xref: /illumos-gate/usr/src/uts/common/fs/zfs/sys/zil.h (revision b19a79ec1a527828a60c4d325ccd8dcbeb2b2e8b)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #ifndef	_SYS_ZIL_H
27 #define	_SYS_ZIL_H
28 
29 #pragma ident	"%Z%%M%	%I%	%E% SMI"
30 
31 #include <sys/types.h>
32 #include <sys/spa.h>
33 #include <sys/zio.h>
34 #include <sys/dmu.h>
35 
36 #ifdef	__cplusplus
37 extern "C" {
38 #endif
39 
40 /*
41  * Intent log format:
42  *
43  * Each objset has its own intent log.  The log header (zil_header_t)
44  * for objset N's intent log is kept in the Nth object of the SPA's
45  * intent_log objset.  The log header points to a chain of log blocks,
46  * each of which contains log records (i.e., transactions) followed by
47  * a log block trailer (zil_trailer_t).  The format of a log record
48  * depends on the record (or transaction) type, but all records begin
49  * with a common structure that defines the type, length, and txg.
50  */
51 
52 /*
53  * Intent log header - this on disk structure holds fields to manage
54  * the log.  All fields are 64 bit to easily handle cross architectures.
55  */
56 typedef struct zil_header {
57 	uint64_t zh_claim_txg;	/* txg in which log blocks were claimed */
58 	uint64_t zh_replay_seq;	/* highest replayed sequence number */
59 	blkptr_t zh_log;	/* log chain */
60 	uint64_t zh_claim_seq;	/* highest claimed sequence number */
61 	uint64_t zh_pad[5];
62 } zil_header_t;
63 
64 /*
65  * Log block trailer - structure at the end of the header and each log block
66  *
67  * The zit_bt contains a zbt_cksum which for the intent log is
68  * the sequence number of this log block. A seq of 0 is invalid.
69  * The zbt_cksum is checked by the SPA against the sequence
70  * number passed in the blk_cksum field of the blkptr_t
71  */
72 typedef struct zil_trailer {
73 	uint64_t zit_pad;
74 	blkptr_t zit_next_blk;	/* next block in chain */
75 	uint64_t zit_nused;	/* bytes in log block used */
76 	zio_block_tail_t zit_bt; /* block trailer */
77 } zil_trailer_t;
78 
79 #define	ZIL_MIN_BLKSZ	4096ULL
80 #define	ZIL_MAX_BLKSZ	SPA_MAXBLOCKSIZE
81 #define	ZIL_BLK_DATA_SZ(lwb)	((lwb)->lwb_sz - sizeof (zil_trailer_t))
82 
83 /*
84  * The words of a log block checksum.
85  */
86 #define	ZIL_ZC_GUID_0	0
87 #define	ZIL_ZC_GUID_1	1
88 #define	ZIL_ZC_OBJSET	2
89 #define	ZIL_ZC_SEQ	3
90 
91 /*
92  * Intent log transaction types and record structures
93  */
94 #define	TX_CREATE	1		/* Create file */
95 #define	TX_MKDIR	2		/* Make directory */
96 #define	TX_MKXATTR	3		/* Make XATTR directory */
97 #define	TX_SYMLINK	4		/* Create symbolic link to a file */
98 #define	TX_REMOVE	5		/* Remove file */
99 #define	TX_RMDIR	6		/* Remove directory */
100 #define	TX_LINK		7		/* Create hard link to a file */
101 #define	TX_RENAME	8		/* Rename a file */
102 #define	TX_WRITE	9		/* File write */
103 #define	TX_TRUNCATE	10		/* Truncate a file */
104 #define	TX_SETATTR	11		/* Set file attributes */
105 #define	TX_ACL		12		/* Set acl */
106 #define	TX_MAX_TYPE	13		/* Max transaction type */
107 
108 /*
109  * Format of log records.
110  * The fields are carefully defined to allow them to be aligned
111  * and sized the same on sparc & intel architectures.
112  * Each log record has a common structure at the beginning.
113  *
114  * Note, lrc_seq holds two different sequence numbers. Whilst in memory
115  * it contains the transaction sequence number.  The log record on
116  * disk holds the sequence number of all log records which is used to
117  * ensure we don't replay the same record.  The two sequence numbers are
118  * different because the transactions can now be pushed out of order.
119  */
120 typedef struct {			/* common log record header */
121 	uint64_t	lrc_txtype;	/* intent log transaction type */
122 	uint64_t	lrc_reclen;	/* transaction record length */
123 	uint64_t	lrc_txg;	/* dmu transaction group number */
124 	uint64_t	lrc_seq;	/* see comment above */
125 } lr_t;
126 
127 typedef struct {
128 	lr_t		lr_common;	/* common portion of log record */
129 	uint64_t	lr_doid;	/* object id of directory */
130 	uint64_t	lr_foid;	/* object id of created file object */
131 	uint64_t	lr_mode;	/* mode of object */
132 	uint64_t	lr_uid;		/* uid of object */
133 	uint64_t	lr_gid;		/* gid of object */
134 	uint64_t	lr_gen;		/* generation (txg of creation) */
135 	uint64_t	lr_crtime[2];	/* creation time */
136 	uint64_t	lr_rdev;	/* rdev of object to create */
137 	/* name of object to create follows this */
138 	/* for symlinks, link content follows name */
139 } lr_create_t;
140 
141 typedef struct {
142 	lr_t		lr_common;	/* common portion of log record */
143 	uint64_t	lr_doid;	/* obj id of directory */
144 	/* name of object to remove follows this */
145 } lr_remove_t;
146 
147 typedef struct {
148 	lr_t		lr_common;	/* common portion of log record */
149 	uint64_t	lr_doid;	/* obj id of directory */
150 	uint64_t	lr_link_obj;	/* obj id of link */
151 	/* name of object to link follows this */
152 } lr_link_t;
153 
154 typedef struct {
155 	lr_t		lr_common;	/* common portion of log record */
156 	uint64_t	lr_sdoid;	/* obj id of source directory */
157 	uint64_t	lr_tdoid;	/* obj id of target directory */
158 	/* 2 strings: names of source and destination follow this */
159 } lr_rename_t;
160 
161 typedef struct {
162 	lr_t		lr_common;	/* common portion of log record */
163 	uint64_t	lr_foid;	/* file object to write */
164 	uint64_t	lr_offset;	/* offset to write to */
165 	uint64_t	lr_length;	/* user data length to write */
166 	uint64_t	lr_blkoff;	/* offset represented by lr_blkptr */
167 	blkptr_t	lr_blkptr;	/* spa block pointer for replay */
168 	/* write data will follow for small writes */
169 } lr_write_t;
170 
171 typedef struct {
172 	lr_t		lr_common;	/* common portion of log record */
173 	uint64_t	lr_foid;	/* object id of file to truncate */
174 	uint64_t	lr_offset;	/* offset to truncate from */
175 	uint64_t	lr_length;	/* length to truncate */
176 } lr_truncate_t;
177 
178 typedef struct {
179 	lr_t		lr_common;	/* common portion of log record */
180 	uint64_t	lr_foid;	/* file object to change attributes */
181 	uint64_t	lr_mask;	/* mask of attributes to set */
182 	uint64_t	lr_mode;	/* mode to set */
183 	uint64_t	lr_uid;		/* uid to set */
184 	uint64_t	lr_gid;		/* gid to set */
185 	uint64_t	lr_size;	/* size to set */
186 	uint64_t	lr_atime[2];	/* access time */
187 	uint64_t	lr_mtime[2];	/* modification time */
188 } lr_setattr_t;
189 
190 typedef struct {
191 	lr_t		lr_common;	/* common portion of log record */
192 	uint64_t	lr_foid;	/* obj id of file */
193 	uint64_t	lr_aclcnt;	/* number of acl entries */
194 	/* lr_aclcnt number of ace_t entries follow this */
195 } lr_acl_t;
196 
197 /*
198  * ZIL structure definitions, interface function prototype and globals.
199  */
200 
201 /*
202  * ZFS intent log transaction structure
203  */
204 typedef enum {
205 	WR_INDIRECT,	/* indirect - a large write (dmu_sync() data */
206 			/* and put blkptr in log, rather than actual data) */
207 	WR_COPIED,	/* immediate - data is copied into lr_write_t */
208 	WR_NEED_COPY,	/* immediate - data needs to be copied if pushed */
209 } itx_wr_state_t;
210 
211 typedef struct itx {
212 	list_node_t	itx_node;	/* linkage on zl_itx_list */
213 	void		*itx_private;	/* type-specific opaque data */
214 	itx_wr_state_t	itx_wr_state;	/* write state */
215 	lr_t		itx_lr;		/* common part of log record */
216 	/* followed by type-specific part of lr_xx_t and its immediate data */
217 } itx_t;
218 
219 typedef void zil_parse_blk_func_t(zilog_t *zilog, blkptr_t *bp, void *arg,
220     uint64_t txg);
221 typedef void zil_parse_lr_func_t(zilog_t *zilog, lr_t *lr, void *arg,
222     uint64_t txg);
223 typedef int zil_replay_func_t();
224 typedef int zil_get_data_t(void *arg, lr_write_t *lr, char *dbuf, zio_t *zio);
225 
226 extern uint64_t	zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
227     zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg);
228 
229 extern void	zil_init(void);
230 extern void	zil_fini(void);
231 
232 extern zilog_t	*zil_alloc(objset_t *os, zil_header_t *zh_phys);
233 extern void	zil_free(zilog_t *zilog);
234 
235 extern zilog_t	*zil_open(objset_t *os, zil_get_data_t *get_data);
236 extern void	zil_close(zilog_t *zilog);
237 
238 extern void	zil_replay(objset_t *os, void *arg, uint64_t *txgp,
239     zil_replay_func_t *replay_func[TX_MAX_TYPE], void (*rm_wait)(void *));
240 extern void	zil_destroy(zilog_t *zilog, boolean_t keep_first);
241 
242 extern itx_t	*zil_itx_create(int txtype, size_t lrsize);
243 extern uint64_t zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx);
244 
245 extern void	zil_commit(zilog_t *zilog, uint64_t seq, uint64_t oid);
246 
247 extern int	zil_claim(char *osname, void *txarg);
248 extern void	zil_sync(zilog_t *zilog, dmu_tx_t *tx);
249 extern void	zil_clean(zilog_t *zilog);
250 extern int	zil_is_committed(zilog_t *zilog);
251 
252 extern int	zil_suspend(zilog_t *zilog);
253 extern void	zil_resume(zilog_t *zilog);
254 
255 extern int zil_disable;
256 
257 #ifdef	__cplusplus
258 }
259 #endif
260 
261 #endif	/* _SYS_ZIL_H */
262