xref: /illumos-gate/usr/src/uts/common/fs/zfs/sys/dnode.h (revision 430755dfc4bf5c4bf947290d2613a6c25d85ea0f)
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 (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
24  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
25  */
26 
27 #ifndef	_SYS_DNODE_H
28 #define	_SYS_DNODE_H
29 
30 #include <sys/zfs_context.h>
31 #include <sys/avl.h>
32 #include <sys/spa.h>
33 #include <sys/txg.h>
34 #include <sys/zio.h>
35 #include <sys/refcount.h>
36 #include <sys/dmu_zfetch.h>
37 #include <sys/zrlock.h>
38 #include <sys/multilist.h>
39 
40 #ifdef	__cplusplus
41 extern "C" {
42 #endif
43 
44 /*
45  * dnode_hold() flags.
46  */
47 #define	DNODE_MUST_BE_ALLOCATED	1
48 #define	DNODE_MUST_BE_FREE	2
49 
50 /*
51  * dnode_next_offset() flags.
52  */
53 #define	DNODE_FIND_HOLE		1
54 #define	DNODE_FIND_BACKWARDS	2
55 #define	DNODE_FIND_HAVELOCK	4
56 
57 /*
58  * Fixed constants.
59  */
60 #define	DNODE_SHIFT		9	/* 512 bytes */
61 #define	DN_MIN_INDBLKSHIFT	12	/* 4k */
62 /*
63  * If we ever increase this value beyond 20, we need to revisit all logic that
64  * does x << level * ebps to handle overflow.  With a 1M indirect block size,
65  * 4 levels of indirect blocks would not be able to guarantee addressing an
66  * entire object, so 5 levels will be used, but 5 * (20 - 7) = 65.
67  */
68 #define	DN_MAX_INDBLKSHIFT	17	/* 128k */
69 #define	DNODE_BLOCK_SHIFT	14	/* 16k */
70 #define	DNODE_CORE_SIZE		64	/* 64 bytes for dnode sans blkptrs */
71 #define	DN_MAX_OBJECT_SHIFT	48	/* 256 trillion (zfs_fid_t limit) */
72 #define	DN_MAX_OFFSET_SHIFT	64	/* 2^64 bytes in a dnode */
73 
74 /*
75  * dnode id flags
76  *
77  * Note: a file will never ever have its
78  * ids moved from bonus->spill
79  * and only in a crypto environment would it be on spill
80  */
81 #define	DN_ID_CHKED_BONUS	0x1
82 #define	DN_ID_CHKED_SPILL	0x2
83 #define	DN_ID_OLD_EXIST		0x4
84 #define	DN_ID_NEW_EXIST		0x8
85 
86 /*
87  * Derived constants.
88  */
89 #define	DNODE_SIZE	(1 << DNODE_SHIFT)
90 #define	DN_MAX_NBLKPTR	((DNODE_SIZE - DNODE_CORE_SIZE) >> SPA_BLKPTRSHIFT)
91 #define	DN_MAX_BONUSLEN	(DNODE_SIZE - DNODE_CORE_SIZE - (1 << SPA_BLKPTRSHIFT))
92 #define	DN_MAX_OBJECT	(1ULL << DN_MAX_OBJECT_SHIFT)
93 #define	DN_ZERO_BONUSLEN	(DN_MAX_BONUSLEN + 1)
94 #define	DN_KILL_SPILLBLK (1)
95 
96 #define	DNODES_PER_BLOCK_SHIFT	(DNODE_BLOCK_SHIFT - DNODE_SHIFT)
97 #define	DNODES_PER_BLOCK	(1ULL << DNODES_PER_BLOCK_SHIFT)
98 
99 /*
100  * This is inaccurate if the indblkshift of the particular object is not the
101  * max.  But it's only used by userland to calculate the zvol reservation.
102  */
103 #define	DNODES_PER_LEVEL_SHIFT	(DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT)
104 #define	DNODES_PER_LEVEL	(1ULL << DNODES_PER_LEVEL_SHIFT)
105 
106 /* The +2 here is a cheesy way to round up */
107 #define	DN_MAX_LEVELS	(2 + ((DN_MAX_OFFSET_SHIFT - SPA_MINBLOCKSHIFT) / \
108 	(DN_MIN_INDBLKSHIFT - SPA_BLKPTRSHIFT)))
109 
110 #define	DN_BONUS(dnp)	((void*)((dnp)->dn_bonus + \
111 	(((dnp)->dn_nblkptr - 1) * sizeof (blkptr_t))))
112 
113 #define	DN_USED_BYTES(dnp) (((dnp)->dn_flags & DNODE_FLAG_USED_BYTES) ? \
114 	(dnp)->dn_used : (dnp)->dn_used << SPA_MINBLOCKSHIFT)
115 
116 #define	EPB(blkshift, typeshift)	(1 << (blkshift - typeshift))
117 
118 struct dmu_buf_impl;
119 struct objset;
120 struct zio;
121 
122 enum dnode_dirtycontext {
123 	DN_UNDIRTIED,
124 	DN_DIRTY_OPEN,
125 	DN_DIRTY_SYNC
126 };
127 
128 /* Is dn_used in bytes?  if not, it's in multiples of SPA_MINBLOCKSIZE */
129 #define	DNODE_FLAG_USED_BYTES		(1<<0)
130 #define	DNODE_FLAG_USERUSED_ACCOUNTED	(1<<1)
131 
132 /* Does dnode have a SA spill blkptr in bonus? */
133 #define	DNODE_FLAG_SPILL_BLKPTR	(1<<2)
134 
135 typedef struct dnode_phys {
136 	uint8_t dn_type;		/* dmu_object_type_t */
137 	uint8_t dn_indblkshift;		/* ln2(indirect block size) */
138 	uint8_t dn_nlevels;		/* 1=dn_blkptr->data blocks */
139 	uint8_t dn_nblkptr;		/* length of dn_blkptr */
140 	uint8_t dn_bonustype;		/* type of data in bonus buffer */
141 	uint8_t	dn_checksum;		/* ZIO_CHECKSUM type */
142 	uint8_t	dn_compress;		/* ZIO_COMPRESS type */
143 	uint8_t dn_flags;		/* DNODE_FLAG_* */
144 	uint16_t dn_datablkszsec;	/* data block size in 512b sectors */
145 	uint16_t dn_bonuslen;		/* length of dn_bonus */
146 	uint8_t dn_pad2[4];
147 
148 	/* accounting is protected by dn_dirty_mtx */
149 	uint64_t dn_maxblkid;		/* largest allocated block ID */
150 	uint64_t dn_used;		/* bytes (or sectors) of disk space */
151 
152 	uint64_t dn_pad3[4];
153 
154 	blkptr_t dn_blkptr[1];
155 	uint8_t dn_bonus[DN_MAX_BONUSLEN - sizeof (blkptr_t)];
156 	blkptr_t dn_spill;
157 } dnode_phys_t;
158 
159 struct dnode {
160 	/*
161 	 * Protects the structure of the dnode, including the number of levels
162 	 * of indirection (dn_nlevels), dn_maxblkid, and dn_next_*
163 	 */
164 	krwlock_t dn_struct_rwlock;
165 
166 	/* Our link on dn_objset->os_dnodes list; protected by os_lock.  */
167 	list_node_t dn_link;
168 
169 	/* immutable: */
170 	struct objset *dn_objset;
171 	uint64_t dn_object;
172 	struct dmu_buf_impl *dn_dbuf;
173 	struct dnode_handle *dn_handle;
174 	dnode_phys_t *dn_phys; /* pointer into dn->dn_dbuf->db.db_data */
175 
176 	/*
177 	 * Copies of stuff in dn_phys.  They're valid in the open
178 	 * context (eg. even before the dnode is first synced).
179 	 * Where necessary, these are protected by dn_struct_rwlock.
180 	 */
181 	dmu_object_type_t dn_type;	/* object type */
182 	uint16_t dn_bonuslen;		/* bonus length */
183 	uint8_t dn_bonustype;		/* bonus type */
184 	uint8_t dn_nblkptr;		/* number of blkptrs (immutable) */
185 	uint8_t dn_checksum;		/* ZIO_CHECKSUM type */
186 	uint8_t dn_compress;		/* ZIO_COMPRESS type */
187 	uint8_t dn_nlevels;
188 	uint8_t dn_indblkshift;
189 	uint8_t dn_datablkshift;	/* zero if blksz not power of 2! */
190 	uint8_t dn_moved;		/* Has this dnode been moved? */
191 	uint16_t dn_datablkszsec;	/* in 512b sectors */
192 	uint32_t dn_datablksz;		/* in bytes */
193 	uint64_t dn_maxblkid;
194 	uint8_t dn_next_type[TXG_SIZE];
195 	uint8_t dn_next_nblkptr[TXG_SIZE];
196 	uint8_t dn_next_nlevels[TXG_SIZE];
197 	uint8_t dn_next_indblkshift[TXG_SIZE];
198 	uint8_t dn_next_bonustype[TXG_SIZE];
199 	uint8_t dn_rm_spillblk[TXG_SIZE];	/* for removing spill blk */
200 	uint16_t dn_next_bonuslen[TXG_SIZE];
201 	uint32_t dn_next_blksz[TXG_SIZE];	/* next block size in bytes */
202 
203 	/* protected by dn_dbufs_mtx; declared here to fill 32-bit hole */
204 	uint32_t dn_dbufs_count;	/* count of dn_dbufs */
205 
206 	/* protected by os_lock: */
207 	multilist_node_t dn_dirty_link[TXG_SIZE]; /* next on dataset's dirty */
208 
209 	/* protected by dn_mtx: */
210 	kmutex_t dn_mtx;
211 	list_t dn_dirty_records[TXG_SIZE];
212 	struct range_tree *dn_free_ranges[TXG_SIZE];
213 	uint64_t dn_allocated_txg;
214 	uint64_t dn_free_txg;
215 	uint64_t dn_assigned_txg;
216 	kcondvar_t dn_notxholds;
217 	enum dnode_dirtycontext dn_dirtyctx;
218 	uint8_t *dn_dirtyctx_firstset;		/* dbg: contents meaningless */
219 
220 	/* protected by own devices */
221 	refcount_t dn_tx_holds;
222 	refcount_t dn_holds;
223 
224 	kmutex_t dn_dbufs_mtx;
225 	/*
226 	 * Descendent dbufs, ordered by dbuf_compare. Note that dn_dbufs
227 	 * can contain multiple dbufs of the same (level, blkid) when a
228 	 * dbuf is marked DB_EVICTING without being removed from
229 	 * dn_dbufs. To maintain the avl invariant that there cannot be
230 	 * duplicate entries, we order the dbufs by an arbitrary value -
231 	 * their address in memory. This means that dn_dbufs cannot be used to
232 	 * directly look up a dbuf. Instead, callers must use avl_walk, have
233 	 * a reference to the dbuf, or look up a non-existant node with
234 	 * db_state = DB_SEARCH (see dbuf_free_range for an example).
235 	 */
236 	avl_tree_t dn_dbufs;
237 
238 	/* protected by dn_struct_rwlock */
239 	struct dmu_buf_impl *dn_bonus;	/* bonus buffer dbuf */
240 
241 	boolean_t dn_have_spill;	/* have spill or are spilling */
242 
243 	/* parent IO for current sync write */
244 	zio_t *dn_zio;
245 
246 	/* used in syncing context */
247 	uint64_t dn_oldused;	/* old phys used bytes */
248 	uint64_t dn_oldflags;	/* old phys dn_flags */
249 	uint64_t dn_olduid, dn_oldgid;
250 	uint64_t dn_newuid, dn_newgid;
251 	int dn_id_flags;
252 
253 	/* holds prefetch structure */
254 	struct zfetch	dn_zfetch;
255 };
256 
257 /*
258  * Adds a level of indirection between the dbuf and the dnode to avoid
259  * iterating descendent dbufs in dnode_move(). Handles are not allocated
260  * individually, but as an array of child dnodes in dnode_hold_impl().
261  */
262 typedef struct dnode_handle {
263 	/* Protects dnh_dnode from modification by dnode_move(). */
264 	zrlock_t dnh_zrlock;
265 	dnode_t *dnh_dnode;
266 } dnode_handle_t;
267 
268 typedef struct dnode_children {
269 	dmu_buf_user_t dnc_dbu;		/* User evict data */
270 	size_t dnc_count;		/* number of children */
271 	dnode_handle_t dnc_children[];	/* sized dynamically */
272 } dnode_children_t;
273 
274 typedef struct free_range {
275 	avl_node_t fr_node;
276 	uint64_t fr_blkid;
277 	uint64_t fr_nblks;
278 } free_range_t;
279 
280 void dnode_special_open(struct objset *dd, dnode_phys_t *dnp,
281     uint64_t object, dnode_handle_t *dnh);
282 void dnode_special_close(dnode_handle_t *dnh);
283 
284 void dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx);
285 void dnode_setbonus_type(dnode_t *dn, dmu_object_type_t, dmu_tx_t *tx);
286 void dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx);
287 
288 int dnode_hold(struct objset *dd, uint64_t object,
289     void *ref, dnode_t **dnp);
290 int dnode_hold_impl(struct objset *dd, uint64_t object, int flag,
291     void *ref, dnode_t **dnp);
292 boolean_t dnode_add_ref(dnode_t *dn, void *ref);
293 void dnode_rele(dnode_t *dn, void *ref);
294 void dnode_rele_and_unlock(dnode_t *dn, void *tag);
295 void dnode_setdirty(dnode_t *dn, dmu_tx_t *tx);
296 void dnode_sync(dnode_t *dn, dmu_tx_t *tx);
297 void dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
298     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx);
299 void dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
300     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx);
301 void dnode_free(dnode_t *dn, dmu_tx_t *tx);
302 void dnode_byteswap(dnode_phys_t *dnp);
303 void dnode_buf_byteswap(void *buf, size_t size);
304 void dnode_verify(dnode_t *dn);
305 int dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx);
306 void dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx);
307 void dnode_diduse_space(dnode_t *dn, int64_t space);
308 void dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t);
309 uint64_t dnode_block_freed(dnode_t *dn, uint64_t blkid);
310 void dnode_init(void);
311 void dnode_fini(void);
312 int dnode_next_offset(dnode_t *dn, int flags, uint64_t *off,
313     int minlvl, uint64_t blkfill, uint64_t txg);
314 void dnode_evict_dbufs(dnode_t *dn);
315 void dnode_evict_bonus(dnode_t *dn);
316 boolean_t dnode_needs_remap(const dnode_t *dn);
317 
318 #define	DNODE_IS_CACHEABLE(_dn)						\
319 	((_dn)->dn_objset->os_primary_cache == ZFS_CACHE_ALL ||		\
320 	(DMU_OT_IS_METADATA((_dn)->dn_type) &&				\
321 	(_dn)->dn_objset->os_primary_cache == ZFS_CACHE_METADATA))
322 
323 #define	DNODE_META_IS_CACHEABLE(_dn)					\
324 	((_dn)->dn_objset->os_primary_cache == ZFS_CACHE_ALL ||		\
325 	(_dn)->dn_objset->os_primary_cache == ZFS_CACHE_METADATA)
326 
327 #ifdef ZFS_DEBUG
328 
329 /*
330  * There should be a ## between the string literal and fmt, to make it
331  * clear that we're joining two strings together, but that piece of shit
332  * gcc doesn't support that preprocessor token.
333  */
334 #define	dprintf_dnode(dn, fmt, ...) do { \
335 	if (zfs_flags & ZFS_DEBUG_DPRINTF) { \
336 	char __db_buf[32]; \
337 	uint64_t __db_obj = (dn)->dn_object; \
338 	if (__db_obj == DMU_META_DNODE_OBJECT) \
339 		(void) strcpy(__db_buf, "mdn"); \
340 	else \
341 		(void) snprintf(__db_buf, sizeof (__db_buf), "%lld", \
342 		    (u_longlong_t)__db_obj);\
343 	dprintf_ds((dn)->dn_objset->os_dsl_dataset, "obj=%s " fmt, \
344 	    __db_buf, __VA_ARGS__); \
345 	} \
346 _NOTE(CONSTCOND) } while (0)
347 
348 #define	DNODE_VERIFY(dn)		dnode_verify(dn)
349 #define	FREE_VERIFY(db, start, end, tx)	free_verify(db, start, end, tx)
350 
351 #else
352 
353 #define	dprintf_dnode(db, fmt, ...)
354 #define	DNODE_VERIFY(dn)
355 #define	FREE_VERIFY(db, start, end, tx)
356 
357 #endif
358 
359 #ifdef	__cplusplus
360 }
361 #endif
362 
363 #endif	/* _SYS_DNODE_H */
364