Print this page
NEX-9673 Add capability to replicate cloned datasets relative to origin
Reviewed by: Alex Deiter <alex.deiter@nexenta.com>
Reviewed by: Sanjay Nadkarni <sanjay.nadkarni@nexenta.com>
NEX-5085 implement async delete for large files
Reviewed by: Saso Kiselkov <saso.kiselkov@nexenta.com>
Reviewed by: Roman Strashkin <roman.strashkin@nexenta.com>
Reviewed by: Sanjay Nadkarni <sanjay.nadkarni@nexenta.com>
Revert "NEX-5085 implement async delete for large files"
This reverts commit 65aa8f42d93fcbd6e0efb3d4883170a20d760611.
Fails regression testing of the zfs test mirror_stress_004.
NEX-5085 implement async delete for large files
Reviewed by: Sanjay Nadkarni <sanjay.nadkarni@nexenta.com>
Reviewed by: Roman Strashkin <roman.strashkin@nexenta.com>
Reviewed by: Kirill Davydychev <kirill.davydychev@nexenta.com>
NEX-7479 Autosnap may dispatch duplicated sync-tasks
Reviewed by: Alex Deiter <alex.deiter@nexenta.com>
Reviewed by: Sanjay Nadkarni <sanjay.nadkarni@nexent.com>
NEX-7543 backout async delete (NEX-5085 and NEX-6151)
Reviewed by: Roman Strashkin <roman.strashkin@nexenta.com>
Reviewed by: Yuri Pankov <yuri.pankov@nexenta.com>
Reviewed by: Sanjay Nadkarni <sanjay.nadkarni@nexenta.com>
NEX-5795 Rename 'wrc' as 'wbc' in the source and in the tech docs
Reviewed by: Alex Aizman <alex.aizman@nexenta.com>
Reviewed by: Sanjay Nadkarni <sanjay.nadkarni@nexenta.com>
Reviewed by: Alek Pinchuk <alek.pinchuk@nexenta.com>
NEX-5078 Want ability to see progress of freeing data and how much is left to free after large file delete patch
Reviewed by: Saso Kiselkov <saso.kiselkov@nexenta.com>
Reviewed by: Sanjay Nadkarni <sanjay.nadkarni@nexenta.com>
NEX-5085 implement async delete for large files
Reviewed by: Marcel Telka <marcel.telka@nexenta.com>
Reviewed by: Saso Kiselkov <saso.kiselkov@nexenta.com>
NEX-5024 Slow performance with a single large file delete
Reviewed by: Saso Kiselkov <saso.kiselkov@nexenta.com>
Reviewed by: Sanjay Nadkarni <sanjay.nadkarni@nexenta.com>
NEX-4830 writecache=off leaks data on special vdev (the data will never migrate)
Reviewed by: Alex Aizman <alex.aizman@nexenta.com>
Reviewed by: Saso Kiselkov <saso.kiselkov@nexenta.com>
5981 Deadlock in dmu_objset_find_dp
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Dan McDonald <danmcd@omniti.com>
Approved by: Robert Mustacchi <rm@joyent.com>
5269 zpool import slow
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <george@delphix.com>
Reviewed by: Dan McDonald <danmcd@omniti.com>
Approved by: Dan McDonald <danmcd@omniti.com>
NEX-4476 WRC: Allow to use write back cache per tree of datasets
Reviewed by: Alek Pinchuk <alek.pinchuk@nexenta.com>
Reviewed by: Alex Aizman <alex.aizman@nexenta.com>
Revert "NEX-4476 WRC: Allow to use write back cache per tree of datasets"
This reverts commit fe97b74444278a6f36fec93179133641296312da.
NEX-4476 WRC: Allow to use write back cache per tree of datasets
Reviewed by: Alek Pinchuk <alek.pinchuk@nexenta.com>
Reviewed by: Alex Aizman <alex.aizman@nexenta.com>
NEX-3964 It should not be allowed to rename a snapshot that its new name is matched to the prefix of in-kernel autosnapshots
Reviewed by: Alek Pinchuk <alek.pinchuk@nexenta.com>
Reviewed by: Josef 'Jeff' Sipek <josef.sipek@nexenta.com>
NEX-3485 Deferred deletes causing loss of service for NFS clients on cluster failover
Reviewed by: Marcel Telka <marcel.telka@nexenta.com>
Reviewed by: Saso Kiselkov <saso.kiselkov@nexenta.com>
NEX-3558 KRRP Integration
NEX-3079 port illumos ARC improvements
Fixup merge results
re #13253 rb4328 ssh: openssl version checking needs updating
re #11441 rb4292 panic in apic_record_rdt_entry on VMware hardware version 9
re #12619, rb4287 Deadlocked zfs txg processing in dsl_sync_task_group_sync()
re #12585 rb4049 ZFS++ work port - refactoring to improve separation of open/closed code, bug fixes, performance improvements - open code
Bug 11205: add missing libzfs_closed_stubs.c to fix opensource-only build.
ZFS plus work: special vdevs, cos, cos/vdev properties


  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  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]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  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 (c) 2014 Integros [integros.com]
  27  * Copyright 2016 Nexenta Systems, Inc.  All rights reserved.
  28  */
  29 

  30 #include <sys/dsl_pool.h>
  31 #include <sys/dsl_dataset.h>
  32 #include <sys/dsl_prop.h>
  33 #include <sys/dsl_dir.h>
  34 #include <sys/dsl_synctask.h>

  35 #include <sys/dsl_scan.h>
  36 #include <sys/dnode.h>
  37 #include <sys/dmu_tx.h>
  38 #include <sys/dmu_objset.h>

  39 #include <sys/arc.h>
  40 #include <sys/zap.h>
  41 #include <sys/zio.h>
  42 #include <sys/zfs_context.h>
  43 #include <sys/fs/zfs.h>
  44 #include <sys/zfs_znode.h>
  45 #include <sys/spa_impl.h>
  46 #include <sys/dsl_deadlist.h>
  47 #include <sys/bptree.h>
  48 #include <sys/zfeature.h>
  49 #include <sys/zil_impl.h>
  50 #include <sys/dsl_userhold.h>
  51 



  52 /*
  53  * ZFS Write Throttle
  54  * ------------------
  55  *
  56  * ZFS must limit the rate of incoming writes to the rate at which it is able
  57  * to sync data modifications to the backend storage. Throttling by too much
  58  * creates an artificial limit; throttling by too little can only be sustained
  59  * for short periods and would lead to highly lumpy performance. On a per-pool
  60  * basis, ZFS tracks the amount of modified (dirty) data. As operations change
  61  * data, the amount of dirty data increases; as ZFS syncs out data, the amount
  62  * of dirty data decreases. When the amount of dirty data exceeds a
  63  * predetermined threshold further modifications are blocked until the amount
  64  * of dirty data decreases (as data is synced out).
  65  *
  66  * The limit on dirty data is tunable, and should be adjusted according to
  67  * both the IO capacity and available memory of the system. The larger the
  68  * window, the more ZFS is able to aggregate and amortize metadata (and data)
  69  * changes. However, memory is a limited resource, and allowing for more dirty
  70  * data comes at the cost of keeping other useful data in memory (for example
  71  * ZFS data cached by the ARC).


 145  * taskq used by zil_clean(); they determine the number of taskq entries
 146  * that are pre-populated when the taskq is first created (via the
 147  * "zfs_zil_clean_taskq_minalloc" tunable) and the maximum number of
 148  * taskq entries that are cached after an on-demand allocation (via the
 149  * "zfs_zil_clean_taskq_maxalloc").
 150  *
 151  * The idea being, we want to try reasonably hard to ensure there will
 152  * already be a taskq entry pre-allocated by the time that it is needed
 153  * by zil_clean(). This way, we can avoid the possibility of an
 154  * on-demand allocation of a new taskq entry from failing, which would
 155  * result in zil_itxg_clean() being called synchronously from zil_clean()
 156  * (which can adversely affect performance of spa_sync()).
 157  *
 158  * Additionally, the number of threads used by the taskq can be
 159  * configured via the "zfs_zil_clean_taskq_nthr_pct" tunable.
 160  */
 161 int zfs_zil_clean_taskq_nthr_pct = 100;
 162 int zfs_zil_clean_taskq_minalloc = 1024;
 163 int zfs_zil_clean_taskq_maxalloc = 1024 * 1024;
 164 






 165 int
 166 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
 167 {
 168         uint64_t obj;
 169         int err;
 170 
 171         err = zap_lookup(dp->dp_meta_objset,
 172             dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
 173             name, sizeof (obj), 1, &obj);
 174         if (err)
 175                 return (err);
 176 
 177         return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
 178 }
 179 
 180 static dsl_pool_t *
 181 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
 182 {
 183         dsl_pool_t *dp;
 184         blkptr_t *bp = spa_get_rootblkptr(spa);
 185 
 186         dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
 187         dp->dp_spa = spa;
 188         dp->dp_meta_rootbp = *bp;
 189         rrw_init(&dp->dp_config_rwlock, B_TRUE);



 190         txg_init(dp, txg);
 191 
 192         txg_list_create(&dp->dp_dirty_datasets, spa,
 193             offsetof(dsl_dataset_t, ds_dirty_link));
 194         txg_list_create(&dp->dp_dirty_zilogs, spa,
 195             offsetof(zilog_t, zl_dirty_link));
 196         txg_list_create(&dp->dp_dirty_dirs, spa,
 197             offsetof(dsl_dir_t, dd_dirty_link));
 198         txg_list_create(&dp->dp_sync_tasks, spa,
 199             offsetof(dsl_sync_task_t, dst_node));
 200 
 201         dp->dp_sync_taskq = taskq_create("dp_sync_taskq",
 202             zfs_sync_taskq_batch_pct, minclsyspri, 1, INT_MAX,
 203             TASKQ_THREADS_CPU_PCT);
 204 
 205         dp->dp_zil_clean_taskq = taskq_create("dp_zil_clean_taskq",
 206             zfs_zil_clean_taskq_nthr_pct, minclsyspri,
 207             zfs_zil_clean_taskq_minalloc,
 208             zfs_zil_clean_taskq_maxalloc,
 209             TASKQ_PREPOPULATE | TASKQ_THREADS_CPU_PCT);
 210 
 211         mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
 212         cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
 213 
 214         dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
 215             1, 4, 0);

 216 
 217         return (dp);
 218 }
 219 
 220 int
 221 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
 222 {
 223         int err;
 224         dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
 225 
 226         err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
 227             &dp->dp_meta_objset);
 228         if (err != 0)
 229                 dsl_pool_close(dp);
 230         else
 231                 *dpp = dp;
 232 
 233         return (err);
 234 }
 235 


 271                 }
 272                 dsl_dir_rele(dd, dp);
 273                 if (err)
 274                         goto out;
 275         }
 276 
 277         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
 278                 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
 279                     &dp->dp_free_dir);
 280                 if (err)
 281                         goto out;
 282 
 283                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 284                     DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
 285                 if (err)
 286                         goto out;
 287                 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
 288                     dp->dp_meta_objset, obj));
 289         }
 290 
 291         if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS)) {
 292                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 293                     DMU_POOL_OBSOLETE_BPOBJ, sizeof (uint64_t), 1, &obj);
 294                 if (err == 0) {
 295                         VERIFY0(bpobj_open(&dp->dp_obsolete_bpobj,
 296                             dp->dp_meta_objset, obj));
 297                 } else if (err == ENOENT) {
 298                         /*
 299                          * We might not have created the remap bpobj yet.

 300                          */
 301                         err = 0;
 302                 } else {
 303                         goto out;
 304                 }
 305         }
 306 
 307         /*
 308          * Note: errors ignored, because the these special dirs, used for
 309          * space accounting, are only created on demand.
 310          */
 311         (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
 312             &dp->dp_leak_dir);
 313 
 314         if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
 315                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 316                     DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
 317                     &dp->dp_bptree_obj);
 318                 if (err != 0)
 319                         goto out;
 320         }
 321 
 322         if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
 323                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 324                     DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
 325                     &dp->dp_empty_bpobj);
 326                 if (err != 0)
 327                         goto out;
 328         }
 329 
 330         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,


 335         if (err)
 336                 goto out;
 337 
 338         err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
 339 
 340 out:
 341         rrw_exit(&dp->dp_config_rwlock, FTAG);
 342         return (err);
 343 }
 344 
 345 void
 346 dsl_pool_close(dsl_pool_t *dp)
 347 {
 348         /*
 349          * Drop our references from dsl_pool_open().
 350          *
 351          * Since we held the origin_snap from "syncing" context (which
 352          * includes pool-opening context), it actually only got a "ref"
 353          * and not a hold, so just drop that here.
 354          */
 355         if (dp->dp_origin_snap != NULL)
 356                 dsl_dataset_rele(dp->dp_origin_snap, dp);
 357         if (dp->dp_mos_dir != NULL)
 358                 dsl_dir_rele(dp->dp_mos_dir, dp);
 359         if (dp->dp_free_dir != NULL)
 360                 dsl_dir_rele(dp->dp_free_dir, dp);
 361         if (dp->dp_leak_dir != NULL)
 362                 dsl_dir_rele(dp->dp_leak_dir, dp);
 363         if (dp->dp_root_dir != NULL)
 364                 dsl_dir_rele(dp->dp_root_dir, dp);
 365 
 366         bpobj_close(&dp->dp_free_bpobj);
 367         bpobj_close(&dp->dp_obsolete_bpobj);
 368 
 369         /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
 370         if (dp->dp_meta_objset != NULL)
 371                 dmu_objset_evict(dp->dp_meta_objset);
 372 
 373         txg_list_destroy(&dp->dp_dirty_datasets);
 374         txg_list_destroy(&dp->dp_dirty_zilogs);
 375         txg_list_destroy(&dp->dp_sync_tasks);
 376         txg_list_destroy(&dp->dp_dirty_dirs);
 377 
 378         taskq_destroy(dp->dp_zil_clean_taskq);
 379         taskq_destroy(dp->dp_sync_taskq);
 380 
 381         /*
 382          * We can't set retry to TRUE since we're explicitly specifying
 383          * a spa to flush. This is good enough; any missed buffers for
 384          * this spa won't cause trouble, and they'll eventually fall
 385          * out of the ARC just like any other unused buffer.
 386          */
 387         arc_flush(dp->dp_spa, FALSE);
 388 
 389         txg_fini(dp);
 390         dsl_scan_fini(dp);
 391         dmu_buf_user_evict_wait();
 392 
 393         rrw_destroy(&dp->dp_config_rwlock);
 394         mutex_destroy(&dp->dp_lock);
 395         taskq_destroy(dp->dp_vnrele_taskq);
 396         if (dp->dp_blkstats != NULL)
 397                 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 398         kmem_free(dp, sizeof (dsl_pool_t));
 399 }
 400 
 401 void
 402 dsl_pool_create_obsolete_bpobj(dsl_pool_t *dp, dmu_tx_t *tx)
 403 {
 404         uint64_t obj;
 405         /*
 406          * Currently, we only create the obsolete_bpobj where there are
 407          * indirect vdevs with referenced mappings.
 408          */
 409         ASSERT(spa_feature_is_active(dp->dp_spa, SPA_FEATURE_DEVICE_REMOVAL));
 410         /* create and open the obsolete_bpobj */
 411         obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
 412         VERIFY0(bpobj_open(&dp->dp_obsolete_bpobj, dp->dp_meta_objset, obj));
 413         VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 414             DMU_POOL_OBSOLETE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
 415         spa_feature_incr(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
 416 }
 417 
 418 void
 419 dsl_pool_destroy_obsolete_bpobj(dsl_pool_t *dp, dmu_tx_t *tx)
 420 {
 421         spa_feature_decr(dp->dp_spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
 422         VERIFY0(zap_remove(dp->dp_meta_objset,
 423             DMU_POOL_DIRECTORY_OBJECT,
 424             DMU_POOL_OBSOLETE_BPOBJ, tx));
 425         bpobj_free(dp->dp_meta_objset,
 426             dp->dp_obsolete_bpobj.bpo_object, tx);
 427         bpobj_close(&dp->dp_obsolete_bpobj);
 428 }
 429 
 430 dsl_pool_t *
 431 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
 432 {
 433         int err;
 434         dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
 435         dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
 436         dsl_dataset_t *ds;
 437         uint64_t obj;
 438 
 439         rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
 440 
 441         /* create and open the MOS (meta-objset) */
 442         dp->dp_meta_objset = dmu_objset_create_impl(spa,
 443             NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
 444 
 445         /* create the pool directory */
 446         err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 447             DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
 448         ASSERT0(err);
 449 


 536                 ASSERT3U(-delta, <=, dp->dp_dirty_total);
 537 
 538         dp->dp_dirty_total += delta;
 539 
 540         /*
 541          * Note: we signal even when increasing dp_dirty_total.
 542          * This ensures forward progress -- each thread wakes the next waiter.
 543          */
 544         if (dp->dp_dirty_total < zfs_dirty_data_max)
 545                 cv_signal(&dp->dp_spaceavail_cv);
 546 }
 547 
 548 void
 549 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
 550 {
 551         zio_t *zio;
 552         dmu_tx_t *tx;
 553         dsl_dir_t *dd;
 554         dsl_dataset_t *ds;
 555         objset_t *mos = dp->dp_meta_objset;

 556         list_t synced_datasets;







 557 















 558         list_create(&synced_datasets, sizeof (dsl_dataset_t),
 559             offsetof(dsl_dataset_t, ds_synced_link));
 560 
 561         tx = dmu_tx_create_assigned(dp, txg);
 562 









 563         /*






































































 564          * Write out all dirty blocks of dirty datasets.
 565          */
 566         zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
 567         while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {

 568                 /*
 569                  * We must not sync any non-MOS datasets twice, because
 570                  * we may have taken a snapshot of them.  However, we
 571                  * may sync newly-created datasets on pass 2.
 572                  */
 573                 ASSERT(!list_link_active(&ds->ds_synced_link));
 574                 list_insert_tail(&synced_datasets, ds);
 575                 dsl_dataset_sync(ds, zio, tx);
 576         }

 577         VERIFY0(zio_wait(zio));
 578 







 579         /*































 580          * We have written all of the accounted dirty data, so our
 581          * dp_space_towrite should now be zero.  However, some seldom-used
 582          * code paths do not adhere to this (e.g. dbuf_undirty(), also
 583          * rounding error in dbuf_write_physdone).
 584          * Shore up the accounting of any dirtied space now.
 585          */
 586         dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
 587 
 588         /*
 589          * Update the long range free counter after
 590          * we're done syncing user data
 591          */
 592         mutex_enter(&dp->dp_lock);
 593         ASSERT(spa_sync_pass(dp->dp_spa) == 1 ||
 594             dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] == 0);


 595         dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] = 0;
 596         mutex_exit(&dp->dp_lock);
 597 
 598         /*
 599          * After the data blocks have been written (ensured by the zio_wait()
 600          * above), update the user/group space accounting.  This happens
 601          * in tasks dispatched to dp_sync_taskq, so wait for them before
 602          * continuing.
 603          */
 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);
 607         }
 608         taskq_wait(dp->dp_sync_taskq);
 609 
 610         /*
 611          * Sync the datasets again to push out the changes due to
 612          * userspace updates.  This must be done before we process the
 613          * sync tasks, so that any snapshots will have the correct
 614          * user accounting information (and we won't get confused
 615          * about which blocks are part of the snapshot).
 616          */

 617         zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
 618         while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
 619                 ASSERT(list_link_active(&ds->ds_synced_link));
 620                 dmu_buf_rele(ds->ds_dbuf, ds);
 621                 dsl_dataset_sync(ds, zio, tx);
 622         }
 623         VERIFY0(zio_wait(zio));
 624 
 625         /*
 626          * Now that the datasets have been completely synced, we can
 627          * clean up our in-memory structures accumulated while syncing:
 628          *
 629          *  - move dead blocks from the pending deadlist to the on-disk deadlist
 630          *  - release hold from dsl_dataset_dirty()
 631          */
 632         while ((ds = list_remove_head(&synced_datasets)) != NULL) {
 633                 dsl_dataset_sync_done(ds, tx);
 634         }
 635         while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
 636                 dsl_dir_sync(dd, tx);


 648                     dp->dp_mos_compressed_delta,
 649                     dp->dp_mos_uncompressed_delta, tx);
 650                 dp->dp_mos_used_delta = 0;
 651                 dp->dp_mos_compressed_delta = 0;
 652                 dp->dp_mos_uncompressed_delta = 0;
 653         }
 654 
 655         if (!multilist_is_empty(mos->os_dirty_dnodes[txg & TXG_MASK])) {
 656                 dsl_pool_sync_mos(dp, tx);
 657         }
 658 
 659         /*
 660          * If we modify a dataset in the same txg that we want to destroy it,
 661          * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
 662          * dsl_dir_destroy_check() will fail if there are unexpected holds.
 663          * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
 664          * and clearing the hold on it) before we process the sync_tasks.
 665          * The MOS data dirtied by the sync_tasks will be synced on the next
 666          * pass.
 667          */

 668         if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
 669                 dsl_sync_task_t *dst;
 670                 /*
 671                  * No more sync tasks should have been added while we
 672                  * were syncing.
 673                  */
 674                 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
 675                 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
 676                         dsl_sync_task_sync(dst, tx);
 677         }
 678 





 679         dmu_tx_commit(tx);
 680 
 681         DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
 682 }
 683 
 684 void
 685 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
 686 {
 687         zilog_t *zilog;
 688 
 689         while (zilog = txg_list_head(&dp->dp_dirty_zilogs, txg)) {
 690                 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
 691                 /*
 692                  * We don't remove the zilog from the dp_dirty_zilogs
 693                  * list until after we've cleaned it. This ensures that
 694                  * callers of zilog_is_dirty() receive an accurate
 695                  * answer when they are racing with the spa sync thread.
 696                  */
 697                 zil_clean(zilog, txg);
 698                 (void) txg_list_remove_this(&dp->dp_dirty_zilogs, zilog, txg);


 722         /*
 723          * If we're trying to assess whether it's OK to do a free,
 724          * cut the reservation in half to allow forward progress
 725          * (e.g. make it possible to rm(1) files from a full pool).
 726          */
 727         space = spa_get_dspace(dp->dp_spa);
 728         resv = spa_get_slop_space(dp->dp_spa);
 729         if (netfree)
 730                 resv >>= 1;
 731 
 732         return (space - resv);
 733 }
 734 
 735 boolean_t
 736 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
 737 {
 738         uint64_t delay_min_bytes =
 739             zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
 740         boolean_t rv;
 741 
 742         mutex_enter(&dp->dp_lock);
 743         if (dp->dp_dirty_total > zfs_dirty_data_sync)
 744                 txg_kick(dp);
 745         rv = (dp->dp_dirty_total > delay_min_bytes);
 746         mutex_exit(&dp->dp_lock);
 747         return (rv);
 748 }
 749 
 750 void
 751 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
 752 {
 753         if (space > 0) {
 754                 mutex_enter(&dp->dp_lock);
 755                 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
 756                 dsl_pool_dirty_delta(dp, space);
 757                 mutex_exit(&dp->dp_lock);
 758         }
 759 }
 760 
 761 void
 762 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
 763 {
 764         ASSERT3S(space, >=, 0);
 765         if (space == 0)
 766                 return;




  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  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]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  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 (c) 2014 Integros [integros.com]
  27  * Copyright 2016 Nexenta Systems, Inc. All rights reserved.
  28  */
  29 
  30 #include <sys/autosnap.h>
  31 #include <sys/dsl_pool.h>
  32 #include <sys/dsl_dataset.h>
  33 #include <sys/dsl_prop.h>
  34 #include <sys/dsl_dir.h>
  35 #include <sys/dsl_synctask.h>
  36 #include <sys/dsl_dataset.h>
  37 #include <sys/dsl_scan.h>
  38 #include <sys/dnode.h>
  39 #include <sys/dmu_tx.h>
  40 #include <sys/dmu_objset.h>
  41 #include <sys/dmu_traverse.h>
  42 #include <sys/arc.h>
  43 #include <sys/zap.h>
  44 #include <sys/zio.h>
  45 #include <sys/zfs_context.h>
  46 #include <sys/fs/zfs.h>
  47 #include <sys/zfs_znode.h>
  48 #include <sys/spa_impl.h>
  49 #include <sys/dsl_deadlist.h>
  50 #include <sys/bptree.h>
  51 #include <sys/zfeature.h>
  52 #include <sys/zil_impl.h>
  53 #include <sys/dsl_userhold.h>
  54 
  55 #include <sys/wbc.h>
  56 #include <sys/time.h>
  57 
  58 /*
  59  * ZFS Write Throttle
  60  * ------------------
  61  *
  62  * ZFS must limit the rate of incoming writes to the rate at which it is able
  63  * to sync data modifications to the backend storage. Throttling by too much
  64  * creates an artificial limit; throttling by too little can only be sustained
  65  * for short periods and would lead to highly lumpy performance. On a per-pool
  66  * basis, ZFS tracks the amount of modified (dirty) data. As operations change
  67  * data, the amount of dirty data increases; as ZFS syncs out data, the amount
  68  * of dirty data decreases. When the amount of dirty data exceeds a
  69  * predetermined threshold further modifications are blocked until the amount
  70  * of dirty data decreases (as data is synced out).
  71  *
  72  * The limit on dirty data is tunable, and should be adjusted according to
  73  * both the IO capacity and available memory of the system. The larger the
  74  * window, the more ZFS is able to aggregate and amortize metadata (and data)
  75  * changes. However, memory is a limited resource, and allowing for more dirty
  76  * data comes at the cost of keeping other useful data in memory (for example
  77  * ZFS data cached by the ARC).


 151  * taskq used by zil_clean(); they determine the number of taskq entries
 152  * that are pre-populated when the taskq is first created (via the
 153  * "zfs_zil_clean_taskq_minalloc" tunable) and the maximum number of
 154  * taskq entries that are cached after an on-demand allocation (via the
 155  * "zfs_zil_clean_taskq_maxalloc").
 156  *
 157  * The idea being, we want to try reasonably hard to ensure there will
 158  * already be a taskq entry pre-allocated by the time that it is needed
 159  * by zil_clean(). This way, we can avoid the possibility of an
 160  * on-demand allocation of a new taskq entry from failing, which would
 161  * result in zil_itxg_clean() being called synchronously from zil_clean()
 162  * (which can adversely affect performance of spa_sync()).
 163  *
 164  * Additionally, the number of threads used by the taskq can be
 165  * configured via the "zfs_zil_clean_taskq_nthr_pct" tunable.
 166  */
 167 int zfs_zil_clean_taskq_nthr_pct = 100;
 168 int zfs_zil_clean_taskq_minalloc = 1024;
 169 int zfs_zil_clean_taskq_maxalloc = 1024 * 1024;
 170 
 171 /*
 172  * Tunable to control max number of tasks available for processing of
 173  * deferred deletes.
 174  */
 175 int zfs_vn_rele_max_tasks = 256;
 176 
 177 int
 178 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
 179 {
 180         uint64_t obj;
 181         int err;
 182 
 183         err = zap_lookup(dp->dp_meta_objset,
 184             dsl_dir_phys(dp->dp_root_dir)->dd_child_dir_zapobj,
 185             name, sizeof (obj), 1, &obj);
 186         if (err)
 187                 return (err);
 188 
 189         return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
 190 }
 191 
 192 static dsl_pool_t *
 193 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
 194 {
 195         dsl_pool_t *dp;
 196         blkptr_t *bp = spa_get_rootblkptr(spa);
 197 
 198         dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
 199         dp->dp_spa = spa;
 200         dp->dp_meta_rootbp = *bp;
 201         rrw_init(&dp->dp_config_rwlock, B_TRUE);
 202 
 203         dp->dp_sync_history[0] = dp->dp_sync_history[1] = 0;
 204 
 205         txg_init(dp, txg);
 206 
 207         txg_list_create(&dp->dp_dirty_datasets, spa,
 208             offsetof(dsl_dataset_t, ds_dirty_link));
 209         txg_list_create(&dp->dp_dirty_zilogs, spa,
 210             offsetof(zilog_t, zl_dirty_link));
 211         txg_list_create(&dp->dp_dirty_dirs, spa,
 212             offsetof(dsl_dir_t, dd_dirty_link));
 213         txg_list_create(&dp->dp_sync_tasks, spa,
 214             offsetof(dsl_sync_task_t, dst_node));
 215 
 216         dp->dp_sync_taskq = taskq_create("dp_sync_taskq",
 217             zfs_sync_taskq_batch_pct, minclsyspri, 1, INT_MAX,
 218             TASKQ_THREADS_CPU_PCT);
 219 
 220         dp->dp_zil_clean_taskq = taskq_create("dp_zil_clean_taskq",
 221             zfs_zil_clean_taskq_nthr_pct, minclsyspri,
 222             zfs_zil_clean_taskq_minalloc,
 223             zfs_zil_clean_taskq_maxalloc,
 224             TASKQ_PREPOPULATE | TASKQ_THREADS_CPU_PCT);
 225 
 226         mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
 227         cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
 228 
 229         dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq",
 230             zfs_vn_rele_max_tasks, minclsyspri,
 231             1, zfs_vn_rele_max_tasks, TASKQ_DYNAMIC);
 232 
 233         return (dp);
 234 }
 235 
 236 int
 237 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
 238 {
 239         int err;
 240         dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
 241 
 242         err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
 243             &dp->dp_meta_objset);
 244         if (err != 0)
 245                 dsl_pool_close(dp);
 246         else
 247                 *dpp = dp;
 248 
 249         return (err);
 250 }
 251 


 287                 }
 288                 dsl_dir_rele(dd, dp);
 289                 if (err)
 290                         goto out;
 291         }
 292 
 293         if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
 294                 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
 295                     &dp->dp_free_dir);
 296                 if (err)
 297                         goto out;
 298 
 299                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 300                     DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
 301                 if (err)
 302                         goto out;
 303                 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
 304                     dp->dp_meta_objset, obj));
 305         }
 306 







 307         /*
 308          * Note: errors ignored, because the leak dir will not exist if we
 309          * have not encountered a leak yet.
 310          */










 311         (void) dsl_pool_open_special_dir(dp, LEAK_DIR_NAME,
 312             &dp->dp_leak_dir);
 313 
 314         if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_ASYNC_DESTROY)) {
 315                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 316                     DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
 317                     &dp->dp_bptree_obj);
 318                 if (err != 0)
 319                         goto out;
 320         }
 321 
 322         if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMPTY_BPOBJ)) {
 323                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 324                     DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
 325                     &dp->dp_empty_bpobj);
 326                 if (err != 0)
 327                         goto out;
 328         }
 329 
 330         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,


 335         if (err)
 336                 goto out;
 337 
 338         err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
 339 
 340 out:
 341         rrw_exit(&dp->dp_config_rwlock, FTAG);
 342         return (err);
 343 }
 344 
 345 void
 346 dsl_pool_close(dsl_pool_t *dp)
 347 {
 348         /*
 349          * Drop our references from dsl_pool_open().
 350          *
 351          * Since we held the origin_snap from "syncing" context (which
 352          * includes pool-opening context), it actually only got a "ref"
 353          * and not a hold, so just drop that here.
 354          */
 355         if (dp->dp_origin_snap)
 356                 dsl_dataset_rele(dp->dp_origin_snap, dp);
 357         if (dp->dp_mos_dir)
 358                 dsl_dir_rele(dp->dp_mos_dir, dp);
 359         if (dp->dp_free_dir)
 360                 dsl_dir_rele(dp->dp_free_dir, dp);
 361         if (dp->dp_leak_dir)
 362                 dsl_dir_rele(dp->dp_leak_dir, dp);
 363         if (dp->dp_root_dir)
 364                 dsl_dir_rele(dp->dp_root_dir, dp);
 365 
 366         bpobj_close(&dp->dp_free_bpobj);

 367 
 368         /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
 369         if (dp->dp_meta_objset)
 370                 dmu_objset_evict(dp->dp_meta_objset);
 371 
 372         txg_list_destroy(&dp->dp_dirty_datasets);
 373         txg_list_destroy(&dp->dp_dirty_zilogs);
 374         txg_list_destroy(&dp->dp_sync_tasks);
 375         txg_list_destroy(&dp->dp_dirty_dirs);
 376 
 377         taskq_destroy(dp->dp_zil_clean_taskq);
 378         taskq_destroy(dp->dp_sync_taskq);
 379 
 380         /*
 381          * We can't set retry to TRUE since we're explicitly specifying
 382          * a spa to flush. This is good enough; any missed buffers for
 383          * this spa won't cause trouble, and they'll eventually fall
 384          * out of the ARC just like any other unused buffer.
 385          */
 386         arc_flush(dp->dp_spa, B_FALSE);

 387         txg_fini(dp);
 388         dsl_scan_fini(dp);
 389         dmu_buf_user_evict_wait();
 390 
 391         rrw_destroy(&dp->dp_config_rwlock);
 392         mutex_destroy(&dp->dp_lock);
 393         taskq_destroy(dp->dp_vnrele_taskq);
 394         if (dp->dp_blkstats)
 395                 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 396         kmem_free(dp, sizeof (dsl_pool_t));
 397 }
 398 





























 399 dsl_pool_t *
 400 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
 401 {
 402         int err;
 403         dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
 404         dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
 405         dsl_dataset_t *ds;
 406         uint64_t obj;
 407 
 408         rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
 409 
 410         /* create and open the MOS (meta-objset) */
 411         dp->dp_meta_objset = dmu_objset_create_impl(spa,
 412             NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
 413 
 414         /* create the pool directory */
 415         err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 416             DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
 417         ASSERT0(err);
 418 


 505                 ASSERT3U(-delta, <=, dp->dp_dirty_total);
 506 
 507         dp->dp_dirty_total += delta;
 508 
 509         /*
 510          * Note: we signal even when increasing dp_dirty_total.
 511          * This ensures forward progress -- each thread wakes the next waiter.
 512          */
 513         if (dp->dp_dirty_total < zfs_dirty_data_max)
 514                 cv_signal(&dp->dp_spaceavail_cv);
 515 }
 516 
 517 void
 518 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
 519 {
 520         zio_t *zio;
 521         dmu_tx_t *tx;
 522         dsl_dir_t *dd;
 523         dsl_dataset_t *ds;
 524         objset_t *mos = dp->dp_meta_objset;
 525         spa_t *spa = dp->dp_spa;
 526         list_t synced_datasets;
 527         dsl_sync_task_t *iter;
 528         boolean_t wbc_skip_txg = B_FALSE;
 529         boolean_t sync_ops = B_FALSE;
 530         boolean_t user_snap = B_FALSE;
 531         zfs_autosnap_t *autosnap = spa_get_autosnap(spa);
 532         boolean_t autosnap_initialized = autosnap->initialized;
 533         char snap[ZFS_MAX_DATASET_NAME_LEN];
 534 
 535         /* check if there are  ny sync ops in the txg */
 536         if (txg_list_head(&dp->dp_sync_tasks, txg) != NULL)
 537                 sync_ops = B_TRUE;
 538 
 539         /* check if there are user snaps in the txg */
 540         for (iter = txg_list_head(&dp->dp_sync_tasks, txg);
 541             iter != NULL;
 542             iter = txg_list_next(&dp->dp_sync_tasks, iter, txg)) {
 543                 if (iter->dst_syncfunc == dsl_dataset_snapshot_sync) {
 544                         user_snap = B_TRUE;
 545                         break;
 546                 }
 547         }
 548 
 549 
 550         list_create(&synced_datasets, sizeof (dsl_dataset_t),
 551             offsetof(dsl_dataset_t, ds_synced_link));
 552 
 553         tx = dmu_tx_create_assigned(dp, txg);
 554 
 555         (void) sprintf(snap, "%s%llu", AUTOSNAP_PREFIX,
 556             (unsigned long long int) txg);
 557 
 558         if (autosnap_initialized && spa->spa_sync_pass == 1) {
 559                 autosnap_zone_t *azone;
 560 
 561                 rrw_enter(&dp->dp_config_rwlock, RW_READER, FTAG);
 562                 mutex_enter(&autosnap->autosnap_lock);
 563 
 564                 /*
 565                  * WBC: the mechanism to ensure all WBC-ed dirty datasets
 566                  * are synchronously auto-snapshotted
 567                  * within (or by) the same TXG sync
 568                  * The "synchronicity" of the rightmost boundary of the WBC
 569                  * window is important to avoid used-space leakages
 570                  * on special vdev.
 571                  * Note that we skip here the WBC-ed datasets that are
 572                  * already fully migrated and don't have data on special
 573                  */
 574 
 575                 for (ds = txg_list_head(&dp->dp_dirty_datasets, txg);
 576                     ds != NULL;
 577                     ds = txg_list_next(&dp->dp_dirty_datasets, ds, txg)) {
 578                         char ds_name[ZFS_MAX_DATASET_NAME_LEN];
 579                         boolean_t wbc_azone;
 580 
 581                         dsl_dataset_name(ds, ds_name);
 582 
 583                         azone = autosnap_find_zone(autosnap, ds_name, B_TRUE);
 584                         if (azone == NULL)
 585                                 continue;
 586 
 587                         if ((azone->flags & AUTOSNAP_CREATOR) == 0)
 588                                 continue;
 589 
 590                         if (azone->created)
 591                                 continue;
 592 
 593                         azone->delayed = B_TRUE;
 594                         azone->dirty = B_TRUE;
 595                         wbc_azone = (azone->flags & AUTOSNAP_WBC) != 0;
 596 
 597                         if (autosnap_confirm_snap(azone, txg)) {
 598                                 if (!wbc_azone && !user_snap && !sync_ops) {
 599                                         autosnap_create_snapshot(azone,
 600                                             snap, dp, txg, tx);
 601                                 }
 602                         } else if (wbc_azone) {
 603                                 wbc_skip_txg = B_TRUE;
 604                         }
 605                 }
 606 
 607                 azone = list_head(&autosnap->autosnap_zones);
 608                 while (azone != NULL) {
 609                         boolean_t wbc_azone =
 610                             ((azone->flags & AUTOSNAP_WBC) != 0);
 611 
 612                         if (user_snap) {
 613                                 azone->delayed = B_TRUE;
 614                         } else if (!azone->dirty && azone->delayed) {
 615                                 if (autosnap_confirm_snap(azone, txg)) {
 616                                         if (!wbc_azone && !user_snap &&
 617                                             !sync_ops) {
 618                                                 autosnap_create_snapshot(azone,
 619                                                     snap, dp, txg, tx);
 620                                         }
 621                                 } else if (wbc_azone) {
 622                                         wbc_skip_txg = B_TRUE;
 623                                 }
 624                         }
 625 
 626                         azone = list_next(&autosnap->autosnap_zones, azone);
 627                 }
 628 
 629                 mutex_exit(&autosnap->autosnap_lock);
 630                 rrw_exit(&dp->dp_config_rwlock, FTAG);
 631         }
 632 
 633 
 634         /*
 635          * Write out all dirty blocks of dirty datasets.
 636          */
 637         zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
 638         while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
 639 
 640                 /*
 641                  * We must not sync any non-MOS datasets twice, because
 642                  * we may have taken a snapshot of them.  However, we
 643                  * may sync newly-created datasets on pass 2.
 644                  */
 645                 ASSERT(!list_link_active(&ds->ds_synced_link));
 646                 list_insert_tail(&synced_datasets, ds);
 647                 dsl_dataset_sync(ds, zio, tx);
 648         }
 649 
 650         VERIFY0(zio_wait(zio));
 651 
 652         if (autosnap_initialized && spa->spa_sync_pass == 1 &&
 653             !user_snap) {
 654                 autosnap_zone_t *azone;
 655 
 656                 rrw_enter(&dp->dp_config_rwlock, RW_READER, FTAG);
 657                 mutex_enter(&autosnap->autosnap_lock);
 658 
 659                 /*
 660                  * At this stage we are walking over all delayed zones
 661                  * to create autosnaps
 662                  */
 663 
 664                 azone = list_head(&autosnap->autosnap_zones);
 665                 while (azone != NULL) {
 666                         boolean_t skip_zone =
 667                             ((azone->flags & AUTOSNAP_CREATOR) == 0);
 668 
 669                         if (azone->delayed && !skip_zone) {
 670                                 boolean_t wbc_azone =
 671                                     ((azone->flags & AUTOSNAP_WBC) != 0);
 672 
 673                                 if ((!wbc_azone || !wbc_skip_txg) &&
 674                                     autosnap_confirm_snap(azone, txg)) {
 675                                         autosnap_create_snapshot(azone,
 676                                             snap, dp, txg, tx);
 677                                 }
 678                         }
 679 
 680                         if (skip_zone)
 681                                 azone->delayed = B_FALSE;
 682 
 683                         azone = list_next(&autosnap->autosnap_zones, azone);
 684                 }
 685 
 686                 mutex_exit(&autosnap->autosnap_lock);
 687                 rrw_exit(&dp->dp_config_rwlock, FTAG);
 688         }
 689 
 690         /*
 691          * We have written all of the accounted dirty data, so our
 692          * dp_space_towrite should now be zero.  However, some seldom-used
 693          * code paths do not adhere to this (e.g. dbuf_undirty(), also
 694          * rounding error in dbuf_write_physdone).
 695          * Shore up the accounting of any dirtied space now.
 696          */
 697         dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
 698 
 699         /*
 700          * Update the long range free counter after
 701          * we're done syncing user data
 702          */
 703         mutex_enter(&dp->dp_lock);
 704         ASSERT(spa_sync_pass(dp->dp_spa) == 1 ||
 705             dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] == 0);
 706         dp->dp_long_freeing_total -=
 707             dp->dp_long_free_dirty_pertxg[txg & TXG_MASK];
 708         dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] = 0;
 709         mutex_exit(&dp->dp_lock);
 710 
 711         /*
 712          * After the data blocks have been written (ensured by the zio_wait()
 713          * above), update the user/group space accounting.  This happens
 714          * in tasks dispatched to dp_sync_taskq, so wait for them before
 715          * continuing.
 716          */
 717         for (ds = list_head(&synced_datasets); ds != NULL;
 718             ds = list_next(&synced_datasets, ds)) {
 719                 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
 720         }
 721         taskq_wait(dp->dp_sync_taskq);
 722 
 723         /*
 724          * Sync the datasets again to push out the changes due to
 725          * userspace updates.  This must be done before we process the
 726          * sync tasks, so that any snapshots will have the correct
 727          * user accounting information (and we won't get confused
 728          * about which blocks are part of the snapshot).
 729          */
 730 
 731         zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
 732         while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
 733                 ASSERT(list_link_active(&ds->ds_synced_link));
 734                 dmu_buf_rele(ds->ds_dbuf, ds);
 735                 dsl_dataset_sync(ds, zio, tx);
 736         }
 737         VERIFY0(zio_wait(zio));
 738 
 739         /*
 740          * Now that the datasets have been completely synced, we can
 741          * clean up our in-memory structures accumulated while syncing:
 742          *
 743          *  - move dead blocks from the pending deadlist to the on-disk deadlist
 744          *  - release hold from dsl_dataset_dirty()
 745          */
 746         while ((ds = list_remove_head(&synced_datasets)) != NULL) {
 747                 dsl_dataset_sync_done(ds, tx);
 748         }
 749         while ((dd = txg_list_remove(&dp->dp_dirty_dirs, txg)) != NULL) {
 750                 dsl_dir_sync(dd, tx);


 762                     dp->dp_mos_compressed_delta,
 763                     dp->dp_mos_uncompressed_delta, tx);
 764                 dp->dp_mos_used_delta = 0;
 765                 dp->dp_mos_compressed_delta = 0;
 766                 dp->dp_mos_uncompressed_delta = 0;
 767         }
 768 
 769         if (!multilist_is_empty(mos->os_dirty_dnodes[txg & TXG_MASK])) {
 770                 dsl_pool_sync_mos(dp, tx);
 771         }
 772 
 773         /*
 774          * If we modify a dataset in the same txg that we want to destroy it,
 775          * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
 776          * dsl_dir_destroy_check() will fail if there are unexpected holds.
 777          * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
 778          * and clearing the hold on it) before we process the sync_tasks.
 779          * The MOS data dirtied by the sync_tasks will be synced on the next
 780          * pass.
 781          */
 782 
 783         if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
 784                 dsl_sync_task_t *dst;
 785                 /*
 786                  * No more sync tasks should have been added while we
 787                  * were syncing.
 788                  */
 789                 ASSERT3U(spa_sync_pass(dp->dp_spa), ==, 1);
 790                 while ((dst = txg_list_remove(&dp->dp_sync_tasks, txg)) != NULL)
 791                         dsl_sync_task_sync(dst, tx);
 792         }
 793 
 794         if (spa_feature_is_active(spa, SPA_FEATURE_WBC)) {
 795                 wbc_trigger_wbcthread(dp->dp_spa,
 796                     ((dp->dp_sync_history[0] + dp->dp_sync_history[1]) / 2));
 797         }
 798 
 799         dmu_tx_commit(tx);
 800 
 801         DTRACE_PROBE2(dsl_pool_sync__done, dsl_pool_t *dp, dp, uint64_t, txg);
 802 }
 803 
 804 void
 805 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
 806 {
 807         zilog_t *zilog;
 808 
 809         while (zilog = txg_list_head(&dp->dp_dirty_zilogs, txg)) {
 810                 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
 811                 /*
 812                  * We don't remove the zilog from the dp_dirty_zilogs
 813                  * list until after we've cleaned it. This ensures that
 814                  * callers of zilog_is_dirty() receive an accurate
 815                  * answer when they are racing with the spa sync thread.
 816                  */
 817                 zil_clean(zilog, txg);
 818                 (void) txg_list_remove_this(&dp->dp_dirty_zilogs, zilog, txg);


 842         /*
 843          * If we're trying to assess whether it's OK to do a free,
 844          * cut the reservation in half to allow forward progress
 845          * (e.g. make it possible to rm(1) files from a full pool).
 846          */
 847         space = spa_get_dspace(dp->dp_spa);
 848         resv = spa_get_slop_space(dp->dp_spa);
 849         if (netfree)
 850                 resv >>= 1;
 851 
 852         return (space - resv);
 853 }
 854 
 855 boolean_t
 856 dsl_pool_need_dirty_delay(dsl_pool_t *dp)
 857 {
 858         uint64_t delay_min_bytes =
 859             zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
 860         boolean_t rv;
 861 

 862         if (dp->dp_dirty_total > zfs_dirty_data_sync)
 863                 txg_kick(dp);
 864         rv = (dp->dp_dirty_total > delay_min_bytes);
 865 
 866         return (rv);
 867 }
 868 
 869 void
 870 dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
 871 {
 872         if (space > 0) {
 873                 mutex_enter(&dp->dp_lock);
 874                 dp->dp_dirty_pertxg[tx->tx_txg & TXG_MASK] += space;
 875                 dsl_pool_dirty_delta(dp, space);
 876                 mutex_exit(&dp->dp_lock);
 877         }
 878 }
 879 
 880 void
 881 dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg)
 882 {
 883         ASSERT3S(space, >=, 0);
 884         if (space == 0)
 885                 return;