xref: /illumos-gate/usr/src/uts/common/fs/zfs/sys/dnode.h (revision f67950b21e185934ccabe311516f4dcbdb00ef79)
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 ids moved from bonus->spill
78  */
79 #define	DN_ID_CHKED_BONUS	0x1
80 #define	DN_ID_CHKED_SPILL	0x2
81 #define	DN_ID_OLD_EXIST		0x4
82 #define	DN_ID_NEW_EXIST		0x8
83 
84 /*
85  * Derived constants.
86  */
87 #define	DNODE_MIN_SIZE		(1 << DNODE_SHIFT)
88 #define	DNODE_MAX_SIZE		(1 << DNODE_BLOCK_SHIFT)
89 #define	DNODE_BLOCK_SIZE	(1 << DNODE_BLOCK_SHIFT)
90 #define	DNODE_MIN_SLOTS		(DNODE_MIN_SIZE >> DNODE_SHIFT)
91 #define	DNODE_MAX_SLOTS		(DNODE_MAX_SIZE >> DNODE_SHIFT)
92 #define	DN_BONUS_SIZE(dnsize)	((dnsize) - DNODE_CORE_SIZE - \
93 	(1 << SPA_BLKPTRSHIFT))
94 #define	DN_SLOTS_TO_BONUSLEN(slots)	DN_BONUS_SIZE((slots) << DNODE_SHIFT)
95 #define	DN_OLD_MAX_BONUSLEN	(DN_BONUS_SIZE(DNODE_MIN_SIZE))
96 #define	DN_MAX_NBLKPTR	((DNODE_MIN_SIZE - DNODE_CORE_SIZE) >> SPA_BLKPTRSHIFT)
97 #define	DN_MAX_OBJECT		(1ULL << DN_MAX_OBJECT_SHIFT)
98 #define	DN_ZERO_BONUSLEN	(DN_BONUS_SIZE(DNODE_MAX_SIZE) + 1)
99 #define	DN_KILL_SPILLBLK	(1)
100 
101 #define	DN_SLOT_UNINIT		((void *)NULL)	/* Uninitialized */
102 #define	DN_SLOT_FREE		((void *)1UL)	/* Free slot */
103 #define	DN_SLOT_ALLOCATED	((void *)2UL)	/* Allocated slot */
104 #define	DN_SLOT_INTERIOR	((void *)3UL)	/* Interior allocated slot */
105 #define	DN_SLOT_IS_PTR(dn)	((void *)dn > DN_SLOT_INTERIOR)
106 #define	DN_SLOT_IS_VALID(dn)	((void *)dn != NULL)
107 
108 #define	DNODES_PER_BLOCK_SHIFT	(DNODE_BLOCK_SHIFT - DNODE_SHIFT)
109 #define	DNODES_PER_BLOCK	(1ULL << DNODES_PER_BLOCK_SHIFT)
110 
111 /*
112  * This is inaccurate if the indblkshift of the particular object is not the
113  * max.  But it's only used by userland to calculate the zvol reservation.
114  */
115 #define	DNODES_PER_LEVEL_SHIFT	(DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT)
116 #define	DNODES_PER_LEVEL	(1ULL << DNODES_PER_LEVEL_SHIFT)
117 
118 /* The +2 here is a cheesy way to round up */
119 #define	DN_MAX_LEVELS	(2 + ((DN_MAX_OFFSET_SHIFT - SPA_MINBLOCKSHIFT) / \
120 	(DN_MIN_INDBLKSHIFT - SPA_BLKPTRSHIFT)))
121 
122 #define	DN_BONUS(dnp)	((void*)((dnp)->dn_bonus + \
123 	(((dnp)->dn_nblkptr - 1) * sizeof (blkptr_t))))
124 #define	DN_MAX_BONUS_LEN(dnp) \
125 	((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) ? \
126 	(uint8_t *)DN_SPILL_BLKPTR(dnp) - (uint8_t *)DN_BONUS(dnp) : \
127 	(uint8_t *)(dnp + (dnp->dn_extra_slots + 1)) - (uint8_t *)DN_BONUS(dnp))
128 
129 #define	DN_USED_BYTES(dnp) (((dnp)->dn_flags & DNODE_FLAG_USED_BYTES) ? \
130 	(dnp)->dn_used : (dnp)->dn_used << SPA_MINBLOCKSHIFT)
131 
132 #define	EPB(blkshift, typeshift)	(1 << (blkshift - typeshift))
133 
134 struct dmu_buf_impl;
135 struct objset;
136 struct zio;
137 
138 enum dnode_dirtycontext {
139 	DN_UNDIRTIED,
140 	DN_DIRTY_OPEN,
141 	DN_DIRTY_SYNC
142 };
143 
144 /* Is dn_used in bytes?  if not, it's in multiples of SPA_MINBLOCKSIZE */
145 #define	DNODE_FLAG_USED_BYTES			(1 << 0)
146 #define	DNODE_FLAG_USERUSED_ACCOUNTED		(1 << 1)
147 
148 /* Does dnode have a SA spill blkptr in bonus? */
149 #define	DNODE_FLAG_SPILL_BLKPTR			(1 << 2)
150 
151 /* User/Group/Project dnode accounting */
152 #define	DNODE_FLAG_USEROBJUSED_ACCOUNTED	(1 << 3)
153 
154 /*
155  * VARIABLE-LENGTH (LARGE) DNODES
156  *
157  * The motivation for variable-length dnodes is to eliminate the overhead
158  * associated with using spill blocks.  Spill blocks are used to store
159  * system attribute data (i.e. file metadata) that does not fit in the
160  * dnode's bonus buffer. By allowing a larger bonus buffer area the use of
161  * a spill block can be avoided.  Spill blocks potentially incur an
162  * additional read I/O for every dnode in a dnode block. As a worst case
163  * example, reading 32 dnodes from a 16k dnode block and all of the spill
164  * blocks could issue 33 separate reads. Now suppose those dnodes have size
165  * 1024 and therefore don't need spill blocks. Then the worst case number
166  * of blocks read is reduced to from 33 to two--one per dnode block.
167  *
168  * ZFS-on-Linux systems that make heavy use of extended attributes benefit
169  * from this feature. In particular, ZFS-on-Linux supports the xattr=sa
170  * dataset property which allows file extended attribute data to be stored
171  * in the dnode bonus buffer as an alternative to the traditional
172  * directory-based format. Workloads such as SELinux and the Lustre
173  * distributed filesystem often store enough xattr data to force spill
174  * blocks when xattr=sa is in effect. Large dnodes may therefore provide a
175  * performance benefit to such systems. Other use cases that benefit from
176  * this feature include files with large ACLs and symbolic links with long
177  * target names.
178  *
179  * The size of a dnode may be a multiple of 512 bytes up to the size of a
180  * dnode block (currently 16384 bytes). The dn_extra_slots field of the
181  * on-disk dnode_phys_t structure describes the size of the physical dnode
182  * on disk. The field represents how many "extra" dnode_phys_t slots a
183  * dnode consumes in its dnode block. This convention results in a value of
184  * 0 for 512 byte dnodes which preserves on-disk format compatibility with
185  * older software which doesn't support large dnodes.
186  *
187  * Similarly, the in-memory dnode_t structure has a dn_num_slots field
188  * to represent the total number of dnode_phys_t slots consumed on disk.
189  * Thus dn->dn_num_slots is 1 greater than the corresponding
190  * dnp->dn_extra_slots. This difference in convention was adopted
191  * because, unlike on-disk structures, backward compatibility is not a
192  * concern for in-memory objects, so we used a more natural way to
193  * represent size for a dnode_t.
194  *
195  * The default size for newly created dnodes is determined by the value of
196  * the "dnodesize" dataset property. By default the property is set to
197  * "legacy" which is compatible with older software. Setting the property
198  * to "auto" will allow the filesystem to choose the most suitable dnode
199  * size. Currently this just sets the default dnode size to 1k, but future
200  * code improvements could dynamically choose a size based on observed
201  * workload patterns. Dnodes of varying sizes can coexist within the same
202  * dataset and even within the same dnode block.
203  */
204 
205 #define	DNODE_CRYPT_PORTABLE_FLAGS_MASK		(DNODE_FLAG_SPILL_BLKPTR)
206 
207 typedef struct dnode_phys {
208 	uint8_t dn_type;		/* dmu_object_type_t */
209 	uint8_t dn_indblkshift;		/* ln2(indirect block size) */
210 	uint8_t dn_nlevels;		/* 1=dn_blkptr->data blocks */
211 	uint8_t dn_nblkptr;		/* length of dn_blkptr */
212 	uint8_t dn_bonustype;		/* type of data in bonus buffer */
213 	uint8_t	dn_checksum;		/* ZIO_CHECKSUM type */
214 	uint8_t	dn_compress;		/* ZIO_COMPRESS type */
215 	uint8_t dn_flags;		/* DNODE_FLAG_* */
216 	uint16_t dn_datablkszsec;	/* data block size in 512b sectors */
217 	uint16_t dn_bonuslen;		/* length of dn_bonus */
218 	uint8_t dn_extra_slots;		/* # of subsequent slots consumed */
219 	uint8_t dn_pad2[3];
220 
221 	/* accounting is protected by dn_dirty_mtx */
222 	uint64_t dn_maxblkid;		/* largest allocated block ID */
223 	uint64_t dn_used;		/* bytes (or sectors) of disk space */
224 
225 	/*
226 	 * Both dn_pad2 and dn_pad3 are protected by the block's MAC. This
227 	 * allows us to protect any fields that might be added here in the
228 	 * future. In either case, developers will want to check
229 	 * zio_crypt_init_uios_dnode() to ensure the new field is being
230 	 * protected properly.
231 	 */
232 	uint64_t dn_pad3[4];
233 	union {
234 		blkptr_t dn_blkptr[1+DN_OLD_MAX_BONUSLEN/sizeof (blkptr_t)];
235 		struct {
236 			blkptr_t __dn_ignore1;
237 			uint8_t dn_bonus[DN_OLD_MAX_BONUSLEN];
238 		};
239 		struct {
240 			blkptr_t __dn_ignore2;
241 			uint8_t __dn_ignore3[DN_OLD_MAX_BONUSLEN -
242 			    sizeof (blkptr_t)];
243 			blkptr_t dn_spill;
244 		};
245 	};
246 } dnode_phys_t;
247 
248 #define	DN_SPILL_BLKPTR(dnp)	((blkptr_t *)((char *)(dnp) +		\
249 	(((dnp)->dn_extra_slots + 1) << DNODE_SHIFT) - (1 << SPA_BLKPTRSHIFT)))
250 
251 struct dnode {
252 	/*
253 	 * Protects the structure of the dnode, including the number of levels
254 	 * of indirection (dn_nlevels), dn_maxblkid, and dn_next_*
255 	 */
256 	krwlock_t dn_struct_rwlock;
257 
258 	/* Our link on dn_objset->os_dnodes list; protected by os_lock.  */
259 	list_node_t dn_link;
260 
261 	/* immutable: */
262 	struct objset *dn_objset;
263 	uint64_t dn_object;
264 	struct dmu_buf_impl *dn_dbuf;
265 	struct dnode_handle *dn_handle;
266 	dnode_phys_t *dn_phys; /* pointer into dn->dn_dbuf->db.db_data */
267 
268 	/*
269 	 * Copies of stuff in dn_phys.  They're valid in the open
270 	 * context (eg. even before the dnode is first synced).
271 	 * Where necessary, these are protected by dn_struct_rwlock.
272 	 */
273 	dmu_object_type_t dn_type;	/* object type */
274 	uint16_t dn_bonuslen;		/* bonus length */
275 	uint8_t dn_bonustype;		/* bonus type */
276 	uint8_t dn_nblkptr;		/* number of blkptrs (immutable) */
277 	uint8_t dn_checksum;		/* ZIO_CHECKSUM type */
278 	uint8_t dn_compress;		/* ZIO_COMPRESS type */
279 	uint8_t dn_nlevels;
280 	uint8_t dn_indblkshift;
281 	uint8_t dn_datablkshift;	/* zero if blksz not power of 2! */
282 	uint8_t dn_moved;		/* Has this dnode been moved? */
283 	uint16_t dn_datablkszsec;	/* in 512b sectors */
284 	uint32_t dn_datablksz;		/* in bytes */
285 	uint64_t dn_maxblkid;
286 	uint8_t dn_next_type[TXG_SIZE];
287 	uint8_t dn_num_slots;		/* metadnode slots consumed on disk */
288 	uint8_t dn_next_nblkptr[TXG_SIZE];
289 	uint8_t dn_next_nlevels[TXG_SIZE];
290 	uint8_t dn_next_indblkshift[TXG_SIZE];
291 	uint8_t dn_next_bonustype[TXG_SIZE];
292 	uint8_t dn_rm_spillblk[TXG_SIZE];	/* for removing spill blk */
293 	uint16_t dn_next_bonuslen[TXG_SIZE];
294 	uint32_t dn_next_blksz[TXG_SIZE];	/* next block size in bytes */
295 	uint64_t dn_next_maxblkid[TXG_SIZE];	/* next maxblkid in bytes */
296 
297 	/* protected by dn_dbufs_mtx; declared here to fill 32-bit hole */
298 	uint32_t dn_dbufs_count;	/* count of dn_dbufs */
299 
300 	/* protected by os_lock: */
301 	multilist_node_t dn_dirty_link[TXG_SIZE]; /* next on dataset's dirty */
302 
303 	/* protected by dn_mtx: */
304 	kmutex_t dn_mtx;
305 	list_t dn_dirty_records[TXG_SIZE];
306 	struct range_tree *dn_free_ranges[TXG_SIZE];
307 	uint64_t dn_allocated_txg;
308 	uint64_t dn_free_txg;
309 	uint64_t dn_assigned_txg;
310 	uint64_t dn_dirty_txg;			/* txg dnode was last dirtied */
311 	kcondvar_t dn_notxholds;
312 	enum dnode_dirtycontext dn_dirtyctx;
313 	uint8_t *dn_dirtyctx_firstset;		/* dbg: contents meaningless */
314 
315 	/* protected by own devices */
316 	zfs_refcount_t dn_tx_holds;
317 	zfs_refcount_t dn_holds;
318 
319 	kmutex_t dn_dbufs_mtx;
320 	/*
321 	 * Descendent dbufs, ordered by dbuf_compare. Note that dn_dbufs
322 	 * can contain multiple dbufs of the same (level, blkid) when a
323 	 * dbuf is marked DB_EVICTING without being removed from
324 	 * dn_dbufs. To maintain the avl invariant that there cannot be
325 	 * duplicate entries, we order the dbufs by an arbitrary value -
326 	 * their address in memory. This means that dn_dbufs cannot be used to
327 	 * directly look up a dbuf. Instead, callers must use avl_walk, have
328 	 * a reference to the dbuf, or look up a non-existant node with
329 	 * db_state = DB_SEARCH (see dbuf_free_range for an example).
330 	 */
331 	avl_tree_t dn_dbufs;
332 
333 	/* protected by dn_struct_rwlock */
334 	struct dmu_buf_impl *dn_bonus;	/* bonus buffer dbuf */
335 
336 	boolean_t dn_have_spill;	/* have spill or are spilling */
337 
338 	/* parent IO for current sync write */
339 	zio_t *dn_zio;
340 
341 	/* used in syncing context */
342 	uint64_t dn_oldused;	/* old phys used bytes */
343 	uint64_t dn_oldflags;	/* old phys dn_flags */
344 	uint64_t dn_olduid, dn_oldgid, dn_oldprojid;
345 	uint64_t dn_newuid, dn_newgid, dn_newprojid;
346 	int dn_id_flags;
347 
348 	/* holds prefetch structure */
349 	struct zfetch	dn_zfetch;
350 };
351 
352 /*
353  * We use this (otherwise unused) bit to indicate if the value of
354  * dn_next_maxblkid[txgoff] is valid to use in dnode_sync().
355  */
356 #define	DMU_NEXT_MAXBLKID_SET		(1ULL << 63)
357 
358 /*
359  * Adds a level of indirection between the dbuf and the dnode to avoid
360  * iterating descendent dbufs in dnode_move(). Handles are not allocated
361  * individually, but as an array of child dnodes in dnode_hold_impl().
362  */
363 typedef struct dnode_handle {
364 	/* Protects dnh_dnode from modification by dnode_move(). */
365 	zrlock_t dnh_zrlock;
366 	dnode_t *dnh_dnode;
367 } dnode_handle_t;
368 
369 typedef struct dnode_children {
370 	dmu_buf_user_t dnc_dbu;		/* User evict data */
371 	size_t dnc_count;		/* number of children */
372 	dnode_handle_t dnc_children[];	/* sized dynamically */
373 } dnode_children_t;
374 
375 typedef struct free_range {
376 	avl_node_t fr_node;
377 	uint64_t fr_blkid;
378 	uint64_t fr_nblks;
379 } free_range_t;
380 
381 void dnode_special_open(struct objset *dd, dnode_phys_t *dnp,
382     uint64_t object, dnode_handle_t *dnh);
383 void dnode_special_close(dnode_handle_t *dnh);
384 
385 void dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx);
386 void dnode_setbonus_type(dnode_t *dn, dmu_object_type_t, dmu_tx_t *tx);
387 void dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx);
388 
389 int dnode_hold(struct objset *dd, uint64_t object,
390     void *ref, dnode_t **dnp);
391 int dnode_hold_impl(struct objset *dd, uint64_t object, int flag, int dn_slots,
392     void *ref, dnode_t **dnp);
393 boolean_t dnode_add_ref(dnode_t *dn, void *ref);
394 void dnode_rele(dnode_t *dn, void *ref);
395 void dnode_rele_and_unlock(dnode_t *dn, void *tag, boolean_t evicting);
396 void dnode_setdirty(dnode_t *dn, dmu_tx_t *tx);
397 void dnode_sync(dnode_t *dn, dmu_tx_t *tx);
398 void dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
399     dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx);
400 void dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
401     dmu_object_type_t bonustype, int bonuslen, int dn_slots,
402     boolean_t keep_spill, dmu_tx_t *tx);
403 void dnode_free(dnode_t *dn, dmu_tx_t *tx);
404 void dnode_byteswap(dnode_phys_t *dnp);
405 void dnode_buf_byteswap(void *buf, size_t size);
406 void dnode_verify(dnode_t *dn);
407 int dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx);
408 int dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx);
409 void dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx);
410 void dnode_diduse_space(dnode_t *dn, int64_t space);
411 void dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx,
412     boolean_t have_read, boolean_t force);
413 uint64_t dnode_block_freed(dnode_t *dn, uint64_t blkid);
414 void dnode_init(void);
415 void dnode_fini(void);
416 int dnode_next_offset(dnode_t *dn, int flags, uint64_t *off,
417     int minlvl, uint64_t blkfill, uint64_t txg);
418 void dnode_evict_dbufs(dnode_t *dn);
419 void dnode_evict_bonus(dnode_t *dn);
420 void dnode_free_interior_slots(dnode_t *dn);
421 boolean_t dnode_needs_remap(const dnode_t *dn);
422 
423 #define	DNODE_IS_DIRTY(_dn)						\
424 	((_dn)->dn_dirty_txg >= spa_syncing_txg((_dn)->dn_objset->os_spa))
425 
426 #define	DNODE_IS_CACHEABLE(_dn)						\
427 	((_dn)->dn_objset->os_primary_cache == ZFS_CACHE_ALL ||		\
428 	(DMU_OT_IS_METADATA((_dn)->dn_type) &&				\
429 	(_dn)->dn_objset->os_primary_cache == ZFS_CACHE_METADATA))
430 
431 #define	DNODE_META_IS_CACHEABLE(_dn)					\
432 	((_dn)->dn_objset->os_primary_cache == ZFS_CACHE_ALL ||		\
433 	(_dn)->dn_objset->os_primary_cache == ZFS_CACHE_METADATA)
434 
435 /*
436  * Used for dnodestats kstat.
437  */
438 typedef struct dnode_stats {
439 	/*
440 	 * Number of failed attempts to hold a meta dnode dbuf.
441 	 */
442 	kstat_named_t dnode_hold_dbuf_hold;
443 	/*
444 	 * Number of failed attempts to read a meta dnode dbuf.
445 	 */
446 	kstat_named_t dnode_hold_dbuf_read;
447 	/*
448 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) was able
449 	 * to hold the requested object number which was allocated.  This is
450 	 * the common case when looking up any allocated object number.
451 	 */
452 	kstat_named_t dnode_hold_alloc_hits;
453 	/*
454 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) was not
455 	 * able to hold the request object number because it was not allocated.
456 	 */
457 	kstat_named_t dnode_hold_alloc_misses;
458 	/*
459 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) was not
460 	 * able to hold the request object number because the object number
461 	 * refers to an interior large dnode slot.
462 	 */
463 	kstat_named_t dnode_hold_alloc_interior;
464 	/*
465 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) needed
466 	 * to retry acquiring slot zrl locks due to contention.
467 	 */
468 	kstat_named_t dnode_hold_alloc_lock_retry;
469 	/*
470 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) did not
471 	 * need to create the dnode because another thread did so after
472 	 * dropping the read lock but before acquiring the write lock.
473 	 */
474 	kstat_named_t dnode_hold_alloc_lock_misses;
475 	/*
476 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) found
477 	 * a free dnode instantiated by dnode_create() but not yet allocated
478 	 * by dnode_allocate().
479 	 */
480 	kstat_named_t dnode_hold_alloc_type_none;
481 	/*
482 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) was able
483 	 * to hold the requested range of free dnode slots.
484 	 */
485 	kstat_named_t dnode_hold_free_hits;
486 	/*
487 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) was not
488 	 * able to hold the requested range of free dnode slots because
489 	 * at least one slot was allocated.
490 	 */
491 	kstat_named_t dnode_hold_free_misses;
492 	/*
493 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) was not
494 	 * able to hold the requested range of free dnode slots because
495 	 * after acquiring the zrl lock at least one slot was allocated.
496 	 */
497 	kstat_named_t dnode_hold_free_lock_misses;
498 	/*
499 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) needed
500 	 * to retry acquiring slot zrl locks due to contention.
501 	 */
502 	kstat_named_t dnode_hold_free_lock_retry;
503 	/*
504 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) requested
505 	 * a range of dnode slots which were held by another thread.
506 	 */
507 	kstat_named_t dnode_hold_free_refcount;
508 	/*
509 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) requested
510 	 * a range of dnode slots which would overflow the dnode_phys_t.
511 	 */
512 	kstat_named_t dnode_hold_free_overflow;
513 	/*
514 	 * Number of times a dnode_hold(...) was attempted on a dnode
515 	 * which had already been unlinked in an earlier txg.
516 	 */
517 	kstat_named_t dnode_hold_free_txg;
518 	/*
519 	 * Number of times dnode_free_interior_slots() needed to retry
520 	 * acquiring a slot zrl lock due to contention.
521 	 */
522 	kstat_named_t dnode_free_interior_lock_retry;
523 	/*
524 	 * Number of new dnodes allocated by dnode_allocate().
525 	 */
526 	kstat_named_t dnode_allocate;
527 	/*
528 	 * Number of dnodes re-allocated by dnode_reallocate().
529 	 */
530 	kstat_named_t dnode_reallocate;
531 	/*
532 	 * Number of meta dnode dbufs evicted.
533 	 */
534 	kstat_named_t dnode_buf_evict;
535 	/*
536 	 * Number of times dmu_object_alloc*() reached the end of the existing
537 	 * object ID chunk and advanced to a new one.
538 	 */
539 	kstat_named_t dnode_alloc_next_chunk;
540 	/*
541 	 * Number of times multiple threads attempted to allocate a dnode
542 	 * from the same block of free dnodes.
543 	 */
544 	kstat_named_t dnode_alloc_race;
545 	/*
546 	 * Number of times dmu_object_alloc*() was forced to advance to the
547 	 * next meta dnode dbuf due to an error from  dmu_object_next().
548 	 */
549 	kstat_named_t dnode_alloc_next_block;
550 	/*
551 	 * Statistics for tracking dnodes which have been moved.
552 	 */
553 	kstat_named_t dnode_move_invalid;
554 	kstat_named_t dnode_move_recheck1;
555 	kstat_named_t dnode_move_recheck2;
556 	kstat_named_t dnode_move_special;
557 	kstat_named_t dnode_move_handle;
558 	kstat_named_t dnode_move_rwlock;
559 	kstat_named_t dnode_move_active;
560 } dnode_stats_t;
561 
562 extern dnode_stats_t dnode_stats;
563 
564 #define	DNODE_STAT_INCR(stat, val) \
565     atomic_add_64(&dnode_stats.stat.value.ui64, (val));
566 #define	DNODE_STAT_BUMP(stat) \
567     DNODE_STAT_INCR(stat, 1);
568 
569 #ifdef ZFS_DEBUG
570 
571 /*
572  * There should be a ## between the string literal and fmt, to make it
573  * clear that we're joining two strings together, but that piece of shit
574  * gcc doesn't support that preprocessor token.
575  */
576 #define	dprintf_dnode(dn, fmt, ...) do { \
577 	if (zfs_flags & ZFS_DEBUG_DPRINTF) { \
578 	char __db_buf[32]; \
579 	uint64_t __db_obj = (dn)->dn_object; \
580 	if (__db_obj == DMU_META_DNODE_OBJECT) \
581 		(void) strcpy(__db_buf, "mdn"); \
582 	else \
583 		(void) snprintf(__db_buf, sizeof (__db_buf), "%lld", \
584 		    (u_longlong_t)__db_obj);\
585 	dprintf_ds((dn)->dn_objset->os_dsl_dataset, "obj=%s " fmt, \
586 	    __db_buf, __VA_ARGS__); \
587 	} \
588 _NOTE(CONSTCOND) } while (0)
589 
590 #define	DNODE_VERIFY(dn)		dnode_verify(dn)
591 #define	FREE_VERIFY(db, start, end, tx)	free_verify(db, start, end, tx)
592 
593 #else
594 
595 #define	dprintf_dnode(db, fmt, ...)
596 #define	DNODE_VERIFY(dn)
597 #define	FREE_VERIFY(db, start, end, tx)
598 
599 #endif
600 
601 #ifdef	__cplusplus
602 }
603 #endif
604 
605 #endif	/* _SYS_DNODE_H */
606