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.
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 os
= dmu_objset_create_impl(dp
->dp_spa
, ds
,
397 dsl_dataset_get_blkptr(ds
), DMU_OST_ZFS
, tx
);
399 zfs_create_fs(os
, kcred
, zplprops
, tx
);
401 dsl_dataset_rele(ds
, FTAG
);
405 rrw_exit(&dp
->dp_config_rwlock
, FTAG
);
411 * Account for the meta-objset space in its placeholder dsl_dir.
414 dsl_pool_mos_diduse_space(dsl_pool_t
*dp
,
415 int64_t used
, int64_t comp
, int64_t uncomp
)
417 ASSERT3U(comp
, ==, uncomp
); /* it's all metadata */
418 mutex_enter(&dp
->dp_lock
);
419 dp
->dp_mos_used_delta
+= used
;
420 dp
->dp_mos_compressed_delta
+= comp
;
421 dp
->dp_mos_uncompressed_delta
+= uncomp
;
422 mutex_exit(&dp
->dp_lock
);
426 deadlist_enqueue_cb(void *arg
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
428 dsl_deadlist_t
*dl
= arg
;
429 dsl_deadlist_insert(dl
, bp
, tx
);
434 dsl_pool_sync_mos(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
436 zio_t
*zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
437 dmu_objset_sync(dp
->dp_meta_objset
, zio
, tx
);
438 VERIFY0(zio_wait(zio
));
439 dprintf_bp(&dp
->dp_meta_rootbp
, "meta objset rootbp is %s", "");
440 spa_set_rootblkptr(dp
->dp_spa
, &dp
->dp_meta_rootbp
);
444 dsl_pool_dirty_delta(dsl_pool_t
*dp
, int64_t delta
)
446 ASSERT(MUTEX_HELD(&dp
->dp_lock
));
449 ASSERT3U(-delta
, <=, dp
->dp_dirty_total
);
451 dp
->dp_dirty_total
+= delta
;
454 * Note: we signal even when increasing dp_dirty_total.
455 * This ensures forward progress -- each thread wakes the next waiter.
457 if (dp
->dp_dirty_total
<= zfs_dirty_data_max
)
458 cv_signal(&dp
->dp_spaceavail_cv
);
462 dsl_pool_sync(dsl_pool_t
*dp
, uint64_t txg
)
468 objset_t
*mos
= dp
->dp_meta_objset
;
469 list_t synced_datasets
;
471 list_create(&synced_datasets
, sizeof (dsl_dataset_t
),
472 offsetof(dsl_dataset_t
, ds_synced_link
));
474 tx
= dmu_tx_create_assigned(dp
, txg
);
477 * Write out all dirty blocks of dirty datasets.
479 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
480 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
482 * We must not sync any non-MOS datasets twice, because
483 * we may have taken a snapshot of them. However, we
484 * may sync newly-created datasets on pass 2.
486 ASSERT(!list_link_active(&ds
->ds_synced_link
));
487 list_insert_tail(&synced_datasets
, ds
);
488 dsl_dataset_sync(ds
, zio
, tx
);
490 VERIFY0(zio_wait(zio
));
493 * We have written all of the accounted dirty data, so our
494 * dp_space_towrite should now be zero. However, some seldom-used
495 * code paths do not adhere to this (e.g. dbuf_undirty(), also
496 * rounding error in dbuf_write_physdone).
497 * Shore up the accounting of any dirtied space now.
499 dsl_pool_undirty_space(dp
, dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], txg
);
502 * After the data blocks have been written (ensured by the zio_wait()
503 * above), update the user/group space accounting.
505 for (ds
= list_head(&synced_datasets
); ds
!= NULL
;
506 ds
= list_next(&synced_datasets
, ds
)) {
507 dmu_objset_do_userquota_updates(ds
->ds_objset
, tx
);
511 * Sync the datasets again to push out the changes due to
512 * userspace updates. This must be done before we process the
513 * sync tasks, so that any snapshots will have the correct
514 * user accounting information (and we won't get confused
515 * about which blocks are part of the snapshot).
517 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
518 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
)) != NULL
) {
519 ASSERT(list_link_active(&ds
->ds_synced_link
));
520 dmu_buf_rele(ds
->ds_dbuf
, ds
);
521 dsl_dataset_sync(ds
, zio
, tx
);
523 VERIFY0(zio_wait(zio
));
526 * Now that the datasets have been completely synced, we can
527 * clean up our in-memory structures accumulated while syncing:
529 * - move dead blocks from the pending deadlist to the on-disk deadlist
530 * - release hold from dsl_dataset_dirty()
532 while ((ds
= list_remove_head(&synced_datasets
)) != NULL
) {
533 objset_t
*os
= ds
->ds_objset
;
534 bplist_iterate(&ds
->ds_pending_deadlist
,
535 deadlist_enqueue_cb
, &ds
->ds_deadlist
, tx
);
536 ASSERT(!dmu_objset_is_dirty(os
, txg
));
537 dmu_buf_rele(ds
->ds_dbuf
, ds
);
539 while ((dd
= txg_list_remove(&dp
->dp_dirty_dirs
, txg
)) != NULL
) {
540 dsl_dir_sync(dd
, tx
);
544 * The MOS's space is accounted for in the pool/$MOS
545 * (dp_mos_dir). We can't modify the mos while we're syncing
546 * it, so we remember the deltas and apply them here.
548 if (dp
->dp_mos_used_delta
!= 0 || dp
->dp_mos_compressed_delta
!= 0 ||
549 dp
->dp_mos_uncompressed_delta
!= 0) {
550 dsl_dir_diduse_space(dp
->dp_mos_dir
, DD_USED_HEAD
,
551 dp
->dp_mos_used_delta
,
552 dp
->dp_mos_compressed_delta
,
553 dp
->dp_mos_uncompressed_delta
, tx
);
554 dp
->dp_mos_used_delta
= 0;
555 dp
->dp_mos_compressed_delta
= 0;
556 dp
->dp_mos_uncompressed_delta
= 0;
559 if (list_head(&mos
->os_dirty_dnodes
[txg
& TXG_MASK
]) != NULL
||
560 list_head(&mos
->os_free_dnodes
[txg
& TXG_MASK
]) != NULL
) {
561 dsl_pool_sync_mos(dp
, tx
);
565 * If we modify a dataset in the same txg that we want to destroy it,
566 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
567 * dsl_dir_destroy_check() will fail if there are unexpected holds.
568 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
569 * and clearing the hold on it) before we process the sync_tasks.
570 * The MOS data dirtied by the sync_tasks will be synced on the next
573 if (!txg_list_empty(&dp
->dp_sync_tasks
, txg
)) {
574 dsl_sync_task_t
*dst
;
576 * No more sync tasks should have been added while we
579 ASSERT3U(spa_sync_pass(dp
->dp_spa
), ==, 1);
580 while ((dst
= txg_list_remove(&dp
->dp_sync_tasks
, txg
)) != NULL
)
581 dsl_sync_task_sync(dst
, tx
);
586 DTRACE_PROBE2(dsl_pool_sync__done
, dsl_pool_t
*dp
, dp
, uint64_t, txg
);
590 dsl_pool_sync_done(dsl_pool_t
*dp
, uint64_t txg
)
594 while (zilog
= txg_list_remove(&dp
->dp_dirty_zilogs
, txg
)) {
595 dsl_dataset_t
*ds
= dmu_objset_ds(zilog
->zl_os
);
596 zil_clean(zilog
, txg
);
597 ASSERT(!dmu_objset_is_dirty(zilog
->zl_os
, txg
));
598 dmu_buf_rele(ds
->ds_dbuf
, zilog
);
600 ASSERT(!dmu_objset_is_dirty(dp
->dp_meta_objset
, txg
));
604 * TRUE if the current thread is the tx_sync_thread or if we
605 * are being called from SPA context during pool initialization.
608 dsl_pool_sync_context(dsl_pool_t
*dp
)
610 return (curthread
== dp
->dp_tx
.tx_sync_thread
||
611 spa_is_initializing(dp
->dp_spa
));
615 dsl_pool_adjustedsize(dsl_pool_t
*dp
, boolean_t netfree
)
617 uint64_t space
, resv
;
620 * If we're trying to assess whether it's OK to do a free,
621 * cut the reservation in half to allow forward progress
622 * (e.g. make it possible to rm(1) files from a full pool).
624 space
= spa_get_dspace(dp
->dp_spa
);
625 resv
= spa_get_slop_space(dp
->dp_spa
);
629 return (space
- resv
);
633 dsl_pool_need_dirty_delay(dsl_pool_t
*dp
)
635 uint64_t delay_min_bytes
=
636 zfs_dirty_data_max
* zfs_delay_min_dirty_percent
/ 100;
639 mutex_enter(&dp
->dp_lock
);
640 if (dp
->dp_dirty_total
> zfs_dirty_data_sync
)
642 rv
= (dp
->dp_dirty_total
> delay_min_bytes
);
643 mutex_exit(&dp
->dp_lock
);
648 dsl_pool_dirty_space(dsl_pool_t
*dp
, int64_t space
, dmu_tx_t
*tx
)
651 mutex_enter(&dp
->dp_lock
);
652 dp
->dp_dirty_pertxg
[tx
->tx_txg
& TXG_MASK
] += space
;
653 dsl_pool_dirty_delta(dp
, space
);
654 mutex_exit(&dp
->dp_lock
);
659 dsl_pool_undirty_space(dsl_pool_t
*dp
, int64_t space
, uint64_t txg
)
661 ASSERT3S(space
, >=, 0);
664 mutex_enter(&dp
->dp_lock
);
665 if (dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] < space
) {
666 /* XXX writing something we didn't dirty? */
667 space
= dp
->dp_dirty_pertxg
[txg
& TXG_MASK
];
669 ASSERT3U(dp
->dp_dirty_pertxg
[txg
& TXG_MASK
], >=, space
);
670 dp
->dp_dirty_pertxg
[txg
& TXG_MASK
] -= space
;
671 ASSERT3U(dp
->dp_dirty_total
, >=, space
);
672 dsl_pool_dirty_delta(dp
, -space
);
673 mutex_exit(&dp
->dp_lock
);
678 upgrade_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*hds
, void *arg
)
681 dsl_dataset_t
*ds
, *prev
= NULL
;
684 err
= dsl_dataset_hold_obj(dp
, hds
->ds_object
, FTAG
, &ds
);
688 while (dsl_dataset_phys(ds
)->ds_prev_snap_obj
!= 0) {
689 err
= dsl_dataset_hold_obj(dp
,
690 dsl_dataset_phys(ds
)->ds_prev_snap_obj
, FTAG
, &prev
);
692 dsl_dataset_rele(ds
, FTAG
);
696 if (dsl_dataset_phys(prev
)->ds_next_snap_obj
!= ds
->ds_object
)
698 dsl_dataset_rele(ds
, FTAG
);
704 prev
= dp
->dp_origin_snap
;
707 * The $ORIGIN can't have any data, or the accounting
710 ASSERT0(dsl_dataset_phys(prev
)->ds_bp
.blk_birth
);
712 /* The origin doesn't get attached to itself */
713 if (ds
->ds_object
== prev
->ds_object
) {
714 dsl_dataset_rele(ds
, FTAG
);
718 dmu_buf_will_dirty(ds
->ds_dbuf
, tx
);
719 dsl_dataset_phys(ds
)->ds_prev_snap_obj
= prev
->ds_object
;
720 dsl_dataset_phys(ds
)->ds_prev_snap_txg
=
721 dsl_dataset_phys(prev
)->ds_creation_txg
;
723 dmu_buf_will_dirty(ds
->ds_dir
->dd_dbuf
, tx
);
724 dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
= prev
->ds_object
;
726 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
727 dsl_dataset_phys(prev
)->ds_num_children
++;
729 if (dsl_dataset_phys(ds
)->ds_next_snap_obj
== 0) {
730 ASSERT(ds
->ds_prev
== NULL
);
731 VERIFY0(dsl_dataset_hold_obj(dp
,
732 dsl_dataset_phys(ds
)->ds_prev_snap_obj
,
737 ASSERT3U(dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
, ==, prev
->ds_object
);
738 ASSERT3U(dsl_dataset_phys(ds
)->ds_prev_snap_obj
, ==, prev
->ds_object
);
740 if (dsl_dataset_phys(prev
)->ds_next_clones_obj
== 0) {
741 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
742 dsl_dataset_phys(prev
)->ds_next_clones_obj
=
743 zap_create(dp
->dp_meta_objset
,
744 DMU_OT_NEXT_CLONES
, DMU_OT_NONE
, 0, tx
);
746 VERIFY0(zap_add_int(dp
->dp_meta_objset
,
747 dsl_dataset_phys(prev
)->ds_next_clones_obj
, ds
->ds_object
, tx
));
749 dsl_dataset_rele(ds
, FTAG
);
750 if (prev
!= dp
->dp_origin_snap
)
751 dsl_dataset_rele(prev
, FTAG
);
756 dsl_pool_upgrade_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
758 ASSERT(dmu_tx_is_syncing(tx
));
759 ASSERT(dp
->dp_origin_snap
!= NULL
);
761 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
, upgrade_clones_cb
,
762 tx
, DS_FIND_CHILDREN
| DS_FIND_SERIALIZE
));
767 upgrade_dir_clones_cb(dsl_pool_t
*dp
, dsl_dataset_t
*ds
, void *arg
)
770 objset_t
*mos
= dp
->dp_meta_objset
;
772 if (dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
!= 0) {
773 dsl_dataset_t
*origin
;
775 VERIFY0(dsl_dataset_hold_obj(dp
,
776 dsl_dir_phys(ds
->ds_dir
)->dd_origin_obj
, FTAG
, &origin
));
778 if (dsl_dir_phys(origin
->ds_dir
)->dd_clones
== 0) {
779 dmu_buf_will_dirty(origin
->ds_dir
->dd_dbuf
, tx
);
780 dsl_dir_phys(origin
->ds_dir
)->dd_clones
=
781 zap_create(mos
, DMU_OT_DSL_CLONES
, DMU_OT_NONE
,
785 VERIFY0(zap_add_int(dp
->dp_meta_objset
,
786 dsl_dir_phys(origin
->ds_dir
)->dd_clones
,
789 dsl_dataset_rele(origin
, FTAG
);
795 dsl_pool_upgrade_dir_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
797 ASSERT(dmu_tx_is_syncing(tx
));
800 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, FREE_DIR_NAME
, tx
);
801 VERIFY0(dsl_pool_open_special_dir(dp
,
802 FREE_DIR_NAME
, &dp
->dp_free_dir
));
805 * We can't use bpobj_alloc(), because spa_version() still
806 * returns the old version, and we need a new-version bpobj with
807 * subobj support. So call dmu_object_alloc() directly.
809 obj
= dmu_object_alloc(dp
->dp_meta_objset
, DMU_OT_BPOBJ
,
810 SPA_OLD_MAXBLOCKSIZE
, DMU_OT_BPOBJ_HDR
, sizeof (bpobj_phys_t
), tx
);
811 VERIFY0(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
812 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
));
813 VERIFY0(bpobj_open(&dp
->dp_free_bpobj
, dp
->dp_meta_objset
, obj
));
815 VERIFY0(dmu_objset_find_dp(dp
, dp
->dp_root_dir_obj
,
816 upgrade_dir_clones_cb
, tx
, DS_FIND_CHILDREN
| DS_FIND_SERIALIZE
));
820 dsl_pool_create_origin(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
825 ASSERT(dmu_tx_is_syncing(tx
));
826 ASSERT(dp
->dp_origin_snap
== NULL
);
827 ASSERT(rrw_held(&dp
->dp_config_rwlock
, RW_WRITER
));
829 /* create the origin dir, ds, & snap-ds */
830 dsobj
= dsl_dataset_create_sync(dp
->dp_root_dir
, ORIGIN_DIR_NAME
,
832 VERIFY0(dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
833 dsl_dataset_snapshot_sync_impl(ds
, ORIGIN_DIR_NAME
, tx
);
834 VERIFY0(dsl_dataset_hold_obj(dp
, dsl_dataset_phys(ds
)->ds_prev_snap_obj
,
835 dp
, &dp
->dp_origin_snap
));
836 dsl_dataset_rele(ds
, FTAG
);
840 dsl_pool_vnrele_taskq(dsl_pool_t
*dp
)
842 return (dp
->dp_vnrele_taskq
);
846 * Walk through the pool-wide zap object of temporary snapshot user holds
850 dsl_pool_clean_tmp_userrefs(dsl_pool_t
*dp
)
854 objset_t
*mos
= dp
->dp_meta_objset
;
855 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
860 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
862 holds
= fnvlist_alloc();
864 for (zap_cursor_init(&zc
, mos
, zapobj
);
865 zap_cursor_retrieve(&zc
, &za
) == 0;
866 zap_cursor_advance(&zc
)) {
870 htag
= strchr(za
.za_name
, '-');
873 if (nvlist_lookup_nvlist(holds
, za
.za_name
, &tags
) != 0) {
874 tags
= fnvlist_alloc();
875 fnvlist_add_boolean(tags
, htag
);
876 fnvlist_add_nvlist(holds
, za
.za_name
, tags
);
879 fnvlist_add_boolean(tags
, htag
);
882 dsl_dataset_user_release_tmp(dp
, holds
);
884 zap_cursor_fini(&zc
);
888 * Create the pool-wide zap object for storing temporary snapshot holds.
891 dsl_pool_user_hold_create_obj(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
893 objset_t
*mos
= dp
->dp_meta_objset
;
895 ASSERT(dp
->dp_tmp_userrefs_obj
== 0);
896 ASSERT(dmu_tx_is_syncing(tx
));
898 dp
->dp_tmp_userrefs_obj
= zap_create_link(mos
, DMU_OT_USERREFS
,
899 DMU_POOL_DIRECTORY_OBJECT
, DMU_POOL_TMP_USERREFS
, tx
);
903 dsl_pool_user_hold_rele_impl(dsl_pool_t
*dp
, uint64_t dsobj
,
904 const char *tag
, uint64_t now
, dmu_tx_t
*tx
, boolean_t holding
)
906 objset_t
*mos
= dp
->dp_meta_objset
;
907 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
911 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
912 ASSERT(dmu_tx_is_syncing(tx
));
915 * If the pool was created prior to SPA_VERSION_USERREFS, the
916 * zap object for temporary holds might not exist yet.
920 dsl_pool_user_hold_create_obj(dp
, tx
);
921 zapobj
= dp
->dp_tmp_userrefs_obj
;
923 return (SET_ERROR(ENOENT
));
927 name
= kmem_asprintf("%llx-%s", (u_longlong_t
)dsobj
, tag
);
929 error
= zap_add(mos
, zapobj
, name
, 8, 1, &now
, tx
);
931 error
= zap_remove(mos
, zapobj
, name
, tx
);
938 * Add a temporary hold for the given dataset object and tag.
941 dsl_pool_user_hold(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
942 uint64_t now
, dmu_tx_t
*tx
)
944 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, now
, tx
, B_TRUE
));
948 * Release a temporary hold for the given dataset object and tag.
951 dsl_pool_user_release(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
954 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, NULL
,
959 * DSL Pool Configuration Lock
961 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
962 * creation / destruction / rename / property setting). It must be held for
963 * read to hold a dataset or dsl_dir. I.e. you must call
964 * dsl_pool_config_enter() or dsl_pool_hold() before calling
965 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
966 * must be held continuously until all datasets and dsl_dirs are released.
968 * The only exception to this rule is that if a "long hold" is placed on
969 * a dataset, then the dp_config_rwlock may be dropped while the dataset
970 * is still held. The long hold will prevent the dataset from being
971 * destroyed -- the destroy will fail with EBUSY. A long hold can be
972 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
973 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
975 * Legitimate long-holders (including owners) should be long-running, cancelable
976 * tasks that should cause "zfs destroy" to fail. This includes DMU
977 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
978 * "zfs send", and "zfs diff". There are several other long-holders whose
979 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
981 * The usual formula for long-holding would be:
984 * ... perform checks ...
985 * dsl_dataset_long_hold()
987 * ... perform long-running task ...
988 * dsl_dataset_long_rele()
991 * Note that when the long hold is released, the dataset is still held but
992 * the pool is not held. The dataset may change arbitrarily during this time
993 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
994 * dataset except release it.
996 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
997 * or modifying operations.
999 * Modifying operations should generally use dsl_sync_task(). The synctask
1000 * infrastructure enforces proper locking strategy with respect to the
1001 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1003 * Read-only operations will manually hold the pool, then the dataset, obtain
1004 * information from the dataset, then release the pool and dataset.
1005 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1010 dsl_pool_hold(const char *name
, void *tag
, dsl_pool_t
**dp
)
1015 error
= spa_open(name
, &spa
, tag
);
1017 *dp
= spa_get_dsl(spa
);
1018 dsl_pool_config_enter(*dp
, tag
);
1024 dsl_pool_rele(dsl_pool_t
*dp
, void *tag
)
1026 dsl_pool_config_exit(dp
, tag
);
1027 spa_close(dp
->dp_spa
, tag
);
1031 dsl_pool_config_enter(dsl_pool_t
*dp
, void *tag
)
1034 * We use a "reentrant" reader-writer lock, but not reentrantly.
1036 * The rrwlock can (with the track_all flag) track all reading threads,
1037 * which is very useful for debugging which code path failed to release
1038 * the lock, and for verifying that the *current* thread does hold
1041 * (Unlike a rwlock, which knows that N threads hold it for
1042 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1043 * if any thread holds it for read, even if this thread doesn't).
1045 ASSERT(!rrw_held(&dp
->dp_config_rwlock
, RW_READER
));
1046 rrw_enter(&dp
->dp_config_rwlock
, RW_READER
, tag
);
1050 dsl_pool_config_enter_prio(dsl_pool_t
*dp
, void *tag
)
1052 ASSERT(!rrw_held(&dp
->dp_config_rwlock
, RW_READER
));
1053 rrw_enter_read_prio(&dp
->dp_config_rwlock
, tag
);
1057 dsl_pool_config_exit(dsl_pool_t
*dp
, void *tag
)
1059 rrw_exit(&dp
->dp_config_rwlock
, tag
);
1063 dsl_pool_config_held(dsl_pool_t
*dp
)
1065 return (RRW_LOCK_HELD(&dp
->dp_config_rwlock
));
1069 dsl_pool_config_held_writer(dsl_pool_t
*dp
)
1071 return (RRW_WRITE_HELD(&dp
->dp_config_rwlock
));