xref: /titanic_50/usr/src/uts/common/fs/zfs/sys/dnode.h (revision 275c9da86e89f8abf71135cf63d9fc23671b2e60)
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 2007 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/zio.h>
36 #include <sys/refcount.h>
37 #include <sys/dmu_zfetch.h>
38 
39 #ifdef	__cplusplus
40 extern "C" {
41 #endif
42 
43 /*
44  * Flags.
45  */
46 #define	DNODE_MUST_BE_ALLOCATED	1
47 #define	DNODE_MUST_BE_FREE	2
48 
49 /*
50  * Fixed constants.
51  */
52 #define	DNODE_SHIFT		9	/* 512 bytes */
53 #define	DN_MIN_INDBLKSHIFT	10	/* 1k */
54 #define	DN_MAX_INDBLKSHIFT	14	/* 16k */
55 #define	DNODE_BLOCK_SHIFT	14	/* 16k */
56 #define	DNODE_CORE_SIZE		64	/* 64 bytes for dnode sans blkptrs */
57 #define	DN_MAX_OBJECT_SHIFT	48	/* 256 trillion (zfs_fid_t limit) */
58 #define	DN_MAX_OFFSET_SHIFT	64	/* 2^64 bytes in a dnode */
59 
60 /*
61  * Derived constants.
62  */
63 #define	DNODE_SIZE	(1 << DNODE_SHIFT)
64 #define	DN_MAX_NBLKPTR	((DNODE_SIZE - DNODE_CORE_SIZE) >> SPA_BLKPTRSHIFT)
65 #define	DN_MAX_BONUSLEN	(DNODE_SIZE - DNODE_CORE_SIZE - (1 << SPA_BLKPTRSHIFT))
66 #define	DN_MAX_OBJECT	(1ULL << DN_MAX_OBJECT_SHIFT)
67 #define	DN_ZERO_BONUSLEN	(DN_MAX_BONUSLEN + 1)
68 
69 #define	DNODES_PER_BLOCK_SHIFT	(DNODE_BLOCK_SHIFT - DNODE_SHIFT)
70 #define	DNODES_PER_BLOCK	(1ULL << DNODES_PER_BLOCK_SHIFT)
71 #define	DNODES_PER_LEVEL_SHIFT	(DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT)
72 
73 /* The +2 here is a cheesy way to round up */
74 #define	DN_MAX_LEVELS	(2 + ((DN_MAX_OFFSET_SHIFT - SPA_MINBLOCKSHIFT) / \
75 	(DN_MIN_INDBLKSHIFT - SPA_BLKPTRSHIFT)))
76 
77 #define	DN_BONUS(dnp)	((void*)((dnp)->dn_bonus + \
78 	(((dnp)->dn_nblkptr - 1) * sizeof (blkptr_t))))
79 
80 #define	DN_USED_BYTES(dnp) (((dnp)->dn_flags & DNODE_FLAG_USED_BYTES) ? \
81 	(dnp)->dn_used : (dnp)->dn_used << SPA_MINBLOCKSHIFT)
82 
83 #define	EPB(blkshift, typeshift)	(1 << (blkshift - typeshift))
84 
85 struct dmu_buf_impl;
86 struct objset_impl;
87 struct zio;
88 
89 enum dnode_dirtycontext {
90 	DN_UNDIRTIED,
91 	DN_DIRTY_OPEN,
92 	DN_DIRTY_SYNC
93 };
94 
95 /* Is dn_used in bytes?  if not, it's in multiples of SPA_MINBLOCKSIZE */
96 #define	DNODE_FLAG_USED_BYTES	(1<<0)
97 
98 typedef struct dnode_phys {
99 	uint8_t dn_type;		/* dmu_object_type_t */
100 	uint8_t dn_indblkshift;		/* ln2(indirect block size) */
101 	uint8_t dn_nlevels;		/* 1=dn_blkptr->data blocks */
102 	uint8_t dn_nblkptr;		/* length of dn_blkptr */
103 	uint8_t dn_bonustype;		/* type of data in bonus buffer */
104 	uint8_t	dn_checksum;		/* ZIO_CHECKSUM type */
105 	uint8_t	dn_compress;		/* ZIO_COMPRESS type */
106 	uint8_t dn_flags;		/* DNODE_FLAG_* */
107 	uint16_t dn_datablkszsec;	/* data block size in 512b sectors */
108 	uint16_t dn_bonuslen;		/* length of dn_bonus */
109 	uint8_t dn_pad2[4];
110 
111 	/* accounting is protected by dn_dirty_mtx */
112 	uint64_t dn_maxblkid;		/* largest allocated block ID */
113 	uint64_t dn_used;		/* bytes (or sectors) of disk space */
114 
115 	uint64_t dn_pad3[4];
116 
117 	blkptr_t dn_blkptr[1];
118 	uint8_t dn_bonus[DN_MAX_BONUSLEN];
119 } dnode_phys_t;
120 
121 typedef struct dnode {
122 	/*
123 	 * dn_struct_rwlock protects the structure of the dnode,
124 	 * including the number of levels of indirection (dn_nlevels),
125 	 * dn_maxblkid, and dn_next_*
126 	 */
127 	krwlock_t dn_struct_rwlock;
128 
129 	/*
130 	 * Our link on dataset's dd_dnodes list.
131 	 * Protected by dd_accounting_mtx.
132 	 */
133 	list_node_t dn_link;
134 
135 	/* immutable: */
136 	struct objset_impl *dn_objset;
137 	uint64_t dn_object;
138 	struct dmu_buf_impl *dn_dbuf;
139 	dnode_phys_t *dn_phys; /* pointer into dn->dn_dbuf->db.db_data */
140 
141 	/*
142 	 * Copies of stuff in dn_phys.  They're valid in the open
143 	 * context (eg. even before the dnode is first synced).
144 	 * Where necessary, these are protected by dn_struct_rwlock.
145 	 */
146 	dmu_object_type_t dn_type;	/* object type */
147 	uint16_t dn_bonuslen;		/* bonus length */
148 	uint8_t dn_bonustype;		/* bonus type */
149 	uint8_t dn_nblkptr;		/* number of blkptrs (immutable) */
150 	uint8_t dn_checksum;		/* ZIO_CHECKSUM type */
151 	uint8_t dn_compress;		/* ZIO_COMPRESS type */
152 	uint8_t dn_nlevels;
153 	uint8_t dn_indblkshift;
154 	uint8_t dn_datablkshift;	/* zero if blksz not power of 2! */
155 	uint16_t dn_datablkszsec;	/* in 512b sectors */
156 	uint32_t dn_datablksz;		/* in bytes */
157 	uint64_t dn_maxblkid;
158 	uint8_t dn_next_nlevels[TXG_SIZE];
159 	uint8_t dn_next_indblkshift[TXG_SIZE];
160 	uint16_t dn_next_bonuslen[TXG_SIZE];
161 	uint32_t dn_next_blksz[TXG_SIZE];	/* next block size in bytes */
162 
163 	/* protected by os_lock: */
164 	list_node_t dn_dirty_link[TXG_SIZE];	/* next on dataset's dirty */
165 
166 	/* protected by dn_mtx: */
167 	kmutex_t dn_mtx;
168 	list_t dn_dirty_records[TXG_SIZE];
169 	avl_tree_t dn_ranges[TXG_SIZE];
170 	uint64_t dn_allocated_txg;
171 	uint64_t dn_free_txg;
172 	uint64_t dn_assigned_txg;
173 	kcondvar_t dn_notxholds;
174 	enum dnode_dirtycontext dn_dirtyctx;
175 	uint8_t *dn_dirtyctx_firstset;		/* dbg: contents meaningless */
176 
177 	/* protected by own devices */
178 	refcount_t dn_tx_holds;
179 	refcount_t dn_holds;
180 
181 	kmutex_t dn_dbufs_mtx;
182 	list_t dn_dbufs;		/* linked list of descendent dbuf_t's */
183 	struct dmu_buf_impl *dn_bonus;	/* bonus buffer dbuf */
184 
185 	/* parent IO for current sync write */
186 	zio_t *dn_zio;
187 
188 	/* holds prefetch structure */
189 	struct zfetch	dn_zfetch;
190 } dnode_t;
191 
192 typedef struct free_range {
193 	avl_node_t fr_node;
194 	uint64_t fr_blkid;
195 	uint64_t fr_nblks;
196 } free_range_t;
197 
198 dnode_t *dnode_special_open(struct objset_impl *dd, dnode_phys_t *dnp,
199     uint64_t object);
200 void dnode_special_close(dnode_t *dn);
201 
202 void dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx);
203 int dnode_hold(struct objset_impl *dd, uint64_t object,
204     void *ref, dnode_t **dnp);
205 int dnode_hold_impl(struct objset_impl *dd, uint64_t object, int flag,
206     void *ref, dnode_t **dnp);
207 boolean_t dnode_add_ref(dnode_t *dn, void *ref);
208 void dnode_rele(dnode_t *dn, void *ref);
209 void dnode_setdirty(dnode_t *dn, dmu_tx_t *tx);
210 void dnode_sync(dnode_t *dn, dmu_tx_t *tx);
211 void dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
212     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx);
213 void dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
214     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx);
215 void dnode_free(dnode_t *dn, dmu_tx_t *tx);
216 void dnode_byteswap(dnode_phys_t *dnp);
217 void dnode_buf_byteswap(void *buf, size_t size);
218 void dnode_verify(dnode_t *dn);
219 int dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx);
220 uint64_t dnode_current_max_length(dnode_t *dn);
221 void dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx);
222 void dnode_clear_range(dnode_t *dn, uint64_t blkid,
223     uint64_t nblks, dmu_tx_t *tx);
224 void dnode_diduse_space(dnode_t *dn, int64_t space);
225 void dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx);
226 void dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx);
227 uint64_t dnode_block_freed(dnode_t *dn, uint64_t blkid);
228 void dnode_init(void);
229 void dnode_fini(void);
230 int dnode_next_offset(dnode_t *dn, boolean_t hole, uint64_t *off, int minlvl,
231     uint64_t blkfill, uint64_t txg);
232 void dnode_evict_dbufs(dnode_t *dn);
233 
234 #ifdef ZFS_DEBUG
235 
236 /*
237  * There should be a ## between the string literal and fmt, to make it
238  * clear that we're joining two strings together, but that piece of shit
239  * gcc doesn't support that preprocessor token.
240  */
241 #define	dprintf_dnode(dn, fmt, ...) do { \
242 	if (zfs_flags & ZFS_DEBUG_DPRINTF) { \
243 	char __db_buf[32]; \
244 	uint64_t __db_obj = (dn)->dn_object; \
245 	if (__db_obj == DMU_META_DNODE_OBJECT) \
246 		(void) strcpy(__db_buf, "mdn"); \
247 	else \
248 		(void) snprintf(__db_buf, sizeof (__db_buf), "%lld", \
249 		    (u_longlong_t)__db_obj);\
250 	dprintf_ds((dn)->dn_objset->os_dsl_dataset, "obj=%s " fmt, \
251 	    __db_buf, __VA_ARGS__); \
252 	} \
253 _NOTE(CONSTCOND) } while (0)
254 
255 #define	DNODE_VERIFY(dn)		dnode_verify(dn)
256 #define	FREE_VERIFY(db, start, end, tx)	free_verify(db, start, end, tx)
257 
258 #else
259 
260 #define	dprintf_dnode(db, fmt, ...)
261 #define	DNODE_VERIFY(dn)
262 #define	FREE_VERIFY(db, start, end, tx)
263 
264 #endif
265 
266 #ifdef	__cplusplus
267 }
268 #endif
269 
270 #endif	/* _SYS_DNODE_H */
271