xref: /illumos-gate/usr/src/uts/common/fs/zfs/sys/dnode.h (revision 7ae7577ce925f555e5e3410f14a3ccc896c842f2)
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 /*
152  * VARIABLE-LENGTH (LARGE) DNODES
153  *
154  * The motivation for variable-length dnodes is to eliminate the overhead
155  * associated with using spill blocks.  Spill blocks are used to store
156  * system attribute data (i.e. file metadata) that does not fit in the
157  * dnode's bonus buffer. By allowing a larger bonus buffer area the use of
158  * a spill block can be avoided.  Spill blocks potentially incur an
159  * additional read I/O for every dnode in a dnode block. As a worst case
160  * example, reading 32 dnodes from a 16k dnode block and all of the spill
161  * blocks could issue 33 separate reads. Now suppose those dnodes have size
162  * 1024 and therefore don't need spill blocks. Then the worst case number
163  * of blocks read is reduced to from 33 to two--one per dnode block.
164  *
165  * ZFS-on-Linux systems that make heavy use of extended attributes benefit
166  * from this feature. In particular, ZFS-on-Linux supports the xattr=sa
167  * dataset property which allows file extended attribute data to be stored
168  * in the dnode bonus buffer as an alternative to the traditional
169  * directory-based format. Workloads such as SELinux and the Lustre
170  * distributed filesystem often store enough xattr data to force spill
171  * blocks when xattr=sa is in effect. Large dnodes may therefore provide a
172  * performance benefit to such systems. Other use cases that benefit from
173  * this feature include files with large ACLs and symbolic links with long
174  * target names.
175  *
176  * The size of a dnode may be a multiple of 512 bytes up to the size of a
177  * dnode block (currently 16384 bytes). The dn_extra_slots field of the
178  * on-disk dnode_phys_t structure describes the size of the physical dnode
179  * on disk. The field represents how many "extra" dnode_phys_t slots a
180  * dnode consumes in its dnode block. This convention results in a value of
181  * 0 for 512 byte dnodes which preserves on-disk format compatibility with
182  * older software which doesn't support large dnodes.
183  *
184  * Similarly, the in-memory dnode_t structure has a dn_num_slots field
185  * to represent the total number of dnode_phys_t slots consumed on disk.
186  * Thus dn->dn_num_slots is 1 greater than the corresponding
187  * dnp->dn_extra_slots. This difference in convention was adopted
188  * because, unlike on-disk structures, backward compatibility is not a
189  * concern for in-memory objects, so we used a more natural way to
190  * represent size for a dnode_t.
191  *
192  * The default size for newly created dnodes is determined by the value of
193  * the "dnodesize" dataset property. By default the property is set to
194  * "legacy" which is compatible with older software. Setting the property
195  * to "auto" will allow the filesystem to choose the most suitable dnode
196  * size. Currently this just sets the default dnode size to 1k, but future
197  * code improvements could dynamically choose a size based on observed
198  * workload patterns. Dnodes of varying sizes can coexist within the same
199  * dataset and even within the same dnode block.
200  */
201 
202 #define	DNODE_CRYPT_PORTABLE_FLAGS_MASK		(DNODE_FLAG_SPILL_BLKPTR)
203 
204 typedef struct dnode_phys {
205 	uint8_t dn_type;		/* dmu_object_type_t */
206 	uint8_t dn_indblkshift;		/* ln2(indirect block size) */
207 	uint8_t dn_nlevels;		/* 1=dn_blkptr->data blocks */
208 	uint8_t dn_nblkptr;		/* length of dn_blkptr */
209 	uint8_t dn_bonustype;		/* type of data in bonus buffer */
210 	uint8_t	dn_checksum;		/* ZIO_CHECKSUM type */
211 	uint8_t	dn_compress;		/* ZIO_COMPRESS type */
212 	uint8_t dn_flags;		/* DNODE_FLAG_* */
213 	uint16_t dn_datablkszsec;	/* data block size in 512b sectors */
214 	uint16_t dn_bonuslen;		/* length of dn_bonus */
215 	uint8_t dn_extra_slots;		/* # of subsequent slots consumed */
216 	uint8_t dn_pad2[3];
217 
218 	/* accounting is protected by dn_dirty_mtx */
219 	uint64_t dn_maxblkid;		/* largest allocated block ID */
220 	uint64_t dn_used;		/* bytes (or sectors) of disk space */
221 
222 	/*
223 	 * Both dn_pad2 and dn_pad3 are protected by the block's MAC. This
224 	 * allows us to protect any fields that might be added here in the
225 	 * future. In either case, developers will want to check
226 	 * zio_crypt_init_uios_dnode() to ensure the new field is being
227 	 * protected properly.
228 	 */
229 	uint64_t dn_pad3[4];
230 	union {
231 		blkptr_t dn_blkptr[1+DN_OLD_MAX_BONUSLEN/sizeof (blkptr_t)];
232 		struct {
233 			blkptr_t __dn_ignore1;
234 			uint8_t dn_bonus[DN_OLD_MAX_BONUSLEN];
235 		};
236 		struct {
237 			blkptr_t __dn_ignore2;
238 			uint8_t __dn_ignore3[DN_OLD_MAX_BONUSLEN -
239 			    sizeof (blkptr_t)];
240 			blkptr_t dn_spill;
241 		};
242 	};
243 } dnode_phys_t;
244 
245 #define	DN_SPILL_BLKPTR(dnp)	((blkptr_t *)((char *)(dnp) +		\
246 	(((dnp)->dn_extra_slots + 1) << DNODE_SHIFT) - (1 << SPA_BLKPTRSHIFT)))
247 
248 struct dnode {
249 	/*
250 	 * Protects the structure of the dnode, including the number of levels
251 	 * of indirection (dn_nlevels), dn_maxblkid, and dn_next_*
252 	 */
253 	krwlock_t dn_struct_rwlock;
254 
255 	/* Our link on dn_objset->os_dnodes list; protected by os_lock.  */
256 	list_node_t dn_link;
257 
258 	/* immutable: */
259 	struct objset *dn_objset;
260 	uint64_t dn_object;
261 	struct dmu_buf_impl *dn_dbuf;
262 	struct dnode_handle *dn_handle;
263 	dnode_phys_t *dn_phys; /* pointer into dn->dn_dbuf->db.db_data */
264 
265 	/*
266 	 * Copies of stuff in dn_phys.  They're valid in the open
267 	 * context (eg. even before the dnode is first synced).
268 	 * Where necessary, these are protected by dn_struct_rwlock.
269 	 */
270 	dmu_object_type_t dn_type;	/* object type */
271 	uint16_t dn_bonuslen;		/* bonus length */
272 	uint8_t dn_bonustype;		/* bonus type */
273 	uint8_t dn_nblkptr;		/* number of blkptrs (immutable) */
274 	uint8_t dn_checksum;		/* ZIO_CHECKSUM type */
275 	uint8_t dn_compress;		/* ZIO_COMPRESS type */
276 	uint8_t dn_nlevels;
277 	uint8_t dn_indblkshift;
278 	uint8_t dn_datablkshift;	/* zero if blksz not power of 2! */
279 	uint8_t dn_moved;		/* Has this dnode been moved? */
280 	uint16_t dn_datablkszsec;	/* in 512b sectors */
281 	uint32_t dn_datablksz;		/* in bytes */
282 	uint64_t dn_maxblkid;
283 	uint8_t dn_next_type[TXG_SIZE];
284 	uint8_t dn_num_slots;		/* metadnode slots consumed on disk */
285 	uint8_t dn_next_nblkptr[TXG_SIZE];
286 	uint8_t dn_next_nlevels[TXG_SIZE];
287 	uint8_t dn_next_indblkshift[TXG_SIZE];
288 	uint8_t dn_next_bonustype[TXG_SIZE];
289 	uint8_t dn_rm_spillblk[TXG_SIZE];	/* for removing spill blk */
290 	uint16_t dn_next_bonuslen[TXG_SIZE];
291 	uint32_t dn_next_blksz[TXG_SIZE];	/* next block size in bytes */
292 	uint64_t dn_next_maxblkid[TXG_SIZE];	/* next maxblkid in bytes */
293 
294 	/* protected by dn_dbufs_mtx; declared here to fill 32-bit hole */
295 	uint32_t dn_dbufs_count;	/* count of dn_dbufs */
296 
297 	/* protected by os_lock: */
298 	multilist_node_t dn_dirty_link[TXG_SIZE]; /* next on dataset's dirty */
299 
300 	/* protected by dn_mtx: */
301 	kmutex_t dn_mtx;
302 	list_t dn_dirty_records[TXG_SIZE];
303 	struct range_tree *dn_free_ranges[TXG_SIZE];
304 	uint64_t dn_allocated_txg;
305 	uint64_t dn_free_txg;
306 	uint64_t dn_assigned_txg;
307 	uint64_t dn_dirty_txg;			/* txg dnode was last dirtied */
308 	kcondvar_t dn_notxholds;
309 	enum dnode_dirtycontext dn_dirtyctx;
310 	uint8_t *dn_dirtyctx_firstset;		/* dbg: contents meaningless */
311 
312 	/* protected by own devices */
313 	zfs_refcount_t dn_tx_holds;
314 	zfs_refcount_t dn_holds;
315 
316 	kmutex_t dn_dbufs_mtx;
317 	/*
318 	 * Descendent dbufs, ordered by dbuf_compare. Note that dn_dbufs
319 	 * can contain multiple dbufs of the same (level, blkid) when a
320 	 * dbuf is marked DB_EVICTING without being removed from
321 	 * dn_dbufs. To maintain the avl invariant that there cannot be
322 	 * duplicate entries, we order the dbufs by an arbitrary value -
323 	 * their address in memory. This means that dn_dbufs cannot be used to
324 	 * directly look up a dbuf. Instead, callers must use avl_walk, have
325 	 * a reference to the dbuf, or look up a non-existant node with
326 	 * db_state = DB_SEARCH (see dbuf_free_range for an example).
327 	 */
328 	avl_tree_t dn_dbufs;
329 
330 	/* protected by dn_struct_rwlock */
331 	struct dmu_buf_impl *dn_bonus;	/* bonus buffer dbuf */
332 
333 	boolean_t dn_have_spill;	/* have spill or are spilling */
334 
335 	/* parent IO for current sync write */
336 	zio_t *dn_zio;
337 
338 	/* used in syncing context */
339 	uint64_t dn_oldused;	/* old phys used bytes */
340 	uint64_t dn_oldflags;	/* old phys dn_flags */
341 	uint64_t dn_olduid, dn_oldgid;
342 	uint64_t dn_newuid, dn_newgid;
343 	int dn_id_flags;
344 
345 	/* holds prefetch structure */
346 	struct zfetch	dn_zfetch;
347 };
348 
349 /*
350  * We use this (otherwise unused) bit to indicate if the value of
351  * dn_next_maxblkid[txgoff] is valid to use in dnode_sync().
352  */
353 #define	DMU_NEXT_MAXBLKID_SET		(1ULL << 63)
354 
355 /*
356  * Adds a level of indirection between the dbuf and the dnode to avoid
357  * iterating descendent dbufs in dnode_move(). Handles are not allocated
358  * individually, but as an array of child dnodes in dnode_hold_impl().
359  */
360 typedef struct dnode_handle {
361 	/* Protects dnh_dnode from modification by dnode_move(). */
362 	zrlock_t dnh_zrlock;
363 	dnode_t *dnh_dnode;
364 } dnode_handle_t;
365 
366 typedef struct dnode_children {
367 	dmu_buf_user_t dnc_dbu;		/* User evict data */
368 	size_t dnc_count;		/* number of children */
369 	dnode_handle_t dnc_children[];	/* sized dynamically */
370 } dnode_children_t;
371 
372 typedef struct free_range {
373 	avl_node_t fr_node;
374 	uint64_t fr_blkid;
375 	uint64_t fr_nblks;
376 } free_range_t;
377 
378 void dnode_special_open(struct objset *dd, dnode_phys_t *dnp,
379     uint64_t object, dnode_handle_t *dnh);
380 void dnode_special_close(dnode_handle_t *dnh);
381 
382 void dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx);
383 void dnode_setbonus_type(dnode_t *dn, dmu_object_type_t, dmu_tx_t *tx);
384 void dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx);
385 
386 int dnode_hold(struct objset *dd, uint64_t object,
387     void *ref, dnode_t **dnp);
388 int dnode_hold_impl(struct objset *dd, uint64_t object, int flag, int dn_slots,
389     void *ref, dnode_t **dnp);
390 boolean_t dnode_add_ref(dnode_t *dn, void *ref);
391 void dnode_rele(dnode_t *dn, void *ref);
392 void dnode_rele_and_unlock(dnode_t *dn, void *tag, boolean_t evicting);
393 void dnode_setdirty(dnode_t *dn, dmu_tx_t *tx);
394 void dnode_sync(dnode_t *dn, dmu_tx_t *tx);
395 void dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
396     dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx);
397 void dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
398     dmu_object_type_t bonustype, int bonuslen, int dn_slots,
399     boolean_t keep_spill, dmu_tx_t *tx);
400 void dnode_free(dnode_t *dn, dmu_tx_t *tx);
401 void dnode_byteswap(dnode_phys_t *dnp);
402 void dnode_buf_byteswap(void *buf, size_t size);
403 void dnode_verify(dnode_t *dn);
404 int dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx);
405 int dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx);
406 void dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx);
407 void dnode_diduse_space(dnode_t *dn, int64_t space);
408 void dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx,
409     boolean_t have_read, boolean_t force);
410 uint64_t dnode_block_freed(dnode_t *dn, uint64_t blkid);
411 void dnode_init(void);
412 void dnode_fini(void);
413 int dnode_next_offset(dnode_t *dn, int flags, uint64_t *off,
414     int minlvl, uint64_t blkfill, uint64_t txg);
415 void dnode_evict_dbufs(dnode_t *dn);
416 void dnode_evict_bonus(dnode_t *dn);
417 void dnode_free_interior_slots(dnode_t *dn);
418 boolean_t dnode_needs_remap(const dnode_t *dn);
419 
420 #define	DNODE_IS_DIRTY(_dn)						\
421 	((_dn)->dn_dirty_txg >= spa_syncing_txg((_dn)->dn_objset->os_spa))
422 
423 #define	DNODE_IS_CACHEABLE(_dn)						\
424 	((_dn)->dn_objset->os_primary_cache == ZFS_CACHE_ALL ||		\
425 	(DMU_OT_IS_METADATA((_dn)->dn_type) &&				\
426 	(_dn)->dn_objset->os_primary_cache == ZFS_CACHE_METADATA))
427 
428 #define	DNODE_META_IS_CACHEABLE(_dn)					\
429 	((_dn)->dn_objset->os_primary_cache == ZFS_CACHE_ALL ||		\
430 	(_dn)->dn_objset->os_primary_cache == ZFS_CACHE_METADATA)
431 
432 /*
433  * Used for dnodestats kstat.
434  */
435 typedef struct dnode_stats {
436 	/*
437 	 * Number of failed attempts to hold a meta dnode dbuf.
438 	 */
439 	kstat_named_t dnode_hold_dbuf_hold;
440 	/*
441 	 * Number of failed attempts to read a meta dnode dbuf.
442 	 */
443 	kstat_named_t dnode_hold_dbuf_read;
444 	/*
445 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) was able
446 	 * to hold the requested object number which was allocated.  This is
447 	 * the common case when looking up any allocated object number.
448 	 */
449 	kstat_named_t dnode_hold_alloc_hits;
450 	/*
451 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) was not
452 	 * able to hold the request object number because it was not allocated.
453 	 */
454 	kstat_named_t dnode_hold_alloc_misses;
455 	/*
456 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) was not
457 	 * able to hold the request object number because the object number
458 	 * refers to an interior large dnode slot.
459 	 */
460 	kstat_named_t dnode_hold_alloc_interior;
461 	/*
462 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) needed
463 	 * to retry acquiring slot zrl locks due to contention.
464 	 */
465 	kstat_named_t dnode_hold_alloc_lock_retry;
466 	/*
467 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) did not
468 	 * need to create the dnode because another thread did so after
469 	 * dropping the read lock but before acquiring the write lock.
470 	 */
471 	kstat_named_t dnode_hold_alloc_lock_misses;
472 	/*
473 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) found
474 	 * a free dnode instantiated by dnode_create() but not yet allocated
475 	 * by dnode_allocate().
476 	 */
477 	kstat_named_t dnode_hold_alloc_type_none;
478 	/*
479 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) was able
480 	 * to hold the requested range of free dnode slots.
481 	 */
482 	kstat_named_t dnode_hold_free_hits;
483 	/*
484 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) was not
485 	 * able to hold the requested range of free dnode slots because
486 	 * at least one slot was allocated.
487 	 */
488 	kstat_named_t dnode_hold_free_misses;
489 	/*
490 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) was not
491 	 * able to hold the requested range of free dnode slots because
492 	 * after acquiring the zrl lock at least one slot was allocated.
493 	 */
494 	kstat_named_t dnode_hold_free_lock_misses;
495 	/*
496 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) needed
497 	 * to retry acquiring slot zrl locks due to contention.
498 	 */
499 	kstat_named_t dnode_hold_free_lock_retry;
500 	/*
501 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) requested
502 	 * a range of dnode slots which were held by another thread.
503 	 */
504 	kstat_named_t dnode_hold_free_refcount;
505 	/*
506 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) requested
507 	 * a range of dnode slots which would overflow the dnode_phys_t.
508 	 */
509 	kstat_named_t dnode_hold_free_overflow;
510 	/*
511 	 * Number of times a dnode_hold(...) was attempted on a dnode
512 	 * which had already been unlinked in an earlier txg.
513 	 */
514 	kstat_named_t dnode_hold_free_txg;
515 	/*
516 	 * Number of times dnode_free_interior_slots() needed to retry
517 	 * acquiring a slot zrl lock due to contention.
518 	 */
519 	kstat_named_t dnode_free_interior_lock_retry;
520 	/*
521 	 * Number of new dnodes allocated by dnode_allocate().
522 	 */
523 	kstat_named_t dnode_allocate;
524 	/*
525 	 * Number of dnodes re-allocated by dnode_reallocate().
526 	 */
527 	kstat_named_t dnode_reallocate;
528 	/*
529 	 * Number of meta dnode dbufs evicted.
530 	 */
531 	kstat_named_t dnode_buf_evict;
532 	/*
533 	 * Number of times dmu_object_alloc*() reached the end of the existing
534 	 * object ID chunk and advanced to a new one.
535 	 */
536 	kstat_named_t dnode_alloc_next_chunk;
537 	/*
538 	 * Number of times multiple threads attempted to allocate a dnode
539 	 * from the same block of free dnodes.
540 	 */
541 	kstat_named_t dnode_alloc_race;
542 	/*
543 	 * Number of times dmu_object_alloc*() was forced to advance to the
544 	 * next meta dnode dbuf due to an error from  dmu_object_next().
545 	 */
546 	kstat_named_t dnode_alloc_next_block;
547 	/*
548 	 * Statistics for tracking dnodes which have been moved.
549 	 */
550 	kstat_named_t dnode_move_invalid;
551 	kstat_named_t dnode_move_recheck1;
552 	kstat_named_t dnode_move_recheck2;
553 	kstat_named_t dnode_move_special;
554 	kstat_named_t dnode_move_handle;
555 	kstat_named_t dnode_move_rwlock;
556 	kstat_named_t dnode_move_active;
557 } dnode_stats_t;
558 
559 extern dnode_stats_t dnode_stats;
560 
561 #define	DNODE_STAT_INCR(stat, val) \
562     atomic_add_64(&dnode_stats.stat.value.ui64, (val));
563 #define	DNODE_STAT_BUMP(stat) \
564     DNODE_STAT_INCR(stat, 1);
565 
566 #ifdef ZFS_DEBUG
567 
568 /*
569  * There should be a ## between the string literal and fmt, to make it
570  * clear that we're joining two strings together, but that piece of shit
571  * gcc doesn't support that preprocessor token.
572  */
573 #define	dprintf_dnode(dn, fmt, ...) do { \
574 	if (zfs_flags & ZFS_DEBUG_DPRINTF) { \
575 	char __db_buf[32]; \
576 	uint64_t __db_obj = (dn)->dn_object; \
577 	if (__db_obj == DMU_META_DNODE_OBJECT) \
578 		(void) strcpy(__db_buf, "mdn"); \
579 	else \
580 		(void) snprintf(__db_buf, sizeof (__db_buf), "%lld", \
581 		    (u_longlong_t)__db_obj);\
582 	dprintf_ds((dn)->dn_objset->os_dsl_dataset, "obj=%s " fmt, \
583 	    __db_buf, __VA_ARGS__); \
584 	} \
585 _NOTE(CONSTCOND) } while (0)
586 
587 #define	DNODE_VERIFY(dn)		dnode_verify(dn)
588 #define	FREE_VERIFY(db, start, end, tx)	free_verify(db, start, end, tx)
589 
590 #else
591 
592 #define	dprintf_dnode(db, fmt, ...)
593 #define	DNODE_VERIFY(dn)
594 #define	FREE_VERIFY(db, start, end, tx)
595 
596 #endif
597 
598 #ifdef	__cplusplus
599 }
600 #endif
601 
602 #endif	/* _SYS_DNODE_H */
603