xref: /titanic_50/usr/src/uts/common/fs/zfs/sys/dnode.h (revision 4899432a4abc50d8c3d81d5faf257b92517b7716)
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_DNODE_H
27 #define	_SYS_DNODE_H
28 
29 #pragma ident	"%Z%%M%	%I%	%E% SMI"
30 
31 #include <sys/zfs_context.h>
32 #include <sys/avl.h>
33 #include <sys/spa.h>
34 #include <sys/txg.h>
35 #include <sys/refcount.h>
36 #include <sys/dmu_zfetch.h>
37 
38 #ifdef	__cplusplus
39 extern "C" {
40 #endif
41 
42 /*
43  * Flags.
44  */
45 #define	DNODE_MUST_BE_ALLOCATED	1
46 #define	DNODE_MUST_BE_FREE	2
47 
48 /*
49  * Fixed constants.
50  */
51 #define	DNODE_SHIFT		9	/* 512 bytes */
52 #define	DN_MIN_INDBLKSHIFT	10	/* 1k */
53 #define	DN_MAX_INDBLKSHIFT	14	/* 16k */
54 #define	DNODE_BLOCK_SHIFT	14	/* 16k */
55 #define	DNODE_CORE_SIZE		64	/* 64 bytes for dnode sans blkptrs */
56 #define	DN_MAX_OBJECT_SHIFT	48	/* 256 trillion (zfs_fid_t limit) */
57 #define	DN_MAX_OFFSET_SHIFT	64	/* 2^64 bytes in a dnode */
58 
59 /*
60  * Derived constants.
61  */
62 #define	DNODE_SIZE	(1 << DNODE_SHIFT)
63 #define	DN_MAX_NBLKPTR	((DNODE_SIZE - DNODE_CORE_SIZE) >> SPA_BLKPTRSHIFT)
64 #define	DN_MAX_BONUSLEN	(DNODE_SIZE - DNODE_CORE_SIZE - (1 << SPA_BLKPTRSHIFT))
65 #define	DN_MAX_OBJECT	(1ULL << DN_MAX_OBJECT_SHIFT)
66 
67 #define	DNODES_PER_BLOCK_SHIFT	(DNODE_BLOCK_SHIFT - DNODE_SHIFT)
68 #define	DNODES_PER_BLOCK	(1ULL << DNODES_PER_BLOCK_SHIFT)
69 #define	DNODES_PER_LEVEL_SHIFT	(DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT)
70 
71 /* The +2 here is a cheesy way to round up */
72 #define	DN_MAX_LEVELS	(2 + ((DN_MAX_OFFSET_SHIFT - SPA_MINBLOCKSHIFT) / \
73 	(DN_MIN_INDBLKSHIFT - SPA_BLKPTRSHIFT)))
74 
75 #define	DN_BONUS(dnp)	((void*)((dnp)->dn_bonus + \
76 	(((dnp)->dn_nblkptr - 1) * sizeof (blkptr_t))))
77 
78 #define	EPB(blkshift, typeshift)	(1 << (blkshift - typeshift))
79 
80 struct dmu_buf_impl;
81 struct objset_impl;
82 struct zio;
83 
84 enum dnode_dirtycontext {
85 	DN_UNDIRTIED,
86 	DN_DIRTY_OPEN,
87 	DN_DIRTY_SYNC
88 };
89 
90 typedef struct dnode_phys {
91 	uint8_t dn_type;		/* dmu_object_type_t */
92 	uint8_t dn_indblkshift;		/* ln2(indirect block size) */
93 	uint8_t dn_nlevels;		/* 1=dn_blkptr->data blocks */
94 	uint8_t dn_nblkptr;		/* length of dn_blkptr */
95 	uint8_t dn_bonustype;		/* type of data in bonus buffer */
96 	uint8_t	dn_checksum;		/* ZIO_CHECKSUM type */
97 	uint8_t	dn_compress;		/* ZIO_COMPRESS type */
98 	uint8_t dn_pad1[1];
99 	uint16_t dn_datablkszsec;	/* data block size in 512b sectors */
100 	uint16_t dn_bonuslen;		/* length of dn_bonus */
101 	uint8_t dn_pad2[4];
102 
103 	/* accounting is protected by dn_dirty_mtx */
104 	uint64_t dn_maxblkid;		/* largest allocated block ID */
105 	uint64_t dn_secphys;		/* 512b sectors of disk space used */
106 
107 	uint64_t dn_pad3[4];
108 
109 	blkptr_t dn_blkptr[1];
110 	uint8_t dn_bonus[DN_MAX_BONUSLEN];
111 } dnode_phys_t;
112 
113 typedef struct dnode {
114 	/*
115 	 * dn_struct_rwlock protects the structure of the dnode,
116 	 * including the number of levels of indirection (dn_nlevels),
117 	 * dn_maxblkid, and dn_next_*
118 	 */
119 	krwlock_t dn_struct_rwlock;
120 
121 	/*
122 	 * Our link on dataset's dd_dnodes list.
123 	 * Protected by dd_accounting_mtx.
124 	 */
125 	list_node_t dn_link;
126 
127 	/* immutable: */
128 	struct objset_impl *dn_objset;
129 	uint64_t dn_object;
130 	struct dmu_buf_impl *dn_dbuf;
131 	dnode_phys_t *dn_phys; /* pointer into dn->dn_dbuf->db.db_data */
132 
133 	/*
134 	 * Copies of stuff in dn_phys.  They're valid in the open
135 	 * context (eg. even before the dnode is first synced).
136 	 * Where necessary, these are protected by dn_struct_rwlock.
137 	 */
138 	dmu_object_type_t dn_type;	/* object type */
139 	uint16_t dn_bonuslen;		/* bonus length */
140 	uint8_t dn_bonustype;		/* bonus type */
141 	uint8_t dn_nblkptr;		/* number of blkptrs (immutable) */
142 	uint8_t dn_checksum;		/* ZIO_CHECKSUM type */
143 	uint8_t dn_compress;		/* ZIO_COMPRESS type */
144 	uint8_t dn_nlevels;
145 	uint8_t dn_indblkshift;
146 	uint8_t dn_datablkshift;	/* zero if blksz not power of 2! */
147 	uint16_t dn_datablkszsec;	/* in 512b sectors */
148 	uint32_t dn_datablksz;		/* in bytes */
149 	uint64_t dn_maxblkid;
150 	uint8_t dn_next_nlevels[TXG_SIZE];
151 	uint8_t dn_next_indblkshift[TXG_SIZE];
152 	uint32_t dn_next_blksz[TXG_SIZE];	/* next block size in bytes */
153 
154 	/* protected by os_lock: */
155 	list_node_t dn_dirty_link[TXG_SIZE];	/* next on dataset's dirty */
156 
157 	/* protected by dn_mtx: */
158 	kmutex_t dn_mtx;
159 	list_t dn_dirty_dbufs[TXG_SIZE];
160 	avl_tree_t dn_ranges[TXG_SIZE];
161 	uint64_t dn_allocated_txg;
162 	uint64_t dn_free_txg;
163 	uint64_t dn_assigned_txg;
164 	struct dmu_tx *dn_assigned_tx;		/* if only one tx cares */
165 	kcondvar_t dn_notxholds;
166 	enum dnode_dirtycontext dn_dirtyctx;
167 	uint8_t *dn_dirtyctx_firstset;		/* dbg: contents meaningless */
168 
169 	/* protected by own devices */
170 	refcount_t dn_tx_holds;
171 	refcount_t dn_holds;
172 
173 	kmutex_t dn_dbufs_mtx;
174 	list_t dn_dbufs;		/* linked list of descendent dbuf_t's */
175 	struct dmu_buf_impl *dn_bonus;	/* bonus buffer dbuf */
176 
177 	/* holds prefetch structure */
178 	struct zfetch	dn_zfetch;
179 } dnode_t;
180 
181 typedef struct free_range {
182 	avl_node_t fr_node;
183 	uint64_t fr_blkid;
184 	uint64_t fr_nblks;
185 } free_range_t;
186 
187 dnode_t *dnode_special_open(struct objset_impl *dd, dnode_phys_t *dnp,
188     uint64_t object);
189 void dnode_special_close(dnode_t *dn);
190 
191 int dnode_hold(struct objset_impl *dd, uint64_t object,
192     void *ref, dnode_t **dnp);
193 int dnode_hold_impl(struct objset_impl *dd, uint64_t object, int flag,
194     void *ref, dnode_t **dnp);
195 void dnode_add_ref(dnode_t *dn, void *ref);
196 void dnode_rele(dnode_t *dn, void *ref);
197 void dnode_setdirty(dnode_t *dn, dmu_tx_t *tx);
198 int dnode_sync(dnode_t *dn, int level, struct zio *zio, dmu_tx_t *tx);
199 void dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
200     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx);
201 void dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
202     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx);
203 void dnode_free(dnode_t *dn, dmu_tx_t *tx);
204 void dnode_byteswap(dnode_phys_t *dnp);
205 void dnode_buf_byteswap(void *buf, size_t size);
206 void dnode_verify(dnode_t *dn);
207 int dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx);
208 uint64_t dnode_current_max_length(dnode_t *dn);
209 uint64_t dnode_max_nonzero_offset(dnode_t *dn);
210 void dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx);
211 void dnode_clear_range(dnode_t *dn, uint64_t blkid,
212     uint64_t nblks, dmu_tx_t *tx);
213 void dnode_diduse_space(dnode_t *dn, int64_t space);
214 void dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx);
215 void dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx);
216 uint64_t dnode_block_freed(dnode_t *dn, uint64_t blkid);
217 void dnode_init(void);
218 void dnode_fini(void);
219 int dnode_next_offset(dnode_t *dn, boolean_t hole, uint64_t *off, int minlvl,
220     uint64_t blkfill);
221 int dnode_evict_dbufs(dnode_t *dn, int try);
222 
223 #ifdef ZFS_DEBUG
224 
225 /*
226  * There should be a ## between the string literal and fmt, to make it
227  * clear that we're joining two strings together, but that piece of shit
228  * gcc doesn't support that preprocessor token.
229  */
230 #define	dprintf_dnode(dn, fmt, ...) do { \
231 	if (zfs_flags & ZFS_DEBUG_DPRINTF) { \
232 	char __db_buf[32]; \
233 	uint64_t __db_obj = (dn)->dn_object; \
234 	if (__db_obj == DMU_META_DNODE_OBJECT) \
235 		(void) strcpy(__db_buf, "mdn"); \
236 	else \
237 		(void) snprintf(__db_buf, sizeof (__db_buf), "%lld", \
238 		    (u_longlong_t)__db_obj);\
239 	dprintf_ds((dn)->dn_objset->os_dsl_dataset, "obj=%s " fmt, \
240 	    __db_buf, __VA_ARGS__); \
241 	} \
242 _NOTE(CONSTCOND) } while (0)
243 
244 #define	DNODE_VERIFY(dn)		dnode_verify(dn)
245 #define	FREE_VERIFY(db, start, end, tx)	free_verify(db, start, end, tx)
246 
247 #else
248 
249 #define	dprintf_dnode(db, fmt, ...)
250 #define	DNODE_VERIFY(dn)
251 #define	FREE_VERIFY(db, start, end, tx)
252 
253 #endif
254 
255 #ifdef	__cplusplus
256 }
257 #endif
258 
259 #endif	/* _SYS_DNODE_H */
260