xref: /illumos-gate/usr/src/uts/common/fs/zfs/sys/dnode.h (revision 28d97a71b8094bcc695c914ba67d41bee3cd3a8a)
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*28d97a71SMark Shellenbaum  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23fa9e4066Sahrens  */
24fa9e4066Sahrens 
25fa9e4066Sahrens #ifndef	_SYS_DNODE_H
26fa9e4066Sahrens #define	_SYS_DNODE_H
27fa9e4066Sahrens 
28fa9e4066Sahrens #include <sys/zfs_context.h>
29fa9e4066Sahrens #include <sys/avl.h>
30fa9e4066Sahrens #include <sys/spa.h>
31fa9e4066Sahrens #include <sys/txg.h>
32c717a561Smaybee #include <sys/zio.h>
33fa9e4066Sahrens #include <sys/refcount.h>
34fa9e4066Sahrens #include <sys/dmu_zfetch.h>
35fa9e4066Sahrens 
36fa9e4066Sahrens #ifdef	__cplusplus
37fa9e4066Sahrens extern "C" {
38fa9e4066Sahrens #endif
39fa9e4066Sahrens 
40fa9e4066Sahrens /*
41cdb0ab79Smaybee  * dnode_hold() flags.
42fa9e4066Sahrens  */
43fa9e4066Sahrens #define	DNODE_MUST_BE_ALLOCATED	1
44fa9e4066Sahrens #define	DNODE_MUST_BE_FREE	2
45fa9e4066Sahrens 
46cdb0ab79Smaybee /*
47cdb0ab79Smaybee  * dnode_next_offset() flags.
48cdb0ab79Smaybee  */
49cdb0ab79Smaybee #define	DNODE_FIND_HOLE		1
50cdb0ab79Smaybee #define	DNODE_FIND_BACKWARDS	2
51cdb0ab79Smaybee #define	DNODE_FIND_HAVELOCK	4
52cdb0ab79Smaybee 
53fa9e4066Sahrens /*
54fa9e4066Sahrens  * Fixed constants.
55fa9e4066Sahrens  */
56fa9e4066Sahrens #define	DNODE_SHIFT		9	/* 512 bytes */
57fa9e4066Sahrens #define	DN_MIN_INDBLKSHIFT	10	/* 1k */
58fa9e4066Sahrens #define	DN_MAX_INDBLKSHIFT	14	/* 16k */
59fa9e4066Sahrens #define	DNODE_BLOCK_SHIFT	14	/* 16k */
60fa9e4066Sahrens #define	DNODE_CORE_SIZE		64	/* 64 bytes for dnode sans blkptrs */
61fa9e4066Sahrens #define	DN_MAX_OBJECT_SHIFT	48	/* 256 trillion (zfs_fid_t limit) */
62fa9e4066Sahrens #define	DN_MAX_OFFSET_SHIFT	64	/* 2^64 bytes in a dnode */
63fa9e4066Sahrens 
640a586ceaSMark Shellenbaum /*
650a586ceaSMark Shellenbaum  * dnode id flags
660a586ceaSMark Shellenbaum  *
670a586ceaSMark Shellenbaum  * Note: a file will never ever have its
680a586ceaSMark Shellenbaum  * ids moved from bonus->spill
690a586ceaSMark Shellenbaum  * and only in a crypto environment would it be on spill
700a586ceaSMark Shellenbaum  */
710a586ceaSMark Shellenbaum #define	DN_ID_CHKED_BONUS	0x1
720a586ceaSMark Shellenbaum #define	DN_ID_CHKED_SPILL	0x2
730a586ceaSMark Shellenbaum #define	DN_ID_OLD_EXIST		0x4
740a586ceaSMark Shellenbaum #define	DN_ID_NEW_EXIST		0x8
750a586ceaSMark Shellenbaum 
76fa9e4066Sahrens /*
77fa9e4066Sahrens  * Derived constants.
78fa9e4066Sahrens  */
79fa9e4066Sahrens #define	DNODE_SIZE	(1 << DNODE_SHIFT)
80fa9e4066Sahrens #define	DN_MAX_NBLKPTR	((DNODE_SIZE - DNODE_CORE_SIZE) >> SPA_BLKPTRSHIFT)
81fa9e4066Sahrens #define	DN_MAX_BONUSLEN	(DNODE_SIZE - DNODE_CORE_SIZE - (1 << SPA_BLKPTRSHIFT))
82ea8dc4b6Seschrock #define	DN_MAX_OBJECT	(1ULL << DN_MAX_OBJECT_SHIFT)
831934e92fSmaybee #define	DN_ZERO_BONUSLEN	(DN_MAX_BONUSLEN + 1)
840a586ceaSMark Shellenbaum #define	DN_KILL_SPILLBLK (1)
85fa9e4066Sahrens 
86fa9e4066Sahrens #define	DNODES_PER_BLOCK_SHIFT	(DNODE_BLOCK_SHIFT - DNODE_SHIFT)
87fa9e4066Sahrens #define	DNODES_PER_BLOCK	(1ULL << DNODES_PER_BLOCK_SHIFT)
88fa9e4066Sahrens #define	DNODES_PER_LEVEL_SHIFT	(DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT)
89c1449561SEric Taylor #define	DNODES_PER_LEVEL	(1ULL << DNODES_PER_LEVEL_SHIFT)
90fa9e4066Sahrens 
91fa9e4066Sahrens /* The +2 here is a cheesy way to round up */
92fa9e4066Sahrens #define	DN_MAX_LEVELS	(2 + ((DN_MAX_OFFSET_SHIFT - SPA_MINBLOCKSHIFT) / \
93fa9e4066Sahrens 	(DN_MIN_INDBLKSHIFT - SPA_BLKPTRSHIFT)))
94fa9e4066Sahrens 
95fa9e4066Sahrens #define	DN_BONUS(dnp)	((void*)((dnp)->dn_bonus + \
96fa9e4066Sahrens 	(((dnp)->dn_nblkptr - 1) * sizeof (blkptr_t))))
97fa9e4066Sahrens 
9899653d4eSeschrock #define	DN_USED_BYTES(dnp) (((dnp)->dn_flags & DNODE_FLAG_USED_BYTES) ? \
9999653d4eSeschrock 	(dnp)->dn_used : (dnp)->dn_used << SPA_MINBLOCKSHIFT)
10099653d4eSeschrock 
101fa9e4066Sahrens #define	EPB(blkshift, typeshift)	(1 << (blkshift - typeshift))
102fa9e4066Sahrens 
103fa9e4066Sahrens struct dmu_buf_impl;
104503ad85cSMatthew Ahrens struct objset;
105fa9e4066Sahrens struct zio;
106fa9e4066Sahrens 
107fa9e4066Sahrens enum dnode_dirtycontext {
108fa9e4066Sahrens 	DN_UNDIRTIED,
109fa9e4066Sahrens 	DN_DIRTY_OPEN,
110fa9e4066Sahrens 	DN_DIRTY_SYNC
111fa9e4066Sahrens };
112fa9e4066Sahrens 
11399653d4eSeschrock /* Is dn_used in bytes?  if not, it's in multiples of SPA_MINBLOCKSIZE */
11414843421SMatthew Ahrens #define	DNODE_FLAG_USED_BYTES		(1<<0)
11514843421SMatthew Ahrens #define	DNODE_FLAG_USERUSED_ACCOUNTED	(1<<1)
11699653d4eSeschrock 
1170a586ceaSMark Shellenbaum /* Does dnode have a SA spill blkptr in bonus? */
1180a586ceaSMark Shellenbaum #define	DNODE_FLAG_SPILL_BLKPTR	(1<<2)
1190a586ceaSMark Shellenbaum 
120fa9e4066Sahrens typedef struct dnode_phys {
121fa9e4066Sahrens 	uint8_t dn_type;		/* dmu_object_type_t */
122fa9e4066Sahrens 	uint8_t dn_indblkshift;		/* ln2(indirect block size) */
123fa9e4066Sahrens 	uint8_t dn_nlevels;		/* 1=dn_blkptr->data blocks */
124fa9e4066Sahrens 	uint8_t dn_nblkptr;		/* length of dn_blkptr */
125fa9e4066Sahrens 	uint8_t dn_bonustype;		/* type of data in bonus buffer */
126fa9e4066Sahrens 	uint8_t	dn_checksum;		/* ZIO_CHECKSUM type */
127fa9e4066Sahrens 	uint8_t	dn_compress;		/* ZIO_COMPRESS type */
12899653d4eSeschrock 	uint8_t dn_flags;		/* DNODE_FLAG_* */
129fa9e4066Sahrens 	uint16_t dn_datablkszsec;	/* data block size in 512b sectors */
130fa9e4066Sahrens 	uint16_t dn_bonuslen;		/* length of dn_bonus */
131fa9e4066Sahrens 	uint8_t dn_pad2[4];
132fa9e4066Sahrens 
133fa9e4066Sahrens 	/* accounting is protected by dn_dirty_mtx */
134fa9e4066Sahrens 	uint64_t dn_maxblkid;		/* largest allocated block ID */
13599653d4eSeschrock 	uint64_t dn_used;		/* bytes (or sectors) of disk space */
136fa9e4066Sahrens 
137fa9e4066Sahrens 	uint64_t dn_pad3[4];
138fa9e4066Sahrens 
139fa9e4066Sahrens 	blkptr_t dn_blkptr[1];
1400a586ceaSMark Shellenbaum 	uint8_t dn_bonus[DN_MAX_BONUSLEN - sizeof (blkptr_t)];
1410a586ceaSMark Shellenbaum 	blkptr_t dn_spill;
142fa9e4066Sahrens } dnode_phys_t;
143fa9e4066Sahrens 
144fa9e4066Sahrens typedef struct dnode {
145fa9e4066Sahrens 	/*
146c543ec06Sahrens 	 * dn_struct_rwlock protects the structure of the dnode,
147c543ec06Sahrens 	 * including the number of levels of indirection (dn_nlevels),
148c543ec06Sahrens 	 * dn_maxblkid, and dn_next_*
149fa9e4066Sahrens 	 */
150fa9e4066Sahrens 	krwlock_t dn_struct_rwlock;
151fa9e4066Sahrens 
15214843421SMatthew Ahrens 	/* Our link on dn_objset->os_dnodes list; protected by os_lock.  */
153fa9e4066Sahrens 	list_node_t dn_link;
154fa9e4066Sahrens 
155fa9e4066Sahrens 	/* immutable: */
156503ad85cSMatthew Ahrens 	struct objset *dn_objset;
157fa9e4066Sahrens 	uint64_t dn_object;
158fa9e4066Sahrens 	struct dmu_buf_impl *dn_dbuf;
159fa9e4066Sahrens 	dnode_phys_t *dn_phys; /* pointer into dn->dn_dbuf->db.db_data */
160fa9e4066Sahrens 
161fa9e4066Sahrens 	/*
162c543ec06Sahrens 	 * Copies of stuff in dn_phys.  They're valid in the open
163c543ec06Sahrens 	 * context (eg. even before the dnode is first synced).
164c543ec06Sahrens 	 * Where necessary, these are protected by dn_struct_rwlock.
165fa9e4066Sahrens 	 */
166c543ec06Sahrens 	dmu_object_type_t dn_type;	/* object type */
167c543ec06Sahrens 	uint16_t dn_bonuslen;		/* bonus length */
168c543ec06Sahrens 	uint8_t dn_bonustype;		/* bonus type */
169fa9e4066Sahrens 	uint8_t dn_nblkptr;		/* number of blkptrs (immutable) */
170fa9e4066Sahrens 	uint8_t dn_checksum;		/* ZIO_CHECKSUM type */
171fa9e4066Sahrens 	uint8_t dn_compress;		/* ZIO_COMPRESS type */
172fa9e4066Sahrens 	uint8_t dn_nlevels;
173fa9e4066Sahrens 	uint8_t dn_indblkshift;
174c543ec06Sahrens 	uint8_t dn_datablkshift;	/* zero if blksz not power of 2! */
175c543ec06Sahrens 	uint16_t dn_datablkszsec;	/* in 512b sectors */
176c543ec06Sahrens 	uint32_t dn_datablksz;		/* in bytes */
177c543ec06Sahrens 	uint64_t dn_maxblkid;
178da03de99SMark Maybee 	uint8_t dn_next_nblkptr[TXG_SIZE];
179fa9e4066Sahrens 	uint8_t dn_next_nlevels[TXG_SIZE];
180fa9e4066Sahrens 	uint8_t dn_next_indblkshift[TXG_SIZE];
1810a586ceaSMark Shellenbaum 	uint8_t dn_next_bonustype[TXG_SIZE];
1820a586ceaSMark Shellenbaum 	uint8_t dn_rm_spillblk[TXG_SIZE];	/* for removing spill blk */
1831934e92fSmaybee 	uint16_t dn_next_bonuslen[TXG_SIZE];
184c543ec06Sahrens 	uint32_t dn_next_blksz[TXG_SIZE];	/* next block size in bytes */
185fa9e4066Sahrens 
186fa9e4066Sahrens 	/* protected by os_lock: */
187fa9e4066Sahrens 	list_node_t dn_dirty_link[TXG_SIZE];	/* next on dataset's dirty */
188fa9e4066Sahrens 
189fa9e4066Sahrens 	/* protected by dn_mtx: */
190fa9e4066Sahrens 	kmutex_t dn_mtx;
191c717a561Smaybee 	list_t dn_dirty_records[TXG_SIZE];
192fa9e4066Sahrens 	avl_tree_t dn_ranges[TXG_SIZE];
193fa9e4066Sahrens 	uint64_t dn_allocated_txg;
194fa9e4066Sahrens 	uint64_t dn_free_txg;
195fa9e4066Sahrens 	uint64_t dn_assigned_txg;
196fa9e4066Sahrens 	kcondvar_t dn_notxholds;
197fa9e4066Sahrens 	enum dnode_dirtycontext dn_dirtyctx;
198fa9e4066Sahrens 	uint8_t *dn_dirtyctx_firstset;		/* dbg: contents meaningless */
199fa9e4066Sahrens 
200fa9e4066Sahrens 	/* protected by own devices */
201fa9e4066Sahrens 	refcount_t dn_tx_holds;
202fa9e4066Sahrens 	refcount_t dn_holds;
203fa9e4066Sahrens 
204fa9e4066Sahrens 	kmutex_t dn_dbufs_mtx;
205fa9e4066Sahrens 	list_t dn_dbufs;		/* linked list of descendent dbuf_t's */
206ea8dc4b6Seschrock 	struct dmu_buf_impl *dn_bonus;	/* bonus buffer dbuf */
2070a586ceaSMark Shellenbaum 	boolean_t dn_have_spill;	/* have spill or are spilling */
208fa9e4066Sahrens 
209c717a561Smaybee 	/* parent IO for current sync write */
210c717a561Smaybee 	zio_t *dn_zio;
211c717a561Smaybee 
21214843421SMatthew Ahrens 	/* used in syncing context */
2130a586ceaSMark Shellenbaum 	uint64_t dn_oldused;	/* old phys used bytes */
2140a586ceaSMark Shellenbaum 	uint64_t dn_oldflags;	/* old phys dn_flags */
2150a586ceaSMark Shellenbaum 	uint64_t dn_olduid, dn_oldgid;
2160a586ceaSMark Shellenbaum 	uint64_t dn_newuid, dn_newgid;
2170a586ceaSMark Shellenbaum 	int dn_id_flags;
21814843421SMatthew Ahrens 
219fa9e4066Sahrens 	/* holds prefetch structure */
220fa9e4066Sahrens 	struct zfetch	dn_zfetch;
221fa9e4066Sahrens } dnode_t;
222fa9e4066Sahrens 
223fa9e4066Sahrens typedef struct free_range {
224fa9e4066Sahrens 	avl_node_t fr_node;
225fa9e4066Sahrens 	uint64_t fr_blkid;
226fa9e4066Sahrens 	uint64_t fr_nblks;
227fa9e4066Sahrens } free_range_t;
228fa9e4066Sahrens 
229503ad85cSMatthew Ahrens dnode_t *dnode_special_open(struct objset *dd, dnode_phys_t *dnp,
230fa9e4066Sahrens     uint64_t object);
231fa9e4066Sahrens void dnode_special_close(dnode_t *dn);
232fa9e4066Sahrens 
2331934e92fSmaybee void dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx);
2340a586ceaSMark Shellenbaum void dnode_setbonus_type(dnode_t *dn, dmu_object_type_t, dmu_tx_t *tx);
2350a586ceaSMark Shellenbaum void dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx);
2360a586ceaSMark Shellenbaum 
237503ad85cSMatthew Ahrens int dnode_hold(struct objset *dd, uint64_t object,
238ea8dc4b6Seschrock     void *ref, dnode_t **dnp);
239503ad85cSMatthew Ahrens int dnode_hold_impl(struct objset *dd, uint64_t object, int flag,
240ea8dc4b6Seschrock     void *ref, dnode_t **dnp);
2411934e92fSmaybee boolean_t dnode_add_ref(dnode_t *dn, void *ref);
242fa9e4066Sahrens void dnode_rele(dnode_t *dn, void *ref);
243fa9e4066Sahrens void dnode_setdirty(dnode_t *dn, dmu_tx_t *tx);
244c717a561Smaybee void dnode_sync(dnode_t *dn, dmu_tx_t *tx);
245fa9e4066Sahrens void dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
246fa9e4066Sahrens     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx);
247fa9e4066Sahrens void dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
248fa9e4066Sahrens     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx);
249fa9e4066Sahrens void dnode_free(dnode_t *dn, dmu_tx_t *tx);
250fa9e4066Sahrens void dnode_byteswap(dnode_phys_t *dnp);
251fa9e4066Sahrens void dnode_buf_byteswap(void *buf, size_t size);
252fa9e4066Sahrens void dnode_verify(dnode_t *dn);
253fa9e4066Sahrens int dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx);
254fa9e4066Sahrens uint64_t dnode_current_max_length(dnode_t *dn);
255fa9e4066Sahrens void dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx);
256fa9e4066Sahrens void dnode_clear_range(dnode_t *dn, uint64_t blkid,
257fa9e4066Sahrens     uint64_t nblks, dmu_tx_t *tx);
258fa9e4066Sahrens void dnode_diduse_space(dnode_t *dn, int64_t space);
259fa9e4066Sahrens void dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx);
2608346f03fSJonathan W Adams void dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t);
261fa9e4066Sahrens uint64_t dnode_block_freed(dnode_t *dn, uint64_t blkid);
262fa9e4066Sahrens void dnode_init(void);
263fa9e4066Sahrens void dnode_fini(void);
264cdb0ab79Smaybee int dnode_next_offset(dnode_t *dn, int flags, uint64_t *off,
265cdb0ab79Smaybee     int minlvl, uint64_t blkfill, uint64_t txg);
2661934e92fSmaybee void dnode_evict_dbufs(dnode_t *dn);
267fa9e4066Sahrens 
268fa9e4066Sahrens #ifdef ZFS_DEBUG
269fa9e4066Sahrens 
270fa9e4066Sahrens /*
271fa9e4066Sahrens  * There should be a ## between the string literal and fmt, to make it
272fa9e4066Sahrens  * clear that we're joining two strings together, but that piece of shit
273fa9e4066Sahrens  * gcc doesn't support that preprocessor token.
274fa9e4066Sahrens  */
275fa9e4066Sahrens #define	dprintf_dnode(dn, fmt, ...) do { \
276fa9e4066Sahrens 	if (zfs_flags & ZFS_DEBUG_DPRINTF) { \
277fa9e4066Sahrens 	char __db_buf[32]; \
278fa9e4066Sahrens 	uint64_t __db_obj = (dn)->dn_object; \
279fa9e4066Sahrens 	if (__db_obj == DMU_META_DNODE_OBJECT) \
280fa9e4066Sahrens 		(void) strcpy(__db_buf, "mdn"); \
281fa9e4066Sahrens 	else \
282fa9e4066Sahrens 		(void) snprintf(__db_buf, sizeof (__db_buf), "%lld", \
283fa9e4066Sahrens 		    (u_longlong_t)__db_obj);\
284fa9e4066Sahrens 	dprintf_ds((dn)->dn_objset->os_dsl_dataset, "obj=%s " fmt, \
285fa9e4066Sahrens 	    __db_buf, __VA_ARGS__); \
286fa9e4066Sahrens 	} \
287fa9e4066Sahrens _NOTE(CONSTCOND) } while (0)
288fa9e4066Sahrens 
2899c9dc39aSek #define	DNODE_VERIFY(dn)		dnode_verify(dn)
2909c9dc39aSek #define	FREE_VERIFY(db, start, end, tx)	free_verify(db, start, end, tx)
2919c9dc39aSek 
292fa9e4066Sahrens #else
293fa9e4066Sahrens 
294fa9e4066Sahrens #define	dprintf_dnode(db, fmt, ...)
2959c9dc39aSek #define	DNODE_VERIFY(dn)
2969c9dc39aSek #define	FREE_VERIFY(db, start, end, tx)
297fa9e4066Sahrens 
298fa9e4066Sahrens #endif
299fa9e4066Sahrens 
300fa9e4066Sahrens #ifdef	__cplusplus
301fa9e4066Sahrens }
302fa9e4066Sahrens #endif
303fa9e4066Sahrens 
304fa9e4066Sahrens #endif	/* _SYS_DNODE_H */
305