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>
52 #include <sys/sysctl.h>
53 #include <sys/types.h>
60 * ZFS must limit the rate of incoming writes to the rate at which it is able
61 * to sync data modifications to the backend storage. Throttling by too much
62 * creates an artificial limit; throttling by too little can only be sustained
63 * for short periods and would lead to highly lumpy performance. On a per-pool
64 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
65 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
66 * of dirty data decreases. When the amount of dirty data exceeds a
67 * predetermined threshold further modifications are blocked until the amount
68 * of dirty data decreases (as data is synced out).
70 * The limit on dirty data is tunable, and should be adjusted according to
71 * both the IO capacity and available memory of the system. The larger the
72 * window, the more ZFS is able to aggregate and amortize metadata (and data)
73 * changes. However, memory is a limited resource, and allowing for more dirty
74 * data comes at the cost of keeping other useful data in memory (for example
75 * ZFS data cached by the ARC).
79 * As buffers are modified dsl_pool_willuse_space() increments both the per-
80 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
81 * dirty space used; dsl_pool_dirty_space() decrements those values as data
82 * is synced out from dsl_pool_sync(). While only the poolwide value is
83 * relevant, the per-txg value is useful for debugging. The tunable
84 * zfs_dirty_data_max determines the dirty space limit. Once that value is
85 * exceeded, new writes are halted until space frees up.
87 * The zfs_dirty_data_sync tunable dictates the threshold at which we
88 * ensure that there is a txg syncing (see the comment in txg.c for a full
89 * description of transaction group stages).
91 * The IO scheduler uses both the dirty space limit and current amount of
92 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
93 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
95 * The delay is also calculated based on the amount of dirty data. See the
96 * comment above dmu_tx_delay() for details.
100 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
101 * capped at zfs_dirty_data_max_max. It can also be overridden in /etc/system.
103 uint64_t zfs_dirty_data_max;
104 uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024;
105 int zfs_dirty_data_max_percent = 10;
108 * If there is at least this much dirty data, push out a txg.
110 uint64_t zfs_dirty_data_sync = 64 * 1024 * 1024;
113 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
114 * and delay each transaction.
115 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
117 int zfs_delay_min_dirty_percent = 60;
120 * This controls how quickly the delay approaches infinity.
121 * Larger values cause it to delay more for a given amount of dirty data.
122 * Therefore larger values will cause there to be less dirty data for a
125 * For the smoothest delay, this value should be about 1 billion divided
126 * by the maximum number of operations per second. This will smoothly
127 * handle between 10x and 1/10th this number.
129 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
130 * multiply in dmu_tx_delay().
132 uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
137 extern int zfs_vdev_async_write_active_max_dirty_percent;
139 SYSCTL_DECL(_vfs_zfs);
141 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_max, CTLFLAG_RWTUN,
142 &zfs_dirty_data_max, 0,
143 "The maximum amount of dirty data in bytes after which new writes are "
144 "halted until space becomes available");
146 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_max_max, CTLFLAG_RDTUN,
147 &zfs_dirty_data_max_max, 0,
148 "The absolute cap on dirty_data_max when auto calculating");
150 static int sysctl_zfs_dirty_data_max_percent(SYSCTL_HANDLER_ARGS);
151 SYSCTL_PROC(_vfs_zfs, OID_AUTO, dirty_data_max_percent,
152 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
153 sysctl_zfs_dirty_data_max_percent, "I",
154 "The percent of physical memory used to auto calculate dirty_data_max");
156 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_sync, CTLFLAG_RWTUN,
157 &zfs_dirty_data_sync, 0,
158 "Force a txg if the number of dirty buffer bytes exceed this value");
160 static int sysctl_zfs_delay_min_dirty_percent(SYSCTL_HANDLER_ARGS);
161 /* No zfs_delay_min_dirty_percent tunable due to limit requirements */
162 SYSCTL_PROC(_vfs_zfs, OID_AUTO, delay_min_dirty_percent,
163 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(int),
164 sysctl_zfs_delay_min_dirty_percent, "I",
165 "The limit of outstanding dirty data before transations are delayed");
167 static int sysctl_zfs_delay_scale(SYSCTL_HANDLER_ARGS);
168 /* No zfs_delay_scale tunable due to limit requirements */
169 SYSCTL_PROC(_vfs_zfs, OID_AUTO, delay_scale,
170 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
171 sysctl_zfs_delay_scale, "QU",
172 "Controls how quickly the delay approaches infinity");
175 sysctl_zfs_dirty_data_max_percent(SYSCTL_HANDLER_ARGS)
179 val = zfs_dirty_data_max_percent;
180 err = sysctl_handle_int(oidp, &val, 0, req);
181 if (err != 0 || req->newptr == NULL)
184 if (val < 0 || val > 100)
187 zfs_dirty_data_max_percent = val;
193 sysctl_zfs_delay_min_dirty_percent(SYSCTL_HANDLER_ARGS)
197 val = zfs_delay_min_dirty_percent;
198 err = sysctl_handle_int(oidp, &val, 0, req);
199 if (err != 0 || req->newptr == NULL)
202 if (val < zfs_vdev_async_write_active_max_dirty_percent)
205 zfs_delay_min_dirty_percent = val;
211 sysctl_zfs_delay_scale(SYSCTL_HANDLER_ARGS)
216 val = zfs_delay_scale;
217 err = sysctl_handle_64(oidp, &val, 0, req);
218 if (err != 0 || req->newptr == NULL)
221 if (val > UINT64_MAX / zfs_dirty_data_max)
224 zfs_delay_scale = val;
230 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
231 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
234 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
239 err = zap_lookup(dp->dp_meta_objset,
240 dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
241 name, sizeof (obj), 1, &obj);
245 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
249 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
252 blkptr_t *bp = spa_get_rootblkptr(spa);
254 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
256 dp->dp_meta_rootbp = *bp;
257 rrw_init(&dp->dp_config_rwlock, B_TRUE);
260 txg_list_create(&dp->dp_dirty_datasets,
261 offsetof(dsl_dataset_t, ds_dirty_link));
262 txg_list_create(&dp->dp_dirty_zilogs,
263 offsetof(zilog_t, zl_dirty_link));
264 txg_list_create(&dp->dp_dirty_dirs,
265 offsetof(dsl_dir_t, dd_dirty_link));
266 txg_list_create(&dp->dp_sync_tasks,
267 offsetof(dsl_sync_task_t, dst_node));
269 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
270 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
272 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
279 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
282 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
284 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
285 &dp->dp_meta_objset);
295 dsl_pool_open(dsl_pool_t *dp)
302 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
303 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
304 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
305 &dp->dp_root_dir_obj);
309 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
310 NULL, dp, &dp->dp_root_dir);
314 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
318 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
319 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
322 err = dsl_dataset_hold_obj(dp,
323 dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds);
325 err = dsl_dataset_hold_obj(dp,
326 dsl_dataset_phys(ds)->ds_prev_snap_obj, dp,
327 &dp->dp_origin_snap);
328 dsl_dataset_rele(ds, FTAG);
330 dsl_dir_rele(dd, dp);
335 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
336 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
341 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
342 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
345 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
346 dp->dp_meta_objset, obj));
350 * Note: errors ignored, because the leak dir will not exist if we
351 * have not encountered a leak yet.
353 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
356 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
357 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
358 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
364 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
365 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
366 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
367 &dp->dp_empty_bpobj);
372 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
373 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
374 &dp->dp_tmp_userrefs_obj);
380 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
383 rrw_exit(&dp->dp_config_rwlock, FTAG);
388 dsl_pool_close(dsl_pool_t *dp)
391 * Drop our references from dsl_pool_open().
393 * Since we held the origin_snap from "syncing" context (which
394 * includes pool-opening context), it actually only got a "ref"
395 * and not a hold, so just drop that here.
397 if (dp->dp_origin_snap)
398 dsl_dataset_rele(dp->dp_origin_snap, dp);
400 dsl_dir_rele(dp->dp_mos_dir, dp);
402 dsl_dir_rele(dp->dp_free_dir, dp);
404 dsl_dir_rele(dp->dp_leak_dir, dp);
406 dsl_dir_rele(dp->dp_root_dir, dp);
408 bpobj_close(&dp->dp_free_bpobj);
410 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
411 if (dp->dp_meta_objset)
412 dmu_objset_evict(dp->dp_meta_objset);
414 txg_list_destroy(&dp->dp_dirty_datasets);
415 txg_list_destroy(&dp->dp_dirty_zilogs);
416 txg_list_destroy(&dp->dp_sync_tasks);
417 txg_list_destroy(&dp->dp_dirty_dirs);
420 * We can't set retry to TRUE since we're explicitly specifying
421 * a spa to flush. This is good enough; any missed buffers for
422 * this spa won't cause trouble, and they'll eventually fall
423 * out of the ARC just like any other unused buffer.
425 arc_flush(dp->dp_spa, FALSE);
429 dmu_buf_user_evict_wait();
431 rrw_destroy(&dp->dp_config_rwlock);
432 mutex_destroy(&dp->dp_lock);
433 taskq_destroy(dp->dp_vnrele_taskq);
435 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
436 kmem_free(dp, sizeof (dsl_pool_t));
440 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
443 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
444 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
449 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
451 /* create and open the MOS (meta-objset) */
452 dp->dp_meta_objset = dmu_objset_create_impl(spa,
453 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
455 /* create the pool directory */
456 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
457 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
460 /* Initialize scan structures */
461 VERIFY0(dsl_scan_init(dp, txg));
463 /* create and open the root dir */
464 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
465 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
466 NULL, dp, &dp->dp_root_dir));
468 /* create and open the meta-objset dir */
469 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
470 VERIFY0(dsl_pool_open_special_dir(dp,
471 MOS_DIR_NAME, &dp->dp_mos_dir));
473 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
474 /* create and open the free dir */
475 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
477 VERIFY0(dsl_pool_open_special_dir(dp,
478 FREE_DIR_NAME, &dp->dp_free_dir));
480 /* create and open the free_bplist */
481 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
482 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
483 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
484 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
485 dp->dp_meta_objset, obj));
488 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
489 dsl_pool_create_origin(dp, tx);
491 /* create the root dataset */
492 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
494 /* create the root objset */
495 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
496 os = dmu_objset_create_impl(dp->dp_spa, ds,
497 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
499 zfs_create_fs(os, kcred, zplprops, tx);
501 dsl_dataset_rele(ds, FTAG);
505 rrw_exit(&dp->dp_config_rwlock, FTAG);
511 * Account for the meta-objset space in its placeholder dsl_dir.
514 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
515 int64_t used, int64_t comp, int64_t uncomp)
517 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
518 mutex_enter(&dp->dp_lock);
519 dp->dp_mos_used_delta += used;
520 dp->dp_mos_compressed_delta += comp;
521 dp->dp_mos_uncompressed_delta += uncomp;
522 mutex_exit(&dp->dp_lock);
526 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
528 dsl_deadlist_t *dl = arg;
529 dsl_deadlist_insert(dl, bp, tx);
534 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
536 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
537 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
538 VERIFY0(zio_wait(zio));
539 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
540 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
544 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
546 ASSERT(MUTEX_HELD(&dp->dp_lock));
549 ASSERT3U(-delta, <=, dp->dp_dirty_total);
551 dp->dp_dirty_total += delta;
554 * Note: we signal even when increasing dp_dirty_total.
555 * This ensures forward progress -- each thread wakes the next waiter.
557 if (dp->dp_dirty_total <= zfs_dirty_data_max)
558 cv_signal(&dp->dp_spaceavail_cv);
562 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
568 objset_t *mos = dp->dp_meta_objset;
569 list_t synced_datasets;
571 list_create(&synced_datasets, sizeof (dsl_dataset_t),
572 offsetof(dsl_dataset_t, ds_synced_link));
574 tx = dmu_tx_create_assigned(dp, txg);
577 * Write out all dirty blocks of dirty datasets.
579 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
580 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
582 * We must not sync any non-MOS datasets twice, because
583 * we may have taken a snapshot of them. However, we
584 * may sync newly-created datasets on pass 2.
586 ASSERT(!list_link_active(&ds->ds_synced_link));
587 list_insert_tail(&synced_datasets, ds);
588 dsl_dataset_sync(ds, zio, tx);
590 VERIFY0(zio_wait(zio));
593 * We have written all of the accounted dirty data, so our
594 * dp_space_towrite should now be zero. However, some seldom-used
595 * code paths do not adhere to this (e.g. dbuf_undirty(), also
596 * rounding error in dbuf_write_physdone).
597 * Shore up the accounting of any dirtied space now.
599 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
602 * After the data blocks have been written (ensured by the zio_wait()
603 * above), update the user/group space accounting.
605 for (ds = list_head(&synced_datasets); ds != NULL;
606 ds = list_next(&synced_datasets, ds)) {
607 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
611 * Sync the datasets again to push out the changes due to
612 * userspace updates. This must be done before we process the
613 * sync tasks, so that any snapshots will have the correct
614 * user accounting information (and we won't get confused
615 * about which blocks are part of the snapshot).
617 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
618 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
619 ASSERT(list_link_active(&ds->ds_synced_link));
620 dmu_buf_rele(ds->ds_dbuf, ds);
621 dsl_dataset_sync(ds, zio, tx);
623 VERIFY0(zio_wait(zio));
626 * Now that the datasets have been completely synced, we can
627 * clean up our in-memory structures accumulated while syncing:
629 * - move dead blocks from the pending deadlist to the on-disk deadlist
630 * - release hold from dsl_dataset_dirty()
632 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
633 objset_t *os = ds->ds_objset;
634 bplist_iterate(&ds->ds_pending_deadlist,
635 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
636 ASSERT(!dmu_objset_is_dirty(os, txg));
637 dmu_buf_rele(ds->ds_dbuf, ds);
639 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
640 dsl_dir_sync(dd, tx);
644 * The MOS's space is accounted for in the pool/$MOS
645 * (dp_mos_dir). We can't modify the mos while we're syncing
646 * it, so we remember the deltas and apply them here.
648 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
649 dp->dp_mos_uncompressed_delta != 0) {
650 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
651 dp->dp_mos_used_delta,
652 dp->dp_mos_compressed_delta,
653 dp->dp_mos_uncompressed_delta, tx);
654 dp->dp_mos_used_delta = 0;
655 dp->dp_mos_compressed_delta = 0;
656 dp->dp_mos_uncompressed_delta = 0;
659 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
660 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
661 dsl_pool_sync_mos(dp, tx);
665 * If we modify a dataset in the same txg that we want to destroy it,
666 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
667 * dsl_dir_destroy_check() will fail if there are unexpected holds.
668 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
669 * and clearing the hold on it) before we process the sync_tasks.
670 * The MOS data dirtied by the sync_tasks will be synced on the next
673 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
674 dsl_sync_task_t *dst;
676 * No more sync tasks should have been added while we
679 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
680 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
681 dsl_sync_task_sync(dst, tx);
686 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
690 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
694 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
695 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
696 zil_clean(zilog, txg);
697 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
698 dmu_buf_rele(ds->ds_dbuf, zilog);
700 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
704 * TRUE if the current thread is the tx_sync_thread or if we
705 * are being called from SPA context during pool initialization.
708 dsl_pool_sync_context(dsl_pool_t *dp)
710 return (curthread == dp->dp_tx.tx_sync_thread ||
711 spa_is_initializing(dp->dp_spa));
715 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
717 uint64_t space, resv;
720 * If we're trying to assess whether it's OK to do a free,
721 * cut the reservation in half to allow forward progress
722 * (e.g. make it possible to rm(1) files from a full pool).
724 space = spa_get_dspace(dp->dp_spa);
725 resv = spa_get_slop_space(dp->dp_spa);
729 return (space - resv);
733 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
735 uint64_t delay_min_bytes =
736 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
739 mutex_enter(&dp->dp_lock);
740 if (dp->dp_dirty_total > zfs_dirty_data_sync)
742 rv = (dp->dp_dirty_total > delay_min_bytes);
743 mutex_exit(&dp->dp_lock);
748 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
751 mutex_enter(&dp->dp_lock);
752 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
753 dsl_pool_dirty_delta(dp, space);
754 mutex_exit(&dp->dp_lock);
759 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
761 ASSERT3S(space, >=, 0);
764 mutex_enter(&dp->dp_lock);
765 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
766 /* XXX writing something we didn't dirty? */
767 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
769 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
770 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
771 ASSERT3U(dp->dp_dirty_total, >=, space);
772 dsl_pool_dirty_delta(dp, -space);
773 mutex_exit(&dp->dp_lock);
778 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
781 dsl_dataset_t *ds, *prev = NULL;
784 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
788 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
789 err = dsl_dataset_hold_obj(dp,
790 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
792 dsl_dataset_rele(ds, FTAG);
796 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
798 dsl_dataset_rele(ds, FTAG);
804 prev = dp->dp_origin_snap;
807 * The $ORIGIN can't have any data, or the accounting
810 ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
812 /* The origin doesn't get attached to itself */
813 if (ds->ds_object == prev->ds_object) {
814 dsl_dataset_rele(ds, FTAG);
818 dmu_buf_will_dirty(ds->ds_dbuf, tx);
819 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
820 dsl_dataset_phys(ds)->ds_prev_snap_txg =
821 dsl_dataset_phys(prev)->ds_creation_txg;
823 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
824 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
826 dmu_buf_will_dirty(prev->ds_dbuf, tx);
827 dsl_dataset_phys(prev)->ds_num_children++;
829 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
830 ASSERT(ds->ds_prev == NULL);
831 VERIFY0(dsl_dataset_hold_obj(dp,
832 dsl_dataset_phys(ds)->ds_prev_snap_obj,
837 ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
838 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
840 if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
841 dmu_buf_will_dirty(prev->ds_dbuf, tx);
842 dsl_dataset_phys(prev)->ds_next_clones_obj =
843 zap_create(dp->dp_meta_objset,
844 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
846 VERIFY0(zap_add_int(dp->dp_meta_objset,
847 dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
849 dsl_dataset_rele(ds, FTAG);
850 if (prev != dp->dp_origin_snap)
851 dsl_dataset_rele(prev, FTAG);
856 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
858 ASSERT(dmu_tx_is_syncing(tx));
859 ASSERT(dp->dp_origin_snap != NULL);
861 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
862 tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
867 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
870 objset_t *mos = dp->dp_meta_objset;
872 if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
873 dsl_dataset_t *origin;
875 VERIFY0(dsl_dataset_hold_obj(dp,
876 dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
878 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
879 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
880 dsl_dir_phys(origin->ds_dir)->dd_clones =
881 zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
885 VERIFY0(zap_add_int(dp->dp_meta_objset,
886 dsl_dir_phys(origin->ds_dir)->dd_clones,
889 dsl_dataset_rele(origin, FTAG);
895 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
897 ASSERT(dmu_tx_is_syncing(tx));
900 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
901 VERIFY0(dsl_pool_open_special_dir(dp,
902 FREE_DIR_NAME, &dp->dp_free_dir));
905 * We can't use bpobj_alloc(), because spa_version() still
906 * returns the old version, and we need a new-version bpobj with
907 * subobj support. So call dmu_object_alloc() directly.
909 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
910 SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
911 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
912 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
913 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
915 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
916 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
920 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
925 ASSERT(dmu_tx_is_syncing(tx));
926 ASSERT(dp->dp_origin_snap == NULL);
927 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
929 /* create the origin dir, ds, & snap-ds */
930 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
932 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
933 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
934 VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
935 dp, &dp->dp_origin_snap));
936 dsl_dataset_rele(ds, FTAG);
940 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
942 return (dp->dp_vnrele_taskq);
946 * Walk through the pool-wide zap object of temporary snapshot user holds
950 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
954 objset_t *mos = dp->dp_meta_objset;
955 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
960 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
962 holds = fnvlist_alloc();
964 for (zap_cursor_init(&zc, mos, zapobj);
965 zap_cursor_retrieve(&zc, &za) == 0;
966 zap_cursor_advance(&zc)) {
970 htag = strchr(za.za_name, '-');
973 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
974 tags = fnvlist_alloc();
975 fnvlist_add_boolean(tags, htag);
976 fnvlist_add_nvlist(holds, za.za_name, tags);
979 fnvlist_add_boolean(tags, htag);
982 dsl_dataset_user_release_tmp(dp, holds);
984 zap_cursor_fini(&zc);
988 * Create the pool-wide zap object for storing temporary snapshot holds.
991 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
993 objset_t *mos = dp->dp_meta_objset;
995 ASSERT(dp->dp_tmp_userrefs_obj == 0);
996 ASSERT(dmu_tx_is_syncing(tx));
998 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
999 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
1003 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
1004 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
1006 objset_t *mos = dp->dp_meta_objset;
1007 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
1011 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1012 ASSERT(dmu_tx_is_syncing(tx));
1015 * If the pool was created prior to SPA_VERSION_USERREFS, the
1016 * zap object for temporary holds might not exist yet.
1020 dsl_pool_user_hold_create_obj(dp, tx);
1021 zapobj = dp->dp_tmp_userrefs_obj;
1023 return (SET_ERROR(ENOENT));
1027 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
1029 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
1031 error = zap_remove(mos, zapobj, name, tx);
1038 * Add a temporary hold for the given dataset object and tag.
1041 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1042 uint64_t now, dmu_tx_t *tx)
1044 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
1048 * Release a temporary hold for the given dataset object and tag.
1051 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1054 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0,
1059 * DSL Pool Configuration Lock
1061 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
1062 * creation / destruction / rename / property setting). It must be held for
1063 * read to hold a dataset or dsl_dir. I.e. you must call
1064 * dsl_pool_config_enter() or dsl_pool_hold() before calling
1065 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
1066 * must be held continuously until all datasets and dsl_dirs are released.
1068 * The only exception to this rule is that if a "long hold" is placed on
1069 * a dataset, then the dp_config_rwlock may be dropped while the dataset
1070 * is still held. The long hold will prevent the dataset from being
1071 * destroyed -- the destroy will fail with EBUSY. A long hold can be
1072 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
1073 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
1075 * Legitimate long-holders (including owners) should be long-running, cancelable
1076 * tasks that should cause "zfs destroy" to fail. This includes DMU
1077 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
1078 * "zfs send", and "zfs diff". There are several other long-holders whose
1079 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
1081 * The usual formula for long-holding would be:
1083 * dsl_dataset_hold()
1084 * ... perform checks ...
1085 * dsl_dataset_long_hold()
1087 * ... perform long-running task ...
1088 * dsl_dataset_long_rele()
1089 * dsl_dataset_rele()
1091 * Note that when the long hold is released, the dataset is still held but
1092 * the pool is not held. The dataset may change arbitrarily during this time
1093 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1094 * dataset except release it.
1096 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1097 * or modifying operations.
1099 * Modifying operations should generally use dsl_sync_task(). The synctask
1100 * infrastructure enforces proper locking strategy with respect to the
1101 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1103 * Read-only operations will manually hold the pool, then the dataset, obtain
1104 * information from the dataset, then release the pool and dataset.
1105 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1110 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1115 error = spa_open(name, &spa, tag);
1117 *dp = spa_get_dsl(spa);
1118 dsl_pool_config_enter(*dp, tag);
1124 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1126 dsl_pool_config_exit(dp, tag);
1127 spa_close(dp->dp_spa, tag);
1131 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1134 * We use a "reentrant" reader-writer lock, but not reentrantly.
1136 * The rrwlock can (with the track_all flag) track all reading threads,
1137 * which is very useful for debugging which code path failed to release
1138 * the lock, and for verifying that the *current* thread does hold
1141 * (Unlike a rwlock, which knows that N threads hold it for
1142 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1143 * if any thread holds it for read, even if this thread doesn't).
1145 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1146 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1150 dsl_pool_config_enter_prio(dsl_pool_t *dp, void *tag)
1152 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1153 rrw_enter_read_prio(&dp->dp_config_rwlock, tag);
1157 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1159 rrw_exit(&dp->dp_config_rwlock, tag);
1163 dsl_pool_config_held(dsl_pool_t *dp)
1165 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1169 dsl_pool_config_held_writer(dsl_pool_t *dp)
1171 return (RRW_WRITE_HELD(&dp->dp_config_rwlock));