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.
28 #include <sys/dsl_pool.h>
29 #include <sys/dsl_dataset.h>
30 #include <sys/dsl_prop.h>
31 #include <sys/dsl_dir.h>
32 #include <sys/dsl_synctask.h>
33 #include <sys/dsl_scan.h>
34 #include <sys/dnode.h>
35 #include <sys/dmu_tx.h>
36 #include <sys/dmu_objset.h>
40 #include <sys/zfs_context.h>
41 #include <sys/fs/zfs.h>
42 #include <sys/zfs_znode.h>
43 #include <sys/spa_impl.h>
44 #include <sys/dsl_deadlist.h>
45 #include <sys/bptree.h>
46 #include <sys/zfeature.h>
47 #include <sys/zil_impl.h>
48 #include <sys/dsl_userhold.h>
51 #include <sys/sysctl.h>
52 #include <sys/types.h>
59 * ZFS must limit the rate of incoming writes to the rate at which it is able
60 * to sync data modifications to the backend storage. Throttling by too much
61 * creates an artificial limit; throttling by too little can only be sustained
62 * for short periods and would lead to highly lumpy performance. On a per-pool
63 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
64 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
65 * of dirty data decreases. When the amount of dirty data exceeds a
66 * predetermined threshold further modifications are blocked until the amount
67 * of dirty data decreases (as data is synced out).
69 * The limit on dirty data is tunable, and should be adjusted according to
70 * both the IO capacity and available memory of the system. The larger the
71 * window, the more ZFS is able to aggregate and amortize metadata (and data)
72 * changes. However, memory is a limited resource, and allowing for more dirty
73 * data comes at the cost of keeping other useful data in memory (for example
74 * ZFS data cached by the ARC).
78 * As buffers are modified dsl_pool_willuse_space() increments both the per-
79 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
80 * dirty space used; dsl_pool_dirty_space() decrements those values as data
81 * is synced out from dsl_pool_sync(). While only the poolwide value is
82 * relevant, the per-txg value is useful for debugging. The tunable
83 * zfs_dirty_data_max determines the dirty space limit. Once that value is
84 * exceeded, new writes are halted until space frees up.
86 * The zfs_dirty_data_sync tunable dictates the threshold at which we
87 * ensure that there is a txg syncing (see the comment in txg.c for a full
88 * description of transaction group stages).
90 * The IO scheduler uses both the dirty space limit and current amount of
91 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
92 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
94 * The delay is also calculated based on the amount of dirty data. See the
95 * comment above dmu_tx_delay() for details.
99 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
100 * capped at zfs_dirty_data_max_max. It can also be overridden in /etc/system.
102 uint64_t zfs_dirty_data_max;
103 uint64_t zfs_dirty_data_max_max = 4ULL * 1024 * 1024 * 1024;
104 int zfs_dirty_data_max_percent = 10;
107 * If there is at least this much dirty data, push out a txg.
109 uint64_t zfs_dirty_data_sync = 64 * 1024 * 1024;
112 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
113 * and delay each transaction.
114 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
116 int zfs_delay_min_dirty_percent = 60;
119 * This controls how quickly the delay approaches infinity.
120 * Larger values cause it to delay more for a given amount of dirty data.
121 * Therefore larger values will cause there to be less dirty data for a
124 * For the smoothest delay, this value should be about 1 billion divided
125 * by the maximum number of operations per second. This will smoothly
126 * handle between 10x and 1/10th this number.
128 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
129 * multiply in dmu_tx_delay().
131 uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
136 extern int zfs_vdev_async_write_active_max_dirty_percent;
138 SYSCTL_DECL(_vfs_zfs);
140 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_max, CTLFLAG_RWTUN,
141 &zfs_dirty_data_max, 0,
142 "The maximum amount of dirty data in bytes after which new writes are "
143 "halted until space becomes available");
145 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_max_max, CTLFLAG_RDTUN,
146 &zfs_dirty_data_max_max, 0,
147 "The absolute cap on dirty_data_max when auto calculating");
149 static int sysctl_zfs_dirty_data_max_percent(SYSCTL_HANDLER_ARGS);
150 SYSCTL_PROC(_vfs_zfs, OID_AUTO, dirty_data_max_percent,
151 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RWTUN, 0, sizeof(int),
152 sysctl_zfs_dirty_data_max_percent, "I",
153 "The percent of physical memory used to auto calculate dirty_data_max");
155 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, dirty_data_sync, CTLFLAG_RWTUN,
156 &zfs_dirty_data_sync, 0,
157 "Force a txg if the number of dirty buffer bytes exceed this value");
159 static int sysctl_zfs_delay_min_dirty_percent(SYSCTL_HANDLER_ARGS);
160 /* No zfs_delay_min_dirty_percent tunable due to limit requirements */
161 SYSCTL_PROC(_vfs_zfs, OID_AUTO, delay_min_dirty_percent,
162 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(int),
163 sysctl_zfs_delay_min_dirty_percent, "I",
164 "The limit of outstanding dirty data before transations are delayed");
166 static int sysctl_zfs_delay_scale(SYSCTL_HANDLER_ARGS);
167 /* No zfs_delay_scale tunable due to limit requirements */
168 SYSCTL_PROC(_vfs_zfs, OID_AUTO, delay_scale,
169 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
170 sysctl_zfs_delay_scale, "QU",
171 "Controls how quickly the delay approaches infinity");
174 sysctl_zfs_dirty_data_max_percent(SYSCTL_HANDLER_ARGS)
178 val = zfs_dirty_data_max_percent;
179 err = sysctl_handle_int(oidp, &val, 0, req);
180 if (err != 0 || req->newptr == NULL)
183 if (val < 0 || val > 100)
186 zfs_dirty_data_max_percent = val;
192 sysctl_zfs_delay_min_dirty_percent(SYSCTL_HANDLER_ARGS)
196 val = zfs_delay_min_dirty_percent;
197 err = sysctl_handle_int(oidp, &val, 0, req);
198 if (err != 0 || req->newptr == NULL)
201 if (val < zfs_vdev_async_write_active_max_dirty_percent)
204 zfs_delay_min_dirty_percent = val;
210 sysctl_zfs_delay_scale(SYSCTL_HANDLER_ARGS)
215 val = zfs_delay_scale;
216 err = sysctl_handle_64(oidp, &val, 0, req);
217 if (err != 0 || req->newptr == NULL)
220 if (val > UINT64_MAX / zfs_dirty_data_max)
223 zfs_delay_scale = val;
229 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
230 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
233 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
238 err = zap_lookup(dp->dp_meta_objset,
239 dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
240 name, sizeof (obj), 1, &obj);
244 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
248 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
251 blkptr_t *bp = spa_get_rootblkptr(spa);
253 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
255 dp->dp_meta_rootbp = *bp;
256 rrw_init(&dp->dp_config_rwlock, B_TRUE);
259 txg_list_create(&dp->dp_dirty_datasets,
260 offsetof(dsl_dataset_t, ds_dirty_link));
261 txg_list_create(&dp->dp_dirty_zilogs,
262 offsetof(zilog_t, zl_dirty_link));
263 txg_list_create(&dp->dp_dirty_dirs,
264 offsetof(dsl_dir_t, dd_dirty_link));
265 txg_list_create(&dp->dp_sync_tasks,
266 offsetof(dsl_sync_task_t, dst_node));
268 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
269 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
271 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
278 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
281 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
283 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
284 &dp->dp_meta_objset);
294 dsl_pool_open(dsl_pool_t *dp)
301 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
302 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
303 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
304 &dp->dp_root_dir_obj);
308 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
309 NULL, dp, &dp->dp_root_dir);
313 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
317 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
318 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
321 err = dsl_dataset_hold_obj(dp,
322 dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds);
324 err = dsl_dataset_hold_obj(dp,
325 dsl_dataset_phys(ds)->ds_prev_snap_obj, dp,
326 &dp->dp_origin_snap);
327 dsl_dataset_rele(ds, FTAG);
329 dsl_dir_rele(dd, dp);
334 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
335 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
340 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
341 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
344 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
345 dp->dp_meta_objset, obj));
349 * Note: errors ignored, because the leak dir will not exist if we
350 * have not encountered a leak yet.
352 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
355 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
356 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
357 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
363 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
364 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
365 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
366 &dp->dp_empty_bpobj);
371 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
372 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
373 &dp->dp_tmp_userrefs_obj);
379 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
382 rrw_exit(&dp->dp_config_rwlock, FTAG);
387 dsl_pool_close(dsl_pool_t *dp)
390 * Drop our references from dsl_pool_open().
392 * Since we held the origin_snap from "syncing" context (which
393 * includes pool-opening context), it actually only got a "ref"
394 * and not a hold, so just drop that here.
396 if (dp->dp_origin_snap)
397 dsl_dataset_rele(dp->dp_origin_snap, dp);
399 dsl_dir_rele(dp->dp_mos_dir, dp);
401 dsl_dir_rele(dp->dp_free_dir, dp);
403 dsl_dir_rele(dp->dp_leak_dir, dp);
405 dsl_dir_rele(dp->dp_root_dir, dp);
407 bpobj_close(&dp->dp_free_bpobj);
409 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
410 if (dp->dp_meta_objset)
411 dmu_objset_evict(dp->dp_meta_objset);
413 txg_list_destroy(&dp->dp_dirty_datasets);
414 txg_list_destroy(&dp->dp_dirty_zilogs);
415 txg_list_destroy(&dp->dp_sync_tasks);
416 txg_list_destroy(&dp->dp_dirty_dirs);
419 * We can't set retry to TRUE since we're explicitly specifying
420 * a spa to flush. This is good enough; any missed buffers for
421 * this spa won't cause trouble, and they'll eventually fall
422 * out of the ARC just like any other unused buffer.
424 arc_flush(dp->dp_spa, FALSE);
428 dmu_buf_user_evict_wait();
430 rrw_destroy(&dp->dp_config_rwlock);
431 mutex_destroy(&dp->dp_lock);
432 taskq_destroy(dp->dp_vnrele_taskq);
434 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
435 kmem_free(dp, sizeof (dsl_pool_t));
439 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
442 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
443 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
448 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
450 /* create and open the MOS (meta-objset) */
451 dp->dp_meta_objset = dmu_objset_create_impl(spa,
452 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
454 /* create the pool directory */
455 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
456 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
459 /* Initialize scan structures */
460 VERIFY0(dsl_scan_init(dp, txg));
462 /* create and open the root dir */
463 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
464 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
465 NULL, dp, &dp->dp_root_dir));
467 /* create and open the meta-objset dir */
468 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
469 VERIFY0(dsl_pool_open_special_dir(dp,
470 MOS_DIR_NAME, &dp->dp_mos_dir));
472 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
473 /* create and open the free dir */
474 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
476 VERIFY0(dsl_pool_open_special_dir(dp,
477 FREE_DIR_NAME, &dp->dp_free_dir));
479 /* create and open the free_bplist */
480 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
481 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
482 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
483 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
484 dp->dp_meta_objset, obj));
487 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
488 dsl_pool_create_origin(dp, tx);
490 /* create the root dataset */
491 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
493 /* create the root objset */
494 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
495 os = dmu_objset_create_impl(dp->dp_spa, ds,
496 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
498 zfs_create_fs(os, kcred, zplprops, tx);
500 dsl_dataset_rele(ds, FTAG);
504 rrw_exit(&dp->dp_config_rwlock, FTAG);
510 * Account for the meta-objset space in its placeholder dsl_dir.
513 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
514 int64_t used, int64_t comp, int64_t uncomp)
516 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
517 mutex_enter(&dp->dp_lock);
518 dp->dp_mos_used_delta += used;
519 dp->dp_mos_compressed_delta += comp;
520 dp->dp_mos_uncompressed_delta += uncomp;
521 mutex_exit(&dp->dp_lock);
525 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
527 dsl_deadlist_t *dl = arg;
528 dsl_deadlist_insert(dl, bp, tx);
533 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
535 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
536 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
537 VERIFY0(zio_wait(zio));
538 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
539 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
543 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
545 ASSERT(MUTEX_HELD(&dp->dp_lock));
548 ASSERT3U(-delta, <=, dp->dp_dirty_total);
550 dp->dp_dirty_total += delta;
553 * Note: we signal even when increasing dp_dirty_total.
554 * This ensures forward progress -- each thread wakes the next waiter.
556 if (dp->dp_dirty_total <= zfs_dirty_data_max)
557 cv_signal(&dp->dp_spaceavail_cv);
561 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
567 objset_t *mos = dp->dp_meta_objset;
568 list_t synced_datasets;
570 list_create(&synced_datasets, sizeof (dsl_dataset_t),
571 offsetof(dsl_dataset_t, ds_synced_link));
573 tx = dmu_tx_create_assigned(dp, txg);
576 * Write out all dirty blocks of dirty datasets.
578 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
579 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
581 * We must not sync any non-MOS datasets twice, because
582 * we may have taken a snapshot of them. However, we
583 * may sync newly-created datasets on pass 2.
585 ASSERT(!list_link_active(&ds->ds_synced_link));
586 list_insert_tail(&synced_datasets, ds);
587 dsl_dataset_sync(ds, zio, tx);
589 VERIFY0(zio_wait(zio));
592 * We have written all of the accounted dirty data, so our
593 * dp_space_towrite should now be zero. However, some seldom-used
594 * code paths do not adhere to this (e.g. dbuf_undirty(), also
595 * rounding error in dbuf_write_physdone).
596 * Shore up the accounting of any dirtied space now.
598 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
601 * After the data blocks have been written (ensured by the zio_wait()
602 * above), update the user/group space accounting.
604 for (ds = list_head(&synced_datasets); ds != NULL;
605 ds = list_next(&synced_datasets, ds)) {
606 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
610 * Sync the datasets again to push out the changes due to
611 * userspace updates. This must be done before we process the
612 * sync tasks, so that any snapshots will have the correct
613 * user accounting information (and we won't get confused
614 * about which blocks are part of the snapshot).
616 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
617 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
618 ASSERT(list_link_active(&ds->ds_synced_link));
619 dmu_buf_rele(ds->ds_dbuf, ds);
620 dsl_dataset_sync(ds, zio, tx);
622 VERIFY0(zio_wait(zio));
625 * Now that the datasets have been completely synced, we can
626 * clean up our in-memory structures accumulated while syncing:
628 * - move dead blocks from the pending deadlist to the on-disk deadlist
629 * - release hold from dsl_dataset_dirty()
631 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
632 objset_t *os = ds->ds_objset;
633 bplist_iterate(&ds->ds_pending_deadlist,
634 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
635 ASSERT(!dmu_objset_is_dirty(os, txg));
636 dmu_buf_rele(ds->ds_dbuf, ds);
638 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
639 dsl_dir_sync(dd, tx);
643 * The MOS's space is accounted for in the pool/$MOS
644 * (dp_mos_dir). We can't modify the mos while we're syncing
645 * it, so we remember the deltas and apply them here.
647 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
648 dp->dp_mos_uncompressed_delta != 0) {
649 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
650 dp->dp_mos_used_delta,
651 dp->dp_mos_compressed_delta,
652 dp->dp_mos_uncompressed_delta, tx);
653 dp->dp_mos_used_delta = 0;
654 dp->dp_mos_compressed_delta = 0;
655 dp->dp_mos_uncompressed_delta = 0;
658 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
659 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
660 dsl_pool_sync_mos(dp, tx);
664 * If we modify a dataset in the same txg that we want to destroy it,
665 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
666 * dsl_dir_destroy_check() will fail if there are unexpected holds.
667 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
668 * and clearing the hold on it) before we process the sync_tasks.
669 * The MOS data dirtied by the sync_tasks will be synced on the next
672 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
673 dsl_sync_task_t *dst;
675 * No more sync tasks should have been added while we
678 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
679 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
680 dsl_sync_task_sync(dst, tx);
685 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
689 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
693 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
694 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
695 zil_clean(zilog, txg);
696 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
697 dmu_buf_rele(ds->ds_dbuf, zilog);
699 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
703 * TRUE if the current thread is the tx_sync_thread or if we
704 * are being called from SPA context during pool initialization.
707 dsl_pool_sync_context(dsl_pool_t *dp)
709 return (curthread == dp->dp_tx.tx_sync_thread ||
710 spa_is_initializing(dp->dp_spa));
714 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
716 uint64_t space, resv;
719 * If we're trying to assess whether it's OK to do a free,
720 * cut the reservation in half to allow forward progress
721 * (e.g. make it possible to rm(1) files from a full pool).
723 space = spa_get_dspace(dp->dp_spa);
724 resv = spa_get_slop_space(dp->dp_spa);
728 return (space - resv);
732 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
734 uint64_t delay_min_bytes =
735 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
738 mutex_enter(&dp->dp_lock);
739 if (dp->dp_dirty_total > zfs_dirty_data_sync)
741 rv = (dp->dp_dirty_total > delay_min_bytes);
742 mutex_exit(&dp->dp_lock);
747 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
750 mutex_enter(&dp->dp_lock);
751 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
752 dsl_pool_dirty_delta(dp, space);
753 mutex_exit(&dp->dp_lock);
758 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
760 ASSERT3S(space, >=, 0);
763 mutex_enter(&dp->dp_lock);
764 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
765 /* XXX writing something we didn't dirty? */
766 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
768 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
769 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
770 ASSERT3U(dp->dp_dirty_total, >=, space);
771 dsl_pool_dirty_delta(dp, -space);
772 mutex_exit(&dp->dp_lock);
777 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
780 dsl_dataset_t *ds, *prev = NULL;
783 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
787 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
788 err = dsl_dataset_hold_obj(dp,
789 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
791 dsl_dataset_rele(ds, FTAG);
795 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
797 dsl_dataset_rele(ds, FTAG);
803 prev = dp->dp_origin_snap;
806 * The $ORIGIN can't have any data, or the accounting
809 ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
811 /* The origin doesn't get attached to itself */
812 if (ds->ds_object == prev->ds_object) {
813 dsl_dataset_rele(ds, FTAG);
817 dmu_buf_will_dirty(ds->ds_dbuf, tx);
818 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
819 dsl_dataset_phys(ds)->ds_prev_snap_txg =
820 dsl_dataset_phys(prev)->ds_creation_txg;
822 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
823 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
825 dmu_buf_will_dirty(prev->ds_dbuf, tx);
826 dsl_dataset_phys(prev)->ds_num_children++;
828 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
829 ASSERT(ds->ds_prev == NULL);
830 VERIFY0(dsl_dataset_hold_obj(dp,
831 dsl_dataset_phys(ds)->ds_prev_snap_obj,
836 ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
837 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
839 if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
840 dmu_buf_will_dirty(prev->ds_dbuf, tx);
841 dsl_dataset_phys(prev)->ds_next_clones_obj =
842 zap_create(dp->dp_meta_objset,
843 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
845 VERIFY0(zap_add_int(dp->dp_meta_objset,
846 dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
848 dsl_dataset_rele(ds, FTAG);
849 if (prev != dp->dp_origin_snap)
850 dsl_dataset_rele(prev, FTAG);
855 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
857 ASSERT(dmu_tx_is_syncing(tx));
858 ASSERT(dp->dp_origin_snap != NULL);
860 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
861 tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
866 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
869 objset_t *mos = dp->dp_meta_objset;
871 if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
872 dsl_dataset_t *origin;
874 VERIFY0(dsl_dataset_hold_obj(dp,
875 dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
877 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
878 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
879 dsl_dir_phys(origin->ds_dir)->dd_clones =
880 zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
884 VERIFY0(zap_add_int(dp->dp_meta_objset,
885 dsl_dir_phys(origin->ds_dir)->dd_clones,
888 dsl_dataset_rele(origin, FTAG);
894 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
896 ASSERT(dmu_tx_is_syncing(tx));
899 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
900 VERIFY0(dsl_pool_open_special_dir(dp,
901 FREE_DIR_NAME, &dp->dp_free_dir));
904 * We can't use bpobj_alloc(), because spa_version() still
905 * returns the old version, and we need a new-version bpobj with
906 * subobj support. So call dmu_object_alloc() directly.
908 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
909 SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
910 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
911 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
912 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
914 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
915 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
919 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
924 ASSERT(dmu_tx_is_syncing(tx));
925 ASSERT(dp->dp_origin_snap == NULL);
926 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
928 /* create the origin dir, ds, & snap-ds */
929 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
931 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
932 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
933 VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
934 dp, &dp->dp_origin_snap));
935 dsl_dataset_rele(ds, FTAG);
939 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
941 return (dp->dp_vnrele_taskq);
945 * Walk through the pool-wide zap object of temporary snapshot user holds
949 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
953 objset_t *mos = dp->dp_meta_objset;
954 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
959 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
961 holds = fnvlist_alloc();
963 for (zap_cursor_init(&zc, mos, zapobj);
964 zap_cursor_retrieve(&zc, &za) == 0;
965 zap_cursor_advance(&zc)) {
969 htag = strchr(za.za_name, '-');
972 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
973 tags = fnvlist_alloc();
974 fnvlist_add_boolean(tags, htag);
975 fnvlist_add_nvlist(holds, za.za_name, tags);
978 fnvlist_add_boolean(tags, htag);
981 dsl_dataset_user_release_tmp(dp, holds);
983 zap_cursor_fini(&zc);
987 * Create the pool-wide zap object for storing temporary snapshot holds.
990 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
992 objset_t *mos = dp->dp_meta_objset;
994 ASSERT(dp->dp_tmp_userrefs_obj == 0);
995 ASSERT(dmu_tx_is_syncing(tx));
997 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
998 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
1002 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
1003 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
1005 objset_t *mos = dp->dp_meta_objset;
1006 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
1010 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
1011 ASSERT(dmu_tx_is_syncing(tx));
1014 * If the pool was created prior to SPA_VERSION_USERREFS, the
1015 * zap object for temporary holds might not exist yet.
1019 dsl_pool_user_hold_create_obj(dp, tx);
1020 zapobj = dp->dp_tmp_userrefs_obj;
1022 return (SET_ERROR(ENOENT));
1026 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
1028 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
1030 error = zap_remove(mos, zapobj, name, tx);
1037 * Add a temporary hold for the given dataset object and tag.
1040 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1041 uint64_t now, dmu_tx_t *tx)
1043 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
1047 * Release a temporary hold for the given dataset object and tag.
1050 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
1053 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0,
1058 * DSL Pool Configuration Lock
1060 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
1061 * creation / destruction / rename / property setting). It must be held for
1062 * read to hold a dataset or dsl_dir. I.e. you must call
1063 * dsl_pool_config_enter() or dsl_pool_hold() before calling
1064 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
1065 * must be held continuously until all datasets and dsl_dirs are released.
1067 * The only exception to this rule is that if a "long hold" is placed on
1068 * a dataset, then the dp_config_rwlock may be dropped while the dataset
1069 * is still held. The long hold will prevent the dataset from being
1070 * destroyed -- the destroy will fail with EBUSY. A long hold can be
1071 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
1072 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
1074 * Legitimate long-holders (including owners) should be long-running, cancelable
1075 * tasks that should cause "zfs destroy" to fail. This includes DMU
1076 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
1077 * "zfs send", and "zfs diff". There are several other long-holders whose
1078 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
1080 * The usual formula for long-holding would be:
1082 * dsl_dataset_hold()
1083 * ... perform checks ...
1084 * dsl_dataset_long_hold()
1086 * ... perform long-running task ...
1087 * dsl_dataset_long_rele()
1088 * dsl_dataset_rele()
1090 * Note that when the long hold is released, the dataset is still held but
1091 * the pool is not held. The dataset may change arbitrarily during this time
1092 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1093 * dataset except release it.
1095 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1096 * or modifying operations.
1098 * Modifying operations should generally use dsl_sync_task(). The synctask
1099 * infrastructure enforces proper locking strategy with respect to the
1100 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1102 * Read-only operations will manually hold the pool, then the dataset, obtain
1103 * information from the dataset, then release the pool and dataset.
1104 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1109 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1114 error = spa_open(name, &spa, tag);
1116 *dp = spa_get_dsl(spa);
1117 dsl_pool_config_enter(*dp, tag);
1123 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1125 dsl_pool_config_exit(dp, tag);
1126 spa_close(dp->dp_spa, tag);
1130 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1133 * We use a "reentrant" reader-writer lock, but not reentrantly.
1135 * The rrwlock can (with the track_all flag) track all reading threads,
1136 * which is very useful for debugging which code path failed to release
1137 * the lock, and for verifying that the *current* thread does hold
1140 * (Unlike a rwlock, which knows that N threads hold it for
1141 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1142 * if any thread holds it for read, even if this thread doesn't).
1144 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1145 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1149 dsl_pool_config_enter_prio(dsl_pool_t *dp, void *tag)
1151 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1152 rrw_enter_read_prio(&dp->dp_config_rwlock, tag);
1156 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1158 rrw_exit(&dp->dp_config_rwlock, tag);
1162 dsl_pool_config_held(dsl_pool_t *dp)
1164 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1168 dsl_pool_config_held_writer(dsl_pool_t *dp)
1170 return (RRW_WRITE_HELD(&dp->dp_config_rwlock));