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) 2012 by Delphix. All rights reserved.
26 #include <sys/dsl_pool.h>
27 #include <sys/dsl_dataset.h>
28 #include <sys/dsl_prop.h>
29 #include <sys/dsl_dir.h>
30 #include <sys/dsl_synctask.h>
31 #include <sys/dsl_scan.h>
32 #include <sys/dnode.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dmu_objset.h>
38 #include <sys/zfs_context.h>
39 #include <sys/fs/zfs.h>
40 #include <sys/zfs_znode.h>
41 #include <sys/spa_impl.h>
42 #include <sys/dsl_deadlist.h>
43 #include <sys/bptree.h>
44 #include <sys/zfeature.h>
45 #include <sys/zil_impl.h>
47 int zfs_no_write_throttle
= 0;
48 int zfs_write_limit_shift
= 3; /* 1/8th of physical memory */
49 int zfs_txg_synctime_ms
= 1000; /* target millisecs to sync a txg */
50 int zfs_txg_history
= 60; /* statistics for the last N txgs */
52 unsigned long zfs_write_limit_min
= 32 << 20; /* min write limit is 32MB */
53 unsigned long zfs_write_limit_max
= 0; /* max data payload per txg */
54 unsigned long zfs_write_limit_inflated
= 0;
55 unsigned long zfs_write_limit_override
= 0;
57 kmutex_t zfs_write_limit_lock
;
59 static pgcnt_t old_physmem
= 0;
62 dsl_pool_txg_history_update(kstat_t
*ksp
, int rw
)
64 dsl_pool_t
*dp
= ksp
->ks_private
;
68 if (rw
== KSTAT_WRITE
)
72 kmem_free(ksp
->ks_data
, ksp
->ks_data_size
);
74 mutex_enter(&dp
->dp_lock
);
76 ksp
->ks_ndata
= dp
->dp_txg_history_size
;
77 ksp
->ks_data_size
= dp
->dp_txg_history_size
* sizeof(kstat_txg_t
);
78 if (ksp
->ks_data_size
> 0)
79 ksp
->ks_data
= kmem_alloc(ksp
->ks_data_size
, KM_PUSHPAGE
);
81 /* Traversed oldest to youngest for the most readable kstat output */
82 for (th
= list_tail(&dp
->dp_txg_history
); th
!= NULL
;
83 th
= list_prev(&dp
->dp_txg_history
, th
)) {
84 mutex_enter(&th
->th_lock
);
85 ASSERT3S(i
+ sizeof(kstat_txg_t
), <=, ksp
->ks_data_size
);
86 memcpy(ksp
->ks_data
+ i
, &th
->th_kstat
, sizeof(kstat_txg_t
));
87 i
+= sizeof(kstat_txg_t
);
88 mutex_exit(&th
->th_lock
);
91 mutex_exit(&dp
->dp_lock
);
97 dsl_pool_txg_history_init(dsl_pool_t
*dp
, uint64_t txg
)
99 char name
[KSTAT_STRLEN
];
101 list_create(&dp
->dp_txg_history
, sizeof (txg_history_t
),
102 offsetof(txg_history_t
, th_link
));
103 dsl_pool_txg_history_add(dp
, txg
);
105 (void) snprintf(name
, KSTAT_STRLEN
, "txgs-%s", spa_name(dp
->dp_spa
));
106 dp
->dp_txg_kstat
= kstat_create("zfs", 0, name
, "misc",
107 KSTAT_TYPE_TXG
, 0, KSTAT_FLAG_VIRTUAL
);
108 if (dp
->dp_txg_kstat
) {
109 dp
->dp_txg_kstat
->ks_data
= NULL
;
110 dp
->dp_txg_kstat
->ks_private
= dp
;
111 dp
->dp_txg_kstat
->ks_update
= dsl_pool_txg_history_update
;
112 kstat_install(dp
->dp_txg_kstat
);
117 dsl_pool_txg_history_destroy(dsl_pool_t
*dp
)
121 if (dp
->dp_txg_kstat
) {
122 if (dp
->dp_txg_kstat
->ks_data
)
123 kmem_free(dp
->dp_txg_kstat
->ks_data
,
124 dp
->dp_txg_kstat
->ks_data_size
);
126 kstat_delete(dp
->dp_txg_kstat
);
129 mutex_enter(&dp
->dp_lock
);
130 while ((th
= list_remove_head(&dp
->dp_txg_history
))) {
131 dp
->dp_txg_history_size
--;
132 mutex_destroy(&th
->th_lock
);
133 kmem_free(th
, sizeof(txg_history_t
));
136 ASSERT3U(dp
->dp_txg_history_size
, ==, 0);
137 list_destroy(&dp
->dp_txg_history
);
138 mutex_exit(&dp
->dp_lock
);
142 dsl_pool_txg_history_add(dsl_pool_t
*dp
, uint64_t txg
)
144 txg_history_t
*th
, *rm
;
146 th
= kmem_zalloc(sizeof(txg_history_t
), KM_SLEEP
);
147 mutex_init(&th
->th_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
148 th
->th_kstat
.txg
= txg
;
149 th
->th_kstat
.state
= TXG_STATE_OPEN
;
150 th
->th_kstat
.birth
= gethrtime();
152 mutex_enter(&dp
->dp_lock
);
154 list_insert_head(&dp
->dp_txg_history
, th
);
155 dp
->dp_txg_history_size
++;
157 while (dp
->dp_txg_history_size
> zfs_txg_history
) {
158 dp
->dp_txg_history_size
--;
159 rm
= list_remove_tail(&dp
->dp_txg_history
);
160 mutex_destroy(&rm
->th_lock
);
161 kmem_free(rm
, sizeof(txg_history_t
));
164 mutex_exit(&dp
->dp_lock
);
170 * Traversed youngest to oldest because lookups are only done for open
171 * or syncing txgs which are guaranteed to be at the head of the list.
172 * The txg_history_t structure will be returned locked.
175 dsl_pool_txg_history_get(dsl_pool_t
*dp
, uint64_t txg
)
179 mutex_enter(&dp
->dp_lock
);
180 for (th
= list_head(&dp
->dp_txg_history
); th
!= NULL
;
181 th
= list_next(&dp
->dp_txg_history
, th
)) {
182 if (th
->th_kstat
.txg
== txg
) {
183 mutex_enter(&th
->th_lock
);
187 mutex_exit(&dp
->dp_lock
);
193 dsl_pool_txg_history_put(txg_history_t
*th
)
195 mutex_exit(&th
->th_lock
);
199 dsl_pool_open_special_dir(dsl_pool_t
*dp
, const char *name
, dsl_dir_t
**ddp
)
204 err
= zap_lookup(dp
->dp_meta_objset
,
205 dp
->dp_root_dir
->dd_phys
->dd_child_dir_zapobj
,
206 name
, sizeof (obj
), 1, &obj
);
210 return (dsl_dir_open_obj(dp
, obj
, name
, dp
, ddp
));
214 dsl_pool_open_impl(spa_t
*spa
, uint64_t txg
)
217 blkptr_t
*bp
= spa_get_rootblkptr(spa
);
219 dp
= kmem_zalloc(sizeof (dsl_pool_t
), KM_SLEEP
);
221 dp
->dp_meta_rootbp
= *bp
;
222 rw_init(&dp
->dp_config_rwlock
, NULL
, RW_DEFAULT
, NULL
);
223 dp
->dp_write_limit
= zfs_write_limit_min
;
226 txg_list_create(&dp
->dp_dirty_datasets
,
227 offsetof(dsl_dataset_t
, ds_dirty_link
));
228 txg_list_create(&dp
->dp_dirty_zilogs
,
229 offsetof(zilog_t
, zl_dirty_link
));
230 txg_list_create(&dp
->dp_dirty_dirs
,
231 offsetof(dsl_dir_t
, dd_dirty_link
));
232 txg_list_create(&dp
->dp_sync_tasks
,
233 offsetof(dsl_sync_task_group_t
, dstg_node
));
235 mutex_init(&dp
->dp_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
237 dp
->dp_iput_taskq
= taskq_create("zfs_iput_taskq", 1, minclsyspri
,
240 dsl_pool_txg_history_init(dp
, txg
);
246 dsl_pool_init(spa_t
*spa
, uint64_t txg
, dsl_pool_t
**dpp
)
249 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
251 err
= dmu_objset_open_impl(spa
, NULL
, &dp
->dp_meta_rootbp
,
252 &dp
->dp_meta_objset
);
262 dsl_pool_open(dsl_pool_t
*dp
)
269 rw_enter(&dp
->dp_config_rwlock
, RW_WRITER
);
270 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
271 DMU_POOL_ROOT_DATASET
, sizeof (uint64_t), 1,
272 &dp
->dp_root_dir_obj
);
276 err
= dsl_dir_open_obj(dp
, dp
->dp_root_dir_obj
,
277 NULL
, dp
, &dp
->dp_root_dir
);
281 err
= dsl_pool_open_special_dir(dp
, MOS_DIR_NAME
, &dp
->dp_mos_dir
);
285 if (spa_version(dp
->dp_spa
) >= SPA_VERSION_ORIGIN
) {
286 err
= dsl_pool_open_special_dir(dp
, ORIGIN_DIR_NAME
, &dd
);
289 err
= dsl_dataset_hold_obj(dp
, dd
->dd_phys
->dd_head_dataset_obj
,
292 err
= dsl_dataset_hold_obj(dp
,
293 ds
->ds_phys
->ds_prev_snap_obj
, dp
,
294 &dp
->dp_origin_snap
);
295 dsl_dataset_rele(ds
, FTAG
);
297 dsl_dir_close(dd
, dp
);
302 if (spa_version(dp
->dp_spa
) >= SPA_VERSION_DEADLISTS
) {
303 err
= dsl_pool_open_special_dir(dp
, FREE_DIR_NAME
,
308 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
309 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
);
312 VERIFY3U(0, ==, bpobj_open(&dp
->dp_free_bpobj
,
313 dp
->dp_meta_objset
, obj
));
316 if (spa_feature_is_active(dp
->dp_spa
,
317 &spa_feature_table
[SPA_FEATURE_ASYNC_DESTROY
])) {
318 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
319 DMU_POOL_BPTREE_OBJ
, sizeof (uint64_t), 1,
325 if (spa_feature_is_active(dp
->dp_spa
,
326 &spa_feature_table
[SPA_FEATURE_EMPTY_BPOBJ
])) {
327 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
328 DMU_POOL_EMPTY_BPOBJ
, sizeof (uint64_t), 1,
329 &dp
->dp_empty_bpobj
);
334 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
335 DMU_POOL_TMP_USERREFS
, sizeof (uint64_t), 1,
336 &dp
->dp_tmp_userrefs_obj
);
342 err
= dsl_scan_init(dp
, dp
->dp_tx
.tx_open_txg
);
345 rw_exit(&dp
->dp_config_rwlock
);
350 dsl_pool_close(dsl_pool_t
*dp
)
352 /* drop our references from dsl_pool_open() */
355 * Since we held the origin_snap from "syncing" context (which
356 * includes pool-opening context), it actually only got a "ref"
357 * and not a hold, so just drop that here.
359 if (dp
->dp_origin_snap
)
360 dsl_dataset_drop_ref(dp
->dp_origin_snap
, dp
);
362 dsl_dir_close(dp
->dp_mos_dir
, dp
);
364 dsl_dir_close(dp
->dp_free_dir
, dp
);
366 dsl_dir_close(dp
->dp_root_dir
, dp
);
368 bpobj_close(&dp
->dp_free_bpobj
);
370 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
371 if (dp
->dp_meta_objset
)
372 dmu_objset_evict(dp
->dp_meta_objset
);
374 txg_list_destroy(&dp
->dp_dirty_datasets
);
375 txg_list_destroy(&dp
->dp_dirty_zilogs
);
376 txg_list_destroy(&dp
->dp_sync_tasks
);
377 txg_list_destroy(&dp
->dp_dirty_dirs
);
379 arc_flush(dp
->dp_spa
);
382 dsl_pool_txg_history_destroy(dp
);
383 rw_destroy(&dp
->dp_config_rwlock
);
384 mutex_destroy(&dp
->dp_lock
);
385 taskq_destroy(dp
->dp_iput_taskq
);
387 kmem_free(dp
->dp_blkstats
, sizeof (zfs_all_blkstats_t
));
388 kmem_free(dp
, sizeof (dsl_pool_t
));
392 dsl_pool_create(spa_t
*spa
, nvlist_t
*zplprops
, uint64_t txg
)
395 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
396 dmu_tx_t
*tx
= dmu_tx_create_assigned(dp
, txg
);
401 /* create and open the MOS (meta-objset) */
402 dp
->dp_meta_objset
= dmu_objset_create_impl(spa
,
403 NULL
, &dp
->dp_meta_rootbp
, DMU_OST_META
, tx
);
405 /* create the pool directory */
406 err
= zap_create_claim(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
407 DMU_OT_OBJECT_DIRECTORY
, DMU_OT_NONE
, 0, tx
);
408 ASSERT3U(err
, ==, 0);
410 /* Initialize scan structures */
411 VERIFY3U(0, ==, dsl_scan_init(dp
, txg
));
413 /* create and open the root dir */
414 dp
->dp_root_dir_obj
= dsl_dir_create_sync(dp
, NULL
, NULL
, tx
);
415 VERIFY(0 == dsl_dir_open_obj(dp
, dp
->dp_root_dir_obj
,
416 NULL
, dp
, &dp
->dp_root_dir
));
418 /* create and open the meta-objset dir */
419 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, MOS_DIR_NAME
, tx
);
420 VERIFY(0 == dsl_pool_open_special_dir(dp
,
421 MOS_DIR_NAME
, &dp
->dp_mos_dir
));
423 if (spa_version(spa
) >= SPA_VERSION_DEADLISTS
) {
424 /* create and open the free dir */
425 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
,
427 VERIFY(0 == dsl_pool_open_special_dir(dp
,
428 FREE_DIR_NAME
, &dp
->dp_free_dir
));
430 /* create and open the free_bplist */
431 obj
= bpobj_alloc(dp
->dp_meta_objset
, SPA_MAXBLOCKSIZE
, tx
);
432 VERIFY(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
433 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
) == 0);
434 VERIFY3U(0, ==, bpobj_open(&dp
->dp_free_bpobj
,
435 dp
->dp_meta_objset
, obj
));
438 if (spa_version(spa
) >= SPA_VERSION_DSL_SCRUB
)
439 dsl_pool_create_origin(dp
, tx
);
441 /* create the root dataset */
442 obj
= dsl_dataset_create_sync_dd(dp
->dp_root_dir
, NULL
, 0, tx
);
444 /* create the root objset */
445 VERIFY(0 == dsl_dataset_hold_obj(dp
, obj
, FTAG
, &ds
));
446 VERIFY(NULL
!= (os
= dmu_objset_create_impl(dp
->dp_spa
, ds
,
447 dsl_dataset_get_blkptr(ds
), DMU_OST_ZFS
, tx
)));
449 zfs_create_fs(os
, kcred
, zplprops
, tx
);
451 dsl_dataset_rele(ds
, FTAG
);
459 * Account for the meta-objset space in its placeholder dsl_dir.
462 dsl_pool_mos_diduse_space(dsl_pool_t
*dp
,
463 int64_t used
, int64_t comp
, int64_t uncomp
)
465 ASSERT3U(comp
, ==, uncomp
); /* it's all metadata */
466 mutex_enter(&dp
->dp_lock
);
467 dp
->dp_mos_used_delta
+= used
;
468 dp
->dp_mos_compressed_delta
+= comp
;
469 dp
->dp_mos_uncompressed_delta
+= uncomp
;
470 mutex_exit(&dp
->dp_lock
);
474 deadlist_enqueue_cb(void *arg
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
476 dsl_deadlist_t
*dl
= arg
;
477 dsl_pool_t
*dp
= dmu_objset_pool(dl
->dl_os
);
478 rw_enter(&dp
->dp_config_rwlock
, RW_READER
);
479 dsl_deadlist_insert(dl
, bp
, tx
);
480 rw_exit(&dp
->dp_config_rwlock
);
485 dsl_pool_sync(dsl_pool_t
*dp
, uint64_t txg
)
491 objset_t
*mos
= dp
->dp_meta_objset
;
492 hrtime_t start
, write_time
;
493 uint64_t data_written
;
495 list_t synced_datasets
;
497 list_create(&synced_datasets
, sizeof (dsl_dataset_t
),
498 offsetof(dsl_dataset_t
, ds_synced_link
));
501 * We need to copy dp_space_towrite() before doing
502 * dsl_sync_task_group_sync(), because
503 * dsl_dataset_snapshot_reserve_space() will increase
504 * dp_space_towrite but not actually write anything.
506 data_written
= dp
->dp_space_towrite
[txg
& TXG_MASK
];
508 tx
= dmu_tx_create_assigned(dp
, txg
);
510 dp
->dp_read_overhead
= 0;
513 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
514 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
))) {
516 * We must not sync any non-MOS datasets twice, because
517 * we may have taken a snapshot of them. However, we
518 * may sync newly-created datasets on pass 2.
520 ASSERT(!list_link_active(&ds
->ds_synced_link
));
521 list_insert_tail(&synced_datasets
, ds
);
522 dsl_dataset_sync(ds
, zio
, tx
);
524 DTRACE_PROBE(pool_sync__1setup
);
527 write_time
= gethrtime() - start
;
529 DTRACE_PROBE(pool_sync__2rootzio
);
532 * After the data blocks have been written (ensured by the zio_wait()
533 * above), update the user/group space accounting.
535 for (ds
= list_head(&synced_datasets
); ds
;
536 ds
= list_next(&synced_datasets
, ds
))
537 dmu_objset_do_userquota_updates(ds
->ds_objset
, tx
);
540 * Sync the datasets again to push out the changes due to
541 * userspace updates. This must be done before we process the
542 * sync tasks, so that any snapshots will have the correct
543 * user accounting information (and we won't get confused
544 * about which blocks are part of the snapshot).
546 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
547 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
))) {
548 ASSERT(list_link_active(&ds
->ds_synced_link
));
549 dmu_buf_rele(ds
->ds_dbuf
, ds
);
550 dsl_dataset_sync(ds
, zio
, tx
);
555 * Now that the datasets have been completely synced, we can
556 * clean up our in-memory structures accumulated while syncing:
558 * - move dead blocks from the pending deadlist to the on-disk deadlist
559 * - clean up zil records
560 * - release hold from dsl_dataset_dirty()
562 while ((ds
= list_remove_head(&synced_datasets
))) {
563 ASSERTV(objset_t
*os
= ds
->ds_objset
);
564 bplist_iterate(&ds
->ds_pending_deadlist
,
565 deadlist_enqueue_cb
, &ds
->ds_deadlist
, tx
);
566 ASSERT(!dmu_objset_is_dirty(os
, txg
));
567 dmu_buf_rele(ds
->ds_dbuf
, ds
);
571 while ((dd
= txg_list_remove(&dp
->dp_dirty_dirs
, txg
)))
572 dsl_dir_sync(dd
, tx
);
573 write_time
+= gethrtime() - start
;
576 * The MOS's space is accounted for in the pool/$MOS
577 * (dp_mos_dir). We can't modify the mos while we're syncing
578 * it, so we remember the deltas and apply them here.
580 if (dp
->dp_mos_used_delta
!= 0 || dp
->dp_mos_compressed_delta
!= 0 ||
581 dp
->dp_mos_uncompressed_delta
!= 0) {
582 dsl_dir_diduse_space(dp
->dp_mos_dir
, DD_USED_HEAD
,
583 dp
->dp_mos_used_delta
,
584 dp
->dp_mos_compressed_delta
,
585 dp
->dp_mos_uncompressed_delta
, tx
);
586 dp
->dp_mos_used_delta
= 0;
587 dp
->dp_mos_compressed_delta
= 0;
588 dp
->dp_mos_uncompressed_delta
= 0;
592 if (list_head(&mos
->os_dirty_dnodes
[txg
& TXG_MASK
]) != NULL
||
593 list_head(&mos
->os_free_dnodes
[txg
& TXG_MASK
]) != NULL
) {
594 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
595 dmu_objset_sync(mos
, zio
, tx
);
598 dprintf_bp(&dp
->dp_meta_rootbp
, "meta objset rootbp is %s", "");
599 spa_set_rootblkptr(dp
->dp_spa
, &dp
->dp_meta_rootbp
);
601 write_time
+= gethrtime() - start
;
602 DTRACE_PROBE2(pool_sync__4io
, hrtime_t
, write_time
,
603 hrtime_t
, dp
->dp_read_overhead
);
604 write_time
-= dp
->dp_read_overhead
;
607 * If we modify a dataset in the same txg that we want to destroy it,
608 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
609 * dsl_dir_destroy_check() will fail if there are unexpected holds.
610 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
611 * and clearing the hold on it) before we process the sync_tasks.
612 * The MOS data dirtied by the sync_tasks will be synced on the next
615 DTRACE_PROBE(pool_sync__3task
);
616 if (!txg_list_empty(&dp
->dp_sync_tasks
, txg
)) {
617 dsl_sync_task_group_t
*dstg
;
619 * No more sync tasks should have been added while we
622 ASSERT(spa_sync_pass(dp
->dp_spa
) == 1);
623 while ((dstg
= txg_list_remove(&dp
->dp_sync_tasks
, txg
)))
624 dsl_sync_task_group_sync(dstg
, tx
);
629 dp
->dp_space_towrite
[txg
& TXG_MASK
] = 0;
630 ASSERT(dp
->dp_tempreserved
[txg
& TXG_MASK
] == 0);
633 * If the write limit max has not been explicitly set, set it
634 * to a fraction of available physical memory (default 1/8th).
635 * Note that we must inflate the limit because the spa
636 * inflates write sizes to account for data replication.
637 * Check this each sync phase to catch changing memory size.
639 if (physmem
!= old_physmem
&& zfs_write_limit_shift
) {
640 mutex_enter(&zfs_write_limit_lock
);
641 old_physmem
= physmem
;
642 zfs_write_limit_max
= ptob(physmem
) >> zfs_write_limit_shift
;
643 zfs_write_limit_inflated
= MAX(zfs_write_limit_min
,
644 spa_get_asize(dp
->dp_spa
, zfs_write_limit_max
));
645 mutex_exit(&zfs_write_limit_lock
);
649 * Attempt to keep the sync time consistent by adjusting the
650 * amount of write traffic allowed into each transaction group.
651 * Weight the throughput calculation towards the current value:
652 * thru = 3/4 old_thru + 1/4 new_thru
654 * Note: write_time is in nanosecs, so write_time/MICROSEC
657 ASSERT(zfs_write_limit_min
> 0);
658 if (data_written
> zfs_write_limit_min
/ 8 && write_time
> MICROSEC
) {
659 uint64_t throughput
= data_written
/ (write_time
/ MICROSEC
);
661 if (dp
->dp_throughput
)
662 dp
->dp_throughput
= throughput
/ 4 +
663 3 * dp
->dp_throughput
/ 4;
665 dp
->dp_throughput
= throughput
;
666 dp
->dp_write_limit
= MIN(zfs_write_limit_inflated
,
667 MAX(zfs_write_limit_min
,
668 dp
->dp_throughput
* zfs_txg_synctime_ms
));
673 dsl_pool_sync_done(dsl_pool_t
*dp
, uint64_t txg
)
678 while ((zilog
= txg_list_remove(&dp
->dp_dirty_zilogs
, txg
))) {
679 ds
= dmu_objset_ds(zilog
->zl_os
);
680 zil_clean(zilog
, txg
);
681 ASSERT(!dmu_objset_is_dirty(zilog
->zl_os
, txg
));
682 dmu_buf_rele(ds
->ds_dbuf
, zilog
);
684 ASSERT(!dmu_objset_is_dirty(dp
->dp_meta_objset
, txg
));
688 * TRUE if the current thread is the tx_sync_thread or if we
689 * are being called from SPA context during pool initialization.
692 dsl_pool_sync_context(dsl_pool_t
*dp
)
694 return (curthread
== dp
->dp_tx
.tx_sync_thread
||
695 spa_is_initializing(dp
->dp_spa
));
699 dsl_pool_adjustedsize(dsl_pool_t
*dp
, boolean_t netfree
)
701 uint64_t space
, resv
;
704 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
706 * XXX The intent log is not accounted for, so it must fit
709 * If we're trying to assess whether it's OK to do a free,
710 * cut the reservation in half to allow forward progress
711 * (e.g. make it possible to rm(1) files from a full pool).
713 space
= spa_get_dspace(dp
->dp_spa
);
714 resv
= MAX(space
>> 6, SPA_MINDEVSIZE
>> 1);
718 return (space
- resv
);
722 dsl_pool_tempreserve_space(dsl_pool_t
*dp
, uint64_t space
, dmu_tx_t
*tx
)
724 uint64_t reserved
= 0;
725 uint64_t write_limit
= (zfs_write_limit_override
?
726 zfs_write_limit_override
: dp
->dp_write_limit
);
728 if (zfs_no_write_throttle
) {
729 atomic_add_64(&dp
->dp_tempreserved
[tx
->tx_txg
& TXG_MASK
],
735 * Check to see if we have exceeded the maximum allowed IO for
736 * this transaction group. We can do this without locks since
737 * a little slop here is ok. Note that we do the reserved check
738 * with only half the requested reserve: this is because the
739 * reserve requests are worst-case, and we really don't want to
740 * throttle based off of worst-case estimates.
742 if (write_limit
> 0) {
743 reserved
= dp
->dp_space_towrite
[tx
->tx_txg
& TXG_MASK
]
744 + dp
->dp_tempreserved
[tx
->tx_txg
& TXG_MASK
] / 2;
746 if (reserved
&& reserved
> write_limit
) {
747 DMU_TX_STAT_BUMP(dmu_tx_write_limit
);
752 atomic_add_64(&dp
->dp_tempreserved
[tx
->tx_txg
& TXG_MASK
], space
);
755 * If this transaction group is over 7/8ths capacity, delay
756 * the caller 1 clock tick. This will slow down the "fill"
757 * rate until the sync process can catch up with us.
759 if (reserved
&& reserved
> (write_limit
- (write_limit
>> 3)))
760 txg_delay(dp
, tx
->tx_txg
, 1);
766 dsl_pool_tempreserve_clear(dsl_pool_t
*dp
, int64_t space
, dmu_tx_t
*tx
)
768 ASSERT(dp
->dp_tempreserved
[tx
->tx_txg
& TXG_MASK
] >= space
);
769 atomic_add_64(&dp
->dp_tempreserved
[tx
->tx_txg
& TXG_MASK
], -space
);
773 dsl_pool_memory_pressure(dsl_pool_t
*dp
)
775 uint64_t space_inuse
= 0;
778 if (dp
->dp_write_limit
== zfs_write_limit_min
)
781 for (i
= 0; i
< TXG_SIZE
; i
++) {
782 space_inuse
+= dp
->dp_space_towrite
[i
];
783 space_inuse
+= dp
->dp_tempreserved
[i
];
785 dp
->dp_write_limit
= MAX(zfs_write_limit_min
,
786 MIN(dp
->dp_write_limit
, space_inuse
/ 4));
790 dsl_pool_willuse_space(dsl_pool_t
*dp
, int64_t space
, dmu_tx_t
*tx
)
793 mutex_enter(&dp
->dp_lock
);
794 dp
->dp_space_towrite
[tx
->tx_txg
& TXG_MASK
] += space
;
795 mutex_exit(&dp
->dp_lock
);
801 upgrade_clones_cb(spa_t
*spa
, uint64_t dsobj
, const char *dsname
, void *arg
)
804 dsl_dataset_t
*ds
, *prev
= NULL
;
806 dsl_pool_t
*dp
= spa_get_dsl(spa
);
808 err
= dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
);
812 while (ds
->ds_phys
->ds_prev_snap_obj
!= 0) {
813 err
= dsl_dataset_hold_obj(dp
, ds
->ds_phys
->ds_prev_snap_obj
,
816 dsl_dataset_rele(ds
, FTAG
);
820 if (prev
->ds_phys
->ds_next_snap_obj
!= ds
->ds_object
)
822 dsl_dataset_rele(ds
, FTAG
);
828 prev
= dp
->dp_origin_snap
;
831 * The $ORIGIN can't have any data, or the accounting
834 ASSERT(prev
->ds_phys
->ds_bp
.blk_birth
== 0);
836 /* The origin doesn't get attached to itself */
837 if (ds
->ds_object
== prev
->ds_object
) {
838 dsl_dataset_rele(ds
, FTAG
);
842 dmu_buf_will_dirty(ds
->ds_dbuf
, tx
);
843 ds
->ds_phys
->ds_prev_snap_obj
= prev
->ds_object
;
844 ds
->ds_phys
->ds_prev_snap_txg
= prev
->ds_phys
->ds_creation_txg
;
846 dmu_buf_will_dirty(ds
->ds_dir
->dd_dbuf
, tx
);
847 ds
->ds_dir
->dd_phys
->dd_origin_obj
= prev
->ds_object
;
849 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
850 prev
->ds_phys
->ds_num_children
++;
852 if (ds
->ds_phys
->ds_next_snap_obj
== 0) {
853 ASSERT(ds
->ds_prev
== NULL
);
854 VERIFY(0 == dsl_dataset_hold_obj(dp
,
855 ds
->ds_phys
->ds_prev_snap_obj
, ds
, &ds
->ds_prev
));
859 ASSERT(ds
->ds_dir
->dd_phys
->dd_origin_obj
== prev
->ds_object
);
860 ASSERT(ds
->ds_phys
->ds_prev_snap_obj
== prev
->ds_object
);
862 if (prev
->ds_phys
->ds_next_clones_obj
== 0) {
863 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
864 prev
->ds_phys
->ds_next_clones_obj
=
865 zap_create(dp
->dp_meta_objset
,
866 DMU_OT_NEXT_CLONES
, DMU_OT_NONE
, 0, tx
);
868 VERIFY(0 == zap_add_int(dp
->dp_meta_objset
,
869 prev
->ds_phys
->ds_next_clones_obj
, ds
->ds_object
, tx
));
871 dsl_dataset_rele(ds
, FTAG
);
872 if (prev
!= dp
->dp_origin_snap
)
873 dsl_dataset_rele(prev
, FTAG
);
878 dsl_pool_upgrade_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
880 ASSERT(dmu_tx_is_syncing(tx
));
881 ASSERT(dp
->dp_origin_snap
!= NULL
);
883 VERIFY3U(0, ==, dmu_objset_find_spa(dp
->dp_spa
, NULL
, upgrade_clones_cb
,
884 tx
, DS_FIND_CHILDREN
));
889 upgrade_dir_clones_cb(spa_t
*spa
, uint64_t dsobj
, const char *dsname
, void *arg
)
893 dsl_pool_t
*dp
= spa_get_dsl(spa
);
894 objset_t
*mos
= dp
->dp_meta_objset
;
896 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
898 if (ds
->ds_dir
->dd_phys
->dd_origin_obj
) {
899 dsl_dataset_t
*origin
;
901 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp
,
902 ds
->ds_dir
->dd_phys
->dd_origin_obj
, FTAG
, &origin
));
904 if (origin
->ds_dir
->dd_phys
->dd_clones
== 0) {
905 dmu_buf_will_dirty(origin
->ds_dir
->dd_dbuf
, tx
);
906 origin
->ds_dir
->dd_phys
->dd_clones
= zap_create(mos
,
907 DMU_OT_DSL_CLONES
, DMU_OT_NONE
, 0, tx
);
910 VERIFY3U(0, ==, zap_add_int(dp
->dp_meta_objset
,
911 origin
->ds_dir
->dd_phys
->dd_clones
, dsobj
, tx
));
913 dsl_dataset_rele(origin
, FTAG
);
916 dsl_dataset_rele(ds
, FTAG
);
921 dsl_pool_upgrade_dir_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
925 ASSERT(dmu_tx_is_syncing(tx
));
927 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, FREE_DIR_NAME
, tx
);
928 VERIFY(0 == dsl_pool_open_special_dir(dp
,
929 FREE_DIR_NAME
, &dp
->dp_free_dir
));
932 * We can't use bpobj_alloc(), because spa_version() still
933 * returns the old version, and we need a new-version bpobj with
934 * subobj support. So call dmu_object_alloc() directly.
936 obj
= dmu_object_alloc(dp
->dp_meta_objset
, DMU_OT_BPOBJ
,
937 SPA_MAXBLOCKSIZE
, DMU_OT_BPOBJ_HDR
, sizeof (bpobj_phys_t
), tx
);
938 VERIFY3U(0, ==, zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
939 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
));
940 VERIFY3U(0, ==, bpobj_open(&dp
->dp_free_bpobj
,
941 dp
->dp_meta_objset
, obj
));
943 VERIFY3U(0, ==, dmu_objset_find_spa(dp
->dp_spa
, NULL
,
944 upgrade_dir_clones_cb
, tx
, DS_FIND_CHILDREN
));
948 dsl_pool_create_origin(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
953 ASSERT(dmu_tx_is_syncing(tx
));
954 ASSERT(dp
->dp_origin_snap
== NULL
);
956 /* create the origin dir, ds, & snap-ds */
957 rw_enter(&dp
->dp_config_rwlock
, RW_WRITER
);
958 dsobj
= dsl_dataset_create_sync(dp
->dp_root_dir
, ORIGIN_DIR_NAME
,
960 VERIFY(0 == dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
961 dsl_dataset_snapshot_sync(ds
, ORIGIN_DIR_NAME
, tx
);
962 VERIFY(0 == dsl_dataset_hold_obj(dp
, ds
->ds_phys
->ds_prev_snap_obj
,
963 dp
, &dp
->dp_origin_snap
));
964 dsl_dataset_rele(ds
, FTAG
);
965 rw_exit(&dp
->dp_config_rwlock
);
969 dsl_pool_iput_taskq(dsl_pool_t
*dp
)
971 return (dp
->dp_iput_taskq
);
975 * Walk through the pool-wide zap object of temporary snapshot user holds
979 dsl_pool_clean_tmp_userrefs(dsl_pool_t
*dp
)
983 objset_t
*mos
= dp
->dp_meta_objset
;
984 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
988 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
990 for (zap_cursor_init(&zc
, mos
, zapobj
);
991 zap_cursor_retrieve(&zc
, &za
) == 0;
992 zap_cursor_advance(&zc
)) {
996 htag
= strchr(za
.za_name
, '-');
999 dsobj
= strtonum(za
.za_name
, NULL
);
1000 (void) dsl_dataset_user_release_tmp(dp
, dsobj
, htag
, B_FALSE
);
1002 zap_cursor_fini(&zc
);
1006 * Create the pool-wide zap object for storing temporary snapshot holds.
1009 dsl_pool_user_hold_create_obj(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
1011 objset_t
*mos
= dp
->dp_meta_objset
;
1013 ASSERT(dp
->dp_tmp_userrefs_obj
== 0);
1014 ASSERT(dmu_tx_is_syncing(tx
));
1016 dp
->dp_tmp_userrefs_obj
= zap_create_link(mos
, DMU_OT_USERREFS
,
1017 DMU_POOL_DIRECTORY_OBJECT
, DMU_POOL_TMP_USERREFS
, tx
);
1021 dsl_pool_user_hold_rele_impl(dsl_pool_t
*dp
, uint64_t dsobj
,
1022 const char *tag
, uint64_t *now
, dmu_tx_t
*tx
, boolean_t holding
)
1024 objset_t
*mos
= dp
->dp_meta_objset
;
1025 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
1029 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
1030 ASSERT(dmu_tx_is_syncing(tx
));
1033 * If the pool was created prior to SPA_VERSION_USERREFS, the
1034 * zap object for temporary holds might not exist yet.
1038 dsl_pool_user_hold_create_obj(dp
, tx
);
1039 zapobj
= dp
->dp_tmp_userrefs_obj
;
1045 name
= kmem_asprintf("%llx-%s", (u_longlong_t
)dsobj
, tag
);
1047 error
= zap_add(mos
, zapobj
, name
, 8, 1, now
, tx
);
1049 error
= zap_remove(mos
, zapobj
, name
, tx
);
1056 * Add a temporary hold for the given dataset object and tag.
1059 dsl_pool_user_hold(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
1060 uint64_t *now
, dmu_tx_t
*tx
)
1062 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, now
, tx
, B_TRUE
));
1066 * Release a temporary hold for the given dataset object and tag.
1069 dsl_pool_user_release(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
1072 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, NULL
,
1076 #if defined(_KERNEL) && defined(HAVE_SPL)
1077 module_param(zfs_no_write_throttle
, int, 0644);
1078 MODULE_PARM_DESC(zfs_no_write_throttle
, "Disable write throttling");
1080 module_param(zfs_write_limit_shift
, int, 0444);
1081 MODULE_PARM_DESC(zfs_write_limit_shift
, "log2(fraction of memory) per txg");
1083 module_param(zfs_txg_synctime_ms
, int, 0644);
1084 MODULE_PARM_DESC(zfs_txg_synctime_ms
, "Target milliseconds between txg sync");
1086 module_param(zfs_txg_history
, int, 0644);
1087 MODULE_PARM_DESC(zfs_txg_history
, "Historic statistics for the last N txgs");
1089 module_param(zfs_write_limit_min
, ulong
, 0444);
1090 MODULE_PARM_DESC(zfs_write_limit_min
, "Min txg write limit");
1092 module_param(zfs_write_limit_max
, ulong
, 0444);
1093 MODULE_PARM_DESC(zfs_write_limit_max
, "Max txg write limit");
1095 module_param(zfs_write_limit_inflated
, ulong
, 0444);
1096 MODULE_PARM_DESC(zfs_write_limit_inflated
, "Inflated txg write limit");
1098 module_param(zfs_write_limit_override
, ulong
, 0444);
1099 MODULE_PARM_DESC(zfs_write_limit_override
, "Override txg write limit");