
);
937 * Add a temporary hold for the given dataset object and tag.
940 dsl_pool_user_hold(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
941 uint64_t now
, dmu_tx_t
*tx
)
943 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, now
, tx
, B_TRUE
));
947 * Release a temporary hold for the given dataset object and tag.
950 dsl_pool_user_release(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
953 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, 0,
958 * DSL Pool Configuration Lock
960 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
961 * creation / destruction / rename / property setting). It must be held for
962 * read to hold a dataset or dsl_dir. I.e. you must call
963 * dsl_pool_config_enter() or dsl_pool_hold() before calling
964 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
965 * must be held continuously until all datasets and dsl_dirs are released.
967 * The only exception to this rule is that if a "long hold" is placed on
968 * a dataset, then the dp_config_rwlock may be dropped while the dataset
969 * is still held. The long hold will prevent the dataset from being
970 * destroyed -- the destroy will fail with EBUSY. A long hold can be
971 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
972 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
974 * Legitimate long-holders (including owners) should be long-running, cancelable
975 * tasks that should cause "zfs destroy" to fail. This includes DMU
976 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
977 * "zfs send", and "zfs diff". There are several other long-holders whose
978 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
980 * The usual formula for long-holding would be:
983 * ... perform checks ...
984 * dsl_dataset_long_hold()
986 * ... perform long-running task ...
987 * dsl_dataset_long_rele()
990 * Note that when the long hold is released, the dataset is still held but
991 * the pool is not held. The dataset may change arbitrarily during this time
992 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
993 * dataset except release it.
995 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
996 * or modifying operations.
998 * Modifying operations should generally use dsl_sync_task(). The synctask
999 * infrastructure enforces proper locking strategy with respect to the
1000 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1002 * Read-only operations will manually hold the pool, then the dataset, obtain
1003 * information from the dataset, then release the pool and dataset.
1004 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1009 dsl_pool_hold(const char *name
, void *tag
, dsl_pool_t
**dp
)
1014 error
= spa_open(name
, &spa
, tag
);
1016 *dp
= spa_get_dsl(spa
);
1017 dsl_pool_config_enter(*dp
, tag
);
1023 dsl_pool_rele(dsl_pool_t
*dp
, void *tag
)
1025 dsl_pool_config_exit(dp
, tag
);
1026 spa_close(dp
->dp_spa
, tag
);
1030 dsl_pool_config_enter(dsl_pool_t
*dp
, void *tag
)
1033 * We use a "reentrant" reader-writer lock, but not reentrantly.
1035 * The rrwlock can (with the track_all flag) track all reading threads,
1036 * which is very useful for debugging which code path failed to release
1037 * the lock, and for verifying that the *current* thread does hold
1040 * (Unlike a rwlock, which knows that N threads hold it for
1041 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1042 * if any thread holds it for read, even if this thread doesn't).
1044 ASSERT(!rrw_held(&dp
->dp_config_rwlock
, RW_READER
));
1045 rrw_enter(&dp
->dp_config_rwlock
, RW_READER
, tag
);
1049 dsl_pool_config_exit(dsl_pool_t
*dp
, void *tag
)
1051 rrw_exit(&dp
->dp_config_rwlock
, tag
);
1055 dsl_pool_config_held(dsl_pool_t
*dp
)
1057 return (RRW_LOCK_HELD(&dp
->dp_config_rwlock
));
1060 #if defined(_KERNEL) && defined(HAVE_SPL)
1061 EXPORT_SYMBOL(dsl_pool_config_enter
);
1062 EXPORT_SYMBOL(dsl_pool_config_exit
);
1064 /* zfs_dirty_data_max_percent only applied at module load in arc_init(). */
1065 module_param(zfs_dirty_data_max_percent
, int, 0444);
1066 MODULE_PARM_DESC(zfs_dirty_data_max_percent
, "percent of ram can be dirty");
1068 /* zfs_dirty_data_max_max_percent only applied at module load in arc_init(). */
1069 module_param(zfs_dirty_data_max_max_percent
, int, 0444);
1070 MODULE_PARM_DESC(zfs_dirty_data_max_max_percent
,
1071 "zfs_dirty_data_max upper bound as % of RAM");
1073 module_param(zfs_delay_min_dirty_percent
, int, 0644);
1074 MODULE_PARM_DESC(zfs_delay_min_dirty_percent
, "transaction delay threshold");
1076 module_param(zfs_dirty_data_max
, ulong
, 0644);
1077 MODULE_PARM_DESC(zfs_dirty_data_max
, "determines the dirty space limit");
1079 /* zfs_dirty_data_max_max only applied at module load in arc_init(). */
1080 module_param(zfs_dirty_data_max_max
, ulong
, 0444);
1081 MODULE_PARM_DESC(zfs_dirty_data_max_max
,
1082 "zfs_dirty_data_max upper bound in bytes");
1084 module_param(zfs_dirty_data_sync
, ulong
, 0644);
1085 MODULE_PARM_DESC(zfs_dirty_data_sync
, "sync txg when this much dirty data");
1087 module_param(zfs_delay_scale
, ulong
, 0644);
1088 MODULE_PARM_DESC(zfs_delay_scale
, "how quickly delay approaches infinity");