Home | History | Annotate | Line # | Download | only in zfs
      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) 2011, 2016 by Delphix. All rights reserved.
     24  */
     25 /* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */
     26 /* Copyright (c) 2013, Joyent, Inc. All rights reserved. */
     27 /* Copyright (c) 2014, Nexenta Systems, Inc. All rights reserved. */
     28 
     29 #include <sys/dmu.h>
     30 #include <sys/dmu_impl.h>
     31 #include <sys/dmu_tx.h>
     32 #include <sys/dbuf.h>
     33 #include <sys/dnode.h>
     34 #include <sys/zfs_context.h>
     35 #include <sys/dmu_objset.h>
     36 #include <sys/dmu_traverse.h>
     37 #include <sys/dsl_dataset.h>
     38 #include <sys/dsl_dir.h>
     39 #include <sys/dsl_pool.h>
     40 #include <sys/dsl_synctask.h>
     41 #include <sys/dsl_prop.h>
     42 #include <sys/dmu_zfetch.h>
     43 #include <sys/zfs_ioctl.h>
     44 #include <sys/zap.h>
     45 #include <sys/zio_checksum.h>
     46 #include <sys/zio_compress.h>
     47 #include <sys/sa.h>
     48 #include <sys/zfeature.h>
     49 #ifdef _KERNEL
     50 #include <sys/racct.h>
     51 #include <sys/vm.h>
     52 #include <sys/zfs_znode.h>
     53 #endif
     54 
     55 /*
     56  * Enable/disable nopwrite feature.
     57  */
     58 int zfs_nopwrite_enabled = 1;
     59 SYSCTL_DECL(_vfs_zfs);
     60 SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN,
     61     &zfs_nopwrite_enabled, 0, "Enable nopwrite feature");
     62 
     63 /*
     64  * Tunable to control percentage of dirtied blocks from frees in one TXG.
     65  * After this threshold is crossed, additional dirty blocks from frees
     66  * wait until the next TXG.
     67  * A value of zero will disable this throttle.
     68  */
     69 uint32_t zfs_per_txg_dirty_frees_percent = 30;
     70 SYSCTL_INT(_vfs_zfs, OID_AUTO, per_txg_dirty_frees_percent, CTLFLAG_RWTUN,
     71 	&zfs_per_txg_dirty_frees_percent, 0, "Percentage of dirtied blocks from frees in one txg");
     72 
     73 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
     74 	{	DMU_BSWAP_UINT8,	TRUE,	"unallocated"		},
     75 	{	DMU_BSWAP_ZAP,		TRUE,	"object directory"	},
     76 	{	DMU_BSWAP_UINT64,	TRUE,	"object array"		},
     77 	{	DMU_BSWAP_UINT8,	TRUE,	"packed nvlist"		},
     78 	{	DMU_BSWAP_UINT64,	TRUE,	"packed nvlist size"	},
     79 	{	DMU_BSWAP_UINT64,	TRUE,	"bpobj"			},
     80 	{	DMU_BSWAP_UINT64,	TRUE,	"bpobj header"		},
     81 	{	DMU_BSWAP_UINT64,	TRUE,	"SPA space map header"	},
     82 	{	DMU_BSWAP_UINT64,	TRUE,	"SPA space map"		},
     83 	{	DMU_BSWAP_UINT64,	TRUE,	"ZIL intent log"	},
     84 	{	DMU_BSWAP_DNODE,	TRUE,	"DMU dnode"		},
     85 	{	DMU_BSWAP_OBJSET,	TRUE,	"DMU objset"		},
     86 	{	DMU_BSWAP_UINT64,	TRUE,	"DSL directory"		},
     87 	{	DMU_BSWAP_ZAP,		TRUE,	"DSL directory child map"},
     88 	{	DMU_BSWAP_ZAP,		TRUE,	"DSL dataset snap map"	},
     89 	{	DMU_BSWAP_ZAP,		TRUE,	"DSL props"		},
     90 	{	DMU_BSWAP_UINT64,	TRUE,	"DSL dataset"		},
     91 	{	DMU_BSWAP_ZNODE,	TRUE,	"ZFS znode"		},
     92 	{	DMU_BSWAP_OLDACL,	TRUE,	"ZFS V0 ACL"		},
     93 	{	DMU_BSWAP_UINT8,	FALSE,	"ZFS plain file"	},
     94 	{	DMU_BSWAP_ZAP,		TRUE,	"ZFS directory"		},
     95 	{	DMU_BSWAP_ZAP,		TRUE,	"ZFS master node"	},
     96 	{	DMU_BSWAP_ZAP,		TRUE,	"ZFS delete queue"	},
     97 	{	DMU_BSWAP_UINT8,	FALSE,	"zvol object"		},
     98 	{	DMU_BSWAP_ZAP,		TRUE,	"zvol prop"		},
     99 	{	DMU_BSWAP_UINT8,	FALSE,	"other uint8[]"		},
    100 	{	DMU_BSWAP_UINT64,	FALSE,	"other uint64[]"	},
    101 	{	DMU_BSWAP_ZAP,		TRUE,	"other ZAP"		},
    102 	{	DMU_BSWAP_ZAP,		TRUE,	"persistent error log"	},
    103 	{	DMU_BSWAP_UINT8,	TRUE,	"SPA history"		},
    104 	{	DMU_BSWAP_UINT64,	TRUE,	"SPA history offsets"	},
    105 	{	DMU_BSWAP_ZAP,		TRUE,	"Pool properties"	},
    106 	{	DMU_BSWAP_ZAP,		TRUE,	"DSL permissions"	},
    107 	{	DMU_BSWAP_ACL,		TRUE,	"ZFS ACL"		},
    108 	{	DMU_BSWAP_UINT8,	TRUE,	"ZFS SYSACL"		},
    109 	{	DMU_BSWAP_UINT8,	TRUE,	"FUID table"		},
    110 	{	DMU_BSWAP_UINT64,	TRUE,	"FUID table size"	},
    111 	{	DMU_BSWAP_ZAP,		TRUE,	"DSL dataset next clones"},
    112 	{	DMU_BSWAP_ZAP,		TRUE,	"scan work queue"	},
    113 	{	DMU_BSWAP_ZAP,		TRUE,	"ZFS user/group used"	},
    114 	{	DMU_BSWAP_ZAP,		TRUE,	"ZFS user/group quota"	},
    115 	{	DMU_BSWAP_ZAP,		TRUE,	"snapshot refcount tags"},
    116 	{	DMU_BSWAP_ZAP,		TRUE,	"DDT ZAP algorithm"	},
    117 	{	DMU_BSWAP_ZAP,		TRUE,	"DDT statistics"	},
    118 	{	DMU_BSWAP_UINT8,	TRUE,	"System attributes"	},
    119 	{	DMU_BSWAP_ZAP,		TRUE,	"SA master node"	},
    120 	{	DMU_BSWAP_ZAP,		TRUE,	"SA attr registration"	},
    121 	{	DMU_BSWAP_ZAP,		TRUE,	"SA attr layouts"	},
    122 	{	DMU_BSWAP_ZAP,		TRUE,	"scan translations"	},
    123 	{	DMU_BSWAP_UINT8,	FALSE,	"deduplicated block"	},
    124 	{	DMU_BSWAP_ZAP,		TRUE,	"DSL deadlist map"	},
    125 	{	DMU_BSWAP_UINT64,	TRUE,	"DSL deadlist map hdr"	},
    126 	{	DMU_BSWAP_ZAP,		TRUE,	"DSL dir clones"	},
    127 	{	DMU_BSWAP_UINT64,	TRUE,	"bpobj subobj"		}
    128 };
    129 
    130 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
    131 	{	byteswap_uint8_array,	"uint8"		},
    132 	{	byteswap_uint16_array,	"uint16"	},
    133 	{	byteswap_uint32_array,	"uint32"	},
    134 	{	byteswap_uint64_array,	"uint64"	},
    135 	{	zap_byteswap,		"zap"		},
    136 	{	dnode_buf_byteswap,	"dnode"		},
    137 	{	dmu_objset_byteswap,	"objset"	},
    138 	{	zfs_znode_byteswap,	"znode"		},
    139 	{	zfs_oldacl_byteswap,	"oldacl"	},
    140 	{	zfs_acl_byteswap,	"acl"		}
    141 };
    142 
    143 int
    144 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
    145     void *tag, dmu_buf_t **dbp)
    146 {
    147 	uint64_t blkid;
    148 	dmu_buf_impl_t *db;
    149 
    150 	blkid = dbuf_whichblock(dn, 0, offset);
    151 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
    152 	db = dbuf_hold(dn, blkid, tag);
    153 	rw_exit(&dn->dn_struct_rwlock);
    154 
    155 	if (db == NULL) {
    156 		*dbp = NULL;
    157 		return (SET_ERROR(EIO));
    158 	}
    159 
    160 	*dbp = &db->db;
    161 	return (0);
    162 }
    163 int
    164 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
    165     void *tag, dmu_buf_t **dbp)
    166 {
    167 	dnode_t *dn;
    168 	uint64_t blkid;
    169 	dmu_buf_impl_t *db;
    170 	int err;
    171 
    172 	err = dnode_hold(os, object, FTAG, &dn);
    173 	if (err)
    174 		return (err);
    175 	blkid = dbuf_whichblock(dn, 0, offset);
    176 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
    177 	db = dbuf_hold(dn, blkid, tag);
    178 	rw_exit(&dn->dn_struct_rwlock);
    179 	dnode_rele(dn, FTAG);
    180 
    181 	if (db == NULL) {
    182 		*dbp = NULL;
    183 		return (SET_ERROR(EIO));
    184 	}
    185 
    186 	*dbp = &db->db;
    187 	return (err);
    188 }
    189 
    190 int
    191 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
    192     void *tag, dmu_buf_t **dbp, int flags)
    193 {
    194 	int err;
    195 	int db_flags = DB_RF_CANFAIL;
    196 
    197 	if (flags & DMU_READ_NO_PREFETCH)
    198 		db_flags |= DB_RF_NOPREFETCH;
    199 
    200 	err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
    201 	if (err == 0) {
    202 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
    203 		err = dbuf_read(db, NULL, db_flags);
    204 		if (err != 0) {
    205 			dbuf_rele(db, tag);
    206 			*dbp = NULL;
    207 		}
    208 	}
    209 
    210 	return (err);
    211 }
    212 
    213 int
    214 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
    215     void *tag, dmu_buf_t **dbp, int flags)
    216 {
    217 	int err;
    218 	int db_flags = DB_RF_CANFAIL;
    219 
    220 	if (flags & DMU_READ_NO_PREFETCH)
    221 		db_flags |= DB_RF_NOPREFETCH;
    222 
    223 	err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
    224 	if (err == 0) {
    225 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
    226 		err = dbuf_read(db, NULL, db_flags);
    227 		if (err != 0) {
    228 			dbuf_rele(db, tag);
    229 			*dbp = NULL;
    230 		}
    231 	}
    232 
    233 	return (err);
    234 }
    235 
    236 int
    237 dmu_bonus_max(void)
    238 {
    239 	return (DN_MAX_BONUSLEN);
    240 }
    241 
    242 int
    243 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
    244 {
    245 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
    246 	dnode_t *dn;
    247 	int error;
    248 
    249 	DB_DNODE_ENTER(db);
    250 	dn = DB_DNODE(db);
    251 
    252 	if (dn->dn_bonus != db) {
    253 		error = SET_ERROR(EINVAL);
    254 	} else if (newsize < 0 || newsize > db_fake->db_size) {
    255 		error = SET_ERROR(EINVAL);
    256 	} else {
    257 		dnode_setbonuslen(dn, newsize, tx);
    258 		error = 0;
    259 	}
    260 
    261 	DB_DNODE_EXIT(db);
    262 	return (error);
    263 }
    264 
    265 int
    266 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
    267 {
    268 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
    269 	dnode_t *dn;
    270 	int error;
    271 
    272 	DB_DNODE_ENTER(db);
    273 	dn = DB_DNODE(db);
    274 
    275 	if (!DMU_OT_IS_VALID(type)) {
    276 		error = SET_ERROR(EINVAL);
    277 	} else if (dn->dn_bonus != db) {
    278 		error = SET_ERROR(EINVAL);
    279 	} else {
    280 		dnode_setbonus_type(dn, type, tx);
    281 		error = 0;
    282 	}
    283 
    284 	DB_DNODE_EXIT(db);
    285 	return (error);
    286 }
    287 
    288 dmu_object_type_t
    289 dmu_get_bonustype(dmu_buf_t *db_fake)
    290 {
    291 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
    292 	dnode_t *dn;
    293 	dmu_object_type_t type;
    294 
    295 	DB_DNODE_ENTER(db);
    296 	dn = DB_DNODE(db);
    297 	type = dn->dn_bonustype;
    298 	DB_DNODE_EXIT(db);
    299 
    300 	return (type);
    301 }
    302 
    303 int
    304 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
    305 {
    306 	dnode_t *dn;
    307 	int error;
    308 
    309 	error = dnode_hold(os, object, FTAG, &dn);
    310 	dbuf_rm_spill(dn, tx);
    311 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
    312 	dnode_rm_spill(dn, tx);
    313 	rw_exit(&dn->dn_struct_rwlock);
    314 	dnode_rele(dn, FTAG);
    315 	return (error);
    316 }
    317 
    318 /*
    319  * returns ENOENT, EIO, or 0.
    320  */
    321 int
    322 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
    323 {
    324 	dnode_t *dn;
    325 	dmu_buf_impl_t *db;
    326 	int error;
    327 
    328 	error = dnode_hold(os, object, FTAG, &dn);
    329 	if (error)
    330 		return (error);
    331 
    332 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
    333 	if (dn->dn_bonus == NULL) {
    334 		rw_exit(&dn->dn_struct_rwlock);
    335 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
    336 		if (dn->dn_bonus == NULL)
    337 			dbuf_create_bonus(dn);
    338 	}
    339 	db = dn->dn_bonus;
    340 
    341 	/* as long as the bonus buf is held, the dnode will be held */
    342 	if (refcount_add(&db->db_holds, tag) == 1) {
    343 		VERIFY(dnode_add_ref(dn, db));
    344 		atomic_inc_32(&dn->dn_dbufs_count);
    345 	}
    346 
    347 	/*
    348 	 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
    349 	 * hold and incrementing the dbuf count to ensure that dnode_move() sees
    350 	 * a dnode hold for every dbuf.
    351 	 */
    352 	rw_exit(&dn->dn_struct_rwlock);
    353 
    354 	dnode_rele(dn, FTAG);
    355 
    356 	VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
    357 
    358 	*dbp = &db->db;
    359 	return (0);
    360 }
    361 
    362 /*
    363  * returns ENOENT, EIO, or 0.
    364  *
    365  * This interface will allocate a blank spill dbuf when a spill blk
    366  * doesn't already exist on the dnode.
    367  *
    368  * if you only want to find an already existing spill db, then
    369  * dmu_spill_hold_existing() should be used.
    370  */
    371 int
    372 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
    373 {
    374 	dmu_buf_impl_t *db = NULL;
    375 	int err;
    376 
    377 	if ((flags & DB_RF_HAVESTRUCT) == 0)
    378 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
    379 
    380 	db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
    381 
    382 	if ((flags & DB_RF_HAVESTRUCT) == 0)
    383 		rw_exit(&dn->dn_struct_rwlock);
    384 
    385 	ASSERT(db != NULL);
    386 	err = dbuf_read(db, NULL, flags);
    387 	if (err == 0)
    388 		*dbp = &db->db;
    389 	else
    390 		dbuf_rele(db, tag);
    391 	return (err);
    392 }
    393 
    394 int
    395 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
    396 {
    397 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
    398 	dnode_t *dn;
    399 	int err;
    400 
    401 	DB_DNODE_ENTER(db);
    402 	dn = DB_DNODE(db);
    403 
    404 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
    405 		err = SET_ERROR(EINVAL);
    406 	} else {
    407 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
    408 
    409 		if (!dn->dn_have_spill) {
    410 			err = SET_ERROR(ENOENT);
    411 		} else {
    412 			err = dmu_spill_hold_by_dnode(dn,
    413 			    DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
    414 		}
    415 
    416 		rw_exit(&dn->dn_struct_rwlock);
    417 	}
    418 
    419 	DB_DNODE_EXIT(db);
    420 	return (err);
    421 }
    422 
    423 int
    424 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
    425 {
    426 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
    427 	dnode_t *dn;
    428 	int err;
    429 
    430 	DB_DNODE_ENTER(db);
    431 	dn = DB_DNODE(db);
    432 	err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
    433 	DB_DNODE_EXIT(db);
    434 
    435 	return (err);
    436 }
    437 
    438 /*
    439  * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
    440  * to take a held dnode rather than <os, object> -- the lookup is wasteful,
    441  * and can induce severe lock contention when writing to several files
    442  * whose dnodes are in the same block.
    443  */
    444 static int
    445 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
    446     boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
    447 {
    448 	dmu_buf_t **dbp;
    449 	uint64_t blkid, nblks, i;
    450 	uint32_t dbuf_flags;
    451 	int err;
    452 	zio_t *zio;
    453 
    454 	ASSERT(length <= DMU_MAX_ACCESS);
    455 
    456 	/*
    457 	 * Note: We directly notify the prefetch code of this read, so that
    458 	 * we can tell it about the multi-block read.  dbuf_read() only knows
    459 	 * about the one block it is accessing.
    460 	 */
    461 	dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
    462 	    DB_RF_NOPREFETCH;
    463 
    464 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
    465 	if (dn->dn_datablkshift) {
    466 		int blkshift = dn->dn_datablkshift;
    467 		nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
    468 		    P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
    469 	} else {
    470 		if (offset + length > dn->dn_datablksz) {
    471 			zfs_panic_recover("zfs: accessing past end of object "
    472 			    "%llx/%llx (size=%u access=%llu+%llu)",
    473 			    (longlong_t)dn->dn_objset->
    474 			    os_dsl_dataset->ds_object,
    475 			    (longlong_t)dn->dn_object, dn->dn_datablksz,
    476 			    (longlong_t)offset, (longlong_t)length);
    477 			rw_exit(&dn->dn_struct_rwlock);
    478 			return (SET_ERROR(EIO));
    479 		}
    480 		nblks = 1;
    481 	}
    482 	dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
    483 
    484 #if defined(_KERNEL) && defined(RACCT)
    485 	if (racct_enable && !read) {
    486 		PROC_LOCK(curproc);
    487 		racct_add_force(curproc, RACCT_WRITEBPS, length);
    488 		racct_add_force(curproc, RACCT_WRITEIOPS, nblks);
    489 		PROC_UNLOCK(curproc);
    490 	}
    491 #endif
    492 
    493 	zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
    494 	blkid = dbuf_whichblock(dn, 0, offset);
    495 	for (i = 0; i < nblks; i++) {
    496 		dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
    497 		if (db == NULL) {
    498 			rw_exit(&dn->dn_struct_rwlock);
    499 			dmu_buf_rele_array(dbp, nblks, tag);
    500 			zio_nowait(zio);
    501 			return (SET_ERROR(EIO));
    502 		}
    503 
    504 		/* initiate async i/o */
    505 		if (read)
    506 			(void) dbuf_read(db, zio, dbuf_flags);
    507 
    508 #ifdef _KERNEL
    509 		else
    510 			curthread->td_ru.ru_oublock++;
    511 #endif
    512 		dbp[i] = &db->db;
    513 	}
    514 
    515 	if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
    516 	    DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
    517 		dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
    518 		    read && DNODE_IS_CACHEABLE(dn));
    519 	}
    520 	rw_exit(&dn->dn_struct_rwlock);
    521 
    522 	/* wait for async i/o */
    523 	err = zio_wait(zio);
    524 	if (err) {
    525 		dmu_buf_rele_array(dbp, nblks, tag);
    526 		return (err);
    527 	}
    528 
    529 	/* wait for other io to complete */
    530 	if (read) {
    531 		for (i = 0; i < nblks; i++) {
    532 			dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
    533 			mutex_enter(&db->db_mtx);
    534 			while (db->db_state == DB_READ ||
    535 			    db->db_state == DB_FILL)
    536 				cv_wait(&db->db_changed, &db->db_mtx);
    537 			if (db->db_state == DB_UNCACHED)
    538 				err = SET_ERROR(EIO);
    539 			mutex_exit(&db->db_mtx);
    540 			if (err) {
    541 				dmu_buf_rele_array(dbp, nblks, tag);
    542 				return (err);
    543 			}
    544 		}
    545 	}
    546 
    547 	*numbufsp = nblks;
    548 	*dbpp = dbp;
    549 	return (0);
    550 }
    551 
    552 static int
    553 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
    554     uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
    555 {
    556 	dnode_t *dn;
    557 	int err;
    558 
    559 	err = dnode_hold(os, object, FTAG, &dn);
    560 	if (err)
    561 		return (err);
    562 
    563 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
    564 	    numbufsp, dbpp, DMU_READ_PREFETCH);
    565 
    566 	dnode_rele(dn, FTAG);
    567 
    568 	return (err);
    569 }
    570 
    571 int
    572 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
    573     uint64_t length, boolean_t read, void *tag, int *numbufsp,
    574     dmu_buf_t ***dbpp)
    575 {
    576 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
    577 	dnode_t *dn;
    578 	int err;
    579 
    580 	DB_DNODE_ENTER(db);
    581 	dn = DB_DNODE(db);
    582 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
    583 	    numbufsp, dbpp, DMU_READ_PREFETCH);
    584 	DB_DNODE_EXIT(db);
    585 
    586 	return (err);
    587 }
    588 
    589 void
    590 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
    591 {
    592 	int i;
    593 	dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
    594 
    595 	if (numbufs == 0)
    596 		return;
    597 
    598 	for (i = 0; i < numbufs; i++) {
    599 		if (dbp[i])
    600 			dbuf_rele(dbp[i], tag);
    601 	}
    602 
    603 	kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
    604 }
    605 
    606 /*
    607  * Issue prefetch i/os for the given blocks.  If level is greater than 0, the
    608  * indirect blocks prefeteched will be those that point to the blocks containing
    609  * the data starting at offset, and continuing to offset + len.
    610  *
    611  * Note that if the indirect blocks above the blocks being prefetched are not in
    612  * cache, they will be asychronously read in.
    613  */
    614 void
    615 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
    616     uint64_t len, zio_priority_t pri)
    617 {
    618 	dnode_t *dn;
    619 	uint64_t blkid;
    620 	int nblks, err;
    621 
    622 	if (len == 0) {  /* they're interested in the bonus buffer */
    623 		dn = DMU_META_DNODE(os);
    624 
    625 		if (object == 0 || object >= DN_MAX_OBJECT)
    626 			return;
    627 
    628 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
    629 		blkid = dbuf_whichblock(dn, level,
    630 		    object * sizeof (dnode_phys_t));
    631 		dbuf_prefetch(dn, level, blkid, pri, 0);
    632 		rw_exit(&dn->dn_struct_rwlock);
    633 		return;
    634 	}
    635 
    636 	/*
    637 	 * XXX - Note, if the dnode for the requested object is not
    638 	 * already cached, we will do a *synchronous* read in the
    639 	 * dnode_hold() call.  The same is true for any indirects.
    640 	 */
    641 	err = dnode_hold(os, object, FTAG, &dn);
    642 	if (err != 0)
    643 		return;
    644 
    645 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
    646 	/*
    647 	 * offset + len - 1 is the last byte we want to prefetch for, and offset
    648 	 * is the first.  Then dbuf_whichblk(dn, level, off + len - 1) is the
    649 	 * last block we want to prefetch, and dbuf_whichblock(dn, level,
    650 	 * offset)  is the first.  Then the number we need to prefetch is the
    651 	 * last - first + 1.
    652 	 */
    653 	if (level > 0 || dn->dn_datablkshift != 0) {
    654 		nblks = dbuf_whichblock(dn, level, offset + len - 1) -
    655 		    dbuf_whichblock(dn, level, offset) + 1;
    656 	} else {
    657 		nblks = (offset < dn->dn_datablksz);
    658 	}
    659 
    660 	if (nblks != 0) {
    661 		blkid = dbuf_whichblock(dn, level, offset);
    662 		for (int i = 0; i < nblks; i++)
    663 			dbuf_prefetch(dn, level, blkid + i, pri, 0);
    664 	}
    665 
    666 	rw_exit(&dn->dn_struct_rwlock);
    667 
    668 	dnode_rele(dn, FTAG);
    669 }
    670 
    671 /*
    672  * Get the next "chunk" of file data to free.  We traverse the file from
    673  * the end so that the file gets shorter over time (if we crashes in the
    674  * middle, this will leave us in a better state).  We find allocated file
    675  * data by simply searching the allocated level 1 indirects.
    676  *
    677  * On input, *start should be the first offset that does not need to be
    678  * freed (e.g. "offset + length").  On return, *start will be the first
    679  * offset that should be freed.
    680  */
    681 static int
    682 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
    683 {
    684 	uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
    685 	/* bytes of data covered by a level-1 indirect block */
    686 	uint64_t iblkrange =
    687 	    dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
    688 
    689 	ASSERT3U(minimum, <=, *start);
    690 
    691 	if (*start - minimum <= iblkrange * maxblks) {
    692 		*start = minimum;
    693 		return (0);
    694 	}
    695 	ASSERT(ISP2(iblkrange));
    696 
    697 	for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
    698 		int err;
    699 
    700 		/*
    701 		 * dnode_next_offset(BACKWARDS) will find an allocated L1
    702 		 * indirect block at or before the input offset.  We must
    703 		 * decrement *start so that it is at the end of the region
    704 		 * to search.
    705 		 */
    706 		(*start)--;
    707 		err = dnode_next_offset(dn,
    708 		    DNODE_FIND_BACKWARDS, start, 2, 1, 0);
    709 
    710 		/* if there are no indirect blocks before start, we are done */
    711 		if (err == ESRCH) {
    712 			*start = minimum;
    713 			break;
    714 		} else if (err != 0) {
    715 			return (err);
    716 		}
    717 
    718 		/* set start to the beginning of this L1 indirect */
    719 		*start = P2ALIGN(*start, iblkrange);
    720 	}
    721 	if (*start < minimum)
    722 		*start = minimum;
    723 	return (0);
    724 }
    725 
    726 static int
    727 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
    728     uint64_t length)
    729 {
    730 	uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
    731 	int err;
    732 	uint64_t dirty_frees_threshold;
    733 	dsl_pool_t *dp = dmu_objset_pool(os);
    734 
    735 	if (offset >= object_size)
    736 		return (0);
    737 
    738 	if (zfs_per_txg_dirty_frees_percent <= 100)
    739 		dirty_frees_threshold =
    740 		    zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
    741 	else
    742 		dirty_frees_threshold = zfs_dirty_data_max / 4;
    743 
    744 	if (length == DMU_OBJECT_END || offset + length > object_size)
    745 		length = object_size - offset;
    746 
    747 	while (length != 0) {
    748 		uint64_t chunk_end, chunk_begin, chunk_len;
    749 		uint64_t long_free_dirty_all_txgs = 0;
    750 		dmu_tx_t *tx;
    751 
    752 		chunk_end = chunk_begin = offset + length;
    753 
    754 		/* move chunk_begin backwards to the beginning of this chunk */
    755 		err = get_next_chunk(dn, &chunk_begin, offset);
    756 		if (err)
    757 			return (err);
    758 		ASSERT3U(chunk_begin, >=, offset);
    759 		ASSERT3U(chunk_begin, <=, chunk_end);
    760 
    761 		chunk_len = chunk_end - chunk_begin;
    762 
    763 		mutex_enter(&dp->dp_lock);
    764 		for (int t = 0; t < TXG_SIZE; t++) {
    765 			long_free_dirty_all_txgs +=
    766 			    dp->dp_long_free_dirty_pertxg[t];
    767 		}
    768 		mutex_exit(&dp->dp_lock);
    769 
    770 		/*
    771 		 * To avoid filling up a TXG with just frees wait for
    772 		 * the next TXG to open before freeing more chunks if
    773 		 * we have reached the threshold of frees
    774 		 */
    775 		if (dirty_frees_threshold != 0 &&
    776 		    long_free_dirty_all_txgs >= dirty_frees_threshold) {
    777 			txg_wait_open(dp, 0);
    778 			continue;
    779 		}
    780 
    781 		tx = dmu_tx_create(os);
    782 		dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
    783 
    784 		/*
    785 		 * Mark this transaction as typically resulting in a net
    786 		 * reduction in space used.
    787 		 */
    788 		dmu_tx_mark_netfree(tx);
    789 		err = dmu_tx_assign(tx, TXG_WAIT);
    790 		if (err) {
    791 			dmu_tx_abort(tx);
    792 			return (err);
    793 		}
    794 
    795 		mutex_enter(&dp->dp_lock);
    796 		dp->dp_long_free_dirty_pertxg[dmu_tx_get_txg(tx) & TXG_MASK] +=
    797 		    chunk_len;
    798 		mutex_exit(&dp->dp_lock);
    799 		DTRACE_PROBE3(free__long__range,
    800 		    uint64_t, long_free_dirty_all_txgs, uint64_t, chunk_len,
    801 		    uint64_t, dmu_tx_get_txg(tx));
    802 		dnode_free_range(dn, chunk_begin, chunk_len, tx);
    803 		dmu_tx_commit(tx);
    804 
    805 		length -= chunk_len;
    806 	}
    807 	return (0);
    808 }
    809 
    810 int
    811 dmu_free_long_range(objset_t *os, uint64_t object,
    812     uint64_t offset, uint64_t length)
    813 {
    814 	dnode_t *dn;
    815 	int err;
    816 
    817 	err = dnode_hold(os, object, FTAG, &dn);
    818 	if (err != 0)
    819 		return (err);
    820 	err = dmu_free_long_range_impl(os, dn, offset, length);
    821 
    822 	/*
    823 	 * It is important to zero out the maxblkid when freeing the entire
    824 	 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
    825 	 * will take the fast path, and (b) dnode_reallocate() can verify
    826 	 * that the entire file has been freed.
    827 	 */
    828 	if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
    829 		dn->dn_maxblkid = 0;
    830 
    831 	dnode_rele(dn, FTAG);
    832 	return (err);
    833 }
    834 
    835 int
    836 dmu_free_long_object(objset_t *os, uint64_t object)
    837 {
    838 	dmu_tx_t *tx;
    839 	int err;
    840 
    841 	err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
    842 	if (err != 0)
    843 		return (err);
    844 
    845 	tx = dmu_tx_create(os);
    846 	dmu_tx_hold_bonus(tx, object);
    847 	dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
    848 	dmu_tx_mark_netfree(tx);
    849 	err = dmu_tx_assign(tx, TXG_WAIT);
    850 	if (err == 0) {
    851 		err = dmu_object_free(os, object, tx);
    852 		dmu_tx_commit(tx);
    853 	} else {
    854 		dmu_tx_abort(tx);
    855 	}
    856 
    857 	return (err);
    858 }
    859 
    860 int
    861 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
    862     uint64_t size, dmu_tx_t *tx)
    863 {
    864 	dnode_t *dn;
    865 	int err = dnode_hold(os, object, FTAG, &dn);
    866 	if (err)
    867 		return (err);
    868 	ASSERT(offset < UINT64_MAX);
    869 	ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
    870 	dnode_free_range(dn, offset, size, tx);
    871 	dnode_rele(dn, FTAG);
    872 	return (0);
    873 }
    874 
    875 int
    876 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
    877     void *buf, uint32_t flags)
    878 {
    879 	dnode_t *dn;
    880 	dmu_buf_t **dbp;
    881 	int numbufs, err;
    882 
    883 	err = dnode_hold(os, object, FTAG, &dn);
    884 	if (err)
    885 		return (err);
    886 
    887 	/*
    888 	 * Deal with odd block sizes, where there can't be data past the first
    889 	 * block.  If we ever do the tail block optimization, we will need to
    890 	 * handle that here as well.
    891 	 */
    892 	if (dn->dn_maxblkid == 0) {
    893 		int newsz = offset > dn->dn_datablksz ? 0 :
    894 		    MIN(size, dn->dn_datablksz - offset);
    895 		bzero((char *)buf + newsz, size - newsz);
    896 		size = newsz;
    897 	}
    898 
    899 	while (size > 0) {
    900 		uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
    901 		int i;
    902 
    903 		/*
    904 		 * NB: we could do this block-at-a-time, but it's nice
    905 		 * to be reading in parallel.
    906 		 */
    907 		err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
    908 		    TRUE, FTAG, &numbufs, &dbp, flags);
    909 		if (err)
    910 			break;
    911 
    912 		for (i = 0; i < numbufs; i++) {
    913 			int tocpy;
    914 			int bufoff;
    915 			dmu_buf_t *db = dbp[i];
    916 
    917 			ASSERT(size > 0);
    918 
    919 			bufoff = offset - db->db_offset;
    920 			tocpy = (int)MIN(db->db_size - bufoff, size);
    921 
    922 			bcopy((char *)db->db_data + bufoff, buf, tocpy);
    923 
    924 			offset += tocpy;
    925 			size -= tocpy;
    926 			buf = (char *)buf + tocpy;
    927 		}
    928 		dmu_buf_rele_array(dbp, numbufs, FTAG);
    929 	}
    930 	dnode_rele(dn, FTAG);
    931 	return (err);
    932 }
    933 
    934 void
    935 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
    936     const void *buf, dmu_tx_t *tx)
    937 {
    938 	dmu_buf_t **dbp;
    939 	int numbufs, i;
    940 
    941 	if (size == 0)
    942 		return;
    943 
    944 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
    945 	    FALSE, FTAG, &numbufs, &dbp));
    946 
    947 	for (i = 0; i < numbufs; i++) {
    948 		int tocpy;
    949 		int bufoff;
    950 		dmu_buf_t *db = dbp[i];
    951 
    952 		ASSERT(size > 0);
    953 
    954 		bufoff = offset - db->db_offset;
    955 		tocpy = (int)MIN(db->db_size - bufoff, size);
    956 
    957 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
    958 
    959 		if (tocpy == db->db_size)
    960 			dmu_buf_will_fill(db, tx);
    961 		else
    962 			dmu_buf_will_dirty(db, tx);
    963 
    964 		bcopy(buf, (char *)db->db_data + bufoff, tocpy);
    965 
    966 		if (tocpy == db->db_size)
    967 			dmu_buf_fill_done(db, tx);
    968 
    969 		offset += tocpy;
    970 		size -= tocpy;
    971 		buf = (char *)buf + tocpy;
    972 	}
    973 	dmu_buf_rele_array(dbp, numbufs, FTAG);
    974 }
    975 
    976 void
    977 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
    978     dmu_tx_t *tx)
    979 {
    980 	dmu_buf_t **dbp;
    981 	int numbufs, i;
    982 
    983 	if (size == 0)
    984 		return;
    985 
    986 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
    987 	    FALSE, FTAG, &numbufs, &dbp));
    988 
    989 	for (i = 0; i < numbufs; i++) {
    990 		dmu_buf_t *db = dbp[i];
    991 
    992 		dmu_buf_will_not_fill(db, tx);
    993 	}
    994 	dmu_buf_rele_array(dbp, numbufs, FTAG);
    995 }
    996 
    997 void
    998 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
    999     void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
   1000     int compressed_size, int byteorder, dmu_tx_t *tx)
   1001 {
   1002 	dmu_buf_t *db;
   1003 
   1004 	ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
   1005 	ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
   1006 	VERIFY0(dmu_buf_hold_noread(os, object, offset,
   1007 	    FTAG, &db));
   1008 
   1009 	dmu_buf_write_embedded(db,
   1010 	    data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
   1011 	    uncompressed_size, compressed_size, byteorder, tx);
   1012 
   1013 	dmu_buf_rele(db, FTAG);
   1014 }
   1015 
   1016 /*
   1017  * DMU support for xuio
   1018  */
   1019 kstat_t *xuio_ksp = NULL;
   1020 
   1021 int
   1022 dmu_xuio_init(xuio_t *xuio, int nblk)
   1023 {
   1024 	dmu_xuio_t *priv;
   1025 	uio_t *uio = &xuio->xu_uio;
   1026 
   1027 	uio->uio_iovcnt = nblk;
   1028 	uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
   1029 
   1030 	priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
   1031 	priv->cnt = nblk;
   1032 	priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
   1033 	priv->iovp = uio->uio_iov;
   1034 	XUIO_XUZC_PRIV(xuio) = priv;
   1035 
   1036 	if (XUIO_XUZC_RW(xuio) == UIO_READ)
   1037 		XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
   1038 	else
   1039 		XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
   1040 
   1041 	return (0);
   1042 }
   1043 
   1044 void
   1045 dmu_xuio_fini(xuio_t *xuio)
   1046 {
   1047 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
   1048 	int nblk = priv->cnt;
   1049 
   1050 	kmem_free(priv->iovp, nblk * sizeof (iovec_t));
   1051 	kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
   1052 	kmem_free(priv, sizeof (dmu_xuio_t));
   1053 
   1054 	if (XUIO_XUZC_RW(xuio) == UIO_READ)
   1055 		XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
   1056 	else
   1057 		XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
   1058 }
   1059 
   1060 /*
   1061  * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
   1062  * and increase priv->next by 1.
   1063  */
   1064 int
   1065 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
   1066 {
   1067 	struct iovec *iov;
   1068 	uio_t *uio = &xuio->xu_uio;
   1069 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
   1070 	int i = priv->next++;
   1071 
   1072 	ASSERT(i < priv->cnt);
   1073 	ASSERT(off + n <= arc_buf_size(abuf));
   1074 	iov = uio->uio_iov + i;
   1075 	iov->iov_base = (char *)abuf->b_data + off;
   1076 	iov->iov_len = n;
   1077 	priv->bufs[i] = abuf;
   1078 	return (0);
   1079 }
   1080 
   1081 int
   1082 dmu_xuio_cnt(xuio_t *xuio)
   1083 {
   1084 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
   1085 	return (priv->cnt);
   1086 }
   1087 
   1088 arc_buf_t *
   1089 dmu_xuio_arcbuf(xuio_t *xuio, int i)
   1090 {
   1091 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
   1092 
   1093 	ASSERT(i < priv->cnt);
   1094 	return (priv->bufs[i]);
   1095 }
   1096 
   1097 void
   1098 dmu_xuio_clear(xuio_t *xuio, int i)
   1099 {
   1100 	dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
   1101 
   1102 	ASSERT(i < priv->cnt);
   1103 	priv->bufs[i] = NULL;
   1104 }
   1105 
   1106 static void
   1107 xuio_stat_init(void)
   1108 {
   1109 	xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
   1110 	    KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
   1111 	    KSTAT_FLAG_VIRTUAL);
   1112 	if (xuio_ksp != NULL) {
   1113 		xuio_ksp->ks_data = &xuio_stats;
   1114 		kstat_install(xuio_ksp);
   1115 	}
   1116 }
   1117 
   1118 static void
   1119 xuio_stat_fini(void)
   1120 {
   1121 	if (xuio_ksp != NULL) {
   1122 		kstat_delete(xuio_ksp);
   1123 		xuio_ksp = NULL;
   1124 	}
   1125 }
   1126 
   1127 void
   1128 xuio_stat_wbuf_copied()
   1129 {
   1130 	XUIOSTAT_BUMP(xuiostat_wbuf_copied);
   1131 }
   1132 
   1133 void
   1134 xuio_stat_wbuf_nocopy()
   1135 {
   1136 	XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
   1137 }
   1138 
   1139 #ifdef _KERNEL
   1140 static int
   1141 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
   1142 {
   1143 	dmu_buf_t **dbp;
   1144 	int numbufs, i, err;
   1145 	xuio_t *xuio = NULL;
   1146 
   1147 	/*
   1148 	 * NB: we could do this block-at-a-time, but it's nice
   1149 	 * to be reading in parallel.
   1150 	 */
   1151 	err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
   1152 	    TRUE, FTAG, &numbufs, &dbp, 0);
   1153 	if (err)
   1154 		return (err);
   1155 
   1156 #ifdef UIO_XUIO
   1157 	if (uio->uio_extflg == UIO_XUIO)
   1158 		xuio = (xuio_t *)uio;
   1159 #endif
   1160 
   1161 	for (i = 0; i < numbufs; i++) {
   1162 		int tocpy;
   1163 		int bufoff;
   1164 		dmu_buf_t *db = dbp[i];
   1165 
   1166 		ASSERT(size > 0);
   1167 
   1168 		bufoff = uio->uio_loffset - db->db_offset;
   1169 		tocpy = (int)MIN(db->db_size - bufoff, size);
   1170 
   1171 		if (xuio) {
   1172 			dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
   1173 			arc_buf_t *dbuf_abuf = dbi->db_buf;
   1174 			arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
   1175 			err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
   1176 			if (!err) {
   1177 				uio->uio_resid -= tocpy;
   1178 				uio->uio_loffset += tocpy;
   1179 			}
   1180 
   1181 			if (abuf == dbuf_abuf)
   1182 				XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
   1183 			else
   1184 				XUIOSTAT_BUMP(xuiostat_rbuf_copied);
   1185 		} else {
   1186 #ifdef illumos
   1187 			err = uiomove((char *)db->db_data + bufoff, tocpy,
   1188 			    UIO_READ, uio);
   1189 #endif
   1190 #ifdef __FreeBSD__
   1191 			err = vn_io_fault_uiomove((char *)db->db_data + bufoff,
   1192 			    tocpy, uio);
   1193 #endif
   1194 #ifdef __NetBSD__
   1195 			err = uiomove((char *)db->db_data + bufoff, tocpy,
   1196 			    UIO_READ, uio);
   1197 #endif
   1198 		}
   1199 		if (err)
   1200 			break;
   1201 
   1202 		size -= tocpy;
   1203 	}
   1204 	dmu_buf_rele_array(dbp, numbufs, FTAG);
   1205 
   1206 	return (err);
   1207 }
   1208 
   1209 /*
   1210  * Read 'size' bytes into the uio buffer.
   1211  * From object zdb->db_object.
   1212  * Starting at offset uio->uio_loffset.
   1213  *
   1214  * If the caller already has a dbuf in the target object
   1215  * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
   1216  * because we don't have to find the dnode_t for the object.
   1217  */
   1218 int
   1219 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
   1220 {
   1221 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
   1222 	dnode_t *dn;
   1223 	int err;
   1224 
   1225 	if (size == 0)
   1226 		return (0);
   1227 
   1228 	DB_DNODE_ENTER(db);
   1229 	dn = DB_DNODE(db);
   1230 	err = dmu_read_uio_dnode(dn, uio, size);
   1231 	DB_DNODE_EXIT(db);
   1232 
   1233 	return (err);
   1234 }
   1235 
   1236 /*
   1237  * Read 'size' bytes into the uio buffer.
   1238  * From the specified object
   1239  * Starting at offset uio->uio_loffset.
   1240  */
   1241 int
   1242 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
   1243 {
   1244 	dnode_t *dn;
   1245 	int err;
   1246 
   1247 	if (size == 0)
   1248 		return (0);
   1249 
   1250 	err = dnode_hold(os, object, FTAG, &dn);
   1251 	if (err)
   1252 		return (err);
   1253 
   1254 	err = dmu_read_uio_dnode(dn, uio, size);
   1255 
   1256 	dnode_rele(dn, FTAG);
   1257 
   1258 	return (err);
   1259 }
   1260 
   1261 static int
   1262 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
   1263 {
   1264 	dmu_buf_t **dbp;
   1265 	int numbufs;
   1266 	int err = 0;
   1267 	int i;
   1268 
   1269 	err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
   1270 	    FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
   1271 	if (err)
   1272 		return (err);
   1273 
   1274 	for (i = 0; i < numbufs; i++) {
   1275 		int tocpy;
   1276 		int bufoff;
   1277 		dmu_buf_t *db = dbp[i];
   1278 
   1279 		ASSERT(size > 0);
   1280 
   1281 		bufoff = uio->uio_loffset - db->db_offset;
   1282 		tocpy = (int)MIN(db->db_size - bufoff, size);
   1283 
   1284 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
   1285 
   1286 		if (tocpy == db->db_size)
   1287 			dmu_buf_will_fill(db, tx);
   1288 		else
   1289 			dmu_buf_will_dirty(db, tx);
   1290 
   1291 #ifdef illumos
   1292 		/*
   1293 		 * XXX uiomove could block forever (eg. nfs-backed
   1294 		 * pages).  There needs to be a uiolockdown() function
   1295 		 * to lock the pages in memory, so that uiomove won't
   1296 		 * block.
   1297 		 */
   1298 		err = uiomove((char *)db->db_data + bufoff, tocpy,
   1299 		    UIO_WRITE, uio);
   1300 #endif
   1301 #ifdef __FreeBSD__
   1302 		err = vn_io_fault_uiomove((char *)db->db_data + bufoff, tocpy,
   1303 		    uio);
   1304 #endif
   1305 #ifdef __NetBSD__
   1306 		err = uiomove((char *)db->db_data + bufoff, tocpy,
   1307 		    UIO_WRITE, uio);
   1308 #endif
   1309 
   1310 		if (tocpy == db->db_size)
   1311 			dmu_buf_fill_done(db, tx);
   1312 
   1313 		if (err)
   1314 			break;
   1315 
   1316 		size -= tocpy;
   1317 	}
   1318 
   1319 	dmu_buf_rele_array(dbp, numbufs, FTAG);
   1320 	return (err);
   1321 }
   1322 
   1323 /*
   1324  * Write 'size' bytes from the uio buffer.
   1325  * To object zdb->db_object.
   1326  * Starting at offset uio->uio_loffset.
   1327  *
   1328  * If the caller already has a dbuf in the target object
   1329  * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
   1330  * because we don't have to find the dnode_t for the object.
   1331  */
   1332 int
   1333 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
   1334     dmu_tx_t *tx)
   1335 {
   1336 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
   1337 	dnode_t *dn;
   1338 	int err;
   1339 
   1340 	if (size == 0)
   1341 		return (0);
   1342 
   1343 	DB_DNODE_ENTER(db);
   1344 	dn = DB_DNODE(db);
   1345 	err = dmu_write_uio_dnode(dn, uio, size, tx);
   1346 	DB_DNODE_EXIT(db);
   1347 
   1348 	return (err);
   1349 }
   1350 
   1351 /*
   1352  * Write 'size' bytes from the uio buffer.
   1353  * To the specified object.
   1354  * Starting at offset uio->uio_loffset.
   1355  */
   1356 int
   1357 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
   1358     dmu_tx_t *tx)
   1359 {
   1360 	dnode_t *dn;
   1361 	int err;
   1362 
   1363 	if (size == 0)
   1364 		return (0);
   1365 
   1366 	err = dnode_hold(os, object, FTAG, &dn);
   1367 	if (err)
   1368 		return (err);
   1369 
   1370 	err = dmu_write_uio_dnode(dn, uio, size, tx);
   1371 
   1372 	dnode_rele(dn, FTAG);
   1373 
   1374 	return (err);
   1375 }
   1376 
   1377 #ifdef illumos
   1378 int
   1379 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
   1380     page_t *pp, dmu_tx_t *tx)
   1381 {
   1382 	dmu_buf_t **dbp;
   1383 	int numbufs, i;
   1384 	int err;
   1385 
   1386 	if (size == 0)
   1387 		return (0);
   1388 
   1389 	err = dmu_buf_hold_array(os, object, offset, size,
   1390 	    FALSE, FTAG, &numbufs, &dbp);
   1391 	if (err)
   1392 		return (err);
   1393 
   1394 	for (i = 0; i < numbufs; i++) {
   1395 		int tocpy, copied, thiscpy;
   1396 		int bufoff;
   1397 		dmu_buf_t *db = dbp[i];
   1398 		caddr_t va;
   1399 
   1400 		ASSERT(size > 0);
   1401 		ASSERT3U(db->db_size, >=, PAGESIZE);
   1402 
   1403 		bufoff = offset - db->db_offset;
   1404 		tocpy = (int)MIN(db->db_size - bufoff, size);
   1405 
   1406 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
   1407 
   1408 		if (tocpy == db->db_size)
   1409 			dmu_buf_will_fill(db, tx);
   1410 		else
   1411 			dmu_buf_will_dirty(db, tx);
   1412 
   1413 		for (copied = 0; copied < tocpy; copied += PAGESIZE) {
   1414 			ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
   1415 			thiscpy = MIN(PAGESIZE, tocpy - copied);
   1416 			va = zfs_map_page(pp, S_READ);
   1417 			bcopy(va, (char *)db->db_data + bufoff, thiscpy);
   1418 			zfs_unmap_page(pp, va);
   1419 			pp = pp->p_next;
   1420 			bufoff += PAGESIZE;
   1421 		}
   1422 
   1423 		if (tocpy == db->db_size)
   1424 			dmu_buf_fill_done(db, tx);
   1425 
   1426 		offset += tocpy;
   1427 		size -= tocpy;
   1428 	}
   1429 	dmu_buf_rele_array(dbp, numbufs, FTAG);
   1430 	return (err);
   1431 }
   1432 #endif /* illumos */
   1433 
   1434 #ifdef __FreeBSD__
   1435 int
   1436 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
   1437     vm_page_t *ma, dmu_tx_t *tx)
   1438 {
   1439 	dmu_buf_t **dbp;
   1440 	struct sf_buf *sf;
   1441 	int numbufs, i;
   1442 	int err;
   1443 
   1444 	if (size == 0)
   1445 		return (0);
   1446 
   1447 	err = dmu_buf_hold_array(os, object, offset, size,
   1448 	    FALSE, FTAG, &numbufs, &dbp);
   1449 	if (err)
   1450 		return (err);
   1451 
   1452 	for (i = 0; i < numbufs; i++) {
   1453 		int tocpy, copied, thiscpy;
   1454 		int bufoff;
   1455 		dmu_buf_t *db = dbp[i];
   1456 		caddr_t va;
   1457 
   1458 		ASSERT(size > 0);
   1459 		ASSERT3U(db->db_size, >=, PAGESIZE);
   1460 
   1461 		bufoff = offset - db->db_offset;
   1462 		tocpy = (int)MIN(db->db_size - bufoff, size);
   1463 
   1464 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
   1465 
   1466 		if (tocpy == db->db_size)
   1467 			dmu_buf_will_fill(db, tx);
   1468 		else
   1469 			dmu_buf_will_dirty(db, tx);
   1470 
   1471 		for (copied = 0; copied < tocpy; copied += PAGESIZE) {
   1472 			ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff);
   1473 			thiscpy = MIN(PAGESIZE, tocpy - copied);
   1474 			va = zfs_map_page(*ma, &sf);
   1475 			bcopy(va, (char *)db->db_data + bufoff, thiscpy);
   1476 			zfs_unmap_page(sf);
   1477 			ma += 1;
   1478 			bufoff += PAGESIZE;
   1479 		}
   1480 
   1481 		if (tocpy == db->db_size)
   1482 			dmu_buf_fill_done(db, tx);
   1483 
   1484 		offset += tocpy;
   1485 		size -= tocpy;
   1486 	}
   1487 	dmu_buf_rele_array(dbp, numbufs, FTAG);
   1488 	return (err);
   1489 }
   1490 #endif	/* __FreeBSD__ */
   1491 
   1492 #ifdef __NetBSD__
   1493 int
   1494 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
   1495     struct vm_page **pgs, dmu_tx_t *tx)
   1496 {
   1497 	dmu_buf_t **dbp;
   1498 	int numbufs, i;
   1499 	int err;
   1500 
   1501 	if (size == 0)
   1502 		return (0);
   1503 
   1504 	err = dmu_buf_hold_array(os, object, offset, size,
   1505 	    FALSE, FTAG, &numbufs, &dbp);
   1506 	if (err)
   1507 		return (err);
   1508 
   1509 	for (i = 0; i < numbufs; i++) {
   1510 		int tocpy, copied, thiscpy;
   1511 		int bufoff;
   1512 		dmu_buf_t *db = dbp[i];
   1513 		caddr_t va;
   1514 
   1515 		ASSERT(size > 0);
   1516 		ASSERT3U(db->db_size, >=, PAGESIZE);
   1517 
   1518 		bufoff = offset - db->db_offset;
   1519 		tocpy = (int)MIN(db->db_size - bufoff, size);
   1520 
   1521 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
   1522 
   1523 		if (tocpy == db->db_size)
   1524 			dmu_buf_will_fill(db, tx);
   1525 		else
   1526 			dmu_buf_will_dirty(db, tx);
   1527 
   1528 		for (copied = 0; copied < tocpy; copied += PAGESIZE) {
   1529 			ASSERT3U((*pgs)->offset, ==, db->db_offset + bufoff);
   1530 			thiscpy = MIN(PAGESIZE, tocpy - copied);
   1531 			va = zfs_map_page(*pgs, S_READ);
   1532 			bcopy(va, (char *)db->db_data + bufoff, thiscpy);
   1533 			zfs_unmap_page(*pgs, va);
   1534 			pgs++;
   1535 			bufoff += PAGESIZE;
   1536 		}
   1537 
   1538 		if (tocpy == db->db_size)
   1539 			dmu_buf_fill_done(db, tx);
   1540 
   1541 		offset += tocpy;
   1542 		size -= tocpy;
   1543 	}
   1544 	dmu_buf_rele_array(dbp, numbufs, FTAG);
   1545 	return (err);
   1546 }
   1547 #endif
   1548 #endif	/* _KERNEL */
   1549 
   1550 /*
   1551  * Allocate a loaned anonymous arc buffer.
   1552  */
   1553 arc_buf_t *
   1554 dmu_request_arcbuf(dmu_buf_t *handle, int size)
   1555 {
   1556 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
   1557 
   1558 	return (arc_loan_buf(db->db_objset->os_spa, size));
   1559 }
   1560 
   1561 /*
   1562  * Free a loaned arc buffer.
   1563  */
   1564 void
   1565 dmu_return_arcbuf(arc_buf_t *buf)
   1566 {
   1567 	arc_return_buf(buf, FTAG);
   1568 	arc_buf_destroy(buf, FTAG);
   1569 }
   1570 
   1571 /*
   1572  * When possible directly assign passed loaned arc buffer to a dbuf.
   1573  * If this is not possible copy the contents of passed arc buf via
   1574  * dmu_write().
   1575  */
   1576 void
   1577 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
   1578     dmu_tx_t *tx)
   1579 {
   1580 	dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
   1581 	dnode_t *dn;
   1582 	dmu_buf_impl_t *db;
   1583 	uint32_t blksz = (uint32_t)arc_buf_size(buf);
   1584 	uint64_t blkid;
   1585 
   1586 	DB_DNODE_ENTER(dbuf);
   1587 	dn = DB_DNODE(dbuf);
   1588 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
   1589 	blkid = dbuf_whichblock(dn, 0, offset);
   1590 	VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
   1591 	rw_exit(&dn->dn_struct_rwlock);
   1592 	DB_DNODE_EXIT(dbuf);
   1593 
   1594 	/*
   1595 	 * We can only assign if the offset is aligned, the arc buf is the
   1596 	 * same size as the dbuf, and the dbuf is not metadata.  It
   1597 	 * can't be metadata because the loaned arc buf comes from the
   1598 	 * user-data kmem arena.
   1599 	 */
   1600 	if (offset == db->db.db_offset && blksz == db->db.db_size &&
   1601 	    DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) {
   1602 #ifdef _KERNEL
   1603 		curthread->td_ru.ru_oublock++;
   1604 #ifdef RACCT
   1605 		if (racct_enable) {
   1606 			PROC_LOCK(curproc);
   1607 			racct_add_force(curproc, RACCT_WRITEBPS, blksz);
   1608 			racct_add_force(curproc, RACCT_WRITEIOPS, 1);
   1609 			PROC_UNLOCK(curproc);
   1610 		}
   1611 #endif /* RACCT */
   1612 #endif /* _KERNEL */
   1613 		dbuf_assign_arcbuf(db, buf, tx);
   1614 		dbuf_rele(db, FTAG);
   1615 	} else {
   1616 		objset_t *os;
   1617 		uint64_t object;
   1618 
   1619 		DB_DNODE_ENTER(dbuf);
   1620 		dn = DB_DNODE(dbuf);
   1621 		os = dn->dn_objset;
   1622 		object = dn->dn_object;
   1623 		DB_DNODE_EXIT(dbuf);
   1624 
   1625 		dbuf_rele(db, FTAG);
   1626 		dmu_write(os, object, offset, blksz, buf->b_data, tx);
   1627 		dmu_return_arcbuf(buf);
   1628 		XUIOSTAT_BUMP(xuiostat_wbuf_copied);
   1629 	}
   1630 }
   1631 
   1632 typedef struct {
   1633 	dbuf_dirty_record_t	*dsa_dr;
   1634 	dmu_sync_cb_t		*dsa_done;
   1635 	zgd_t			*dsa_zgd;
   1636 	dmu_tx_t		*dsa_tx;
   1637 } dmu_sync_arg_t;
   1638 
   1639 /* ARGSUSED */
   1640 static void
   1641 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
   1642 {
   1643 	dmu_sync_arg_t *dsa = varg;
   1644 	dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
   1645 	blkptr_t *bp = zio->io_bp;
   1646 
   1647 	if (zio->io_error == 0) {
   1648 		if (BP_IS_HOLE(bp)) {
   1649 			/*
   1650 			 * A block of zeros may compress to a hole, but the
   1651 			 * block size still needs to be known for replay.
   1652 			 */
   1653 			BP_SET_LSIZE(bp, db->db_size);
   1654 		} else if (!BP_IS_EMBEDDED(bp)) {
   1655 			ASSERT(BP_GET_LEVEL(bp) == 0);
   1656 			bp->blk_fill = 1;
   1657 		}
   1658 	}
   1659 }
   1660 
   1661 static void
   1662 dmu_sync_late_arrival_ready(zio_t *zio)
   1663 {
   1664 	dmu_sync_ready(zio, NULL, zio->io_private);
   1665 }
   1666 
   1667 /* ARGSUSED */
   1668 static void
   1669 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
   1670 {
   1671 	dmu_sync_arg_t *dsa = varg;
   1672 	dbuf_dirty_record_t *dr = dsa->dsa_dr;
   1673 	dmu_buf_impl_t *db = dr->dr_dbuf;
   1674 
   1675 	mutex_enter(&db->db_mtx);
   1676 	ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
   1677 	if (zio->io_error == 0) {
   1678 		dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
   1679 		if (dr->dt.dl.dr_nopwrite) {
   1680 			blkptr_t *bp = zio->io_bp;
   1681 			blkptr_t *bp_orig = &zio->io_bp_orig;
   1682 			uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
   1683 
   1684 			ASSERT(BP_EQUAL(bp, bp_orig));
   1685 			ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
   1686 			ASSERT(zio_checksum_table[chksum].ci_flags &
   1687 			    ZCHECKSUM_FLAG_NOPWRITE);
   1688 		}
   1689 		dr->dt.dl.dr_overridden_by = *zio->io_bp;
   1690 		dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
   1691 		dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
   1692 
   1693 		/*
   1694 		 * Old style holes are filled with all zeros, whereas
   1695 		 * new-style holes maintain their lsize, type, level,
   1696 		 * and birth time (see zio_write_compress). While we
   1697 		 * need to reset the BP_SET_LSIZE() call that happened
   1698 		 * in dmu_sync_ready for old style holes, we do *not*
   1699 		 * want to wipe out the information contained in new
   1700 		 * style holes. Thus, only zero out the block pointer if
   1701 		 * it's an old style hole.
   1702 		 */
   1703 		if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
   1704 		    dr->dt.dl.dr_overridden_by.blk_birth == 0)
   1705 			BP_ZERO(&dr->dt.dl.dr_overridden_by);
   1706 	} else {
   1707 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
   1708 	}
   1709 	cv_broadcast(&db->db_changed);
   1710 	mutex_exit(&db->db_mtx);
   1711 
   1712 	dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
   1713 
   1714 	kmem_free(dsa, sizeof (*dsa));
   1715 }
   1716 
   1717 static void
   1718 dmu_sync_late_arrival_done(zio_t *zio)
   1719 {
   1720 	blkptr_t *bp = zio->io_bp;
   1721 	dmu_sync_arg_t *dsa = zio->io_private;
   1722 	blkptr_t *bp_orig = &zio->io_bp_orig;
   1723 
   1724 	if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
   1725 		/*
   1726 		 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
   1727 		 * then there is nothing to do here. Otherwise, free the
   1728 		 * newly allocated block in this txg.
   1729 		 */
   1730 		if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
   1731 			ASSERT(BP_EQUAL(bp, bp_orig));
   1732 		} else {
   1733 			ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
   1734 			ASSERT(zio->io_bp->blk_birth == zio->io_txg);
   1735 			ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
   1736 			zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
   1737 		}
   1738 	}
   1739 
   1740 	dmu_tx_commit(dsa->dsa_tx);
   1741 
   1742 	dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
   1743 
   1744 	kmem_free(dsa, sizeof (*dsa));
   1745 }
   1746 
   1747 static int
   1748 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
   1749     zio_prop_t *zp, zbookmark_phys_t *zb)
   1750 {
   1751 	dmu_sync_arg_t *dsa;
   1752 	dmu_tx_t *tx;
   1753 
   1754 	tx = dmu_tx_create(os);
   1755 	dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
   1756 	if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
   1757 		dmu_tx_abort(tx);
   1758 		/* Make zl_get_data do txg_waited_synced() */
   1759 		return (SET_ERROR(EIO));
   1760 	}
   1761 
   1762 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
   1763 	dsa->dsa_dr = NULL;
   1764 	dsa->dsa_done = done;
   1765 	dsa->dsa_zgd = zgd;
   1766 	dsa->dsa_tx = tx;
   1767 
   1768 	zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx),
   1769 	    zgd->zgd_bp, zgd->zgd_db->db_data, zgd->zgd_db->db_size,
   1770 	    zp, dmu_sync_late_arrival_ready, NULL,
   1771 	    NULL, dmu_sync_late_arrival_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
   1772 	    ZIO_FLAG_CANFAIL, zb));
   1773 
   1774 	return (0);
   1775 }
   1776 
   1777 /*
   1778  * Intent log support: sync the block associated with db to disk.
   1779  * N.B. and XXX: the caller is responsible for making sure that the
   1780  * data isn't changing while dmu_sync() is writing it.
   1781  *
   1782  * Return values:
   1783  *
   1784  *	EEXIST: this txg has already been synced, so there's nothing to do.
   1785  *		The caller should not log the write.
   1786  *
   1787  *	ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
   1788  *		The caller should not log the write.
   1789  *
   1790  *	EALREADY: this block is already in the process of being synced.
   1791  *		The caller should track its progress (somehow).
   1792  *
   1793  *	EIO: could not do the I/O.
   1794  *		The caller should do a txg_wait_synced().
   1795  *
   1796  *	0: the I/O has been initiated.
   1797  *		The caller should log this blkptr in the done callback.
   1798  *		It is possible that the I/O will fail, in which case
   1799  *		the error will be reported to the done callback and
   1800  *		propagated to pio from zio_done().
   1801  */
   1802 int
   1803 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
   1804 {
   1805 	blkptr_t *bp = zgd->zgd_bp;
   1806 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
   1807 	objset_t *os = db->db_objset;
   1808 	dsl_dataset_t *ds = os->os_dsl_dataset;
   1809 	dbuf_dirty_record_t *dr;
   1810 	dmu_sync_arg_t *dsa;
   1811 	zbookmark_phys_t zb;
   1812 	zio_prop_t zp;
   1813 	dnode_t *dn;
   1814 
   1815 	ASSERT(pio != NULL);
   1816 	ASSERT(txg != 0);
   1817 
   1818 	SET_BOOKMARK(&zb, ds->ds_object,
   1819 	    db->db.db_object, db->db_level, db->db_blkid);
   1820 
   1821 	DB_DNODE_ENTER(db);
   1822 	dn = DB_DNODE(db);
   1823 	dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
   1824 	DB_DNODE_EXIT(db);
   1825 
   1826 	/*
   1827 	 * If we're frozen (running ziltest), we always need to generate a bp.
   1828 	 */
   1829 	if (txg > spa_freeze_txg(os->os_spa))
   1830 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
   1831 
   1832 	/*
   1833 	 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
   1834 	 * and us.  If we determine that this txg is not yet syncing,
   1835 	 * but it begins to sync a moment later, that's OK because the
   1836 	 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
   1837 	 */
   1838 	mutex_enter(&db->db_mtx);
   1839 
   1840 	if (txg <= spa_last_synced_txg(os->os_spa)) {
   1841 		/*
   1842 		 * This txg has already synced.  There's nothing to do.
   1843 		 */
   1844 		mutex_exit(&db->db_mtx);
   1845 		return (SET_ERROR(EEXIST));
   1846 	}
   1847 
   1848 	if (txg <= spa_syncing_txg(os->os_spa)) {
   1849 		/*
   1850 		 * This txg is currently syncing, so we can't mess with
   1851 		 * the dirty record anymore; just write a new log block.
   1852 		 */
   1853 		mutex_exit(&db->db_mtx);
   1854 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
   1855 	}
   1856 
   1857 	dr = db->db_last_dirty;
   1858 	while (dr && dr->dr_txg != txg)
   1859 		dr = dr->dr_next;
   1860 
   1861 	if (dr == NULL) {
   1862 		/*
   1863 		 * There's no dr for this dbuf, so it must have been freed.
   1864 		 * There's no need to log writes to freed blocks, so we're done.
   1865 		 */
   1866 		mutex_exit(&db->db_mtx);
   1867 		return (SET_ERROR(ENOENT));
   1868 	}
   1869 
   1870 	ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
   1871 
   1872 	/*
   1873 	 * Assume the on-disk data is X, the current syncing data (in
   1874 	 * txg - 1) is Y, and the current in-memory data is Z (currently
   1875 	 * in dmu_sync).
   1876 	 *
   1877 	 * We usually want to perform a nopwrite if X and Z are the
   1878 	 * same.  However, if Y is different (i.e. the BP is going to
   1879 	 * change before this write takes effect), then a nopwrite will
   1880 	 * be incorrect - we would override with X, which could have
   1881 	 * been freed when Y was written.
   1882 	 *
   1883 	 * (Note that this is not a concern when we are nop-writing from
   1884 	 * syncing context, because X and Y must be identical, because
   1885 	 * all previous txgs have been synced.)
   1886 	 *
   1887 	 * Therefore, we disable nopwrite if the current BP could change
   1888 	 * before this TXG.  There are two ways it could change: by
   1889 	 * being dirty (dr_next is non-NULL), or by being freed
   1890 	 * (dnode_block_freed()).  This behavior is verified by
   1891 	 * zio_done(), which VERIFYs that the override BP is identical
   1892 	 * to the on-disk BP.
   1893 	 */
   1894 	DB_DNODE_ENTER(db);
   1895 	dn = DB_DNODE(db);
   1896 	if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
   1897 		zp.zp_nopwrite = B_FALSE;
   1898 	DB_DNODE_EXIT(db);
   1899 
   1900 	ASSERT(dr->dr_txg == txg);
   1901 	if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
   1902 	    dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
   1903 		/*
   1904 		 * We have already issued a sync write for this buffer,
   1905 		 * or this buffer has already been synced.  It could not
   1906 		 * have been dirtied since, or we would have cleared the state.
   1907 		 */
   1908 		mutex_exit(&db->db_mtx);
   1909 		return (SET_ERROR(EALREADY));
   1910 	}
   1911 
   1912 	ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
   1913 	dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
   1914 	mutex_exit(&db->db_mtx);
   1915 
   1916 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
   1917 	dsa->dsa_dr = dr;
   1918 	dsa->dsa_done = done;
   1919 	dsa->dsa_zgd = zgd;
   1920 	dsa->dsa_tx = NULL;
   1921 
   1922 	zio_nowait(arc_write(pio, os->os_spa, txg,
   1923 	    bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
   1924 	    &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
   1925 	    ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
   1926 
   1927 	return (0);
   1928 }
   1929 
   1930 int
   1931 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
   1932     dmu_tx_t *tx)
   1933 {
   1934 	dnode_t *dn;
   1935 	int err;
   1936 
   1937 	err = dnode_hold(os, object, FTAG, &dn);
   1938 	if (err)
   1939 		return (err);
   1940 	err = dnode_set_blksz(dn, size, ibs, tx);
   1941 	dnode_rele(dn, FTAG);
   1942 	return (err);
   1943 }
   1944 
   1945 void
   1946 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
   1947     dmu_tx_t *tx)
   1948 {
   1949 	dnode_t *dn;
   1950 
   1951 	/*
   1952 	 * Send streams include each object's checksum function.  This
   1953 	 * check ensures that the receiving system can understand the
   1954 	 * checksum function transmitted.
   1955 	 */
   1956 	ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
   1957 
   1958 	VERIFY0(dnode_hold(os, object, FTAG, &dn));
   1959 	ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
   1960 	dn->dn_checksum = checksum;
   1961 	dnode_setdirty(dn, tx);
   1962 	dnode_rele(dn, FTAG);
   1963 }
   1964 
   1965 void
   1966 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
   1967     dmu_tx_t *tx)
   1968 {
   1969 	dnode_t *dn;
   1970 
   1971 	/*
   1972 	 * Send streams include each object's compression function.  This
   1973 	 * check ensures that the receiving system can understand the
   1974 	 * compression function transmitted.
   1975 	 */
   1976 	ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
   1977 
   1978 	VERIFY0(dnode_hold(os, object, FTAG, &dn));
   1979 	dn->dn_compress = compress;
   1980 	dnode_setdirty(dn, tx);
   1981 	dnode_rele(dn, FTAG);
   1982 }
   1983 
   1984 int zfs_mdcomp_disable = 0;
   1985 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RWTUN,
   1986     &zfs_mdcomp_disable, 0, "Disable metadata compression");
   1987 
   1988 /*
   1989  * When the "redundant_metadata" property is set to "most", only indirect
   1990  * blocks of this level and higher will have an additional ditto block.
   1991  */
   1992 int zfs_redundant_metadata_most_ditto_level = 2;
   1993 
   1994 void
   1995 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
   1996 {
   1997 	dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
   1998 	boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
   1999 	    (wp & WP_SPILL));
   2000 	enum zio_checksum checksum = os->os_checksum;
   2001 	enum zio_compress compress = os->os_compress;
   2002 	enum zio_checksum dedup_checksum = os->os_dedup_checksum;
   2003 	boolean_t dedup = B_FALSE;
   2004 	boolean_t nopwrite = B_FALSE;
   2005 	boolean_t dedup_verify = os->os_dedup_verify;
   2006 	int copies = os->os_copies;
   2007 
   2008 	/*
   2009 	 * We maintain different write policies for each of the following
   2010 	 * types of data:
   2011 	 *	 1. metadata
   2012 	 *	 2. preallocated blocks (i.e. level-0 blocks of a dump device)
   2013 	 *	 3. all other level 0 blocks
   2014 	 */
   2015 	if (ismd) {
   2016 		if (zfs_mdcomp_disable) {
   2017 			compress = ZIO_COMPRESS_EMPTY;
   2018 		} else {
   2019 			/*
   2020 			 * XXX -- we should design a compression algorithm
   2021 			 * that specializes in arrays of bps.
   2022 			 */
   2023 			compress = zio_compress_select(os->os_spa,
   2024 			    ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
   2025 		}
   2026 
   2027 		/*
   2028 		 * Metadata always gets checksummed.  If the data
   2029 		 * checksum is multi-bit correctable, and it's not a
   2030 		 * ZBT-style checksum, then it's suitable for metadata
   2031 		 * as well.  Otherwise, the metadata checksum defaults
   2032 		 * to fletcher4.
   2033 		 */
   2034 		if (!(zio_checksum_table[checksum].ci_flags &
   2035 		    ZCHECKSUM_FLAG_METADATA) ||
   2036 		    (zio_checksum_table[checksum].ci_flags &
   2037 		    ZCHECKSUM_FLAG_EMBEDDED))
   2038 			checksum = ZIO_CHECKSUM_FLETCHER_4;
   2039 
   2040 		if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
   2041 		    (os->os_redundant_metadata ==
   2042 		    ZFS_REDUNDANT_METADATA_MOST &&
   2043 		    (level >= zfs_redundant_metadata_most_ditto_level ||
   2044 		    DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
   2045 			copies++;
   2046 	} else if (wp & WP_NOFILL) {
   2047 		ASSERT(level == 0);
   2048 
   2049 		/*
   2050 		 * If we're writing preallocated blocks, we aren't actually
   2051 		 * writing them so don't set any policy properties.  These
   2052 		 * blocks are currently only used by an external subsystem
   2053 		 * outside of zfs (i.e. dump) and not written by the zio
   2054 		 * pipeline.
   2055 		 */
   2056 		compress = ZIO_COMPRESS_OFF;
   2057 		checksum = ZIO_CHECKSUM_NOPARITY;
   2058 	} else {
   2059 		compress = zio_compress_select(os->os_spa, dn->dn_compress,
   2060 		    compress);
   2061 
   2062 		checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
   2063 		    zio_checksum_select(dn->dn_checksum, checksum) :
   2064 		    dedup_checksum;
   2065 
   2066 		/*
   2067 		 * Determine dedup setting.  If we are in dmu_sync(),
   2068 		 * we won't actually dedup now because that's all
   2069 		 * done in syncing context; but we do want to use the
   2070 		 * dedup checkum.  If the checksum is not strong
   2071 		 * enough to ensure unique signatures, force
   2072 		 * dedup_verify.
   2073 		 */
   2074 		if (dedup_checksum != ZIO_CHECKSUM_OFF) {
   2075 			dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
   2076 			if (!(zio_checksum_table[checksum].ci_flags &
   2077 			    ZCHECKSUM_FLAG_DEDUP))
   2078 				dedup_verify = B_TRUE;
   2079 		}
   2080 
   2081 		/*
   2082 		 * Enable nopwrite if we have secure enough checksum
   2083 		 * algorithm (see comment in zio_nop_write) and
   2084 		 * compression is enabled.  We don't enable nopwrite if
   2085 		 * dedup is enabled as the two features are mutually
   2086 		 * exclusive.
   2087 		 */
   2088 		nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
   2089 		    ZCHECKSUM_FLAG_NOPWRITE) &&
   2090 		    compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
   2091 	}
   2092 
   2093 	zp->zp_checksum = checksum;
   2094 	zp->zp_compress = compress;
   2095 	zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
   2096 	zp->zp_level = level;
   2097 	zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
   2098 	zp->zp_dedup = dedup;
   2099 	zp->zp_dedup_verify = dedup && dedup_verify;
   2100 	zp->zp_nopwrite = nopwrite;
   2101 }
   2102 
   2103 int
   2104 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
   2105 {
   2106 	dnode_t *dn;
   2107 	int err;
   2108 
   2109 	/*
   2110 	 * Sync any current changes before
   2111 	 * we go trundling through the block pointers.
   2112 	 */
   2113 	err = dmu_object_wait_synced(os, object);
   2114 	if (err) {
   2115 		return (err);
   2116 	}
   2117 
   2118 	err = dnode_hold(os, object, FTAG, &dn);
   2119 	if (err) {
   2120 		return (err);
   2121 	}
   2122 
   2123 	err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
   2124 	dnode_rele(dn, FTAG);
   2125 
   2126 	return (err);
   2127 }
   2128 
   2129 /*
   2130  * Given the ZFS object, if it contains any dirty nodes
   2131  * this function flushes all dirty blocks to disk. This
   2132  * ensures the DMU object info is updated. A more efficient
   2133  * future version might just find the TXG with the maximum
   2134  * ID and wait for that to be synced.
   2135  */
   2136 int
   2137 dmu_object_wait_synced(objset_t *os, uint64_t object)
   2138 {
   2139 	dnode_t *dn;
   2140 	int error, i;
   2141 
   2142 	error = dnode_hold(os, object, FTAG, &dn);
   2143 	if (error) {
   2144 		return (error);
   2145 	}
   2146 
   2147 	for (i = 0; i < TXG_SIZE; i++) {
   2148 		if (list_link_active(&dn->dn_dirty_link[i]) ||
   2149 		    !list_is_empty(&dn->dn_dirty_records[i])) {
   2150 			break;
   2151 		}
   2152 	}
   2153 	dnode_rele(dn, FTAG);
   2154 	if (i != TXG_SIZE) {
   2155 		txg_wait_synced(dmu_objset_pool(os), 0);
   2156 	}
   2157 
   2158 	return (0);
   2159 }
   2160 
   2161 void
   2162 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
   2163 {
   2164 	dnode_phys_t *dnp;
   2165 
   2166 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
   2167 	mutex_enter(&dn->dn_mtx);
   2168 
   2169 	dnp = dn->dn_phys;
   2170 
   2171 	doi->doi_data_block_size = dn->dn_datablksz;
   2172 	doi->doi_metadata_block_size = dn->dn_indblkshift ?
   2173 	    1ULL << dn->dn_indblkshift : 0;
   2174 	doi->doi_type = dn->dn_type;
   2175 	doi->doi_bonus_type = dn->dn_bonustype;
   2176 	doi->doi_bonus_size = dn->dn_bonuslen;
   2177 	doi->doi_indirection = dn->dn_nlevels;
   2178 	doi->doi_checksum = dn->dn_checksum;
   2179 	doi->doi_compress = dn->dn_compress;
   2180 	doi->doi_nblkptr = dn->dn_nblkptr;
   2181 	doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
   2182 	doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
   2183 	doi->doi_fill_count = 0;
   2184 	for (int i = 0; i < dnp->dn_nblkptr; i++)
   2185 		doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
   2186 
   2187 	mutex_exit(&dn->dn_mtx);
   2188 	rw_exit(&dn->dn_struct_rwlock);
   2189 }
   2190 
   2191 /*
   2192  * Get information on a DMU object.
   2193  * If doi is NULL, just indicates whether the object exists.
   2194  */
   2195 int
   2196 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
   2197 {
   2198 	dnode_t *dn;
   2199 	int err = dnode_hold(os, object, FTAG, &dn);
   2200 
   2201 	if (err)
   2202 		return (err);
   2203 
   2204 	if (doi != NULL)
   2205 		dmu_object_info_from_dnode(dn, doi);
   2206 
   2207 	dnode_rele(dn, FTAG);
   2208 	return (0);
   2209 }
   2210 
   2211 /*
   2212  * As above, but faster; can be used when you have a held dbuf in hand.
   2213  */
   2214 void
   2215 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
   2216 {
   2217 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
   2218 
   2219 	DB_DNODE_ENTER(db);
   2220 	dmu_object_info_from_dnode(DB_DNODE(db), doi);
   2221 	DB_DNODE_EXIT(db);
   2222 }
   2223 
   2224 /*
   2225  * Faster still when you only care about the size.
   2226  * This is specifically optimized for zfs_getattr().
   2227  */
   2228 void
   2229 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
   2230     u_longlong_t *nblk512)
   2231 {
   2232 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
   2233 	dnode_t *dn;
   2234 
   2235 	DB_DNODE_ENTER(db);
   2236 	dn = DB_DNODE(db);
   2237 
   2238 	*blksize = dn->dn_datablksz;
   2239 	/* add 1 for dnode space */
   2240 	*nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
   2241 	    SPA_MINBLOCKSHIFT) + 1;
   2242 	DB_DNODE_EXIT(db);
   2243 }
   2244 
   2245 void
   2246 byteswap_uint64_array(void *vbuf, size_t size)
   2247 {
   2248 	uint64_t *buf = vbuf;
   2249 	size_t count = size >> 3;
   2250 	int i;
   2251 
   2252 	ASSERT((size & 7) == 0);
   2253 
   2254 	for (i = 0; i < count; i++)
   2255 		buf[i] = BSWAP_64(buf[i]);
   2256 }
   2257 
   2258 void
   2259 byteswap_uint32_array(void *vbuf, size_t size)
   2260 {
   2261 	uint32_t *buf = vbuf;
   2262 	size_t count = size >> 2;
   2263 	int i;
   2264 
   2265 	ASSERT((size & 3) == 0);
   2266 
   2267 	for (i = 0; i < count; i++)
   2268 		buf[i] = BSWAP_32(buf[i]);
   2269 }
   2270 
   2271 void
   2272 byteswap_uint16_array(void *vbuf, size_t size)
   2273 {
   2274 	uint16_t *buf = vbuf;
   2275 	size_t count = size >> 1;
   2276 	int i;
   2277 
   2278 	ASSERT((size & 1) == 0);
   2279 
   2280 	for (i = 0; i < count; i++)
   2281 		buf[i] = BSWAP_16(buf[i]);
   2282 }
   2283 
   2284 /* ARGSUSED */
   2285 void
   2286 byteswap_uint8_array(void *vbuf, size_t size)
   2287 {
   2288 }
   2289 
   2290 void
   2291 dmu_init(void)
   2292 {
   2293 	zfs_dbgmsg_init();
   2294 	sa_cache_init();
   2295 	xuio_stat_init();
   2296 	dmu_objset_init();
   2297 	dnode_init();
   2298 	zfetch_init();
   2299 	zio_compress_init();
   2300 	l2arc_init();
   2301 	arc_init();
   2302 	dbuf_init();
   2303 }
   2304 
   2305 void
   2306 dmu_fini(void)
   2307 {
   2308 	arc_fini(); /* arc depends on l2arc, so arc must go first */
   2309 	l2arc_fini();
   2310 	zfetch_fini();
   2311 	zio_compress_fini();
   2312 	dbuf_fini();
   2313 	dnode_fini();
   2314 	dmu_objset_fini();
   2315 	xuio_stat_fini();
   2316 	sa_cache_fini();
   2317 	zfs_dbgmsg_fini();
   2318 }
   2319