xref: /freebsd-src/sys/contrib/openzfs/module/zfs/dmu.c (revision 17aab35a77a1b1bf02fc85bb8ffadccb0ca5006d)
1eda14cbcSMatt Macy /*
2eda14cbcSMatt Macy  * CDDL HEADER START
3eda14cbcSMatt Macy  *
4eda14cbcSMatt Macy  * The contents of this file are subject to the terms of the
5eda14cbcSMatt Macy  * Common Development and Distribution License (the "License").
6eda14cbcSMatt Macy  * You may not use this file except in compliance with the License.
7eda14cbcSMatt Macy  *
8eda14cbcSMatt Macy  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9271171e0SMartin Matuska  * or https://opensource.org/licenses/CDDL-1.0.
10eda14cbcSMatt Macy  * See the License for the specific language governing permissions
11eda14cbcSMatt Macy  * and limitations under the License.
12eda14cbcSMatt Macy  *
13eda14cbcSMatt Macy  * When distributing Covered Code, include this CDDL HEADER in each
14eda14cbcSMatt Macy  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15eda14cbcSMatt Macy  * If applicable, add the following below this CDDL HEADER, with the
16eda14cbcSMatt Macy  * fields enclosed by brackets "[]" replaced with your own identifying
17eda14cbcSMatt Macy  * information: Portions Copyright [yyyy] [name of copyright owner]
18eda14cbcSMatt Macy  *
19eda14cbcSMatt Macy  * CDDL HEADER END
20eda14cbcSMatt Macy  */
21eda14cbcSMatt Macy /*
22eda14cbcSMatt Macy  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23eda14cbcSMatt Macy  * Copyright (c) 2011, 2020 by Delphix. All rights reserved.
24eda14cbcSMatt Macy  * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25eda14cbcSMatt Macy  * Copyright (c) 2013, Joyent, Inc. All rights reserved.
26eda14cbcSMatt Macy  * Copyright (c) 2016, Nexenta Systems, Inc. All rights reserved.
27eda14cbcSMatt Macy  * Copyright (c) 2015 by Chunwei Chen. All rights reserved.
28eda14cbcSMatt Macy  * Copyright (c) 2019 Datto Inc.
29ce4dcb97SMartin Matuska  * Copyright (c) 2019, 2023, Klara Inc.
30eda14cbcSMatt Macy  * Copyright (c) 2019, Allan Jude
31dbd5678dSMartin Matuska  * Copyright (c) 2022 Hewlett Packard Enterprise Development LP.
322a58b312SMartin Matuska  * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek
33eda14cbcSMatt Macy  */
34eda14cbcSMatt Macy 
35eda14cbcSMatt Macy #include <sys/dmu.h>
36eda14cbcSMatt Macy #include <sys/dmu_impl.h>
37eda14cbcSMatt Macy #include <sys/dmu_tx.h>
38eda14cbcSMatt Macy #include <sys/dbuf.h>
39eda14cbcSMatt Macy #include <sys/dnode.h>
40eda14cbcSMatt Macy #include <sys/zfs_context.h>
41eda14cbcSMatt Macy #include <sys/dmu_objset.h>
42eda14cbcSMatt Macy #include <sys/dmu_traverse.h>
43eda14cbcSMatt Macy #include <sys/dsl_dataset.h>
44eda14cbcSMatt Macy #include <sys/dsl_dir.h>
45eda14cbcSMatt Macy #include <sys/dsl_pool.h>
46eda14cbcSMatt Macy #include <sys/dsl_synctask.h>
47eda14cbcSMatt Macy #include <sys/dsl_prop.h>
48eda14cbcSMatt Macy #include <sys/dmu_zfetch.h>
49eda14cbcSMatt Macy #include <sys/zfs_ioctl.h>
50eda14cbcSMatt Macy #include <sys/zap.h>
51eda14cbcSMatt Macy #include <sys/zio_checksum.h>
52eda14cbcSMatt Macy #include <sys/zio_compress.h>
53eda14cbcSMatt Macy #include <sys/sa.h>
54eda14cbcSMatt Macy #include <sys/zfeature.h>
55eda14cbcSMatt Macy #include <sys/abd.h>
562a58b312SMartin Matuska #include <sys/brt.h>
57eda14cbcSMatt Macy #include <sys/trace_zfs.h>
58ba27dd8bSMartin Matuska #include <sys/zfs_racct.h>
59eda14cbcSMatt Macy #include <sys/zfs_rlock.h>
60eda14cbcSMatt Macy #ifdef _KERNEL
61eda14cbcSMatt Macy #include <sys/vmsystm.h>
62eda14cbcSMatt Macy #include <sys/zfs_znode.h>
63eda14cbcSMatt Macy #endif
64eda14cbcSMatt Macy 
65eda14cbcSMatt Macy /*
66eda14cbcSMatt Macy  * Enable/disable nopwrite feature.
67eda14cbcSMatt Macy  */
68e92ffd9bSMartin Matuska static int zfs_nopwrite_enabled = 1;
69eda14cbcSMatt Macy 
70eda14cbcSMatt Macy /*
71eda14cbcSMatt Macy  * Tunable to control percentage of dirtied L1 blocks from frees allowed into
72eda14cbcSMatt Macy  * one TXG. After this threshold is crossed, additional dirty blocks from frees
73eda14cbcSMatt Macy  * will wait until the next TXG.
74eda14cbcSMatt Macy  * A value of zero will disable this throttle.
75eda14cbcSMatt Macy  */
76dbd5678dSMartin Matuska static uint_t zfs_per_txg_dirty_frees_percent = 30;
77eda14cbcSMatt Macy 
78eda14cbcSMatt Macy /*
79681ce946SMartin Matuska  * Enable/disable forcing txg sync when dirty checking for holes with lseek().
80681ce946SMartin Matuska  * By default this is enabled to ensure accurate hole reporting, it can result
81681ce946SMartin Matuska  * in a significant performance penalty for lseek(SEEK_HOLE) heavy workloads.
82681ce946SMartin Matuska  * Disabling this option will result in holes never being reported in dirty
83681ce946SMartin Matuska  * files which is always safe.
84eda14cbcSMatt Macy  */
85e92ffd9bSMartin Matuska static int zfs_dmu_offset_next_sync = 1;
86eda14cbcSMatt Macy 
87eda14cbcSMatt Macy /*
88eda14cbcSMatt Macy  * Limit the amount we can prefetch with one call to this amount.  This
89eda14cbcSMatt Macy  * helps to limit the amount of memory that can be used by prefetching.
90eda14cbcSMatt Macy  * Larger objects should be prefetched a bit at a time.
91eda14cbcSMatt Macy  */
92315ee00fSMartin Matuska #ifdef _ILP32
93315ee00fSMartin Matuska uint_t dmu_prefetch_max = 8 * 1024 * 1024;
94315ee00fSMartin Matuska #else
95be181ee2SMartin Matuska uint_t dmu_prefetch_max = 8 * SPA_MAXBLOCKSIZE;
96315ee00fSMartin Matuska #endif
97eda14cbcSMatt Macy 
98e2df9bb4SMartin Matuska /*
99e2df9bb4SMartin Matuska  * Override copies= for dedup state objects. 0 means the traditional behaviour
100e2df9bb4SMartin Matuska  * (ie the default for the containing objset ie 3 for the MOS).
101e2df9bb4SMartin Matuska  */
102e2df9bb4SMartin Matuska uint_t dmu_ddt_copies = 0;
103e2df9bb4SMartin Matuska 
104eda14cbcSMatt Macy const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
105eda14cbcSMatt Macy 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, FALSE, "unallocated"		},
106eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "object directory"	},
107eda14cbcSMatt Macy 	{DMU_BSWAP_UINT64, TRUE,  TRUE,  FALSE, "object array"		},
108eda14cbcSMatt Macy 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, FALSE, "packed nvlist"		},
109eda14cbcSMatt Macy 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "packed nvlist size"	},
110eda14cbcSMatt Macy 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "bpobj"			},
111eda14cbcSMatt Macy 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "bpobj header"		},
112eda14cbcSMatt Macy 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "SPA space map header"	},
113eda14cbcSMatt Macy 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "SPA space map"		},
114eda14cbcSMatt Macy 	{DMU_BSWAP_UINT64, TRUE,  FALSE, TRUE,  "ZIL intent log"	},
115eda14cbcSMatt Macy 	{DMU_BSWAP_DNODE,  TRUE,  FALSE, TRUE,  "DMU dnode"		},
116eda14cbcSMatt Macy 	{DMU_BSWAP_OBJSET, TRUE,  TRUE,  FALSE, "DMU objset"		},
117eda14cbcSMatt Macy 	{DMU_BSWAP_UINT64, TRUE,  TRUE,  FALSE, "DSL directory"		},
118eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL directory child map"},
119eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL dataset snap map"	},
120eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL props"		},
121eda14cbcSMatt Macy 	{DMU_BSWAP_UINT64, TRUE,  TRUE,  FALSE, "DSL dataset"		},
122eda14cbcSMatt Macy 	{DMU_BSWAP_ZNODE,  TRUE,  FALSE, FALSE, "ZFS znode"		},
123eda14cbcSMatt Macy 	{DMU_BSWAP_OLDACL, TRUE,  FALSE, TRUE,  "ZFS V0 ACL"		},
124eda14cbcSMatt Macy 	{DMU_BSWAP_UINT8,  FALSE, FALSE, TRUE,  "ZFS plain file"	},
125eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,  "ZFS directory"		},
126eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "ZFS master node"	},
127eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,  "ZFS delete queue"	},
128eda14cbcSMatt Macy 	{DMU_BSWAP_UINT8,  FALSE, FALSE, TRUE,  "zvol object"		},
129eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "zvol prop"		},
130eda14cbcSMatt Macy 	{DMU_BSWAP_UINT8,  FALSE, FALSE, TRUE,  "other uint8[]"		},
131eda14cbcSMatt Macy 	{DMU_BSWAP_UINT64, FALSE, FALSE, TRUE,  "other uint64[]"	},
132eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "other ZAP"		},
133eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "persistent error log"	},
134eda14cbcSMatt Macy 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, FALSE, "SPA history"		},
135eda14cbcSMatt Macy 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "SPA history offsets"	},
136eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "Pool properties"	},
137eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL permissions"	},
138eda14cbcSMatt Macy 	{DMU_BSWAP_ACL,    TRUE,  FALSE, TRUE,  "ZFS ACL"		},
139eda14cbcSMatt Macy 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, TRUE,  "ZFS SYSACL"		},
140eda14cbcSMatt Macy 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, TRUE,  "FUID table"		},
141eda14cbcSMatt Macy 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "FUID table size"	},
142eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL dataset next clones"},
143eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "scan work queue"	},
144eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,  "ZFS user/group/project used" },
145eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,  "ZFS user/group/project quota"},
146eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "snapshot refcount tags"},
147eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "DDT ZAP algorithm"	},
148eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "DDT statistics"	},
149eda14cbcSMatt Macy 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, TRUE,	"System attributes"	},
150eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,	"SA master node"	},
151eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,	"SA attr registration"	},
152eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, TRUE,	"SA attr layouts"	},
153eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  FALSE, FALSE, "scan translations"	},
154eda14cbcSMatt Macy 	{DMU_BSWAP_UINT8,  FALSE, FALSE, TRUE,  "deduplicated block"	},
155eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL deadlist map"	},
156eda14cbcSMatt Macy 	{DMU_BSWAP_UINT64, TRUE,  TRUE,  FALSE, "DSL deadlist map hdr"	},
157eda14cbcSMatt Macy 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "DSL dir clones"	},
158eda14cbcSMatt Macy 	{DMU_BSWAP_UINT64, TRUE,  FALSE, FALSE, "bpobj subobj"		}
159eda14cbcSMatt Macy };
160eda14cbcSMatt Macy 
161be181ee2SMartin Matuska dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
162eda14cbcSMatt Macy 	{	byteswap_uint8_array,	"uint8"		},
163eda14cbcSMatt Macy 	{	byteswap_uint16_array,	"uint16"	},
164eda14cbcSMatt Macy 	{	byteswap_uint32_array,	"uint32"	},
165eda14cbcSMatt Macy 	{	byteswap_uint64_array,	"uint64"	},
166eda14cbcSMatt Macy 	{	zap_byteswap,		"zap"		},
167eda14cbcSMatt Macy 	{	dnode_buf_byteswap,	"dnode"		},
168eda14cbcSMatt Macy 	{	dmu_objset_byteswap,	"objset"	},
169eda14cbcSMatt Macy 	{	zfs_znode_byteswap,	"znode"		},
170eda14cbcSMatt Macy 	{	zfs_oldacl_byteswap,	"oldacl"	},
171eda14cbcSMatt Macy 	{	zfs_acl_byteswap,	"acl"		}
172eda14cbcSMatt Macy };
173eda14cbcSMatt Macy 
174315ee00fSMartin Matuska int
175eda14cbcSMatt Macy dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
176a0b956f5SMartin Matuska     const void *tag, dmu_buf_t **dbp)
177eda14cbcSMatt Macy {
178eda14cbcSMatt Macy 	uint64_t blkid;
179eda14cbcSMatt Macy 	dmu_buf_impl_t *db;
180eda14cbcSMatt Macy 
181eda14cbcSMatt Macy 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
182eda14cbcSMatt Macy 	blkid = dbuf_whichblock(dn, 0, offset);
183eda14cbcSMatt Macy 	db = dbuf_hold(dn, blkid, tag);
184eda14cbcSMatt Macy 	rw_exit(&dn->dn_struct_rwlock);
185eda14cbcSMatt Macy 
186eda14cbcSMatt Macy 	if (db == NULL) {
187eda14cbcSMatt Macy 		*dbp = NULL;
188eda14cbcSMatt Macy 		return (SET_ERROR(EIO));
189eda14cbcSMatt Macy 	}
190eda14cbcSMatt Macy 
191eda14cbcSMatt Macy 	*dbp = &db->db;
192eda14cbcSMatt Macy 	return (0);
193eda14cbcSMatt Macy }
194315ee00fSMartin Matuska 
195eda14cbcSMatt Macy int
196eda14cbcSMatt Macy dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
197a0b956f5SMartin Matuska     const void *tag, dmu_buf_t **dbp)
198eda14cbcSMatt Macy {
199eda14cbcSMatt Macy 	dnode_t *dn;
200eda14cbcSMatt Macy 	uint64_t blkid;
201eda14cbcSMatt Macy 	dmu_buf_impl_t *db;
202eda14cbcSMatt Macy 	int err;
203eda14cbcSMatt Macy 
204eda14cbcSMatt Macy 	err = dnode_hold(os, object, FTAG, &dn);
205eda14cbcSMatt Macy 	if (err)
206eda14cbcSMatt Macy 		return (err);
207eda14cbcSMatt Macy 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
208eda14cbcSMatt Macy 	blkid = dbuf_whichblock(dn, 0, offset);
209eda14cbcSMatt Macy 	db = dbuf_hold(dn, blkid, tag);
210eda14cbcSMatt Macy 	rw_exit(&dn->dn_struct_rwlock);
211eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
212eda14cbcSMatt Macy 
213eda14cbcSMatt Macy 	if (db == NULL) {
214eda14cbcSMatt Macy 		*dbp = NULL;
215eda14cbcSMatt Macy 		return (SET_ERROR(EIO));
216eda14cbcSMatt Macy 	}
217eda14cbcSMatt Macy 
218eda14cbcSMatt Macy 	*dbp = &db->db;
219eda14cbcSMatt Macy 	return (err);
220eda14cbcSMatt Macy }
221eda14cbcSMatt Macy 
222eda14cbcSMatt Macy int
223eda14cbcSMatt Macy dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
224a0b956f5SMartin Matuska     const void *tag, dmu_buf_t **dbp, int flags)
225eda14cbcSMatt Macy {
226eda14cbcSMatt Macy 	int err;
227eda14cbcSMatt Macy 	int db_flags = DB_RF_CANFAIL;
228eda14cbcSMatt Macy 
229eda14cbcSMatt Macy 	if (flags & DMU_READ_NO_PREFETCH)
230eda14cbcSMatt Macy 		db_flags |= DB_RF_NOPREFETCH;
231eda14cbcSMatt Macy 	if (flags & DMU_READ_NO_DECRYPT)
232eda14cbcSMatt Macy 		db_flags |= DB_RF_NO_DECRYPT;
233eda14cbcSMatt Macy 
234eda14cbcSMatt Macy 	err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
235eda14cbcSMatt Macy 	if (err == 0) {
236eda14cbcSMatt Macy 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
237eda14cbcSMatt Macy 		err = dbuf_read(db, NULL, db_flags);
238eda14cbcSMatt Macy 		if (err != 0) {
239eda14cbcSMatt Macy 			dbuf_rele(db, tag);
240eda14cbcSMatt Macy 			*dbp = NULL;
241eda14cbcSMatt Macy 		}
242eda14cbcSMatt Macy 	}
243eda14cbcSMatt Macy 
244eda14cbcSMatt Macy 	return (err);
245eda14cbcSMatt Macy }
246eda14cbcSMatt Macy 
247eda14cbcSMatt Macy int
248eda14cbcSMatt Macy dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
249a0b956f5SMartin Matuska     const void *tag, dmu_buf_t **dbp, int flags)
250eda14cbcSMatt Macy {
251eda14cbcSMatt Macy 	int err;
252eda14cbcSMatt Macy 	int db_flags = DB_RF_CANFAIL;
253eda14cbcSMatt Macy 
254eda14cbcSMatt Macy 	if (flags & DMU_READ_NO_PREFETCH)
255eda14cbcSMatt Macy 		db_flags |= DB_RF_NOPREFETCH;
256eda14cbcSMatt Macy 	if (flags & DMU_READ_NO_DECRYPT)
257eda14cbcSMatt Macy 		db_flags |= DB_RF_NO_DECRYPT;
258eda14cbcSMatt Macy 
259eda14cbcSMatt Macy 	err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
260eda14cbcSMatt Macy 	if (err == 0) {
261eda14cbcSMatt Macy 		dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
262eda14cbcSMatt Macy 		err = dbuf_read(db, NULL, db_flags);
263eda14cbcSMatt Macy 		if (err != 0) {
264eda14cbcSMatt Macy 			dbuf_rele(db, tag);
265eda14cbcSMatt Macy 			*dbp = NULL;
266eda14cbcSMatt Macy 		}
267eda14cbcSMatt Macy 	}
268eda14cbcSMatt Macy 
269eda14cbcSMatt Macy 	return (err);
270eda14cbcSMatt Macy }
271eda14cbcSMatt Macy 
272eda14cbcSMatt Macy int
273eda14cbcSMatt Macy dmu_bonus_max(void)
274eda14cbcSMatt Macy {
275eda14cbcSMatt Macy 	return (DN_OLD_MAX_BONUSLEN);
276eda14cbcSMatt Macy }
277eda14cbcSMatt Macy 
278eda14cbcSMatt Macy int
279eda14cbcSMatt Macy dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
280eda14cbcSMatt Macy {
281eda14cbcSMatt Macy 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
282eda14cbcSMatt Macy 	dnode_t *dn;
283eda14cbcSMatt Macy 	int error;
284eda14cbcSMatt Macy 
285ce4dcb97SMartin Matuska 	if (newsize < 0 || newsize > db_fake->db_size)
286ce4dcb97SMartin Matuska 		return (SET_ERROR(EINVAL));
287ce4dcb97SMartin Matuska 
288eda14cbcSMatt Macy 	DB_DNODE_ENTER(db);
289eda14cbcSMatt Macy 	dn = DB_DNODE(db);
290eda14cbcSMatt Macy 
291eda14cbcSMatt Macy 	if (dn->dn_bonus != db) {
292eda14cbcSMatt Macy 		error = SET_ERROR(EINVAL);
293eda14cbcSMatt Macy 	} else {
294eda14cbcSMatt Macy 		dnode_setbonuslen(dn, newsize, tx);
295eda14cbcSMatt Macy 		error = 0;
296eda14cbcSMatt Macy 	}
297eda14cbcSMatt Macy 
298eda14cbcSMatt Macy 	DB_DNODE_EXIT(db);
299eda14cbcSMatt Macy 	return (error);
300eda14cbcSMatt Macy }
301eda14cbcSMatt Macy 
302eda14cbcSMatt Macy int
303eda14cbcSMatt Macy dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
304eda14cbcSMatt Macy {
305eda14cbcSMatt Macy 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
306eda14cbcSMatt Macy 	dnode_t *dn;
307eda14cbcSMatt Macy 	int error;
308eda14cbcSMatt Macy 
309ce4dcb97SMartin Matuska 	if (!DMU_OT_IS_VALID(type))
310ce4dcb97SMartin Matuska 		return (SET_ERROR(EINVAL));
311ce4dcb97SMartin Matuska 
312eda14cbcSMatt Macy 	DB_DNODE_ENTER(db);
313eda14cbcSMatt Macy 	dn = DB_DNODE(db);
314eda14cbcSMatt Macy 
315ce4dcb97SMartin Matuska 	if (dn->dn_bonus != db) {
316eda14cbcSMatt Macy 		error = SET_ERROR(EINVAL);
317eda14cbcSMatt Macy 	} else {
318eda14cbcSMatt Macy 		dnode_setbonus_type(dn, type, tx);
319eda14cbcSMatt Macy 		error = 0;
320eda14cbcSMatt Macy 	}
321eda14cbcSMatt Macy 
322eda14cbcSMatt Macy 	DB_DNODE_EXIT(db);
323eda14cbcSMatt Macy 	return (error);
324eda14cbcSMatt Macy }
325eda14cbcSMatt Macy 
326eda14cbcSMatt Macy dmu_object_type_t
327eda14cbcSMatt Macy dmu_get_bonustype(dmu_buf_t *db_fake)
328eda14cbcSMatt Macy {
329eda14cbcSMatt Macy 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
330eda14cbcSMatt Macy 	dmu_object_type_t type;
331eda14cbcSMatt Macy 
332eda14cbcSMatt Macy 	DB_DNODE_ENTER(db);
333ce4dcb97SMartin Matuska 	type = DB_DNODE(db)->dn_bonustype;
334eda14cbcSMatt Macy 	DB_DNODE_EXIT(db);
335eda14cbcSMatt Macy 
336eda14cbcSMatt Macy 	return (type);
337eda14cbcSMatt Macy }
338eda14cbcSMatt Macy 
339eda14cbcSMatt Macy int
340eda14cbcSMatt Macy dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
341eda14cbcSMatt Macy {
342eda14cbcSMatt Macy 	dnode_t *dn;
343eda14cbcSMatt Macy 	int error;
344eda14cbcSMatt Macy 
345eda14cbcSMatt Macy 	error = dnode_hold(os, object, FTAG, &dn);
346eda14cbcSMatt Macy 	dbuf_rm_spill(dn, tx);
347eda14cbcSMatt Macy 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
348eda14cbcSMatt Macy 	dnode_rm_spill(dn, tx);
349eda14cbcSMatt Macy 	rw_exit(&dn->dn_struct_rwlock);
350eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
351eda14cbcSMatt Macy 	return (error);
352eda14cbcSMatt Macy }
353eda14cbcSMatt Macy 
354eda14cbcSMatt Macy /*
355eda14cbcSMatt Macy  * Lookup and hold the bonus buffer for the provided dnode.  If the dnode
356eda14cbcSMatt Macy  * has not yet been allocated a new bonus dbuf a will be allocated.
357eda14cbcSMatt Macy  * Returns ENOENT, EIO, or 0.
358eda14cbcSMatt Macy  */
359a0b956f5SMartin Matuska int dmu_bonus_hold_by_dnode(dnode_t *dn, const void *tag, dmu_buf_t **dbp,
360eda14cbcSMatt Macy     uint32_t flags)
361eda14cbcSMatt Macy {
362eda14cbcSMatt Macy 	dmu_buf_impl_t *db;
363eda14cbcSMatt Macy 	int error;
364eda14cbcSMatt Macy 	uint32_t db_flags = DB_RF_MUST_SUCCEED;
365eda14cbcSMatt Macy 
366eda14cbcSMatt Macy 	if (flags & DMU_READ_NO_PREFETCH)
367eda14cbcSMatt Macy 		db_flags |= DB_RF_NOPREFETCH;
368eda14cbcSMatt Macy 	if (flags & DMU_READ_NO_DECRYPT)
369eda14cbcSMatt Macy 		db_flags |= DB_RF_NO_DECRYPT;
370eda14cbcSMatt Macy 
371eda14cbcSMatt Macy 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
372eda14cbcSMatt Macy 	if (dn->dn_bonus == NULL) {
3732a58b312SMartin Matuska 		if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
374eda14cbcSMatt Macy 			rw_exit(&dn->dn_struct_rwlock);
375eda14cbcSMatt Macy 			rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3762a58b312SMartin Matuska 		}
377eda14cbcSMatt Macy 		if (dn->dn_bonus == NULL)
378eda14cbcSMatt Macy 			dbuf_create_bonus(dn);
379eda14cbcSMatt Macy 	}
380eda14cbcSMatt Macy 	db = dn->dn_bonus;
381eda14cbcSMatt Macy 
382eda14cbcSMatt Macy 	/* as long as the bonus buf is held, the dnode will be held */
383eda14cbcSMatt Macy 	if (zfs_refcount_add(&db->db_holds, tag) == 1) {
384eda14cbcSMatt Macy 		VERIFY(dnode_add_ref(dn, db));
385eda14cbcSMatt Macy 		atomic_inc_32(&dn->dn_dbufs_count);
386eda14cbcSMatt Macy 	}
387eda14cbcSMatt Macy 
388eda14cbcSMatt Macy 	/*
389eda14cbcSMatt Macy 	 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
390eda14cbcSMatt Macy 	 * hold and incrementing the dbuf count to ensure that dnode_move() sees
391eda14cbcSMatt Macy 	 * a dnode hold for every dbuf.
392eda14cbcSMatt Macy 	 */
393eda14cbcSMatt Macy 	rw_exit(&dn->dn_struct_rwlock);
394eda14cbcSMatt Macy 
395eda14cbcSMatt Macy 	error = dbuf_read(db, NULL, db_flags);
396eda14cbcSMatt Macy 	if (error) {
397eda14cbcSMatt Macy 		dnode_evict_bonus(dn);
398eda14cbcSMatt Macy 		dbuf_rele(db, tag);
399eda14cbcSMatt Macy 		*dbp = NULL;
400eda14cbcSMatt Macy 		return (error);
401eda14cbcSMatt Macy 	}
402eda14cbcSMatt Macy 
403eda14cbcSMatt Macy 	*dbp = &db->db;
404eda14cbcSMatt Macy 	return (0);
405eda14cbcSMatt Macy }
406eda14cbcSMatt Macy 
407eda14cbcSMatt Macy int
408a0b956f5SMartin Matuska dmu_bonus_hold(objset_t *os, uint64_t object, const void *tag, dmu_buf_t **dbp)
409eda14cbcSMatt Macy {
410eda14cbcSMatt Macy 	dnode_t *dn;
411eda14cbcSMatt Macy 	int error;
412eda14cbcSMatt Macy 
413eda14cbcSMatt Macy 	error = dnode_hold(os, object, FTAG, &dn);
414eda14cbcSMatt Macy 	if (error)
415eda14cbcSMatt Macy 		return (error);
416eda14cbcSMatt Macy 
417eda14cbcSMatt Macy 	error = dmu_bonus_hold_by_dnode(dn, tag, dbp, DMU_READ_NO_PREFETCH);
418eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
419eda14cbcSMatt Macy 
420eda14cbcSMatt Macy 	return (error);
421eda14cbcSMatt Macy }
422eda14cbcSMatt Macy 
423eda14cbcSMatt Macy /*
424eda14cbcSMatt Macy  * returns ENOENT, EIO, or 0.
425eda14cbcSMatt Macy  *
426eda14cbcSMatt Macy  * This interface will allocate a blank spill dbuf when a spill blk
427eda14cbcSMatt Macy  * doesn't already exist on the dnode.
428eda14cbcSMatt Macy  *
429eda14cbcSMatt Macy  * if you only want to find an already existing spill db, then
430eda14cbcSMatt Macy  * dmu_spill_hold_existing() should be used.
431eda14cbcSMatt Macy  */
432eda14cbcSMatt Macy int
433a0b956f5SMartin Matuska dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, const void *tag,
434a0b956f5SMartin Matuska     dmu_buf_t **dbp)
435eda14cbcSMatt Macy {
436eda14cbcSMatt Macy 	dmu_buf_impl_t *db = NULL;
437eda14cbcSMatt Macy 	int err;
438eda14cbcSMatt Macy 
439eda14cbcSMatt Macy 	if ((flags & DB_RF_HAVESTRUCT) == 0)
440eda14cbcSMatt Macy 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
441eda14cbcSMatt Macy 
442eda14cbcSMatt Macy 	db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
443eda14cbcSMatt Macy 
444eda14cbcSMatt Macy 	if ((flags & DB_RF_HAVESTRUCT) == 0)
445eda14cbcSMatt Macy 		rw_exit(&dn->dn_struct_rwlock);
446eda14cbcSMatt Macy 
447eda14cbcSMatt Macy 	if (db == NULL) {
448eda14cbcSMatt Macy 		*dbp = NULL;
449eda14cbcSMatt Macy 		return (SET_ERROR(EIO));
450eda14cbcSMatt Macy 	}
451eda14cbcSMatt Macy 	err = dbuf_read(db, NULL, flags);
452eda14cbcSMatt Macy 	if (err == 0)
453eda14cbcSMatt Macy 		*dbp = &db->db;
454eda14cbcSMatt Macy 	else {
455eda14cbcSMatt Macy 		dbuf_rele(db, tag);
456eda14cbcSMatt Macy 		*dbp = NULL;
457eda14cbcSMatt Macy 	}
458eda14cbcSMatt Macy 	return (err);
459eda14cbcSMatt Macy }
460eda14cbcSMatt Macy 
461eda14cbcSMatt Macy int
462a0b956f5SMartin Matuska dmu_spill_hold_existing(dmu_buf_t *bonus, const void *tag, dmu_buf_t **dbp)
463eda14cbcSMatt Macy {
464eda14cbcSMatt Macy 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
465eda14cbcSMatt Macy 	dnode_t *dn;
466eda14cbcSMatt Macy 	int err;
467eda14cbcSMatt Macy 
468eda14cbcSMatt Macy 	DB_DNODE_ENTER(db);
469eda14cbcSMatt Macy 	dn = DB_DNODE(db);
470eda14cbcSMatt Macy 
471eda14cbcSMatt Macy 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
472eda14cbcSMatt Macy 		err = SET_ERROR(EINVAL);
473eda14cbcSMatt Macy 	} else {
474eda14cbcSMatt Macy 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
475eda14cbcSMatt Macy 
476eda14cbcSMatt Macy 		if (!dn->dn_have_spill) {
477eda14cbcSMatt Macy 			err = SET_ERROR(ENOENT);
478eda14cbcSMatt Macy 		} else {
479eda14cbcSMatt Macy 			err = dmu_spill_hold_by_dnode(dn,
480eda14cbcSMatt Macy 			    DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
481eda14cbcSMatt Macy 		}
482eda14cbcSMatt Macy 
483eda14cbcSMatt Macy 		rw_exit(&dn->dn_struct_rwlock);
484eda14cbcSMatt Macy 	}
485eda14cbcSMatt Macy 
486eda14cbcSMatt Macy 	DB_DNODE_EXIT(db);
487eda14cbcSMatt Macy 	return (err);
488eda14cbcSMatt Macy }
489eda14cbcSMatt Macy 
490eda14cbcSMatt Macy int
491a0b956f5SMartin Matuska dmu_spill_hold_by_bonus(dmu_buf_t *bonus, uint32_t flags, const void *tag,
492eda14cbcSMatt Macy     dmu_buf_t **dbp)
493eda14cbcSMatt Macy {
494eda14cbcSMatt Macy 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
495eda14cbcSMatt Macy 	int err;
496eda14cbcSMatt Macy 	uint32_t db_flags = DB_RF_CANFAIL;
497eda14cbcSMatt Macy 
498eda14cbcSMatt Macy 	if (flags & DMU_READ_NO_DECRYPT)
499eda14cbcSMatt Macy 		db_flags |= DB_RF_NO_DECRYPT;
500eda14cbcSMatt Macy 
501eda14cbcSMatt Macy 	DB_DNODE_ENTER(db);
502ce4dcb97SMartin Matuska 	err = dmu_spill_hold_by_dnode(DB_DNODE(db), db_flags, tag, dbp);
503eda14cbcSMatt Macy 	DB_DNODE_EXIT(db);
504eda14cbcSMatt Macy 
505eda14cbcSMatt Macy 	return (err);
506eda14cbcSMatt Macy }
507eda14cbcSMatt Macy 
508eda14cbcSMatt Macy /*
509eda14cbcSMatt Macy  * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
510eda14cbcSMatt Macy  * to take a held dnode rather than <os, object> -- the lookup is wasteful,
511eda14cbcSMatt Macy  * and can induce severe lock contention when writing to several files
512eda14cbcSMatt Macy  * whose dnodes are in the same block.
513eda14cbcSMatt Macy  */
514eda14cbcSMatt Macy int
515eda14cbcSMatt Macy dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
516a0b956f5SMartin Matuska     boolean_t read, const void *tag, int *numbufsp, dmu_buf_t ***dbpp,
517a0b956f5SMartin Matuska     uint32_t flags)
518eda14cbcSMatt Macy {
519eda14cbcSMatt Macy 	dmu_buf_t **dbp;
520f9693befSMartin Matuska 	zstream_t *zs = NULL;
521eda14cbcSMatt Macy 	uint64_t blkid, nblks, i;
522eda14cbcSMatt Macy 	uint32_t dbuf_flags;
523eda14cbcSMatt Macy 	int err;
5247877fdebSMatt Macy 	zio_t *zio = NULL;
525f9693befSMartin Matuska 	boolean_t missed = B_FALSE;
526eda14cbcSMatt Macy 
5272a58b312SMartin Matuska 	ASSERT(!read || length <= DMU_MAX_ACCESS);
528eda14cbcSMatt Macy 
529eda14cbcSMatt Macy 	/*
530eda14cbcSMatt Macy 	 * Note: We directly notify the prefetch code of this read, so that
531eda14cbcSMatt Macy 	 * we can tell it about the multi-block read.  dbuf_read() only knows
532eda14cbcSMatt Macy 	 * about the one block it is accessing.
533eda14cbcSMatt Macy 	 */
534eda14cbcSMatt Macy 	dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
535eda14cbcSMatt Macy 	    DB_RF_NOPREFETCH;
536eda14cbcSMatt Macy 
537271171e0SMartin Matuska 	if ((flags & DMU_READ_NO_DECRYPT) != 0)
538271171e0SMartin Matuska 		dbuf_flags |= DB_RF_NO_DECRYPT;
539271171e0SMartin Matuska 
540eda14cbcSMatt Macy 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
541eda14cbcSMatt Macy 	if (dn->dn_datablkshift) {
542eda14cbcSMatt Macy 		int blkshift = dn->dn_datablkshift;
543eda14cbcSMatt Macy 		nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
544aca928a5SMartin Matuska 		    P2ALIGN_TYPED(offset, 1ULL << blkshift, uint64_t))
545aca928a5SMartin Matuska 		    >> blkshift;
546eda14cbcSMatt Macy 	} else {
547eda14cbcSMatt Macy 		if (offset + length > dn->dn_datablksz) {
548eda14cbcSMatt Macy 			zfs_panic_recover("zfs: accessing past end of object "
549eda14cbcSMatt Macy 			    "%llx/%llx (size=%u access=%llu+%llu)",
550eda14cbcSMatt Macy 			    (longlong_t)dn->dn_objset->
551eda14cbcSMatt Macy 			    os_dsl_dataset->ds_object,
552eda14cbcSMatt Macy 			    (longlong_t)dn->dn_object, dn->dn_datablksz,
553eda14cbcSMatt Macy 			    (longlong_t)offset, (longlong_t)length);
554eda14cbcSMatt Macy 			rw_exit(&dn->dn_struct_rwlock);
555eda14cbcSMatt Macy 			return (SET_ERROR(EIO));
556eda14cbcSMatt Macy 		}
557eda14cbcSMatt Macy 		nblks = 1;
558eda14cbcSMatt Macy 	}
559eda14cbcSMatt Macy 	dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
560eda14cbcSMatt Macy 
5617877fdebSMatt Macy 	if (read)
5627877fdebSMatt Macy 		zio = zio_root(dn->dn_objset->os_spa, NULL, NULL,
5637877fdebSMatt Macy 		    ZIO_FLAG_CANFAIL);
564eda14cbcSMatt Macy 	blkid = dbuf_whichblock(dn, 0, offset);
565315ee00fSMartin Matuska 	if ((flags & DMU_READ_NO_PREFETCH) == 0) {
566f9693befSMartin Matuska 		/*
567f9693befSMartin Matuska 		 * Prepare the zfetch before initiating the demand reads, so
568f9693befSMartin Matuska 		 * that if multiple threads block on same indirect block, we
569f9693befSMartin Matuska 		 * base predictions on the original less racy request order.
570f9693befSMartin Matuska 		 */
57115f0b8c3SMartin Matuska 		zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks, read,
57215f0b8c3SMartin Matuska 		    B_TRUE);
573f9693befSMartin Matuska 	}
574eda14cbcSMatt Macy 	for (i = 0; i < nblks; i++) {
575eda14cbcSMatt Macy 		dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
576eda14cbcSMatt Macy 		if (db == NULL) {
5771719886fSMartin Matuska 			if (zs) {
5781719886fSMartin Matuska 				dmu_zfetch_run(&dn->dn_zfetch, zs, missed,
5791719886fSMartin Matuska 				    B_TRUE);
5801719886fSMartin Matuska 			}
581eda14cbcSMatt Macy 			rw_exit(&dn->dn_struct_rwlock);
582eda14cbcSMatt Macy 			dmu_buf_rele_array(dbp, nblks, tag);
5837877fdebSMatt Macy 			if (read)
584eda14cbcSMatt Macy 				zio_nowait(zio);
585eda14cbcSMatt Macy 			return (SET_ERROR(EIO));
586eda14cbcSMatt Macy 		}
587eda14cbcSMatt Macy 
588f9693befSMartin Matuska 		/*
589f9693befSMartin Matuska 		 * Initiate async demand data read.
590f9693befSMartin Matuska 		 * We check the db_state after calling dbuf_read() because
591f9693befSMartin Matuska 		 * (1) dbuf_read() may change the state to CACHED due to a
592f9693befSMartin Matuska 		 * hit in the ARC, and (2) on a cache miss, a child will
593f9693befSMartin Matuska 		 * have been added to "zio" but not yet completed, so the
594f9693befSMartin Matuska 		 * state will not yet be CACHED.
595f9693befSMartin Matuska 		 */
596f9693befSMartin Matuska 		if (read) {
59715f0b8c3SMartin Matuska 			if (i == nblks - 1 && blkid + i < dn->dn_maxblkid &&
59815f0b8c3SMartin Matuska 			    offset + length < db->db.db_offset +
59915f0b8c3SMartin Matuska 			    db->db.db_size) {
60015f0b8c3SMartin Matuska 				if (offset <= db->db.db_offset)
60115f0b8c3SMartin Matuska 					dbuf_flags |= DB_RF_PARTIAL_FIRST;
60215f0b8c3SMartin Matuska 				else
60315f0b8c3SMartin Matuska 					dbuf_flags |= DB_RF_PARTIAL_MORE;
60415f0b8c3SMartin Matuska 			}
605eda14cbcSMatt Macy 			(void) dbuf_read(db, zio, dbuf_flags);
606f9693befSMartin Matuska 			if (db->db_state != DB_CACHED)
607f9693befSMartin Matuska 				missed = B_TRUE;
608f9693befSMartin Matuska 		}
609eda14cbcSMatt Macy 		dbp[i] = &db->db;
610eda14cbcSMatt Macy 	}
611eda14cbcSMatt Macy 
6127a7741afSMartin Matuska 	/*
6137a7741afSMartin Matuska 	 * If we are doing O_DIRECT we still hold the dbufs, even for reads,
6147a7741afSMartin Matuska 	 * but we do not issue any reads here. We do not want to account for
6157a7741afSMartin Matuska 	 * writes in this case.
6167a7741afSMartin Matuska 	 *
6177a7741afSMartin Matuska 	 * O_DIRECT write/read accounting takes place in
6187a7741afSMartin Matuska 	 * dmu_{write/read}_abd().
6197a7741afSMartin Matuska 	 */
6207a7741afSMartin Matuska 	if (!read && ((flags & DMU_DIRECTIO) == 0))
6217a7741afSMartin Matuska 		zfs_racct_write(dn->dn_objset->os_spa, length, nblks, flags);
622ba27dd8bSMartin Matuska 
623f9693befSMartin Matuska 	if (zs)
6241719886fSMartin Matuska 		dmu_zfetch_run(&dn->dn_zfetch, zs, missed, B_TRUE);
625eda14cbcSMatt Macy 	rw_exit(&dn->dn_struct_rwlock);
626eda14cbcSMatt Macy 
6277877fdebSMatt Macy 	if (read) {
6287877fdebSMatt Macy 		/* wait for async read i/o */
629eda14cbcSMatt Macy 		err = zio_wait(zio);
630eda14cbcSMatt Macy 		if (err) {
631eda14cbcSMatt Macy 			dmu_buf_rele_array(dbp, nblks, tag);
632eda14cbcSMatt Macy 			return (err);
633eda14cbcSMatt Macy 		}
634eda14cbcSMatt Macy 
635eda14cbcSMatt Macy 		/* wait for other io to complete */
636eda14cbcSMatt Macy 		for (i = 0; i < nblks; i++) {
637eda14cbcSMatt Macy 			dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
638eda14cbcSMatt Macy 			mutex_enter(&db->db_mtx);
639eda14cbcSMatt Macy 			while (db->db_state == DB_READ ||
640eda14cbcSMatt Macy 			    db->db_state == DB_FILL)
641eda14cbcSMatt Macy 				cv_wait(&db->db_changed, &db->db_mtx);
642eda14cbcSMatt Macy 			if (db->db_state == DB_UNCACHED)
643eda14cbcSMatt Macy 				err = SET_ERROR(EIO);
644eda14cbcSMatt Macy 			mutex_exit(&db->db_mtx);
645eda14cbcSMatt Macy 			if (err) {
646eda14cbcSMatt Macy 				dmu_buf_rele_array(dbp, nblks, tag);
647eda14cbcSMatt Macy 				return (err);
648eda14cbcSMatt Macy 			}
649eda14cbcSMatt Macy 		}
650eda14cbcSMatt Macy 	}
651eda14cbcSMatt Macy 
652eda14cbcSMatt Macy 	*numbufsp = nblks;
653eda14cbcSMatt Macy 	*dbpp = dbp;
654eda14cbcSMatt Macy 	return (0);
655eda14cbcSMatt Macy }
656eda14cbcSMatt Macy 
6576ba2210eSMartin Matuska int
658eda14cbcSMatt Macy dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
659a0b956f5SMartin Matuska     uint64_t length, int read, const void *tag, int *numbufsp,
660a0b956f5SMartin Matuska     dmu_buf_t ***dbpp)
661eda14cbcSMatt Macy {
662eda14cbcSMatt Macy 	dnode_t *dn;
663eda14cbcSMatt Macy 	int err;
664eda14cbcSMatt Macy 
665eda14cbcSMatt Macy 	err = dnode_hold(os, object, FTAG, &dn);
666eda14cbcSMatt Macy 	if (err)
667eda14cbcSMatt Macy 		return (err);
668eda14cbcSMatt Macy 
669eda14cbcSMatt Macy 	err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
670eda14cbcSMatt Macy 	    numbufsp, dbpp, DMU_READ_PREFETCH);
671eda14cbcSMatt Macy 
672eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
673eda14cbcSMatt Macy 
674eda14cbcSMatt Macy 	return (err);
675eda14cbcSMatt Macy }
676eda14cbcSMatt Macy 
677eda14cbcSMatt Macy int
678eda14cbcSMatt Macy dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
679a0b956f5SMartin Matuska     uint64_t length, boolean_t read, const void *tag, int *numbufsp,
680eda14cbcSMatt Macy     dmu_buf_t ***dbpp)
681eda14cbcSMatt Macy {
682eda14cbcSMatt Macy 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
683eda14cbcSMatt Macy 	int err;
684eda14cbcSMatt Macy 
685eda14cbcSMatt Macy 	DB_DNODE_ENTER(db);
686ce4dcb97SMartin Matuska 	err = dmu_buf_hold_array_by_dnode(DB_DNODE(db), offset, length, read,
687ce4dcb97SMartin Matuska 	    tag, numbufsp, dbpp, DMU_READ_PREFETCH);
688eda14cbcSMatt Macy 	DB_DNODE_EXIT(db);
689eda14cbcSMatt Macy 
690eda14cbcSMatt Macy 	return (err);
691eda14cbcSMatt Macy }
692eda14cbcSMatt Macy 
693eda14cbcSMatt Macy void
694a0b956f5SMartin Matuska dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, const void *tag)
695eda14cbcSMatt Macy {
696eda14cbcSMatt Macy 	int i;
697eda14cbcSMatt Macy 	dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
698eda14cbcSMatt Macy 
699eda14cbcSMatt Macy 	if (numbufs == 0)
700eda14cbcSMatt Macy 		return;
701eda14cbcSMatt Macy 
702eda14cbcSMatt Macy 	for (i = 0; i < numbufs; i++) {
703eda14cbcSMatt Macy 		if (dbp[i])
704eda14cbcSMatt Macy 			dbuf_rele(dbp[i], tag);
705eda14cbcSMatt Macy 	}
706eda14cbcSMatt Macy 
707eda14cbcSMatt Macy 	kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
708eda14cbcSMatt Macy }
709eda14cbcSMatt Macy 
710eda14cbcSMatt Macy /*
711315ee00fSMartin Matuska  * Issue prefetch I/Os for the given blocks.  If level is greater than 0, the
712eda14cbcSMatt Macy  * indirect blocks prefetched will be those that point to the blocks containing
713315ee00fSMartin Matuska  * the data starting at offset, and continuing to offset + len.  If the range
714ce4dcb97SMartin Matuska  * is too long, prefetch the first dmu_prefetch_max bytes as requested, while
715315ee00fSMartin Matuska  * for the rest only a higher level, also fitting within dmu_prefetch_max.  It
716315ee00fSMartin Matuska  * should primarily help random reads, since for long sequential reads there is
717315ee00fSMartin Matuska  * a speculative prefetcher.
718eda14cbcSMatt Macy  *
719eda14cbcSMatt Macy  * Note that if the indirect blocks above the blocks being prefetched are not
720315ee00fSMartin Matuska  * in cache, they will be asynchronously read in.  Dnode read by dnode_hold()
721315ee00fSMartin Matuska  * is currently synchronous.
722eda14cbcSMatt Macy  */
723eda14cbcSMatt Macy void
724eda14cbcSMatt Macy dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
725eda14cbcSMatt Macy     uint64_t len, zio_priority_t pri)
726eda14cbcSMatt Macy {
727eda14cbcSMatt Macy 	dnode_t *dn;
728eda14cbcSMatt Macy 
729315ee00fSMartin Matuska 	if (dmu_prefetch_max == 0 || len == 0) {
730315ee00fSMartin Matuska 		dmu_prefetch_dnode(os, object, pri);
731eda14cbcSMatt Macy 		return;
732eda14cbcSMatt Macy 	}
733eda14cbcSMatt Macy 
734315ee00fSMartin Matuska 	if (dnode_hold(os, object, FTAG, &dn) != 0)
735eda14cbcSMatt Macy 		return;
736eda14cbcSMatt Macy 
737783d3ff6SMartin Matuska 	dmu_prefetch_by_dnode(dn, level, offset, len, pri);
738783d3ff6SMartin Matuska 
739783d3ff6SMartin Matuska 	dnode_rele(dn, FTAG);
740783d3ff6SMartin Matuska }
741783d3ff6SMartin Matuska 
742783d3ff6SMartin Matuska void
743783d3ff6SMartin Matuska dmu_prefetch_by_dnode(dnode_t *dn, int64_t level, uint64_t offset,
744783d3ff6SMartin Matuska     uint64_t len, zio_priority_t pri)
745783d3ff6SMartin Matuska {
746783d3ff6SMartin Matuska 	int64_t level2 = level;
747783d3ff6SMartin Matuska 	uint64_t start, end, start2, end2;
748783d3ff6SMartin Matuska 
749eda14cbcSMatt Macy 	/*
750315ee00fSMartin Matuska 	 * Depending on len we may do two prefetches: blocks [start, end) at
751315ee00fSMartin Matuska 	 * level, and following blocks [start2, end2) at higher level2.
752eda14cbcSMatt Macy 	 */
753eda14cbcSMatt Macy 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
754315ee00fSMartin Matuska 	if (dn->dn_datablkshift != 0) {
755315ee00fSMartin Matuska 		/*
756315ee00fSMartin Matuska 		 * The object has multiple blocks.  Calculate the full range
757315ee00fSMartin Matuska 		 * of blocks [start, end2) and then split it into two parts,
758315ee00fSMartin Matuska 		 * so that the first [start, end) fits into dmu_prefetch_max.
759315ee00fSMartin Matuska 		 */
760315ee00fSMartin Matuska 		start = dbuf_whichblock(dn, level, offset);
761315ee00fSMartin Matuska 		end2 = dbuf_whichblock(dn, level, offset + len - 1) + 1;
762315ee00fSMartin Matuska 		uint8_t ibs = dn->dn_indblkshift;
763315ee00fSMartin Matuska 		uint8_t bs = (level == 0) ? dn->dn_datablkshift : ibs;
764315ee00fSMartin Matuska 		uint_t limit = P2ROUNDUP(dmu_prefetch_max, 1 << bs) >> bs;
765315ee00fSMartin Matuska 		start2 = end = MIN(end2, start + limit);
766315ee00fSMartin Matuska 
767315ee00fSMartin Matuska 		/*
768315ee00fSMartin Matuska 		 * Find level2 where [start2, end2) fits into dmu_prefetch_max.
769315ee00fSMartin Matuska 		 */
770315ee00fSMartin Matuska 		uint8_t ibps = ibs - SPA_BLKPTRSHIFT;
771315ee00fSMartin Matuska 		limit = P2ROUNDUP(dmu_prefetch_max, 1 << ibs) >> ibs;
772315ee00fSMartin Matuska 		do {
773315ee00fSMartin Matuska 			level2++;
774315ee00fSMartin Matuska 			start2 = P2ROUNDUP(start2, 1 << ibps) >> ibps;
775315ee00fSMartin Matuska 			end2 = P2ROUNDUP(end2, 1 << ibps) >> ibps;
776315ee00fSMartin Matuska 		} while (end2 - start2 > limit);
777eda14cbcSMatt Macy 	} else {
778315ee00fSMartin Matuska 		/* There is only one block.  Prefetch it or nothing. */
779315ee00fSMartin Matuska 		start = start2 = end2 = 0;
780315ee00fSMartin Matuska 		end = start + (level == 0 && offset < dn->dn_datablksz);
781eda14cbcSMatt Macy 	}
782eda14cbcSMatt Macy 
783315ee00fSMartin Matuska 	for (uint64_t i = start; i < end; i++)
784315ee00fSMartin Matuska 		dbuf_prefetch(dn, level, i, pri, 0);
785315ee00fSMartin Matuska 	for (uint64_t i = start2; i < end2; i++)
786315ee00fSMartin Matuska 		dbuf_prefetch(dn, level2, i, pri, 0);
787eda14cbcSMatt Macy 	rw_exit(&dn->dn_struct_rwlock);
788eda14cbcSMatt Macy }
789eda14cbcSMatt Macy 
790ce4dcb97SMartin Matuska typedef struct {
791ce4dcb97SMartin Matuska 	kmutex_t	dpa_lock;
792ce4dcb97SMartin Matuska 	kcondvar_t	dpa_cv;
793ce4dcb97SMartin Matuska 	uint64_t	dpa_pending_io;
794ce4dcb97SMartin Matuska } dmu_prefetch_arg_t;
795ce4dcb97SMartin Matuska 
796ce4dcb97SMartin Matuska static void
797ce4dcb97SMartin Matuska dmu_prefetch_done(void *arg, uint64_t level, uint64_t blkid, boolean_t issued)
798ce4dcb97SMartin Matuska {
799ce4dcb97SMartin Matuska 	(void) level; (void) blkid; (void)issued;
800ce4dcb97SMartin Matuska 	dmu_prefetch_arg_t *dpa = arg;
801ce4dcb97SMartin Matuska 
802ce4dcb97SMartin Matuska 	ASSERT0(level);
803ce4dcb97SMartin Matuska 
804ce4dcb97SMartin Matuska 	mutex_enter(&dpa->dpa_lock);
805ce4dcb97SMartin Matuska 	ASSERT3U(dpa->dpa_pending_io, >, 0);
806ce4dcb97SMartin Matuska 	if (--dpa->dpa_pending_io == 0)
807ce4dcb97SMartin Matuska 		cv_broadcast(&dpa->dpa_cv);
808ce4dcb97SMartin Matuska 	mutex_exit(&dpa->dpa_lock);
809ce4dcb97SMartin Matuska }
810ce4dcb97SMartin Matuska 
811ce4dcb97SMartin Matuska static void
812ce4dcb97SMartin Matuska dmu_prefetch_wait_by_dnode(dnode_t *dn, uint64_t offset, uint64_t len)
813ce4dcb97SMartin Matuska {
814ce4dcb97SMartin Matuska 	dmu_prefetch_arg_t dpa;
815ce4dcb97SMartin Matuska 
816ce4dcb97SMartin Matuska 	mutex_init(&dpa.dpa_lock, NULL, MUTEX_DEFAULT, NULL);
817ce4dcb97SMartin Matuska 	cv_init(&dpa.dpa_cv, NULL, CV_DEFAULT, NULL);
818ce4dcb97SMartin Matuska 
819ce4dcb97SMartin Matuska 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
820ce4dcb97SMartin Matuska 
821ce4dcb97SMartin Matuska 	uint64_t start = dbuf_whichblock(dn, 0, offset);
822ce4dcb97SMartin Matuska 	uint64_t end = dbuf_whichblock(dn, 0, offset + len - 1) + 1;
823ce4dcb97SMartin Matuska 	dpa.dpa_pending_io = end - start;
824ce4dcb97SMartin Matuska 
825ce4dcb97SMartin Matuska 	for (uint64_t blk = start; blk < end; blk++) {
826ce4dcb97SMartin Matuska 		(void) dbuf_prefetch_impl(dn, 0, blk, ZIO_PRIORITY_ASYNC_READ,
827ce4dcb97SMartin Matuska 		    0, dmu_prefetch_done, &dpa);
828ce4dcb97SMartin Matuska 	}
829ce4dcb97SMartin Matuska 
830ce4dcb97SMartin Matuska 	rw_exit(&dn->dn_struct_rwlock);
831ce4dcb97SMartin Matuska 
832ce4dcb97SMartin Matuska 	/* wait for prefetch L0 reads to finish */
833ce4dcb97SMartin Matuska 	mutex_enter(&dpa.dpa_lock);
834ce4dcb97SMartin Matuska 	while (dpa.dpa_pending_io > 0) {
835ce4dcb97SMartin Matuska 		cv_wait(&dpa.dpa_cv, &dpa.dpa_lock);
836ce4dcb97SMartin Matuska 
837ce4dcb97SMartin Matuska 	}
838ce4dcb97SMartin Matuska 	mutex_exit(&dpa.dpa_lock);
839ce4dcb97SMartin Matuska 
840ce4dcb97SMartin Matuska 	mutex_destroy(&dpa.dpa_lock);
841ce4dcb97SMartin Matuska 	cv_destroy(&dpa.dpa_cv);
842ce4dcb97SMartin Matuska }
843ce4dcb97SMartin Matuska 
844ce4dcb97SMartin Matuska /*
845ce4dcb97SMartin Matuska  * Issue prefetch I/Os for the given L0 block range and wait for the I/O
846ce4dcb97SMartin Matuska  * to complete. This does not enforce dmu_prefetch_max and will prefetch
847ce4dcb97SMartin Matuska  * the entire range. The blocks are read from disk into the ARC but no
848ce4dcb97SMartin Matuska  * decompression occurs (i.e., the dbuf cache is not required).
849ce4dcb97SMartin Matuska  */
850ce4dcb97SMartin Matuska int
851ce4dcb97SMartin Matuska dmu_prefetch_wait(objset_t *os, uint64_t object, uint64_t offset, uint64_t size)
852ce4dcb97SMartin Matuska {
853ce4dcb97SMartin Matuska 	dnode_t *dn;
854ce4dcb97SMartin Matuska 	int err = 0;
855ce4dcb97SMartin Matuska 
856ce4dcb97SMartin Matuska 	err = dnode_hold(os, object, FTAG, &dn);
857ce4dcb97SMartin Matuska 	if (err != 0)
858ce4dcb97SMartin Matuska 		return (err);
859ce4dcb97SMartin Matuska 
860ce4dcb97SMartin Matuska 	/*
861ce4dcb97SMartin Matuska 	 * Chunk the requests (16 indirects worth) so that we can be interrupted
862ce4dcb97SMartin Matuska 	 */
863ce4dcb97SMartin Matuska 	uint64_t chunksize;
864ce4dcb97SMartin Matuska 	if (dn->dn_indblkshift) {
865ce4dcb97SMartin Matuska 		uint64_t nbps = bp_span_in_blocks(dn->dn_indblkshift, 1);
866ce4dcb97SMartin Matuska 		chunksize = (nbps * 16) << dn->dn_datablkshift;
867ce4dcb97SMartin Matuska 	} else {
868ce4dcb97SMartin Matuska 		chunksize = dn->dn_datablksz;
869ce4dcb97SMartin Matuska 	}
870ce4dcb97SMartin Matuska 
871ce4dcb97SMartin Matuska 	while (size > 0) {
872ce4dcb97SMartin Matuska 		uint64_t mylen = MIN(size, chunksize);
873ce4dcb97SMartin Matuska 
874ce4dcb97SMartin Matuska 		dmu_prefetch_wait_by_dnode(dn, offset, mylen);
875ce4dcb97SMartin Matuska 
876ce4dcb97SMartin Matuska 		offset += mylen;
877ce4dcb97SMartin Matuska 		size -= mylen;
878ce4dcb97SMartin Matuska 
879ce4dcb97SMartin Matuska 		if (issig()) {
880ce4dcb97SMartin Matuska 			err = SET_ERROR(EINTR);
881ce4dcb97SMartin Matuska 			break;
882ce4dcb97SMartin Matuska 		}
883ce4dcb97SMartin Matuska 	}
884ce4dcb97SMartin Matuska 
885ce4dcb97SMartin Matuska 	dnode_rele(dn, FTAG);
886ce4dcb97SMartin Matuska 
887ce4dcb97SMartin Matuska 	return (err);
888ce4dcb97SMartin Matuska }
889ce4dcb97SMartin Matuska 
890eda14cbcSMatt Macy /*
891315ee00fSMartin Matuska  * Issue prefetch I/Os for the given object's dnode.
892315ee00fSMartin Matuska  */
893315ee00fSMartin Matuska void
894315ee00fSMartin Matuska dmu_prefetch_dnode(objset_t *os, uint64_t object, zio_priority_t pri)
895315ee00fSMartin Matuska {
896315ee00fSMartin Matuska 	if (object == 0 || object >= DN_MAX_OBJECT)
897315ee00fSMartin Matuska 		return;
898315ee00fSMartin Matuska 
899315ee00fSMartin Matuska 	dnode_t *dn = DMU_META_DNODE(os);
900315ee00fSMartin Matuska 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
901315ee00fSMartin Matuska 	uint64_t blkid = dbuf_whichblock(dn, 0, object * sizeof (dnode_phys_t));
902315ee00fSMartin Matuska 	dbuf_prefetch(dn, 0, blkid, pri, 0);
903315ee00fSMartin Matuska 	rw_exit(&dn->dn_struct_rwlock);
904315ee00fSMartin Matuska }
905315ee00fSMartin Matuska 
906315ee00fSMartin Matuska /*
907eda14cbcSMatt Macy  * Get the next "chunk" of file data to free.  We traverse the file from
9087a7741afSMartin Matuska  * the end so that the file gets shorter over time (if we crash in the
909eda14cbcSMatt Macy  * middle, this will leave us in a better state).  We find allocated file
910eda14cbcSMatt Macy  * data by simply searching the allocated level 1 indirects.
911eda14cbcSMatt Macy  *
912eda14cbcSMatt Macy  * On input, *start should be the first offset that does not need to be
913eda14cbcSMatt Macy  * freed (e.g. "offset + length").  On return, *start will be the first
914eda14cbcSMatt Macy  * offset that should be freed and l1blks is set to the number of level 1
915eda14cbcSMatt Macy  * indirect blocks found within the chunk.
916eda14cbcSMatt Macy  */
917eda14cbcSMatt Macy static int
918eda14cbcSMatt Macy get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks)
919eda14cbcSMatt Macy {
920eda14cbcSMatt Macy 	uint64_t blks;
921eda14cbcSMatt Macy 	uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
922eda14cbcSMatt Macy 	/* bytes of data covered by a level-1 indirect block */
923eda14cbcSMatt Macy 	uint64_t iblkrange = (uint64_t)dn->dn_datablksz *
924eda14cbcSMatt Macy 	    EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
925eda14cbcSMatt Macy 
926eda14cbcSMatt Macy 	ASSERT3U(minimum, <=, *start);
927eda14cbcSMatt Macy 
928ce4dcb97SMartin Matuska 	/* dn_nlevels == 1 means we don't have any L1 blocks */
929ce4dcb97SMartin Matuska 	if (dn->dn_nlevels <= 1) {
930ce4dcb97SMartin Matuska 		*l1blks = 0;
931ce4dcb97SMartin Matuska 		*start = minimum;
932ce4dcb97SMartin Matuska 		return (0);
933ce4dcb97SMartin Matuska 	}
934ce4dcb97SMartin Matuska 
935eda14cbcSMatt Macy 	/*
936eda14cbcSMatt Macy 	 * Check if we can free the entire range assuming that all of the
937eda14cbcSMatt Macy 	 * L1 blocks in this range have data. If we can, we use this
938eda14cbcSMatt Macy 	 * worst case value as an estimate so we can avoid having to look
939eda14cbcSMatt Macy 	 * at the object's actual data.
940eda14cbcSMatt Macy 	 */
941eda14cbcSMatt Macy 	uint64_t total_l1blks =
942eda14cbcSMatt Macy 	    (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) /
943eda14cbcSMatt Macy 	    iblkrange;
944eda14cbcSMatt Macy 	if (total_l1blks <= maxblks) {
945eda14cbcSMatt Macy 		*l1blks = total_l1blks;
946eda14cbcSMatt Macy 		*start = minimum;
947eda14cbcSMatt Macy 		return (0);
948eda14cbcSMatt Macy 	}
949eda14cbcSMatt Macy 	ASSERT(ISP2(iblkrange));
950eda14cbcSMatt Macy 
951eda14cbcSMatt Macy 	for (blks = 0; *start > minimum && blks < maxblks; blks++) {
952eda14cbcSMatt Macy 		int err;
953eda14cbcSMatt Macy 
954eda14cbcSMatt Macy 		/*
955eda14cbcSMatt Macy 		 * dnode_next_offset(BACKWARDS) will find an allocated L1
956eda14cbcSMatt Macy 		 * indirect block at or before the input offset.  We must
957eda14cbcSMatt Macy 		 * decrement *start so that it is at the end of the region
958eda14cbcSMatt Macy 		 * to search.
959eda14cbcSMatt Macy 		 */
960eda14cbcSMatt Macy 		(*start)--;
961eda14cbcSMatt Macy 
962eda14cbcSMatt Macy 		err = dnode_next_offset(dn,
963eda14cbcSMatt Macy 		    DNODE_FIND_BACKWARDS, start, 2, 1, 0);
964eda14cbcSMatt Macy 
965eda14cbcSMatt Macy 		/* if there are no indirect blocks before start, we are done */
966eda14cbcSMatt Macy 		if (err == ESRCH) {
967eda14cbcSMatt Macy 			*start = minimum;
968eda14cbcSMatt Macy 			break;
969eda14cbcSMatt Macy 		} else if (err != 0) {
970eda14cbcSMatt Macy 			*l1blks = blks;
971eda14cbcSMatt Macy 			return (err);
972eda14cbcSMatt Macy 		}
973eda14cbcSMatt Macy 
974eda14cbcSMatt Macy 		/* set start to the beginning of this L1 indirect */
975aca928a5SMartin Matuska 		*start = P2ALIGN_TYPED(*start, iblkrange, uint64_t);
976eda14cbcSMatt Macy 	}
977eda14cbcSMatt Macy 	if (*start < minimum)
978eda14cbcSMatt Macy 		*start = minimum;
979eda14cbcSMatt Macy 	*l1blks = blks;
980eda14cbcSMatt Macy 
981eda14cbcSMatt Macy 	return (0);
982eda14cbcSMatt Macy }
983eda14cbcSMatt Macy 
984eda14cbcSMatt Macy /*
985eda14cbcSMatt Macy  * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
986eda14cbcSMatt Macy  * otherwise return false.
987eda14cbcSMatt Macy  * Used below in dmu_free_long_range_impl() to enable abort when unmounting
988eda14cbcSMatt Macy  */
989eda14cbcSMatt Macy static boolean_t
990eda14cbcSMatt Macy dmu_objset_zfs_unmounting(objset_t *os)
991eda14cbcSMatt Macy {
992eda14cbcSMatt Macy #ifdef _KERNEL
993eda14cbcSMatt Macy 	if (dmu_objset_type(os) == DMU_OST_ZFS)
994eda14cbcSMatt Macy 		return (zfs_get_vfs_flag_unmounted(os));
995e92ffd9bSMartin Matuska #else
996e92ffd9bSMartin Matuska 	(void) os;
997eda14cbcSMatt Macy #endif
998eda14cbcSMatt Macy 	return (B_FALSE);
999eda14cbcSMatt Macy }
1000eda14cbcSMatt Macy 
1001eda14cbcSMatt Macy static int
1002eda14cbcSMatt Macy dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
1003eda14cbcSMatt Macy     uint64_t length)
1004eda14cbcSMatt Macy {
1005eda14cbcSMatt Macy 	uint64_t object_size;
1006eda14cbcSMatt Macy 	int err;
1007eda14cbcSMatt Macy 	uint64_t dirty_frees_threshold;
1008eda14cbcSMatt Macy 	dsl_pool_t *dp = dmu_objset_pool(os);
1009eda14cbcSMatt Macy 
1010eda14cbcSMatt Macy 	if (dn == NULL)
1011eda14cbcSMatt Macy 		return (SET_ERROR(EINVAL));
1012eda14cbcSMatt Macy 
1013eda14cbcSMatt Macy 	object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1014eda14cbcSMatt Macy 	if (offset >= object_size)
1015eda14cbcSMatt Macy 		return (0);
1016eda14cbcSMatt Macy 
1017eda14cbcSMatt Macy 	if (zfs_per_txg_dirty_frees_percent <= 100)
1018eda14cbcSMatt Macy 		dirty_frees_threshold =
1019eda14cbcSMatt Macy 		    zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
1020eda14cbcSMatt Macy 	else
1021eda14cbcSMatt Macy 		dirty_frees_threshold = zfs_dirty_data_max / 20;
1022eda14cbcSMatt Macy 
1023eda14cbcSMatt Macy 	if (length == DMU_OBJECT_END || offset + length > object_size)
1024eda14cbcSMatt Macy 		length = object_size - offset;
1025eda14cbcSMatt Macy 
1026eda14cbcSMatt Macy 	while (length != 0) {
1027eda14cbcSMatt Macy 		uint64_t chunk_end, chunk_begin, chunk_len;
1028eda14cbcSMatt Macy 		uint64_t l1blks;
1029eda14cbcSMatt Macy 		dmu_tx_t *tx;
1030eda14cbcSMatt Macy 
1031eda14cbcSMatt Macy 		if (dmu_objset_zfs_unmounting(dn->dn_objset))
1032eda14cbcSMatt Macy 			return (SET_ERROR(EINTR));
1033eda14cbcSMatt Macy 
1034eda14cbcSMatt Macy 		chunk_end = chunk_begin = offset + length;
1035eda14cbcSMatt Macy 
1036eda14cbcSMatt Macy 		/* move chunk_begin backwards to the beginning of this chunk */
1037eda14cbcSMatt Macy 		err = get_next_chunk(dn, &chunk_begin, offset, &l1blks);
1038eda14cbcSMatt Macy 		if (err)
1039eda14cbcSMatt Macy 			return (err);
1040eda14cbcSMatt Macy 		ASSERT3U(chunk_begin, >=, offset);
1041eda14cbcSMatt Macy 		ASSERT3U(chunk_begin, <=, chunk_end);
1042eda14cbcSMatt Macy 
1043eda14cbcSMatt Macy 		chunk_len = chunk_end - chunk_begin;
1044eda14cbcSMatt Macy 
1045eda14cbcSMatt Macy 		tx = dmu_tx_create(os);
1046eda14cbcSMatt Macy 		dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
1047eda14cbcSMatt Macy 
1048eda14cbcSMatt Macy 		/*
1049eda14cbcSMatt Macy 		 * Mark this transaction as typically resulting in a net
1050eda14cbcSMatt Macy 		 * reduction in space used.
1051eda14cbcSMatt Macy 		 */
1052eda14cbcSMatt Macy 		dmu_tx_mark_netfree(tx);
1053eda14cbcSMatt Macy 		err = dmu_tx_assign(tx, TXG_WAIT);
1054eda14cbcSMatt Macy 		if (err) {
1055eda14cbcSMatt Macy 			dmu_tx_abort(tx);
1056eda14cbcSMatt Macy 			return (err);
1057eda14cbcSMatt Macy 		}
1058eda14cbcSMatt Macy 
1059eda14cbcSMatt Macy 		uint64_t txg = dmu_tx_get_txg(tx);
1060eda14cbcSMatt Macy 
1061eda14cbcSMatt Macy 		mutex_enter(&dp->dp_lock);
1062eda14cbcSMatt Macy 		uint64_t long_free_dirty =
1063eda14cbcSMatt Macy 		    dp->dp_long_free_dirty_pertxg[txg & TXG_MASK];
1064eda14cbcSMatt Macy 		mutex_exit(&dp->dp_lock);
1065eda14cbcSMatt Macy 
1066eda14cbcSMatt Macy 		/*
1067eda14cbcSMatt Macy 		 * To avoid filling up a TXG with just frees, wait for
1068eda14cbcSMatt Macy 		 * the next TXG to open before freeing more chunks if
1069eda14cbcSMatt Macy 		 * we have reached the threshold of frees.
1070eda14cbcSMatt Macy 		 */
1071eda14cbcSMatt Macy 		if (dirty_frees_threshold != 0 &&
1072eda14cbcSMatt Macy 		    long_free_dirty >= dirty_frees_threshold) {
1073eda14cbcSMatt Macy 			DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay);
1074eda14cbcSMatt Macy 			dmu_tx_commit(tx);
1075eda14cbcSMatt Macy 			txg_wait_open(dp, 0, B_TRUE);
1076eda14cbcSMatt Macy 			continue;
1077eda14cbcSMatt Macy 		}
1078eda14cbcSMatt Macy 
1079eda14cbcSMatt Macy 		/*
1080eda14cbcSMatt Macy 		 * In order to prevent unnecessary write throttling, for each
1081eda14cbcSMatt Macy 		 * TXG, we track the cumulative size of L1 blocks being dirtied
1082eda14cbcSMatt Macy 		 * in dnode_free_range() below. We compare this number to a
1083eda14cbcSMatt Macy 		 * tunable threshold, past which we prevent new L1 dirty freeing
1084eda14cbcSMatt Macy 		 * blocks from being added into the open TXG. See
1085eda14cbcSMatt Macy 		 * dmu_free_long_range_impl() for details. The threshold
1086eda14cbcSMatt Macy 		 * prevents write throttle activation due to dirty freeing L1
1087eda14cbcSMatt Macy 		 * blocks taking up a large percentage of zfs_dirty_data_max.
1088eda14cbcSMatt Macy 		 */
1089eda14cbcSMatt Macy 		mutex_enter(&dp->dp_lock);
1090eda14cbcSMatt Macy 		dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] +=
1091eda14cbcSMatt Macy 		    l1blks << dn->dn_indblkshift;
1092eda14cbcSMatt Macy 		mutex_exit(&dp->dp_lock);
1093eda14cbcSMatt Macy 		DTRACE_PROBE3(free__long__range,
1094eda14cbcSMatt Macy 		    uint64_t, long_free_dirty, uint64_t, chunk_len,
1095eda14cbcSMatt Macy 		    uint64_t, txg);
1096eda14cbcSMatt Macy 		dnode_free_range(dn, chunk_begin, chunk_len, tx);
1097eda14cbcSMatt Macy 
1098eda14cbcSMatt Macy 		dmu_tx_commit(tx);
1099eda14cbcSMatt Macy 
1100eda14cbcSMatt Macy 		length -= chunk_len;
1101eda14cbcSMatt Macy 	}
1102eda14cbcSMatt Macy 	return (0);
1103eda14cbcSMatt Macy }
1104eda14cbcSMatt Macy 
1105eda14cbcSMatt Macy int
1106eda14cbcSMatt Macy dmu_free_long_range(objset_t *os, uint64_t object,
1107eda14cbcSMatt Macy     uint64_t offset, uint64_t length)
1108eda14cbcSMatt Macy {
1109eda14cbcSMatt Macy 	dnode_t *dn;
1110eda14cbcSMatt Macy 	int err;
1111eda14cbcSMatt Macy 
1112eda14cbcSMatt Macy 	err = dnode_hold(os, object, FTAG, &dn);
1113eda14cbcSMatt Macy 	if (err != 0)
1114eda14cbcSMatt Macy 		return (err);
1115eda14cbcSMatt Macy 	err = dmu_free_long_range_impl(os, dn, offset, length);
1116eda14cbcSMatt Macy 
1117eda14cbcSMatt Macy 	/*
1118eda14cbcSMatt Macy 	 * It is important to zero out the maxblkid when freeing the entire
1119eda14cbcSMatt Macy 	 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
1120eda14cbcSMatt Macy 	 * will take the fast path, and (b) dnode_reallocate() can verify
1121eda14cbcSMatt Macy 	 * that the entire file has been freed.
1122eda14cbcSMatt Macy 	 */
1123eda14cbcSMatt Macy 	if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
1124eda14cbcSMatt Macy 		dn->dn_maxblkid = 0;
1125eda14cbcSMatt Macy 
1126eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
1127eda14cbcSMatt Macy 	return (err);
1128eda14cbcSMatt Macy }
1129eda14cbcSMatt Macy 
1130eda14cbcSMatt Macy int
1131eda14cbcSMatt Macy dmu_free_long_object(objset_t *os, uint64_t object)
1132eda14cbcSMatt Macy {
1133eda14cbcSMatt Macy 	dmu_tx_t *tx;
1134eda14cbcSMatt Macy 	int err;
1135eda14cbcSMatt Macy 
1136eda14cbcSMatt Macy 	err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
1137eda14cbcSMatt Macy 	if (err != 0)
1138eda14cbcSMatt Macy 		return (err);
1139eda14cbcSMatt Macy 
1140eda14cbcSMatt Macy 	tx = dmu_tx_create(os);
1141eda14cbcSMatt Macy 	dmu_tx_hold_bonus(tx, object);
1142eda14cbcSMatt Macy 	dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
1143eda14cbcSMatt Macy 	dmu_tx_mark_netfree(tx);
1144eda14cbcSMatt Macy 	err = dmu_tx_assign(tx, TXG_WAIT);
1145eda14cbcSMatt Macy 	if (err == 0) {
1146eda14cbcSMatt Macy 		err = dmu_object_free(os, object, tx);
1147eda14cbcSMatt Macy 		dmu_tx_commit(tx);
1148eda14cbcSMatt Macy 	} else {
1149eda14cbcSMatt Macy 		dmu_tx_abort(tx);
1150eda14cbcSMatt Macy 	}
1151eda14cbcSMatt Macy 
1152eda14cbcSMatt Macy 	return (err);
1153eda14cbcSMatt Macy }
1154eda14cbcSMatt Macy 
1155eda14cbcSMatt Macy int
1156eda14cbcSMatt Macy dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
1157eda14cbcSMatt Macy     uint64_t size, dmu_tx_t *tx)
1158eda14cbcSMatt Macy {
1159eda14cbcSMatt Macy 	dnode_t *dn;
1160eda14cbcSMatt Macy 	int err = dnode_hold(os, object, FTAG, &dn);
1161eda14cbcSMatt Macy 	if (err)
1162eda14cbcSMatt Macy 		return (err);
1163eda14cbcSMatt Macy 	ASSERT(offset < UINT64_MAX);
1164eda14cbcSMatt Macy 	ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset);
1165eda14cbcSMatt Macy 	dnode_free_range(dn, offset, size, tx);
1166eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
1167eda14cbcSMatt Macy 	return (0);
1168eda14cbcSMatt Macy }
1169eda14cbcSMatt Macy 
1170eda14cbcSMatt Macy static int
1171eda14cbcSMatt Macy dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
1172eda14cbcSMatt Macy     void *buf, uint32_t flags)
1173eda14cbcSMatt Macy {
1174eda14cbcSMatt Macy 	dmu_buf_t **dbp;
1175eda14cbcSMatt Macy 	int numbufs, err = 0;
1176eda14cbcSMatt Macy 
1177eda14cbcSMatt Macy 	/*
1178eda14cbcSMatt Macy 	 * Deal with odd block sizes, where there can't be data past the first
1179eda14cbcSMatt Macy 	 * block. If we ever do the tail block optimization, we will need to
1180eda14cbcSMatt Macy 	 * handle that here as well.
1181eda14cbcSMatt Macy 	 */
1182eda14cbcSMatt Macy 	if (dn->dn_maxblkid == 0) {
1183eda14cbcSMatt Macy 		uint64_t newsz = offset > dn->dn_datablksz ? 0 :
1184eda14cbcSMatt Macy 		    MIN(size, dn->dn_datablksz - offset);
1185da5137abSMartin Matuska 		memset((char *)buf + newsz, 0, size - newsz);
1186eda14cbcSMatt Macy 		size = newsz;
1187eda14cbcSMatt Macy 	}
1188eda14cbcSMatt Macy 
11897a7741afSMartin Matuska 	if (size == 0)
11907a7741afSMartin Matuska 		return (0);
11917a7741afSMartin Matuska 
11927a7741afSMartin Matuska 	/* Allow Direct I/O when requested and properly aligned */
11937a7741afSMartin Matuska 	if ((flags & DMU_DIRECTIO) && zfs_dio_page_aligned(buf) &&
11947a7741afSMartin Matuska 	    zfs_dio_aligned(offset, size, PAGESIZE)) {
11957a7741afSMartin Matuska 		abd_t *data = abd_get_from_buf(buf, size);
11967a7741afSMartin Matuska 		err = dmu_read_abd(dn, offset, size, data, flags);
11977a7741afSMartin Matuska 		abd_free(data);
11987a7741afSMartin Matuska 		return (err);
11997a7741afSMartin Matuska 	}
12007a7741afSMartin Matuska 
1201eda14cbcSMatt Macy 	while (size > 0) {
1202eda14cbcSMatt Macy 		uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
1203eda14cbcSMatt Macy 		int i;
1204eda14cbcSMatt Macy 
1205eda14cbcSMatt Macy 		/*
1206eda14cbcSMatt Macy 		 * NB: we could do this block-at-a-time, but it's nice
1207eda14cbcSMatt Macy 		 * to be reading in parallel.
1208eda14cbcSMatt Macy 		 */
1209eda14cbcSMatt Macy 		err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
1210eda14cbcSMatt Macy 		    TRUE, FTAG, &numbufs, &dbp, flags);
1211eda14cbcSMatt Macy 		if (err)
1212eda14cbcSMatt Macy 			break;
1213eda14cbcSMatt Macy 
1214eda14cbcSMatt Macy 		for (i = 0; i < numbufs; i++) {
1215eda14cbcSMatt Macy 			uint64_t tocpy;
1216eda14cbcSMatt Macy 			int64_t bufoff;
1217eda14cbcSMatt Macy 			dmu_buf_t *db = dbp[i];
1218eda14cbcSMatt Macy 
1219eda14cbcSMatt Macy 			ASSERT(size > 0);
1220eda14cbcSMatt Macy 
1221eda14cbcSMatt Macy 			bufoff = offset - db->db_offset;
1222eda14cbcSMatt Macy 			tocpy = MIN(db->db_size - bufoff, size);
1223eda14cbcSMatt Macy 
1224*17aab35aSMartin Matuska 			ASSERT(db->db_data != NULL);
1225eda14cbcSMatt Macy 			(void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
1226eda14cbcSMatt Macy 
1227eda14cbcSMatt Macy 			offset += tocpy;
1228eda14cbcSMatt Macy 			size -= tocpy;
1229eda14cbcSMatt Macy 			buf = (char *)buf + tocpy;
1230eda14cbcSMatt Macy 		}
1231eda14cbcSMatt Macy 		dmu_buf_rele_array(dbp, numbufs, FTAG);
1232eda14cbcSMatt Macy 	}
1233eda14cbcSMatt Macy 	return (err);
1234eda14cbcSMatt Macy }
1235eda14cbcSMatt Macy 
1236eda14cbcSMatt Macy int
1237eda14cbcSMatt Macy dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1238eda14cbcSMatt Macy     void *buf, uint32_t flags)
1239eda14cbcSMatt Macy {
1240eda14cbcSMatt Macy 	dnode_t *dn;
1241eda14cbcSMatt Macy 	int err;
1242eda14cbcSMatt Macy 
1243eda14cbcSMatt Macy 	err = dnode_hold(os, object, FTAG, &dn);
1244eda14cbcSMatt Macy 	if (err != 0)
1245eda14cbcSMatt Macy 		return (err);
1246eda14cbcSMatt Macy 
1247eda14cbcSMatt Macy 	err = dmu_read_impl(dn, offset, size, buf, flags);
1248eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
1249eda14cbcSMatt Macy 	return (err);
1250eda14cbcSMatt Macy }
1251eda14cbcSMatt Macy 
1252eda14cbcSMatt Macy int
1253eda14cbcSMatt Macy dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
1254eda14cbcSMatt Macy     uint32_t flags)
1255eda14cbcSMatt Macy {
1256eda14cbcSMatt Macy 	return (dmu_read_impl(dn, offset, size, buf, flags));
1257eda14cbcSMatt Macy }
1258eda14cbcSMatt Macy 
1259eda14cbcSMatt Macy static void
1260eda14cbcSMatt Macy dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
1261eda14cbcSMatt Macy     const void *buf, dmu_tx_t *tx)
1262eda14cbcSMatt Macy {
1263eda14cbcSMatt Macy 	int i;
1264eda14cbcSMatt Macy 
1265eda14cbcSMatt Macy 	for (i = 0; i < numbufs; i++) {
1266eda14cbcSMatt Macy 		uint64_t tocpy;
1267eda14cbcSMatt Macy 		int64_t bufoff;
1268eda14cbcSMatt Macy 		dmu_buf_t *db = dbp[i];
1269eda14cbcSMatt Macy 
1270eda14cbcSMatt Macy 		ASSERT(size > 0);
1271eda14cbcSMatt Macy 
1272eda14cbcSMatt Macy 		bufoff = offset - db->db_offset;
1273eda14cbcSMatt Macy 		tocpy = MIN(db->db_size - bufoff, size);
1274eda14cbcSMatt Macy 
1275eda14cbcSMatt Macy 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1276eda14cbcSMatt Macy 
1277eda14cbcSMatt Macy 		if (tocpy == db->db_size)
1278188408daSMartin Matuska 			dmu_buf_will_fill(db, tx, B_FALSE);
1279eda14cbcSMatt Macy 		else
1280eda14cbcSMatt Macy 			dmu_buf_will_dirty(db, tx);
1281eda14cbcSMatt Macy 
1282*17aab35aSMartin Matuska 		ASSERT(db->db_data != NULL);
1283eda14cbcSMatt Macy 		(void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
1284eda14cbcSMatt Macy 
1285eda14cbcSMatt Macy 		if (tocpy == db->db_size)
1286188408daSMartin Matuska 			dmu_buf_fill_done(db, tx, B_FALSE);
1287eda14cbcSMatt Macy 
1288eda14cbcSMatt Macy 		offset += tocpy;
1289eda14cbcSMatt Macy 		size -= tocpy;
1290eda14cbcSMatt Macy 		buf = (char *)buf + tocpy;
1291eda14cbcSMatt Macy 	}
1292eda14cbcSMatt Macy }
1293eda14cbcSMatt Macy 
1294eda14cbcSMatt Macy void
1295eda14cbcSMatt Macy dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1296eda14cbcSMatt Macy     const void *buf, dmu_tx_t *tx)
1297eda14cbcSMatt Macy {
1298eda14cbcSMatt Macy 	dmu_buf_t **dbp;
1299eda14cbcSMatt Macy 	int numbufs;
1300eda14cbcSMatt Macy 
1301eda14cbcSMatt Macy 	if (size == 0)
1302eda14cbcSMatt Macy 		return;
1303eda14cbcSMatt Macy 
1304eda14cbcSMatt Macy 	VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1305eda14cbcSMatt Macy 	    FALSE, FTAG, &numbufs, &dbp));
1306eda14cbcSMatt Macy 	dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1307eda14cbcSMatt Macy 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1308eda14cbcSMatt Macy }
1309eda14cbcSMatt Macy 
1310eda14cbcSMatt Macy /*
13117a7741afSMartin Matuska  * This interface is not used internally by ZFS but is provided for
13127a7741afSMartin Matuska  * use by Lustre which is built on the DMU interfaces.
1313eda14cbcSMatt Macy  */
13147a7741afSMartin Matuska int
13157a7741afSMartin Matuska dmu_write_by_dnode_flags(dnode_t *dn, uint64_t offset, uint64_t size,
13167a7741afSMartin Matuska     const void *buf, dmu_tx_t *tx, uint32_t flags)
1317eda14cbcSMatt Macy {
1318eda14cbcSMatt Macy 	dmu_buf_t **dbp;
1319eda14cbcSMatt Macy 	int numbufs;
13207a7741afSMartin Matuska 	int error;
1321eda14cbcSMatt Macy 
1322eda14cbcSMatt Macy 	if (size == 0)
13237a7741afSMartin Matuska 		return (0);
13247a7741afSMartin Matuska 
13257a7741afSMartin Matuska 	/* Allow Direct I/O when requested and properly aligned */
13267a7741afSMartin Matuska 	if ((flags & DMU_DIRECTIO) && zfs_dio_page_aligned((void *)buf) &&
13277a7741afSMartin Matuska 	    zfs_dio_aligned(offset, size, dn->dn_datablksz)) {
13287a7741afSMartin Matuska 		abd_t *data = abd_get_from_buf((void *)buf, size);
13297a7741afSMartin Matuska 		error = dmu_write_abd(dn, offset, size, data, DMU_DIRECTIO, tx);
13307a7741afSMartin Matuska 		abd_free(data);
13317a7741afSMartin Matuska 		return (error);
13327a7741afSMartin Matuska 	}
1333eda14cbcSMatt Macy 
1334eda14cbcSMatt Macy 	VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1335eda14cbcSMatt Macy 	    FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1336eda14cbcSMatt Macy 	dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1337eda14cbcSMatt Macy 	dmu_buf_rele_array(dbp, numbufs, FTAG);
13387a7741afSMartin Matuska 	return (0);
13397a7741afSMartin Matuska }
13407a7741afSMartin Matuska 
13417a7741afSMartin Matuska int
13427a7741afSMartin Matuska dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
13437a7741afSMartin Matuska     const void *buf, dmu_tx_t *tx)
13447a7741afSMartin Matuska {
13457a7741afSMartin Matuska 	return (dmu_write_by_dnode_flags(dn, offset, size, buf, tx, 0));
1346eda14cbcSMatt Macy }
1347eda14cbcSMatt Macy 
1348eda14cbcSMatt Macy void
1349eda14cbcSMatt Macy dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1350eda14cbcSMatt Macy     dmu_tx_t *tx)
1351eda14cbcSMatt Macy {
1352eda14cbcSMatt Macy 	dmu_buf_t **dbp;
1353eda14cbcSMatt Macy 	int numbufs, i;
1354eda14cbcSMatt Macy 
1355eda14cbcSMatt Macy 	if (size == 0)
1356eda14cbcSMatt Macy 		return;
1357eda14cbcSMatt Macy 
1358eda14cbcSMatt Macy 	VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1359eda14cbcSMatt Macy 	    FALSE, FTAG, &numbufs, &dbp));
1360eda14cbcSMatt Macy 
1361eda14cbcSMatt Macy 	for (i = 0; i < numbufs; i++) {
1362eda14cbcSMatt Macy 		dmu_buf_t *db = dbp[i];
1363eda14cbcSMatt Macy 
1364eda14cbcSMatt Macy 		dmu_buf_will_not_fill(db, tx);
1365eda14cbcSMatt Macy 	}
1366eda14cbcSMatt Macy 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1367eda14cbcSMatt Macy }
1368eda14cbcSMatt Macy 
1369eda14cbcSMatt Macy void
1370eda14cbcSMatt Macy dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1371eda14cbcSMatt Macy     void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1372eda14cbcSMatt Macy     int compressed_size, int byteorder, dmu_tx_t *tx)
1373eda14cbcSMatt Macy {
1374eda14cbcSMatt Macy 	dmu_buf_t *db;
1375eda14cbcSMatt Macy 
1376eda14cbcSMatt Macy 	ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1377eda14cbcSMatt Macy 	ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1378eda14cbcSMatt Macy 	VERIFY0(dmu_buf_hold_noread(os, object, offset,
1379eda14cbcSMatt Macy 	    FTAG, &db));
1380eda14cbcSMatt Macy 
1381eda14cbcSMatt Macy 	dmu_buf_write_embedded(db,
1382eda14cbcSMatt Macy 	    data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1383eda14cbcSMatt Macy 	    uncompressed_size, compressed_size, byteorder, tx);
1384eda14cbcSMatt Macy 
1385eda14cbcSMatt Macy 	dmu_buf_rele(db, FTAG);
1386eda14cbcSMatt Macy }
1387eda14cbcSMatt Macy 
1388eda14cbcSMatt Macy void
1389eda14cbcSMatt Macy dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1390eda14cbcSMatt Macy     dmu_tx_t *tx)
1391eda14cbcSMatt Macy {
1392eda14cbcSMatt Macy 	int numbufs, i;
1393eda14cbcSMatt Macy 	dmu_buf_t **dbp;
1394eda14cbcSMatt Macy 
1395eda14cbcSMatt Macy 	VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG,
1396eda14cbcSMatt Macy 	    &numbufs, &dbp));
1397eda14cbcSMatt Macy 	for (i = 0; i < numbufs; i++)
1398eda14cbcSMatt Macy 		dmu_buf_redact(dbp[i], tx);
1399eda14cbcSMatt Macy 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1400eda14cbcSMatt Macy }
1401eda14cbcSMatt Macy 
1402eda14cbcSMatt Macy #ifdef _KERNEL
1403eda14cbcSMatt Macy int
1404184c1b94SMartin Matuska dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size)
1405eda14cbcSMatt Macy {
1406eda14cbcSMatt Macy 	dmu_buf_t **dbp;
1407eda14cbcSMatt Macy 	int numbufs, i, err;
1408eda14cbcSMatt Macy 
14097a7741afSMartin Matuska 	if (uio->uio_extflg & UIO_DIRECT)
14107a7741afSMartin Matuska 		return (dmu_read_uio_direct(dn, uio, size));
14117a7741afSMartin Matuska 
1412eda14cbcSMatt Macy 	/*
1413eda14cbcSMatt Macy 	 * NB: we could do this block-at-a-time, but it's nice
1414eda14cbcSMatt Macy 	 * to be reading in parallel.
1415eda14cbcSMatt Macy 	 */
1416184c1b94SMartin Matuska 	err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1417eda14cbcSMatt Macy 	    TRUE, FTAG, &numbufs, &dbp, 0);
1418eda14cbcSMatt Macy 	if (err)
1419eda14cbcSMatt Macy 		return (err);
1420eda14cbcSMatt Macy 
1421eda14cbcSMatt Macy 	for (i = 0; i < numbufs; i++) {
1422eda14cbcSMatt Macy 		uint64_t tocpy;
1423eda14cbcSMatt Macy 		int64_t bufoff;
1424eda14cbcSMatt Macy 		dmu_buf_t *db = dbp[i];
1425eda14cbcSMatt Macy 
1426eda14cbcSMatt Macy 		ASSERT(size > 0);
1427eda14cbcSMatt Macy 
1428184c1b94SMartin Matuska 		bufoff = zfs_uio_offset(uio) - db->db_offset;
1429eda14cbcSMatt Macy 		tocpy = MIN(db->db_size - bufoff, size);
1430eda14cbcSMatt Macy 
1431*17aab35aSMartin Matuska 		ASSERT(db->db_data != NULL);
1432184c1b94SMartin Matuska 		err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy,
1433eda14cbcSMatt Macy 		    UIO_READ, uio);
1434184c1b94SMartin Matuska 
1435eda14cbcSMatt Macy 		if (err)
1436eda14cbcSMatt Macy 			break;
1437eda14cbcSMatt Macy 
1438eda14cbcSMatt Macy 		size -= tocpy;
1439eda14cbcSMatt Macy 	}
1440eda14cbcSMatt Macy 	dmu_buf_rele_array(dbp, numbufs, FTAG);
1441eda14cbcSMatt Macy 
1442eda14cbcSMatt Macy 	return (err);
1443eda14cbcSMatt Macy }
1444eda14cbcSMatt Macy 
1445eda14cbcSMatt Macy /*
1446eda14cbcSMatt Macy  * Read 'size' bytes into the uio buffer.
1447eda14cbcSMatt Macy  * From object zdb->db_object.
1448184c1b94SMartin Matuska  * Starting at zfs_uio_offset(uio).
1449eda14cbcSMatt Macy  *
1450eda14cbcSMatt Macy  * If the caller already has a dbuf in the target object
1451eda14cbcSMatt Macy  * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1452eda14cbcSMatt Macy  * because we don't have to find the dnode_t for the object.
1453eda14cbcSMatt Macy  */
1454eda14cbcSMatt Macy int
1455184c1b94SMartin Matuska dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size)
1456eda14cbcSMatt Macy {
1457eda14cbcSMatt Macy 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1458eda14cbcSMatt Macy 	int err;
1459eda14cbcSMatt Macy 
1460eda14cbcSMatt Macy 	if (size == 0)
1461eda14cbcSMatt Macy 		return (0);
1462eda14cbcSMatt Macy 
1463eda14cbcSMatt Macy 	DB_DNODE_ENTER(db);
1464ce4dcb97SMartin Matuska 	err = dmu_read_uio_dnode(DB_DNODE(db), uio, size);
1465eda14cbcSMatt Macy 	DB_DNODE_EXIT(db);
1466eda14cbcSMatt Macy 
1467eda14cbcSMatt Macy 	return (err);
1468eda14cbcSMatt Macy }
1469eda14cbcSMatt Macy 
1470eda14cbcSMatt Macy /*
1471eda14cbcSMatt Macy  * Read 'size' bytes into the uio buffer.
1472eda14cbcSMatt Macy  * From the specified object
1473184c1b94SMartin Matuska  * Starting at offset zfs_uio_offset(uio).
1474eda14cbcSMatt Macy  */
1475eda14cbcSMatt Macy int
1476184c1b94SMartin Matuska dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size)
1477eda14cbcSMatt Macy {
1478eda14cbcSMatt Macy 	dnode_t *dn;
1479eda14cbcSMatt Macy 	int err;
1480eda14cbcSMatt Macy 
1481eda14cbcSMatt Macy 	if (size == 0)
1482eda14cbcSMatt Macy 		return (0);
1483eda14cbcSMatt Macy 
1484eda14cbcSMatt Macy 	err = dnode_hold(os, object, FTAG, &dn);
1485eda14cbcSMatt Macy 	if (err)
1486eda14cbcSMatt Macy 		return (err);
1487eda14cbcSMatt Macy 
1488eda14cbcSMatt Macy 	err = dmu_read_uio_dnode(dn, uio, size);
1489eda14cbcSMatt Macy 
1490eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
1491eda14cbcSMatt Macy 
1492eda14cbcSMatt Macy 	return (err);
1493eda14cbcSMatt Macy }
1494eda14cbcSMatt Macy 
1495eda14cbcSMatt Macy int
1496184c1b94SMartin Matuska dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx)
1497eda14cbcSMatt Macy {
1498eda14cbcSMatt Macy 	dmu_buf_t **dbp;
1499eda14cbcSMatt Macy 	int numbufs;
1500eda14cbcSMatt Macy 	int err = 0;
15017a7741afSMartin Matuska 	uint64_t write_size;
1502eda14cbcSMatt Macy 
15037a7741afSMartin Matuska top:
15047a7741afSMartin Matuska 	write_size = size;
15057a7741afSMartin Matuska 
15067a7741afSMartin Matuska 	/*
15077a7741afSMartin Matuska 	 * We only allow Direct I/O writes to happen if we are block
15087a7741afSMartin Matuska 	 * sized aligned. Otherwise, we pass the write off to the ARC.
15097a7741afSMartin Matuska 	 */
15107a7741afSMartin Matuska 	if ((uio->uio_extflg & UIO_DIRECT) &&
15117a7741afSMartin Matuska 	    (write_size >= dn->dn_datablksz)) {
15127a7741afSMartin Matuska 		if (zfs_dio_aligned(zfs_uio_offset(uio), write_size,
15137a7741afSMartin Matuska 		    dn->dn_datablksz)) {
15147a7741afSMartin Matuska 			return (dmu_write_uio_direct(dn, uio, size, tx));
15157a7741afSMartin Matuska 		} else if (write_size > dn->dn_datablksz &&
15167a7741afSMartin Matuska 		    zfs_dio_offset_aligned(zfs_uio_offset(uio),
15177a7741afSMartin Matuska 		    dn->dn_datablksz)) {
15187a7741afSMartin Matuska 			write_size =
15197a7741afSMartin Matuska 			    dn->dn_datablksz * (write_size / dn->dn_datablksz);
15207a7741afSMartin Matuska 			err = dmu_write_uio_direct(dn, uio, write_size, tx);
15217a7741afSMartin Matuska 			if (err == 0) {
15227a7741afSMartin Matuska 				size -= write_size;
15237a7741afSMartin Matuska 				goto top;
15247a7741afSMartin Matuska 			} else {
15257a7741afSMartin Matuska 				return (err);
15267a7741afSMartin Matuska 			}
15277a7741afSMartin Matuska 		} else {
15287a7741afSMartin Matuska 			write_size =
15297a7741afSMartin Matuska 			    P2PHASE(zfs_uio_offset(uio), dn->dn_datablksz);
15307a7741afSMartin Matuska 		}
15317a7741afSMartin Matuska 	}
15327a7741afSMartin Matuska 
15337a7741afSMartin Matuska 	err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), write_size,
1534eda14cbcSMatt Macy 	    FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1535eda14cbcSMatt Macy 	if (err)
1536eda14cbcSMatt Macy 		return (err);
1537eda14cbcSMatt Macy 
15387a7741afSMartin Matuska 	for (int i = 0; i < numbufs; i++) {
1539eda14cbcSMatt Macy 		uint64_t tocpy;
1540eda14cbcSMatt Macy 		int64_t bufoff;
1541eda14cbcSMatt Macy 		dmu_buf_t *db = dbp[i];
1542eda14cbcSMatt Macy 
15437a7741afSMartin Matuska 		ASSERT(write_size > 0);
1544eda14cbcSMatt Macy 
1545188408daSMartin Matuska 		offset_t off = zfs_uio_offset(uio);
1546188408daSMartin Matuska 		bufoff = off - db->db_offset;
15477a7741afSMartin Matuska 		tocpy = MIN(db->db_size - bufoff, write_size);
1548eda14cbcSMatt Macy 
1549eda14cbcSMatt Macy 		ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1550eda14cbcSMatt Macy 
1551eda14cbcSMatt Macy 		if (tocpy == db->db_size)
1552188408daSMartin Matuska 			dmu_buf_will_fill(db, tx, B_TRUE);
1553eda14cbcSMatt Macy 		else
1554eda14cbcSMatt Macy 			dmu_buf_will_dirty(db, tx);
1555eda14cbcSMatt Macy 
1556*17aab35aSMartin Matuska 		ASSERT(db->db_data != NULL);
1557184c1b94SMartin Matuska 		err = zfs_uio_fault_move((char *)db->db_data + bufoff,
1558184c1b94SMartin Matuska 		    tocpy, UIO_WRITE, uio);
1559184c1b94SMartin Matuska 
1560188408daSMartin Matuska 		if (tocpy == db->db_size && dmu_buf_fill_done(db, tx, err)) {
1561188408daSMartin Matuska 			/* The fill was reverted.  Undo any uio progress. */
1562188408daSMartin Matuska 			zfs_uio_advance(uio, off - zfs_uio_offset(uio));
1563188408daSMartin Matuska 		}
1564eda14cbcSMatt Macy 
1565eda14cbcSMatt Macy 		if (err)
1566eda14cbcSMatt Macy 			break;
1567eda14cbcSMatt Macy 
15687a7741afSMartin Matuska 		write_size -= tocpy;
1569eda14cbcSMatt Macy 		size -= tocpy;
1570eda14cbcSMatt Macy 	}
1571eda14cbcSMatt Macy 
15727a7741afSMartin Matuska 	IMPLY(err == 0, write_size == 0);
15737a7741afSMartin Matuska 
1574eda14cbcSMatt Macy 	dmu_buf_rele_array(dbp, numbufs, FTAG);
15757a7741afSMartin Matuska 
15767a7741afSMartin Matuska 	if ((uio->uio_extflg & UIO_DIRECT) && size > 0) {
15777a7741afSMartin Matuska 		goto top;
15787a7741afSMartin Matuska 	}
15797a7741afSMartin Matuska 
1580eda14cbcSMatt Macy 	return (err);
1581eda14cbcSMatt Macy }
1582eda14cbcSMatt Macy 
1583eda14cbcSMatt Macy /*
1584eda14cbcSMatt Macy  * Write 'size' bytes from the uio buffer.
1585eda14cbcSMatt Macy  * To object zdb->db_object.
1586184c1b94SMartin Matuska  * Starting at offset zfs_uio_offset(uio).
1587eda14cbcSMatt Macy  *
1588eda14cbcSMatt Macy  * If the caller already has a dbuf in the target object
1589eda14cbcSMatt Macy  * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1590eda14cbcSMatt Macy  * because we don't have to find the dnode_t for the object.
1591eda14cbcSMatt Macy  */
1592eda14cbcSMatt Macy int
1593184c1b94SMartin Matuska dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size,
1594eda14cbcSMatt Macy     dmu_tx_t *tx)
1595eda14cbcSMatt Macy {
1596eda14cbcSMatt Macy 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1597eda14cbcSMatt Macy 	int err;
1598eda14cbcSMatt Macy 
1599eda14cbcSMatt Macy 	if (size == 0)
1600eda14cbcSMatt Macy 		return (0);
1601eda14cbcSMatt Macy 
1602eda14cbcSMatt Macy 	DB_DNODE_ENTER(db);
1603ce4dcb97SMartin Matuska 	err = dmu_write_uio_dnode(DB_DNODE(db), uio, size, tx);
1604eda14cbcSMatt Macy 	DB_DNODE_EXIT(db);
1605eda14cbcSMatt Macy 
1606eda14cbcSMatt Macy 	return (err);
1607eda14cbcSMatt Macy }
1608eda14cbcSMatt Macy 
1609eda14cbcSMatt Macy /*
1610eda14cbcSMatt Macy  * Write 'size' bytes from the uio buffer.
1611eda14cbcSMatt Macy  * To the specified object.
1612184c1b94SMartin Matuska  * Starting at offset zfs_uio_offset(uio).
1613eda14cbcSMatt Macy  */
1614eda14cbcSMatt Macy int
1615184c1b94SMartin Matuska dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size,
1616eda14cbcSMatt Macy     dmu_tx_t *tx)
1617eda14cbcSMatt Macy {
1618eda14cbcSMatt Macy 	dnode_t *dn;
1619eda14cbcSMatt Macy 	int err;
1620eda14cbcSMatt Macy 
1621eda14cbcSMatt Macy 	if (size == 0)
1622eda14cbcSMatt Macy 		return (0);
1623eda14cbcSMatt Macy 
1624eda14cbcSMatt Macy 	err = dnode_hold(os, object, FTAG, &dn);
1625eda14cbcSMatt Macy 	if (err)
1626eda14cbcSMatt Macy 		return (err);
1627eda14cbcSMatt Macy 
1628eda14cbcSMatt Macy 	err = dmu_write_uio_dnode(dn, uio, size, tx);
1629eda14cbcSMatt Macy 
1630eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
1631eda14cbcSMatt Macy 
1632eda14cbcSMatt Macy 	return (err);
1633eda14cbcSMatt Macy }
1634eda14cbcSMatt Macy #endif /* _KERNEL */
1635eda14cbcSMatt Macy 
1636ce4dcb97SMartin Matuska static void
1637ce4dcb97SMartin Matuska dmu_cached_bps(spa_t *spa, blkptr_t *bps, uint_t nbps,
1638ce4dcb97SMartin Matuska     uint64_t *l1sz, uint64_t *l2sz)
1639ce4dcb97SMartin Matuska {
1640ce4dcb97SMartin Matuska 	int cached_flags;
1641ce4dcb97SMartin Matuska 
1642ce4dcb97SMartin Matuska 	if (bps == NULL)
1643ce4dcb97SMartin Matuska 		return;
1644ce4dcb97SMartin Matuska 
1645ce4dcb97SMartin Matuska 	for (size_t blk_off = 0; blk_off < nbps; blk_off++) {
1646ce4dcb97SMartin Matuska 		blkptr_t *bp = &bps[blk_off];
1647ce4dcb97SMartin Matuska 
1648ce4dcb97SMartin Matuska 		if (BP_IS_HOLE(bp))
1649ce4dcb97SMartin Matuska 			continue;
1650ce4dcb97SMartin Matuska 
1651ce4dcb97SMartin Matuska 		cached_flags = arc_cached(spa, bp);
1652ce4dcb97SMartin Matuska 		if (cached_flags == 0)
1653ce4dcb97SMartin Matuska 			continue;
1654ce4dcb97SMartin Matuska 
1655ce4dcb97SMartin Matuska 		if ((cached_flags & (ARC_CACHED_IN_L1 | ARC_CACHED_IN_L2)) ==
1656ce4dcb97SMartin Matuska 		    ARC_CACHED_IN_L2)
1657ce4dcb97SMartin Matuska 			*l2sz += BP_GET_LSIZE(bp);
1658ce4dcb97SMartin Matuska 		else
1659ce4dcb97SMartin Matuska 			*l1sz += BP_GET_LSIZE(bp);
1660ce4dcb97SMartin Matuska 	}
1661ce4dcb97SMartin Matuska }
1662ce4dcb97SMartin Matuska 
1663ce4dcb97SMartin Matuska /*
1664ce4dcb97SMartin Matuska  * Estimate DMU object cached size.
1665ce4dcb97SMartin Matuska  */
1666ce4dcb97SMartin Matuska int
1667ce4dcb97SMartin Matuska dmu_object_cached_size(objset_t *os, uint64_t object,
1668ce4dcb97SMartin Matuska     uint64_t *l1sz, uint64_t *l2sz)
1669ce4dcb97SMartin Matuska {
1670ce4dcb97SMartin Matuska 	dnode_t *dn;
1671ce4dcb97SMartin Matuska 	dmu_object_info_t doi;
1672ce4dcb97SMartin Matuska 	int err = 0;
1673ce4dcb97SMartin Matuska 
1674ce4dcb97SMartin Matuska 	*l1sz = *l2sz = 0;
1675ce4dcb97SMartin Matuska 
1676ce4dcb97SMartin Matuska 	if (dnode_hold(os, object, FTAG, &dn) != 0)
1677ce4dcb97SMartin Matuska 		return (0);
1678ce4dcb97SMartin Matuska 
1679ce4dcb97SMartin Matuska 	if (dn->dn_nlevels < 2) {
1680ce4dcb97SMartin Matuska 		dnode_rele(dn, FTAG);
1681ce4dcb97SMartin Matuska 		return (0);
1682ce4dcb97SMartin Matuska 	}
1683ce4dcb97SMartin Matuska 
1684ce4dcb97SMartin Matuska 	dmu_object_info_from_dnode(dn, &doi);
1685ce4dcb97SMartin Matuska 
1686ce4dcb97SMartin Matuska 	for (uint64_t off = 0; off < doi.doi_max_offset;
1687ce4dcb97SMartin Matuska 	    off += dmu_prefetch_max) {
1688ce4dcb97SMartin Matuska 		/* dbuf_read doesn't prefetch L1 blocks. */
1689ce4dcb97SMartin Matuska 		dmu_prefetch_by_dnode(dn, 1, off,
1690ce4dcb97SMartin Matuska 		    dmu_prefetch_max, ZIO_PRIORITY_SYNC_READ);
1691ce4dcb97SMartin Matuska 	}
1692ce4dcb97SMartin Matuska 
1693ce4dcb97SMartin Matuska 	/*
1694ce4dcb97SMartin Matuska 	 * Hold all valid L1 blocks, asking ARC the status of each BP
1695ce4dcb97SMartin Matuska 	 * contained in each such L1 block.
1696ce4dcb97SMartin Matuska 	 */
1697ce4dcb97SMartin Matuska 	uint_t nbps = bp_span_in_blocks(dn->dn_indblkshift, 1);
1698ce4dcb97SMartin Matuska 	uint64_t l1blks = 1 + (dn->dn_maxblkid / nbps);
1699ce4dcb97SMartin Matuska 
1700ce4dcb97SMartin Matuska 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1701ce4dcb97SMartin Matuska 	for (uint64_t blk = 0; blk < l1blks; blk++) {
1702ce4dcb97SMartin Matuska 		dmu_buf_impl_t *db = NULL;
1703ce4dcb97SMartin Matuska 
1704ce4dcb97SMartin Matuska 		if (issig()) {
1705ce4dcb97SMartin Matuska 			/*
1706ce4dcb97SMartin Matuska 			 * On interrupt, get out, and bubble up EINTR
1707ce4dcb97SMartin Matuska 			 */
1708ce4dcb97SMartin Matuska 			err = EINTR;
1709ce4dcb97SMartin Matuska 			break;
1710ce4dcb97SMartin Matuska 		}
1711ce4dcb97SMartin Matuska 
1712ce4dcb97SMartin Matuska 		/*
1713ce4dcb97SMartin Matuska 		 * If we get an i/o error here, the L1 can't be read,
1714ce4dcb97SMartin Matuska 		 * and nothing under it could be cached, so we just
1715ce4dcb97SMartin Matuska 		 * continue. Ignoring the error from dbuf_hold_impl
1716ce4dcb97SMartin Matuska 		 * or from dbuf_read is then a reasonable choice.
1717ce4dcb97SMartin Matuska 		 */
1718ce4dcb97SMartin Matuska 		err = dbuf_hold_impl(dn, 1, blk, B_TRUE, B_FALSE, FTAG, &db);
1719ce4dcb97SMartin Matuska 		if (err != 0) {
1720ce4dcb97SMartin Matuska 			/*
1721ce4dcb97SMartin Matuska 			 * ignore error and continue
1722ce4dcb97SMartin Matuska 			 */
1723ce4dcb97SMartin Matuska 			err = 0;
1724ce4dcb97SMartin Matuska 			continue;
1725ce4dcb97SMartin Matuska 		}
1726ce4dcb97SMartin Matuska 
1727ce4dcb97SMartin Matuska 		err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1728ce4dcb97SMartin Matuska 		if (err == 0) {
1729ce4dcb97SMartin Matuska 			dmu_cached_bps(dmu_objset_spa(os), db->db.db_data,
1730ce4dcb97SMartin Matuska 			    nbps, l1sz, l2sz);
1731ce4dcb97SMartin Matuska 		}
1732ce4dcb97SMartin Matuska 		/*
1733ce4dcb97SMartin Matuska 		 * error may be ignored, and we continue
1734ce4dcb97SMartin Matuska 		 */
1735ce4dcb97SMartin Matuska 		err = 0;
1736ce4dcb97SMartin Matuska 		dbuf_rele(db, FTAG);
1737ce4dcb97SMartin Matuska 	}
1738ce4dcb97SMartin Matuska 	rw_exit(&dn->dn_struct_rwlock);
1739ce4dcb97SMartin Matuska 
1740ce4dcb97SMartin Matuska 	dnode_rele(dn, FTAG);
1741ce4dcb97SMartin Matuska 	return (err);
1742ce4dcb97SMartin Matuska }
1743ce4dcb97SMartin Matuska 
1744eda14cbcSMatt Macy /*
1745eda14cbcSMatt Macy  * Allocate a loaned anonymous arc buffer.
1746eda14cbcSMatt Macy  */
1747eda14cbcSMatt Macy arc_buf_t *
1748eda14cbcSMatt Macy dmu_request_arcbuf(dmu_buf_t *handle, int size)
1749eda14cbcSMatt Macy {
1750eda14cbcSMatt Macy 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1751eda14cbcSMatt Macy 
1752eda14cbcSMatt Macy 	return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1753eda14cbcSMatt Macy }
1754eda14cbcSMatt Macy 
1755eda14cbcSMatt Macy /*
1756eda14cbcSMatt Macy  * Free a loaned arc buffer.
1757eda14cbcSMatt Macy  */
1758eda14cbcSMatt Macy void
1759eda14cbcSMatt Macy dmu_return_arcbuf(arc_buf_t *buf)
1760eda14cbcSMatt Macy {
1761eda14cbcSMatt Macy 	arc_return_buf(buf, FTAG);
1762eda14cbcSMatt Macy 	arc_buf_destroy(buf, FTAG);
1763eda14cbcSMatt Macy }
1764eda14cbcSMatt Macy 
1765eda14cbcSMatt Macy /*
17667877fdebSMatt Macy  * A "lightweight" write is faster than a regular write (e.g.
17677877fdebSMatt Macy  * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the
17687877fdebSMatt Macy  * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t.  However, the
17697877fdebSMatt Macy  * data can not be read or overwritten until the transaction's txg has been
17707877fdebSMatt Macy  * synced.  This makes it appropriate for workloads that are known to be
17717877fdebSMatt Macy  * (temporarily) write-only, like "zfs receive".
17727877fdebSMatt Macy  *
17737877fdebSMatt Macy  * A single block is written, starting at the specified offset in bytes.  If
17747877fdebSMatt Macy  * the call is successful, it returns 0 and the provided abd has been
17757877fdebSMatt Macy  * consumed (the caller should not free it).
17767877fdebSMatt Macy  */
17777877fdebSMatt Macy int
17787877fdebSMatt Macy dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd,
1779dbd5678dSMartin Matuska     const zio_prop_t *zp, zio_flag_t flags, dmu_tx_t *tx)
17807877fdebSMatt Macy {
17817877fdebSMatt Macy 	dbuf_dirty_record_t *dr =
17827877fdebSMatt Macy 	    dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx);
17837877fdebSMatt Macy 	if (dr == NULL)
17847877fdebSMatt Macy 		return (SET_ERROR(EIO));
17857877fdebSMatt Macy 	dr->dt.dll.dr_abd = abd;
17867877fdebSMatt Macy 	dr->dt.dll.dr_props = *zp;
17877877fdebSMatt Macy 	dr->dt.dll.dr_flags = flags;
17887877fdebSMatt Macy 	return (0);
17897877fdebSMatt Macy }
17907877fdebSMatt Macy 
17917877fdebSMatt Macy /*
1792eda14cbcSMatt Macy  * When possible directly assign passed loaned arc buffer to a dbuf.
1793eda14cbcSMatt Macy  * If this is not possible copy the contents of passed arc buf via
1794eda14cbcSMatt Macy  * dmu_write().
1795eda14cbcSMatt Macy  */
1796eda14cbcSMatt Macy int
1797eda14cbcSMatt Macy dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1798eda14cbcSMatt Macy     dmu_tx_t *tx)
1799eda14cbcSMatt Macy {
1800eda14cbcSMatt Macy 	dmu_buf_impl_t *db;
1801eda14cbcSMatt Macy 	objset_t *os = dn->dn_objset;
1802eda14cbcSMatt Macy 	uint64_t object = dn->dn_object;
1803eda14cbcSMatt Macy 	uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1804eda14cbcSMatt Macy 	uint64_t blkid;
1805eda14cbcSMatt Macy 
1806eda14cbcSMatt Macy 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
1807eda14cbcSMatt Macy 	blkid = dbuf_whichblock(dn, 0, offset);
1808eda14cbcSMatt Macy 	db = dbuf_hold(dn, blkid, FTAG);
18095fb307d2SMartin Matuska 	rw_exit(&dn->dn_struct_rwlock);
1810eda14cbcSMatt Macy 	if (db == NULL)
1811eda14cbcSMatt Macy 		return (SET_ERROR(EIO));
1812eda14cbcSMatt Macy 
1813eda14cbcSMatt Macy 	/*
18147877fdebSMatt Macy 	 * We can only assign if the offset is aligned and the arc buf is the
18157877fdebSMatt Macy 	 * same size as the dbuf.
1816eda14cbcSMatt Macy 	 */
1817eda14cbcSMatt Macy 	if (offset == db->db.db_offset && blksz == db->db.db_size) {
18187a7741afSMartin Matuska 		zfs_racct_write(os->os_spa, blksz, 1, 0);
1819eda14cbcSMatt Macy 		dbuf_assign_arcbuf(db, buf, tx);
1820eda14cbcSMatt Macy 		dbuf_rele(db, FTAG);
1821eda14cbcSMatt Macy 	} else {
1822eda14cbcSMatt Macy 		/* compressed bufs must always be assignable to their dbuf */
1823eda14cbcSMatt Macy 		ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1824eda14cbcSMatt Macy 		ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1825eda14cbcSMatt Macy 
1826eda14cbcSMatt Macy 		dbuf_rele(db, FTAG);
1827eda14cbcSMatt Macy 		dmu_write(os, object, offset, blksz, buf->b_data, tx);
1828eda14cbcSMatt Macy 		dmu_return_arcbuf(buf);
1829eda14cbcSMatt Macy 	}
1830eda14cbcSMatt Macy 
1831eda14cbcSMatt Macy 	return (0);
1832eda14cbcSMatt Macy }
1833eda14cbcSMatt Macy 
1834eda14cbcSMatt Macy int
1835eda14cbcSMatt Macy dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1836eda14cbcSMatt Macy     dmu_tx_t *tx)
1837eda14cbcSMatt Macy {
1838eda14cbcSMatt Macy 	int err;
1839ce4dcb97SMartin Matuska 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1840eda14cbcSMatt Macy 
1841ce4dcb97SMartin Matuska 	DB_DNODE_ENTER(db);
1842ce4dcb97SMartin Matuska 	err = dmu_assign_arcbuf_by_dnode(DB_DNODE(db), offset, buf, tx);
1843ce4dcb97SMartin Matuska 	DB_DNODE_EXIT(db);
1844eda14cbcSMatt Macy 
1845eda14cbcSMatt Macy 	return (err);
1846eda14cbcSMatt Macy }
1847eda14cbcSMatt Macy 
18487a7741afSMartin Matuska void
1849eda14cbcSMatt Macy dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1850eda14cbcSMatt Macy {
1851e92ffd9bSMartin Matuska 	(void) buf;
1852eda14cbcSMatt Macy 	dmu_sync_arg_t *dsa = varg;
1853eda14cbcSMatt Macy 
1854eda14cbcSMatt Macy 	if (zio->io_error == 0) {
18557a7741afSMartin Matuska 		dbuf_dirty_record_t *dr = dsa->dsa_dr;
18567a7741afSMartin Matuska 		blkptr_t *bp = zio->io_bp;
18577a7741afSMartin Matuska 
1858eda14cbcSMatt Macy 		if (BP_IS_HOLE(bp)) {
18597a7741afSMartin Matuska 			dmu_buf_t *db = NULL;
18607a7741afSMartin Matuska 			if (dr)
18617a7741afSMartin Matuska 				db = &(dr->dr_dbuf->db);
18627a7741afSMartin Matuska 			else
18637a7741afSMartin Matuska 				db = dsa->dsa_zgd->zgd_db;
1864eda14cbcSMatt Macy 			/*
1865eda14cbcSMatt Macy 			 * A block of zeros may compress to a hole, but the
1866eda14cbcSMatt Macy 			 * block size still needs to be known for replay.
1867eda14cbcSMatt Macy 			 */
1868eda14cbcSMatt Macy 			BP_SET_LSIZE(bp, db->db_size);
1869eda14cbcSMatt Macy 		} else if (!BP_IS_EMBEDDED(bp)) {
1870eda14cbcSMatt Macy 			ASSERT(BP_GET_LEVEL(bp) == 0);
1871eda14cbcSMatt Macy 			BP_SET_FILL(bp, 1);
1872eda14cbcSMatt Macy 		}
1873eda14cbcSMatt Macy 	}
1874eda14cbcSMatt Macy }
1875eda14cbcSMatt Macy 
1876eda14cbcSMatt Macy static void
1877eda14cbcSMatt Macy dmu_sync_late_arrival_ready(zio_t *zio)
1878eda14cbcSMatt Macy {
1879eda14cbcSMatt Macy 	dmu_sync_ready(zio, NULL, zio->io_private);
1880eda14cbcSMatt Macy }
1881eda14cbcSMatt Macy 
18827a7741afSMartin Matuska void
1883eda14cbcSMatt Macy dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1884eda14cbcSMatt Macy {
1885e92ffd9bSMartin Matuska 	(void) buf;
1886eda14cbcSMatt Macy 	dmu_sync_arg_t *dsa = varg;
1887eda14cbcSMatt Macy 	dbuf_dirty_record_t *dr = dsa->dsa_dr;
1888eda14cbcSMatt Macy 	dmu_buf_impl_t *db = dr->dr_dbuf;
1889eda14cbcSMatt Macy 	zgd_t *zgd = dsa->dsa_zgd;
1890eda14cbcSMatt Macy 
1891eda14cbcSMatt Macy 	/*
1892eda14cbcSMatt Macy 	 * Record the vdev(s) backing this blkptr so they can be flushed after
1893eda14cbcSMatt Macy 	 * the writes for the lwb have completed.
1894eda14cbcSMatt Macy 	 */
18957a7741afSMartin Matuska 	if (zgd && zio->io_error == 0) {
1896eda14cbcSMatt Macy 		zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1897eda14cbcSMatt Macy 	}
1898eda14cbcSMatt Macy 
1899eda14cbcSMatt Macy 	mutex_enter(&db->db_mtx);
1900eda14cbcSMatt Macy 	ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1901eda14cbcSMatt Macy 	if (zio->io_error == 0) {
19025c65a0a9SMartin Matuska 		ASSERT0(dr->dt.dl.dr_has_raw_params);
1903eda14cbcSMatt Macy 		dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1904eda14cbcSMatt Macy 		if (dr->dt.dl.dr_nopwrite) {
1905eda14cbcSMatt Macy 			blkptr_t *bp = zio->io_bp;
1906eda14cbcSMatt Macy 			blkptr_t *bp_orig = &zio->io_bp_orig;
1907eda14cbcSMatt Macy 			uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1908eda14cbcSMatt Macy 
1909eda14cbcSMatt Macy 			ASSERT(BP_EQUAL(bp, bp_orig));
1910eda14cbcSMatt Macy 			VERIFY(BP_EQUAL(bp, db->db_blkptr));
1911eda14cbcSMatt Macy 			ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1912eda14cbcSMatt Macy 			VERIFY(zio_checksum_table[chksum].ci_flags &
1913eda14cbcSMatt Macy 			    ZCHECKSUM_FLAG_NOPWRITE);
1914eda14cbcSMatt Macy 		}
1915eda14cbcSMatt Macy 		dr->dt.dl.dr_overridden_by = *zio->io_bp;
1916eda14cbcSMatt Macy 		dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1917eda14cbcSMatt Macy 		dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1918eda14cbcSMatt Macy 
1919eda14cbcSMatt Macy 		/*
1920eda14cbcSMatt Macy 		 * Old style holes are filled with all zeros, whereas
1921eda14cbcSMatt Macy 		 * new-style holes maintain their lsize, type, level,
1922eda14cbcSMatt Macy 		 * and birth time (see zio_write_compress). While we
1923eda14cbcSMatt Macy 		 * need to reset the BP_SET_LSIZE() call that happened
1924eda14cbcSMatt Macy 		 * in dmu_sync_ready for old style holes, we do *not*
1925eda14cbcSMatt Macy 		 * want to wipe out the information contained in new
1926eda14cbcSMatt Macy 		 * style holes. Thus, only zero out the block pointer if
1927eda14cbcSMatt Macy 		 * it's an old style hole.
1928eda14cbcSMatt Macy 		 */
1929eda14cbcSMatt Macy 		if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1930783d3ff6SMartin Matuska 		    BP_GET_LOGICAL_BIRTH(&dr->dt.dl.dr_overridden_by) == 0)
1931eda14cbcSMatt Macy 			BP_ZERO(&dr->dt.dl.dr_overridden_by);
1932eda14cbcSMatt Macy 	} else {
1933eda14cbcSMatt Macy 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1934eda14cbcSMatt Macy 	}
19357a7741afSMartin Matuska 
1936eda14cbcSMatt Macy 	cv_broadcast(&db->db_changed);
1937eda14cbcSMatt Macy 	mutex_exit(&db->db_mtx);
1938eda14cbcSMatt Macy 
19397a7741afSMartin Matuska 	if (dsa->dsa_done)
1940eda14cbcSMatt Macy 		dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1941eda14cbcSMatt Macy 
1942eda14cbcSMatt Macy 	kmem_free(dsa, sizeof (*dsa));
1943eda14cbcSMatt Macy }
1944eda14cbcSMatt Macy 
1945eda14cbcSMatt Macy static void
1946eda14cbcSMatt Macy dmu_sync_late_arrival_done(zio_t *zio)
1947eda14cbcSMatt Macy {
1948eda14cbcSMatt Macy 	blkptr_t *bp = zio->io_bp;
1949eda14cbcSMatt Macy 	dmu_sync_arg_t *dsa = zio->io_private;
1950eda14cbcSMatt Macy 	zgd_t *zgd = dsa->dsa_zgd;
1951eda14cbcSMatt Macy 
1952eda14cbcSMatt Macy 	if (zio->io_error == 0) {
1953eda14cbcSMatt Macy 		/*
1954eda14cbcSMatt Macy 		 * Record the vdev(s) backing this blkptr so they can be
1955eda14cbcSMatt Macy 		 * flushed after the writes for the lwb have completed.
1956eda14cbcSMatt Macy 		 */
1957eda14cbcSMatt Macy 		zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1958eda14cbcSMatt Macy 
1959eda14cbcSMatt Macy 		if (!BP_IS_HOLE(bp)) {
1960eda14cbcSMatt Macy 			blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig;
1961eda14cbcSMatt Macy 			ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1962eda14cbcSMatt Macy 			ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1963783d3ff6SMartin Matuska 			ASSERT(BP_GET_LOGICAL_BIRTH(zio->io_bp) == zio->io_txg);
1964eda14cbcSMatt Macy 			ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1965eda14cbcSMatt Macy 			zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1966eda14cbcSMatt Macy 		}
1967eda14cbcSMatt Macy 	}
1968eda14cbcSMatt Macy 
1969eda14cbcSMatt Macy 	dmu_tx_commit(dsa->dsa_tx);
1970eda14cbcSMatt Macy 
1971eda14cbcSMatt Macy 	dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1972eda14cbcSMatt Macy 
1973184c1b94SMartin Matuska 	abd_free(zio->io_abd);
1974eda14cbcSMatt Macy 	kmem_free(dsa, sizeof (*dsa));
1975eda14cbcSMatt Macy }
1976eda14cbcSMatt Macy 
1977eda14cbcSMatt Macy static int
1978eda14cbcSMatt Macy dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1979eda14cbcSMatt Macy     zio_prop_t *zp, zbookmark_phys_t *zb)
1980eda14cbcSMatt Macy {
1981eda14cbcSMatt Macy 	dmu_sync_arg_t *dsa;
1982eda14cbcSMatt Macy 	dmu_tx_t *tx;
1983315ee00fSMartin Matuska 	int error;
1984315ee00fSMartin Matuska 
1985315ee00fSMartin Matuska 	error = dbuf_read((dmu_buf_impl_t *)zgd->zgd_db, NULL,
1986315ee00fSMartin Matuska 	    DB_RF_CANFAIL | DB_RF_NOPREFETCH);
1987315ee00fSMartin Matuska 	if (error != 0)
1988315ee00fSMartin Matuska 		return (error);
1989eda14cbcSMatt Macy 
1990eda14cbcSMatt Macy 	tx = dmu_tx_create(os);
1991eda14cbcSMatt Macy 	dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1992315ee00fSMartin Matuska 	/*
1993315ee00fSMartin Matuska 	 * This transaction does not produce any dirty data or log blocks, so
1994315ee00fSMartin Matuska 	 * it should not be throttled.  All other cases wait for TXG sync, by
1995315ee00fSMartin Matuska 	 * which time the log block we are writing will be obsolete, so we can
1996315ee00fSMartin Matuska 	 * skip waiting and just return error here instead.
1997315ee00fSMartin Matuska 	 */
1998315ee00fSMartin Matuska 	if (dmu_tx_assign(tx, TXG_NOWAIT | TXG_NOTHROTTLE) != 0) {
1999eda14cbcSMatt Macy 		dmu_tx_abort(tx);
2000eda14cbcSMatt Macy 		/* Make zl_get_data do txg_waited_synced() */
2001eda14cbcSMatt Macy 		return (SET_ERROR(EIO));
2002eda14cbcSMatt Macy 	}
2003eda14cbcSMatt Macy 
2004eda14cbcSMatt Macy 	/*
2005eda14cbcSMatt Macy 	 * In order to prevent the zgd's lwb from being free'd prior to
2006eda14cbcSMatt Macy 	 * dmu_sync_late_arrival_done() being called, we have to ensure
2007eda14cbcSMatt Macy 	 * the lwb's "max txg" takes this tx's txg into account.
2008eda14cbcSMatt Macy 	 */
2009eda14cbcSMatt Macy 	zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
2010eda14cbcSMatt Macy 
2011eda14cbcSMatt Macy 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
2012eda14cbcSMatt Macy 	dsa->dsa_dr = NULL;
2013eda14cbcSMatt Macy 	dsa->dsa_done = done;
2014eda14cbcSMatt Macy 	dsa->dsa_zgd = zgd;
2015eda14cbcSMatt Macy 	dsa->dsa_tx = tx;
2016eda14cbcSMatt Macy 
2017eda14cbcSMatt Macy 	/*
2018eda14cbcSMatt Macy 	 * Since we are currently syncing this txg, it's nontrivial to
2019eda14cbcSMatt Macy 	 * determine what BP to nopwrite against, so we disable nopwrite.
2020eda14cbcSMatt Macy 	 *
2021eda14cbcSMatt Macy 	 * When syncing, the db_blkptr is initially the BP of the previous
2022eda14cbcSMatt Macy 	 * txg.  We can not nopwrite against it because it will be changed
2023eda14cbcSMatt Macy 	 * (this is similar to the non-late-arrival case where the dbuf is
2024eda14cbcSMatt Macy 	 * dirty in a future txg).
2025eda14cbcSMatt Macy 	 *
2026eda14cbcSMatt Macy 	 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
2027eda14cbcSMatt Macy 	 * We can not nopwrite against it because although the BP will not
2028eda14cbcSMatt Macy 	 * (typically) be changed, the data has not yet been persisted to this
2029eda14cbcSMatt Macy 	 * location.
2030eda14cbcSMatt Macy 	 *
2031eda14cbcSMatt Macy 	 * Finally, when dbuf_write_done() is called, it is theoretically
2032eda14cbcSMatt Macy 	 * possible to always nopwrite, because the data that was written in
2033eda14cbcSMatt Macy 	 * this txg is the same data that we are trying to write.  However we
2034eda14cbcSMatt Macy 	 * would need to check that this dbuf is not dirty in any future
2035eda14cbcSMatt Macy 	 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
2036eda14cbcSMatt Macy 	 * don't nopwrite in this case.
2037eda14cbcSMatt Macy 	 */
2038eda14cbcSMatt Macy 	zp->zp_nopwrite = B_FALSE;
2039eda14cbcSMatt Macy 
2040eda14cbcSMatt Macy 	zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
2041eda14cbcSMatt Macy 	    abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
2042eda14cbcSMatt Macy 	    zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
2043b7198dcfSMartin Matuska 	    dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done,
2044eda14cbcSMatt Macy 	    dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
2045eda14cbcSMatt Macy 
2046eda14cbcSMatt Macy 	return (0);
2047eda14cbcSMatt Macy }
2048eda14cbcSMatt Macy 
2049eda14cbcSMatt Macy /*
2050eda14cbcSMatt Macy  * Intent log support: sync the block associated with db to disk.
2051eda14cbcSMatt Macy  * N.B. and XXX: the caller is responsible for making sure that the
2052eda14cbcSMatt Macy  * data isn't changing while dmu_sync() is writing it.
2053eda14cbcSMatt Macy  *
2054eda14cbcSMatt Macy  * Return values:
2055eda14cbcSMatt Macy  *
2056eda14cbcSMatt Macy  *	EEXIST: this txg has already been synced, so there's nothing to do.
2057eda14cbcSMatt Macy  *		The caller should not log the write.
2058eda14cbcSMatt Macy  *
2059eda14cbcSMatt Macy  *	ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
2060eda14cbcSMatt Macy  *		The caller should not log the write.
2061eda14cbcSMatt Macy  *
2062eda14cbcSMatt Macy  *	EALREADY: this block is already in the process of being synced.
2063eda14cbcSMatt Macy  *		The caller should track its progress (somehow).
2064eda14cbcSMatt Macy  *
2065eda14cbcSMatt Macy  *	EIO: could not do the I/O.
2066eda14cbcSMatt Macy  *		The caller should do a txg_wait_synced().
2067eda14cbcSMatt Macy  *
2068eda14cbcSMatt Macy  *	0: the I/O has been initiated.
2069eda14cbcSMatt Macy  *		The caller should log this blkptr in the done callback.
2070eda14cbcSMatt Macy  *		It is possible that the I/O will fail, in which case
2071eda14cbcSMatt Macy  *		the error will be reported to the done callback and
2072eda14cbcSMatt Macy  *		propagated to pio from zio_done().
2073eda14cbcSMatt Macy  */
2074eda14cbcSMatt Macy int
2075eda14cbcSMatt Macy dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
2076eda14cbcSMatt Macy {
2077eda14cbcSMatt Macy 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
2078eda14cbcSMatt Macy 	objset_t *os = db->db_objset;
2079eda14cbcSMatt Macy 	dsl_dataset_t *ds = os->os_dsl_dataset;
2080eda14cbcSMatt Macy 	dbuf_dirty_record_t *dr, *dr_next;
2081eda14cbcSMatt Macy 	dmu_sync_arg_t *dsa;
2082eda14cbcSMatt Macy 	zbookmark_phys_t zb;
2083eda14cbcSMatt Macy 	zio_prop_t zp;
2084eda14cbcSMatt Macy 
2085eda14cbcSMatt Macy 	ASSERT(pio != NULL);
2086eda14cbcSMatt Macy 	ASSERT(txg != 0);
2087eda14cbcSMatt Macy 
2088eda14cbcSMatt Macy 	SET_BOOKMARK(&zb, ds->ds_object,
2089eda14cbcSMatt Macy 	    db->db.db_object, db->db_level, db->db_blkid);
2090eda14cbcSMatt Macy 
2091eda14cbcSMatt Macy 	DB_DNODE_ENTER(db);
2092ce4dcb97SMartin Matuska 	dmu_write_policy(os, DB_DNODE(db), db->db_level, WP_DMU_SYNC, &zp);
2093eda14cbcSMatt Macy 	DB_DNODE_EXIT(db);
2094eda14cbcSMatt Macy 
2095eda14cbcSMatt Macy 	/*
2096eda14cbcSMatt Macy 	 * If we're frozen (running ziltest), we always need to generate a bp.
2097eda14cbcSMatt Macy 	 */
2098eda14cbcSMatt Macy 	if (txg > spa_freeze_txg(os->os_spa))
2099eda14cbcSMatt Macy 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2100eda14cbcSMatt Macy 
2101eda14cbcSMatt Macy 	/*
2102eda14cbcSMatt Macy 	 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
2103eda14cbcSMatt Macy 	 * and us.  If we determine that this txg is not yet syncing,
2104eda14cbcSMatt Macy 	 * but it begins to sync a moment later, that's OK because the
2105eda14cbcSMatt Macy 	 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
2106eda14cbcSMatt Macy 	 */
2107eda14cbcSMatt Macy 	mutex_enter(&db->db_mtx);
2108eda14cbcSMatt Macy 
2109eda14cbcSMatt Macy 	if (txg <= spa_last_synced_txg(os->os_spa)) {
2110eda14cbcSMatt Macy 		/*
2111eda14cbcSMatt Macy 		 * This txg has already synced.  There's nothing to do.
2112eda14cbcSMatt Macy 		 */
2113eda14cbcSMatt Macy 		mutex_exit(&db->db_mtx);
2114eda14cbcSMatt Macy 		return (SET_ERROR(EEXIST));
2115eda14cbcSMatt Macy 	}
2116eda14cbcSMatt Macy 
2117eda14cbcSMatt Macy 	if (txg <= spa_syncing_txg(os->os_spa)) {
2118eda14cbcSMatt Macy 		/*
2119eda14cbcSMatt Macy 		 * This txg is currently syncing, so we can't mess with
2120eda14cbcSMatt Macy 		 * the dirty record anymore; just write a new log block.
2121eda14cbcSMatt Macy 		 */
2122eda14cbcSMatt Macy 		mutex_exit(&db->db_mtx);
2123eda14cbcSMatt Macy 		return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2124eda14cbcSMatt Macy 	}
2125eda14cbcSMatt Macy 
2126eda14cbcSMatt Macy 	dr = dbuf_find_dirty_eq(db, txg);
2127eda14cbcSMatt Macy 
2128eda14cbcSMatt Macy 	if (dr == NULL) {
2129eda14cbcSMatt Macy 		/*
2130eda14cbcSMatt Macy 		 * There's no dr for this dbuf, so it must have been freed.
2131eda14cbcSMatt Macy 		 * There's no need to log writes to freed blocks, so we're done.
2132eda14cbcSMatt Macy 		 */
2133eda14cbcSMatt Macy 		mutex_exit(&db->db_mtx);
2134eda14cbcSMatt Macy 		return (SET_ERROR(ENOENT));
2135eda14cbcSMatt Macy 	}
2136eda14cbcSMatt Macy 
2137eda14cbcSMatt Macy 	dr_next = list_next(&db->db_dirty_records, dr);
2138eda14cbcSMatt Macy 	ASSERT(dr_next == NULL || dr_next->dr_txg < txg);
2139eda14cbcSMatt Macy 
2140eda14cbcSMatt Macy 	if (db->db_blkptr != NULL) {
2141eda14cbcSMatt Macy 		/*
2142eda14cbcSMatt Macy 		 * We need to fill in zgd_bp with the current blkptr so that
2143eda14cbcSMatt Macy 		 * the nopwrite code can check if we're writing the same
2144eda14cbcSMatt Macy 		 * data that's already on disk.  We can only nopwrite if we
2145eda14cbcSMatt Macy 		 * are sure that after making the copy, db_blkptr will not
2146eda14cbcSMatt Macy 		 * change until our i/o completes.  We ensure this by
2147eda14cbcSMatt Macy 		 * holding the db_mtx, and only allowing nopwrite if the
2148eda14cbcSMatt Macy 		 * block is not already dirty (see below).  This is verified
2149eda14cbcSMatt Macy 		 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
2150eda14cbcSMatt Macy 		 * not changed.
2151eda14cbcSMatt Macy 		 */
2152eda14cbcSMatt Macy 		*zgd->zgd_bp = *db->db_blkptr;
2153eda14cbcSMatt Macy 	}
2154eda14cbcSMatt Macy 
2155eda14cbcSMatt Macy 	/*
2156eda14cbcSMatt Macy 	 * Assume the on-disk data is X, the current syncing data (in
2157eda14cbcSMatt Macy 	 * txg - 1) is Y, and the current in-memory data is Z (currently
2158eda14cbcSMatt Macy 	 * in dmu_sync).
2159eda14cbcSMatt Macy 	 *
2160eda14cbcSMatt Macy 	 * We usually want to perform a nopwrite if X and Z are the
2161eda14cbcSMatt Macy 	 * same.  However, if Y is different (i.e. the BP is going to
2162eda14cbcSMatt Macy 	 * change before this write takes effect), then a nopwrite will
2163eda14cbcSMatt Macy 	 * be incorrect - we would override with X, which could have
2164eda14cbcSMatt Macy 	 * been freed when Y was written.
2165eda14cbcSMatt Macy 	 *
2166eda14cbcSMatt Macy 	 * (Note that this is not a concern when we are nop-writing from
2167eda14cbcSMatt Macy 	 * syncing context, because X and Y must be identical, because
2168eda14cbcSMatt Macy 	 * all previous txgs have been synced.)
2169eda14cbcSMatt Macy 	 *
2170eda14cbcSMatt Macy 	 * Therefore, we disable nopwrite if the current BP could change
2171eda14cbcSMatt Macy 	 * before this TXG.  There are two ways it could change: by
2172eda14cbcSMatt Macy 	 * being dirty (dr_next is non-NULL), or by being freed
2173eda14cbcSMatt Macy 	 * (dnode_block_freed()).  This behavior is verified by
2174eda14cbcSMatt Macy 	 * zio_done(), which VERIFYs that the override BP is identical
2175eda14cbcSMatt Macy 	 * to the on-disk BP.
2176eda14cbcSMatt Macy 	 */
2177ce4dcb97SMartin Matuska 	if (dr_next != NULL) {
2178ce4dcb97SMartin Matuska 		zp.zp_nopwrite = B_FALSE;
2179ce4dcb97SMartin Matuska 	} else {
2180eda14cbcSMatt Macy 		DB_DNODE_ENTER(db);
2181ce4dcb97SMartin Matuska 		if (dnode_block_freed(DB_DNODE(db), db->db_blkid))
2182eda14cbcSMatt Macy 			zp.zp_nopwrite = B_FALSE;
2183eda14cbcSMatt Macy 		DB_DNODE_EXIT(db);
2184ce4dcb97SMartin Matuska 	}
2185eda14cbcSMatt Macy 
2186eda14cbcSMatt Macy 	ASSERT(dr->dr_txg == txg);
2187eda14cbcSMatt Macy 	if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
2188eda14cbcSMatt Macy 	    dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2189eda14cbcSMatt Macy 		/*
2190eda14cbcSMatt Macy 		 * We have already issued a sync write for this buffer,
2191eda14cbcSMatt Macy 		 * or this buffer has already been synced.  It could not
2192eda14cbcSMatt Macy 		 * have been dirtied since, or we would have cleared the state.
2193eda14cbcSMatt Macy 		 */
2194eda14cbcSMatt Macy 		mutex_exit(&db->db_mtx);
2195eda14cbcSMatt Macy 		return (SET_ERROR(EALREADY));
2196eda14cbcSMatt Macy 	}
2197eda14cbcSMatt Macy 
21985c65a0a9SMartin Matuska 	ASSERT0(dr->dt.dl.dr_has_raw_params);
2199eda14cbcSMatt Macy 	ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2200eda14cbcSMatt Macy 	dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
2201eda14cbcSMatt Macy 	mutex_exit(&db->db_mtx);
2202eda14cbcSMatt Macy 
2203eda14cbcSMatt Macy 	dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
2204eda14cbcSMatt Macy 	dsa->dsa_dr = dr;
2205eda14cbcSMatt Macy 	dsa->dsa_done = done;
2206eda14cbcSMatt Macy 	dsa->dsa_zgd = zgd;
2207eda14cbcSMatt Macy 	dsa->dsa_tx = NULL;
2208eda14cbcSMatt Macy 
220915f0b8c3SMartin Matuska 	zio_nowait(arc_write(pio, os->os_spa, txg, zgd->zgd_bp,
22107a7741afSMartin Matuska 	    dr->dt.dl.dr_data, !DBUF_IS_CACHEABLE(db),
22117a7741afSMartin Matuska 	    dbuf_is_l2cacheable(db, NULL), &zp, dmu_sync_ready, NULL,
22127a7741afSMartin Matuska 	    dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL,
22137a7741afSMartin Matuska 	    &zb));
2214eda14cbcSMatt Macy 
2215eda14cbcSMatt Macy 	return (0);
2216eda14cbcSMatt Macy }
2217eda14cbcSMatt Macy 
2218eda14cbcSMatt Macy int
2219eda14cbcSMatt Macy dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx)
2220eda14cbcSMatt Macy {
2221eda14cbcSMatt Macy 	dnode_t *dn;
2222eda14cbcSMatt Macy 	int err;
2223eda14cbcSMatt Macy 
2224eda14cbcSMatt Macy 	err = dnode_hold(os, object, FTAG, &dn);
2225eda14cbcSMatt Macy 	if (err)
2226eda14cbcSMatt Macy 		return (err);
2227eda14cbcSMatt Macy 	err = dnode_set_nlevels(dn, nlevels, tx);
2228eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
2229eda14cbcSMatt Macy 	return (err);
2230eda14cbcSMatt Macy }
2231eda14cbcSMatt Macy 
2232eda14cbcSMatt Macy int
2233eda14cbcSMatt Macy dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
2234eda14cbcSMatt Macy     dmu_tx_t *tx)
2235eda14cbcSMatt Macy {
2236eda14cbcSMatt Macy 	dnode_t *dn;
2237eda14cbcSMatt Macy 	int err;
2238eda14cbcSMatt Macy 
2239eda14cbcSMatt Macy 	err = dnode_hold(os, object, FTAG, &dn);
2240eda14cbcSMatt Macy 	if (err)
2241eda14cbcSMatt Macy 		return (err);
2242eda14cbcSMatt Macy 	err = dnode_set_blksz(dn, size, ibs, tx);
2243eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
2244eda14cbcSMatt Macy 	return (err);
2245eda14cbcSMatt Macy }
2246eda14cbcSMatt Macy 
2247eda14cbcSMatt Macy int
2248eda14cbcSMatt Macy dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid,
2249eda14cbcSMatt Macy     dmu_tx_t *tx)
2250eda14cbcSMatt Macy {
2251eda14cbcSMatt Macy 	dnode_t *dn;
2252eda14cbcSMatt Macy 	int err;
2253eda14cbcSMatt Macy 
2254eda14cbcSMatt Macy 	err = dnode_hold(os, object, FTAG, &dn);
2255eda14cbcSMatt Macy 	if (err)
2256eda14cbcSMatt Macy 		return (err);
2257eda14cbcSMatt Macy 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2258eda14cbcSMatt Macy 	dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE);
2259eda14cbcSMatt Macy 	rw_exit(&dn->dn_struct_rwlock);
2260eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
2261eda14cbcSMatt Macy 	return (0);
2262eda14cbcSMatt Macy }
2263eda14cbcSMatt Macy 
2264eda14cbcSMatt Macy void
2265eda14cbcSMatt Macy dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
2266eda14cbcSMatt Macy     dmu_tx_t *tx)
2267eda14cbcSMatt Macy {
2268eda14cbcSMatt Macy 	dnode_t *dn;
2269eda14cbcSMatt Macy 
2270eda14cbcSMatt Macy 	/*
2271eda14cbcSMatt Macy 	 * Send streams include each object's checksum function.  This
2272eda14cbcSMatt Macy 	 * check ensures that the receiving system can understand the
2273eda14cbcSMatt Macy 	 * checksum function transmitted.
2274eda14cbcSMatt Macy 	 */
2275eda14cbcSMatt Macy 	ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
2276eda14cbcSMatt Macy 
2277eda14cbcSMatt Macy 	VERIFY0(dnode_hold(os, object, FTAG, &dn));
2278eda14cbcSMatt Macy 	ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
2279eda14cbcSMatt Macy 	dn->dn_checksum = checksum;
2280eda14cbcSMatt Macy 	dnode_setdirty(dn, tx);
2281eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
2282eda14cbcSMatt Macy }
2283eda14cbcSMatt Macy 
2284eda14cbcSMatt Macy void
2285eda14cbcSMatt Macy dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
2286eda14cbcSMatt Macy     dmu_tx_t *tx)
2287eda14cbcSMatt Macy {
2288eda14cbcSMatt Macy 	dnode_t *dn;
2289eda14cbcSMatt Macy 
2290eda14cbcSMatt Macy 	/*
2291eda14cbcSMatt Macy 	 * Send streams include each object's compression function.  This
2292eda14cbcSMatt Macy 	 * check ensures that the receiving system can understand the
2293eda14cbcSMatt Macy 	 * compression function transmitted.
2294eda14cbcSMatt Macy 	 */
2295eda14cbcSMatt Macy 	ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
2296eda14cbcSMatt Macy 
2297eda14cbcSMatt Macy 	VERIFY0(dnode_hold(os, object, FTAG, &dn));
2298eda14cbcSMatt Macy 	dn->dn_compress = compress;
2299eda14cbcSMatt Macy 	dnode_setdirty(dn, tx);
2300eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
2301eda14cbcSMatt Macy }
2302eda14cbcSMatt Macy 
2303eda14cbcSMatt Macy /*
2304eda14cbcSMatt Macy  * When the "redundant_metadata" property is set to "most", only indirect
2305eda14cbcSMatt Macy  * blocks of this level and higher will have an additional ditto block.
2306eda14cbcSMatt Macy  */
2307e92ffd9bSMartin Matuska static const int zfs_redundant_metadata_most_ditto_level = 2;
2308eda14cbcSMatt Macy 
2309eda14cbcSMatt Macy void
2310eda14cbcSMatt Macy dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
2311eda14cbcSMatt Macy {
2312eda14cbcSMatt Macy 	dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2313eda14cbcSMatt Macy 	boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2314eda14cbcSMatt Macy 	    (wp & WP_SPILL));
2315eda14cbcSMatt Macy 	enum zio_checksum checksum = os->os_checksum;
2316eda14cbcSMatt Macy 	enum zio_compress compress = os->os_compress;
2317eda14cbcSMatt Macy 	uint8_t complevel = os->os_complevel;
2318eda14cbcSMatt Macy 	enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2319eda14cbcSMatt Macy 	boolean_t dedup = B_FALSE;
2320eda14cbcSMatt Macy 	boolean_t nopwrite = B_FALSE;
2321eda14cbcSMatt Macy 	boolean_t dedup_verify = os->os_dedup_verify;
2322eda14cbcSMatt Macy 	boolean_t encrypt = B_FALSE;
2323eda14cbcSMatt Macy 	int copies = os->os_copies;
2324eda14cbcSMatt Macy 
2325eda14cbcSMatt Macy 	/*
2326eda14cbcSMatt Macy 	 * We maintain different write policies for each of the following
2327eda14cbcSMatt Macy 	 * types of data:
2328eda14cbcSMatt Macy 	 *	 1. metadata
2329eda14cbcSMatt Macy 	 *	 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2330eda14cbcSMatt Macy 	 *	 3. all other level 0 blocks
2331eda14cbcSMatt Macy 	 */
2332eda14cbcSMatt Macy 	if (ismd) {
2333eda14cbcSMatt Macy 		/*
2334eda14cbcSMatt Macy 		 * XXX -- we should design a compression algorithm
2335eda14cbcSMatt Macy 		 * that specializes in arrays of bps.
2336eda14cbcSMatt Macy 		 */
2337eda14cbcSMatt Macy 		compress = zio_compress_select(os->os_spa,
2338eda14cbcSMatt Macy 		    ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2339eda14cbcSMatt Macy 
2340eda14cbcSMatt Macy 		/*
2341eda14cbcSMatt Macy 		 * Metadata always gets checksummed.  If the data
2342eda14cbcSMatt Macy 		 * checksum is multi-bit correctable, and it's not a
2343eda14cbcSMatt Macy 		 * ZBT-style checksum, then it's suitable for metadata
2344eda14cbcSMatt Macy 		 * as well.  Otherwise, the metadata checksum defaults
2345eda14cbcSMatt Macy 		 * to fletcher4.
2346eda14cbcSMatt Macy 		 */
2347eda14cbcSMatt Macy 		if (!(zio_checksum_table[checksum].ci_flags &
2348eda14cbcSMatt Macy 		    ZCHECKSUM_FLAG_METADATA) ||
2349eda14cbcSMatt Macy 		    (zio_checksum_table[checksum].ci_flags &
2350eda14cbcSMatt Macy 		    ZCHECKSUM_FLAG_EMBEDDED))
2351eda14cbcSMatt Macy 			checksum = ZIO_CHECKSUM_FLETCHER_4;
2352eda14cbcSMatt Macy 
2353dbd5678dSMartin Matuska 		switch (os->os_redundant_metadata) {
2354dbd5678dSMartin Matuska 		case ZFS_REDUNDANT_METADATA_ALL:
2355eda14cbcSMatt Macy 			copies++;
2356dbd5678dSMartin Matuska 			break;
2357dbd5678dSMartin Matuska 		case ZFS_REDUNDANT_METADATA_MOST:
2358dbd5678dSMartin Matuska 			if (level >= zfs_redundant_metadata_most_ditto_level ||
2359dbd5678dSMartin Matuska 			    DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))
2360dbd5678dSMartin Matuska 				copies++;
2361dbd5678dSMartin Matuska 			break;
2362dbd5678dSMartin Matuska 		case ZFS_REDUNDANT_METADATA_SOME:
2363dbd5678dSMartin Matuska 			if (DMU_OT_IS_CRITICAL(type))
2364dbd5678dSMartin Matuska 				copies++;
2365dbd5678dSMartin Matuska 			break;
2366dbd5678dSMartin Matuska 		case ZFS_REDUNDANT_METADATA_NONE:
2367dbd5678dSMartin Matuska 			break;
2368dbd5678dSMartin Matuska 		}
2369e2df9bb4SMartin Matuska 
2370e2df9bb4SMartin Matuska 		if (dmu_ddt_copies > 0) {
2371e2df9bb4SMartin Matuska 			/*
2372e2df9bb4SMartin Matuska 			 * If this tuneable is set, and this is a write for a
2373e2df9bb4SMartin Matuska 			 * dedup entry store (zap or log), then we treat it
2374e2df9bb4SMartin Matuska 			 * something like ZFS_REDUNDANT_METADATA_MOST on a
2375e2df9bb4SMartin Matuska 			 * regular dataset: this many copies, and one more for
2376e2df9bb4SMartin Matuska 			 * "higher" indirect blocks. This specific exception is
2377e2df9bb4SMartin Matuska 			 * necessary because dedup objects are stored in the
2378e2df9bb4SMartin Matuska 			 * MOS, which always has the highest possible copies.
2379e2df9bb4SMartin Matuska 			 */
2380e2df9bb4SMartin Matuska 			dmu_object_type_t stype =
2381e2df9bb4SMartin Matuska 			    dn ? dn->dn_storage_type : DMU_OT_NONE;
2382e2df9bb4SMartin Matuska 			if (stype == DMU_OT_NONE)
2383e2df9bb4SMartin Matuska 				stype = type;
2384e2df9bb4SMartin Matuska 			if (stype == DMU_OT_DDT_ZAP) {
2385e2df9bb4SMartin Matuska 				copies = dmu_ddt_copies;
2386e2df9bb4SMartin Matuska 				if (level >=
2387e2df9bb4SMartin Matuska 				    zfs_redundant_metadata_most_ditto_level)
2388e2df9bb4SMartin Matuska 					copies++;
2389e2df9bb4SMartin Matuska 			}
2390e2df9bb4SMartin Matuska 		}
2391eda14cbcSMatt Macy 	} else if (wp & WP_NOFILL) {
2392eda14cbcSMatt Macy 		ASSERT(level == 0);
2393eda14cbcSMatt Macy 
2394eda14cbcSMatt Macy 		/*
2395eda14cbcSMatt Macy 		 * If we're writing preallocated blocks, we aren't actually
2396eda14cbcSMatt Macy 		 * writing them so don't set any policy properties.  These
2397eda14cbcSMatt Macy 		 * blocks are currently only used by an external subsystem
2398eda14cbcSMatt Macy 		 * outside of zfs (i.e. dump) and not written by the zio
2399eda14cbcSMatt Macy 		 * pipeline.
2400eda14cbcSMatt Macy 		 */
2401eda14cbcSMatt Macy 		compress = ZIO_COMPRESS_OFF;
2402eda14cbcSMatt Macy 		checksum = ZIO_CHECKSUM_OFF;
2403eda14cbcSMatt Macy 	} else {
2404eda14cbcSMatt Macy 		compress = zio_compress_select(os->os_spa, dn->dn_compress,
2405eda14cbcSMatt Macy 		    compress);
2406eda14cbcSMatt Macy 		complevel = zio_complevel_select(os->os_spa, compress,
2407eda14cbcSMatt Macy 		    complevel, complevel);
2408eda14cbcSMatt Macy 
2409eda14cbcSMatt Macy 		checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2410eda14cbcSMatt Macy 		    zio_checksum_select(dn->dn_checksum, checksum) :
2411eda14cbcSMatt Macy 		    dedup_checksum;
2412eda14cbcSMatt Macy 
2413eda14cbcSMatt Macy 		/*
2414eda14cbcSMatt Macy 		 * Determine dedup setting.  If we are in dmu_sync(),
2415eda14cbcSMatt Macy 		 * we won't actually dedup now because that's all
2416eda14cbcSMatt Macy 		 * done in syncing context; but we do want to use the
2417eda14cbcSMatt Macy 		 * dedup checksum.  If the checksum is not strong
2418eda14cbcSMatt Macy 		 * enough to ensure unique signatures, force
2419eda14cbcSMatt Macy 		 * dedup_verify.
2420eda14cbcSMatt Macy 		 */
2421eda14cbcSMatt Macy 		if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2422eda14cbcSMatt Macy 			dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2423eda14cbcSMatt Macy 			if (!(zio_checksum_table[checksum].ci_flags &
2424eda14cbcSMatt Macy 			    ZCHECKSUM_FLAG_DEDUP))
2425eda14cbcSMatt Macy 				dedup_verify = B_TRUE;
2426eda14cbcSMatt Macy 		}
2427eda14cbcSMatt Macy 
2428eda14cbcSMatt Macy 		/*
2429eda14cbcSMatt Macy 		 * Enable nopwrite if we have secure enough checksum
2430eda14cbcSMatt Macy 		 * algorithm (see comment in zio_nop_write) and
2431eda14cbcSMatt Macy 		 * compression is enabled.  We don't enable nopwrite if
2432eda14cbcSMatt Macy 		 * dedup is enabled as the two features are mutually
2433eda14cbcSMatt Macy 		 * exclusive.
2434eda14cbcSMatt Macy 		 */
2435eda14cbcSMatt Macy 		nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2436eda14cbcSMatt Macy 		    ZCHECKSUM_FLAG_NOPWRITE) &&
2437eda14cbcSMatt Macy 		    compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2438eda14cbcSMatt Macy 	}
2439eda14cbcSMatt Macy 
2440eda14cbcSMatt Macy 	/*
2441eda14cbcSMatt Macy 	 * All objects in an encrypted objset are protected from modification
2442eda14cbcSMatt Macy 	 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2443eda14cbcSMatt Macy 	 * in the bp, so we cannot use all copies. Encrypted objects are also
2444eda14cbcSMatt Macy 	 * not subject to nopwrite since writing the same data will still
2445eda14cbcSMatt Macy 	 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2446eda14cbcSMatt Macy 	 * to avoid ambiguity in the dedup code since the DDT does not store
2447eda14cbcSMatt Macy 	 * object types.
2448eda14cbcSMatt Macy 	 */
2449eda14cbcSMatt Macy 	if (os->os_encrypted && (wp & WP_NOFILL) == 0) {
2450eda14cbcSMatt Macy 		encrypt = B_TRUE;
2451eda14cbcSMatt Macy 
2452eda14cbcSMatt Macy 		if (DMU_OT_IS_ENCRYPTED(type)) {
2453eda14cbcSMatt Macy 			copies = MIN(copies, SPA_DVAS_PER_BP - 1);
2454eda14cbcSMatt Macy 			nopwrite = B_FALSE;
2455eda14cbcSMatt Macy 		} else {
2456eda14cbcSMatt Macy 			dedup = B_FALSE;
2457eda14cbcSMatt Macy 		}
2458eda14cbcSMatt Macy 
2459eda14cbcSMatt Macy 		if (level <= 0 &&
2460eda14cbcSMatt Macy 		    (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) {
2461eda14cbcSMatt Macy 			compress = ZIO_COMPRESS_EMPTY;
2462eda14cbcSMatt Macy 		}
2463eda14cbcSMatt Macy 	}
2464eda14cbcSMatt Macy 
2465eda14cbcSMatt Macy 	zp->zp_compress = compress;
2466eda14cbcSMatt Macy 	zp->zp_complevel = complevel;
2467eda14cbcSMatt Macy 	zp->zp_checksum = checksum;
2468eda14cbcSMatt Macy 	zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2469eda14cbcSMatt Macy 	zp->zp_level = level;
2470eda14cbcSMatt Macy 	zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2471eda14cbcSMatt Macy 	zp->zp_dedup = dedup;
2472eda14cbcSMatt Macy 	zp->zp_dedup_verify = dedup && dedup_verify;
2473eda14cbcSMatt Macy 	zp->zp_nopwrite = nopwrite;
2474eda14cbcSMatt Macy 	zp->zp_encrypt = encrypt;
2475eda14cbcSMatt Macy 	zp->zp_byteorder = ZFS_HOST_BYTEORDER;
24767a7741afSMartin Matuska 	zp->zp_direct_write = (wp & WP_DIRECT_WR) ? B_TRUE : B_FALSE;
2477da5137abSMartin Matuska 	memset(zp->zp_salt, 0, ZIO_DATA_SALT_LEN);
2478da5137abSMartin Matuska 	memset(zp->zp_iv, 0, ZIO_DATA_IV_LEN);
2479da5137abSMartin Matuska 	memset(zp->zp_mac, 0, ZIO_DATA_MAC_LEN);
2480eda14cbcSMatt Macy 	zp->zp_zpl_smallblk = DMU_OT_IS_FILE(zp->zp_type) ?
2481eda14cbcSMatt Macy 	    os->os_zpl_special_smallblock : 0;
2482ce4dcb97SMartin Matuska 	zp->zp_storage_type = dn ? dn->dn_storage_type : DMU_OT_NONE;
2483eda14cbcSMatt Macy 
2484eda14cbcSMatt Macy 	ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2485eda14cbcSMatt Macy }
2486eda14cbcSMatt Macy 
2487eda14cbcSMatt Macy /*
24882a58b312SMartin Matuska  * Reports the location of data and holes in an object.  In order to
24892a58b312SMartin Matuska  * accurately report holes all dirty data must be synced to disk.  This
24902a58b312SMartin Matuska  * causes extremely poor performance when seeking for holes in a dirty file.
24912a58b312SMartin Matuska  * As a compromise, only provide hole data when the dnode is clean.  When
24922a58b312SMartin Matuska  * a dnode is dirty report the dnode as having no holes by returning EBUSY
24932a58b312SMartin Matuska  * which is always safe to do.
2494eda14cbcSMatt Macy  */
2495eda14cbcSMatt Macy int
2496eda14cbcSMatt Macy dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2497eda14cbcSMatt Macy {
2498eda14cbcSMatt Macy 	dnode_t *dn;
24992a58b312SMartin Matuska 	int restarted = 0, err;
2500eda14cbcSMatt Macy 
250181b22a98SMartin Matuska restart:
2502eda14cbcSMatt Macy 	err = dnode_hold(os, object, FTAG, &dn);
2503eda14cbcSMatt Macy 	if (err)
2504eda14cbcSMatt Macy 		return (err);
2505eda14cbcSMatt Macy 
250681b22a98SMartin Matuska 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
2507eda14cbcSMatt Macy 
250881b22a98SMartin Matuska 	if (dnode_is_dirty(dn)) {
2509eda14cbcSMatt Macy 		/*
251081b22a98SMartin Matuska 		 * If the zfs_dmu_offset_next_sync module option is enabled
25112a58b312SMartin Matuska 		 * then hole reporting has been requested.  Dirty dnodes
25122a58b312SMartin Matuska 		 * must be synced to disk to accurately report holes.
251381b22a98SMartin Matuska 		 *
25142a58b312SMartin Matuska 		 * Provided a RL_READER rangelock spanning 0-UINT64_MAX is
25152a58b312SMartin Matuska 		 * held by the caller only a single restart will be required.
25162a58b312SMartin Matuska 		 * We tolerate callers which do not hold the rangelock by
25172a58b312SMartin Matuska 		 * returning EBUSY and not reporting holes after one restart.
2518eda14cbcSMatt Macy 		 */
251981b22a98SMartin Matuska 		if (zfs_dmu_offset_next_sync) {
252081b22a98SMartin Matuska 			rw_exit(&dn->dn_struct_rwlock);
2521eda14cbcSMatt Macy 			dnode_rele(dn, FTAG);
25222a58b312SMartin Matuska 
25232a58b312SMartin Matuska 			if (restarted)
25242a58b312SMartin Matuska 				return (SET_ERROR(EBUSY));
25252a58b312SMartin Matuska 
2526eda14cbcSMatt Macy 			txg_wait_synced(dmu_objset_pool(os), 0);
25272a58b312SMartin Matuska 			restarted = 1;
252881b22a98SMartin Matuska 			goto restart;
2529eda14cbcSMatt Macy 		}
2530eda14cbcSMatt Macy 
2531eda14cbcSMatt Macy 		err = SET_ERROR(EBUSY);
253281b22a98SMartin Matuska 	} else {
253381b22a98SMartin Matuska 		err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK |
253481b22a98SMartin Matuska 		    (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
253581b22a98SMartin Matuska 	}
2536eda14cbcSMatt Macy 
253781b22a98SMartin Matuska 	rw_exit(&dn->dn_struct_rwlock);
2538eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
2539eda14cbcSMatt Macy 
2540eda14cbcSMatt Macy 	return (err);
2541eda14cbcSMatt Macy }
2542eda14cbcSMatt Macy 
25432a58b312SMartin Matuska int
25442a58b312SMartin Matuska dmu_read_l0_bps(objset_t *os, uint64_t object, uint64_t offset, uint64_t length,
2545e639e0d2SMartin Matuska     blkptr_t *bps, size_t *nbpsp)
25462a58b312SMartin Matuska {
25472a58b312SMartin Matuska 	dmu_buf_t **dbp, *dbuf;
25482a58b312SMartin Matuska 	dmu_buf_impl_t *db;
25492a58b312SMartin Matuska 	blkptr_t *bp;
25502a58b312SMartin Matuska 	int error, numbufs;
25512a58b312SMartin Matuska 
25522a58b312SMartin Matuska 	error = dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG,
25532a58b312SMartin Matuska 	    &numbufs, &dbp);
25542a58b312SMartin Matuska 	if (error != 0) {
25552a58b312SMartin Matuska 		if (error == ESRCH) {
25562a58b312SMartin Matuska 			error = SET_ERROR(ENXIO);
25572a58b312SMartin Matuska 		}
25582a58b312SMartin Matuska 		return (error);
25592a58b312SMartin Matuska 	}
25602a58b312SMartin Matuska 
25612a58b312SMartin Matuska 	ASSERT3U(numbufs, <=, *nbpsp);
25622a58b312SMartin Matuska 
25632a58b312SMartin Matuska 	for (int i = 0; i < numbufs; i++) {
25642a58b312SMartin Matuska 		dbuf = dbp[i];
25652a58b312SMartin Matuska 		db = (dmu_buf_impl_t *)dbuf;
25662a58b312SMartin Matuska 
25672a58b312SMartin Matuska 		mutex_enter(&db->db_mtx);
25682a58b312SMartin Matuska 
25692a58b312SMartin Matuska 		if (!list_is_empty(&db->db_dirty_records)) {
25702a58b312SMartin Matuska 			dbuf_dirty_record_t *dr;
25712a58b312SMartin Matuska 
25722a58b312SMartin Matuska 			dr = list_head(&db->db_dirty_records);
25732a58b312SMartin Matuska 			if (dr->dt.dl.dr_brtwrite) {
25742a58b312SMartin Matuska 				/*
25752a58b312SMartin Matuska 				 * This is very special case where we clone a
25762a58b312SMartin Matuska 				 * block and in the same transaction group we
25772a58b312SMartin Matuska 				 * read its BP (most likely to clone the clone).
25782a58b312SMartin Matuska 				 */
25792a58b312SMartin Matuska 				bp = &dr->dt.dl.dr_overridden_by;
25802a58b312SMartin Matuska 			} else {
25812a58b312SMartin Matuska 				/*
25822a58b312SMartin Matuska 				 * The block was modified in the same
25832a58b312SMartin Matuska 				 * transaction group.
25842a58b312SMartin Matuska 				 */
25852a58b312SMartin Matuska 				mutex_exit(&db->db_mtx);
25862a58b312SMartin Matuska 				error = SET_ERROR(EAGAIN);
25872a58b312SMartin Matuska 				goto out;
25882a58b312SMartin Matuska 			}
25892a58b312SMartin Matuska 		} else {
25902a58b312SMartin Matuska 			bp = db->db_blkptr;
25912a58b312SMartin Matuska 		}
25922a58b312SMartin Matuska 
25932a58b312SMartin Matuska 		mutex_exit(&db->db_mtx);
25942a58b312SMartin Matuska 
25952a58b312SMartin Matuska 		if (bp == NULL) {
25962a58b312SMartin Matuska 			/*
2597783d3ff6SMartin Matuska 			 * The file size was increased, but the block was never
2598783d3ff6SMartin Matuska 			 * written, otherwise we would either have the block
2599783d3ff6SMartin Matuska 			 * pointer or the dirty record and would not get here.
2600783d3ff6SMartin Matuska 			 * It is effectively a hole, so report it as such.
26012a58b312SMartin Matuska 			 */
2602783d3ff6SMartin Matuska 			BP_ZERO(&bps[i]);
2603783d3ff6SMartin Matuska 			continue;
26042a58b312SMartin Matuska 		}
26052a58b312SMartin Matuska 		/*
26062a58b312SMartin Matuska 		 * Make sure we clone only data blocks.
26072a58b312SMartin Matuska 		 */
26082a58b312SMartin Matuska 		if (BP_IS_METADATA(bp) && !BP_IS_HOLE(bp)) {
26092a58b312SMartin Matuska 			error = SET_ERROR(EINVAL);
26102a58b312SMartin Matuska 			goto out;
26112a58b312SMartin Matuska 		}
26122a58b312SMartin Matuska 
26133494f7c0SMartin Matuska 		/*
26143494f7c0SMartin Matuska 		 * If the block was allocated in transaction group that is not
26153494f7c0SMartin Matuska 		 * yet synced, we could clone it, but we couldn't write this
26163494f7c0SMartin Matuska 		 * operation into ZIL, or it may be impossible to replay, since
26173494f7c0SMartin Matuska 		 * the block may appear not yet allocated at that point.
26183494f7c0SMartin Matuska 		 */
2619783d3ff6SMartin Matuska 		if (BP_GET_BIRTH(bp) > spa_freeze_txg(os->os_spa)) {
26203494f7c0SMartin Matuska 			error = SET_ERROR(EINVAL);
26213494f7c0SMartin Matuska 			goto out;
26223494f7c0SMartin Matuska 		}
2623783d3ff6SMartin Matuska 		if (BP_GET_BIRTH(bp) > spa_last_synced_txg(os->os_spa)) {
26243494f7c0SMartin Matuska 			error = SET_ERROR(EAGAIN);
26253494f7c0SMartin Matuska 			goto out;
26263494f7c0SMartin Matuska 		}
26273494f7c0SMartin Matuska 
26282a58b312SMartin Matuska 		bps[i] = *bp;
26292a58b312SMartin Matuska 	}
26302a58b312SMartin Matuska 
26312a58b312SMartin Matuska 	*nbpsp = numbufs;
26322a58b312SMartin Matuska out:
26332a58b312SMartin Matuska 	dmu_buf_rele_array(dbp, numbufs, FTAG);
26342a58b312SMartin Matuska 
26352a58b312SMartin Matuska 	return (error);
26362a58b312SMartin Matuska }
26372a58b312SMartin Matuska 
2638e639e0d2SMartin Matuska int
26392a58b312SMartin Matuska dmu_brt_clone(objset_t *os, uint64_t object, uint64_t offset, uint64_t length,
2640525fe93dSMartin Matuska     dmu_tx_t *tx, const blkptr_t *bps, size_t nbps)
26412a58b312SMartin Matuska {
26422a58b312SMartin Matuska 	spa_t *spa;
26432a58b312SMartin Matuska 	dmu_buf_t **dbp, *dbuf;
26442a58b312SMartin Matuska 	dmu_buf_impl_t *db;
26452a58b312SMartin Matuska 	struct dirty_leaf *dl;
26462a58b312SMartin Matuska 	dbuf_dirty_record_t *dr;
26472a58b312SMartin Matuska 	const blkptr_t *bp;
2648e639e0d2SMartin Matuska 	int error = 0, i, numbufs;
26492a58b312SMartin Matuska 
26502a58b312SMartin Matuska 	spa = os->os_spa;
26512a58b312SMartin Matuska 
26522a58b312SMartin Matuska 	VERIFY0(dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG,
26532a58b312SMartin Matuska 	    &numbufs, &dbp));
26542a58b312SMartin Matuska 	ASSERT3U(nbps, ==, numbufs);
26552a58b312SMartin Matuska 
2656e639e0d2SMartin Matuska 	/*
2657e639e0d2SMartin Matuska 	 * Before we start cloning make sure that the dbufs sizes match new BPs
2658e639e0d2SMartin Matuska 	 * sizes. If they don't, that's a no-go, as we are not able to shrink
2659e639e0d2SMartin Matuska 	 * dbufs.
2660e639e0d2SMartin Matuska 	 */
2661e639e0d2SMartin Matuska 	for (i = 0; i < numbufs; i++) {
26622a58b312SMartin Matuska 		dbuf = dbp[i];
26632a58b312SMartin Matuska 		db = (dmu_buf_impl_t *)dbuf;
26642a58b312SMartin Matuska 		bp = &bps[i];
26652a58b312SMartin Matuska 
26665c65a0a9SMartin Matuska 		ASSERT3U(db->db.db_object, !=, DMU_META_DNODE_OBJECT);
26672a58b312SMartin Matuska 		ASSERT0(db->db_level);
26682a58b312SMartin Matuska 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2669e639e0d2SMartin Matuska 		ASSERT(db->db_blkid != DMU_SPILL_BLKID);
26702a58b312SMartin Matuska 
2671e639e0d2SMartin Matuska 		if (!BP_IS_HOLE(bp) && BP_GET_LSIZE(bp) != dbuf->db_size) {
2672e639e0d2SMartin Matuska 			error = SET_ERROR(EXDEV);
2673e639e0d2SMartin Matuska 			goto out;
2674e639e0d2SMartin Matuska 		}
26752a58b312SMartin Matuska 	}
26762a58b312SMartin Matuska 
2677e639e0d2SMartin Matuska 	for (i = 0; i < numbufs; i++) {
2678e639e0d2SMartin Matuska 		dbuf = dbp[i];
2679e639e0d2SMartin Matuska 		db = (dmu_buf_impl_t *)dbuf;
2680e639e0d2SMartin Matuska 		bp = &bps[i];
26812a58b312SMartin Matuska 
26827a7741afSMartin Matuska 		dmu_buf_will_clone_or_dio(dbuf, tx);
26832a58b312SMartin Matuska 
26842a58b312SMartin Matuska 		mutex_enter(&db->db_mtx);
26852a58b312SMartin Matuska 
26862a58b312SMartin Matuska 		dr = list_head(&db->db_dirty_records);
26872a58b312SMartin Matuska 		VERIFY(dr != NULL);
26882a58b312SMartin Matuska 		ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
26892a58b312SMartin Matuska 		dl = &dr->dt.dl;
26905c65a0a9SMartin Matuska 		ASSERT0(dl->dr_has_raw_params);
26912a58b312SMartin Matuska 		dl->dr_overridden_by = *bp;
2692783d3ff6SMartin Matuska 		if (!BP_IS_HOLE(bp) || BP_GET_LOGICAL_BIRTH(bp) != 0) {
2693783d3ff6SMartin Matuska 			if (!BP_IS_EMBEDDED(bp)) {
2694783d3ff6SMartin Matuska 				BP_SET_BIRTH(&dl->dr_overridden_by, dr->dr_txg,
2695783d3ff6SMartin Matuska 				    BP_GET_BIRTH(bp));
2696783d3ff6SMartin Matuska 			} else {
2697783d3ff6SMartin Matuska 				BP_SET_LOGICAL_BIRTH(&dl->dr_overridden_by,
2698783d3ff6SMartin Matuska 				    dr->dr_txg);
2699783d3ff6SMartin Matuska 			}
2700783d3ff6SMartin Matuska 		}
27012a58b312SMartin Matuska 		dl->dr_brtwrite = B_TRUE;
27022a58b312SMartin Matuska 		dl->dr_override_state = DR_OVERRIDDEN;
27032a58b312SMartin Matuska 
27042a58b312SMartin Matuska 		mutex_exit(&db->db_mtx);
27052a58b312SMartin Matuska 
27062a58b312SMartin Matuska 		/*
27072a58b312SMartin Matuska 		 * When data in embedded into BP there is no need to create
27082a58b312SMartin Matuska 		 * BRT entry as there is no data block. Just copy the BP as
27092a58b312SMartin Matuska 		 * it contains the data.
27102a58b312SMartin Matuska 		 */
2711525fe93dSMartin Matuska 		if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
27122a58b312SMartin Matuska 			brt_pending_add(spa, bp, tx);
27132a58b312SMartin Matuska 		}
27142a58b312SMartin Matuska 	}
2715e639e0d2SMartin Matuska out:
27162a58b312SMartin Matuska 	dmu_buf_rele_array(dbp, numbufs, FTAG);
2717e639e0d2SMartin Matuska 
2718e639e0d2SMartin Matuska 	return (error);
27192a58b312SMartin Matuska }
27202a58b312SMartin Matuska 
2721eda14cbcSMatt Macy void
2722eda14cbcSMatt Macy __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2723eda14cbcSMatt Macy {
2724eda14cbcSMatt Macy 	dnode_phys_t *dnp = dn->dn_phys;
2725eda14cbcSMatt Macy 
2726eda14cbcSMatt Macy 	doi->doi_data_block_size = dn->dn_datablksz;
2727eda14cbcSMatt Macy 	doi->doi_metadata_block_size = dn->dn_indblkshift ?
2728eda14cbcSMatt Macy 	    1ULL << dn->dn_indblkshift : 0;
2729eda14cbcSMatt Macy 	doi->doi_type = dn->dn_type;
2730eda14cbcSMatt Macy 	doi->doi_bonus_type = dn->dn_bonustype;
2731eda14cbcSMatt Macy 	doi->doi_bonus_size = dn->dn_bonuslen;
2732eda14cbcSMatt Macy 	doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2733eda14cbcSMatt Macy 	doi->doi_indirection = dn->dn_nlevels;
2734eda14cbcSMatt Macy 	doi->doi_checksum = dn->dn_checksum;
2735eda14cbcSMatt Macy 	doi->doi_compress = dn->dn_compress;
2736eda14cbcSMatt Macy 	doi->doi_nblkptr = dn->dn_nblkptr;
2737eda14cbcSMatt Macy 	doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2738eda14cbcSMatt Macy 	doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2739eda14cbcSMatt Macy 	doi->doi_fill_count = 0;
2740eda14cbcSMatt Macy 	for (int i = 0; i < dnp->dn_nblkptr; i++)
2741eda14cbcSMatt Macy 		doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2742eda14cbcSMatt Macy }
2743eda14cbcSMatt Macy 
2744eda14cbcSMatt Macy void
2745eda14cbcSMatt Macy dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2746eda14cbcSMatt Macy {
2747eda14cbcSMatt Macy 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
2748eda14cbcSMatt Macy 	mutex_enter(&dn->dn_mtx);
2749eda14cbcSMatt Macy 
2750eda14cbcSMatt Macy 	__dmu_object_info_from_dnode(dn, doi);
2751eda14cbcSMatt Macy 
2752eda14cbcSMatt Macy 	mutex_exit(&dn->dn_mtx);
2753eda14cbcSMatt Macy 	rw_exit(&dn->dn_struct_rwlock);
2754eda14cbcSMatt Macy }
2755eda14cbcSMatt Macy 
2756eda14cbcSMatt Macy /*
2757eda14cbcSMatt Macy  * Get information on a DMU object.
2758eda14cbcSMatt Macy  * If doi is NULL, just indicates whether the object exists.
2759eda14cbcSMatt Macy  */
2760eda14cbcSMatt Macy int
2761eda14cbcSMatt Macy dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2762eda14cbcSMatt Macy {
2763eda14cbcSMatt Macy 	dnode_t *dn;
2764eda14cbcSMatt Macy 	int err = dnode_hold(os, object, FTAG, &dn);
2765eda14cbcSMatt Macy 
2766eda14cbcSMatt Macy 	if (err)
2767eda14cbcSMatt Macy 		return (err);
2768eda14cbcSMatt Macy 
2769eda14cbcSMatt Macy 	if (doi != NULL)
2770eda14cbcSMatt Macy 		dmu_object_info_from_dnode(dn, doi);
2771eda14cbcSMatt Macy 
2772eda14cbcSMatt Macy 	dnode_rele(dn, FTAG);
2773eda14cbcSMatt Macy 	return (0);
2774eda14cbcSMatt Macy }
2775eda14cbcSMatt Macy 
2776eda14cbcSMatt Macy /*
2777eda14cbcSMatt Macy  * As above, but faster; can be used when you have a held dbuf in hand.
2778eda14cbcSMatt Macy  */
2779eda14cbcSMatt Macy void
2780eda14cbcSMatt Macy dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2781eda14cbcSMatt Macy {
2782eda14cbcSMatt Macy 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2783eda14cbcSMatt Macy 
2784eda14cbcSMatt Macy 	DB_DNODE_ENTER(db);
2785eda14cbcSMatt Macy 	dmu_object_info_from_dnode(DB_DNODE(db), doi);
2786eda14cbcSMatt Macy 	DB_DNODE_EXIT(db);
2787eda14cbcSMatt Macy }
2788eda14cbcSMatt Macy 
2789eda14cbcSMatt Macy /*
2790eda14cbcSMatt Macy  * Faster still when you only care about the size.
2791eda14cbcSMatt Macy  */
2792eda14cbcSMatt Macy void
2793eda14cbcSMatt Macy dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2794eda14cbcSMatt Macy     u_longlong_t *nblk512)
2795eda14cbcSMatt Macy {
2796eda14cbcSMatt Macy 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2797eda14cbcSMatt Macy 	dnode_t *dn;
2798eda14cbcSMatt Macy 
2799eda14cbcSMatt Macy 	DB_DNODE_ENTER(db);
2800eda14cbcSMatt Macy 	dn = DB_DNODE(db);
2801eda14cbcSMatt Macy 
2802eda14cbcSMatt Macy 	*blksize = dn->dn_datablksz;
2803eda14cbcSMatt Macy 	/* add in number of slots used for the dnode itself */
2804eda14cbcSMatt Macy 	*nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2805eda14cbcSMatt Macy 	    SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2806eda14cbcSMatt Macy 	DB_DNODE_EXIT(db);
2807eda14cbcSMatt Macy }
2808eda14cbcSMatt Macy 
2809eda14cbcSMatt Macy void
2810eda14cbcSMatt Macy dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2811eda14cbcSMatt Macy {
2812eda14cbcSMatt Macy 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2813eda14cbcSMatt Macy 
2814eda14cbcSMatt Macy 	DB_DNODE_ENTER(db);
2815ce4dcb97SMartin Matuska 	*dnsize = DB_DNODE(db)->dn_num_slots << DNODE_SHIFT;
2816eda14cbcSMatt Macy 	DB_DNODE_EXIT(db);
2817eda14cbcSMatt Macy }
2818eda14cbcSMatt Macy 
2819eda14cbcSMatt Macy void
2820eda14cbcSMatt Macy byteswap_uint64_array(void *vbuf, size_t size)
2821eda14cbcSMatt Macy {
2822eda14cbcSMatt Macy 	uint64_t *buf = vbuf;
2823eda14cbcSMatt Macy 	size_t count = size >> 3;
2824eda14cbcSMatt Macy 	int i;
2825eda14cbcSMatt Macy 
2826eda14cbcSMatt Macy 	ASSERT((size & 7) == 0);
2827eda14cbcSMatt Macy 
2828eda14cbcSMatt Macy 	for (i = 0; i < count; i++)
2829eda14cbcSMatt Macy 		buf[i] = BSWAP_64(buf[i]);
2830eda14cbcSMatt Macy }
2831eda14cbcSMatt Macy 
2832eda14cbcSMatt Macy void
2833eda14cbcSMatt Macy byteswap_uint32_array(void *vbuf, size_t size)
2834eda14cbcSMatt Macy {
2835eda14cbcSMatt Macy 	uint32_t *buf = vbuf;
2836eda14cbcSMatt Macy 	size_t count = size >> 2;
2837eda14cbcSMatt Macy 	int i;
2838eda14cbcSMatt Macy 
2839eda14cbcSMatt Macy 	ASSERT((size & 3) == 0);
2840eda14cbcSMatt Macy 
2841eda14cbcSMatt Macy 	for (i = 0; i < count; i++)
2842eda14cbcSMatt Macy 		buf[i] = BSWAP_32(buf[i]);
2843eda14cbcSMatt Macy }
2844eda14cbcSMatt Macy 
2845eda14cbcSMatt Macy void
2846eda14cbcSMatt Macy byteswap_uint16_array(void *vbuf, size_t size)
2847eda14cbcSMatt Macy {
2848eda14cbcSMatt Macy 	uint16_t *buf = vbuf;
2849eda14cbcSMatt Macy 	size_t count = size >> 1;
2850eda14cbcSMatt Macy 	int i;
2851eda14cbcSMatt Macy 
2852eda14cbcSMatt Macy 	ASSERT((size & 1) == 0);
2853eda14cbcSMatt Macy 
2854eda14cbcSMatt Macy 	for (i = 0; i < count; i++)
2855eda14cbcSMatt Macy 		buf[i] = BSWAP_16(buf[i]);
2856eda14cbcSMatt Macy }
2857eda14cbcSMatt Macy 
2858eda14cbcSMatt Macy void
2859eda14cbcSMatt Macy byteswap_uint8_array(void *vbuf, size_t size)
2860eda14cbcSMatt Macy {
2861e92ffd9bSMartin Matuska 	(void) vbuf, (void) size;
2862eda14cbcSMatt Macy }
2863eda14cbcSMatt Macy 
2864eda14cbcSMatt Macy void
2865eda14cbcSMatt Macy dmu_init(void)
2866eda14cbcSMatt Macy {
2867eda14cbcSMatt Macy 	abd_init();
2868eda14cbcSMatt Macy 	zfs_dbgmsg_init();
2869eda14cbcSMatt Macy 	sa_cache_init();
2870eda14cbcSMatt Macy 	dmu_objset_init();
2871eda14cbcSMatt Macy 	dnode_init();
2872eda14cbcSMatt Macy 	zfetch_init();
2873eda14cbcSMatt Macy 	dmu_tx_init();
2874eda14cbcSMatt Macy 	l2arc_init();
2875eda14cbcSMatt Macy 	arc_init();
2876eda14cbcSMatt Macy 	dbuf_init();
2877eda14cbcSMatt Macy }
2878eda14cbcSMatt Macy 
2879eda14cbcSMatt Macy void
2880eda14cbcSMatt Macy dmu_fini(void)
2881eda14cbcSMatt Macy {
2882eda14cbcSMatt Macy 	arc_fini(); /* arc depends on l2arc, so arc must go first */
2883eda14cbcSMatt Macy 	l2arc_fini();
2884eda14cbcSMatt Macy 	dmu_tx_fini();
2885eda14cbcSMatt Macy 	zfetch_fini();
2886eda14cbcSMatt Macy 	dbuf_fini();
2887eda14cbcSMatt Macy 	dnode_fini();
2888eda14cbcSMatt Macy 	dmu_objset_fini();
2889eda14cbcSMatt Macy 	sa_cache_fini();
2890eda14cbcSMatt Macy 	zfs_dbgmsg_fini();
2891eda14cbcSMatt Macy 	abd_fini();
2892eda14cbcSMatt Macy }
2893eda14cbcSMatt Macy 
2894eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_bonus_hold);
2895eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_bonus_hold_by_dnode);
2896eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2897eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_buf_rele_array);
2898eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_prefetch);
2899783d3ff6SMartin Matuska EXPORT_SYMBOL(dmu_prefetch_by_dnode);
2900783d3ff6SMartin Matuska EXPORT_SYMBOL(dmu_prefetch_dnode);
2901eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_free_range);
2902eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_free_long_range);
2903eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_free_long_object);
2904eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_read);
2905eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_read_by_dnode);
29067a7741afSMartin Matuska EXPORT_SYMBOL(dmu_read_uio);
29077a7741afSMartin Matuska EXPORT_SYMBOL(dmu_read_uio_dbuf);
29087a7741afSMartin Matuska EXPORT_SYMBOL(dmu_read_uio_dnode);
2909eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_write);
2910eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_write_by_dnode);
29117a7741afSMartin Matuska EXPORT_SYMBOL(dmu_write_by_dnode_flags);
29127a7741afSMartin Matuska EXPORT_SYMBOL(dmu_write_uio);
29137a7741afSMartin Matuska EXPORT_SYMBOL(dmu_write_uio_dbuf);
29147a7741afSMartin Matuska EXPORT_SYMBOL(dmu_write_uio_dnode);
2915eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_prealloc);
2916eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_object_info);
2917eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_object_info_from_dnode);
2918eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_object_info_from_db);
2919eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_object_size_from_db);
2920eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2921eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_object_set_nlevels);
2922eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_object_set_blocksize);
2923eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_object_set_maxblkid);
2924eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_object_set_checksum);
2925eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_object_set_compress);
2926ac0bf12eSMatt Macy EXPORT_SYMBOL(dmu_offset_next);
2927eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_write_policy);
2928eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_sync);
2929eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_request_arcbuf);
2930eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_return_arcbuf);
2931eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode);
2932eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf);
2933eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_buf_hold);
2934eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_ot);
2935eda14cbcSMatt Macy 
2936eda14cbcSMatt Macy ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW,
2937eda14cbcSMatt Macy 	"Enable NOP writes");
2938eda14cbcSMatt Macy 
2939dbd5678dSMartin Matuska ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, UINT, ZMOD_RW,
2940eda14cbcSMatt Macy 	"Percentage of dirtied blocks from frees in one TXG");
2941eda14cbcSMatt Macy 
2942eda14cbcSMatt Macy ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW,
2943eda14cbcSMatt Macy 	"Enable forcing txg sync to find holes");
2944eda14cbcSMatt Macy 
2945be181ee2SMartin Matuska ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, UINT, ZMOD_RW,
2946eda14cbcSMatt Macy 	"Limit one prefetch call to this size");
2947e2df9bb4SMartin Matuska 
2948e2df9bb4SMartin Matuska ZFS_MODULE_PARAM(zfs, , dmu_ddt_copies, UINT, ZMOD_RW,
2949e2df9bb4SMartin Matuska 	"Override copies= for dedup objects");
2950