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) 2013 by Delphix. All rights reserved.
26 #include <sys/dsl_pool.h>
27 #include <sys/dsl_dataset.h>
28 #include <sys/dsl_prop.h>
29 #include <sys/dsl_dir.h>
30 #include <sys/dsl_synctask.h>
31 #include <sys/dsl_scan.h>
32 #include <sys/dnode.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dmu_objset.h>
38 #include <sys/zfs_context.h>
39 #include <sys/fs/zfs.h>
40 #include <sys/zfs_znode.h>
41 #include <sys/spa_impl.h>
42 #include <sys/dsl_deadlist.h>
43 #include <sys/bptree.h>
44 #include <sys/zfeature.h>
45 #include <sys/zil_impl.h>
46 #include <sys/dsl_userhold.h>
48 int zfs_no_write_throttle = 0;
49 int zfs_write_limit_shift = 3; /* 1/8th of physical memory */
50 int zfs_txg_synctime_ms = 1000; /* target millisecs to sync a txg */
52 uint64_t zfs_write_limit_min = 32 << 20; /* min write limit is 32MB */
53 uint64_t zfs_write_limit_max = 0; /* max data payload per txg */
54 uint64_t zfs_write_limit_inflated = 0;
55 uint64_t zfs_write_limit_override = 0;
57 kmutex_t zfs_write_limit_lock;
59 static pgcnt_t old_physmem = 0;
61 SYSCTL_DECL(_vfs_zfs);
62 TUNABLE_INT("vfs.zfs.no_write_throttle", &zfs_no_write_throttle);
63 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_write_throttle, CTLFLAG_RDTUN,
64 &zfs_no_write_throttle, 0, "");
65 TUNABLE_INT("vfs.zfs.write_limit_shift", &zfs_write_limit_shift);
66 SYSCTL_INT(_vfs_zfs, OID_AUTO, write_limit_shift, CTLFLAG_RDTUN,
67 &zfs_write_limit_shift, 0, "2^N of physical memory");
68 SYSCTL_DECL(_vfs_zfs_txg);
69 TUNABLE_INT("vfs.zfs.txg.synctime_ms", &zfs_txg_synctime_ms);
70 SYSCTL_INT(_vfs_zfs_txg, OID_AUTO, synctime_ms, CTLFLAG_RDTUN,
71 &zfs_txg_synctime_ms, 0, "Target milliseconds to sync a txg");
73 TUNABLE_QUAD("vfs.zfs.write_limit_min", &zfs_write_limit_min);
74 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_min, CTLFLAG_RDTUN,
75 &zfs_write_limit_min, 0, "Minimum write limit");
76 TUNABLE_QUAD("vfs.zfs.write_limit_max", &zfs_write_limit_max);
77 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_max, CTLFLAG_RDTUN,
78 &zfs_write_limit_max, 0, "Maximum data payload per txg");
79 TUNABLE_QUAD("vfs.zfs.write_limit_inflated", &zfs_write_limit_inflated);
80 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_inflated, CTLFLAG_RDTUN,
81 &zfs_write_limit_inflated, 0, "Maximum size of the dynamic write limit");
82 TUNABLE_QUAD("vfs.zfs.write_limit_override", &zfs_write_limit_override);
83 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, write_limit_override, CTLFLAG_RDTUN,
84 &zfs_write_limit_override, 0,
85 "Force a txg if dirty buffers exceed this value (bytes)");
88 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
93 err = zap_lookup(dp->dp_meta_objset,
94 dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
95 name, sizeof (obj), 1, &obj);
99 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
103 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
106 blkptr_t *bp = spa_get_rootblkptr(spa);
108 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
110 dp->dp_meta_rootbp = *bp;
111 rrw_init(&dp->dp_config_rwlock, B_TRUE);
112 dp->dp_write_limit = zfs_write_limit_min;
115 txg_list_create(&dp->dp_dirty_datasets,
116 offsetof(dsl_dataset_t, ds_dirty_link));
117 txg_list_create(&dp->dp_dirty_zilogs,
118 offsetof(zilog_t, zl_dirty_link));
119 txg_list_create(&dp->dp_dirty_dirs,
120 offsetof(dsl_dir_t, dd_dirty_link));
121 txg_list_create(&dp->dp_sync_tasks,
122 offsetof(dsl_sync_task_t, dst_node));
124 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
126 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
133 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
136 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
138 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
139 &dp->dp_meta_objset);
149 dsl_pool_open(dsl_pool_t *dp)
156 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
157 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
158 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
159 &dp->dp_root_dir_obj);
163 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
164 NULL, dp, &dp->dp_root_dir);
168 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
172 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
173 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
176 err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
179 err = dsl_dataset_hold_obj(dp,
180 ds->ds_phys->ds_prev_snap_obj, dp,
181 &dp->dp_origin_snap);
182 dsl_dataset_rele(ds, FTAG);
184 dsl_dir_rele(dd, dp);
189 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
190 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
195 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
196 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
199 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
200 dp->dp_meta_objset, obj));
203 if (spa_feature_is_active(dp->dp_spa,
204 &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
205 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
206 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
212 if (spa_feature_is_active(dp->dp_spa,
213 &spa_feature_table[SPA_FEATURE_EMPTY_BPOBJ])) {
214 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
215 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
216 &dp->dp_empty_bpobj);
221 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
222 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
223 &dp->dp_tmp_userrefs_obj);
229 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
232 rrw_exit(&dp->dp_config_rwlock, FTAG);
237 dsl_pool_close(dsl_pool_t *dp)
239 /* drop our references from dsl_pool_open() */
242 * Since we held the origin_snap from "syncing" context (which
243 * includes pool-opening context), it actually only got a "ref"
244 * and not a hold, so just drop that here.
246 if (dp->dp_origin_snap)
247 dsl_dataset_rele(dp->dp_origin_snap, dp);
249 dsl_dir_rele(dp->dp_mos_dir, dp);
251 dsl_dir_rele(dp->dp_free_dir, dp);
253 dsl_dir_rele(dp->dp_root_dir, dp);
255 bpobj_close(&dp->dp_free_bpobj);
257 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
258 if (dp->dp_meta_objset)
259 dmu_objset_evict(dp->dp_meta_objset);
261 txg_list_destroy(&dp->dp_dirty_datasets);
262 txg_list_destroy(&dp->dp_dirty_zilogs);
263 txg_list_destroy(&dp->dp_sync_tasks);
264 txg_list_destroy(&dp->dp_dirty_dirs);
266 arc_flush(dp->dp_spa);
269 rrw_destroy(&dp->dp_config_rwlock);
270 mutex_destroy(&dp->dp_lock);
271 taskq_destroy(dp->dp_vnrele_taskq);
273 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
274 kmem_free(dp, sizeof (dsl_pool_t));
278 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
281 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
282 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
287 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
289 /* create and open the MOS (meta-objset) */
290 dp->dp_meta_objset = dmu_objset_create_impl(spa,
291 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
293 /* create the pool directory */
294 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
295 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
298 /* Initialize scan structures */
299 VERIFY0(dsl_scan_init(dp, txg));
301 /* create and open the root dir */
302 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
303 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
304 NULL, dp, &dp->dp_root_dir));
306 /* create and open the meta-objset dir */
307 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
308 VERIFY0(dsl_pool_open_special_dir(dp,
309 MOS_DIR_NAME, &dp->dp_mos_dir));
311 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
312 /* create and open the free dir */
313 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
315 VERIFY0(dsl_pool_open_special_dir(dp,
316 FREE_DIR_NAME, &dp->dp_free_dir));
318 /* create and open the free_bplist */
319 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
320 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
321 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
322 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
323 dp->dp_meta_objset, obj));
326 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
327 dsl_pool_create_origin(dp, tx);
329 /* create the root dataset */
330 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
332 /* create the root objset */
333 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
334 os = dmu_objset_create_impl(dp->dp_spa, ds,
335 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
337 zfs_create_fs(os, kcred, zplprops, tx);
339 dsl_dataset_rele(ds, FTAG);
343 rrw_exit(&dp->dp_config_rwlock, FTAG);
349 * Account for the meta-objset space in its placeholder dsl_dir.
352 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
353 int64_t used, int64_t comp, int64_t uncomp)
355 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
356 mutex_enter(&dp->dp_lock);
357 dp->dp_mos_used_delta += used;
358 dp->dp_mos_compressed_delta += comp;
359 dp->dp_mos_uncompressed_delta += uncomp;
360 mutex_exit(&dp->dp_lock);
364 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
366 dsl_deadlist_t *dl = arg;
367 dsl_deadlist_insert(dl, bp, tx);
372 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
378 objset_t *mos = dp->dp_meta_objset;
379 hrtime_t start, write_time;
380 uint64_t data_written;
382 list_t synced_datasets;
384 list_create(&synced_datasets, sizeof (dsl_dataset_t),
385 offsetof(dsl_dataset_t, ds_synced_link));
388 * We need to copy dp_space_towrite() before doing
389 * dsl_sync_task_sync(), because
390 * dsl_dataset_snapshot_reserve_space() will increase
391 * dp_space_towrite but not actually write anything.
393 data_written = dp->dp_space_towrite[txg & TXG_MASK];
395 tx = dmu_tx_create_assigned(dp, txg);
397 dp->dp_read_overhead = 0;
400 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
401 while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
403 * We must not sync any non-MOS datasets twice, because
404 * we may have taken a snapshot of them. However, we
405 * may sync newly-created datasets on pass 2.
407 ASSERT(!list_link_active(&ds->ds_synced_link));
408 list_insert_tail(&synced_datasets, ds);
409 dsl_dataset_sync(ds, zio, tx);
411 DTRACE_PROBE(pool_sync__1setup);
414 write_time = gethrtime() - start;
416 DTRACE_PROBE(pool_sync__2rootzio);
419 * After the data blocks have been written (ensured by the zio_wait()
420 * above), update the user/group space accounting.
422 for (ds = list_head(&synced_datasets); ds;
423 ds = list_next(&synced_datasets, ds))
424 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
427 * Sync the datasets again to push out the changes due to
428 * userspace updates. This must be done before we process the
429 * sync tasks, so that any snapshots will have the correct
430 * user accounting information (and we won't get confused
431 * about which blocks are part of the snapshot).
433 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
434 while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
435 ASSERT(list_link_active(&ds->ds_synced_link));
436 dmu_buf_rele(ds->ds_dbuf, ds);
437 dsl_dataset_sync(ds, zio, tx);
442 * Now that the datasets have been completely synced, we can
443 * clean up our in-memory structures accumulated while syncing:
445 * - move dead blocks from the pending deadlist to the on-disk deadlist
446 * - release hold from dsl_dataset_dirty()
448 while (ds = list_remove_head(&synced_datasets)) {
449 objset_t *os = ds->ds_objset;
450 bplist_iterate(&ds->ds_pending_deadlist,
451 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
452 ASSERT(!dmu_objset_is_dirty(os, txg));
453 dmu_buf_rele(ds->ds_dbuf, ds);
457 while (dd = txg_list_remove(&dp->dp_dirty_dirs, txg))
458 dsl_dir_sync(dd, tx);
459 write_time += gethrtime() - start;
462 * The MOS's space is accounted for in the pool/$MOS
463 * (dp_mos_dir). We can't modify the mos while we're syncing
464 * it, so we remember the deltas and apply them here.
466 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
467 dp->dp_mos_uncompressed_delta != 0) {
468 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
469 dp->dp_mos_used_delta,
470 dp->dp_mos_compressed_delta,
471 dp->dp_mos_uncompressed_delta, tx);
472 dp->dp_mos_used_delta = 0;
473 dp->dp_mos_compressed_delta = 0;
474 dp->dp_mos_uncompressed_delta = 0;
478 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
479 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
480 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
481 dmu_objset_sync(mos, zio, tx);
484 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
485 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
487 write_time += gethrtime() - start;
488 DTRACE_PROBE2(pool_sync__4io, hrtime_t, write_time,
489 hrtime_t, dp->dp_read_overhead);
490 write_time -= dp->dp_read_overhead;
493 * If we modify a dataset in the same txg that we want to destroy it,
494 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
495 * dsl_dir_destroy_check() will fail if there are unexpected holds.
496 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
497 * and clearing the hold on it) before we process the sync_tasks.
498 * The MOS data dirtied by the sync_tasks will be synced on the next
501 DTRACE_PROBE(pool_sync__3task);
502 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
503 dsl_sync_task_t *dst;
505 * No more sync tasks should have been added while we
508 ASSERT(spa_sync_pass(dp->dp_spa) == 1);
509 while (dst = txg_list_remove(&dp->dp_sync_tasks, txg))
510 dsl_sync_task_sync(dst, tx);
515 dp->dp_space_towrite[txg & TXG_MASK] = 0;
516 ASSERT(dp->dp_tempreserved[txg & TXG_MASK] == 0);
519 * If the write limit max has not been explicitly set, set it
520 * to a fraction of available physical memory (default 1/8th).
521 * Note that we must inflate the limit because the spa
522 * inflates write sizes to account for data replication.
523 * Check this each sync phase to catch changing memory size.
525 if (physmem != old_physmem && zfs_write_limit_shift) {
526 mutex_enter(&zfs_write_limit_lock);
527 old_physmem = physmem;
528 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
529 zfs_write_limit_inflated = MAX(zfs_write_limit_min,
530 spa_get_asize(dp->dp_spa, zfs_write_limit_max));
531 mutex_exit(&zfs_write_limit_lock);
535 * Attempt to keep the sync time consistent by adjusting the
536 * amount of write traffic allowed into each transaction group.
537 * Weight the throughput calculation towards the current value:
538 * thru = 3/4 old_thru + 1/4 new_thru
540 * Note: write_time is in nanosecs, so write_time/MICROSEC
543 ASSERT(zfs_write_limit_min > 0);
544 if (data_written > zfs_write_limit_min / 8 && write_time > MICROSEC) {
545 uint64_t throughput = data_written / (write_time / MICROSEC);
547 if (dp->dp_throughput)
548 dp->dp_throughput = throughput / 4 +
549 3 * dp->dp_throughput / 4;
551 dp->dp_throughput = throughput;
552 dp->dp_write_limit = MIN(zfs_write_limit_inflated,
553 MAX(zfs_write_limit_min,
554 dp->dp_throughput * zfs_txg_synctime_ms));
559 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
564 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
565 ds = dmu_objset_ds(zilog->zl_os);
566 zil_clean(zilog, txg);
567 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
568 dmu_buf_rele(ds->ds_dbuf, zilog);
570 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
574 * TRUE if the current thread is the tx_sync_thread or if we
575 * are being called from SPA context during pool initialization.
578 dsl_pool_sync_context(dsl_pool_t *dp)
580 return (curthread == dp->dp_tx.tx_sync_thread ||
581 spa_is_initializing(dp->dp_spa));
585 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
587 uint64_t space, resv;
590 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
592 * XXX The intent log is not accounted for, so it must fit
595 * If we're trying to assess whether it's OK to do a free,
596 * cut the reservation in half to allow forward progress
597 * (e.g. make it possible to rm(1) files from a full pool).
599 space = spa_get_dspace(dp->dp_spa);
600 resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
604 return (space - resv);
608 dsl_pool_tempreserve_space(dsl_pool_t *dp, uint64_t space, dmu_tx_t *tx)
610 uint64_t reserved = 0;
611 uint64_t write_limit = (zfs_write_limit_override ?
612 zfs_write_limit_override : dp->dp_write_limit);
614 if (zfs_no_write_throttle) {
615 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK],
621 * Check to see if we have exceeded the maximum allowed IO for
622 * this transaction group. We can do this without locks since
623 * a little slop here is ok. Note that we do the reserved check
624 * with only half the requested reserve: this is because the
625 * reserve requests are worst-case, and we really don't want to
626 * throttle based off of worst-case estimates.
628 if (write_limit > 0) {
629 reserved = dp->dp_space_towrite[tx->tx_txg & TXG_MASK]
630 + dp->dp_tempreserved[tx->tx_txg & TXG_MASK] / 2;
632 if (reserved && reserved > write_limit)
633 return (SET_ERROR(ERESTART));
636 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], space);
639 * If this transaction group is over 7/8ths capacity, delay
640 * the caller 1 clock tick. This will slow down the "fill"
641 * rate until the sync process can catch up with us.
643 if (reserved && reserved > (write_limit - (write_limit >> 3)))
644 txg_delay(dp, tx->tx_txg, 1);
650 dsl_pool_tempreserve_clear(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
652 ASSERT(dp->dp_tempreserved[tx->tx_txg & TXG_MASK] >= space);
653 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], -space);
657 dsl_pool_memory_pressure(dsl_pool_t *dp)
659 uint64_t space_inuse = 0;
662 if (dp->dp_write_limit == zfs_write_limit_min)
665 for (i = 0; i < TXG_SIZE; i++) {
666 space_inuse += dp->dp_space_towrite[i];
667 space_inuse += dp->dp_tempreserved[i];
669 dp->dp_write_limit = MAX(zfs_write_limit_min,
670 MIN(dp->dp_write_limit, space_inuse / 4));
674 dsl_pool_willuse_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
677 mutex_enter(&dp->dp_lock);
678 dp->dp_space_towrite[tx->tx_txg & TXG_MASK] += space;
679 mutex_exit(&dp->dp_lock);
685 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
688 dsl_dataset_t *ds, *prev = NULL;
691 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
695 while (ds->ds_phys->ds_prev_snap_obj != 0) {
696 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
699 dsl_dataset_rele(ds, FTAG);
703 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
705 dsl_dataset_rele(ds, FTAG);
711 prev = dp->dp_origin_snap;
714 * The $ORIGIN can't have any data, or the accounting
717 ASSERT0(prev->ds_phys->ds_bp.blk_birth);
719 /* The origin doesn't get attached to itself */
720 if (ds->ds_object == prev->ds_object) {
721 dsl_dataset_rele(ds, FTAG);
725 dmu_buf_will_dirty(ds->ds_dbuf, tx);
726 ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
727 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
729 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
730 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
732 dmu_buf_will_dirty(prev->ds_dbuf, tx);
733 prev->ds_phys->ds_num_children++;
735 if (ds->ds_phys->ds_next_snap_obj == 0) {
736 ASSERT(ds->ds_prev == NULL);
737 VERIFY0(dsl_dataset_hold_obj(dp,
738 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
742 ASSERT3U(ds->ds_dir->dd_phys->dd_origin_obj, ==, prev->ds_object);
743 ASSERT3U(ds->ds_phys->ds_prev_snap_obj, ==, prev->ds_object);
745 if (prev->ds_phys->ds_next_clones_obj == 0) {
746 dmu_buf_will_dirty(prev->ds_dbuf, tx);
747 prev->ds_phys->ds_next_clones_obj =
748 zap_create(dp->dp_meta_objset,
749 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
751 VERIFY0(zap_add_int(dp->dp_meta_objset,
752 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
754 dsl_dataset_rele(ds, FTAG);
755 if (prev != dp->dp_origin_snap)
756 dsl_dataset_rele(prev, FTAG);
761 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
763 ASSERT(dmu_tx_is_syncing(tx));
764 ASSERT(dp->dp_origin_snap != NULL);
766 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
767 tx, DS_FIND_CHILDREN));
772 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
775 objset_t *mos = dp->dp_meta_objset;
777 if (ds->ds_dir->dd_phys->dd_origin_obj != 0) {
778 dsl_dataset_t *origin;
780 VERIFY0(dsl_dataset_hold_obj(dp,
781 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
783 if (origin->ds_dir->dd_phys->dd_clones == 0) {
784 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
785 origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
786 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
789 VERIFY0(zap_add_int(dp->dp_meta_objset,
790 origin->ds_dir->dd_phys->dd_clones, ds->ds_object, tx));
792 dsl_dataset_rele(origin, FTAG);
798 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
800 ASSERT(dmu_tx_is_syncing(tx));
803 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
804 VERIFY0(dsl_pool_open_special_dir(dp,
805 FREE_DIR_NAME, &dp->dp_free_dir));
808 * We can't use bpobj_alloc(), because spa_version() still
809 * returns the old version, and we need a new-version bpobj with
810 * subobj support. So call dmu_object_alloc() directly.
812 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
813 SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
814 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
815 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
816 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
818 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
819 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
823 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
828 ASSERT(dmu_tx_is_syncing(tx));
829 ASSERT(dp->dp_origin_snap == NULL);
830 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
832 /* create the origin dir, ds, & snap-ds */
833 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
835 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
836 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
837 VERIFY0(dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
838 dp, &dp->dp_origin_snap));
839 dsl_dataset_rele(ds, FTAG);
843 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
845 return (dp->dp_vnrele_taskq);
849 * Walk through the pool-wide zap object of temporary snapshot user holds
853 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
857 objset_t *mos = dp->dp_meta_objset;
858 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
862 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
864 for (zap_cursor_init(&zc, mos, zapobj);
865 zap_cursor_retrieve(&zc, &za) == 0;
866 zap_cursor_advance(&zc)) {
870 htag = strchr(za.za_name, '-');
873 dsobj = strtonum(za.za_name, NULL);
874 dsl_dataset_user_release_tmp(dp, dsobj, htag);
876 zap_cursor_fini(&zc);
880 * Create the pool-wide zap object for storing temporary snapshot holds.
883 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
885 objset_t *mos = dp->dp_meta_objset;
887 ASSERT(dp->dp_tmp_userrefs_obj == 0);
888 ASSERT(dmu_tx_is_syncing(tx));
890 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
891 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
895 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
896 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
898 objset_t *mos = dp->dp_meta_objset;
899 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
903 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
904 ASSERT(dmu_tx_is_syncing(tx));
907 * If the pool was created prior to SPA_VERSION_USERREFS, the
908 * zap object for temporary holds might not exist yet.
912 dsl_pool_user_hold_create_obj(dp, tx);
913 zapobj = dp->dp_tmp_userrefs_obj;
915 return (SET_ERROR(ENOENT));
919 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
921 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
923 error = zap_remove(mos, zapobj, name, tx);
930 * Add a temporary hold for the given dataset object and tag.
933 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
934 uint64_t now, dmu_tx_t *tx)
936 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
940 * Release a temporary hold for the given dataset object and tag.
943 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
946 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, 0,
951 * DSL Pool Configuration Lock
953 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
954 * creation / destruction / rename / property setting). It must be held for
955 * read to hold a dataset or dsl_dir. I.e. you must call
956 * dsl_pool_config_enter() or dsl_pool_hold() before calling
957 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
958 * must be held continuously until all datasets and dsl_dirs are released.
960 * The only exception to this rule is that if a "long hold" is placed on
961 * a dataset, then the dp_config_rwlock may be dropped while the dataset
962 * is still held. The long hold will prevent the dataset from being
963 * destroyed -- the destroy will fail with EBUSY. A long hold can be
964 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
965 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
967 * Legitimate long-holders (including owners) should be long-running, cancelable
968 * tasks that should cause "zfs destroy" to fail. This includes DMU
969 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
970 * "zfs send", and "zfs diff". There are several other long-holders whose
971 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
973 * The usual formula for long-holding would be:
976 * ... perform checks ...
977 * dsl_dataset_long_hold()
979 * ... perform long-running task ...
980 * dsl_dataset_long_rele()
983 * Note that when the long hold is released, the dataset is still held but
984 * the pool is not held. The dataset may change arbitrarily during this time
985 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
986 * dataset except release it.
988 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
989 * or modifying operations.
991 * Modifying operations should generally use dsl_sync_task(). The synctask
992 * infrastructure enforces proper locking strategy with respect to the
993 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
995 * Read-only operations will manually hold the pool, then the dataset, obtain
996 * information from the dataset, then release the pool and dataset.
997 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
1002 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
1007 error = spa_open(name, &spa, tag);
1009 *dp = spa_get_dsl(spa);
1010 dsl_pool_config_enter(*dp, tag);
1016 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1018 dsl_pool_config_exit(dp, tag);
1019 spa_close(dp->dp_spa, tag);
1023 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1026 * We use a "reentrant" reader-writer lock, but not reentrantly.
1028 * The rrwlock can (with the track_all flag) track all reading threads,
1029 * which is very useful for debugging which code path failed to release
1030 * the lock, and for verifying that the *current* thread does hold
1033 * (Unlike a rwlock, which knows that N threads hold it for
1034 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1035 * if any thread holds it for read, even if this thread doesn't).
1037 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1038 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1042 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1044 rrw_exit(&dp->dp_config_rwlock, tag);
1048 dsl_pool_config_held(dsl_pool_t *dp)
1050 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));