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, 2017 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 2016 Nexenta Systems, Inc. All rights reserved.
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>
50 #include <sys/trace_txg.h>
57 * ZFS must limit the rate of incoming writes to the rate at which it is able
58 * to sync data modifications to the backend storage. Throttling by too much
59 * creates an artificial limit; throttling by too little can only be sustained
60 * for short periods and would lead to highly lumpy performance. On a per-pool
61 * basis, ZFS tracks the amount of modified (dirty) data. As operations change
62 * data, the amount of dirty data increases; as ZFS syncs out data, the amount
63 * of dirty data decreases. When the amount of dirty data exceeds a
64 * predetermined threshold further modifications are blocked until the amount
65 * of dirty data decreases (as data is synced out).
67 * The limit on dirty data is tunable, and should be adjusted according to
68 * both the IO capacity and available memory of the system. The larger the
69 * window, the more ZFS is able to aggregate and amortize metadata (and data)
70 * changes. However, memory is a limited resource, and allowing for more dirty
71 * data comes at the cost of keeping other useful data in memory (for example
72 * ZFS data cached by the ARC).
76 * As buffers are modified dsl_pool_willuse_space() increments both the per-
77 * txg (dp_dirty_pertxg[]) and poolwide (dp_dirty_total) accounting of
78 * dirty space used; dsl_pool_dirty_space() decrements those values as data
79 * is synced out from dsl_pool_sync(). While only the poolwide value is
80 * relevant, the per-txg value is useful for debugging. The tunable
81 * zfs_dirty_data_max determines the dirty space limit. Once that value is
82 * exceeded, new writes are halted until space frees up.
84 * The zfs_dirty_data_sync tunable dictates the threshold at which we
85 * ensure that there is a txg syncing (see the comment in txg.c for a full
86 * description of transaction group stages).
88 * The IO scheduler uses both the dirty space limit and current amount of
89 * dirty data as inputs. Those values affect the number of concurrent IOs ZFS
90 * issues. See the comment in vdev_queue.c for details of the IO scheduler.
92 * The delay is also calculated based on the amount of dirty data. See the
93 * comment above dmu_tx_delay() for details.
97 * zfs_dirty_data_max will be set to zfs_dirty_data_max_percent% of all memory,
98 * capped at zfs_dirty_data_max_max. It can also be overridden with a module
101 unsigned long zfs_dirty_data_max = 0;
102 unsigned long zfs_dirty_data_max_max = 0;
103 int zfs_dirty_data_max_percent = 10;
104 int zfs_dirty_data_max_max_percent = 25;
107 * If there is at least this much dirty data, push out a txg.
109 unsigned long 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 unsigned long zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
134 * This determines the number of threads used by the dp_sync_taskq.
136 int zfs_sync_taskq_batch_pct = 75;
139 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
144 err = zap_lookup(dp->dp_meta_objset,
145 dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
146 name, sizeof (obj), 1, &obj);
150 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
154 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
157 blkptr_t *bp = spa_get_rootblkptr(spa);
159 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
161 dp->dp_meta_rootbp = *bp;
162 rrw_init(&dp->dp_config_rwlock, B_TRUE);
166 txg_list_create(&dp->dp_dirty_datasets, spa,
167 offsetof(dsl_dataset_t, ds_dirty_link));
168 txg_list_create(&dp->dp_dirty_zilogs, spa,
169 offsetof(zilog_t, zl_dirty_link));
170 txg_list_create(&dp->dp_dirty_dirs, spa,
171 offsetof(dsl_dir_t, dd_dirty_link));
172 txg_list_create(&dp->dp_sync_tasks, spa,
173 offsetof(dsl_sync_task_t, dst_node));
175 dp->dp_sync_taskq = taskq_create("dp_sync_taskq",
176 zfs_sync_taskq_batch_pct, minclsyspri, 1, INT_MAX,
177 TASKQ_THREADS_CPU_PCT);
179 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
180 cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
182 dp->dp_iput_taskq = taskq_create("z_iput", max_ncpus, defclsyspri,
183 max_ncpus * 8, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
189 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
192 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
195 * Initialize the caller's dsl_pool_t structure before we actually open
196 * the meta objset. This is done because a self-healing write zio may
197 * be issued as part of dmu_objset_open_impl() and the spa needs its
198 * dsl_pool_t initialized in order to handle the write.
202 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
203 &dp->dp_meta_objset);
213 dsl_pool_open(dsl_pool_t *dp)
220 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
221 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
222 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
223 &dp->dp_root_dir_obj);
227 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
228 NULL, dp, &dp->dp_root_dir);
232 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
236 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
237 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
240 err = dsl_dataset_hold_obj(dp,
241 dsl_dir_phys(dd)->dd_head_dataset_obj, FTAG, &ds);
243 err = dsl_dataset_hold_obj(dp,
244 dsl_dataset_phys(ds)->ds_prev_snap_obj, dp,
245 &dp->dp_origin_snap);
246 dsl_dataset_rele(ds, FTAG);
248 dsl_dir_rele(dd, dp);
253 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
254 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
259 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
260 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
263 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
264 dp->dp_meta_objset, obj));
268 * Note: errors ignored, because the leak dir will not exist if we
269 * have not encountered a leak yet.
271 (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
274 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
275 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
276 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
282 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
283 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
284 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
285 &dp->dp_empty_bpobj);
290 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
291 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
292 &dp->dp_tmp_userrefs_obj);
298 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
301 rrw_exit(&dp->dp_config_rwlock, FTAG);
306 dsl_pool_close(dsl_pool_t *dp)
309 * Drop our references from dsl_pool_open().
311 * Since we held the origin_snap from "syncing" context (which
312 * includes pool-opening context), it actually only got a "ref"
313 * and not a hold, so just drop that here.
315 if (dp->dp_origin_snap)
316 dsl_dataset_rele(dp->dp_origin_snap, dp);
318 dsl_dir_rele(dp->dp_mos_dir, dp);
320 dsl_dir_rele(dp->dp_free_dir, dp);
322 dsl_dir_rele(dp->dp_leak_dir, dp);
324 dsl_dir_rele(dp->dp_root_dir, dp);
326 bpobj_close(&dp->dp_free_bpobj);
328 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
329 if (dp->dp_meta_objset)
330 dmu_objset_evict(dp->dp_meta_objset);
332 txg_list_destroy(&dp->dp_dirty_datasets);
333 txg_list_destroy(&dp->dp_dirty_zilogs);
334 txg_list_destroy(&dp->dp_sync_tasks);
335 txg_list_destroy(&dp->dp_dirty_dirs);
337 taskq_destroy(dp->dp_sync_taskq);
340 * We can't set retry to TRUE since we're explicitly specifying
341 * a spa to flush. This is good enough; any missed buffers for
342 * this spa won't cause trouble, and they'll eventually fall
343 * out of the ARC just like any other unused buffer.
345 arc_flush(dp->dp_spa, FALSE);
347 mmp_fini(dp->dp_spa);
350 dmu_buf_user_evict_wait();
352 rrw_destroy(&dp->dp_config_rwlock);
353 mutex_destroy(&dp->dp_lock);
354 cv_destroy(&dp->dp_spaceavail_cv);
355 taskq_destroy(dp->dp_iput_taskq);
357 vmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
358 kmem_free(dp, sizeof (dsl_pool_t));
362 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
365 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
366 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
371 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
373 /* create and open the MOS (meta-objset) */
374 dp->dp_meta_objset = dmu_objset_create_impl(spa,
375 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
377 /* create the pool directory */
378 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
379 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
382 /* Initialize scan structures */
383 VERIFY0(dsl_scan_init(dp, txg));
385 /* create and open the root dir */
386 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
387 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
388 NULL, dp, &dp->dp_root_dir));
390 /* create and open the meta-objset dir */
391 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
392 VERIFY0(dsl_pool_open_special_dir(dp,
393 MOS_DIR_NAME, &dp->dp_mos_dir));
395 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
396 /* create and open the free dir */
397 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
399 VERIFY0(dsl_pool_open_special_dir(dp,
400 FREE_DIR_NAME, &dp->dp_free_dir));
402 /* create and open the free_bplist */
403 obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
404 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
405 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
406 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
407 dp->dp_meta_objset, obj));
410 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
411 dsl_pool_create_origin(dp, tx);
413 /* create the root dataset */
414 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
416 /* create the root objset */
417 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
418 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
419 VERIFY(NULL != (os = dmu_objset_create_impl(dp->dp_spa, ds,
420 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx)));
421 rrw_exit(&ds->ds_bp_rwlock, FTAG);
423 zfs_create_fs(os, kcred, zplprops, tx);
425 dsl_dataset_rele(ds, FTAG);
429 rrw_exit(&dp->dp_config_rwlock, FTAG);
435 * Account for the meta-objset space in its placeholder dsl_dir.
438 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
439 int64_t used, int64_t comp, int64_t uncomp)
441 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
442 mutex_enter(&dp->dp_lock);
443 dp->dp_mos_used_delta += used;
444 dp->dp_mos_compressed_delta += comp;
445 dp->dp_mos_uncompressed_delta += uncomp;
446 mutex_exit(&dp->dp_lock);
450 dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
452 zio_t *zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
453 dmu_objset_sync(dp->dp_meta_objset, zio, tx);
454 VERIFY0(zio_wait(zio));
455 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
456 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
460 dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
462 ASSERT(MUTEX_HELD(&dp->dp_lock));
465 ASSERT3U(-delta, <=, dp->dp_dirty_total);
467 dp->dp_dirty_total += delta;
470 * Note: we signal even when increasing dp_dirty_total.
471 * This ensures forward progress -- each thread wakes the next waiter.
473 if (dp->dp_dirty_total < zfs_dirty_data_max)
474 cv_signal(&dp->dp_spaceavail_cv);
478 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
484 objset_t *mos = dp->dp_meta_objset;
485 list_t synced_datasets;
487 list_create(&synced_datasets, sizeof (dsl_dataset_t),
488 offsetof(dsl_dataset_t, ds_synced_link));
490 tx = dmu_tx_create_assigned(dp, txg);
493 * Write out all dirty blocks of dirty datasets.
495 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
496 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
498 * We must not sync any non-MOS datasets twice, because
499 * we may have taken a snapshot of them. However, we
500 * may sync newly-created datasets on pass 2.
502 ASSERT(!list_link_active(&ds->ds_synced_link));
503 list_insert_tail(&synced_datasets, ds);
504 dsl_dataset_sync(ds, zio, tx);
506 VERIFY0(zio_wait(zio));
509 * We have written all of the accounted dirty data, so our
510 * dp_space_towrite should now be zero. However, some seldom-used
511 * code paths do not adhere to this (e.g. dbuf_undirty(), also
512 * rounding error in dbuf_write_physdone).
513 * Shore up the accounting of any dirtied space now.
515 dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
518 * Update the long range free counter after
519 * we're done syncing user data
521 mutex_enter(&dp->dp_lock);
522 ASSERT(spa_sync_pass(dp->dp_spa) == 1 ||
523 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] == 0);
524 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] = 0;
525 mutex_exit(&dp->dp_lock);
528 * After the data blocks have been written (ensured by the zio_wait()
529 * above), update the user/group space accounting. This happens
530 * in tasks dispatched to dp_sync_taskq, so wait for them before
533 for (ds = list_head(&synced_datasets); ds != NULL;
534 ds = list_next(&synced_datasets, ds)) {
535 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
537 taskq_wait(dp->dp_sync_taskq);
540 * Sync the datasets again to push out the changes due to
541 * userspace updates. This must be done before we process the
542 * sync tasks, so that any snapshots will have the correct
543 * user accounting information (and we won't get confused
544 * about which blocks are part of the snapshot).
546 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
547 while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
548 ASSERT(list_link_active(&ds->ds_synced_link));
549 dmu_buf_rele(ds->ds_dbuf, ds);
550 dsl_dataset_sync(ds, zio, tx);
552 VERIFY0(zio_wait(zio));
555 * Now that the datasets have been completely synced, we can
556 * clean up our in-memory structures accumulated while syncing:
558 * - move dead blocks from the pending deadlist to the on-disk deadlist
559 * - release hold from dsl_dataset_dirty()
561 while ((ds = list_remove_head(&synced_datasets)) != NULL) {
562 dsl_dataset_sync_done(ds, tx);
565 while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
566 dsl_dir_sync(dd, tx);
570 * The MOS's space is accounted for in the pool/$MOS
571 * (dp_mos_dir). We can't modify the mos while we're syncing
572 * it, so we remember the deltas and apply them here.
574 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
575 dp->dp_mos_uncompressed_delta != 0) {
576 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
577 dp->dp_mos_used_delta,
578 dp->dp_mos_compressed_delta,
579 dp->dp_mos_uncompressed_delta, tx);
580 dp->dp_mos_used_delta = 0;
581 dp->dp_mos_compressed_delta = 0;
582 dp->dp_mos_uncompressed_delta = 0;
585 if (!multilist_is_empty(mos->os_dirty_dnodes[txg & TXG_MASK])) {
586 dsl_pool_sync_mos(dp, tx);
590 * If we modify a dataset in the same txg that we want to destroy it,
591 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
592 * dsl_dir_destroy_check() will fail if there are unexpected holds.
593 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
594 * and clearing the hold on it) before we process the sync_tasks.
595 * The MOS data dirtied by the sync_tasks will be synced on the next
598 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
599 dsl_sync_task_t *dst;
601 * No more sync tasks should have been added while we
604 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
605 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
606 dsl_sync_task_sync(dst, tx);
611 DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
615 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
619 while ((zilog = txg_list_head(&dp->dp_dirty_zilogs, txg))) {
620 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
622 * We don't remove the zilog from the dp_dirty_zilogs
623 * list until after we've cleaned it. This ensures that
624 * callers of zilog_is_dirty() receive an accurate
625 * answer when they are racing with the spa sync thread.
627 zil_clean(zilog, txg);
628 (void) txg_list_remove_this(&dp->dp_dirty_zilogs, zilog, txg);
629 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
630 dmu_buf_rele(ds->ds_dbuf, zilog);
632 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
636 * TRUE if the current thread is the tx_sync_thread or if we
637 * are being called from SPA context during pool initialization.
640 dsl_pool_sync_context(dsl_pool_t *dp)
642 return (curthread == dp->dp_tx.tx_sync_thread ||
643 spa_is_initializing(dp->dp_spa) ||
644 taskq_member(dp->dp_sync_taskq, curthread));
648 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
650 uint64_t space, resv;
653 * If we're trying to assess whether it's OK to do a free,
654 * cut the reservation in half to allow forward progress
655 * (e.g. make it possible to rm(1) files from a full pool).
657 space = spa_get_dspace(dp->dp_spa);
658 resv = spa_get_slop_space(dp->dp_spa);
662 return (space - resv);
666 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
668 uint64_t delay_min_bytes =
669 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
672 mutex_enter(&dp->dp_lock);
673 if (dp->dp_dirty_total > zfs_dirty_data_sync)
675 rv = (dp->dp_dirty_total > delay_min_bytes);
676 mutex_exit(&dp->dp_lock);
681 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
684 mutex_enter(&dp->dp_lock);
685 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
686 dsl_pool_dirty_delta(dp, space);
687 mutex_exit(&dp->dp_lock);
692 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
694 ASSERT3S(space, >=, 0);
698 mutex_enter(&dp->dp_lock);
699 if (dp->dp_dirty_pertxg[txg & TXG_MASK] < space) {
700 /* XXX writing something we didn't dirty? */
701 space = dp->dp_dirty_pertxg[txg & TXG_MASK];
703 ASSERT3U(dp->dp_dirty_pertxg[txg & TXG_MASK], >=, space);
704 dp->dp_dirty_pertxg[txg & TXG_MASK] -= space;
705 ASSERT3U(dp->dp_dirty_total, >=, space);
706 dsl_pool_dirty_delta(dp, -space);
707 mutex_exit(&dp->dp_lock);
712 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
715 dsl_dataset_t *ds, *prev = NULL;
718 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
722 while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
723 err = dsl_dataset_hold_obj(dp,
724 dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
726 dsl_dataset_rele(ds, FTAG);
730 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object)
732 dsl_dataset_rele(ds, FTAG);
738 prev = dp->dp_origin_snap;
741 * The $ORIGIN can't have any data, or the accounting
744 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
745 ASSERT0(dsl_dataset_phys(prev)->ds_bp.blk_birth);
746 rrw_exit(&ds->ds_bp_rwlock, FTAG);
748 /* The origin doesn't get attached to itself */
749 if (ds->ds_object == prev->ds_object) {
750 dsl_dataset_rele(ds, FTAG);
754 dmu_buf_will_dirty(ds->ds_dbuf, tx);
755 dsl_dataset_phys(ds)->ds_prev_snap_obj = prev->ds_object;
756 dsl_dataset_phys(ds)->ds_prev_snap_txg =
757 dsl_dataset_phys(prev)->ds_creation_txg;
759 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
760 dsl_dir_phys(ds->ds_dir)->dd_origin_obj = prev->ds_object;
762 dmu_buf_will_dirty(prev->ds_dbuf, tx);
763 dsl_dataset_phys(prev)->ds_num_children++;
765 if (dsl_dataset_phys(ds)->ds_next_snap_obj == 0) {
766 ASSERT(ds->ds_prev == NULL);
767 VERIFY0(dsl_dataset_hold_obj(dp,
768 dsl_dataset_phys(ds)->ds_prev_snap_obj,
773 ASSERT3U(dsl_dir_phys(ds->ds_dir)->dd_origin_obj, ==, prev->ds_object);
774 ASSERT3U(dsl_dataset_phys(ds)->ds_prev_snap_obj, ==, prev->ds_object);
776 if (dsl_dataset_phys(prev)->ds_next_clones_obj == 0) {
777 dmu_buf_will_dirty(prev->ds_dbuf, tx);
778 dsl_dataset_phys(prev)->ds_next_clones_obj =
779 zap_create(dp->dp_meta_objset,
780 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
782 VERIFY0(zap_add_int(dp->dp_meta_objset,
783 dsl_dataset_phys(prev)->ds_next_clones_obj, ds->ds_object, tx));
785 dsl_dataset_rele(ds, FTAG);
786 if (prev != dp->dp_origin_snap)
787 dsl_dataset_rele(prev, FTAG);
792 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
794 ASSERT(dmu_tx_is_syncing(tx));
795 ASSERT(dp->dp_origin_snap != NULL);
797 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
798 tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
803 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
806 objset_t *mos = dp->dp_meta_objset;
808 if (dsl_dir_phys(ds->ds_dir)->dd_origin_obj != 0) {
809 dsl_dataset_t *origin;
811 VERIFY0(dsl_dataset_hold_obj(dp,
812 dsl_dir_phys(ds->ds_dir)->dd_origin_obj, FTAG, &origin));
814 if (dsl_dir_phys(origin->ds_dir)->dd_clones == 0) {
815 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
816 dsl_dir_phys(origin->ds_dir)->dd_clones =
817 zap_create(mos, DMU_OT_DSL_CLONES, DMU_OT_NONE,
821 VERIFY0(zap_add_int(dp->dp_meta_objset,
822 dsl_dir_phys(origin->ds_dir)->dd_clones,
825 dsl_dataset_rele(origin, FTAG);
831 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
835 ASSERT(dmu_tx_is_syncing(tx));
837 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
838 VERIFY0(dsl_pool_open_special_dir(dp,
839 FREE_DIR_NAME, &dp->dp_free_dir));
842 * We can't use bpobj_alloc(), because spa_version() still
843 * returns the old version, and we need a new-version bpobj with
844 * subobj support. So call dmu_object_alloc() directly.
846 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
847 SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
848 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
849 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
850 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
852 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
853 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN | DS_FIND_SERIALIZE));
857 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
862 ASSERT(dmu_tx_is_syncing(tx));
863 ASSERT(dp->dp_origin_snap == NULL);
864 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
866 /* create the origin dir, ds, & snap-ds */
867 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
869 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
870 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
871 VERIFY0(dsl_dataset_hold_obj(dp, dsl_dataset_phys(ds)->ds_prev_snap_obj,
872 dp, &dp->dp_origin_snap));
873 dsl_dataset_rele(ds, FTAG);
877 dsl_pool_iput_taskq(dsl_pool_t *dp)
879 return (dp->dp_iput_taskq);
883 * Walk through the pool-wide zap object of temporary snapshot user holds
887 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
891 objset_t *mos = dp->dp_meta_objset;
892 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
897 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
899 holds = fnvlist_alloc();
901 for (zap_cursor_init(&zc, mos, zapobj);
902 zap_cursor_retrieve(&zc, &za) == 0;
903 zap_cursor_advance(&zc)) {
907 htag = strchr(za.za_name, '-');
910 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
911 tags = fnvlist_alloc();
912 fnvlist_add_boolean(tags, htag);
913 fnvlist_add_nvlist(holds, za.za_name, tags);
916 fnvlist_add_boolean(tags, htag);
919 dsl_dataset_user_release_tmp(dp, holds);
921 zap_cursor_fini(&zc);
925 * Create the pool-wide zap object for storing temporary snapshot holds.
928 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
930 objset_t *mos = dp->dp_meta_objset;
932 ASSERT(dp->dp_tmp_userrefs_obj == 0);
933 ASSERT(dmu_tx_is_syncing(tx));
935 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
936 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
940 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
941 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
943 objset_t *mos = dp->dp_meta_objset;
944 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
948 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
949 ASSERT(dmu_tx_is_syncing(tx));
952 * If the pool was created prior to SPA_VERSION_USERREFS, the
953 * zap object for temporary holds might not exist yet.
957 dsl_pool_user_hold_create_obj(dp, tx);
958 zapobj = dp->dp_tmp_userrefs_obj;
960 return (SET_ERROR(ENOENT));
964 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
966 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
968 error = zap_remove(mos, zapobj, name, tx);
975 * Add a temporary hold for the given dataset object and tag.
978 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
979 uint64_t now, dmu_tx_t *tx)
981 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
985 * Release a temporary hold for the given dataset object and tag.
988 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
991 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0,
996 * DSL Pool Configuration Lock
998 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
999 * creation / destruction / rename / property setting). It must be held for
1000 * read to hold a dataset or dsl_dir. I.e. you must call
1001 * dsl_pool_config_enter() or dsl_pool_hold() before calling
1002 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
1003 * must be held continuously until all datasets and dsl_dirs are released.
1005 * The only exception to this rule is that if a "long hold" is placed on
1006 * a dataset, then the dp_config_rwlock may be dropped while the dataset
1007 * is still held. The long hold will prevent the dataset from being
1008 * destroyed -- the destroy will fail with EBUSY. A long hold can be
1009 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
1010 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
1012 * Legitimate long-holders (including owners) should be long-running, cancelable
1013 * tasks that should cause "zfs destroy" to fail. This includes DMU
1014 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
1015 * "zfs send", and "zfs diff". There are several other long-holders whose
1016 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
1018 * The usual formula for long-holding would be:
1020 * dsl_dataset_hold()
1021 * ... perform checks ...
1022 * dsl_dataset_long_hold()
1024 * ... perform long-running task ...
1025 * dsl_dataset_long_rele()
1026 * dsl_dataset_rele()
1028 * Note that when the long hold is released, the dataset is still held but
1029 * the pool is not held. The dataset may change arbitrarily during this time
1030 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
1031 * dataset except release it.
1033 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
1034 * or modifying operations.
1036 * Modifying operations should generally use dsl_sync_task(). The synctask
1037 * infrastructure enforces proper locking strategy with respect to the
1038 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
1040 * Read-only operations will manually hold the pool, then the dataset, obtain
1041 * information from the dataset, then release the pool and dataset.
1042 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1047 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1052 error = spa_open(name, &spa, tag);
1054 *dp = spa_get_dsl(spa);
1055 dsl_pool_config_enter(*dp, tag);
1061 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1063 dsl_pool_config_exit(dp, tag);
1064 spa_close(dp->dp_spa, tag);
1068 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1071 * We use a "reentrant" reader-writer lock, but not reentrantly.
1073 * The rrwlock can (with the track_all flag) track all reading threads,
1074 * which is very useful for debugging which code path failed to release
1075 * the lock, and for verifying that the *current* thread does hold
1078 * (Unlike a rwlock, which knows that N threads hold it for
1079 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1080 * if any thread holds it for read, even if this thread doesn't).
1082 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1083 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1087 dsl_pool_config_enter_prio(dsl_pool_t *dp, void *tag)
1089 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1090 rrw_enter_read_prio(&dp->dp_config_rwlock, tag);
1094 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1096 rrw_exit(&dp->dp_config_rwlock, tag);
1100 dsl_pool_config_held(dsl_pool_t *dp)
1102 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1106 dsl_pool_config_held_writer(dsl_pool_t *dp)
1108 return (RRW_WRITE_HELD(&dp->dp_config_rwlock));
1111 #if defined(_KERNEL) && defined(HAVE_SPL)
1112 EXPORT_SYMBOL(dsl_pool_config_enter);
1113 EXPORT_SYMBOL(dsl_pool_config_exit);
1116 /* zfs_dirty_data_max_percent only applied at module load in arc_init(). */
1117 module_param(zfs_dirty_data_max_percent, int, 0444);
1118 MODULE_PARM_DESC(zfs_dirty_data_max_percent, "percent of ram can be dirty");
1120 /* zfs_dirty_data_max_max_percent only applied at module load in arc_init(). */
1121 module_param(zfs_dirty_data_max_max_percent, int, 0444);
1122 MODULE_PARM_DESC(zfs_dirty_data_max_max_percent,
1123 "zfs_dirty_data_max upper bound as % of RAM");
1125 module_param(zfs_delay_min_dirty_percent, int, 0644);
1126 MODULE_PARM_DESC(zfs_delay_min_dirty_percent, "transaction delay threshold");
1128 module_param(zfs_dirty_data_max, ulong, 0644);
1129 MODULE_PARM_DESC(zfs_dirty_data_max, "determines the dirty space limit");
1131 /* zfs_dirty_data_max_max only applied at module load in arc_init(). */
1132 module_param(zfs_dirty_data_max_max, ulong, 0444);
1133 MODULE_PARM_DESC(zfs_dirty_data_max_max,
1134 "zfs_dirty_data_max upper bound in bytes");
1136 module_param(zfs_dirty_data_sync, ulong, 0644);
1137 MODULE_PARM_DESC(zfs_dirty_data_sync, "sync txg when this much dirty data");
1139 module_param(zfs_delay_scale, ulong, 0644);
1140 MODULE_PARM_DESC(zfs_delay_scale, "how quickly delay approaches infinity");
1142 module_param(zfs_sync_taskq_batch_pct, int, 0644);
1143 MODULE_PARM_DESC(zfs_sync_taskq_batch_pct,
1144 "max percent of CPUs that are used to sync dirty data");