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.
25 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
28 #include <sys/dsl_pool.h>
29 #include <sys/dsl_dataset.h>
30 #include <sys/dsl_prop.h>
31 #include <sys/dsl_dir.h>
32 #include <sys/dsl_synctask.h>
33 #include <sys/dsl_scan.h>
34 #include <sys/dnode.h>
35 #include <sys/dmu_tx.h>
36 #include <sys/dmu_objset.h>
40 #include <sys/zfs_context.h>
41 #include <sys/fs/zfs.h>
42 #include <sys/zfs_znode.h>
43 #include <sys/spa_impl.h>
44 #include <sys/dsl_deadlist.h>
45 #include <sys/bptree.h>
46 #include <sys/zfeature.h>
47 #include <sys/zil_impl.h>
48 #include <sys/dsl_userhold.h>
49 #include <sys/trace_txg.h>
55 * ZFS must limit the rate of incoming writes to the rate at which it is able
56 * to sync data modifications to the backend storage. Throttling by too much
57 * creates an artificial limit; throttling by too little can only be sustained
58 * for short periods and would lead to highly lumpy performance. On a per-pool
59 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
60 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
61 * of dirty data decreases. When the amount of dirty data exceeds a
62 * predetermined threshold further modifications are blocked until the amount
63 * of dirty data decreases (as data is synced out).
65 * The limit on dirty data is tunable, and should be adjusted according to
66 * both the IO capacity and available memory of the system. The larger the
67 * window, the more ZFS is able to aggregate and amortize metadata (and data)
68 * changes. However, memory is a limited resource, and allowing for more dirty
69 * data comes at the cost of keeping other useful data in memory (for example
70 * ZFS data cached by the ARC).
74 * As buffers are modified dsl_pool_willuse_space() increments both the per-
75 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
76 * dirty space used; dsl_pool_dirty_space() decrements those values as data
77 * is synced out from dsl_pool_sync(). While only the poolwide value is
78 * relevant, the per-txg value is useful for debugging. The tunable
79 * zfs_dirty_data_max determines the dirty space limit. Once that value is
80 * exceeded, new writes are halted until space frees up.
82 * The zfs_dirty_data_sync tunable dictates the threshold at which we
83 * ensure that there is a txg syncing (see the comment in txg.c for a full
84 * description of transaction group stages).
86 * The IO scheduler uses both the dirty space limit and current amount of
87 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
88 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
90 * The delay is also calculated based on the amount of dirty data. See the
91 * comment above dmu_tx_delay() for details.
95 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
96 * capped at zfs_dirty_data_max_max. It can also be overridden with a module
99 unsigned long zfs_dirty_data_max
= 0;
100 unsigned long zfs_dirty_data_max_max
= 0;
101 int zfs_dirty_data_max_percent
= 10;
102 int zfs_dirty_data_max_max_percent
= 25;
105 * If there is at least this much dirty data, push out a txg.
107 unsigned long zfs_dirty_data_sync
= 64 * 1024 * 1024;
110 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
111 * and delay each transaction.
112 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
114 int zfs_delay_min_dirty_percent
= 60;
117 * This controls how quickly the delay approaches infinity.
118 * Larger values cause it to delay more for a given amount of dirty data.
119 * Therefore larger values will cause there to be less dirty data for a
122 * For the smoothest delay, this value should be about 1 billion divided
123 * by the maximum number of operations per second. This will smoothly
124 * handle between 10x and 1/10th this number.
126 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
127 * multiply in dmu_tx_delay().
129 unsigned long zfs_delay_scale
= 1000 * 1000 * 1000 / 2000;
131 hrtime_t zfs_throttle_delay
= MSEC2NSEC(10);
132 hrtime_t zfs_throttle_resolution
= MSEC2NSEC(10);
135 dsl_pool_open_special_dir(dsl_pool_t
*dp
, const char *name
, dsl_dir_t
**ddp
)
140 err
= zap_lookup(dp
->dp_meta_objset
,
141 dsl_dir_phys(dp
->dp_root_dir
)->dd_child_dir_zapobj
,
142 name
, sizeof (obj
), 1, &obj
);
146 return (dsl_dir_hold_obj(dp
, obj
, name
, dp
, ddp
));
150 dsl_pool_open_impl(spa_t
*spa
, uint64_t txg
)
153 blkptr_t
*bp
= spa_get_rootblkptr(spa
);
155 dp
= kmem_zalloc(sizeof (dsl_pool_t
), KM_SLEEP
);
157 dp
->dp_meta_rootbp
= *bp
;
158 rrw_init(&dp
->dp_config_rwlock
, B_TRUE
);
161 txg_list_create(&dp
->dp_dirty_datasets
,
162 offsetof(dsl_dataset_t
, ds_dirty_link
));
163 txg_list_create(&dp
->dp_dirty_zilogs
,
164 offsetof(zilog_t
, zl_dirty_link
));
165 txg_list_create(&dp
->dp_dirty_dirs
,
166 offsetof(dsl_dir_t
, dd_dirty_link
));
167 txg_list_create(&dp
->dp_sync_tasks
,
168 offsetof(dsl_sync_task_t
, dst_node
));
170 mutex_init(&dp
->dp_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
171 cv_init(&dp
->dp_spaceavail_cv
, NULL
, CV_DEFAULT
, NULL
);
173 dp
->dp_iput_taskq
= taskq_create("z_iput", max_ncpus
, defclsyspri
,
174 max_ncpus
* 8, INT_MAX
, TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
);
180 dsl_pool_init(spa_t
*spa
, uint64_t txg
, dsl_pool_t
**dpp
)
183 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
186 * Initialize the caller's dsl_pool_t structure before we actually open
187 * the meta objset. This is done because a self-healing write zio may
188 * be issued as part of dmu_objset_open_impl() and the spa needs its
189 * dsl_pool_t initialized in order to handle the write.
193 err
= dmu_objset_open_impl(spa
, NULL
, &dp
->dp_meta_rootbp
,
194 &dp
->dp_meta_objset
);
204 dsl_pool_open(dsl_pool_t
*dp
)
211 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
212 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
213 DMU_POOL_ROOT_DATASET
, sizeof (uint64_t), 1,
214 &dp
->dp_root_dir_obj
);
218 err
= dsl_dir_hold_obj(dp
, dp
->dp_root_dir_obj
,
219 NULL
, dp
, &dp
->dp_root_dir
);
223 err
= dsl_pool_open_special_dir(dp
, MOS_DIR_NAME
, &dp
->dp_mos_dir
);
227 if (spa_version(dp
->dp_spa
) >= SPA_VERSION_ORIGIN
) {
228 err
= dsl_pool_open_special_dir(dp
, ORIGIN_DIR_NAME
, &dd
);
231 err
= dsl_dataset_hold_obj(dp
,
232 dsl_dir_phys(dd
)->dd_head_dataset_obj
, FTAG
, &ds
);
234 err
= dsl_dataset_hold_obj(dp
,
235 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, dp
,
236 &dp
->dp_origin_snap
);
237 dsl_dataset_rele(ds
, FTAG
);
239 dsl_dir_rele(dd
, dp
);
244 if (spa_version(dp
->dp_spa
) >= SPA_VERSION_DEADLISTS
) {
245 err
= dsl_pool_open_special_dir(dp
, FREE_DIR_NAME
,
250 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
251 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
);
254 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
,
255 dp
->dp_meta_objset
, obj
));
259 * Note: errors ignored, because the leak dir will not exist if we
260 * have not encountered a leak yet.
262 (void) dsl_pool_open_special_dir(dp
, LEAK_DIR_NAME
,
265 if (spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_ASYNC_DESTROY
)) {
266 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
267 DMU_POOL_BPTREE_OBJ
, sizeof (uint64_t), 1,
273 if (spa_feature_is_active(dp
->dp_spa
, SPA_FEATURE_EMPTY_BPOBJ
)) {
274 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
275 DMU_POOL_EMPTY_BPOBJ
, sizeof (uint64_t), 1,
276 &dp
->dp_empty_bpobj
);
281 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
282 DMU_POOL_TMP_USERREFS
, sizeof (uint64_t), 1,
283 &dp
->dp_tmp_userrefs_obj
);
289 err
= dsl_scan_init(dp
, dp
->dp_tx
.tx_open_txg
);
292 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
297 dsl_pool_close(dsl_pool_t
*dp
)
300 * Drop our references from dsl_pool_open().
302 * Since we held the origin_snap from "syncing" context (which
303 * includes pool-opening context), it actually only got a "ref"
304 * and not a hold, so just drop that here.
306 if (dp
->dp_origin_snap
)
307 dsl_dataset_rele(dp
->dp_origin_snap
, dp
);
309 dsl_dir_rele(dp
->dp_mos_dir
, dp
);
311 dsl_dir_rele(dp
->dp_free_dir
, dp
);
313 dsl_dir_rele(dp
->dp_leak_dir
, dp
);
315 dsl_dir_rele(dp
->dp_root_dir
, dp
);
317 bpobj_close(&dp
->dp_free_bpobj
);
319 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
320 if (dp
->dp_meta_objset
)
321 dmu_objset_evict(dp
->dp_meta_objset
);
323 txg_list_destroy(&dp
->dp_dirty_datasets
);
324 txg_list_destroy(&dp
->dp_dirty_zilogs
);
325 txg_list_destroy(&dp
->dp_sync_tasks
);
326 txg_list_destroy(&dp
->dp_dirty_dirs
);
329 * We can't set retry to TRUE since we're explicitly specifying
330 * a spa to flush. This is good enough; any missed buffers for
331 * this spa won't cause trouble, and they'll eventually fall
332 * out of the ARC just like any other unused buffer.
334 arc_flush(dp
->dp_spa
, FALSE
);
338 dmu_buf_user_evict_wait();
340 rrw_destroy(&dp
->dp_config_rwlock
);
341 mutex_destroy(&dp
->dp_lock
);
342 taskq_destroy(dp
->dp_iput_taskq
);
344 vmem_free(dp
->dp_blkstats
, sizeof (zfs_all_blkstats_t
));
345 kmem_free(dp
, sizeof (dsl_pool_t
));
349 dsl_pool_create(spa_t
*spa
, nvlist_t
*zplprops
, uint64_t txg
)
352 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
353 dmu_tx_t
*tx
= dmu_tx_create_assigned(dp
, txg
);
358 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
360 /* create and open the MOS (meta-objset) */
361 dp
->dp_meta_objset
= dmu_objset_create_impl(spa
,
362 NULL
, &dp
->dp_meta_rootbp
, DMU_OST_META
, tx
);
364 /* create the pool directory */
365 err
= zap_create_claim(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
366 DMU_OT_OBJECT_DIRECTORY
, DMU_OT_NONE
, 0, tx
);
369 /* Initialize scan structures */
370 VERIFY0(dsl_scan_init(dp
, txg
));
372 /* create and open the root dir */
373 dp
->dp_root_dir_obj
= dsl_dir_create_sync(dp
, NULL
, NULL
, tx
);
374 VERIFY0(dsl_dir_hold_obj(dp
, dp
->dp_root_dir_obj
,
375 NULL
, dp
, &dp
->dp_root_dir
));
377 /* create and open the meta-objset dir */
378 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, MOS_DIR_NAME
, tx
);
379 VERIFY0(dsl_pool_open_special_dir(dp
,
380 MOS_DIR_NAME
, &dp
->dp_mos_dir
));
382 if (spa_version(spa
) >= SPA_VERSION_DEADLISTS
) {
383 /* create and open the free dir */
384 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
,
386 VERIFY0(dsl_pool_open_special_dir(dp
,
387 FREE_DIR_NAME
, &dp
->dp_free_dir
));
389 /* create and open the free_bplist */
390 obj
= bpobj_alloc(dp
->dp_meta_objset
, SPA_OLD_MAXBLOCKSIZE
, tx
);
391 VERIFY(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
392 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
) == 0);
393 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
,
394 dp
->dp_meta_objset
, obj
));
397 if (spa_version(spa
) >= SPA_VERSION_DSL_SCRUB
)
398 dsl_pool_create_origin(dp
, tx
);
400 /* create the root dataset */
401 obj
= dsl_dataset_create_sync_dd(dp
->dp_root_dir
, NULL
, 0, tx
);
403 /* create the root objset */
404 VERIFY0(dsl_dataset_hold_obj(dp
, obj
, FTAG
, &ds
));
405 VERIFY(NULL
!= (os
= dmu_objset_create_impl(dp
->dp_spa
, ds
,
406 dsl_dataset_get_blkptr(ds
), DMU_OST_ZFS
, tx
)));
408 zfs_create_fs(os
, kcred
, zplprops
, tx
);
410 dsl_dataset_rele(ds
, FTAG
);
414 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
420 * Account for the meta-objset space in its placeholder dsl_dir.
423 dsl_pool_mos_diduse_space(dsl_pool_t
*dp
,
424 int64_t used
, int64_t comp
, int64_t uncomp
)
426 ASSERT3U(comp
, ==, uncomp
); /* it's all metadata */
427 mutex_enter(&dp
->dp_lock
);
428 dp
->dp_mos_used_delta
+= used
;
429 dp
->dp_mos_compressed_delta
+= comp
;
430 dp
->dp_mos_uncompressed_delta
+= uncomp
;
431 mutex_exit(&dp
->dp_lock
);
435 deadlist_enqueue_cb(void *arg
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
437 dsl_deadlist_t
*dl
= arg
;
438 dsl_deadlist_insert(dl
, bp
, tx
);
443 dsl_pool_sync_mos(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
445 zio_t
*zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
446 dmu_objset_sync(dp
->dp_meta_objset
, zio
, tx
);
447 VERIFY0(zio_wait(zio
));
448 dprintf_bp(&dp
->dp_meta_rootbp
, "meta objset rootbp is %s", "");
449 spa_set_rootblkptr(dp
->dp_spa
, &dp
->dp_meta_rootbp
);
453 dsl_pool_dirty_delta(dsl_pool_t
*dp
, int64_t delta
)
455 ASSERT(MUTEX_HELD(&dp
->dp_lock
));
458 ASSERT3U(-delta
, <=, dp
->dp_dirty_total
);
460 dp
->dp_dirty_total
+= delta
;
463 * Note: we signal even when increasing dp_dirty_total.
464 * This ensures forward progress -- each thread wakes the next waiter.
466 if (dp
->dp_dirty_total
<= zfs_dirty_data_max
)
467 cv_signal(&dp
->dp_spaceavail_cv
);
471 dsl_pool_sync(dsl_pool_t
*dp
, uint64_t txg
)
477 objset_t
*mos
= dp
->dp_meta_objset
;
478 list_t synced_datasets
;
480 list_create(&synced_datasets
, sizeof (dsl_dataset_t
),
481 offsetof(dsl_dataset_t
, ds_synced_link
));
483 tx
= dmu_tx_create_assigned(dp
, txg
);
486 * Write out all dirty blocks of dirty datasets.
488 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
489 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
491 * We must not sync any non-MOS datasets twice, because
492 * we may have taken a snapshot of them. However, we
493 * may sync newly-created datasets on pass 2.
495 ASSERT(!list_link_active(&ds
->ds_synced_link
));
496 list_insert_tail(&synced_datasets
, ds
);
497 dsl_dataset_sync(ds
, zio
, tx
);
499 VERIFY0(zio_wait(zio
));
502 * We have written all of the accounted dirty data, so our
503 * dp_space_towrite should now be zero. However, some seldom-used
504 * code paths do not adhere to this (e.g. dbuf_undirty(), also
505 * rounding error in dbuf_write_physdone).
506 * Shore up the accounting of any dirtied space now.
508 dsl_pool_undirty_space(dp
, dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], txg
);
511 * After the data blocks have been written (ensured by the zio_wait()
512 * above), update the user/group space accounting.
514 for (ds
= list_head(&synced_datasets
); ds
!= NULL
;
515 ds
= list_next(&synced_datasets
, ds
)) {
516 dmu_objset_do_userquota_updates(ds
->ds_objset
, tx
);
520 * Sync the datasets again to push out the changes due to
521 * userspace updates. This must be done before we process the
522 * sync tasks, so that any snapshots will have the correct
523 * user accounting information (and we won't get confused
524 * about which blocks are part of the snapshot).
526 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
527 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
528 ASSERT(list_link_active(&ds
->ds_synced_link
));
529 dmu_buf_rele(ds
->ds_dbuf
, ds
);
530 dsl_dataset_sync(ds
, zio
, tx
);
532 VERIFY0(zio_wait(zio
));
535 * Now that the datasets have been completely synced, we can
536 * clean up our in-memory structures accumulated while syncing:
538 * - move dead blocks from the pending deadlist to the on-disk deadlist
539 * - release hold from dsl_dataset_dirty()
541 while ((ds
= list_remove_head(&synced_datasets
)) != NULL
) {
542 ASSERTV(objset_t
*os
= ds
->ds_objset
);
543 bplist_iterate(&ds
->ds_pending_deadlist
,
544 deadlist_enqueue_cb
, &ds
->ds_deadlist
, tx
);
545 ASSERT(!dmu_objset_is_dirty(os
, txg
));
546 dmu_buf_rele(ds
->ds_dbuf
, ds
);
549 while ((dd
= txg_list_remove(&dp
->dp_dirty_dirs
, txg
)) != NULL
) {
550 dsl_dir_sync(dd
, tx
);
554 * The MOS's space is accounted for in the pool/$MOS
555 * (dp_mos_dir). We can't modify the mos while we're syncing
556 * it, so we remember the deltas and apply them here.
558 if (dp
->dp_mos_used_delta
!= 0 || dp
->dp_mos_compressed_delta
!= 0 ||
559 dp
->dp_mos_uncompressed_delta
!= 0) {
560 dsl_dir_diduse_space(dp
->dp_mos_dir
, DD_USED_HEAD
,
561 dp
->dp_mos_used_delta
,
562 dp
->dp_mos_compressed_delta
,
563 dp
->dp_mos_uncompressed_delta
, tx
);
564 dp
->dp_mos_used_delta
= 0;
565 dp
->dp_mos_compressed_delta
= 0;
566 dp
->dp_mos_uncompressed_delta
= 0;
569 if (list_head(&mos
->os_dirty_dnodes
[txg
& TXG_MASK
]) != NULL
||
570 list_head(&mos
->os_free_dnodes
[txg
& TXG_MASK
]) != NULL
) {
571 dsl_pool_sync_mos(dp
, tx
);
575 * If we modify a dataset in the same txg that we want to destroy it,
576 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
577 * dsl_dir_destroy_check() will fail if there are unexpected holds.
578 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
579 * and clearing the hold on it) before we process the sync_tasks.
580 * The MOS data dirtied by the sync_tasks will be synced on the next
583 if (!txg_list_empty(&dp
->dp_sync_tasks
, txg
)) {
584 dsl_sync_task_t
*dst
;
586 * No more sync tasks should have been added while we
589 ASSERT3U(spa_sync_pass(dp
->dp_spa
), ==, 1);
590 while ((dst
= txg_list_remove(&dp
->dp_sync_tasks
, txg
)) != NULL
)
591 dsl_sync_task_sync(dst
, tx
);
596 DTRACE_PROBE2(dsl_pool_sync__done
, dsl_pool_t
*dp
, dp
, uint64_t, txg
);
600 dsl_pool_sync_done(dsl_pool_t
*dp
, uint64_t txg
)
604 while ((zilog
= txg_list_remove(&dp
->dp_dirty_zilogs
, txg
))) {
605 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
606 zil_clean(zilog
, txg
);
607 ASSERT(!dmu_objset_is_dirty(zilog
->zl_os
, txg
));
608 dmu_buf_rele(ds
->ds_dbuf
, zilog
);
610 ASSERT(!dmu_objset_is_dirty(dp
->dp_meta_objset
, txg
));
614 * TRUE if the current thread is the tx_sync_thread or if we
615 * are being called from SPA context during pool initialization.
618 dsl_pool_sync_context(dsl_pool_t
*dp
)
620 return (curthread
== dp
->dp_tx
.tx_sync_thread
||
621 spa_is_initializing(dp
->dp_spa
));
625 dsl_pool_adjustedsize(dsl_pool_t
*dp
, boolean_t netfree
)
627 uint64_t space
, resv
;
630 * If we're trying to assess whether it's OK to do a free,
631 * cut the reservation in half to allow forward progress
632 * (e.g. make it possible to rm(1) files from a full pool).
634 space
= spa_get_dspace(dp
->dp_spa
);
635 resv
= spa_get_slop_space(dp
->dp_spa
);
639 return (space
- resv
);
643 dsl_pool_need_dirty_delay(dsl_pool_t
*dp
)
645 uint64_t delay_min_bytes
=
646 zfs_dirty_data_max
* zfs_delay_min_dirty_percent
/ 100;
649 mutex_enter(&dp
->dp_lock
);
650 if (dp
->dp_dirty_total
> zfs_dirty_data_sync
)
652 rv
= (dp
->dp_dirty_total
> delay_min_bytes
);
653 mutex_exit(&dp
->dp_lock
);
658 dsl_pool_dirty_space(dsl_pool_t
*dp
, int64_t space
, dmu_tx_t
*tx
)
661 mutex_enter(&dp
->dp_lock
);
662 dp
->dp_dirty_pertxg
[tx
->tx_txg
& TXG_MASK
] += space
;
663 dsl_pool_dirty_delta(dp
, space
);
664 mutex_exit(&dp
->dp_lock
);
669 dsl_pool_undirty_space(dsl_pool_t
*dp
, int64_t space
, uint64_t txg
)
671 ASSERT3S(space
, >=, 0);
675 mutex_enter(&dp
->dp_lock
);
676 if (dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] < space
) {
677 /* XXX writing something we didn't dirty? */
678 space
= dp
->dp_dirty_pertxg
[txg
& TXG_MASK
];
680 ASSERT3U(dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], >=, space
);
681 dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] -= space
;
682 ASSERT3U(dp
->dp_dirty_total
, >=, space
);
683 dsl_pool_dirty_delta(dp
, -space
);
684 mutex_exit(&dp
->dp_lock
);
689 upgrade_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*hds
, void *arg
)
692 dsl_dataset_t
*ds
, *prev
= NULL
;
695 err
= dsl_dataset_hold_obj(dp
, hds
->ds_object
, FTAG
, &ds
);
699 while (dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= 0) {
700 err
= dsl_dataset_hold_obj(dp
,
701 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, FTAG
, &prev
);
703 dsl_dataset_rele(ds
, FTAG
);
707 if (dsl_dataset_phys(prev
)->ds_next_snap_obj
!= ds
->ds_object
)
709 dsl_dataset_rele(ds
, FTAG
);
715 prev
= dp
->dp_origin_snap
;
718 * The $ORIGIN can't have any data, or the accounting
721 ASSERT0(dsl_dataset_phys(prev
)->ds_bp
.blk_birth
);
723 /* The origin doesn't get attached to itself */
724 if (ds
->ds_object
== prev
->ds_object
) {
725 dsl_dataset_rele(ds
, FTAG
);
729 dmu_buf_will_dirty(ds
->ds_dbuf
, tx
);
730 dsl_dataset_phys(ds
)->ds_prev_snap_obj
= prev
->ds_object
;
731 dsl_dataset_phys(ds
)->ds_prev_snap_txg
=
732 dsl_dataset_phys(prev
)->ds_creation_txg
;
734 dmu_buf_will_dirty(ds
->ds_dir
->dd_dbuf
, tx
);
735 dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
= prev
->ds_object
;
737 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
738 dsl_dataset_phys(prev
)->ds_num_children
++;
740 if (dsl_dataset_phys(ds
)->ds_next_snap_obj
== 0) {
741 ASSERT(ds
->ds_prev
== NULL
);
742 VERIFY0(dsl_dataset_hold_obj(dp
,
743 dsl_dataset_phys(ds
)->ds_prev_snap_obj
,
748 ASSERT3U(dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
, ==, prev
->ds_object
);
749 ASSERT3U(dsl_dataset_phys(ds
)->ds_prev_snap_obj
, ==, prev
->ds_object
);
751 if (dsl_dataset_phys(prev
)->ds_next_clones_obj
== 0) {
752 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
753 dsl_dataset_phys(prev
)->ds_next_clones_obj
=
754 zap_create(dp
->dp_meta_objset
,
755 DMU_OT_NEXT_CLONES
, DMU_OT_NONE
, 0, tx
);
757 VERIFY0(zap_add_int(dp
->dp_meta_objset
,
758 dsl_dataset_phys(prev
)->ds_next_clones_obj
, ds
->ds_object
, tx
));
760 dsl_dataset_rele(ds
, FTAG
);
761 if (prev
!= dp
->dp_origin_snap
)
762 dsl_dataset_rele(prev
, FTAG
);
767 dsl_pool_upgrade_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
769 ASSERT(dmu_tx_is_syncing(tx
));
770 ASSERT(dp
->dp_origin_snap
!= NULL
);
772 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
, upgrade_clones_cb
,
773 tx
, DS_FIND_CHILDREN
| DS_FIND_SERIALIZE
));
778 upgrade_dir_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *arg
)
781 objset_t
*mos
= dp
->dp_meta_objset
;
783 if (dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
!= 0) {
784 dsl_dataset_t
*origin
;
786 VERIFY0(dsl_dataset_hold_obj(dp
,
787 dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
, FTAG
, &origin
));
789 if (dsl_dir_phys(origin
->ds_dir
)->dd_clones
== 0) {
790 dmu_buf_will_dirty(origin
->ds_dir
->dd_dbuf
, tx
);
791 dsl_dir_phys(origin
->ds_dir
)->dd_clones
=
792 zap_create(mos
, DMU_OT_DSL_CLONES
, DMU_OT_NONE
,
796 VERIFY0(zap_add_int(dp
->dp_meta_objset
,
797 dsl_dir_phys(origin
->ds_dir
)->dd_clones
,
800 dsl_dataset_rele(origin
, FTAG
);
806 dsl_pool_upgrade_dir_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
810 ASSERT(dmu_tx_is_syncing(tx
));
812 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, FREE_DIR_NAME
, tx
);
813 VERIFY0(dsl_pool_open_special_dir(dp
,
814 FREE_DIR_NAME
, &dp
->dp_free_dir
));
817 * We can't use bpobj_alloc(), because spa_version() still
818 * returns the old version, and we need a new-version bpobj with
819 * subobj support. So call dmu_object_alloc() directly.
821 obj
= dmu_object_alloc(dp
->dp_meta_objset
, DMU_OT_BPOBJ
,
822 SPA_OLD_MAXBLOCKSIZE
, DMU_OT_BPOBJ_HDR
, sizeof (bpobj_phys_t
), tx
);
823 VERIFY0(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
824 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
));
825 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
, dp
->dp_meta_objset
, obj
));
827 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
,
828 upgrade_dir_clones_cb
, tx
, DS_FIND_CHILDREN
| DS_FIND_SERIALIZE
));
832 dsl_pool_create_origin(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
837 ASSERT(dmu_tx_is_syncing(tx
));
838 ASSERT(dp
->dp_origin_snap
== NULL
);
839 ASSERT(rrw_held(&dp
->dp_config_rwlock
, RW_WRITER
));
841 /* create the origin dir, ds, & snap-ds */
842 dsobj
= dsl_dataset_create_sync(dp
->dp_root_dir
, ORIGIN_DIR_NAME
,
844 VERIFY0(dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
845 dsl_dataset_snapshot_sync_impl(ds
, ORIGIN_DIR_NAME
, tx
);
846 VERIFY0(dsl_dataset_hold_obj(dp
, dsl_dataset_phys(ds
)->ds_prev_snap_obj
,
847 dp
, &dp
->dp_origin_snap
));
848 dsl_dataset_rele(ds
, FTAG
);
852 dsl_pool_iput_taskq(dsl_pool_t
*dp
)
854 return (dp
->dp_iput_taskq
);
858 * Walk through the pool-wide zap object of temporary snapshot user holds
862 dsl_pool_clean_tmp_userrefs(dsl_pool_t
*dp
)
866 objset_t
*mos
= dp
->dp_meta_objset
;
867 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
872 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
874 holds
= fnvlist_alloc();
876 for (zap_cursor_init(&zc
, mos
, zapobj
);
877 zap_cursor_retrieve(&zc
, &za
) == 0;
878 zap_cursor_advance(&zc
)) {
882 htag
= strchr(za
.za_name
, '-');
885 if (nvlist_lookup_nvlist(holds
, za
.za_name
, &tags
) != 0) {
886 tags
= fnvlist_alloc();
887 fnvlist_add_boolean(tags
, htag
);
888 fnvlist_add_nvlist(holds
, za
.za_name
, tags
);
891 fnvlist_add_boolean(tags
, htag
);
894 dsl_dataset_user_release_tmp(dp
, holds
);
896 zap_cursor_fini(&zc
);
900 * Create the pool-wide zap object for storing temporary snapshot holds.
903 dsl_pool_user_hold_create_obj(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
905 objset_t
*mos
= dp
->dp_meta_objset
;
907 ASSERT(dp
->dp_tmp_userrefs_obj
== 0);
908 ASSERT(dmu_tx_is_syncing(tx
));
910 dp
->dp_tmp_userrefs_obj
= zap_create_link(mos
, DMU_OT_USERREFS
,
911 DMU_POOL_DIRECTORY_OBJECT
, DMU_POOL_TMP_USERREFS
, tx
);
915 dsl_pool_user_hold_rele_impl(dsl_pool_t
*dp
, uint64_t dsobj
,
916 const char *tag
, uint64_t now
, dmu_tx_t
*tx
, boolean_t holding
)
918 objset_t
*mos
= dp
->dp_meta_objset
;
919 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
923 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
924 ASSERT(dmu_tx_is_syncing(tx
));
927 * If the pool was created prior to SPA_VERSION_USERREFS, the
928 * zap object for temporary holds might not exist yet.
932 dsl_pool_user_hold_create_obj(dp
, tx
);
933 zapobj
= dp
->dp_tmp_userrefs_obj
;
935 return (SET_ERROR(ENOENT
));
939 name
= kmem_asprintf("%llx-%s", (u_longlong_t
)dsobj
, tag
);
941 error
= zap_add(mos
, zapobj
, name
, 8, 1, &now
, tx
);
943 error
= zap_remove(mos
, zapobj
, name
, tx
);
950 * Add a temporary hold for the given dataset object and tag.
953 dsl_pool_user_hold(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
954 uint64_t now
, dmu_tx_t
*tx
)
956 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, now
, tx
, B_TRUE
));
960 * Release a temporary hold for the given dataset object and tag.
963 dsl_pool_user_release(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
966 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, 0,
971 * DSL Pool Configuration Lock
973 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
974 * creation / destruction / rename / property setting). It must be held for
975 * read to hold a dataset or dsl_dir. I.e. you must call
976 * dsl_pool_config_enter() or dsl_pool_hold() before calling
977 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
978 * must be held continuously until all datasets and dsl_dirs are released.
980 * The only exception to this rule is that if a "long hold" is placed on
981 * a dataset, then the dp_config_rwlock may be dropped while the dataset
982 * is still held. The long hold will prevent the dataset from being
983 * destroyed -- the destroy will fail with EBUSY. A long hold can be
984 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
985 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
987 * Legitimate long-holders (including owners) should be long-running, cancelable
988 * tasks that should cause "zfs destroy" to fail. This includes DMU
989 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
990 * "zfs send", and "zfs diff". There are several other long-holders whose
991 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
993 * The usual formula for long-holding would be:
996 * ... perform checks ...
997 * dsl_dataset_long_hold()
999 * ... perform long-running task ...
1000 * dsl_dataset_long_rele()
1001 * dsl_dataset_rele()
1003 * Note that when the long hold is released, the dataset is still held but
1004 * the pool is not held. The dataset may change arbitrarily during this time
1005 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1006 * dataset except release it.
1008 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1009 * or modifying operations.
1011 * Modifying operations should generally use dsl_sync_task(). The synctask
1012 * infrastructure enforces proper locking strategy with respect to the
1013 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1015 * Read-only operations will manually hold the pool, then the dataset, obtain
1016 * information from the dataset, then release the pool and dataset.
1017 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1022 dsl_pool_hold(const char *name
, void *tag
, dsl_pool_t
**dp
)
1027 error
= spa_open(name
, &spa
, tag
);
1029 *dp
= spa_get_dsl(spa
);
1030 dsl_pool_config_enter(*dp
, tag
);
1036 dsl_pool_rele(dsl_pool_t
*dp
, void *tag
)
1038 dsl_pool_config_exit(dp
, tag
);
1039 spa_close(dp
->dp_spa
, tag
);
1043 dsl_pool_config_enter(dsl_pool_t
*dp
, void *tag
)
1046 * We use a "reentrant" reader-writer lock, but not reentrantly.
1048 * The rrwlock can (with the track_all flag) track all reading threads,
1049 * which is very useful for debugging which code path failed to release
1050 * the lock, and for verifying that the *current* thread does hold
1053 * (Unlike a rwlock, which knows that N threads hold it for
1054 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1055 * if any thread holds it for read, even if this thread doesn't).
1057 ASSERT(!rrw_held(&dp
->dp_config_rwlock
, RW_READER
));
1058 rrw_enter(&dp
->dp_config_rwlock
, RW_READER
, tag
);
1062 dsl_pool_config_enter_prio(dsl_pool_t
*dp
, void *tag
)
1064 ASSERT(!rrw_held(&dp
->dp_config_rwlock
, RW_READER
));
1065 rrw_enter_read_prio(&dp
->dp_config_rwlock
, tag
);
1069 dsl_pool_config_exit(dsl_pool_t
*dp
, void *tag
)
1071 rrw_exit(&dp
->dp_config_rwlock
, tag
);
1075 dsl_pool_config_held(dsl_pool_t
*dp
)
1077 return (RRW_LOCK_HELD(&dp
->dp_config_rwlock
));
1081 dsl_pool_config_held_writer(dsl_pool_t
*dp
)
1083 return (RRW_WRITE_HELD(&dp
->dp_config_rwlock
));
1086 #if defined(_KERNEL) && defined(HAVE_SPL)
1087 EXPORT_SYMBOL(dsl_pool_config_enter
);
1088 EXPORT_SYMBOL(dsl_pool_config_exit
);
1090 /* zfs_dirty_data_max_percent only applied at module load in arc_init(). */
1091 module_param(zfs_dirty_data_max_percent
, int, 0444);
1092 MODULE_PARM_DESC(zfs_dirty_data_max_percent
, "percent of ram can be dirty");
1094 /* zfs_dirty_data_max_max_percent only applied at module load in arc_init(). */
1095 module_param(zfs_dirty_data_max_max_percent
, int, 0444);
1096 MODULE_PARM_DESC(zfs_dirty_data_max_max_percent
,
1097 "zfs_dirty_data_max upper bound as % of RAM");
1099 module_param(zfs_delay_min_dirty_percent
, int, 0644);
1100 MODULE_PARM_DESC(zfs_delay_min_dirty_percent
, "transaction delay threshold");
1102 module_param(zfs_dirty_data_max
, ulong
, 0644);
1103 MODULE_PARM_DESC(zfs_dirty_data_max
, "determines the dirty space limit");
1105 /* zfs_dirty_data_max_max only applied at module load in arc_init(). */
1106 module_param(zfs_dirty_data_max_max
, ulong
, 0444);
1107 MODULE_PARM_DESC(zfs_dirty_data_max_max
,
1108 "zfs_dirty_data_max upper bound in bytes");
1110 module_param(zfs_dirty_data_sync
, ulong
, 0644);
1111 MODULE_PARM_DESC(zfs_dirty_data_sync
, "sync txg when this much dirty data");
1113 module_param(zfs_delay_scale
, ulong
, 0644);
1114 MODULE_PARM_DESC(zfs_delay_scale
, "how quickly delay approaches infinity");