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>
49 #include <sys/trace_txg.h>
55 * ZFS must limit the rate of incoming writes to the rate at which it is able
56 * to sync data modifications to the backend storage. Throttling by too much
57 * creates an artificial limit; throttling by too little can only be sustained
58 * for short periods and would lead to highly lumpy performance. On a per-pool
59 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
60 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
61 * of dirty data decreases. When the amount of dirty data exceeds a
62 * predetermined threshold further modifications are blocked until the amount
63 * of dirty data decreases (as data is synced out).
65 * The limit on dirty data is tunable, and should be adjusted according to
66 * both the IO capacity and available memory of the system. The larger the
67 * window, the more ZFS is able to aggregate and amortize metadata (and data)
68 * changes. However, memory is a limited resource, and allowing for more dirty
69 * data comes at the cost of keeping other useful data in memory (for example
70 * ZFS data cached by the ARC).
74 * As buffers are modified dsl_pool_willuse_space() increments both the per-
75 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
76 * dirty space used; dsl_pool_dirty_space() decrements those values as data
77 * is synced out from dsl_pool_sync(). While only the poolwide value is
78 * relevant, the per-txg value is useful for debugging. The tunable
79 * zfs_dirty_data_max determines the dirty space limit. Once that value is
80 * exceeded, new writes are halted until space frees up.
82 * The zfs_dirty_data_sync tunable dictates the threshold at which we
83 * ensure that there is a txg syncing (see the comment in txg.c for a full
84 * description of transaction group stages).
86 * The IO scheduler uses both the dirty space limit and current amount of
87 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
88 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
90 * The delay is also calculated based on the amount of dirty data. See the
91 * comment above dmu_tx_delay() for details.
95 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
96 * capped at zfs_dirty_data_max_max. It can also be overridden with a module
99 unsigned long zfs_dirty_data_max = 0;
100 unsigned long zfs_dirty_data_max_max = 0;
101 int zfs_dirty_data_max_percent = 10;
102 int zfs_dirty_data_max_max_percent = 25;
105 * If there is at least this much dirty data, push out a txg.
107 unsigned long zfs_dirty_data_sync = 64 * 1024 * 1024;
110 * Once there is this amount of dirty data, the dmu_tx_delay() will kick in
111 * and delay each transaction.
112 * This value should be >= zfs_vdev_async_write_active_max_dirty_percent.
114 int zfs_delay_min_dirty_percent = 60;
117 * This controls how quickly the delay approaches infinity.
118 * Larger values cause it to delay more for a given amount of dirty data.
119 * Therefore larger values will cause there to be less dirty data for a
122 * For the smoothest delay, this value should be about 1 billion divided
123 * by the maximum number of operations per second. This will smoothly
124 * handle between 10x and 1/10th this number.
126 * Note: zfs_delay_scale * zfs_dirty_data_max must be < 2^64, due to the
127 * multiply in dmu_tx_delay().
129 unsigned long zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
131 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
132 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
135 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
140 err = zap_lookup(dp->dp_meta_objset,
141 dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
142 name, sizeof (obj), 1, &obj);
146 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
150 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
153 blkptr_t *bp = spa_get_rootblkptr(spa);
155 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
157 dp->dp_meta_rootbp = *bp;
158 rrw_init(&dp->dp_config_rwlock, B_TRUE);
161 txg_list_create(&dp->dp_dirty_datasets,
162 offsetof(dsl_dataset_t, ds_dirty_link));
163 txg_list_create(&dp->dp_dirty_zilogs,
164 offsetof(zilog_t, zl_dirty_link));
165 txg_list_create(&dp->dp_dirty_dirs,
166 offsetof(dsl_dir_t, dd_dirty_link));
167 txg_list_create(&dp->dp_sync_tasks,
168 offsetof(dsl_sync_task_t, dst_node));
170 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
171 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
173 dp->dp_iput_taskq = taskq_create("z_iput", max_ncpus, minclsyspri,
174 max_ncpus * 8, INT_MAX, TASKQ_PREPOPULATE);
180 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
183 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
185 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
186 &dp->dp_meta_objset);
196 dsl_pool_open(dsl_pool_t *dp)
203 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
204 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
205 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
206 &dp->dp_root_dir_obj);
210 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
211 NULL, dp, &dp->dp_root_dir);
215 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
219 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
220 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
223 err = dsl_dataset_hold_obj(dp,
224 dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds);
226 err = dsl_dataset_hold_obj(dp,
227 dsl_dataset_phys(ds)->ds_prev_snap_obj, dp,
228 &dp->dp_origin_snap);
229 dsl_dataset_rele(ds, FTAG);
231 dsl_dir_rele(dd, dp);
236 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
237 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
242 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
243 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
246 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
247 dp->dp_meta_objset, obj));
251 * Note: errors ignored, because the leak dir will not exist if we
252 * have not encountered a leak yet.
254 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
257 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
258 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
259 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
265 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
266 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
267 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
268 &dp->dp_empty_bpobj);
273 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
274 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
275 &dp->dp_tmp_userrefs_obj);
281 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
284 rrw_exit(&dp->dp_config_rwlock, FTAG);
289 dsl_pool_close(dsl_pool_t *dp)
292 * Drop our references from dsl_pool_open().
294 * Since we held the origin_snap from "syncing" context (which
295 * includes pool-opening context), it actually only got a "ref"
296 * and not a hold, so just drop that here.
298 if (dp->dp_origin_snap)
299 dsl_dataset_rele(dp->dp_origin_snap, dp);
301 dsl_dir_rele(dp->dp_mos_dir, dp);
303 dsl_dir_rele(dp->dp_free_dir, dp);
305 dsl_dir_rele(dp->dp_leak_dir, dp);
307 dsl_dir_rele(dp->dp_root_dir, dp);
309 bpobj_close(&dp->dp_free_bpobj);
311 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
312 if (dp->dp_meta_objset)
313 dmu_objset_evict(dp->dp_meta_objset);
315 txg_list_destroy(&dp->dp_dirty_datasets);
316 txg_list_destroy(&dp->dp_dirty_zilogs);
317 txg_list_destroy(&dp->dp_sync_tasks);
318 txg_list_destroy(&dp->dp_dirty_dirs);
321 * We can't set retry to TRUE since we're explicitly specifying
322 * a spa to flush. This is good enough; any missed buffers for
323 * this spa won't cause trouble, and they'll eventually fall
324 * out of the ARC just like any other unused buffer.
326 arc_flush(dp->dp_spa, FALSE);
330 dmu_buf_user_evict_wait();
332 rrw_destroy(&dp->dp_config_rwlock);
333 mutex_destroy(&dp->dp_lock);
334 taskq_destroy(dp->dp_iput_taskq);
336 vmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
337 kmem_free(dp, sizeof (dsl_pool_t));
341 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
344 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
345 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
350 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
352 /* create and open the MOS (meta-objset) */
353 dp->dp_meta_objset = dmu_objset_create_impl(spa,
354 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
356 /* create the pool directory */
357 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
358 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
361 /* Initialize scan structures */
362 VERIFY0(dsl_scan_init(dp, txg));
364 /* create and open the root dir */
365 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
366 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
367 NULL, dp, &dp->dp_root_dir));
369 /* create and open the meta-objset dir */
370 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
371 VERIFY0(dsl_pool_open_special_dir(dp,
372 MOS_DIR_NAME, &dp->dp_mos_dir));
374 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
375 /* create and open the free dir */
376 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
378 VERIFY0(dsl_pool_open_special_dir(dp,
379 FREE_DIR_NAME, &dp->dp_free_dir));
381 /* create and open the free_bplist */
382 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
383 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
384 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
385 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
386 dp->dp_meta_objset, obj));
389 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
390 dsl_pool_create_origin(dp, tx);
392 /* create the root dataset */
393 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
395 /* create the root objset */
396 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
397 VERIFY(NULL != (os = dmu_objset_create_impl(dp->dp_spa, ds,
398 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx)));
400 zfs_create_fs(os, kcred, zplprops, tx);
402 dsl_dataset_rele(ds, FTAG);
406 rrw_exit(&dp->dp_config_rwlock, FTAG);
412 * Account for the meta-objset space in its placeholder dsl_dir.
415 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
416 int64_t used, int64_t comp, int64_t uncomp)
418 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
419 mutex_enter(&dp->dp_lock);
420 dp->dp_mos_used_delta += used;
421 dp->dp_mos_compressed_delta += comp;
422 dp->dp_mos_uncompressed_delta += uncomp;
423 mutex_exit(&dp->dp_lock);
427 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
429 dsl_deadlist_t *dl = arg;
430 dsl_deadlist_insert(dl, bp, tx);
435 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
437 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
438 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
439 VERIFY0(zio_wait(zio));
440 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
441 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
445 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
447 ASSERT(MUTEX_HELD(&dp->dp_lock));
450 ASSERT3U(-delta, <=, dp->dp_dirty_total);
452 dp->dp_dirty_total += delta;
455 * Note: we signal even when increasing dp_dirty_total.
456 * This ensures forward progress -- each thread wakes the next waiter.
458 if (dp->dp_dirty_total <= zfs_dirty_data_max)
459 cv_signal(&dp->dp_spaceavail_cv);
463 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
469 objset_t *mos = dp->dp_meta_objset;
470 list_t synced_datasets;
472 list_create(&synced_datasets, sizeof (dsl_dataset_t),
473 offsetof(dsl_dataset_t, ds_synced_link));
475 tx = dmu_tx_create_assigned(dp, txg);
478 * Write out all dirty blocks of dirty datasets.
480 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
481 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
483 * We must not sync any non-MOS datasets twice, because
484 * we may have taken a snapshot of them. However, we
485 * may sync newly-created datasets on pass 2.
487 ASSERT(!list_link_active(&ds->ds_synced_link));
488 list_insert_tail(&synced_datasets, ds);
489 dsl_dataset_sync(ds, zio, tx);
491 VERIFY0(zio_wait(zio));
494 * We have written all of the accounted dirty data, so our
495 * dp_space_towrite should now be zero. However, some seldom-used
496 * code paths do not adhere to this (e.g. dbuf_undirty(), also
497 * rounding error in dbuf_write_physdone).
498 * Shore up the accounting of any dirtied space now.
500 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
503 * After the data blocks have been written (ensured by the zio_wait()
504 * above), update the user/group space accounting.
506 for (ds = list_head(&synced_datasets); ds != NULL;
507 ds = list_next(&synced_datasets, ds)) {
508 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
512 * Sync the datasets again to push out the changes due to
513 * userspace updates. This must be done before we process the
514 * sync tasks, so that any snapshots will have the correct
515 * user accounting information (and we won't get confused
516 * about which blocks are part of the snapshot).
518 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
519 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
520 ASSERT(list_link_active(&ds->ds_synced_link));
521 dmu_buf_rele(ds->ds_dbuf, ds);
522 dsl_dataset_sync(ds, zio, tx);
524 VERIFY0(zio_wait(zio));
527 * Now that the datasets have been completely synced, we can
528 * clean up our in-memory structures accumulated while syncing:
530 * - move dead blocks from the pending deadlist to the on-disk deadlist
531 * - release hold from dsl_dataset_dirty()
533 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
534 ASSERTV(objset_t *os = ds->ds_objset);
535 bplist_iterate(&ds->ds_pending_deadlist,
536 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
537 ASSERT(!dmu_objset_is_dirty(os, txg));
538 dmu_buf_rele(ds->ds_dbuf, ds);
541 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
542 dsl_dir_sync(dd, tx);
546 * The MOS's space is accounted for in the pool/$MOS
547 * (dp_mos_dir). We can't modify the mos while we're syncing
548 * it, so we remember the deltas and apply them here.
550 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
551 dp->dp_mos_uncompressed_delta != 0) {
552 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
553 dp->dp_mos_used_delta,
554 dp->dp_mos_compressed_delta,
555 dp->dp_mos_uncompressed_delta, tx);
556 dp->dp_mos_used_delta = 0;
557 dp->dp_mos_compressed_delta = 0;
558 dp->dp_mos_uncompressed_delta = 0;
561 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
562 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
563 dsl_pool_sync_mos(dp, tx);
567 * If we modify a dataset in the same txg that we want to destroy it,
568 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
569 * dsl_dir_destroy_check() will fail if there are unexpected holds.
570 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
571 * and clearing the hold on it) before we process the sync_tasks.
572 * The MOS data dirtied by the sync_tasks will be synced on the next
575 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
576 dsl_sync_task_t *dst;
578 * No more sync tasks should have been added while we
581 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
582 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
583 dsl_sync_task_sync(dst, tx);
588 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
592 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
596 while ((zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg))) {
597 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
598 zil_clean(zilog, txg);
599 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
600 dmu_buf_rele(ds->ds_dbuf, zilog);
602 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
606 * TRUE if the current thread is the tx_sync_thread or if we
607 * are being called from SPA context during pool initialization.
610 dsl_pool_sync_context(dsl_pool_t *dp)
612 return (curthread == dp->dp_tx.tx_sync_thread ||
613 spa_is_initializing(dp->dp_spa));
617 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
619 uint64_t space, resv;
622 * If we're trying to assess whether it's OK to do a free,
623 * cut the reservation in half to allow forward progress
624 * (e.g. make it possible to rm(1) files from a full pool).
626 space = spa_get_dspace(dp->dp_spa);
627 resv = spa_get_slop_space(dp->dp_spa);
631 return (space - resv);
635 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
637 uint64_t delay_min_bytes =
638 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
641 mutex_enter(&dp->dp_lock);
642 if (dp->dp_dirty_total > zfs_dirty_data_sync)
644 rv = (dp->dp_dirty_total > delay_min_bytes);
645 mutex_exit(&dp->dp_lock);
650 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
653 mutex_enter(&dp->dp_lock);
654 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
655 dsl_pool_dirty_delta(dp, space);
656 mutex_exit(&dp->dp_lock);
661 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
663 ASSERT3S(space, >=, 0);
667 mutex_enter(&dp->dp_lock);
668 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
669 /* XXX writing something we didn't dirty? */
670 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
672 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
673 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
674 ASSERT3U(dp->dp_dirty_total, >=, space);
675 dsl_pool_dirty_delta(dp, -space);
676 mutex_exit(&dp->dp_lock);
681 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
684 dsl_dataset_t *ds, *prev = NULL;
687 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
691 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
692 err = dsl_dataset_hold_obj(dp,
693 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
695 dsl_dataset_rele(ds, FTAG);
699 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
701 dsl_dataset_rele(ds, FTAG);
707 prev = dp->dp_origin_snap;
710 * The $ORIGIN can't have any data, or the accounting
713 ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
715 /* The origin doesn't get attached to itself */
716 if (ds->ds_object == prev->ds_object) {
717 dsl_dataset_rele(ds, FTAG);
721 dmu_buf_will_dirty(ds->ds_dbuf, tx);
722 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
723 dsl_dataset_phys(ds)->ds_prev_snap_txg =
724 dsl_dataset_phys(prev)->ds_creation_txg;
726 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
727 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
729 dmu_buf_will_dirty(prev->ds_dbuf, tx);
730 dsl_dataset_phys(prev)->ds_num_children++;
732 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
733 ASSERT(ds->ds_prev == NULL);
734 VERIFY0(dsl_dataset_hold_obj(dp,
735 dsl_dataset_phys(ds)->ds_prev_snap_obj,
740 ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
741 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
743 if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
744 dmu_buf_will_dirty(prev->ds_dbuf, tx);
745 dsl_dataset_phys(prev)->ds_next_clones_obj =
746 zap_create(dp->dp_meta_objset,
747 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
749 VERIFY0(zap_add_int(dp->dp_meta_objset,
750 dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
752 dsl_dataset_rele(ds, FTAG);
753 if (prev != dp->dp_origin_snap)
754 dsl_dataset_rele(prev, FTAG);
759 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
761 ASSERT(dmu_tx_is_syncing(tx));
762 ASSERT(dp->dp_origin_snap != NULL);
764 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
765 tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
770 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
773 objset_t *mos = dp->dp_meta_objset;
775 if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
776 dsl_dataset_t *origin;
778 VERIFY0(dsl_dataset_hold_obj(dp,
779 dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
781 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
782 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
783 dsl_dir_phys(origin->ds_dir)->dd_clones =
784 zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
788 VERIFY0(zap_add_int(dp->dp_meta_objset,
789 dsl_dir_phys(origin->ds_dir)->dd_clones,
792 dsl_dataset_rele(origin, FTAG);
798 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
802 ASSERT(dmu_tx_is_syncing(tx));
804 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
805 VERIFY0(dsl_pool_open_special_dir(dp,
806 FREE_DIR_NAME, &dp->dp_free_dir));
809 * We can't use bpobj_alloc(), because spa_version() still
810 * returns the old version, and we need a new-version bpobj with
811 * subobj support. So call dmu_object_alloc() directly.
813 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
814 SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
815 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
816 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
817 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
819 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
820 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
824 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
829 ASSERT(dmu_tx_is_syncing(tx));
830 ASSERT(dp->dp_origin_snap == NULL);
831 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
833 /* create the origin dir, ds, & snap-ds */
834 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
836 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
837 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
838 VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
839 dp, &dp->dp_origin_snap));
840 dsl_dataset_rele(ds, FTAG);
844 dsl_pool_iput_taskq(dsl_pool_t *dp)
846 return (dp->dp_iput_taskq);
850 * Walk through the pool-wide zap object of temporary snapshot user holds
854 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
858 objset_t *mos = dp->dp_meta_objset;
859 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
864 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
866 holds = fnvlist_alloc();
868 for (zap_cursor_init(&zc, mos, zapobj);
869 zap_cursor_retrieve(&zc, &za) == 0;
870 zap_cursor_advance(&zc)) {
874 htag = strchr(za.za_name, '-');
877 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
878 tags = fnvlist_alloc();
879 fnvlist_add_boolean(tags, htag);
880 fnvlist_add_nvlist(holds, za.za_name, tags);
883 fnvlist_add_boolean(tags, htag);
886 dsl_dataset_user_release_tmp(dp, holds);
888 zap_cursor_fini(&zc);
892 * Create the pool-wide zap object for storing temporary snapshot holds.
895 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
897 objset_t *mos = dp->dp_meta_objset;
899 ASSERT(dp->dp_tmp_userrefs_obj == 0);
900 ASSERT(dmu_tx_is_syncing(tx));
902 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
903 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
907 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
908 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
910 objset_t *mos = dp->dp_meta_objset;
911 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
915 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
916 ASSERT(dmu_tx_is_syncing(tx));
919 * If the pool was created prior to SPA_VERSION_USERREFS, the
920 * zap object for temporary holds might not exist yet.
924 dsl_pool_user_hold_create_obj(dp, tx);
925 zapobj = dp->dp_tmp_userrefs_obj;
927 return (SET_ERROR(ENOENT));
931 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
933 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
935 error = zap_remove(mos, zapobj, name, tx);
942 * Add a temporary hold for the given dataset object and tag.
945 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
946 uint64_t now, dmu_tx_t *tx)
948 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
952 * Release a temporary hold for the given dataset object and tag.
955 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
958 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0,
963 * DSL Pool Configuration Lock
965 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
966 * creation / destruction / rename / property setting). It must be held for
967 * read to hold a dataset or dsl_dir. I.e. you must call
968 * dsl_pool_config_enter() or dsl_pool_hold() before calling
969 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
970 * must be held continuously until all datasets and dsl_dirs are released.
972 * The only exception to this rule is that if a "long hold" is placed on
973 * a dataset, then the dp_config_rwlock may be dropped while the dataset
974 * is still held. The long hold will prevent the dataset from being
975 * destroyed -- the destroy will fail with EBUSY. A long hold can be
976 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
977 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
979 * Legitimate long-holders (including owners) should be long-running, cancelable
980 * tasks that should cause "zfs destroy" to fail. This includes DMU
981 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
982 * "zfs send", and "zfs diff". There are several other long-holders whose
983 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
985 * The usual formula for long-holding would be:
988 * ... perform checks ...
989 * dsl_dataset_long_hold()
991 * ... perform long-running task ...
992 * dsl_dataset_long_rele()
995 * Note that when the long hold is released, the dataset is still held but
996 * the pool is not held. The dataset may change arbitrarily during this time
997 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
998 * dataset except release it.
1000 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1001 * or modifying operations.
1003 * Modifying operations should generally use dsl_sync_task(). The synctask
1004 * infrastructure enforces proper locking strategy with respect to the
1005 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1007 * Read-only operations will manually hold the pool, then the dataset, obtain
1008 * information from the dataset, then release the pool and dataset.
1009 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1014 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1019 error = spa_open(name, &spa, tag);
1021 *dp = spa_get_dsl(spa);
1022 dsl_pool_config_enter(*dp, tag);
1028 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1030 dsl_pool_config_exit(dp, tag);
1031 spa_close(dp->dp_spa, tag);
1035 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1038 * We use a "reentrant" reader-writer lock, but not reentrantly.
1040 * The rrwlock can (with the track_all flag) track all reading threads,
1041 * which is very useful for debugging which code path failed to release
1042 * the lock, and for verifying that the *current* thread does hold
1045 * (Unlike a rwlock, which knows that N threads hold it for
1046 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1047 * if any thread holds it for read, even if this thread doesn't).
1049 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1050 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1054 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1056 rrw_exit(&dp->dp_config_rwlock, tag);
1060 dsl_pool_config_held(dsl_pool_t *dp)
1062 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1066 dsl_pool_config_held_writer(dsl_pool_t *dp)
1068 return (RRW_WRITE_HELD(&dp->dp_config_rwlock));
1071 #if defined(_KERNEL) && defined(HAVE_SPL)
1072 EXPORT_SYMBOL(dsl_pool_config_enter);
1073 EXPORT_SYMBOL(dsl_pool_config_exit);
1075 /* zfs_dirty_data_max_percent only applied at module load in arc_init(). */
1076 module_param(zfs_dirty_data_max_percent, int, 0444);
1077 MODULE_PARM_DESC(zfs_dirty_data_max_percent, "percent of ram can be dirty");
1079 /* zfs_dirty_data_max_max_percent only applied at module load in arc_init(). */
1080 module_param(zfs_dirty_data_max_max_percent, int, 0444);
1081 MODULE_PARM_DESC(zfs_dirty_data_max_max_percent,
1082 "zfs_dirty_data_max upper bound as % of RAM");
1084 module_param(zfs_delay_min_dirty_percent, int, 0644);
1085 MODULE_PARM_DESC(zfs_delay_min_dirty_percent, "transaction delay threshold");
1087 module_param(zfs_dirty_data_max, ulong, 0644);
1088 MODULE_PARM_DESC(zfs_dirty_data_max, "determines the dirty space limit");
1090 /* zfs_dirty_data_max_max only applied at module load in arc_init(). */
1091 module_param(zfs_dirty_data_max_max, ulong, 0444);
1092 MODULE_PARM_DESC(zfs_dirty_data_max_max,
1093 "zfs_dirty_data_max upper bound in bytes");
1095 module_param(zfs_dirty_data_sync, ulong, 0644);
1096 MODULE_PARM_DESC(zfs_dirty_data_sync, "sync txg when this much dirty data");
1098 module_param(zfs_delay_scale, ulong, 0644);
1099 MODULE_PARM_DESC(zfs_delay_scale, "how quickly delay approaches infinity");