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, 2016 by Delphix. All rights reserved.
24 * Copyright (c) 2013 Steven Hartland. All rights reserved.
25 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
26 * Copyright (c) 2014 Integros [integros.com]
29 #include <sys/dsl_pool.h>
30 #include <sys/dsl_dataset.h>
31 #include <sys/dsl_prop.h>
32 #include <sys/dsl_dir.h>
33 #include <sys/dsl_synctask.h>
34 #include <sys/dsl_scan.h>
35 #include <sys/dnode.h>
36 #include <sys/dmu_tx.h>
37 #include <sys/dmu_objset.h>
41 #include <sys/zfs_context.h>
42 #include <sys/fs/zfs.h>
43 #include <sys/zfs_znode.h>
44 #include <sys/spa_impl.h>
45 #include <sys/dsl_deadlist.h>
46 #include <sys/bptree.h>
47 #include <sys/zfeature.h>
48 #include <sys/zil_impl.h>
49 #include <sys/dsl_userhold.h>
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 in /etc/system.
98 uint64_t zfs_dirty_data_max
;
99 uint64_t zfs_dirty_data_max_max
= 4ULL * 1024 * 1024 * 1024;
100 int zfs_dirty_data_max_percent
= 10;
103 * If there is at least this much dirty data, push out a txg.
105 uint64_t 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 uint64_t 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_vnrele_taskq
= taskq_create("zfs_vn_rele_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
);
320 * We can't set retry to TRUE since we're explicitly specifying
321 * a spa to flush. This is good enough; any missed buffers for
322 * this spa won't cause trouble, and they'll eventually fall
323 * out of the ARC just like any other unused buffer.
325 arc_flush(dp
->dp_spa
, FALSE
);
329 dmu_buf_user_evict_wait();
331 rrw_destroy(&dp
->dp_config_rwlock
);
332 mutex_destroy(&dp
->dp_lock
);
333 taskq_destroy(dp
->dp_vnrele_taskq
);
335 kmem_free(dp
->dp_blkstats
, sizeof (zfs_all_blkstats_t
));
336 kmem_free(dp
, sizeof (dsl_pool_t
));
340 dsl_pool_create(spa_t
*spa
, nvlist_t
*zplprops
, uint64_t txg
)
343 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
344 dmu_tx_t
*tx
= dmu_tx_create_assigned(dp
, txg
);
349 rrw_enter(&dp
->dp_config_rwlock
, RW_WRITER
, FTAG
);
351 /* create and open the MOS (meta-objset) */
352 dp
->dp_meta_objset
= dmu_objset_create_impl(spa
,
353 NULL
, &dp
->dp_meta_rootbp
, DMU_OST_META
, tx
);
355 /* create the pool directory */
356 err
= zap_create_claim(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
357 DMU_OT_OBJECT_DIRECTORY
, DMU_OT_NONE
, 0, tx
);
360 /* Initialize scan structures */
361 VERIFY0(dsl_scan_init(dp
, txg
));
363 /* create and open the root dir */
364 dp
->dp_root_dir_obj
= dsl_dir_create_sync(dp
, NULL
, NULL
, tx
);
365 VERIFY0(dsl_dir_hold_obj(dp
, dp
->dp_root_dir_obj
,
366 NULL
, dp
, &dp
->dp_root_dir
));
368 /* create and open the meta-objset dir */
369 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, MOS_DIR_NAME
, tx
);
370 VERIFY0(dsl_pool_open_special_dir(dp
,
371 MOS_DIR_NAME
, &dp
->dp_mos_dir
));
373 if (spa_version(spa
) >= SPA_VERSION_DEADLISTS
) {
374 /* create and open the free dir */
375 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
,
377 VERIFY0(dsl_pool_open_special_dir(dp
,
378 FREE_DIR_NAME
, &dp
->dp_free_dir
));
380 /* create and open the free_bplist */
381 obj
= bpobj_alloc(dp
->dp_meta_objset
, SPA_OLD_MAXBLOCKSIZE
, tx
);
382 VERIFY(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
383 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
) == 0);
384 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
,
385 dp
->dp_meta_objset
, obj
));
388 if (spa_version(spa
) >= SPA_VERSION_DSL_SCRUB
)
389 dsl_pool_create_origin(dp
, tx
);
391 /* create the root dataset */
392 obj
= dsl_dataset_create_sync_dd(dp
->dp_root_dir
, NULL
, 0, tx
);
394 /* create the root objset */
395 VERIFY0(dsl_dataset_hold_obj(dp
, obj
, FTAG
, &ds
));
396 rrw_enter(&ds
->ds_bp_rwlock
, RW_READER
, FTAG
);
397 os
= dmu_objset_create_impl(dp
->dp_spa
, ds
,
398 dsl_dataset_get_blkptr(ds
), DMU_OST_ZFS
, tx
);
399 rrw_exit(&ds
->ds_bp_rwlock
, FTAG
);
401 zfs_create_fs(os
, kcred
, zplprops
, tx
);
403 dsl_dataset_rele(ds
, FTAG
);
407 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
413 * Account for the meta-objset space in its placeholder dsl_dir.
416 dsl_pool_mos_diduse_space(dsl_pool_t
*dp
,
417 int64_t used
, int64_t comp
, int64_t uncomp
)
419 ASSERT3U(comp
, ==, uncomp
); /* it's all metadata */
420 mutex_enter(&dp
->dp_lock
);
421 dp
->dp_mos_used_delta
+= used
;
422 dp
->dp_mos_compressed_delta
+= comp
;
423 dp
->dp_mos_uncompressed_delta
+= uncomp
;
424 mutex_exit(&dp
->dp_lock
);
428 deadlist_enqueue_cb(void *arg
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
430 dsl_deadlist_t
*dl
= arg
;
431 dsl_deadlist_insert(dl
, bp
, tx
);
436 dsl_pool_sync_mos(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
438 zio_t
*zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
439 dmu_objset_sync(dp
->dp_meta_objset
, zio
, tx
);
440 VERIFY0(zio_wait(zio
));
441 dprintf_bp(&dp
->dp_meta_rootbp
, "meta objset rootbp is %s", "");
442 spa_set_rootblkptr(dp
->dp_spa
, &dp
->dp_meta_rootbp
);
446 dsl_pool_dirty_delta(dsl_pool_t
*dp
, int64_t delta
)
448 ASSERT(MUTEX_HELD(&dp
->dp_lock
));
451 ASSERT3U(-delta
, <=, dp
->dp_dirty_total
);
453 dp
->dp_dirty_total
+= delta
;
456 * Note: we signal even when increasing dp_dirty_total.
457 * This ensures forward progress -- each thread wakes the next waiter.
459 if (dp
->dp_dirty_total
<= zfs_dirty_data_max
)
460 cv_signal(&dp
->dp_spaceavail_cv
);
464 dsl_pool_sync(dsl_pool_t
*dp
, uint64_t txg
)
470 objset_t
*mos
= dp
->dp_meta_objset
;
471 list_t synced_datasets
;
473 list_create(&synced_datasets
, sizeof (dsl_dataset_t
),
474 offsetof(dsl_dataset_t
, ds_synced_link
));
476 tx
= dmu_tx_create_assigned(dp
, txg
);
479 * Write out all dirty blocks of dirty datasets.
481 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
482 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
484 * We must not sync any non-MOS datasets twice, because
485 * we may have taken a snapshot of them. However, we
486 * may sync newly-created datasets on pass 2.
488 ASSERT(!list_link_active(&ds
->ds_synced_link
));
489 list_insert_tail(&synced_datasets
, ds
);
490 dsl_dataset_sync(ds
, zio
, tx
);
492 VERIFY0(zio_wait(zio
));
495 * We have written all of the accounted dirty data, so our
496 * dp_space_towrite should now be zero. However, some seldom-used
497 * code paths do not adhere to this (e.g. dbuf_undirty(), also
498 * rounding error in dbuf_write_physdone).
499 * Shore up the accounting of any dirtied space now.
501 dsl_pool_undirty_space(dp
, dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], txg
);
504 * After the data blocks have been written (ensured by the zio_wait()
505 * above), update the user/group space accounting.
507 for (ds
= list_head(&synced_datasets
); ds
!= NULL
;
508 ds
= list_next(&synced_datasets
, ds
)) {
509 dmu_objset_do_userquota_updates(ds
->ds_objset
, tx
);
513 * Sync the datasets again to push out the changes due to
514 * userspace updates. This must be done before we process the
515 * sync tasks, so that any snapshots will have the correct
516 * user accounting information (and we won't get confused
517 * about which blocks are part of the snapshot).
519 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
520 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
521 ASSERT(list_link_active(&ds
->ds_synced_link
));
522 dmu_buf_rele(ds
->ds_dbuf
, ds
);
523 dsl_dataset_sync(ds
, zio
, tx
);
525 VERIFY0(zio_wait(zio
));
528 * Now that the datasets have been completely synced, we can
529 * clean up our in-memory structures accumulated while syncing:
531 * - move dead blocks from the pending deadlist to the on-disk deadlist
532 * - release hold from dsl_dataset_dirty()
534 while ((ds
= list_remove_head(&synced_datasets
)) != NULL
) {
535 objset_t
*os
= ds
->ds_objset
;
536 bplist_iterate(&ds
->ds_pending_deadlist
,
537 deadlist_enqueue_cb
, &ds
->ds_deadlist
, tx
);
538 ASSERT(!dmu_objset_is_dirty(os
, txg
));
539 dmu_buf_rele(ds
->ds_dbuf
, ds
);
541 while ((dd
= txg_list_remove(&dp
->dp_dirty_dirs
, txg
)) != NULL
) {
542 dsl_dir_sync(dd
, tx
);
546 * The MOS's space is accounted for in the pool/$MOS
547 * (dp_mos_dir). We can't modify the mos while we're syncing
548 * it, so we remember the deltas and apply them here.
550 if (dp
->dp_mos_used_delta
!= 0 || dp
->dp_mos_compressed_delta
!= 0 ||
551 dp
->dp_mos_uncompressed_delta
!= 0) {
552 dsl_dir_diduse_space(dp
->dp_mos_dir
, DD_USED_HEAD
,
553 dp
->dp_mos_used_delta
,
554 dp
->dp_mos_compressed_delta
,
555 dp
->dp_mos_uncompressed_delta
, tx
);
556 dp
->dp_mos_used_delta
= 0;
557 dp
->dp_mos_compressed_delta
= 0;
558 dp
->dp_mos_uncompressed_delta
= 0;
561 if (list_head(&mos
->os_dirty_dnodes
[txg
& TXG_MASK
]) != NULL
||
562 list_head(&mos
->os_free_dnodes
[txg
& TXG_MASK
]) != NULL
) {
563 dsl_pool_sync_mos(dp
, tx
);
567 * If we modify a dataset in the same txg that we want to destroy it,
568 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
569 * dsl_dir_destroy_check() will fail if there are unexpected holds.
570 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
571 * and clearing the hold on it) before we process the sync_tasks.
572 * The MOS data dirtied by the sync_tasks will be synced on the next
575 if (!txg_list_empty(&dp
->dp_sync_tasks
, txg
)) {
576 dsl_sync_task_t
*dst
;
578 * No more sync tasks should have been added while we
581 ASSERT3U(spa_sync_pass(dp
->dp_spa
), ==, 1);
582 while ((dst
= txg_list_remove(&dp
->dp_sync_tasks
, txg
)) != NULL
)
583 dsl_sync_task_sync(dst
, tx
);
588 DTRACE_PROBE2(dsl_pool_sync__done
, dsl_pool_t
*dp
, dp
, uint64_t, txg
);
592 dsl_pool_sync_done(dsl_pool_t
*dp
, uint64_t txg
)
596 while (zilog
= txg_list_head(&dp
->dp_dirty_zilogs
, txg
)) {
597 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
599 * We don't remove the zilog from the dp_dirty_zilogs
600 * list until after we've cleaned it. This ensures that
601 * callers of zilog_is_dirty() receive an accurate
602 * answer when they are racing with the spa sync thread.
604 zil_clean(zilog
, txg
);
605 (void) txg_list_remove_this(&dp
->dp_dirty_zilogs
, zilog
, txg
);
606 ASSERT(!dmu_objset_is_dirty(zilog
->zl_os
, txg
));
607 dmu_buf_rele(ds
->ds_dbuf
, zilog
);
609 ASSERT(!dmu_objset_is_dirty(dp
->dp_meta_objset
, txg
));
613 * TRUE if the current thread is the tx_sync_thread or if we
614 * are being called from SPA context during pool initialization.
617 dsl_pool_sync_context(dsl_pool_t
*dp
)
619 return (curthread
== dp
->dp_tx
.tx_sync_thread
||
620 spa_is_initializing(dp
->dp_spa
));
624 dsl_pool_adjustedsize(dsl_pool_t
*dp
, boolean_t netfree
)
626 uint64_t space
, resv
;
629 * If we're trying to assess whether it's OK to do a free,
630 * cut the reservation in half to allow forward progress
631 * (e.g. make it possible to rm(1) files from a full pool).
633 space
= spa_get_dspace(dp
->dp_spa
);
634 resv
= spa_get_slop_space(dp
->dp_spa
);
638 return (space
- resv
);
642 dsl_pool_need_dirty_delay(dsl_pool_t
*dp
)
644 uint64_t delay_min_bytes
=
645 zfs_dirty_data_max
* zfs_delay_min_dirty_percent
/ 100;
648 mutex_enter(&dp
->dp_lock
);
649 if (dp
->dp_dirty_total
> zfs_dirty_data_sync
)
651 rv
= (dp
->dp_dirty_total
> delay_min_bytes
);
652 mutex_exit(&dp
->dp_lock
);
657 dsl_pool_dirty_space(dsl_pool_t
*dp
, int64_t space
, dmu_tx_t
*tx
)
660 mutex_enter(&dp
->dp_lock
);
661 dp
->dp_dirty_pertxg
[tx
->tx_txg
& TXG_MASK
] += space
;
662 dsl_pool_dirty_delta(dp
, space
);
663 mutex_exit(&dp
->dp_lock
);
668 dsl_pool_undirty_space(dsl_pool_t
*dp
, int64_t space
, uint64_t txg
)
670 ASSERT3S(space
, >=, 0);
673 mutex_enter(&dp
->dp_lock
);
674 if (dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] < space
) {
675 /* XXX writing something we didn't dirty? */
676 space
= dp
->dp_dirty_pertxg
[txg
& TXG_MASK
];
678 ASSERT3U(dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], >=, space
);
679 dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] -= space
;
680 ASSERT3U(dp
->dp_dirty_total
, >=, space
);
681 dsl_pool_dirty_delta(dp
, -space
);
682 mutex_exit(&dp
->dp_lock
);
687 upgrade_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*hds
, void *arg
)
690 dsl_dataset_t
*ds
, *prev
= NULL
;
693 err
= dsl_dataset_hold_obj(dp
, hds
->ds_object
, FTAG
, &ds
);
697 while (dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= 0) {
698 err
= dsl_dataset_hold_obj(dp
,
699 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, FTAG
, &prev
);
701 dsl_dataset_rele(ds
, FTAG
);
705 if (dsl_dataset_phys(prev
)->ds_next_snap_obj
!= ds
->ds_object
)
707 dsl_dataset_rele(ds
, FTAG
);
713 prev
= dp
->dp_origin_snap
;
716 * The $ORIGIN can't have any data, or the accounting
719 rrw_enter(&ds
->ds_bp_rwlock
, RW_READER
, FTAG
);
720 ASSERT0(dsl_dataset_phys(prev
)->ds_bp
.blk_birth
);
721 rrw_exit(&ds
->ds_bp_rwlock
, FTAG
);
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
)
808 ASSERT(dmu_tx_is_syncing(tx
));
811 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, FREE_DIR_NAME
, tx
);
812 VERIFY0(dsl_pool_open_special_dir(dp
,
813 FREE_DIR_NAME
, &dp
->dp_free_dir
));
816 * We can't use bpobj_alloc(), because spa_version() still
817 * returns the old version, and we need a new-version bpobj with
818 * subobj support. So call dmu_object_alloc() directly.
820 obj
= dmu_object_alloc(dp
->dp_meta_objset
, DMU_OT_BPOBJ
,
821 SPA_OLD_MAXBLOCKSIZE
, DMU_OT_BPOBJ_HDR
, sizeof (bpobj_phys_t
), tx
);
822 VERIFY0(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
823 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
));
824 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
, dp
->dp_meta_objset
, obj
));
826 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
,
827 upgrade_dir_clones_cb
, tx
, DS_FIND_CHILDREN
| DS_FIND_SERIALIZE
));
831 dsl_pool_create_origin(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
836 ASSERT(dmu_tx_is_syncing(tx
));
837 ASSERT(dp
->dp_origin_snap
== NULL
);
838 ASSERT(rrw_held(&dp
->dp_config_rwlock
, RW_WRITER
));
840 /* create the origin dir, ds, & snap-ds */
841 dsobj
= dsl_dataset_create_sync(dp
->dp_root_dir
, ORIGIN_DIR_NAME
,
843 VERIFY0(dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
844 dsl_dataset_snapshot_sync_impl(ds
, ORIGIN_DIR_NAME
, tx
);
845 VERIFY0(dsl_dataset_hold_obj(dp
, dsl_dataset_phys(ds
)->ds_prev_snap_obj
,
846 dp
, &dp
->dp_origin_snap
));
847 dsl_dataset_rele(ds
, FTAG
);
851 dsl_pool_vnrele_taskq(dsl_pool_t
*dp
)
853 return (dp
->dp_vnrele_taskq
);
857 * Walk through the pool-wide zap object of temporary snapshot user holds
861 dsl_pool_clean_tmp_userrefs(dsl_pool_t
*dp
)
865 objset_t
*mos
= dp
->dp_meta_objset
;
866 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
871 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
873 holds
= fnvlist_alloc();
875 for (zap_cursor_init(&zc
, mos
, zapobj
);
876 zap_cursor_retrieve(&zc
, &za
) == 0;
877 zap_cursor_advance(&zc
)) {
881 htag
= strchr(za
.za_name
, '-');
884 if (nvlist_lookup_nvlist(holds
, za
.za_name
, &tags
) != 0) {
885 tags
= fnvlist_alloc();
886 fnvlist_add_boolean(tags
, htag
);
887 fnvlist_add_nvlist(holds
, za
.za_name
, tags
);
890 fnvlist_add_boolean(tags
, htag
);
893 dsl_dataset_user_release_tmp(dp
, holds
);
895 zap_cursor_fini(&zc
);
899 * Create the pool-wide zap object for storing temporary snapshot holds.
902 dsl_pool_user_hold_create_obj(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
904 objset_t
*mos
= dp
->dp_meta_objset
;
906 ASSERT(dp
->dp_tmp_userrefs_obj
== 0);
907 ASSERT(dmu_tx_is_syncing(tx
));
909 dp
->dp_tmp_userrefs_obj
= zap_create_link(mos
, DMU_OT_USERREFS
,
910 DMU_POOL_DIRECTORY_OBJECT
, DMU_POOL_TMP_USERREFS
, tx
);
914 dsl_pool_user_hold_rele_impl(dsl_pool_t
*dp
, uint64_t dsobj
,
915 const char *tag
, uint64_t now
, dmu_tx_t
*tx
, boolean_t holding
)
917 objset_t
*mos
= dp
->dp_meta_objset
;
918 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
922 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
923 ASSERT(dmu_tx_is_syncing(tx
));
926 * If the pool was created prior to SPA_VERSION_USERREFS, the
927 * zap object for temporary holds might not exist yet.
931 dsl_pool_user_hold_create_obj(dp
, tx
);
932 zapobj
= dp
->dp_tmp_userrefs_obj
;
934 return (SET_ERROR(ENOENT
));
938 name
= kmem_asprintf("%llx-%s", (u_longlong_t
)dsobj
, tag
);
940 error
= zap_add(mos
, zapobj
, name
, 8, 1, &now
, tx
);
942 error
= zap_remove(mos
, zapobj
, name
, tx
);
949 * Add a temporary hold for the given dataset object and tag.
952 dsl_pool_user_hold(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
953 uint64_t now
, dmu_tx_t
*tx
)
955 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, now
, tx
, B_TRUE
));
959 * Release a temporary hold for the given dataset object and tag.
962 dsl_pool_user_release(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
965 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, NULL
,
970 * DSL Pool Configuration Lock
972 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
973 * creation / destruction / rename / property setting). It must be held for
974 * read to hold a dataset or dsl_dir. I.e. you must call
975 * dsl_pool_config_enter() or dsl_pool_hold() before calling
976 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
977 * must be held continuously until all datasets and dsl_dirs are released.
979 * The only exception to this rule is that if a "long hold" is placed on
980 * a dataset, then the dp_config_rwlock may be dropped while the dataset
981 * is still held. The long hold will prevent the dataset from being
982 * destroyed -- the destroy will fail with EBUSY. A long hold can be
983 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
984 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
986 * Legitimate long-holders (including owners) should be long-running, cancelable
987 * tasks that should cause "zfs destroy" to fail. This includes DMU
988 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
989 * "zfs send", and "zfs diff". There are several other long-holders whose
990 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
992 * The usual formula for long-holding would be:
995 * ... perform checks ...
996 * dsl_dataset_long_hold()
998 * ... perform long-running task ...
999 * dsl_dataset_long_rele()
1000 * dsl_dataset_rele()
1002 * Note that when the long hold is released, the dataset is still held but
1003 * the pool is not held. The dataset may change arbitrarily during this time
1004 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1005 * dataset except release it.
1007 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1008 * or modifying operations.
1010 * Modifying operations should generally use dsl_sync_task(). The synctask
1011 * infrastructure enforces proper locking strategy with respect to the
1012 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1014 * Read-only operations will manually hold the pool, then the dataset, obtain
1015 * information from the dataset, then release the pool and dataset.
1016 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1021 dsl_pool_hold(const char *name
, void *tag
, dsl_pool_t
**dp
)
1026 error
= spa_open(name
, &spa
, tag
);
1028 *dp
= spa_get_dsl(spa
);
1029 dsl_pool_config_enter(*dp
, tag
);
1035 dsl_pool_rele(dsl_pool_t
*dp
, void *tag
)
1037 dsl_pool_config_exit(dp
, tag
);
1038 spa_close(dp
->dp_spa
, tag
);
1042 dsl_pool_config_enter(dsl_pool_t
*dp
, void *tag
)
1045 * We use a "reentrant" reader-writer lock, but not reentrantly.
1047 * The rrwlock can (with the track_all flag) track all reading threads,
1048 * which is very useful for debugging which code path failed to release
1049 * the lock, and for verifying that the *current* thread does hold
1052 * (Unlike a rwlock, which knows that N threads hold it for
1053 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1054 * if any thread holds it for read, even if this thread doesn't).
1056 ASSERT(!rrw_held(&dp
->dp_config_rwlock
, RW_READER
));
1057 rrw_enter(&dp
->dp_config_rwlock
, RW_READER
, tag
);
1061 dsl_pool_config_enter_prio(dsl_pool_t
*dp
, void *tag
)
1063 ASSERT(!rrw_held(&dp
->dp_config_rwlock
, RW_READER
));
1064 rrw_enter_read_prio(&dp
->dp_config_rwlock
, tag
);
1068 dsl_pool_config_exit(dsl_pool_t
*dp
, void *tag
)
1070 rrw_exit(&dp
->dp_config_rwlock
, tag
);
1074 dsl_pool_config_held(dsl_pool_t
*dp
)
1076 return (RRW_LOCK_HELD(&dp
->dp_config_rwlock
));
1080 dsl_pool_config_held_writer(dsl_pool_t
*dp
)
1082 return (RRW_WRITE_HELD(&dp
->dp_config_rwlock
));