xref: /illumos-gate/usr/src/uts/common/fs/zfs/sys/dnode.h (revision cdb0ab79ea1af7b8fc339a04d4bf7426dc77ec4e)
1fa9e4066Sahrens /*
2fa9e4066Sahrens  * CDDL HEADER START
3fa9e4066Sahrens  *
4fa9e4066Sahrens  * The contents of this file are subject to the terms of the
5ea8dc4b6Seschrock  * Common Development and Distribution License (the "License").
6ea8dc4b6Seschrock  * You may not use this file except in compliance with the License.
7fa9e4066Sahrens  *
8fa9e4066Sahrens  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9fa9e4066Sahrens  * or http://www.opensolaris.org/os/licensing.
10fa9e4066Sahrens  * See the License for the specific language governing permissions
11fa9e4066Sahrens  * and limitations under the License.
12fa9e4066Sahrens  *
13fa9e4066Sahrens  * When distributing Covered Code, include this CDDL HEADER in each
14fa9e4066Sahrens  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15fa9e4066Sahrens  * If applicable, add the following below this CDDL HEADER, with the
16fa9e4066Sahrens  * fields enclosed by brackets "[]" replaced with your own identifying
17fa9e4066Sahrens  * information: Portions Copyright [yyyy] [name of copyright owner]
18fa9e4066Sahrens  *
19fa9e4066Sahrens  * CDDL HEADER END
20fa9e4066Sahrens  */
21fa9e4066Sahrens /*
22*cdb0ab79Smaybee  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
23fa9e4066Sahrens  * Use is subject to license terms.
24fa9e4066Sahrens  */
25fa9e4066Sahrens 
26fa9e4066Sahrens #ifndef	_SYS_DNODE_H
27fa9e4066Sahrens #define	_SYS_DNODE_H
28fa9e4066Sahrens 
29fa9e4066Sahrens #pragma ident	"%Z%%M%	%I%	%E% SMI"
30fa9e4066Sahrens 
31fa9e4066Sahrens #include <sys/zfs_context.h>
32fa9e4066Sahrens #include <sys/avl.h>
33fa9e4066Sahrens #include <sys/spa.h>
34fa9e4066Sahrens #include <sys/txg.h>
35c717a561Smaybee #include <sys/zio.h>
36fa9e4066Sahrens #include <sys/refcount.h>
37fa9e4066Sahrens #include <sys/dmu_zfetch.h>
38fa9e4066Sahrens 
39fa9e4066Sahrens #ifdef	__cplusplus
40fa9e4066Sahrens extern "C" {
41fa9e4066Sahrens #endif
42fa9e4066Sahrens 
43fa9e4066Sahrens /*
44*cdb0ab79Smaybee  * dnode_hold() flags.
45fa9e4066Sahrens  */
46fa9e4066Sahrens #define	DNODE_MUST_BE_ALLOCATED	1
47fa9e4066Sahrens #define	DNODE_MUST_BE_FREE	2
48fa9e4066Sahrens 
49*cdb0ab79Smaybee /*
50*cdb0ab79Smaybee  * dnode_next_offset() flags.
51*cdb0ab79Smaybee  */
52*cdb0ab79Smaybee #define	DNODE_FIND_HOLE		1
53*cdb0ab79Smaybee #define	DNODE_FIND_BACKWARDS	2
54*cdb0ab79Smaybee #define	DNODE_FIND_HAVELOCK	4
55*cdb0ab79Smaybee 
56fa9e4066Sahrens /*
57fa9e4066Sahrens  * Fixed constants.
58fa9e4066Sahrens  */
59fa9e4066Sahrens #define	DNODE_SHIFT		9	/* 512 bytes */
60fa9e4066Sahrens #define	DN_MIN_INDBLKSHIFT	10	/* 1k */
61fa9e4066Sahrens #define	DN_MAX_INDBLKSHIFT	14	/* 16k */
62fa9e4066Sahrens #define	DNODE_BLOCK_SHIFT	14	/* 16k */
63fa9e4066Sahrens #define	DNODE_CORE_SIZE		64	/* 64 bytes for dnode sans blkptrs */
64fa9e4066Sahrens #define	DN_MAX_OBJECT_SHIFT	48	/* 256 trillion (zfs_fid_t limit) */
65fa9e4066Sahrens #define	DN_MAX_OFFSET_SHIFT	64	/* 2^64 bytes in a dnode */
66fa9e4066Sahrens 
67fa9e4066Sahrens /*
68fa9e4066Sahrens  * Derived constants.
69fa9e4066Sahrens  */
70fa9e4066Sahrens #define	DNODE_SIZE	(1 << DNODE_SHIFT)
71fa9e4066Sahrens #define	DN_MAX_NBLKPTR	((DNODE_SIZE - DNODE_CORE_SIZE) >> SPA_BLKPTRSHIFT)
72fa9e4066Sahrens #define	DN_MAX_BONUSLEN	(DNODE_SIZE - DNODE_CORE_SIZE - (1 << SPA_BLKPTRSHIFT))
73ea8dc4b6Seschrock #define	DN_MAX_OBJECT	(1ULL << DN_MAX_OBJECT_SHIFT)
741934e92fSmaybee #define	DN_ZERO_BONUSLEN	(DN_MAX_BONUSLEN + 1)
75fa9e4066Sahrens 
76fa9e4066Sahrens #define	DNODES_PER_BLOCK_SHIFT	(DNODE_BLOCK_SHIFT - DNODE_SHIFT)
77fa9e4066Sahrens #define	DNODES_PER_BLOCK	(1ULL << DNODES_PER_BLOCK_SHIFT)
78fa9e4066Sahrens #define	DNODES_PER_LEVEL_SHIFT	(DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT)
79fa9e4066Sahrens 
80fa9e4066Sahrens /* The +2 here is a cheesy way to round up */
81fa9e4066Sahrens #define	DN_MAX_LEVELS	(2 + ((DN_MAX_OFFSET_SHIFT - SPA_MINBLOCKSHIFT) / \
82fa9e4066Sahrens 	(DN_MIN_INDBLKSHIFT - SPA_BLKPTRSHIFT)))
83fa9e4066Sahrens 
84fa9e4066Sahrens #define	DN_BONUS(dnp)	((void*)((dnp)->dn_bonus + \
85fa9e4066Sahrens 	(((dnp)->dn_nblkptr - 1) * sizeof (blkptr_t))))
86fa9e4066Sahrens 
8799653d4eSeschrock #define	DN_USED_BYTES(dnp) (((dnp)->dn_flags & DNODE_FLAG_USED_BYTES) ? \
8899653d4eSeschrock 	(dnp)->dn_used : (dnp)->dn_used << SPA_MINBLOCKSHIFT)
8999653d4eSeschrock 
90fa9e4066Sahrens #define	EPB(blkshift, typeshift)	(1 << (blkshift - typeshift))
91fa9e4066Sahrens 
92fa9e4066Sahrens struct dmu_buf_impl;
93fa9e4066Sahrens struct objset_impl;
94fa9e4066Sahrens struct zio;
95fa9e4066Sahrens 
96fa9e4066Sahrens enum dnode_dirtycontext {
97fa9e4066Sahrens 	DN_UNDIRTIED,
98fa9e4066Sahrens 	DN_DIRTY_OPEN,
99fa9e4066Sahrens 	DN_DIRTY_SYNC
100fa9e4066Sahrens };
101fa9e4066Sahrens 
10299653d4eSeschrock /* Is dn_used in bytes?  if not, it's in multiples of SPA_MINBLOCKSIZE */
10399653d4eSeschrock #define	DNODE_FLAG_USED_BYTES	(1<<0)
10499653d4eSeschrock 
105fa9e4066Sahrens typedef struct dnode_phys {
106fa9e4066Sahrens 	uint8_t dn_type;		/* dmu_object_type_t */
107fa9e4066Sahrens 	uint8_t dn_indblkshift;		/* ln2(indirect block size) */
108fa9e4066Sahrens 	uint8_t dn_nlevels;		/* 1=dn_blkptr->data blocks */
109fa9e4066Sahrens 	uint8_t dn_nblkptr;		/* length of dn_blkptr */
110fa9e4066Sahrens 	uint8_t dn_bonustype;		/* type of data in bonus buffer */
111fa9e4066Sahrens 	uint8_t	dn_checksum;		/* ZIO_CHECKSUM type */
112fa9e4066Sahrens 	uint8_t	dn_compress;		/* ZIO_COMPRESS type */
11399653d4eSeschrock 	uint8_t dn_flags;		/* DNODE_FLAG_* */
114fa9e4066Sahrens 	uint16_t dn_datablkszsec;	/* data block size in 512b sectors */
115fa9e4066Sahrens 	uint16_t dn_bonuslen;		/* length of dn_bonus */
116fa9e4066Sahrens 	uint8_t dn_pad2[4];
117fa9e4066Sahrens 
118fa9e4066Sahrens 	/* accounting is protected by dn_dirty_mtx */
119fa9e4066Sahrens 	uint64_t dn_maxblkid;		/* largest allocated block ID */
12099653d4eSeschrock 	uint64_t dn_used;		/* bytes (or sectors) of disk space */
121fa9e4066Sahrens 
122fa9e4066Sahrens 	uint64_t dn_pad3[4];
123fa9e4066Sahrens 
124fa9e4066Sahrens 	blkptr_t dn_blkptr[1];
125fa9e4066Sahrens 	uint8_t dn_bonus[DN_MAX_BONUSLEN];
126fa9e4066Sahrens } dnode_phys_t;
127fa9e4066Sahrens 
128fa9e4066Sahrens typedef struct dnode {
129fa9e4066Sahrens 	/*
130c543ec06Sahrens 	 * dn_struct_rwlock protects the structure of the dnode,
131c543ec06Sahrens 	 * including the number of levels of indirection (dn_nlevels),
132c543ec06Sahrens 	 * dn_maxblkid, and dn_next_*
133fa9e4066Sahrens 	 */
134fa9e4066Sahrens 	krwlock_t dn_struct_rwlock;
135fa9e4066Sahrens 
136fa9e4066Sahrens 	/*
137fa9e4066Sahrens 	 * Our link on dataset's dd_dnodes list.
138fa9e4066Sahrens 	 * Protected by dd_accounting_mtx.
139fa9e4066Sahrens 	 */
140fa9e4066Sahrens 	list_node_t dn_link;
141fa9e4066Sahrens 
142fa9e4066Sahrens 	/* immutable: */
143fa9e4066Sahrens 	struct objset_impl *dn_objset;
144fa9e4066Sahrens 	uint64_t dn_object;
145fa9e4066Sahrens 	struct dmu_buf_impl *dn_dbuf;
146fa9e4066Sahrens 	dnode_phys_t *dn_phys; /* pointer into dn->dn_dbuf->db.db_data */
147fa9e4066Sahrens 
148fa9e4066Sahrens 	/*
149c543ec06Sahrens 	 * Copies of stuff in dn_phys.  They're valid in the open
150c543ec06Sahrens 	 * context (eg. even before the dnode is first synced).
151c543ec06Sahrens 	 * Where necessary, these are protected by dn_struct_rwlock.
152fa9e4066Sahrens 	 */
153c543ec06Sahrens 	dmu_object_type_t dn_type;	/* object type */
154c543ec06Sahrens 	uint16_t dn_bonuslen;		/* bonus length */
155c543ec06Sahrens 	uint8_t dn_bonustype;		/* bonus type */
156fa9e4066Sahrens 	uint8_t dn_nblkptr;		/* number of blkptrs (immutable) */
157fa9e4066Sahrens 	uint8_t dn_checksum;		/* ZIO_CHECKSUM type */
158fa9e4066Sahrens 	uint8_t dn_compress;		/* ZIO_COMPRESS type */
159fa9e4066Sahrens 	uint8_t dn_nlevels;
160fa9e4066Sahrens 	uint8_t dn_indblkshift;
161c543ec06Sahrens 	uint8_t dn_datablkshift;	/* zero if blksz not power of 2! */
162c543ec06Sahrens 	uint16_t dn_datablkszsec;	/* in 512b sectors */
163c543ec06Sahrens 	uint32_t dn_datablksz;		/* in bytes */
164c543ec06Sahrens 	uint64_t dn_maxblkid;
165fa9e4066Sahrens 	uint8_t dn_next_nlevels[TXG_SIZE];
166fa9e4066Sahrens 	uint8_t dn_next_indblkshift[TXG_SIZE];
1671934e92fSmaybee 	uint16_t dn_next_bonuslen[TXG_SIZE];
168c543ec06Sahrens 	uint32_t dn_next_blksz[TXG_SIZE];	/* next block size in bytes */
169fa9e4066Sahrens 
170fa9e4066Sahrens 	/* protected by os_lock: */
171fa9e4066Sahrens 	list_node_t dn_dirty_link[TXG_SIZE];	/* next on dataset's dirty */
172fa9e4066Sahrens 
173fa9e4066Sahrens 	/* protected by dn_mtx: */
174fa9e4066Sahrens 	kmutex_t dn_mtx;
175c717a561Smaybee 	list_t dn_dirty_records[TXG_SIZE];
176fa9e4066Sahrens 	avl_tree_t dn_ranges[TXG_SIZE];
177fa9e4066Sahrens 	uint64_t dn_allocated_txg;
178fa9e4066Sahrens 	uint64_t dn_free_txg;
179fa9e4066Sahrens 	uint64_t dn_assigned_txg;
180fa9e4066Sahrens 	kcondvar_t dn_notxholds;
181fa9e4066Sahrens 	enum dnode_dirtycontext dn_dirtyctx;
182fa9e4066Sahrens 	uint8_t *dn_dirtyctx_firstset;		/* dbg: contents meaningless */
183fa9e4066Sahrens 
184fa9e4066Sahrens 	/* protected by own devices */
185fa9e4066Sahrens 	refcount_t dn_tx_holds;
186fa9e4066Sahrens 	refcount_t dn_holds;
187fa9e4066Sahrens 
188fa9e4066Sahrens 	kmutex_t dn_dbufs_mtx;
189fa9e4066Sahrens 	list_t dn_dbufs;		/* linked list of descendent dbuf_t's */
190ea8dc4b6Seschrock 	struct dmu_buf_impl *dn_bonus;	/* bonus buffer dbuf */
191fa9e4066Sahrens 
192c717a561Smaybee 	/* parent IO for current sync write */
193c717a561Smaybee 	zio_t *dn_zio;
194c717a561Smaybee 
195fa9e4066Sahrens 	/* holds prefetch structure */
196fa9e4066Sahrens 	struct zfetch	dn_zfetch;
197fa9e4066Sahrens } dnode_t;
198fa9e4066Sahrens 
199fa9e4066Sahrens typedef struct free_range {
200fa9e4066Sahrens 	avl_node_t fr_node;
201fa9e4066Sahrens 	uint64_t fr_blkid;
202fa9e4066Sahrens 	uint64_t fr_nblks;
203fa9e4066Sahrens } free_range_t;
204fa9e4066Sahrens 
205fa9e4066Sahrens dnode_t *dnode_special_open(struct objset_impl *dd, dnode_phys_t *dnp,
206fa9e4066Sahrens     uint64_t object);
207fa9e4066Sahrens void dnode_special_close(dnode_t *dn);
208fa9e4066Sahrens 
2091934e92fSmaybee void dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx);
210ea8dc4b6Seschrock int dnode_hold(struct objset_impl *dd, uint64_t object,
211ea8dc4b6Seschrock     void *ref, dnode_t **dnp);
212ea8dc4b6Seschrock int dnode_hold_impl(struct objset_impl *dd, uint64_t object, int flag,
213ea8dc4b6Seschrock     void *ref, dnode_t **dnp);
2141934e92fSmaybee boolean_t dnode_add_ref(dnode_t *dn, void *ref);
215fa9e4066Sahrens void dnode_rele(dnode_t *dn, void *ref);
216fa9e4066Sahrens void dnode_setdirty(dnode_t *dn, dmu_tx_t *tx);
217c717a561Smaybee void dnode_sync(dnode_t *dn, dmu_tx_t *tx);
218fa9e4066Sahrens void dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
219fa9e4066Sahrens     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx);
220fa9e4066Sahrens void dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
221fa9e4066Sahrens     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx);
222fa9e4066Sahrens void dnode_free(dnode_t *dn, dmu_tx_t *tx);
223fa9e4066Sahrens void dnode_byteswap(dnode_phys_t *dnp);
224fa9e4066Sahrens void dnode_buf_byteswap(void *buf, size_t size);
225fa9e4066Sahrens void dnode_verify(dnode_t *dn);
226fa9e4066Sahrens int dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx);
227fa9e4066Sahrens uint64_t dnode_current_max_length(dnode_t *dn);
228fa9e4066Sahrens void dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx);
229fa9e4066Sahrens void dnode_clear_range(dnode_t *dn, uint64_t blkid,
230fa9e4066Sahrens     uint64_t nblks, dmu_tx_t *tx);
231fa9e4066Sahrens void dnode_diduse_space(dnode_t *dn, int64_t space);
232fa9e4066Sahrens void dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx);
233fa9e4066Sahrens void dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx);
234fa9e4066Sahrens uint64_t dnode_block_freed(dnode_t *dn, uint64_t blkid);
235fa9e4066Sahrens void dnode_init(void);
236fa9e4066Sahrens void dnode_fini(void);
237*cdb0ab79Smaybee int dnode_next_offset(dnode_t *dn, int flags, uint64_t *off,
238*cdb0ab79Smaybee     int minlvl, uint64_t blkfill, uint64_t txg);
2391934e92fSmaybee void dnode_evict_dbufs(dnode_t *dn);
240fa9e4066Sahrens 
241fa9e4066Sahrens #ifdef ZFS_DEBUG
242fa9e4066Sahrens 
243fa9e4066Sahrens /*
244fa9e4066Sahrens  * There should be a ## between the string literal and fmt, to make it
245fa9e4066Sahrens  * clear that we're joining two strings together, but that piece of shit
246fa9e4066Sahrens  * gcc doesn't support that preprocessor token.
247fa9e4066Sahrens  */
248fa9e4066Sahrens #define	dprintf_dnode(dn, fmt, ...) do { \
249fa9e4066Sahrens 	if (zfs_flags & ZFS_DEBUG_DPRINTF) { \
250fa9e4066Sahrens 	char __db_buf[32]; \
251fa9e4066Sahrens 	uint64_t __db_obj = (dn)->dn_object; \
252fa9e4066Sahrens 	if (__db_obj == DMU_META_DNODE_OBJECT) \
253fa9e4066Sahrens 		(void) strcpy(__db_buf, "mdn"); \
254fa9e4066Sahrens 	else \
255fa9e4066Sahrens 		(void) snprintf(__db_buf, sizeof (__db_buf), "%lld", \
256fa9e4066Sahrens 		    (u_longlong_t)__db_obj);\
257fa9e4066Sahrens 	dprintf_ds((dn)->dn_objset->os_dsl_dataset, "obj=%s " fmt, \
258fa9e4066Sahrens 	    __db_buf, __VA_ARGS__); \
259fa9e4066Sahrens 	} \
260fa9e4066Sahrens _NOTE(CONSTCOND) } while (0)
261fa9e4066Sahrens 
2629c9dc39aSek #define	DNODE_VERIFY(dn)		dnode_verify(dn)
2639c9dc39aSek #define	FREE_VERIFY(db, start, end, tx)	free_verify(db, start, end, tx)
2649c9dc39aSek 
265fa9e4066Sahrens #else
266fa9e4066Sahrens 
267fa9e4066Sahrens #define	dprintf_dnode(db, fmt, ...)
2689c9dc39aSek #define	DNODE_VERIFY(dn)
2699c9dc39aSek #define	FREE_VERIFY(db, start, end, tx)
270fa9e4066Sahrens 
271fa9e4066Sahrens #endif
272fa9e4066Sahrens 
273fa9e4066Sahrens #ifdef	__cplusplus
274fa9e4066Sahrens }
275fa9e4066Sahrens #endif
276fa9e4066Sahrens 
277fa9e4066Sahrens #endif	/* _SYS_DNODE_H */
278