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, 2018 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_MIN_SIZE		(1 << DNODE_SHIFT)
90 #define	DNODE_MAX_SIZE		(1 << DNODE_BLOCK_SHIFT)
91 #define	DNODE_BLOCK_SIZE	(1 << DNODE_BLOCK_SHIFT)
92 #define	DNODE_MIN_SLOTS		(DNODE_MIN_SIZE >> DNODE_SHIFT)
93 #define	DNODE_MAX_SLOTS		(DNODE_MAX_SIZE >> DNODE_SHIFT)
94 #define	DN_BONUS_SIZE(dnsize)	((dnsize) - DNODE_CORE_SIZE - \
95 	(1 << SPA_BLKPTRSHIFT))
96 #define	DN_SLOTS_TO_BONUSLEN(slots)	DN_BONUS_SIZE((slots) << DNODE_SHIFT)
97 #define	DN_OLD_MAX_BONUSLEN	(DN_BONUS_SIZE(DNODE_MIN_SIZE))
98 #define	DN_MAX_NBLKPTR	((DNODE_MIN_SIZE - DNODE_CORE_SIZE) >> SPA_BLKPTRSHIFT)
99 #define	DN_MAX_OBJECT	(1ULL << DN_MAX_OBJECT_SHIFT)
100 #define	DN_ZERO_BONUSLEN	(DN_BONUS_SIZE(DNODE_MAX_SIZE) + 1)
101 #define	DN_KILL_SPILLBLK (1)
102 
103 #define	DNODES_PER_BLOCK_SHIFT	(DNODE_BLOCK_SHIFT - DNODE_SHIFT)
104 #define	DNODES_PER_BLOCK	(1ULL << DNODES_PER_BLOCK_SHIFT)
105 
106 /*
107  * This is inaccurate if the indblkshift of the particular object is not the
108  * max.  But it's only used by userland to calculate the zvol reservation.
109  */
110 #define	DNODES_PER_LEVEL_SHIFT	(DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT)
111 #define	DNODES_PER_LEVEL	(1ULL << DNODES_PER_LEVEL_SHIFT)
112 
113 /* The +2 here is a cheesy way to round up */
114 #define	DN_MAX_LEVELS	(2 + ((DN_MAX_OFFSET_SHIFT - SPA_MINBLOCKSHIFT) / \
115 	(DN_MIN_INDBLKSHIFT - SPA_BLKPTRSHIFT)))
116 
117 #define	DN_BONUS(dnp)	((void*)((dnp)->dn_bonus + \
118 	(((dnp)->dn_nblkptr - 1) * sizeof (blkptr_t))))
119 #define	DN_MAX_BONUS_LEN(dnp) \
120 	((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) ? \
121 	(uint8_t *)DN_SPILL_BLKPTR(dnp) - (uint8_t *)DN_BONUS(dnp) : \
122 	(uint8_t *)(dnp + (dnp->dn_extra_slots + 1)) - (uint8_t *)DN_BONUS(dnp))
123 
124 #define	DN_USED_BYTES(dnp) (((dnp)->dn_flags & DNODE_FLAG_USED_BYTES) ? \
125 	(dnp)->dn_used : (dnp)->dn_used << SPA_MINBLOCKSHIFT)
126 
127 #define	EPB(blkshift, typeshift)	(1 << (blkshift - typeshift))
128 
129 struct dmu_buf_impl;
130 struct objset;
131 struct zio;
132 
133 enum dnode_dirtycontext {
134 	DN_UNDIRTIED,
135 	DN_DIRTY_OPEN,
136 	DN_DIRTY_SYNC
137 };
138 
139 /* Is dn_used in bytes?  if not, it's in multiples of SPA_MINBLOCKSIZE */
140 #define	DNODE_FLAG_USED_BYTES		(1<<0)
141 #define	DNODE_FLAG_USERUSED_ACCOUNTED	(1<<1)
142 
143 /* Does dnode have a SA spill blkptr in bonus? */
144 #define	DNODE_FLAG_SPILL_BLKPTR	(1<<2)
145 
146 typedef struct dnode_phys {
147 	uint8_t dn_type;		/* dmu_object_type_t */
148 	uint8_t dn_indblkshift;		/* ln2(indirect block size) */
149 	uint8_t dn_nlevels;		/* 1=dn_blkptr->data blocks */
150 	uint8_t dn_nblkptr;		/* length of dn_blkptr */
151 	uint8_t dn_bonustype;		/* type of data in bonus buffer */
152 	uint8_t	dn_checksum;		/* ZIO_CHECKSUM type */
153 	uint8_t	dn_compress;		/* ZIO_COMPRESS type */
154 	uint8_t dn_flags;		/* DNODE_FLAG_* */
155 	uint16_t dn_datablkszsec;	/* data block size in 512b sectors */
156 	uint16_t dn_bonuslen;		/* length of dn_bonus */
157 	uint8_t dn_extra_slots;		/* # of subsequent slots consumed */
158 	uint8_t dn_pad2[3];
159 
160 	/* accounting is protected by dn_dirty_mtx */
161 	uint64_t dn_maxblkid;		/* largest allocated block ID */
162 	uint64_t dn_used;		/* bytes (or sectors) of disk space */
163 
164 	/*
165 	 * Both dn_pad2 and dn_pad3 are protected by the block's MAC. This
166 	 * allows us to protect any fields that might be added here in the
167 	 * future. In either case, developers will want to check
168 	 * zio_crypt_init_uios_dnode() to ensure the new field is being
169 	 * protected properly.
170 	 */
171 	uint64_t dn_pad3[4];
172 	/*
173 	 * The tail region is 448 bytes for a 512 byte dnode, and
174 	 * correspondingly larger for larger dnode sizes. The spill
175 	 * block pointer, when present, is always at the end of the tail
176 	 * region. There are three ways this space may be used, using
177 	 * a 512 byte dnode for this diagram:
178 	 *
179 	 * 0       64      128     192     256     320     384     448 (offset)
180 	 * +---------------+---------------+---------------+-------+
181 	 * | dn_blkptr[0]  | dn_blkptr[1]  | dn_blkptr[2]  | /     |
182 	 * +---------------+---------------+---------------+-------+
183 	 * | dn_blkptr[0]  | dn_bonus[0..319]                      |
184 	 * +---------------+-----------------------+---------------+
185 	 * | dn_blkptr[0]  | dn_bonus[0..191]      | dn_spill      |
186 	 * +---------------+-----------------------+---------------+
187 	 */
188 	union {
189 		blkptr_t dn_blkptr[1+DN_OLD_MAX_BONUSLEN/sizeof (blkptr_t)];
190 		struct {
191 			blkptr_t __dn_ignore1;
192 			uint8_t dn_bonus[DN_OLD_MAX_BONUSLEN];
193 		};
194 		struct {
195 			blkptr_t __dn_ignore2;
196 			uint8_t __dn_ignore3[DN_OLD_MAX_BONUSLEN -
197 			    sizeof (blkptr_t)];
198 			blkptr_t dn_spill;
199  		};
200 	};
201 } dnode_phys_t;
202 
203 #define	DN_SPILL_BLKPTR(dnp)	(blkptr_t *)((char *)(dnp) + \
204 	(((dnp)->dn_extra_slots + 1) << DNODE_SHIFT) - (1 << SPA_BLKPTRSHIFT))
205 
206 struct dnode {
207 	/*
208 	 * Protects the structure of the dnode, including the number of levels
209 	 * of indirection (dn_nlevels), dn_maxblkid, and dn_next_*
210 	 */
211 	krwlock_t dn_struct_rwlock;
212 
213 	/* Our link on dn_objset->os_dnodes list; protected by os_lock.  */
214 	list_node_t dn_link;
215 
216 	/* immutable: */
217 	struct objset *dn_objset;
218 	uint64_t dn_object;
219 	struct dmu_buf_impl *dn_dbuf;
220 	struct dnode_handle *dn_handle;
221 	dnode_phys_t *dn_phys; /* pointer into dn->dn_dbuf->db.db_data */
222 
223 	/*
224 	 * Copies of stuff in dn_phys.  They're valid in the open
225 	 * context (eg. even before the dnode is first synced).
226 	 * Where necessary, these are protected by dn_struct_rwlock.
227 	 */
228 	dmu_object_type_t dn_type;	/* object type */
229 	uint16_t dn_bonuslen;		/* bonus length */
230 	uint8_t dn_bonustype;		/* bonus type */
231 	uint8_t dn_nblkptr;		/* number of blkptrs (immutable) */
232 	uint8_t dn_checksum;		/* ZIO_CHECKSUM type */
233 	uint8_t dn_compress;		/* ZIO_COMPRESS type */
234 	uint8_t dn_nlevels;
235 	uint8_t dn_indblkshift;
236 	uint8_t dn_datablkshift;	/* zero if blksz not power of 2! */
237 	uint8_t dn_moved;		/* Has this dnode been moved? */
238 	uint16_t dn_datablkszsec;	/* in 512b sectors */
239 	uint32_t dn_datablksz;		/* in bytes */
240 	uint64_t dn_maxblkid;
241 	uint8_t dn_next_type[TXG_SIZE];
242 	uint8_t dn_num_slots;		/* metadnode slots consumed on disk */
243 	uint8_t dn_next_nblkptr[TXG_SIZE];
244 	uint8_t dn_next_nlevels[TXG_SIZE];
245 	uint8_t dn_next_indblkshift[TXG_SIZE];
246 	uint8_t dn_next_bonustype[TXG_SIZE];
247 	uint8_t dn_rm_spillblk[TXG_SIZE];	/* for removing spill blk */
248 	uint16_t dn_next_bonuslen[TXG_SIZE];
249 	uint32_t dn_next_blksz[TXG_SIZE];	/* next block size in bytes */
250 
251 	/* protected by dn_dbufs_mtx; declared here to fill 32-bit hole */
252 	uint32_t dn_dbufs_count;	/* count of dn_dbufs */
253 
254 	/* protected by os_lock: */
255 	multilist_node_t dn_dirty_link[TXG_SIZE]; /* next on dataset's dirty */
256 
257 	/* protected by dn_mtx: */
258 	kmutex_t dn_mtx;
259 	list_t dn_dirty_records[TXG_SIZE];
260 	struct range_tree *dn_free_ranges[TXG_SIZE];
261 	uint64_t dn_allocated_txg;
262 	uint64_t dn_free_txg;
263 	uint64_t dn_assigned_txg;
264 	kcondvar_t dn_notxholds;
265 	enum dnode_dirtycontext dn_dirtyctx;
266 	uint8_t *dn_dirtyctx_firstset;		/* dbg: contents meaningless */
267 
268 	/* protected by own devices */
269 	refcount_t dn_tx_holds;
270 	refcount_t dn_holds;
271 
272 	kmutex_t dn_dbufs_mtx;
273 	/*
274 	 * Descendent dbufs, ordered by dbuf_compare. Note that dn_dbufs
275 	 * can contain multiple dbufs of the same (level, blkid) when a
276 	 * dbuf is marked DB_EVICTING without being removed from
277 	 * dn_dbufs. To maintain the avl invariant that there cannot be
278 	 * duplicate entries, we order the dbufs by an arbitrary value -
279 	 * their address in memory. This means that dn_dbufs cannot be used to
280 	 * directly look up a dbuf. Instead, callers must use avl_walk, have
281 	 * a reference to the dbuf, or look up a non-existant node with
282 	 * db_state = DB_SEARCH (see dbuf_free_range for an example).
283 	 */
284 	avl_tree_t dn_dbufs;
285 
286 	/* protected by dn_struct_rwlock */
287 	struct dmu_buf_impl *dn_bonus;	/* bonus buffer dbuf */
288 
289 	boolean_t dn_have_spill;	/* have spill or are spilling */
290 
291 	/* parent IO for current sync write */
292 	zio_t *dn_zio;
293 
294 	/* used in syncing context */
295 	uint64_t dn_oldused;	/* old phys used bytes */
296 	uint64_t dn_oldflags;	/* old phys dn_flags */
297 	uint64_t dn_olduid, dn_oldgid;
298 	uint64_t dn_newuid, dn_newgid;
299 	int dn_id_flags;
300 
301 	/* holds prefetch structure */
302 	struct zfetch	dn_zfetch;
303 };
304 
305 /*
306  * Adds a level of indirection between the dbuf and the dnode to avoid
307  * iterating descendent dbufs in dnode_move(). Handles are not allocated
308  * individually, but as an array of child dnodes in dnode_hold_impl().
309  */
310 typedef struct dnode_handle {
311 	/* Protects dnh_dnode from modification by dnode_move(). */
312 	zrlock_t dnh_zrlock;
313 	dnode_t *dnh_dnode;
314 } dnode_handle_t;
315 
316 typedef struct dnode_children {
317 	dmu_buf_user_t dnc_dbu;		/* User evict data */
318 	size_t dnc_count;		/* number of children */
319 	dnode_handle_t dnc_children[];	/* sized dynamically */
320 } dnode_children_t;
321 
322 typedef struct free_range {
323 	avl_node_t fr_node;
324 	uint64_t fr_blkid;
325 	uint64_t fr_nblks;
326 } free_range_t;
327 
328 void dnode_special_open(struct objset *dd, dnode_phys_t *dnp,
329     uint64_t object, dnode_handle_t *dnh);
330 void dnode_special_close(dnode_handle_t *dnh);
331 
332 void dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx);
333 void dnode_setbonus_type(dnode_t *dn, dmu_object_type_t, dmu_tx_t *tx);
334 void dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx);
335 
336 int dnode_hold(struct objset *dd, uint64_t object,
337     void *ref, dnode_t **dnp);
338 int dnode_hold_impl(struct objset *dd, uint64_t object, int flag, int dn_slots,
339     void *ref, dnode_t **dnp);
340 boolean_t dnode_add_ref(dnode_t *dn, void *ref);
341 void dnode_rele(dnode_t *dn, void *ref);
342 void dnode_rele_and_unlock(dnode_t *dn, void *tag, boolean_t evicting);
343 void dnode_setdirty(dnode_t *dn, dmu_tx_t *tx);
344 void dnode_sync(dnode_t *dn, dmu_tx_t *tx);
345 void dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
346     dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx);
347 void dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
348     dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx);
349 void dnode_free(dnode_t *dn, dmu_tx_t *tx);
350 void dnode_byteswap(dnode_phys_t *dnp);
351 void dnode_buf_byteswap(void *buf, size_t size);
352 void dnode_verify(dnode_t *dn);
353 int dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx);
354 void dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx);
355 void dnode_diduse_space(dnode_t *dn, int64_t space);
356 void dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t);
357 uint64_t dnode_block_freed(dnode_t *dn, uint64_t blkid);
358 void dnode_init(void);
359 void dnode_fini(void);
360 int dnode_next_offset(dnode_t *dn, int flags, uint64_t *off,
361     int minlvl, uint64_t blkfill, uint64_t txg);
362 void dnode_evict_dbufs(dnode_t *dn);
363 void dnode_evict_bonus(dnode_t *dn);
364 boolean_t dnode_needs_remap(const dnode_t *dn);
365 
366 #define	DNODE_IS_CACHEABLE(_dn)						\
367 	((_dn)->dn_objset->os_primary_cache == ZFS_CACHE_ALL ||		\
368 	(DMU_OT_IS_METADATA((_dn)->dn_type) &&				\
369 	(_dn)->dn_objset->os_primary_cache == ZFS_CACHE_METADATA))
370 
371 #define	DNODE_META_IS_CACHEABLE(_dn)					\
372 	((_dn)->dn_objset->os_primary_cache == ZFS_CACHE_ALL ||		\
373 	(_dn)->dn_objset->os_primary_cache == ZFS_CACHE_METADATA)
374 
375 #ifdef ZFS_DEBUG
376 
377 /*
378  * There should be a ## between the string literal and fmt, to make it
379  * clear that we're joining two strings together, but that piece of shit
380  * gcc doesn't support that preprocessor token.
381  */
382 #define	dprintf_dnode(dn, fmt, ...) do { \
383 	if (zfs_flags & ZFS_DEBUG_DPRINTF) { \
384 	char __db_buf[32]; \
385 	uint64_t __db_obj = (dn)->dn_object; \
386 	if (__db_obj == DMU_META_DNODE_OBJECT) \
387 		(void) strcpy(__db_buf, "mdn"); \
388 	else \
389 		(void) snprintf(__db_buf, sizeof (__db_buf), "%lld", \
390 		    (u_longlong_t)__db_obj);\
391 	dprintf_ds((dn)->dn_objset->os_dsl_dataset, "obj=%s " fmt, \
392 	    __db_buf, __VA_ARGS__); \
393 	} \
394 _NOTE(CONSTCOND) } while (0)
395 
396 #define	DNODE_VERIFY(dn)		dnode_verify(dn)
397 #define	FREE_VERIFY(db, start, end, tx)	free_verify(db, start, end, tx)
398 
399 #else
400 
401 #define	dprintf_dnode(db, fmt, ...)
402 #define	DNODE_VERIFY(dn)
403 #define	FREE_VERIFY(db, start, end, tx)
404 
405 #endif
406 
407 #ifdef	__cplusplus
408 }
409 #endif
410 
411 #endif	/* _SYS_DNODE_H */
412