1 /*
   2  * CDDL HEADER START
   3  *
   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.
   7  *
   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.
  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) 2012 by Delphix. All rights reserved.
  24  * Copyright (c) 2012, Joyent, Inc. All rights reserved.
  25  */
  26 
  27 /* Portions Copyright 2010 Robert Milkowski */
  28 
  29 #include <sys/cred.h>
  30 #include <sys/zfs_context.h>
  31 #include <sys/dmu_objset.h>
  32 #include <sys/dsl_dir.h>
  33 #include <sys/dsl_dataset.h>
  34 #include <sys/dsl_prop.h>
  35 #include <sys/dsl_pool.h>
  36 #include <sys/dsl_synctask.h>
  37 #include <sys/dsl_deleg.h>
  38 #include <sys/dnode.h>
  39 #include <sys/dbuf.h>
  40 #include <sys/zvol.h>
  41 #include <sys/dmu_tx.h>
  42 #include <sys/zap.h>
  43 #include <sys/zil.h>
  44 #include <sys/dmu_impl.h>
  45 #include <sys/zfs_ioctl.h>
  46 #include <sys/sa.h>
  47 #include <sys/zfs_onexit.h>
  48 
  49 /*
  50  * Needed to close a window in dnode_move() that allows the objset to be freed
  51  * before it can be safely accessed.
  52  */
  53 krwlock_t os_lock;
  54 
  55 void
  56 dmu_objset_init(void)
  57 {
  58         rw_init(&os_lock, NULL, RW_DEFAULT, NULL);
  59 }
  60 
  61 void
  62 dmu_objset_fini(void)
  63 {
  64         rw_destroy(&os_lock);
  65 }
  66 
  67 spa_t *
  68 dmu_objset_spa(objset_t *os)
  69 {
  70         return (os->os_spa);
  71 }
  72 
  73 zilog_t *
  74 dmu_objset_zil(objset_t *os)
  75 {
  76         return (os->os_zil);
  77 }
  78 
  79 dsl_pool_t *
  80 dmu_objset_pool(objset_t *os)
  81 {
  82         dsl_dataset_t *ds;
  83 
  84         if ((ds = os->os_dsl_dataset) != NULL && ds->ds_dir)
  85                 return (ds->ds_dir->dd_pool);
  86         else
  87                 return (spa_get_dsl(os->os_spa));
  88 }
  89 
  90 dsl_dataset_t *
  91 dmu_objset_ds(objset_t *os)
  92 {
  93         return (os->os_dsl_dataset);
  94 }
  95 
  96 dmu_objset_type_t
  97 dmu_objset_type(objset_t *os)
  98 {
  99         return (os->os_phys->os_type);
 100 }
 101 
 102 void
 103 dmu_objset_name(objset_t *os, char *buf)
 104 {
 105         dsl_dataset_name(os->os_dsl_dataset, buf);
 106 }
 107 
 108 uint64_t
 109 dmu_objset_id(objset_t *os)
 110 {
 111         dsl_dataset_t *ds = os->os_dsl_dataset;
 112 
 113         return (ds ? ds->ds_object : 0);
 114 }
 115 
 116 uint64_t
 117 dmu_objset_syncprop(objset_t *os)
 118 {
 119         return (os->os_sync);
 120 }
 121 
 122 uint64_t
 123 dmu_objset_logbias(objset_t *os)
 124 {
 125         return (os->os_logbias);
 126 }
 127 
 128 static void
 129 checksum_changed_cb(void *arg, uint64_t newval)
 130 {
 131         objset_t *os = arg;
 132 
 133         /*
 134          * Inheritance should have been done by now.
 135          */
 136         ASSERT(newval != ZIO_CHECKSUM_INHERIT);
 137 
 138         os->os_checksum = zio_checksum_select(newval, ZIO_CHECKSUM_ON_VALUE);
 139 }
 140 
 141 static void
 142 compression_changed_cb(void *arg, uint64_t newval)
 143 {
 144         objset_t *os = arg;
 145 
 146         /*
 147          * Inheritance and range checking should have been done by now.
 148          */
 149         ASSERT(newval != ZIO_COMPRESS_INHERIT);
 150 
 151         os->os_compress = zio_compress_select(newval, ZIO_COMPRESS_ON_VALUE);
 152 }
 153 
 154 static void
 155 copies_changed_cb(void *arg, uint64_t newval)
 156 {
 157         objset_t *os = arg;
 158 
 159         /*
 160          * Inheritance and range checking should have been done by now.
 161          */
 162         ASSERT(newval > 0);
 163         ASSERT(newval <= spa_max_replication(os->os_spa));
 164 
 165         os->os_copies = newval;
 166 }
 167 
 168 static void
 169 dedup_changed_cb(void *arg, uint64_t newval)
 170 {
 171         objset_t *os = arg;
 172         spa_t *spa = os->os_spa;
 173         enum zio_checksum checksum;
 174 
 175         /*
 176          * Inheritance should have been done by now.
 177          */
 178         ASSERT(newval != ZIO_CHECKSUM_INHERIT);
 179 
 180         checksum = zio_checksum_dedup_select(spa, newval, ZIO_CHECKSUM_OFF);
 181 
 182         os->os_dedup_checksum = checksum & ZIO_CHECKSUM_MASK;
 183         os->os_dedup_verify = !!(checksum & ZIO_CHECKSUM_VERIFY);
 184 }
 185 
 186 static void
 187 primary_cache_changed_cb(void *arg, uint64_t newval)
 188 {
 189         objset_t *os = arg;
 190 
 191         /*
 192          * Inheritance and range checking should have been done by now.
 193          */
 194         ASSERT(newval == ZFS_CACHE_ALL || newval == ZFS_CACHE_NONE ||
 195             newval == ZFS_CACHE_METADATA);
 196 
 197         os->os_primary_cache = newval;
 198 }
 199 
 200 static void
 201 secondary_cache_changed_cb(void *arg, uint64_t newval)
 202 {
 203         objset_t *os = arg;
 204 
 205         /*
 206          * Inheritance and range checking should have been done by now.
 207          */
 208         ASSERT(newval == ZFS_CACHE_ALL || newval == ZFS_CACHE_NONE ||
 209             newval == ZFS_CACHE_METADATA);
 210 
 211         os->os_secondary_cache = newval;
 212 }
 213 
 214 static void
 215 sync_changed_cb(void *arg, uint64_t newval)
 216 {
 217         objset_t *os = arg;
 218 
 219         /*
 220          * Inheritance and range checking should have been done by now.
 221          */
 222         ASSERT(newval == ZFS_SYNC_STANDARD || newval == ZFS_SYNC_ALWAYS ||
 223             newval == ZFS_SYNC_DISABLED);
 224 
 225         os->os_sync = newval;
 226         if (os->os_zil)
 227                 zil_set_sync(os->os_zil, newval);
 228 }
 229 
 230 static void
 231 logbias_changed_cb(void *arg, uint64_t newval)
 232 {
 233         objset_t *os = arg;
 234 
 235         ASSERT(newval == ZFS_LOGBIAS_LATENCY ||
 236             newval == ZFS_LOGBIAS_THROUGHPUT);
 237         os->os_logbias = newval;
 238         if (os->os_zil)
 239                 zil_set_logbias(os->os_zil, newval);
 240 }
 241 
 242 void
 243 dmu_objset_byteswap(void *buf, size_t size)
 244 {
 245         objset_phys_t *osp = buf;
 246 
 247         ASSERT(size == OBJSET_OLD_PHYS_SIZE || size == sizeof (objset_phys_t));
 248         dnode_byteswap(&osp->os_meta_dnode);
 249         byteswap_uint64_array(&osp->os_zil_header, sizeof (zil_header_t));
 250         osp->os_type = BSWAP_64(osp->os_type);
 251         osp->os_flags = BSWAP_64(osp->os_flags);
 252         if (size == sizeof (objset_phys_t)) {
 253                 dnode_byteswap(&osp->os_userused_dnode);
 254                 dnode_byteswap(&osp->os_groupused_dnode);
 255         }
 256 }
 257 
 258 int
 259 dmu_objset_open_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
 260     objset_t **osp)
 261 {
 262         objset_t *os;
 263         int i, err;
 264 
 265         ASSERT(ds == NULL || MUTEX_HELD(&ds->ds_opening_lock));
 266 
 267         os = kmem_zalloc(sizeof (objset_t), KM_SLEEP);
 268         os->os_dsl_dataset = ds;
 269         os->os_spa = spa;
 270         os->os_rootbp = bp;
 271         if (!BP_IS_HOLE(os->os_rootbp)) {
 272                 uint32_t aflags = ARC_WAIT;
 273                 zbookmark_t zb;
 274                 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
 275                     ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
 276 
 277                 if (DMU_OS_IS_L2CACHEABLE(os))
 278                         aflags |= ARC_L2CACHE;
 279 
 280                 dprintf_bp(os->os_rootbp, "reading %s", "");
 281                 /*
 282                  * XXX when bprewrite scrub can change the bp,
 283                  * and this is called from dmu_objset_open_ds_os, the bp
 284                  * could change, and we'll need a lock.
 285                  */
 286                 err = dsl_read_nolock(NULL, spa, os->os_rootbp,
 287                     arc_getbuf_func, &os->os_phys_buf,
 288                     ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL, &aflags, &zb);
 289                 if (err) {
 290                         kmem_free(os, sizeof (objset_t));
 291                         /* convert checksum errors into IO errors */
 292                         if (err == ECKSUM)
 293                                 err = EIO;
 294                         return (err);
 295                 }
 296 
 297                 /* Increase the blocksize if we are permitted. */
 298                 if (spa_version(spa) >= SPA_VERSION_USERSPACE &&
 299                     arc_buf_size(os->os_phys_buf) < sizeof (objset_phys_t)) {
 300                         arc_buf_t *buf = arc_buf_alloc(spa,
 301                             sizeof (objset_phys_t), &os->os_phys_buf,
 302                             ARC_BUFC_METADATA);
 303                         bzero(buf->b_data, sizeof (objset_phys_t));
 304                         bcopy(os->os_phys_buf->b_data, buf->b_data,
 305                             arc_buf_size(os->os_phys_buf));
 306                         (void) arc_buf_remove_ref(os->os_phys_buf,
 307                             &os->os_phys_buf);
 308                         os->os_phys_buf = buf;
 309                 }
 310 
 311                 os->os_phys = os->os_phys_buf->b_data;
 312                 os->os_flags = os->os_phys->os_flags;
 313         } else {
 314                 int size = spa_version(spa) >= SPA_VERSION_USERSPACE ?
 315                     sizeof (objset_phys_t) : OBJSET_OLD_PHYS_SIZE;
 316                 os->os_phys_buf = arc_buf_alloc(spa, size,
 317                     &os->os_phys_buf, ARC_BUFC_METADATA);
 318                 os->os_phys = os->os_phys_buf->b_data;
 319                 bzero(os->os_phys, size);
 320         }
 321 
 322         /*
 323          * Note: the changed_cb will be called once before the register
 324          * func returns, thus changing the checksum/compression from the
 325          * default (fletcher2/off).  Snapshots don't need to know about
 326          * checksum/compression/copies.
 327          */
 328         if (ds) {
 329                 err = dsl_prop_register(ds, "primarycache",
 330                     primary_cache_changed_cb, os);
 331                 if (err == 0)
 332                         err = dsl_prop_register(ds, "secondarycache",
 333                             secondary_cache_changed_cb, os);
 334                 if (!dsl_dataset_is_snapshot(ds)) {
 335                         if (err == 0)
 336                                 err = dsl_prop_register(ds, "checksum",
 337                                     checksum_changed_cb, os);
 338                         if (err == 0)
 339                                 err = dsl_prop_register(ds, "compression",
 340                                     compression_changed_cb, os);
 341                         if (err == 0)
 342                                 err = dsl_prop_register(ds, "copies",
 343                                     copies_changed_cb, os);
 344                         if (err == 0)
 345                                 err = dsl_prop_register(ds, "dedup",
 346                                     dedup_changed_cb, os);
 347                         if (err == 0)
 348                                 err = dsl_prop_register(ds, "logbias",
 349                                     logbias_changed_cb, os);
 350                         if (err == 0)
 351                                 err = dsl_prop_register(ds, "sync",
 352                                     sync_changed_cb, os);
 353                 }
 354                 if (err) {
 355                         VERIFY(arc_buf_remove_ref(os->os_phys_buf,
 356                             &os->os_phys_buf) == 1);
 357                         kmem_free(os, sizeof (objset_t));
 358                         return (err);
 359                 }
 360         } else if (ds == NULL) {
 361                 /* It's the meta-objset. */
 362                 os->os_checksum = ZIO_CHECKSUM_FLETCHER_4;
 363                 os->os_compress = ZIO_COMPRESS_LZJB;
 364                 os->os_copies = spa_max_replication(spa);
 365                 os->os_dedup_checksum = ZIO_CHECKSUM_OFF;
 366                 os->os_dedup_verify = 0;
 367                 os->os_logbias = 0;
 368                 os->os_sync = 0;
 369                 os->os_primary_cache = ZFS_CACHE_ALL;
 370                 os->os_secondary_cache = ZFS_CACHE_ALL;
 371         }
 372 
 373         if (ds == NULL || !dsl_dataset_is_snapshot(ds))
 374                 os->os_zil_header = os->os_phys->os_zil_header;
 375         os->os_zil = zil_alloc(os, &os->os_zil_header);
 376 
 377         for (i = 0; i < TXG_SIZE; i++) {
 378                 list_create(&os->os_dirty_dnodes[i], sizeof (dnode_t),
 379                     offsetof(dnode_t, dn_dirty_link[i]));
 380                 list_create(&os->os_free_dnodes[i], sizeof (dnode_t),
 381                     offsetof(dnode_t, dn_dirty_link[i]));
 382         }
 383         list_create(&os->os_dnodes, sizeof (dnode_t),
 384             offsetof(dnode_t, dn_link));
 385         list_create(&os->os_downgraded_dbufs, sizeof (dmu_buf_impl_t),
 386             offsetof(dmu_buf_impl_t, db_link));
 387 
 388         mutex_init(&os->os_lock, NULL, MUTEX_DEFAULT, NULL);
 389         mutex_init(&os->os_obj_lock, NULL, MUTEX_DEFAULT, NULL);
 390         mutex_init(&os->os_user_ptr_lock, NULL, MUTEX_DEFAULT, NULL);
 391 
 392         DMU_META_DNODE(os) = dnode_special_open(os,
 393             &os->os_phys->os_meta_dnode, DMU_META_DNODE_OBJECT,
 394             &os->os_meta_dnode);
 395         if (arc_buf_size(os->os_phys_buf) >= sizeof (objset_phys_t)) {
 396                 DMU_USERUSED_DNODE(os) = dnode_special_open(os,
 397                     &os->os_phys->os_userused_dnode, DMU_USERUSED_OBJECT,
 398                     &os->os_userused_dnode);
 399                 DMU_GROUPUSED_DNODE(os) = dnode_special_open(os,
 400                     &os->os_phys->os_groupused_dnode, DMU_GROUPUSED_OBJECT,
 401                     &os->os_groupused_dnode);
 402         }
 403 
 404         /*
 405          * We should be the only thread trying to do this because we
 406          * have ds_opening_lock
 407          */
 408         if (ds) {
 409                 mutex_enter(&ds->ds_lock);
 410                 ASSERT(ds->ds_objset == NULL);
 411                 ds->ds_objset = os;
 412                 mutex_exit(&ds->ds_lock);
 413         }
 414 
 415         *osp = os;
 416         return (0);
 417 }
 418 
 419 int
 420 dmu_objset_from_ds(dsl_dataset_t *ds, objset_t **osp)
 421 {
 422         int err = 0;
 423 
 424         mutex_enter(&ds->ds_opening_lock);
 425         *osp = ds->ds_objset;
 426         if (*osp == NULL) {
 427                 err = dmu_objset_open_impl(dsl_dataset_get_spa(ds),
 428                     ds, dsl_dataset_get_blkptr(ds), osp);
 429         }
 430         mutex_exit(&ds->ds_opening_lock);
 431         return (err);
 432 }
 433 
 434 /* called from zpl */
 435 int
 436 dmu_objset_hold(const char *name, void *tag, objset_t **osp)
 437 {
 438         dsl_dataset_t *ds;
 439         int err;
 440 
 441         err = dsl_dataset_hold(name, tag, &ds);
 442         if (err)
 443                 return (err);
 444 
 445         err = dmu_objset_from_ds(ds, osp);
 446         if (err)
 447                 dsl_dataset_rele(ds, tag);
 448 
 449         return (err);
 450 }
 451 
 452 /* called from zpl */
 453 int
 454 dmu_objset_own(const char *name, dmu_objset_type_t type,
 455     boolean_t readonly, void *tag, objset_t **osp)
 456 {
 457         dsl_dataset_t *ds;
 458         int err;
 459 
 460         err = dsl_dataset_own(name, B_FALSE, tag, &ds);
 461         if (err)
 462                 return (err);
 463 
 464         err = dmu_objset_from_ds(ds, osp);
 465         if (err) {
 466                 dsl_dataset_disown(ds, tag);
 467         } else if (type != DMU_OST_ANY && type != (*osp)->os_phys->os_type) {
 468                 dmu_objset_disown(*osp, tag);
 469                 return (EINVAL);
 470         } else if (!readonly && dsl_dataset_is_snapshot(ds)) {
 471                 dmu_objset_disown(*osp, tag);
 472                 return (EROFS);
 473         }
 474         return (err);
 475 }
 476 
 477 void
 478 dmu_objset_rele(objset_t *os, void *tag)
 479 {
 480         dsl_dataset_rele(os->os_dsl_dataset, tag);
 481 }
 482 
 483 void
 484 dmu_objset_disown(objset_t *os, void *tag)
 485 {
 486         dsl_dataset_disown(os->os_dsl_dataset, tag);
 487 }
 488 
 489 int
 490 dmu_objset_evict_dbufs(objset_t *os)
 491 {
 492         dnode_t *dn;
 493 
 494         mutex_enter(&os->os_lock);
 495 
 496         /* process the mdn last, since the other dnodes have holds on it */
 497         list_remove(&os->os_dnodes, DMU_META_DNODE(os));
 498         list_insert_tail(&os->os_dnodes, DMU_META_DNODE(os));
 499 
 500         /*
 501          * Find the first dnode with holds.  We have to do this dance
 502          * because dnode_add_ref() only works if you already have a
 503          * hold.  If there are no holds then it has no dbufs so OK to
 504          * skip.
 505          */
 506         for (dn = list_head(&os->os_dnodes);
 507             dn && !dnode_add_ref(dn, FTAG);
 508             dn = list_next(&os->os_dnodes, dn))
 509                 continue;
 510 
 511         while (dn) {
 512                 dnode_t *next_dn = dn;
 513 
 514                 do {
 515                         next_dn = list_next(&os->os_dnodes, next_dn);
 516                 } while (next_dn && !dnode_add_ref(next_dn, FTAG));
 517 
 518                 mutex_exit(&os->os_lock);
 519                 dnode_evict_dbufs(dn);
 520                 dnode_rele(dn, FTAG);
 521                 mutex_enter(&os->os_lock);
 522                 dn = next_dn;
 523         }
 524         dn = list_head(&os->os_dnodes);
 525         mutex_exit(&os->os_lock);
 526         return (dn != DMU_META_DNODE(os));
 527 }
 528 
 529 void
 530 dmu_objset_evict(objset_t *os)
 531 {
 532         dsl_dataset_t *ds = os->os_dsl_dataset;
 533 
 534         for (int t = 0; t < TXG_SIZE; t++)
 535                 ASSERT(!dmu_objset_is_dirty(os, t));
 536 
 537         if (ds) {
 538                 if (!dsl_dataset_is_snapshot(ds)) {
 539                         VERIFY(0 == dsl_prop_unregister(ds, "checksum",
 540                             checksum_changed_cb, os));
 541                         VERIFY(0 == dsl_prop_unregister(ds, "compression",
 542                             compression_changed_cb, os));
 543                         VERIFY(0 == dsl_prop_unregister(ds, "copies",
 544                             copies_changed_cb, os));
 545                         VERIFY(0 == dsl_prop_unregister(ds, "dedup",
 546                             dedup_changed_cb, os));
 547                         VERIFY(0 == dsl_prop_unregister(ds, "logbias",
 548                             logbias_changed_cb, os));
 549                         VERIFY(0 == dsl_prop_unregister(ds, "sync",
 550                             sync_changed_cb, os));
 551                 }
 552                 VERIFY(0 == dsl_prop_unregister(ds, "primarycache",
 553                     primary_cache_changed_cb, os));
 554                 VERIFY(0 == dsl_prop_unregister(ds, "secondarycache",
 555                     secondary_cache_changed_cb, os));
 556         }
 557 
 558         if (os->os_sa)
 559                 sa_tear_down(os);
 560 
 561         /*
 562          * We should need only a single pass over the dnode list, since
 563          * nothing can be added to the list at this point.
 564          */
 565         (void) dmu_objset_evict_dbufs(os);
 566 
 567         dnode_special_close(&os->os_meta_dnode);
 568         if (DMU_USERUSED_DNODE(os)) {
 569                 dnode_special_close(&os->os_userused_dnode);
 570                 dnode_special_close(&os->os_groupused_dnode);
 571         }
 572         zil_free(os->os_zil);
 573 
 574         ASSERT3P(list_head(&os->os_dnodes), ==, NULL);
 575 
 576         VERIFY(arc_buf_remove_ref(os->os_phys_buf, &os->os_phys_buf) == 1);
 577 
 578         /*
 579          * This is a barrier to prevent the objset from going away in
 580          * dnode_move() until we can safely ensure that the objset is still in
 581          * use. We consider the objset valid before the barrier and invalid
 582          * after the barrier.
 583          */
 584         rw_enter(&os_lock, RW_READER);
 585         rw_exit(&os_lock);
 586 
 587         mutex_destroy(&os->os_lock);
 588         mutex_destroy(&os->os_obj_lock);
 589         mutex_destroy(&os->os_user_ptr_lock);
 590         kmem_free(os, sizeof (objset_t));
 591 }
 592 
 593 timestruc_t
 594 dmu_objset_snap_cmtime(objset_t *os)
 595 {
 596         return (dsl_dir_snap_cmtime(os->os_dsl_dataset->ds_dir));
 597 }
 598 
 599 /* called from dsl for meta-objset */
 600 objset_t *
 601 dmu_objset_create_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
 602     dmu_objset_type_t type, dmu_tx_t *tx)
 603 {
 604         objset_t *os;
 605         dnode_t *mdn;
 606 
 607         ASSERT(dmu_tx_is_syncing(tx));
 608         if (ds != NULL)
 609                 VERIFY(0 == dmu_objset_from_ds(ds, &os));
 610         else
 611                 VERIFY(0 == dmu_objset_open_impl(spa, NULL, bp, &os));
 612 
 613         mdn = DMU_META_DNODE(os);
 614 
 615         dnode_allocate(mdn, DMU_OT_DNODE, 1 << DNODE_BLOCK_SHIFT,
 616             DN_MAX_INDBLKSHIFT, DMU_OT_NONE, 0, tx);
 617 
 618         /*
 619          * We don't want to have to increase the meta-dnode's nlevels
 620          * later, because then we could do it in quescing context while
 621          * we are also accessing it in open context.
 622          *
 623          * This precaution is not necessary for the MOS (ds == NULL),
 624          * because the MOS is only updated in syncing context.
 625          * This is most fortunate: the MOS is the only objset that
 626          * needs to be synced multiple times as spa_sync() iterates
 627          * to convergence, so minimizing its dn_nlevels matters.
 628          */
 629         if (ds != NULL) {
 630                 int levels = 1;
 631 
 632                 /*
 633                  * Determine the number of levels necessary for the meta-dnode
 634                  * to contain DN_MAX_OBJECT dnodes.
 635                  */
 636                 while ((uint64_t)mdn->dn_nblkptr << (mdn->dn_datablkshift +
 637                     (levels - 1) * (mdn->dn_indblkshift - SPA_BLKPTRSHIFT)) <
 638                     DN_MAX_OBJECT * sizeof (dnode_phys_t))
 639                         levels++;
 640 
 641                 mdn->dn_next_nlevels[tx->tx_txg & TXG_MASK] =
 642                     mdn->dn_nlevels = levels;
 643         }
 644 
 645         ASSERT(type != DMU_OST_NONE);
 646         ASSERT(type != DMU_OST_ANY);
 647         ASSERT(type < DMU_OST_NUMTYPES);
 648         os->os_phys->os_type = type;
 649         if (dmu_objset_userused_enabled(os)) {
 650                 os->os_phys->os_flags |= OBJSET_FLAG_USERACCOUNTING_COMPLETE;
 651                 os->os_flags = os->os_phys->os_flags;
 652         }
 653 
 654         dsl_dataset_dirty(ds, tx);
 655 
 656         return (os);
 657 }
 658 
 659 struct oscarg {
 660         void (*userfunc)(objset_t *os, void *arg, cred_t *cr, dmu_tx_t *tx);
 661         void *userarg;
 662         dsl_dataset_t *clone_origin;
 663         const char *lastname;
 664         dmu_objset_type_t type;
 665         uint64_t flags;
 666         cred_t *cr;
 667 };
 668 
 669 /*ARGSUSED*/
 670 static int
 671 dmu_objset_create_check(void *arg1, void *arg2, dmu_tx_t *tx)
 672 {
 673         dsl_dir_t *dd = arg1;
 674         struct oscarg *oa = arg2;
 675         objset_t *mos = dd->dd_pool->dp_meta_objset;
 676         int err;
 677         uint64_t ddobj;
 678 
 679         err = zap_lookup(mos, dd->dd_phys->dd_child_dir_zapobj,
 680             oa->lastname, sizeof (uint64_t), 1, &ddobj);
 681         if (err != ENOENT)
 682                 return (err ? err : EEXIST);
 683 
 684         if (oa->clone_origin != NULL) {
 685                 /* You can't clone across pools. */
 686                 if (oa->clone_origin->ds_dir->dd_pool != dd->dd_pool)
 687                         return (EXDEV);
 688 
 689                 /* You can only clone snapshots, not the head datasets. */
 690                 if (!dsl_dataset_is_snapshot(oa->clone_origin))
 691                         return (EINVAL);
 692         }
 693 
 694         return (dsl_dir_fscount_check(dd, 1, NULL, oa->cr));
 695 }
 696 
 697 static void
 698 dmu_objset_create_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 699 {
 700         dsl_dir_t *dd = arg1;
 701         spa_t *spa = dd->dd_pool->dp_spa;
 702         struct oscarg *oa = arg2;
 703         uint64_t obj;
 704         dsl_dataset_t *ds;
 705         blkptr_t *bp;
 706 
 707         ASSERT(dmu_tx_is_syncing(tx));
 708 
 709         dsl_dir_fscount_adjust(dd, tx, 1, B_TRUE);
 710 
 711         obj = dsl_dataset_create_sync(dd, oa->lastname,
 712             oa->clone_origin, oa->flags, oa->cr, tx);
 713 
 714         VERIFY3U(0, ==, dsl_dataset_hold_obj(dd->dd_pool, obj, FTAG, &ds));
 715         bp = dsl_dataset_get_blkptr(ds);
 716         if (BP_IS_HOLE(bp)) {
 717                 objset_t *os =
 718                     dmu_objset_create_impl(spa, ds, bp, oa->type, tx);
 719 
 720                 if (oa->userfunc)
 721                         oa->userfunc(os, oa->userarg, oa->cr, tx);
 722         }
 723 
 724         if (oa->clone_origin == NULL) {
 725                 spa_history_log_internal_ds(ds, "create", tx, "");
 726         } else {
 727                 char namebuf[MAXNAMELEN];
 728                 dsl_dataset_name(oa->clone_origin, namebuf);
 729                 spa_history_log_internal_ds(ds, "clone", tx,
 730                     "origin=%s (%llu)", namebuf, oa->clone_origin->ds_object);
 731         }
 732         dsl_dataset_rele(ds, FTAG);
 733 }
 734 
 735 int
 736 dmu_objset_create(const char *name, dmu_objset_type_t type, uint64_t flags,
 737     void (*func)(objset_t *os, void *arg, cred_t *cr, dmu_tx_t *tx), void *arg)
 738 {
 739         dsl_dir_t *pdd;
 740         const char *tail;
 741         int err = 0;
 742         struct oscarg oa = { 0 };
 743 
 744         ASSERT(strchr(name, '@') == NULL);
 745         err = dsl_dir_open(name, FTAG, &pdd, &tail);
 746         if (err)
 747                 return (err);
 748         if (tail == NULL) {
 749                 dsl_dir_close(pdd, FTAG);
 750                 return (EEXIST);
 751         }
 752 
 753         oa.userfunc = func;
 754         oa.userarg = arg;
 755         oa.lastname = tail;
 756         oa.type = type;
 757         oa.flags = flags;
 758         oa.cr = CRED();
 759 
 760         err = dsl_sync_task_do(pdd->dd_pool, dmu_objset_create_check,
 761             dmu_objset_create_sync, pdd, &oa, 5);
 762         dsl_dir_close(pdd, FTAG);
 763         return (err);
 764 }
 765 
 766 int
 767 dmu_objset_clone(const char *name, dsl_dataset_t *clone_origin, uint64_t flags)
 768 {
 769         dsl_dir_t *pdd;
 770         const char *tail;
 771         int err = 0;
 772         struct oscarg oa = { 0 };
 773 
 774         ASSERT(strchr(name, '@') == NULL);
 775         err = dsl_dir_open(name, FTAG, &pdd, &tail);
 776         if (err)
 777                 return (err);
 778         if (tail == NULL) {
 779                 dsl_dir_close(pdd, FTAG);
 780                 return (EEXIST);
 781         }
 782 
 783         oa.lastname = tail;
 784         oa.clone_origin = clone_origin;
 785         oa.flags = flags;
 786         oa.cr = CRED();
 787 
 788         err = dsl_sync_task_do(pdd->dd_pool, dmu_objset_create_check,
 789             dmu_objset_create_sync, pdd, &oa, 5);
 790         dsl_dir_close(pdd, FTAG);
 791         return (err);
 792 }
 793 
 794 int
 795 dmu_objset_destroy(const char *name, boolean_t defer)
 796 {
 797         dsl_dataset_t *ds;
 798         int error;
 799 
 800         error = dsl_dataset_own(name, B_TRUE, FTAG, &ds);
 801         if (error == 0) {
 802                 error = dsl_dataset_destroy(ds, FTAG, defer);
 803                 /* dsl_dataset_destroy() closes the ds. */
 804         }
 805 
 806         return (error);
 807 }
 808 
 809 typedef struct snapallarg {
 810         dsl_sync_task_group_t *saa_dstg;
 811         boolean_t saa_needsuspend;
 812         nvlist_t *saa_props;
 813         cred_t *saa_cr;
 814 
 815         /* the following are used only if 'temporary' is set: */
 816         boolean_t saa_temporary;
 817         const char *saa_htag;
 818         struct dsl_ds_holdarg *saa_ha;
 819         dsl_dataset_t *saa_newds;
 820 } snapallarg_t;
 821 
 822 typedef struct snaponearg {
 823         const char *soa_longname; /* long snap name */
 824         const char *soa_snapname; /* short snap name */
 825         uint64_t soa_tot_cnt;
 826         snapallarg_t *soa_saa;
 827 } snaponearg_t;
 828 
 829 static int
 830 snapshot_check(void *arg1, void *arg2, dmu_tx_t *tx)
 831 {
 832         objset_t *os = arg1;
 833         snaponearg_t *soa = arg2;
 834         snapallarg_t *saa = soa->soa_saa;
 835         int error;
 836 
 837         /* The props have already been checked by zfs_check_userprops(). */
 838 
 839         error = dsl_dataset_snapshot_check(os->os_dsl_dataset,
 840             soa->soa_snapname, soa->soa_tot_cnt, tx, saa->saa_cr);
 841         if (error)
 842                 return (error);
 843 
 844         if (saa->saa_temporary) {
 845                 /*
 846                  * Ideally we would just call
 847                  * dsl_dataset_user_hold_check() and
 848                  * dsl_dataset_destroy_check() here.  However the
 849                  * dataset we want to hold and destroy is the snapshot
 850                  * that we just confirmed we can create, but it won't
 851                  * exist until after these checks are run.  Do any
 852                  * checks we can here and if more checks are added to
 853                  * those routines in the future, similar checks may be
 854                  * necessary here.
 855                  */
 856                 if (spa_version(os->os_spa) < SPA_VERSION_USERREFS)
 857                         return (ENOTSUP);
 858                 /*
 859                  * Not checking number of tags because the tag will be
 860                  * unique, as it will be the only tag.
 861                  */
 862                 if (strlen(saa->saa_htag) + MAX_TAG_PREFIX_LEN >= MAXNAMELEN)
 863                         return (E2BIG);
 864 
 865                 saa->saa_ha = kmem_alloc(sizeof (struct dsl_ds_holdarg),
 866                     KM_SLEEP);
 867                 saa->saa_ha->temphold = B_TRUE;
 868                 saa->saa_ha->htag = saa->saa_htag;
 869         }
 870         return (error);
 871 }
 872 
 873 static void
 874 snapshot_sync(void *arg1, void *arg2, dmu_tx_t *tx)
 875 {
 876         objset_t *os = arg1;
 877         dsl_dataset_t *ds = os->os_dsl_dataset;
 878         snaponearg_t *soa = arg2;
 879         snapallarg_t *saa = soa->soa_saa;
 880 
 881         dsl_dataset_snapshot_sync(ds, soa->soa_snapname, tx);
 882 
 883         if (saa->saa_props != NULL) {
 884                 dsl_props_arg_t pa;
 885                 pa.pa_props = saa->saa_props;
 886                 pa.pa_source = ZPROP_SRC_LOCAL;
 887                 dsl_props_set_sync(ds->ds_prev, &pa, tx);
 888         }
 889 
 890         if (saa->saa_temporary) {
 891                 struct dsl_ds_destroyarg da;
 892 
 893                 dsl_dataset_user_hold_sync(ds->ds_prev, saa->saa_ha, tx);
 894                 kmem_free(saa->saa_ha, sizeof (struct dsl_ds_holdarg));
 895                 saa->saa_ha = NULL;
 896                 saa->saa_newds = ds->ds_prev;
 897 
 898                 da.ds = ds->ds_prev;
 899                 da.defer = B_TRUE;
 900                 dsl_dataset_destroy_sync(&da, FTAG, tx);
 901         }
 902 }
 903 
 904 static int
 905 snapshot_one_impl(const char *snapname, void *arg, uint64_t cnt)
 906 {
 907         char fsname[MAXPATHLEN];
 908         snapallarg_t *saa = arg;
 909         snaponearg_t *soa;
 910         objset_t *os;
 911         int err;
 912 
 913         (void) strlcpy(fsname, snapname, sizeof (fsname));
 914         strchr(fsname, '@')[0] = '\0';
 915 
 916         err = dmu_objset_hold(fsname, saa, &os);
 917         if (err != 0)
 918                 return (err);
 919 
 920         /*
 921          * If the objset is in an inconsistent state (eg, in the process
 922          * of being destroyed), don't snapshot it.
 923          */
 924         if (os->os_dsl_dataset->ds_phys->ds_flags & DS_FLAG_INCONSISTENT) {
 925                 dmu_objset_rele(os, saa);
 926                 return (EBUSY);
 927         }
 928 
 929         if (saa->saa_needsuspend) {
 930                 err = zil_suspend(dmu_objset_zil(os));
 931                 if (err) {
 932                         dmu_objset_rele(os, saa);
 933                         return (err);
 934                 }
 935         }
 936 
 937         soa = kmem_zalloc(sizeof (*soa), KM_SLEEP);
 938         soa->soa_saa = saa;
 939         soa->soa_longname = snapname;
 940         soa->soa_snapname = strchr(snapname, '@') + 1;
 941         soa->soa_tot_cnt = cnt;
 942 
 943         dsl_sync_task_create(saa->saa_dstg, snapshot_check, snapshot_sync,
 944             os, soa, 3);
 945 
 946         return (0);
 947 }
 948 
 949 /*
 950  * The snapshots must all be in the same pool.
 951  */
 952 int
 953 dmu_objset_snapshot(nvlist_t *snaps, nvlist_t *props, nvlist_t *errors)
 954 {
 955         dsl_sync_task_t *dst;
 956         snapallarg_t saa = { 0 };
 957         spa_t *spa;
 958         int rv = 0;
 959         int err;
 960         nvpair_t *pair;
 961         nvlist_t *cnt_track = NULL;
 962         char *pdelim;
 963         uint64_t val;
 964         char nm[MAXPATHLEN];
 965 
 966         pair = nvlist_next_nvpair(snaps, NULL);
 967         if (pair == NULL)
 968                 return (0);
 969 
 970         err = spa_open(nvpair_name(pair), &spa, FTAG);
 971         if (err)
 972                 return (err);
 973         saa.saa_dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
 974         saa.saa_props = props;
 975         saa.saa_needsuspend = (spa_version(spa) < SPA_VERSION_FAST_SNAP);
 976         saa.saa_cr = CRED();
 977 
 978         /*
 979          * Pre-compute how many total new snapshots will be created for each
 980          * level in the tree and below. This is needed for validating the
 981          * snapshot limit when taking a recursive snapshot.
 982          *
 983          * The problem is that the counts are not actually adjusted when
 984          * we are checking, only when we finally sync. For a single snapshot,
 985          * this is easy, the count will increase by 1 at each node up the tree,
 986          * but its more complicated for recursive snapshots. Since we are
 987          * validating each snapshot independently we need to be sure that we
 988          * are validating the complete count for the entire set of snapshots.
 989          * We do this by rolling up the counts for each component of the name
 990          * into an nvlist then we'll use that count in the validation of each
 991          * individual snapshot.
 992          *
 993          * We validated the snapshot names in zfs_ioc_snapshot so we know they
 994          * have a '@'.
 995          */
 996         cnt_track = fnvlist_alloc();
 997 
 998         for (pair = nvlist_next_nvpair(snaps, NULL); pair != NULL;
 999             pair = nvlist_next_nvpair(snaps, pair)) {
1000                 (void) strlcpy(nm, nvpair_name(pair), sizeof (nm));
1001                 pdelim = strchr(nm, '@');
1002                 *pdelim = '\0';
1003 
1004                 do {
1005                         if (nvlist_lookup_uint64(cnt_track, nm, &val) == 0) {
1006                                 /* update existing entry */
1007                                 fnvlist_add_uint64(cnt_track, nm, val + 1);
1008                         } else {
1009                                 /* add to list */
1010                                 fnvlist_add_uint64(cnt_track, nm, 1);
1011                         }
1012 
1013                         pdelim = strrchr(nm, '/');
1014                         if (pdelim != NULL)
1015                                 *pdelim = '\0';
1016                 } while (pdelim != NULL);
1017         }
1018 
1019         /*
1020          * We've calculated the counts, now validate.
1021          */
1022         for (pair = nvlist_next_nvpair(snaps, NULL); pair != NULL;
1023             pair = nvlist_next_nvpair(snaps, pair)) {
1024                 (void) strlcpy(nm, nvpair_name(pair), sizeof (nm));
1025                 pdelim = strchr(nm, '@');
1026                 *pdelim = '\0';
1027 
1028                 val = fnvlist_lookup_uint64(cnt_track, nm);
1029                 err = snapshot_one_impl(nvpair_name(pair), &saa, val);
1030                 if (err != 0) {
1031                         if (errors != NULL) {
1032                                 fnvlist_add_int32(errors,
1033                                     nvpair_name(pair), err);
1034                         }
1035                         rv = err;
1036                 }
1037         }
1038 
1039         nvlist_free(cnt_track);
1040 
1041         /*
1042          * If any call to snapshot_one_impl() failed, don't execute the
1043          * sync task.  The error handling code below will clean up the
1044          * snaponearg_t from any successful calls to
1045          * snapshot_one_impl().
1046          */
1047         if (rv == 0)
1048                 err = dsl_sync_task_group_wait(saa.saa_dstg);
1049         if (err != 0)
1050                 rv = err;
1051 
1052         for (dst = list_head(&saa.saa_dstg->dstg_tasks); dst;
1053             dst = list_next(&saa.saa_dstg->dstg_tasks, dst)) {
1054                 objset_t *os = dst->dst_arg1;
1055                 snaponearg_t *soa = dst->dst_arg2;
1056                 if (dst->dst_err != 0) {
1057                         if (errors != NULL) {
1058                                 fnvlist_add_int32(errors,
1059                                     soa->soa_longname, dst->dst_err);
1060                         }
1061                         rv = dst->dst_err;
1062                 }
1063 
1064                 if (saa.saa_needsuspend)
1065                         zil_resume(dmu_objset_zil(os));
1066                 dmu_objset_rele(os, &saa);
1067                 kmem_free(soa, sizeof (*soa));
1068         }
1069 
1070         dsl_sync_task_group_destroy(saa.saa_dstg);
1071         spa_close(spa, FTAG);
1072         return (rv);
1073 }
1074 
1075 int
1076 dmu_objset_snapshot_one(const char *fsname, const char *snapname)
1077 {
1078         int err;
1079         char *longsnap = kmem_asprintf("%s@%s", fsname, snapname);
1080         nvlist_t *snaps = fnvlist_alloc();
1081 
1082         fnvlist_add_boolean(snaps, longsnap);
1083         err = dmu_objset_snapshot(snaps, NULL, NULL);
1084         fnvlist_free(snaps);
1085         strfree(longsnap);
1086         return (err);
1087 }
1088 
1089 int
1090 dmu_objset_snapshot_tmp(const char *snapname, const char *tag, int cleanup_fd)
1091 {
1092         dsl_sync_task_t *dst;
1093         snapallarg_t saa = { 0 };
1094         spa_t *spa;
1095         minor_t minor;
1096         int err;
1097 
1098         err = spa_open(snapname, &spa, FTAG);
1099         if (err)
1100                 return (err);
1101         saa.saa_dstg = dsl_sync_task_group_create(spa_get_dsl(spa));
1102         saa.saa_htag = tag;
1103         saa.saa_needsuspend = (spa_version(spa) < SPA_VERSION_FAST_SNAP);
1104         saa.saa_temporary = B_TRUE;
1105 
1106         if (cleanup_fd < 0) {
1107                 spa_close(spa, FTAG);
1108                 return (EINVAL);
1109         }
1110         if ((err = zfs_onexit_fd_hold(cleanup_fd, &minor)) != 0) {
1111                 spa_close(spa, FTAG);
1112                 return (err);
1113         }
1114 
1115         err = snapshot_one_impl(snapname, &saa, 1);
1116 
1117         if (err == 0)
1118                 err = dsl_sync_task_group_wait(saa.saa_dstg);
1119 
1120         for (dst = list_head(&saa.saa_dstg->dstg_tasks); dst;
1121             dst = list_next(&saa.saa_dstg->dstg_tasks, dst)) {
1122                 objset_t *os = dst->dst_arg1;
1123                 dsl_register_onexit_hold_cleanup(saa.saa_newds, tag, minor);
1124                 if (saa.saa_needsuspend)
1125                         zil_resume(dmu_objset_zil(os));
1126                 dmu_objset_rele(os, &saa);
1127         }
1128 
1129         zfs_onexit_fd_rele(cleanup_fd);
1130         dsl_sync_task_group_destroy(saa.saa_dstg);
1131         spa_close(spa, FTAG);
1132         return (err);
1133 }
1134 
1135 
1136 static void
1137 dmu_objset_sync_dnodes(list_t *list, list_t *newlist, dmu_tx_t *tx)
1138 {
1139         dnode_t *dn;
1140 
1141         while (dn = list_head(list)) {
1142                 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
1143                 ASSERT(dn->dn_dbuf->db_data_pending);
1144                 /*
1145                  * Initialize dn_zio outside dnode_sync() because the
1146                  * meta-dnode needs to set it ouside dnode_sync().
1147                  */
1148                 dn->dn_zio = dn->dn_dbuf->db_data_pending->dr_zio;
1149                 ASSERT(dn->dn_zio);
1150 
1151                 ASSERT3U(dn->dn_nlevels, <=, DN_MAX_LEVELS);
1152                 list_remove(list, dn);
1153 
1154                 if (newlist) {
1155                         (void) dnode_add_ref(dn, newlist);
1156                         list_insert_tail(newlist, dn);
1157                 }
1158 
1159                 dnode_sync(dn, tx);
1160         }
1161 }
1162 
1163 /* ARGSUSED */
1164 static void
1165 dmu_objset_write_ready(zio_t *zio, arc_buf_t *abuf, void *arg)
1166 {
1167         blkptr_t *bp = zio->io_bp;
1168         objset_t *os = arg;
1169         dnode_phys_t *dnp = &os->os_phys->os_meta_dnode;
1170 
1171         ASSERT(bp == os->os_rootbp);
1172         ASSERT(BP_GET_TYPE(bp) == DMU_OT_OBJSET);
1173         ASSERT(BP_GET_LEVEL(bp) == 0);
1174 
1175         /*
1176          * Update rootbp fill count: it should be the number of objects
1177          * allocated in the object set (not counting the "special"
1178          * objects that are stored in the objset_phys_t -- the meta
1179          * dnode and user/group accounting objects).
1180          */
1181         bp->blk_fill = 0;
1182         for (int i = 0; i < dnp->dn_nblkptr; i++)
1183                 bp->blk_fill += dnp->dn_blkptr[i].blk_fill;
1184 }
1185 
1186 /* ARGSUSED */
1187 static void
1188 dmu_objset_write_done(zio_t *zio, arc_buf_t *abuf, void *arg)
1189 {
1190         blkptr_t *bp = zio->io_bp;
1191         blkptr_t *bp_orig = &zio->io_bp_orig;
1192         objset_t *os = arg;
1193 
1194         if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
1195                 ASSERT(BP_EQUAL(bp, bp_orig));
1196         } else {
1197                 dsl_dataset_t *ds = os->os_dsl_dataset;
1198                 dmu_tx_t *tx = os->os_synctx;
1199 
1200                 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
1201                 dsl_dataset_block_born(ds, bp, tx);
1202         }
1203 }
1204 
1205 /* called from dsl */
1206 void
1207 dmu_objset_sync(objset_t *os, zio_t *pio, dmu_tx_t *tx)
1208 {
1209         int txgoff;
1210         zbookmark_t zb;
1211         zio_prop_t zp;
1212         zio_t *zio;
1213         list_t *list;
1214         list_t *newlist = NULL;
1215         dbuf_dirty_record_t *dr;
1216 
1217         dprintf_ds(os->os_dsl_dataset, "txg=%llu\n", tx->tx_txg);
1218 
1219         ASSERT(dmu_tx_is_syncing(tx));
1220         /* XXX the write_done callback should really give us the tx... */
1221         os->os_synctx = tx;
1222 
1223         if (os->os_dsl_dataset == NULL) {
1224                 /*
1225                  * This is the MOS.  If we have upgraded,
1226                  * spa_max_replication() could change, so reset
1227                  * os_copies here.
1228                  */
1229                 os->os_copies = spa_max_replication(os->os_spa);
1230         }
1231 
1232         /*
1233          * Create the root block IO
1234          */
1235         SET_BOOKMARK(&zb, os->os_dsl_dataset ?
1236             os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1237             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
1238         VERIFY3U(0, ==, arc_release_bp(os->os_phys_buf, &os->os_phys_buf,
1239             os->os_rootbp, os->os_spa, &zb));
1240 
1241         dmu_write_policy(os, NULL, 0, 0, &zp);
1242 
1243         zio = arc_write(pio, os->os_spa, tx->tx_txg,
1244             os->os_rootbp, os->os_phys_buf, DMU_OS_IS_L2CACHEABLE(os), &zp,
1245             dmu_objset_write_ready, dmu_objset_write_done, os,
1246             ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
1247 
1248         /*
1249          * Sync special dnodes - the parent IO for the sync is the root block
1250          */
1251         DMU_META_DNODE(os)->dn_zio = zio;
1252         dnode_sync(DMU_META_DNODE(os), tx);
1253 
1254         os->os_phys->os_flags = os->os_flags;
1255 
1256         if (DMU_USERUSED_DNODE(os) &&
1257             DMU_USERUSED_DNODE(os)->dn_type != DMU_OT_NONE) {
1258                 DMU_USERUSED_DNODE(os)->dn_zio = zio;
1259                 dnode_sync(DMU_USERUSED_DNODE(os), tx);
1260                 DMU_GROUPUSED_DNODE(os)->dn_zio = zio;
1261                 dnode_sync(DMU_GROUPUSED_DNODE(os), tx);
1262         }
1263 
1264         txgoff = tx->tx_txg & TXG_MASK;
1265 
1266         if (dmu_objset_userused_enabled(os)) {
1267                 newlist = &os->os_synced_dnodes;
1268                 /*
1269                  * We must create the list here because it uses the
1270                  * dn_dirty_link[] of this txg.
1271                  */
1272                 list_create(newlist, sizeof (dnode_t),
1273                     offsetof(dnode_t, dn_dirty_link[txgoff]));
1274         }
1275 
1276         dmu_objset_sync_dnodes(&os->os_free_dnodes[txgoff], newlist, tx);
1277         dmu_objset_sync_dnodes(&os->os_dirty_dnodes[txgoff], newlist, tx);
1278 
1279         list = &DMU_META_DNODE(os)->dn_dirty_records[txgoff];
1280         while (dr = list_head(list)) {
1281                 ASSERT(dr->dr_dbuf->db_level == 0);
1282                 list_remove(list, dr);
1283                 if (dr->dr_zio)
1284                         zio_nowait(dr->dr_zio);
1285         }
1286         /*
1287          * Free intent log blocks up to this tx.
1288          */
1289         zil_sync(os->os_zil, tx);
1290         os->os_phys->os_zil_header = os->os_zil_header;
1291         zio_nowait(zio);
1292 }
1293 
1294 boolean_t
1295 dmu_objset_is_dirty(objset_t *os, uint64_t txg)
1296 {
1297         return (!list_is_empty(&os->os_dirty_dnodes[txg & TXG_MASK]) ||
1298             !list_is_empty(&os->os_free_dnodes[txg & TXG_MASK]));
1299 }
1300 
1301 static objset_used_cb_t *used_cbs[DMU_OST_NUMTYPES];
1302 
1303 void
1304 dmu_objset_register_type(dmu_objset_type_t ost, objset_used_cb_t *cb)
1305 {
1306         used_cbs[ost] = cb;
1307 }
1308 
1309 boolean_t
1310 dmu_objset_userused_enabled(objset_t *os)
1311 {
1312         return (spa_version(os->os_spa) >= SPA_VERSION_USERSPACE &&
1313             used_cbs[os->os_phys->os_type] != NULL &&
1314             DMU_USERUSED_DNODE(os) != NULL);
1315 }
1316 
1317 static void
1318 do_userquota_update(objset_t *os, uint64_t used, uint64_t flags,
1319     uint64_t user, uint64_t group, boolean_t subtract, dmu_tx_t *tx)
1320 {
1321         if ((flags & DNODE_FLAG_USERUSED_ACCOUNTED)) {
1322                 int64_t delta = DNODE_SIZE + used;
1323                 if (subtract)
1324                         delta = -delta;
1325                 VERIFY3U(0, ==, zap_increment_int(os, DMU_USERUSED_OBJECT,
1326                     user, delta, tx));
1327                 VERIFY3U(0, ==, zap_increment_int(os, DMU_GROUPUSED_OBJECT,
1328                     group, delta, tx));
1329         }
1330 }
1331 
1332 void
1333 dmu_objset_do_userquota_updates(objset_t *os, dmu_tx_t *tx)
1334 {
1335         dnode_t *dn;
1336         list_t *list = &os->os_synced_dnodes;
1337 
1338         ASSERT(list_head(list) == NULL || dmu_objset_userused_enabled(os));
1339 
1340         while (dn = list_head(list)) {
1341                 int flags;
1342                 ASSERT(!DMU_OBJECT_IS_SPECIAL(dn->dn_object));
1343                 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE ||
1344                     dn->dn_phys->dn_flags &
1345                     DNODE_FLAG_USERUSED_ACCOUNTED);
1346 
1347                 /* Allocate the user/groupused objects if necessary. */
1348                 if (DMU_USERUSED_DNODE(os)->dn_type == DMU_OT_NONE) {
1349                         VERIFY(0 == zap_create_claim(os,
1350                             DMU_USERUSED_OBJECT,
1351                             DMU_OT_USERGROUP_USED, DMU_OT_NONE, 0, tx));
1352                         VERIFY(0 == zap_create_claim(os,
1353                             DMU_GROUPUSED_OBJECT,
1354                             DMU_OT_USERGROUP_USED, DMU_OT_NONE, 0, tx));
1355                 }
1356 
1357                 /*
1358                  * We intentionally modify the zap object even if the
1359                  * net delta is zero.  Otherwise
1360                  * the block of the zap obj could be shared between
1361                  * datasets but need to be different between them after
1362                  * a bprewrite.
1363                  */
1364 
1365                 flags = dn->dn_id_flags;
1366                 ASSERT(flags);
1367                 if (flags & DN_ID_OLD_EXIST)  {
1368                         do_userquota_update(os, dn->dn_oldused, dn->dn_oldflags,
1369                             dn->dn_olduid, dn->dn_oldgid, B_TRUE, tx);
1370                 }
1371                 if (flags & DN_ID_NEW_EXIST) {
1372                         do_userquota_update(os, DN_USED_BYTES(dn->dn_phys),
1373                             dn->dn_phys->dn_flags,  dn->dn_newuid,
1374                             dn->dn_newgid, B_FALSE, tx);
1375                 }
1376 
1377                 mutex_enter(&dn->dn_mtx);
1378                 dn->dn_oldused = 0;
1379                 dn->dn_oldflags = 0;
1380                 if (dn->dn_id_flags & DN_ID_NEW_EXIST) {
1381                         dn->dn_olduid = dn->dn_newuid;
1382                         dn->dn_oldgid = dn->dn_newgid;
1383                         dn->dn_id_flags |= DN_ID_OLD_EXIST;
1384                         if (dn->dn_bonuslen == 0)
1385                                 dn->dn_id_flags |= DN_ID_CHKED_SPILL;
1386                         else
1387                                 dn->dn_id_flags |= DN_ID_CHKED_BONUS;
1388                 }
1389                 dn->dn_id_flags &= ~(DN_ID_NEW_EXIST);
1390                 mutex_exit(&dn->dn_mtx);
1391 
1392                 list_remove(list, dn);
1393                 dnode_rele(dn, list);
1394         }
1395 }
1396 
1397 /*
1398  * Returns a pointer to data to find uid/gid from
1399  *
1400  * If a dirty record for transaction group that is syncing can't
1401  * be found then NULL is returned.  In the NULL case it is assumed
1402  * the uid/gid aren't changing.
1403  */
1404 static void *
1405 dmu_objset_userquota_find_data(dmu_buf_impl_t *db, dmu_tx_t *tx)
1406 {
1407         dbuf_dirty_record_t *dr, **drp;
1408         void *data;
1409 
1410         if (db->db_dirtycnt == 0)
1411                 return (db->db.db_data);  /* Nothing is changing */
1412 
1413         for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1414                 if (dr->dr_txg == tx->tx_txg)
1415                         break;
1416 
1417         if (dr == NULL) {
1418                 data = NULL;
1419         } else {
1420                 dnode_t *dn;
1421 
1422                 DB_DNODE_ENTER(dr->dr_dbuf);
1423                 dn = DB_DNODE(dr->dr_dbuf);
1424 
1425                 if (dn->dn_bonuslen == 0 &&
1426                     dr->dr_dbuf->db_blkid == DMU_SPILL_BLKID)
1427                         data = dr->dt.dl.dr_data->b_data;
1428                 else
1429                         data = dr->dt.dl.dr_data;
1430 
1431                 DB_DNODE_EXIT(dr->dr_dbuf);
1432         }
1433 
1434         return (data);
1435 }
1436 
1437 void
1438 dmu_objset_userquota_get_ids(dnode_t *dn, boolean_t before, dmu_tx_t *tx)
1439 {
1440         objset_t *os = dn->dn_objset;
1441         void *data = NULL;
1442         dmu_buf_impl_t *db = NULL;
1443         uint64_t *user, *group;
1444         int flags = dn->dn_id_flags;
1445         int error;
1446         boolean_t have_spill = B_FALSE;
1447 
1448         if (!dmu_objset_userused_enabled(dn->dn_objset))
1449                 return;
1450 
1451         if (before && (flags & (DN_ID_CHKED_BONUS|DN_ID_OLD_EXIST|
1452             DN_ID_CHKED_SPILL)))
1453                 return;
1454 
1455         if (before && dn->dn_bonuslen != 0)
1456                 data = DN_BONUS(dn->dn_phys);
1457         else if (!before && dn->dn_bonuslen != 0) {
1458                 if (dn->dn_bonus) {
1459                         db = dn->dn_bonus;
1460                         mutex_enter(&db->db_mtx);
1461                         data = dmu_objset_userquota_find_data(db, tx);
1462                 } else {
1463                         data = DN_BONUS(dn->dn_phys);
1464                 }
1465         } else if (dn->dn_bonuslen == 0 && dn->dn_bonustype == DMU_OT_SA) {
1466                         int rf = 0;
1467 
1468                         if (RW_WRITE_HELD(&dn->dn_struct_rwlock))
1469                                 rf |= DB_RF_HAVESTRUCT;
1470                         error = dmu_spill_hold_by_dnode(dn,
1471                             rf | DB_RF_MUST_SUCCEED,
1472                             FTAG, (dmu_buf_t **)&db);
1473                         ASSERT(error == 0);
1474                         mutex_enter(&db->db_mtx);
1475                         data = (before) ? db->db.db_data :
1476                             dmu_objset_userquota_find_data(db, tx);
1477                         have_spill = B_TRUE;
1478         } else {
1479                 mutex_enter(&dn->dn_mtx);
1480                 dn->dn_id_flags |= DN_ID_CHKED_BONUS;
1481                 mutex_exit(&dn->dn_mtx);
1482                 return;
1483         }
1484 
1485         if (before) {
1486                 ASSERT(data);
1487                 user = &dn->dn_olduid;
1488                 group = &dn->dn_oldgid;
1489         } else if (data) {
1490                 user = &dn->dn_newuid;
1491                 group = &dn->dn_newgid;
1492         }
1493 
1494         /*
1495          * Must always call the callback in case the object
1496          * type has changed and that type isn't an object type to track
1497          */
1498         error = used_cbs[os->os_phys->os_type](dn->dn_bonustype, data,
1499             user, group);
1500 
1501         /*
1502          * Preserve existing uid/gid when the callback can't determine
1503          * what the new uid/gid are and the callback returned EEXIST.
1504          * The EEXIST error tells us to just use the existing uid/gid.
1505          * If we don't know what the old values are then just assign
1506          * them to 0, since that is a new file  being created.
1507          */
1508         if (!before && data == NULL && error == EEXIST) {
1509                 if (flags & DN_ID_OLD_EXIST) {
1510                         dn->dn_newuid = dn->dn_olduid;
1511                         dn->dn_newgid = dn->dn_oldgid;
1512                 } else {
1513                         dn->dn_newuid = 0;
1514                         dn->dn_newgid = 0;
1515                 }
1516                 error = 0;
1517         }
1518 
1519         if (db)
1520                 mutex_exit(&db->db_mtx);
1521 
1522         mutex_enter(&dn->dn_mtx);
1523         if (error == 0 && before)
1524                 dn->dn_id_flags |= DN_ID_OLD_EXIST;
1525         if (error == 0 && !before)
1526                 dn->dn_id_flags |= DN_ID_NEW_EXIST;
1527 
1528         if (have_spill) {
1529                 dn->dn_id_flags |= DN_ID_CHKED_SPILL;
1530         } else {
1531                 dn->dn_id_flags |= DN_ID_CHKED_BONUS;
1532         }
1533         mutex_exit(&dn->dn_mtx);
1534         if (have_spill)
1535                 dmu_buf_rele((dmu_buf_t *)db, FTAG);
1536 }
1537 
1538 boolean_t
1539 dmu_objset_userspace_present(objset_t *os)
1540 {
1541         return (os->os_phys->os_flags &
1542             OBJSET_FLAG_USERACCOUNTING_COMPLETE);
1543 }
1544 
1545 int
1546 dmu_objset_userspace_upgrade(objset_t *os)
1547 {
1548         uint64_t obj;
1549         int err = 0;
1550 
1551         if (dmu_objset_userspace_present(os))
1552                 return (0);
1553         if (!dmu_objset_userused_enabled(os))
1554                 return (ENOTSUP);
1555         if (dmu_objset_is_snapshot(os))
1556                 return (EINVAL);
1557 
1558         /*
1559          * We simply need to mark every object dirty, so that it will be
1560          * synced out and now accounted.  If this is called
1561          * concurrently, or if we already did some work before crashing,
1562          * that's fine, since we track each object's accounted state
1563          * independently.
1564          */
1565 
1566         for (obj = 0; err == 0; err = dmu_object_next(os, &obj, FALSE, 0)) {
1567                 dmu_tx_t *tx;
1568                 dmu_buf_t *db;
1569                 int objerr;
1570 
1571                 if (issig(JUSTLOOKING) && issig(FORREAL))
1572                         return (EINTR);
1573 
1574                 objerr = dmu_bonus_hold(os, obj, FTAG, &db);
1575                 if (objerr)
1576                         continue;
1577                 tx = dmu_tx_create(os);
1578                 dmu_tx_hold_bonus(tx, obj);
1579                 objerr = dmu_tx_assign(tx, TXG_WAIT);
1580                 if (objerr) {
1581                         dmu_tx_abort(tx);
1582                         continue;
1583                 }
1584                 dmu_buf_will_dirty(db, tx);
1585                 dmu_buf_rele(db, FTAG);
1586                 dmu_tx_commit(tx);
1587         }
1588 
1589         os->os_flags |= OBJSET_FLAG_USERACCOUNTING_COMPLETE;
1590         txg_wait_synced(dmu_objset_pool(os), 0);
1591         return (0);
1592 }
1593 
1594 void
1595 dmu_objset_space(objset_t *os, uint64_t *refdbytesp, uint64_t *availbytesp,
1596     uint64_t *usedobjsp, uint64_t *availobjsp)
1597 {
1598         dsl_dataset_space(os->os_dsl_dataset, refdbytesp, availbytesp,
1599             usedobjsp, availobjsp);
1600 }
1601 
1602 uint64_t
1603 dmu_objset_fsid_guid(objset_t *os)
1604 {
1605         return (dsl_dataset_fsid_guid(os->os_dsl_dataset));
1606 }
1607 
1608 void
1609 dmu_objset_fast_stat(objset_t *os, dmu_objset_stats_t *stat)
1610 {
1611         stat->dds_type = os->os_phys->os_type;
1612         if (os->os_dsl_dataset)
1613                 dsl_dataset_fast_stat(os->os_dsl_dataset, stat);
1614 }
1615 
1616 void
1617 dmu_objset_stats(objset_t *os, nvlist_t *nv)
1618 {
1619         ASSERT(os->os_dsl_dataset ||
1620             os->os_phys->os_type == DMU_OST_META);
1621 
1622         if (os->os_dsl_dataset != NULL)
1623                 dsl_dataset_stats(os->os_dsl_dataset, nv);
1624 
1625         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_TYPE,
1626             os->os_phys->os_type);
1627         dsl_prop_nvlist_add_uint64(nv, ZFS_PROP_USERACCOUNTING,
1628             dmu_objset_userspace_present(os));
1629 }
1630 
1631 int
1632 dmu_objset_is_snapshot(objset_t *os)
1633 {
1634         if (os->os_dsl_dataset != NULL)
1635                 return (dsl_dataset_is_snapshot(os->os_dsl_dataset));
1636         else
1637                 return (B_FALSE);
1638 }
1639 
1640 int
1641 dmu_snapshot_realname(objset_t *os, char *name, char *real, int maxlen,
1642     boolean_t *conflict)
1643 {
1644         dsl_dataset_t *ds = os->os_dsl_dataset;
1645         uint64_t ignored;
1646 
1647         if (ds->ds_phys->ds_snapnames_zapobj == 0)
1648                 return (ENOENT);
1649 
1650         return (zap_lookup_norm(ds->ds_dir->dd_pool->dp_meta_objset,
1651             ds->ds_phys->ds_snapnames_zapobj, name, 8, 1, &ignored, MT_FIRST,
1652             real, maxlen, conflict));
1653 }
1654 
1655 int
1656 dmu_snapshot_list_next(objset_t *os, int namelen, char *name,
1657     uint64_t *idp, uint64_t *offp, boolean_t *case_conflict)
1658 {
1659         dsl_dataset_t *ds = os->os_dsl_dataset;
1660         zap_cursor_t cursor;
1661         zap_attribute_t attr;
1662 
1663         if (ds->ds_phys->ds_snapnames_zapobj == 0)
1664                 return (ENOENT);
1665 
1666         zap_cursor_init_serialized(&cursor,
1667             ds->ds_dir->dd_pool->dp_meta_objset,
1668             ds->ds_phys->ds_snapnames_zapobj, *offp);
1669 
1670         if (zap_cursor_retrieve(&cursor, &attr) != 0) {
1671                 zap_cursor_fini(&cursor);
1672                 return (ENOENT);
1673         }
1674 
1675         if (strlen(attr.za_name) + 1 > namelen) {
1676                 zap_cursor_fini(&cursor);
1677                 return (ENAMETOOLONG);
1678         }
1679 
1680         (void) strcpy(name, attr.za_name);
1681         if (idp)
1682                 *idp = attr.za_first_integer;
1683         if (case_conflict)
1684                 *case_conflict = attr.za_normalization_conflict;
1685         zap_cursor_advance(&cursor);
1686         *offp = zap_cursor_serialize(&cursor);
1687         zap_cursor_fini(&cursor);
1688 
1689         return (0);
1690 }
1691 
1692 int
1693 dmu_dir_list_next(objset_t *os, int namelen, char *name,
1694     uint64_t *idp, uint64_t *offp)
1695 {
1696         dsl_dir_t *dd = os->os_dsl_dataset->ds_dir;
1697         zap_cursor_t cursor;
1698         zap_attribute_t attr;
1699 
1700         /* there is no next dir on a snapshot! */
1701         if (os->os_dsl_dataset->ds_object !=
1702             dd->dd_phys->dd_head_dataset_obj)
1703                 return (ENOENT);
1704 
1705         zap_cursor_init_serialized(&cursor,
1706             dd->dd_pool->dp_meta_objset,
1707             dd->dd_phys->dd_child_dir_zapobj, *offp);
1708 
1709         if (zap_cursor_retrieve(&cursor, &attr) != 0) {
1710                 zap_cursor_fini(&cursor);
1711                 return (ENOENT);
1712         }
1713 
1714         if (strlen(attr.za_name) + 1 > namelen) {
1715                 zap_cursor_fini(&cursor);
1716                 return (ENAMETOOLONG);
1717         }
1718 
1719         (void) strcpy(name, attr.za_name);
1720         if (idp)
1721                 *idp = attr.za_first_integer;
1722         zap_cursor_advance(&cursor);
1723         *offp = zap_cursor_serialize(&cursor);
1724         zap_cursor_fini(&cursor);
1725 
1726         return (0);
1727 }
1728 
1729 struct findarg {
1730         int (*func)(const char *, void *);
1731         void *arg;
1732 };
1733 
1734 /* ARGSUSED */
1735 static int
1736 findfunc(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
1737 {
1738         struct findarg *fa = arg;
1739         return (fa->func(dsname, fa->arg));
1740 }
1741 
1742 /*
1743  * Find all objsets under name, and for each, call 'func(child_name, arg)'.
1744  * Perhaps change all callers to use dmu_objset_find_spa()?
1745  */
1746 int
1747 dmu_objset_find(char *name, int func(const char *, void *), void *arg,
1748     int flags)
1749 {
1750         struct findarg fa;
1751         fa.func = func;
1752         fa.arg = arg;
1753         return (dmu_objset_find_spa(NULL, name, findfunc, &fa, flags));
1754 }
1755 
1756 /*
1757  * Find all objsets under name, call func on each
1758  */
1759 int
1760 dmu_objset_find_spa(spa_t *spa, const char *name,
1761     int func(spa_t *, uint64_t, const char *, void *), void *arg, int flags)
1762 {
1763         dsl_dir_t *dd;
1764         dsl_pool_t *dp;
1765         dsl_dataset_t *ds;
1766         zap_cursor_t zc;
1767         zap_attribute_t *attr;
1768         char *child;
1769         uint64_t thisobj;
1770         int err;
1771 
1772         if (name == NULL)
1773                 name = spa_name(spa);
1774         err = dsl_dir_open_spa(spa, name, FTAG, &dd, NULL);
1775         if (err)
1776                 return (err);
1777 
1778         /* Don't visit hidden ($MOS & $ORIGIN) objsets. */
1779         if (dd->dd_myname[0] == '$') {
1780                 dsl_dir_close(dd, FTAG);
1781                 return (0);
1782         }
1783 
1784         thisobj = dd->dd_phys->dd_head_dataset_obj;
1785         attr = kmem_alloc(sizeof (zap_attribute_t), KM_SLEEP);
1786         dp = dd->dd_pool;
1787 
1788         /*
1789          * Iterate over all children.
1790          */
1791         if (flags & DS_FIND_CHILDREN) {
1792                 for (zap_cursor_init(&zc, dp->dp_meta_objset,
1793                     dd->dd_phys->dd_child_dir_zapobj);
1794                     zap_cursor_retrieve(&zc, attr) == 0;
1795                     (void) zap_cursor_advance(&zc)) {
1796                         ASSERT(attr->za_integer_length == sizeof (uint64_t));
1797                         ASSERT(attr->za_num_integers == 1);
1798 
1799                         child = kmem_asprintf("%s/%s", name, attr->za_name);
1800                         err = dmu_objset_find_spa(spa, child, func, arg, flags);
1801                         strfree(child);
1802                         if (err)
1803                                 break;
1804                 }
1805                 zap_cursor_fini(&zc);
1806 
1807                 if (err) {
1808                         dsl_dir_close(dd, FTAG);
1809                         kmem_free(attr, sizeof (zap_attribute_t));
1810                         return (err);
1811                 }
1812         }
1813 
1814         /*
1815          * Iterate over all snapshots.
1816          */
1817         if (flags & DS_FIND_SNAPSHOTS) {
1818                 if (!dsl_pool_sync_context(dp))
1819                         rw_enter(&dp->dp_config_rwlock, RW_READER);
1820                 err = dsl_dataset_hold_obj(dp, thisobj, FTAG, &ds);
1821                 if (!dsl_pool_sync_context(dp))
1822                         rw_exit(&dp->dp_config_rwlock);
1823 
1824                 if (err == 0) {
1825                         uint64_t snapobj = ds->ds_phys->ds_snapnames_zapobj;
1826                         dsl_dataset_rele(ds, FTAG);
1827 
1828                         for (zap_cursor_init(&zc, dp->dp_meta_objset, snapobj);
1829                             zap_cursor_retrieve(&zc, attr) == 0;
1830                             (void) zap_cursor_advance(&zc)) {
1831                                 ASSERT(attr->za_integer_length ==
1832                                     sizeof (uint64_t));
1833                                 ASSERT(attr->za_num_integers == 1);
1834 
1835                                 child = kmem_asprintf("%s@%s",
1836                                     name, attr->za_name);
1837                                 err = func(spa, attr->za_first_integer,
1838                                     child, arg);
1839                                 strfree(child);
1840                                 if (err)
1841                                         break;
1842                         }
1843                         zap_cursor_fini(&zc);
1844                 }
1845         }
1846 
1847         dsl_dir_close(dd, FTAG);
1848         kmem_free(attr, sizeof (zap_attribute_t));
1849 
1850         if (err)
1851                 return (err);
1852 
1853         /*
1854          * Apply to self if appropriate.
1855          */
1856         err = func(spa, thisobj, name, arg);
1857         return (err);
1858 }
1859 
1860 /* ARGSUSED */
1861 int
1862 dmu_objset_prefetch(const char *name, void *arg)
1863 {
1864         dsl_dataset_t *ds;
1865 
1866         if (dsl_dataset_hold(name, FTAG, &ds))
1867                 return (0);
1868 
1869         if (!BP_IS_HOLE(&ds->ds_phys->ds_bp)) {
1870                 mutex_enter(&ds->ds_opening_lock);
1871                 if (ds->ds_objset == NULL) {
1872                         uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1873                         zbookmark_t zb;
1874 
1875                         SET_BOOKMARK(&zb, ds->ds_object, ZB_ROOT_OBJECT,
1876                             ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
1877 
1878                         (void) dsl_read_nolock(NULL, dsl_dataset_get_spa(ds),
1879                             &ds->ds_phys->ds_bp, NULL, NULL,
1880                             ZIO_PRIORITY_ASYNC_READ,
1881                             ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1882                             &aflags, &zb);
1883                 }
1884                 mutex_exit(&ds->ds_opening_lock);
1885         }
1886 
1887         dsl_dataset_rele(ds, FTAG);
1888         return (0);
1889 }
1890 
1891 void
1892 dmu_objset_set_user(objset_t *os, void *user_ptr)
1893 {
1894         ASSERT(MUTEX_HELD(&os->os_user_ptr_lock));
1895         os->os_user_ptr = user_ptr;
1896 }
1897 
1898 void *
1899 dmu_objset_get_user(objset_t *os)
1900 {
1901         ASSERT(MUTEX_HELD(&os->os_user_ptr_lock));
1902         return (os->os_user_ptr);
1903 }