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 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>
44 int zfs_no_write_throttle
= 0;
45 int zfs_write_limit_shift
= 3; /* 1/8th of physical memory */
46 int zfs_txg_synctime_ms
= 1000; /* target millisecs to sync a txg */
47 int zfs_txg_history
= 60; /* statistics for the last N txgs */
49 unsigned long zfs_write_limit_min
= 32 << 20; /* min write limit is 32MB */
50 unsigned long zfs_write_limit_max
= 0; /* max data payload per txg */
51 unsigned long zfs_write_limit_inflated
= 0;
52 unsigned long zfs_write_limit_override
= 0;
54 kmutex_t zfs_write_limit_lock
;
56 static pgcnt_t old_physmem
= 0;
59 dsl_pool_txg_history_update(kstat_t
*ksp
, int rw
)
61 dsl_pool_t
*dp
= ksp
->ks_private
;
65 if (rw
== KSTAT_WRITE
)
69 kmem_free(ksp
->ks_data
, ksp
->ks_data_size
);
71 mutex_enter(&dp
->dp_lock
);
73 ksp
->ks_ndata
= dp
->dp_txg_history_size
;
74 ksp
->ks_data_size
= dp
->dp_txg_history_size
* sizeof(kstat_txg_t
);
75 if (ksp
->ks_data_size
> 0)
76 ksp
->ks_data
= kmem_alloc(ksp
->ks_data_size
, KM_PUSHPAGE
);
78 /* Traversed oldest to youngest for the most readable kstat output */
79 for (th
= list_tail(&dp
->dp_txg_history
); th
!= NULL
;
80 th
= list_prev(&dp
->dp_txg_history
, th
)) {
81 mutex_enter(&th
->th_lock
);
82 ASSERT3S(i
+ sizeof(kstat_txg_t
), <=, ksp
->ks_data_size
);
83 memcpy(ksp
->ks_data
+ i
, &th
->th_kstat
, sizeof(kstat_txg_t
));
84 i
+= sizeof(kstat_txg_t
);
85 mutex_exit(&th
->th_lock
);
88 mutex_exit(&dp
->dp_lock
);
94 dsl_pool_txg_history_init(dsl_pool_t
*dp
, uint64_t txg
)
96 char name
[KSTAT_STRLEN
];
98 list_create(&dp
->dp_txg_history
, sizeof (txg_history_t
),
99 offsetof(txg_history_t
, th_link
));
100 dsl_pool_txg_history_add(dp
, txg
);
102 (void) snprintf(name
, KSTAT_STRLEN
, "txgs-%s", spa_name(dp
->dp_spa
));
103 dp
->dp_txg_kstat
= kstat_create("zfs", 0, name
, "misc",
104 KSTAT_TYPE_TXG
, 0, KSTAT_FLAG_VIRTUAL
);
105 if (dp
->dp_txg_kstat
) {
106 dp
->dp_txg_kstat
->ks_data
= NULL
;
107 dp
->dp_txg_kstat
->ks_private
= dp
;
108 dp
->dp_txg_kstat
->ks_update
= dsl_pool_txg_history_update
;
109 kstat_install(dp
->dp_txg_kstat
);
114 dsl_pool_txg_history_destroy(dsl_pool_t
*dp
)
118 if (dp
->dp_txg_kstat
) {
119 if (dp
->dp_txg_kstat
->ks_data
)
120 kmem_free(dp
->dp_txg_kstat
->ks_data
,
121 dp
->dp_txg_kstat
->ks_data_size
);
123 kstat_delete(dp
->dp_txg_kstat
);
126 mutex_enter(&dp
->dp_lock
);
127 while ((th
= list_remove_head(&dp
->dp_txg_history
))) {
128 dp
->dp_txg_history_size
--;
129 mutex_destroy(&th
->th_lock
);
130 kmem_free(th
, sizeof(txg_history_t
));
133 ASSERT3U(dp
->dp_txg_history_size
, ==, 0);
134 list_destroy(&dp
->dp_txg_history
);
135 mutex_exit(&dp
->dp_lock
);
139 dsl_pool_txg_history_add(dsl_pool_t
*dp
, uint64_t txg
)
141 txg_history_t
*th
, *rm
;
143 th
= kmem_zalloc(sizeof(txg_history_t
), KM_SLEEP
);
144 mutex_init(&th
->th_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
145 th
->th_kstat
.txg
= txg
;
146 th
->th_kstat
.state
= TXG_STATE_OPEN
;
147 th
->th_kstat
.birth
= gethrtime();
149 mutex_enter(&dp
->dp_lock
);
151 list_insert_head(&dp
->dp_txg_history
, th
);
152 dp
->dp_txg_history_size
++;
154 while (dp
->dp_txg_history_size
> zfs_txg_history
) {
155 dp
->dp_txg_history_size
--;
156 rm
= list_remove_tail(&dp
->dp_txg_history
);
157 mutex_destroy(&rm
->th_lock
);
158 kmem_free(rm
, sizeof(txg_history_t
));
161 mutex_exit(&dp
->dp_lock
);
167 * Traversed youngest to oldest because lookups are only done for open
168 * or syncing txgs which are guaranteed to be at the head of the list.
169 * The txg_history_t structure will be returned locked.
172 dsl_pool_txg_history_get(dsl_pool_t
*dp
, uint64_t txg
)
176 mutex_enter(&dp
->dp_lock
);
177 for (th
= list_head(&dp
->dp_txg_history
); th
!= NULL
;
178 th
= list_next(&dp
->dp_txg_history
, th
)) {
179 if (th
->th_kstat
.txg
== txg
) {
180 mutex_enter(&th
->th_lock
);
184 mutex_exit(&dp
->dp_lock
);
190 dsl_pool_txg_history_put(txg_history_t
*th
)
192 mutex_exit(&th
->th_lock
);
196 dsl_pool_open_special_dir(dsl_pool_t
*dp
, const char *name
, dsl_dir_t
**ddp
)
201 err
= zap_lookup(dp
->dp_meta_objset
,
202 dp
->dp_root_dir
->dd_phys
->dd_child_dir_zapobj
,
203 name
, sizeof (obj
), 1, &obj
);
207 return (dsl_dir_open_obj(dp
, obj
, name
, dp
, ddp
));
211 dsl_pool_open_impl(spa_t
*spa
, uint64_t txg
)
214 blkptr_t
*bp
= spa_get_rootblkptr(spa
);
216 dp
= kmem_zalloc(sizeof (dsl_pool_t
), KM_SLEEP
);
218 dp
->dp_meta_rootbp
= *bp
;
219 rw_init(&dp
->dp_config_rwlock
, NULL
, RW_DEFAULT
, NULL
);
220 dp
->dp_write_limit
= zfs_write_limit_min
;
223 txg_list_create(&dp
->dp_dirty_datasets
,
224 offsetof(dsl_dataset_t
, ds_dirty_link
));
225 txg_list_create(&dp
->dp_dirty_dirs
,
226 offsetof(dsl_dir_t
, dd_dirty_link
));
227 txg_list_create(&dp
->dp_sync_tasks
,
228 offsetof(dsl_sync_task_group_t
, dstg_node
));
229 list_create(&dp
->dp_synced_datasets
, sizeof (dsl_dataset_t
),
230 offsetof(dsl_dataset_t
, ds_synced_link
));
232 mutex_init(&dp
->dp_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
234 dp
->dp_iput_taskq
= taskq_create("zfs_iput_taskq", 1, minclsyspri
,
237 dsl_pool_txg_history_init(dp
, txg
);
243 dsl_pool_open(spa_t
*spa
, uint64_t txg
, dsl_pool_t
**dpp
)
246 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
251 rw_enter(&dp
->dp_config_rwlock
, RW_WRITER
);
252 err
= dmu_objset_open_impl(spa
, NULL
, &dp
->dp_meta_rootbp
,
253 &dp
->dp_meta_objset
);
257 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
258 DMU_POOL_ROOT_DATASET
, sizeof (uint64_t), 1,
259 &dp
->dp_root_dir_obj
);
263 err
= dsl_dir_open_obj(dp
, dp
->dp_root_dir_obj
,
264 NULL
, dp
, &dp
->dp_root_dir
);
268 err
= dsl_pool_open_special_dir(dp
, MOS_DIR_NAME
, &dp
->dp_mos_dir
);
272 if (spa_version(spa
) >= SPA_VERSION_ORIGIN
) {
273 err
= dsl_pool_open_special_dir(dp
, ORIGIN_DIR_NAME
, &dd
);
276 err
= dsl_dataset_hold_obj(dp
, dd
->dd_phys
->dd_head_dataset_obj
,
279 err
= dsl_dataset_hold_obj(dp
,
280 ds
->ds_phys
->ds_prev_snap_obj
, dp
,
281 &dp
->dp_origin_snap
);
282 dsl_dataset_rele(ds
, FTAG
);
284 dsl_dir_close(dd
, dp
);
289 if (spa_version(spa
) >= SPA_VERSION_DEADLISTS
) {
290 err
= dsl_pool_open_special_dir(dp
, FREE_DIR_NAME
,
295 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
296 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
);
299 VERIFY3U(0, ==, bpobj_open(&dp
->dp_free_bpobj
,
300 dp
->dp_meta_objset
, obj
));
303 err
= zap_lookup(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
304 DMU_POOL_TMP_USERREFS
, sizeof (uint64_t), 1,
305 &dp
->dp_tmp_userrefs_obj
);
311 err
= dsl_scan_init(dp
, txg
);
314 rw_exit(&dp
->dp_config_rwlock
);
324 dsl_pool_close(dsl_pool_t
*dp
)
326 /* drop our references from dsl_pool_open() */
329 * Since we held the origin_snap from "syncing" context (which
330 * includes pool-opening context), it actually only got a "ref"
331 * and not a hold, so just drop that here.
333 if (dp
->dp_origin_snap
)
334 dsl_dataset_drop_ref(dp
->dp_origin_snap
, dp
);
336 dsl_dir_close(dp
->dp_mos_dir
, dp
);
338 dsl_dir_close(dp
->dp_free_dir
, dp
);
340 dsl_dir_close(dp
->dp_root_dir
, dp
);
342 bpobj_close(&dp
->dp_free_bpobj
);
344 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
345 if (dp
->dp_meta_objset
)
346 dmu_objset_evict(dp
->dp_meta_objset
);
348 txg_list_destroy(&dp
->dp_dirty_datasets
);
349 txg_list_destroy(&dp
->dp_sync_tasks
);
350 txg_list_destroy(&dp
->dp_dirty_dirs
);
351 list_destroy(&dp
->dp_synced_datasets
);
353 arc_flush(dp
->dp_spa
);
356 dsl_pool_txg_history_destroy(dp
);
357 rw_destroy(&dp
->dp_config_rwlock
);
358 mutex_destroy(&dp
->dp_lock
);
359 taskq_destroy(dp
->dp_iput_taskq
);
361 kmem_free(dp
->dp_blkstats
, sizeof (zfs_all_blkstats_t
));
362 kmem_free(dp
, sizeof (dsl_pool_t
));
366 dsl_pool_create(spa_t
*spa
, nvlist_t
*zplprops
, uint64_t txg
)
369 dsl_pool_t
*dp
= dsl_pool_open_impl(spa
, txg
);
370 dmu_tx_t
*tx
= dmu_tx_create_assigned(dp
, txg
);
375 /* create and open the MOS (meta-objset) */
376 dp
->dp_meta_objset
= dmu_objset_create_impl(spa
,
377 NULL
, &dp
->dp_meta_rootbp
, DMU_OST_META
, tx
);
379 /* create the pool directory */
380 err
= zap_create_claim(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
381 DMU_OT_OBJECT_DIRECTORY
, DMU_OT_NONE
, 0, tx
);
382 ASSERT3U(err
, ==, 0);
384 /* Initialize scan structures */
385 VERIFY3U(0, ==, dsl_scan_init(dp
, txg
));
387 /* create and open the root dir */
388 dp
->dp_root_dir_obj
= dsl_dir_create_sync(dp
, NULL
, NULL
, tx
);
389 VERIFY(0 == dsl_dir_open_obj(dp
, dp
->dp_root_dir_obj
,
390 NULL
, dp
, &dp
->dp_root_dir
));
392 /* create and open the meta-objset dir */
393 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, MOS_DIR_NAME
, tx
);
394 VERIFY(0 == dsl_pool_open_special_dir(dp
,
395 MOS_DIR_NAME
, &dp
->dp_mos_dir
));
397 if (spa_version(spa
) >= SPA_VERSION_DEADLISTS
) {
398 /* create and open the free dir */
399 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
,
401 VERIFY(0 == dsl_pool_open_special_dir(dp
,
402 FREE_DIR_NAME
, &dp
->dp_free_dir
));
404 /* create and open the free_bplist */
405 obj
= bpobj_alloc(dp
->dp_meta_objset
, SPA_MAXBLOCKSIZE
, tx
);
406 VERIFY(zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
407 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
) == 0);
408 VERIFY3U(0, ==, bpobj_open(&dp
->dp_free_bpobj
,
409 dp
->dp_meta_objset
, obj
));
412 if (spa_version(spa
) >= SPA_VERSION_DSL_SCRUB
)
413 dsl_pool_create_origin(dp
, tx
);
415 /* create the root dataset */
416 obj
= dsl_dataset_create_sync_dd(dp
->dp_root_dir
, NULL
, 0, tx
);
418 /* create the root objset */
419 VERIFY(0 == dsl_dataset_hold_obj(dp
, obj
, FTAG
, &ds
));
420 VERIFY(NULL
!= (os
= dmu_objset_create_impl(dp
->dp_spa
, ds
,
421 dsl_dataset_get_blkptr(ds
), DMU_OST_ZFS
, tx
)));
423 zfs_create_fs(os
, kcred
, zplprops
, tx
);
425 dsl_dataset_rele(ds
, FTAG
);
433 deadlist_enqueue_cb(void *arg
, const blkptr_t
*bp
, dmu_tx_t
*tx
)
435 dsl_deadlist_t
*dl
= arg
;
436 dsl_pool_t
*dp
= dmu_objset_pool(dl
->dl_os
);
437 rw_enter(&dp
->dp_config_rwlock
, RW_READER
);
438 dsl_deadlist_insert(dl
, bp
, tx
);
439 rw_exit(&dp
->dp_config_rwlock
);
444 dsl_pool_sync(dsl_pool_t
*dp
, uint64_t txg
)
450 dsl_sync_task_group_t
*dstg
;
451 objset_t
*mos
= dp
->dp_meta_objset
;
452 hrtime_t start
, write_time
;
453 uint64_t data_written
;
457 * We need to copy dp_space_towrite() before doing
458 * dsl_sync_task_group_sync(), because
459 * dsl_dataset_snapshot_reserve_space() will increase
460 * dp_space_towrite but not actually write anything.
462 data_written
= dp
->dp_space_towrite
[txg
& TXG_MASK
];
464 tx
= dmu_tx_create_assigned(dp
, txg
);
466 dp
->dp_read_overhead
= 0;
469 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
470 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
))) {
472 * We must not sync any non-MOS datasets twice, because
473 * we may have taken a snapshot of them. However, we
474 * may sync newly-created datasets on pass 2.
476 ASSERT(!list_link_active(&ds
->ds_synced_link
));
477 list_insert_tail(&dp
->dp_synced_datasets
, ds
);
478 dsl_dataset_sync(ds
, zio
, tx
);
480 DTRACE_PROBE(pool_sync__1setup
);
483 write_time
= gethrtime() - start
;
485 DTRACE_PROBE(pool_sync__2rootzio
);
487 for (ds
= list_head(&dp
->dp_synced_datasets
); ds
;
488 ds
= list_next(&dp
->dp_synced_datasets
, ds
))
489 dmu_objset_do_userquota_updates(ds
->ds_objset
, tx
);
492 * Sync the datasets again to push out the changes due to
493 * userspace updates. This must be done before we process the
494 * sync tasks, because that could cause a snapshot of a dataset
495 * whose ds_bp will be rewritten when we do this 2nd sync.
497 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
498 while ((ds
= txg_list_remove(&dp
->dp_dirty_datasets
, txg
))) {
499 ASSERT(list_link_active(&ds
->ds_synced_link
));
500 dmu_buf_rele(ds
->ds_dbuf
, ds
);
501 dsl_dataset_sync(ds
, zio
, tx
);
506 * Move dead blocks from the pending deadlist to the on-disk
509 for (ds
= list_head(&dp
->dp_synced_datasets
); ds
;
510 ds
= list_next(&dp
->dp_synced_datasets
, ds
)) {
511 bplist_iterate(&ds
->ds_pending_deadlist
,
512 deadlist_enqueue_cb
, &ds
->ds_deadlist
, tx
);
515 while ((dstg
= txg_list_remove(&dp
->dp_sync_tasks
, txg
))) {
517 * No more sync tasks should have been added while we
520 ASSERT(spa_sync_pass(dp
->dp_spa
) == 1);
521 dsl_sync_task_group_sync(dstg
, tx
);
523 DTRACE_PROBE(pool_sync__3task
);
526 while ((dd
= txg_list_remove(&dp
->dp_dirty_dirs
, txg
)))
527 dsl_dir_sync(dd
, tx
);
528 write_time
+= gethrtime() - start
;
531 if (list_head(&mos
->os_dirty_dnodes
[txg
& TXG_MASK
]) != NULL
||
532 list_head(&mos
->os_free_dnodes
[txg
& TXG_MASK
]) != NULL
) {
533 zio
= zio_root(dp
->dp_spa
, NULL
, NULL
, ZIO_FLAG_MUSTSUCCEED
);
534 dmu_objset_sync(mos
, zio
, tx
);
537 dprintf_bp(&dp
->dp_meta_rootbp
, "meta objset rootbp is %s", "");
538 spa_set_rootblkptr(dp
->dp_spa
, &dp
->dp_meta_rootbp
);
540 write_time
+= gethrtime() - start
;
541 DTRACE_PROBE2(pool_sync__4io
, hrtime_t
, write_time
,
542 hrtime_t
, dp
->dp_read_overhead
);
543 write_time
-= dp
->dp_read_overhead
;
547 dp
->dp_space_towrite
[txg
& TXG_MASK
] = 0;
548 ASSERT(dp
->dp_tempreserved
[txg
& TXG_MASK
] == 0);
551 * If the write limit max has not been explicitly set, set it
552 * to a fraction of available physical memory (default 1/8th).
553 * Note that we must inflate the limit because the spa
554 * inflates write sizes to account for data replication.
555 * Check this each sync phase to catch changing memory size.
557 if (physmem
!= old_physmem
&& zfs_write_limit_shift
) {
558 mutex_enter(&zfs_write_limit_lock
);
559 old_physmem
= physmem
;
560 zfs_write_limit_max
= ptob(physmem
) >> zfs_write_limit_shift
;
561 zfs_write_limit_inflated
= MAX(zfs_write_limit_min
,
562 spa_get_asize(dp
->dp_spa
, zfs_write_limit_max
));
563 mutex_exit(&zfs_write_limit_lock
);
567 * Attempt to keep the sync time consistent by adjusting the
568 * amount of write traffic allowed into each transaction group.
569 * Weight the throughput calculation towards the current value:
570 * thru = 3/4 old_thru + 1/4 new_thru
572 * Note: write_time is in nanosecs, so write_time/MICROSEC
575 ASSERT(zfs_write_limit_min
> 0);
576 if (data_written
> zfs_write_limit_min
/ 8 && write_time
> MICROSEC
) {
577 uint64_t throughput
= data_written
/ (write_time
/ MICROSEC
);
579 if (dp
->dp_throughput
)
580 dp
->dp_throughput
= throughput
/ 4 +
581 3 * dp
->dp_throughput
/ 4;
583 dp
->dp_throughput
= throughput
;
584 dp
->dp_write_limit
= MIN(zfs_write_limit_inflated
,
585 MAX(zfs_write_limit_min
,
586 dp
->dp_throughput
* zfs_txg_synctime_ms
));
591 dsl_pool_sync_done(dsl_pool_t
*dp
, uint64_t txg
)
596 while ((ds
= list_head(&dp
->dp_synced_datasets
))) {
597 list_remove(&dp
->dp_synced_datasets
, ds
);
599 zil_clean(os
->os_zil
, txg
);
600 ASSERT(!dmu_objset_is_dirty(os
, txg
));
601 dmu_buf_rele(ds
->ds_dbuf
, ds
);
603 ASSERT(!dmu_objset_is_dirty(dp
->dp_meta_objset
, txg
));
607 * TRUE if the current thread is the tx_sync_thread or if we
608 * are being called from SPA context during pool initialization.
611 dsl_pool_sync_context(dsl_pool_t
*dp
)
613 return (curthread
== dp
->dp_tx
.tx_sync_thread
||
614 spa_get_dsl(dp
->dp_spa
) == NULL
);
618 dsl_pool_adjustedsize(dsl_pool_t
*dp
, boolean_t netfree
)
620 uint64_t space
, resv
;
623 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
625 * XXX The intent log is not accounted for, so it must fit
628 * If we're trying to assess whether it's OK to do a free,
629 * cut the reservation in half to allow forward progress
630 * (e.g. make it possible to rm(1) files from a full pool).
632 space
= spa_get_dspace(dp
->dp_spa
);
633 resv
= MAX(space
>> 6, SPA_MINDEVSIZE
>> 1);
637 return (space
- resv
);
641 dsl_pool_tempreserve_space(dsl_pool_t
*dp
, uint64_t space
, dmu_tx_t
*tx
)
643 uint64_t reserved
= 0;
644 uint64_t write_limit
= (zfs_write_limit_override
?
645 zfs_write_limit_override
: dp
->dp_write_limit
);
647 if (zfs_no_write_throttle
) {
648 atomic_add_64(&dp
->dp_tempreserved
[tx
->tx_txg
& TXG_MASK
],
654 * Check to see if we have exceeded the maximum allowed IO for
655 * this transaction group. We can do this without locks since
656 * a little slop here is ok. Note that we do the reserved check
657 * with only half the requested reserve: this is because the
658 * reserve requests are worst-case, and we really don't want to
659 * throttle based off of worst-case estimates.
661 if (write_limit
> 0) {
662 reserved
= dp
->dp_space_towrite
[tx
->tx_txg
& TXG_MASK
]
663 + dp
->dp_tempreserved
[tx
->tx_txg
& TXG_MASK
] / 2;
665 if (reserved
&& reserved
> write_limit
) {
666 DMU_TX_STAT_BUMP(dmu_tx_write_limit
);
671 atomic_add_64(&dp
->dp_tempreserved
[tx
->tx_txg
& TXG_MASK
], space
);
674 * If this transaction group is over 7/8ths capacity, delay
675 * the caller 1 clock tick. This will slow down the "fill"
676 * rate until the sync process can catch up with us.
678 if (reserved
&& reserved
> (write_limit
- (write_limit
>> 3)))
679 txg_delay(dp
, tx
->tx_txg
, 1);
685 dsl_pool_tempreserve_clear(dsl_pool_t
*dp
, int64_t space
, dmu_tx_t
*tx
)
687 ASSERT(dp
->dp_tempreserved
[tx
->tx_txg
& TXG_MASK
] >= space
);
688 atomic_add_64(&dp
->dp_tempreserved
[tx
->tx_txg
& TXG_MASK
], -space
);
692 dsl_pool_memory_pressure(dsl_pool_t
*dp
)
694 uint64_t space_inuse
= 0;
697 if (dp
->dp_write_limit
== zfs_write_limit_min
)
700 for (i
= 0; i
< TXG_SIZE
; i
++) {
701 space_inuse
+= dp
->dp_space_towrite
[i
];
702 space_inuse
+= dp
->dp_tempreserved
[i
];
704 dp
->dp_write_limit
= MAX(zfs_write_limit_min
,
705 MIN(dp
->dp_write_limit
, space_inuse
/ 4));
709 dsl_pool_willuse_space(dsl_pool_t
*dp
, int64_t space
, dmu_tx_t
*tx
)
712 mutex_enter(&dp
->dp_lock
);
713 dp
->dp_space_towrite
[tx
->tx_txg
& TXG_MASK
] += space
;
714 mutex_exit(&dp
->dp_lock
);
720 upgrade_clones_cb(spa_t
*spa
, uint64_t dsobj
, const char *dsname
, void *arg
)
723 dsl_dataset_t
*ds
, *prev
= NULL
;
725 dsl_pool_t
*dp
= spa_get_dsl(spa
);
727 err
= dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
);
731 while (ds
->ds_phys
->ds_prev_snap_obj
!= 0) {
732 err
= dsl_dataset_hold_obj(dp
, ds
->ds_phys
->ds_prev_snap_obj
,
735 dsl_dataset_rele(ds
, FTAG
);
739 if (prev
->ds_phys
->ds_next_snap_obj
!= ds
->ds_object
)
741 dsl_dataset_rele(ds
, FTAG
);
747 prev
= dp
->dp_origin_snap
;
750 * The $ORIGIN can't have any data, or the accounting
753 ASSERT(prev
->ds_phys
->ds_bp
.blk_birth
== 0);
755 /* The origin doesn't get attached to itself */
756 if (ds
->ds_object
== prev
->ds_object
) {
757 dsl_dataset_rele(ds
, FTAG
);
761 dmu_buf_will_dirty(ds
->ds_dbuf
, tx
);
762 ds
->ds_phys
->ds_prev_snap_obj
= prev
->ds_object
;
763 ds
->ds_phys
->ds_prev_snap_txg
= prev
->ds_phys
->ds_creation_txg
;
765 dmu_buf_will_dirty(ds
->ds_dir
->dd_dbuf
, tx
);
766 ds
->ds_dir
->dd_phys
->dd_origin_obj
= prev
->ds_object
;
768 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
769 prev
->ds_phys
->ds_num_children
++;
771 if (ds
->ds_phys
->ds_next_snap_obj
== 0) {
772 ASSERT(ds
->ds_prev
== NULL
);
773 VERIFY(0 == dsl_dataset_hold_obj(dp
,
774 ds
->ds_phys
->ds_prev_snap_obj
, ds
, &ds
->ds_prev
));
778 ASSERT(ds
->ds_dir
->dd_phys
->dd_origin_obj
== prev
->ds_object
);
779 ASSERT(ds
->ds_phys
->ds_prev_snap_obj
== prev
->ds_object
);
781 if (prev
->ds_phys
->ds_next_clones_obj
== 0) {
782 dmu_buf_will_dirty(prev
->ds_dbuf
, tx
);
783 prev
->ds_phys
->ds_next_clones_obj
=
784 zap_create(dp
->dp_meta_objset
,
785 DMU_OT_NEXT_CLONES
, DMU_OT_NONE
, 0, tx
);
787 VERIFY(0 == zap_add_int(dp
->dp_meta_objset
,
788 prev
->ds_phys
->ds_next_clones_obj
, ds
->ds_object
, tx
));
790 dsl_dataset_rele(ds
, FTAG
);
791 if (prev
!= dp
->dp_origin_snap
)
792 dsl_dataset_rele(prev
, FTAG
);
797 dsl_pool_upgrade_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
799 ASSERT(dmu_tx_is_syncing(tx
));
800 ASSERT(dp
->dp_origin_snap
!= NULL
);
802 VERIFY3U(0, ==, dmu_objset_find_spa(dp
->dp_spa
, NULL
, upgrade_clones_cb
,
803 tx
, DS_FIND_CHILDREN
));
808 upgrade_dir_clones_cb(spa_t
*spa
, uint64_t dsobj
, const char *dsname
, void *arg
)
812 dsl_pool_t
*dp
= spa_get_dsl(spa
);
813 objset_t
*mos
= dp
->dp_meta_objset
;
815 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
817 if (ds
->ds_dir
->dd_phys
->dd_origin_obj
) {
818 dsl_dataset_t
*origin
;
820 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp
,
821 ds
->ds_dir
->dd_phys
->dd_origin_obj
, FTAG
, &origin
));
823 if (origin
->ds_dir
->dd_phys
->dd_clones
== 0) {
824 dmu_buf_will_dirty(origin
->ds_dir
->dd_dbuf
, tx
);
825 origin
->ds_dir
->dd_phys
->dd_clones
= zap_create(mos
,
826 DMU_OT_DSL_CLONES
, DMU_OT_NONE
, 0, tx
);
829 VERIFY3U(0, ==, zap_add_int(dp
->dp_meta_objset
,
830 origin
->ds_dir
->dd_phys
->dd_clones
, dsobj
, tx
));
832 dsl_dataset_rele(origin
, FTAG
);
835 dsl_dataset_rele(ds
, FTAG
);
840 dsl_pool_upgrade_dir_clones(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
844 ASSERT(dmu_tx_is_syncing(tx
));
846 (void) dsl_dir_create_sync(dp
, dp
->dp_root_dir
, FREE_DIR_NAME
, tx
);
847 VERIFY(0 == dsl_pool_open_special_dir(dp
,
848 FREE_DIR_NAME
, &dp
->dp_free_dir
));
851 * We can't use bpobj_alloc(), because spa_version() still
852 * returns the old version, and we need a new-version bpobj with
853 * subobj support. So call dmu_object_alloc() directly.
855 obj
= dmu_object_alloc(dp
->dp_meta_objset
, DMU_OT_BPOBJ
,
856 SPA_MAXBLOCKSIZE
, DMU_OT_BPOBJ_HDR
, sizeof (bpobj_phys_t
), tx
);
857 VERIFY3U(0, ==, zap_add(dp
->dp_meta_objset
, DMU_POOL_DIRECTORY_OBJECT
,
858 DMU_POOL_FREE_BPOBJ
, sizeof (uint64_t), 1, &obj
, tx
));
859 VERIFY3U(0, ==, bpobj_open(&dp
->dp_free_bpobj
,
860 dp
->dp_meta_objset
, obj
));
862 VERIFY3U(0, ==, dmu_objset_find_spa(dp
->dp_spa
, NULL
,
863 upgrade_dir_clones_cb
, tx
, DS_FIND_CHILDREN
));
867 dsl_pool_create_origin(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
872 ASSERT(dmu_tx_is_syncing(tx
));
873 ASSERT(dp
->dp_origin_snap
== NULL
);
875 /* create the origin dir, ds, & snap-ds */
876 rw_enter(&dp
->dp_config_rwlock
, RW_WRITER
);
877 dsobj
= dsl_dataset_create_sync(dp
->dp_root_dir
, ORIGIN_DIR_NAME
,
879 VERIFY(0 == dsl_dataset_hold_obj(dp
, dsobj
, FTAG
, &ds
));
880 dsl_dataset_snapshot_sync(ds
, ORIGIN_DIR_NAME
, tx
);
881 VERIFY(0 == dsl_dataset_hold_obj(dp
, ds
->ds_phys
->ds_prev_snap_obj
,
882 dp
, &dp
->dp_origin_snap
));
883 dsl_dataset_rele(ds
, FTAG
);
884 rw_exit(&dp
->dp_config_rwlock
);
888 dsl_pool_iput_taskq(dsl_pool_t
*dp
)
890 return (dp
->dp_iput_taskq
);
894 * Walk through the pool-wide zap object of temporary snapshot user holds
898 dsl_pool_clean_tmp_userrefs(dsl_pool_t
*dp
)
902 objset_t
*mos
= dp
->dp_meta_objset
;
903 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
907 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
909 for (zap_cursor_init(&zc
, mos
, zapobj
);
910 zap_cursor_retrieve(&zc
, &za
) == 0;
911 zap_cursor_advance(&zc
)) {
915 htag
= strchr(za
.za_name
, '-');
918 dsobj
= strtonum(za
.za_name
, NULL
);
919 (void) dsl_dataset_user_release_tmp(dp
, dsobj
, htag
, B_FALSE
);
921 zap_cursor_fini(&zc
);
925 * Create the pool-wide zap object for storing temporary snapshot holds.
928 dsl_pool_user_hold_create_obj(dsl_pool_t
*dp
, dmu_tx_t
*tx
)
930 objset_t
*mos
= dp
->dp_meta_objset
;
932 ASSERT(dp
->dp_tmp_userrefs_obj
== 0);
933 ASSERT(dmu_tx_is_syncing(tx
));
935 dp
->dp_tmp_userrefs_obj
= zap_create(mos
, DMU_OT_USERREFS
,
938 VERIFY(zap_add(mos
, DMU_POOL_DIRECTORY_OBJECT
, DMU_POOL_TMP_USERREFS
,
939 sizeof (uint64_t), 1, &dp
->dp_tmp_userrefs_obj
, tx
) == 0);
943 dsl_pool_user_hold_rele_impl(dsl_pool_t
*dp
, uint64_t dsobj
,
944 const char *tag
, uint64_t *now
, dmu_tx_t
*tx
, boolean_t holding
)
946 objset_t
*mos
= dp
->dp_meta_objset
;
947 uint64_t zapobj
= dp
->dp_tmp_userrefs_obj
;
951 ASSERT(spa_version(dp
->dp_spa
) >= SPA_VERSION_USERREFS
);
952 ASSERT(dmu_tx_is_syncing(tx
));
955 * If the pool was created prior to SPA_VERSION_USERREFS, the
956 * zap object for temporary holds might not exist yet.
960 dsl_pool_user_hold_create_obj(dp
, tx
);
961 zapobj
= dp
->dp_tmp_userrefs_obj
;
967 name
= kmem_asprintf("%llx-%s", (u_longlong_t
)dsobj
, tag
);
969 error
= zap_add(mos
, zapobj
, name
, 8, 1, now
, tx
);
971 error
= zap_remove(mos
, zapobj
, name
, tx
);
978 * Add a temporary hold for the given dataset object and tag.
981 dsl_pool_user_hold(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
982 uint64_t *now
, dmu_tx_t
*tx
)
984 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, now
, tx
, B_TRUE
));
988 * Release a temporary hold for the given dataset object and tag.
991 dsl_pool_user_release(dsl_pool_t
*dp
, uint64_t dsobj
, const char *tag
,
994 return (dsl_pool_user_hold_rele_impl(dp
, dsobj
, tag
, NULL
,
998 #if defined(_KERNEL) && defined(HAVE_SPL)
999 module_param(zfs_no_write_throttle
, int, 0644);
1000 MODULE_PARM_DESC(zfs_no_write_throttle
, "Disable write throttling");
1002 module_param(zfs_write_limit_shift
, int, 0444);
1003 MODULE_PARM_DESC(zfs_write_limit_shift
, "log2(fraction of memory) per txg");
1005 module_param(zfs_txg_synctime_ms
, int, 0644);
1006 MODULE_PARM_DESC(zfs_txg_synctime_ms
, "Target milliseconds between txg sync");
1008 module_param(zfs_txg_history
, int, 0644);
1009 MODULE_PARM_DESC(zfs_txg_history
, "Historic statistics for the last N txgs");
1011 module_param(zfs_write_limit_min
, ulong
, 0444);
1012 MODULE_PARM_DESC(zfs_write_limit_min
, "Min txg write limit");
1014 module_param(zfs_write_limit_max
, ulong
, 0444);
1015 MODULE_PARM_DESC(zfs_write_limit_max
, "Max txg write limit");
1017 module_param(zfs_write_limit_inflated
, ulong
, 0444);
1018 MODULE_PARM_DESC(zfs_write_limit_inflated
, "Inflated txg write limit");
1020 module_param(zfs_write_limit_override
, ulong
, 0444);
1021 MODULE_PARM_DESC(zfs_write_limit_override
, "Override txg write limit");