xref: /illumos-gate/usr/src/uts/common/fs/zfs/sys/dnode.h (revision b390f3a9)
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 #define	DNODE_DRY_RUN		4
50 
51 /*
52  * dnode_next_offset() flags.
53  */
54 #define	DNODE_FIND_HOLE		1
55 #define	DNODE_FIND_BACKWARDS	2
56 #define	DNODE_FIND_HAVELOCK	4
57 
58 /*
59  * Fixed constants.
60  */
61 #define	DNODE_SHIFT		9	/* 512 bytes */
62 #define	DN_MIN_INDBLKSHIFT	12	/* 4k */
63 /*
64  * If we ever increase this value beyond 20, we need to revisit all logic that
65  * does x << level * ebps to handle overflow.  With a 1M indirect block size,
66  * 4 levels of indirect blocks would not be able to guarantee addressing an
67  * entire object, so 5 levels will be used, but 5 * (20 - 7) = 65.
68  */
69 #define	DN_MAX_INDBLKSHIFT	17	/* 128k */
70 #define	DNODE_BLOCK_SHIFT	14	/* 16k */
71 #define	DNODE_CORE_SIZE		64	/* 64 bytes for dnode sans blkptrs */
72 #define	DN_MAX_OBJECT_SHIFT	48	/* 256 trillion (zfs_fid_t limit) */
73 #define	DN_MAX_OFFSET_SHIFT	64	/* 2^64 bytes in a dnode */
74 
75 /*
76  * dnode id flags
77  *
78  * Note: a file will never ever have its ids moved from bonus->spill
79  */
80 #define	DN_ID_CHKED_BONUS	0x1
81 #define	DN_ID_CHKED_SPILL	0x2
82 #define	DN_ID_OLD_EXIST		0x4
83 #define	DN_ID_NEW_EXIST		0x8
84 
85 /*
86  * Derived constants.
87  */
88 #define	DNODE_MIN_SIZE		(1 << DNODE_SHIFT)
89 #define	DNODE_MAX_SIZE		(1 << DNODE_BLOCK_SHIFT)
90 #define	DNODE_BLOCK_SIZE	(1 << DNODE_BLOCK_SHIFT)
91 #define	DNODE_MIN_SLOTS		(DNODE_MIN_SIZE >> DNODE_SHIFT)
92 #define	DNODE_MAX_SLOTS		(DNODE_MAX_SIZE >> DNODE_SHIFT)
93 #define	DN_BONUS_SIZE(dnsize)	((dnsize) - DNODE_CORE_SIZE - \
94 	(1 << SPA_BLKPTRSHIFT))
95 #define	DN_SLOTS_TO_BONUSLEN(slots)	DN_BONUS_SIZE((slots) << DNODE_SHIFT)
96 #define	DN_OLD_MAX_BONUSLEN	(DN_BONUS_SIZE(DNODE_MIN_SIZE))
97 #define	DN_MAX_NBLKPTR	((DNODE_MIN_SIZE - DNODE_CORE_SIZE) >> SPA_BLKPTRSHIFT)
98 #define	DN_MAX_OBJECT		(1ULL << DN_MAX_OBJECT_SHIFT)
99 #define	DN_ZERO_BONUSLEN	(DN_BONUS_SIZE(DNODE_MAX_SIZE) + 1)
100 #define	DN_KILL_SPILLBLK	(1)
101 
102 #define	DN_SLOT_UNINIT		((void *)NULL)	/* Uninitialized */
103 #define	DN_SLOT_FREE		((void *)1UL)	/* Free slot */
104 #define	DN_SLOT_ALLOCATED	((void *)2UL)	/* Allocated slot */
105 #define	DN_SLOT_INTERIOR	((void *)3UL)	/* Interior allocated slot */
106 #define	DN_SLOT_IS_PTR(dn)	((void *)dn > DN_SLOT_INTERIOR)
107 #define	DN_SLOT_IS_VALID(dn)	((void *)dn != NULL)
108 
109 #define	DNODES_PER_BLOCK_SHIFT	(DNODE_BLOCK_SHIFT - DNODE_SHIFT)
110 #define	DNODES_PER_BLOCK	(1ULL << DNODES_PER_BLOCK_SHIFT)
111 
112 /*
113  * This is inaccurate if the indblkshift of the particular object is not the
114  * max.  But it's only used by userland to calculate the zvol reservation.
115  */
116 #define	DNODES_PER_LEVEL_SHIFT	(DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT)
117 #define	DNODES_PER_LEVEL	(1ULL << DNODES_PER_LEVEL_SHIFT)
118 
119 /* The +2 here is a cheesy way to round up */
120 #define	DN_MAX_LEVELS	(2 + ((DN_MAX_OFFSET_SHIFT - SPA_MINBLOCKSHIFT) / \
121 	(DN_MIN_INDBLKSHIFT - SPA_BLKPTRSHIFT)))
122 
123 #define	DN_BONUS(dnp)	((void*)((dnp)->dn_bonus + \
124 	(((dnp)->dn_nblkptr - 1) * sizeof (blkptr_t))))
125 #define	DN_MAX_BONUS_LEN(dnp) \
126 	((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) ? \
127 	(uint8_t *)DN_SPILL_BLKPTR(dnp) - (uint8_t *)DN_BONUS(dnp) : \
128 	(uint8_t *)(dnp + (dnp->dn_extra_slots + 1)) - (uint8_t *)DN_BONUS(dnp))
129 
130 #define	DN_USED_BYTES(dnp) (((dnp)->dn_flags & DNODE_FLAG_USED_BYTES) ? \
131 	(dnp)->dn_used : (dnp)->dn_used << SPA_MINBLOCKSHIFT)
132 
133 #define	EPB(blkshift, typeshift)	(1 << (blkshift - typeshift))
134 
135 struct dmu_buf_impl;
136 struct objset;
137 struct zio;
138 
139 enum dnode_dirtycontext {
140 	DN_UNDIRTIED,
141 	DN_DIRTY_OPEN,
142 	DN_DIRTY_SYNC
143 };
144 
145 /* Is dn_used in bytes?  if not, it's in multiples of SPA_MINBLOCKSIZE */
146 #define	DNODE_FLAG_USED_BYTES			(1 << 0)
147 #define	DNODE_FLAG_USERUSED_ACCOUNTED		(1 << 1)
148 
149 /* Does dnode have a SA spill blkptr in bonus? */
150 #define	DNODE_FLAG_SPILL_BLKPTR			(1 << 2)
151 
152 /* User/Group/Project dnode accounting */
153 #define	DNODE_FLAG_USEROBJUSED_ACCOUNTED	(1 << 3)
154 
155 /*
156  * VARIABLE-LENGTH (LARGE) DNODES
157  *
158  * The motivation for variable-length dnodes is to eliminate the overhead
159  * associated with using spill blocks.  Spill blocks are used to store
160  * system attribute data (i.e. file metadata) that does not fit in the
161  * dnode's bonus buffer. By allowing a larger bonus buffer area the use of
162  * a spill block can be avoided.  Spill blocks potentially incur an
163  * additional read I/O for every dnode in a dnode block. As a worst case
164  * example, reading 32 dnodes from a 16k dnode block and all of the spill
165  * blocks could issue 33 separate reads. Now suppose those dnodes have size
166  * 1024 and therefore don't need spill blocks. Then the worst case number
167  * of blocks read is reduced to from 33 to two--one per dnode block.
168  *
169  * ZFS-on-Linux systems that make heavy use of extended attributes benefit
170  * from this feature. In particular, ZFS-on-Linux supports the xattr=sa
171  * dataset property which allows file extended attribute data to be stored
172  * in the dnode bonus buffer as an alternative to the traditional
173  * directory-based format. Workloads such as SELinux and the Lustre
174  * distributed filesystem often store enough xattr data to force spill
175  * blocks when xattr=sa is in effect. Large dnodes may therefore provide a
176  * performance benefit to such systems. Other use cases that benefit from
177  * this feature include files with large ACLs and symbolic links with long
178  * target names.
179  *
180  * The size of a dnode may be a multiple of 512 bytes up to the size of a
181  * dnode block (currently 16384 bytes). The dn_extra_slots field of the
182  * on-disk dnode_phys_t structure describes the size of the physical dnode
183  * on disk. The field represents how many "extra" dnode_phys_t slots a
184  * dnode consumes in its dnode block. This convention results in a value of
185  * 0 for 512 byte dnodes which preserves on-disk format compatibility with
186  * older software which doesn't support large dnodes.
187  *
188  * Similarly, the in-memory dnode_t structure has a dn_num_slots field
189  * to represent the total number of dnode_phys_t slots consumed on disk.
190  * Thus dn->dn_num_slots is 1 greater than the corresponding
191  * dnp->dn_extra_slots. This difference in convention was adopted
192  * because, unlike on-disk structures, backward compatibility is not a
193  * concern for in-memory objects, so we used a more natural way to
194  * represent size for a dnode_t.
195  *
196  * The default size for newly created dnodes is determined by the value of
197  * the "dnodesize" dataset property. By default the property is set to
198  * "legacy" which is compatible with older software. Setting the property
199  * to "auto" will allow the filesystem to choose the most suitable dnode
200  * size. Currently this just sets the default dnode size to 1k, but future
201  * code improvements could dynamically choose a size based on observed
202  * workload patterns. Dnodes of varying sizes can coexist within the same
203  * dataset and even within the same dnode block.
204  */
205 
206 #define	DNODE_CRYPT_PORTABLE_FLAGS_MASK		(DNODE_FLAG_SPILL_BLKPTR)
207 
208 typedef struct dnode_phys {
209 	uint8_t dn_type;		/* dmu_object_type_t */
210 	uint8_t dn_indblkshift;		/* ln2(indirect block size) */
211 	uint8_t dn_nlevels;		/* 1=dn_blkptr->data blocks */
212 	uint8_t dn_nblkptr;		/* length of dn_blkptr */
213 	uint8_t dn_bonustype;		/* type of data in bonus buffer */
214 	uint8_t	dn_checksum;		/* ZIO_CHECKSUM type */
215 	uint8_t	dn_compress;		/* ZIO_COMPRESS type */
216 	uint8_t dn_flags;		/* DNODE_FLAG_* */
217 	uint16_t dn_datablkszsec;	/* data block size in 512b sectors */
218 	uint16_t dn_bonuslen;		/* length of dn_bonus */
219 	uint8_t dn_extra_slots;		/* # of subsequent slots consumed */
220 	uint8_t dn_pad2[3];
221 
222 	/* accounting is protected by dn_dirty_mtx */
223 	uint64_t dn_maxblkid;		/* largest allocated block ID */
224 	uint64_t dn_used;		/* bytes (or sectors) of disk space */
225 
226 	/*
227 	 * Both dn_pad2 and dn_pad3 are protected by the block's MAC. This
228 	 * allows us to protect any fields that might be added here in the
229 	 * future. In either case, developers will want to check
230 	 * zio_crypt_init_uios_dnode() to ensure the new field is being
231 	 * protected properly.
232 	 */
233 	uint64_t dn_pad3[4];
234 	union {
235 		blkptr_t dn_blkptr[1+DN_OLD_MAX_BONUSLEN/sizeof (blkptr_t)];
236 		struct {
237 			blkptr_t __dn_ignore1;
238 			uint8_t dn_bonus[DN_OLD_MAX_BONUSLEN];
239 		};
240 		struct {
241 			blkptr_t __dn_ignore2;
242 			uint8_t __dn_ignore3[DN_OLD_MAX_BONUSLEN -
243 			    sizeof (blkptr_t)];
244 			blkptr_t dn_spill;
245 		};
246 	};
247 } dnode_phys_t;
248 
249 #define	DN_SPILL_BLKPTR(dnp)	((blkptr_t *)((char *)(dnp) +		\
250 	(((dnp)->dn_extra_slots + 1) << DNODE_SHIFT) - (1 << SPA_BLKPTRSHIFT)))
251 
252 struct dnode {
253 	/*
254 	 * Protects the structure of the dnode, including the number of levels
255 	 * of indirection (dn_nlevels), dn_maxblkid, and dn_next_*
256 	 */
257 	krwlock_t dn_struct_rwlock;
258 
259 	/* Our link on dn_objset->os_dnodes list; protected by os_lock.  */
260 	list_node_t dn_link;
261 
262 	/* immutable: */
263 	struct objset *dn_objset;
264 	uint64_t dn_object;
265 	struct dmu_buf_impl *dn_dbuf;
266 	struct dnode_handle *dn_handle;
267 	dnode_phys_t *dn_phys; /* pointer into dn->dn_dbuf->db.db_data */
268 
269 	/*
270 	 * Copies of stuff in dn_phys.  They're valid in the open
271 	 * context (eg. even before the dnode is first synced).
272 	 * Where necessary, these are protected by dn_struct_rwlock.
273 	 */
274 	dmu_object_type_t dn_type;	/* object type */
275 	uint16_t dn_bonuslen;		/* bonus length */
276 	uint8_t dn_bonustype;		/* bonus type */
277 	uint8_t dn_nblkptr;		/* number of blkptrs (immutable) */
278 	uint8_t dn_checksum;		/* ZIO_CHECKSUM type */
279 	uint8_t dn_compress;		/* ZIO_COMPRESS type */
280 	uint8_t dn_nlevels;
281 	uint8_t dn_indblkshift;
282 	uint8_t dn_datablkshift;	/* zero if blksz not power of 2! */
283 	uint8_t dn_moved;		/* Has this dnode been moved? */
284 	uint16_t dn_datablkszsec;	/* in 512b sectors */
285 	uint32_t dn_datablksz;		/* in bytes */
286 	uint64_t dn_maxblkid;
287 	uint8_t dn_next_type[TXG_SIZE];
288 	uint8_t dn_num_slots;		/* metadnode slots consumed on disk */
289 	uint8_t dn_next_nblkptr[TXG_SIZE];
290 	uint8_t dn_next_nlevels[TXG_SIZE];
291 	uint8_t dn_next_indblkshift[TXG_SIZE];
292 	uint8_t dn_next_bonustype[TXG_SIZE];
293 	uint8_t dn_rm_spillblk[TXG_SIZE];	/* for removing spill blk */
294 	uint16_t dn_next_bonuslen[TXG_SIZE];
295 	uint32_t dn_next_blksz[TXG_SIZE];	/* next block size in bytes */
296 	uint64_t dn_next_maxblkid[TXG_SIZE];	/* next maxblkid in bytes */
297 
298 	/* protected by dn_dbufs_mtx; declared here to fill 32-bit hole */
299 	uint32_t dn_dbufs_count;	/* count of dn_dbufs */
300 
301 	/* protected by os_lock: */
302 	multilist_node_t dn_dirty_link[TXG_SIZE]; /* next on dataset's dirty */
303 
304 	/* protected by dn_mtx: */
305 	kmutex_t dn_mtx;
306 	list_t dn_dirty_records[TXG_SIZE];
307 	struct range_tree *dn_free_ranges[TXG_SIZE];
308 	uint64_t dn_allocated_txg;
309 	uint64_t dn_free_txg;
310 	uint64_t dn_assigned_txg;
311 	uint64_t dn_dirty_txg;			/* txg dnode was last dirtied */
312 	kcondvar_t dn_notxholds;
313 	kcondvar_t dn_nodnholds;
314 	enum dnode_dirtycontext dn_dirtyctx;
315 	uint8_t *dn_dirtyctx_firstset;		/* dbg: contents meaningless */
316 
317 	/* protected by own devices */
318 	zfs_refcount_t dn_tx_holds;
319 	zfs_refcount_t dn_holds;
320 
321 	kmutex_t dn_dbufs_mtx;
322 	/*
323 	 * Descendent dbufs, ordered by dbuf_compare. Note that dn_dbufs
324 	 * can contain multiple dbufs of the same (level, blkid) when a
325 	 * dbuf is marked DB_EVICTING without being removed from
326 	 * dn_dbufs. To maintain the avl invariant that there cannot be
327 	 * duplicate entries, we order the dbufs by an arbitrary value -
328 	 * their address in memory. This means that dn_dbufs cannot be used to
329 	 * directly look up a dbuf. Instead, callers must use avl_walk, have
330 	 * a reference to the dbuf, or look up a non-existant node with
331 	 * db_state = DB_SEARCH (see dbuf_free_range for an example).
332 	 */
333 	avl_tree_t dn_dbufs;
334 
335 	/* protected by dn_struct_rwlock */
336 	struct dmu_buf_impl *dn_bonus;	/* bonus buffer dbuf */
337 
338 	boolean_t dn_have_spill;	/* have spill or are spilling */
339 
340 	/* parent IO for current sync write */
341 	zio_t *dn_zio;
342 
343 	/* used in syncing context */
344 	uint64_t dn_oldused;	/* old phys used bytes */
345 	uint64_t dn_oldflags;	/* old phys dn_flags */
346 	uint64_t dn_olduid, dn_oldgid, dn_oldprojid;
347 	uint64_t dn_newuid, dn_newgid, dn_newprojid;
348 	int dn_id_flags;
349 
350 	/* holds prefetch structure */
351 	struct zfetch	dn_zfetch;
352 };
353 
354 /*
355  * We use this (otherwise unused) bit to indicate if the value of
356  * dn_next_maxblkid[txgoff] is valid to use in dnode_sync().
357  */
358 #define	DMU_NEXT_MAXBLKID_SET		(1ULL << 63)
359 
360 /*
361  * Adds a level of indirection between the dbuf and the dnode to avoid
362  * iterating descendent dbufs in dnode_move(). Handles are not allocated
363  * individually, but as an array of child dnodes in dnode_hold_impl().
364  */
365 typedef struct dnode_handle {
366 	/* Protects dnh_dnode from modification by dnode_move(). */
367 	zrlock_t dnh_zrlock;
368 	dnode_t *dnh_dnode;
369 } dnode_handle_t;
370 
371 typedef struct dnode_children {
372 	dmu_buf_user_t dnc_dbu;		/* User evict data */
373 	size_t dnc_count;		/* number of children */
374 	dnode_handle_t dnc_children[];	/* sized dynamically */
375 } dnode_children_t;
376 
377 typedef struct free_range {
378 	avl_node_t fr_node;
379 	uint64_t fr_blkid;
380 	uint64_t fr_nblks;
381 } free_range_t;
382 
383 void dnode_special_open(struct objset *dd, dnode_phys_t *dnp,
384     uint64_t object, dnode_handle_t *dnh);
385 void dnode_special_close(dnode_handle_t *dnh);
386 
387 void dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx);
388 void dnode_setbonus_type(dnode_t *dn, dmu_object_type_t, dmu_tx_t *tx);
389 void dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx);
390 
391 int dnode_hold(struct objset *dd, uint64_t object,
392     void *ref, dnode_t **dnp);
393 int dnode_hold_impl(struct objset *dd, uint64_t object, int flag, int dn_slots,
394     void *ref, dnode_t **dnp);
395 boolean_t dnode_add_ref(dnode_t *dn, void *ref);
396 void dnode_rele(dnode_t *dn, void *ref);
397 void dnode_rele_and_unlock(dnode_t *dn, void *tag, boolean_t evicting);
398 int dnode_try_claim(objset_t *os, uint64_t object, int slots);
399 void dnode_setdirty(dnode_t *dn, dmu_tx_t *tx);
400 void dnode_sync(dnode_t *dn, dmu_tx_t *tx);
401 void dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
402     dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx);
403 void dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
404     dmu_object_type_t bonustype, int bonuslen, int dn_slots,
405     boolean_t keep_spill, dmu_tx_t *tx);
406 void dnode_free(dnode_t *dn, dmu_tx_t *tx);
407 void dnode_byteswap(dnode_phys_t *dnp);
408 void dnode_buf_byteswap(void *buf, size_t size);
409 void dnode_verify(dnode_t *dn);
410 int dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx);
411 int dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx);
412 void dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx);
413 void dnode_diduse_space(dnode_t *dn, int64_t space);
414 void dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx,
415     boolean_t have_read, boolean_t force);
416 uint64_t dnode_block_freed(dnode_t *dn, uint64_t blkid);
417 void dnode_init(void);
418 void dnode_fini(void);
419 int dnode_next_offset(dnode_t *dn, int flags, uint64_t *off,
420     int minlvl, uint64_t blkfill, uint64_t txg);
421 void dnode_evict_dbufs(dnode_t *dn);
422 void dnode_evict_bonus(dnode_t *dn);
423 void dnode_free_interior_slots(dnode_t *dn);
424 boolean_t dnode_needs_remap(const dnode_t *dn);
425 
426 #define	DNODE_IS_DIRTY(_dn)						\
427 	((_dn)->dn_dirty_txg >= spa_syncing_txg((_dn)->dn_objset->os_spa))
428 
429 #define	DNODE_IS_CACHEABLE(_dn)						\
430 	((_dn)->dn_objset->os_primary_cache == ZFS_CACHE_ALL ||		\
431 	(DMU_OT_IS_METADATA((_dn)->dn_type) &&				\
432 	(_dn)->dn_objset->os_primary_cache == ZFS_CACHE_METADATA))
433 
434 #define	DNODE_META_IS_CACHEABLE(_dn)					\
435 	((_dn)->dn_objset->os_primary_cache == ZFS_CACHE_ALL ||		\
436 	(_dn)->dn_objset->os_primary_cache == ZFS_CACHE_METADATA)
437 
438 /*
439  * Used for dnodestats kstat.
440  */
441 typedef struct dnode_stats {
442 	/*
443 	 * Number of failed attempts to hold a meta dnode dbuf.
444 	 */
445 	kstat_named_t dnode_hold_dbuf_hold;
446 	/*
447 	 * Number of failed attempts to read a meta dnode dbuf.
448 	 */
449 	kstat_named_t dnode_hold_dbuf_read;
450 	/*
451 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) was able
452 	 * to hold the requested object number which was allocated.  This is
453 	 * the common case when looking up any allocated object number.
454 	 */
455 	kstat_named_t dnode_hold_alloc_hits;
456 	/*
457 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) was not
458 	 * able to hold the request object number because it was not allocated.
459 	 */
460 	kstat_named_t dnode_hold_alloc_misses;
461 	/*
462 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) was not
463 	 * able to hold the request object number because the object number
464 	 * refers to an interior large dnode slot.
465 	 */
466 	kstat_named_t dnode_hold_alloc_interior;
467 	/*
468 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) needed
469 	 * to retry acquiring slot zrl locks due to contention.
470 	 */
471 	kstat_named_t dnode_hold_alloc_lock_retry;
472 	/*
473 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) did not
474 	 * need to create the dnode because another thread did so after
475 	 * dropping the read lock but before acquiring the write lock.
476 	 */
477 	kstat_named_t dnode_hold_alloc_lock_misses;
478 	/*
479 	 * Number of times dnode_hold(..., DNODE_MUST_BE_ALLOCATED) found
480 	 * a free dnode instantiated by dnode_create() but not yet allocated
481 	 * by dnode_allocate().
482 	 */
483 	kstat_named_t dnode_hold_alloc_type_none;
484 	/*
485 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) was able
486 	 * to hold the requested range of free dnode slots.
487 	 */
488 	kstat_named_t dnode_hold_free_hits;
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 	 * at least one slot was allocated.
493 	 */
494 	kstat_named_t dnode_hold_free_misses;
495 	/*
496 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) was not
497 	 * able to hold the requested range of free dnode slots because
498 	 * after acquiring the zrl lock at least one slot was allocated.
499 	 */
500 	kstat_named_t dnode_hold_free_lock_misses;
501 	/*
502 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) needed
503 	 * to retry acquiring slot zrl locks due to contention.
504 	 */
505 	kstat_named_t dnode_hold_free_lock_retry;
506 	/*
507 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) requested
508 	 * a range of dnode slots which were held by another thread.
509 	 */
510 	kstat_named_t dnode_hold_free_refcount;
511 	/*
512 	 * Number of times dnode_hold(..., DNODE_MUST_BE_FREE) requested
513 	 * a range of dnode slots which would overflow the dnode_phys_t.
514 	 */
515 	kstat_named_t dnode_hold_free_overflow;
516 	/*
517 	 * Number of times dnode_free_interior_slots() needed to retry
518 	 * acquiring a slot zrl lock due to contention.
519 	 */
520 	kstat_named_t dnode_free_interior_lock_retry;
521 	/*
522 	 * Number of new dnodes allocated by dnode_allocate().
523 	 */
524 	kstat_named_t dnode_allocate;
525 	/*
526 	 * Number of dnodes re-allocated by dnode_reallocate().
527 	 */
528 	kstat_named_t dnode_reallocate;
529 	/*
530 	 * Number of meta dnode dbufs evicted.
531 	 */
532 	kstat_named_t dnode_buf_evict;
533 	/*
534 	 * Number of times dmu_object_alloc*() reached the end of the existing
535 	 * object ID chunk and advanced to a new one.
536 	 */
537 	kstat_named_t dnode_alloc_next_chunk;
538 	/*
539 	 * Number of times multiple threads attempted to allocate a dnode
540 	 * from the same block of free dnodes.
541 	 */
542 	kstat_named_t dnode_alloc_race;
543 	/*
544 	 * Number of times dmu_object_alloc*() was forced to advance to the
545 	 * next meta dnode dbuf due to an error from  dmu_object_next().
546 	 */
547 	kstat_named_t dnode_alloc_next_block;
548 	/*
549 	 * Statistics for tracking dnodes which have been moved.
550 	 */
551 	kstat_named_t dnode_move_invalid;
552 	kstat_named_t dnode_move_recheck1;
553 	kstat_named_t dnode_move_recheck2;
554 	kstat_named_t dnode_move_special;
555 	kstat_named_t dnode_move_handle;
556 	kstat_named_t dnode_move_rwlock;
557 	kstat_named_t dnode_move_active;
558 } dnode_stats_t;
559 
560 extern dnode_stats_t dnode_stats;
561 
562 #define	DNODE_STAT_INCR(stat, val) \
563     atomic_add_64(&dnode_stats.stat.value.ui64, (val));
564 #define	DNODE_STAT_BUMP(stat) \
565     DNODE_STAT_INCR(stat, 1);
566 
567 #ifdef ZFS_DEBUG
568 
569 /*
570  * There should be a ## between the string literal and fmt, to make it
571  * clear that we're joining two strings together, but that piece of shit
572  * gcc doesn't support that preprocessor token.
573  */
574 #define	dprintf_dnode(dn, fmt, ...) do { \
575 	if (zfs_flags & ZFS_DEBUG_DPRINTF) { \
576 	char __db_buf[32]; \
577 	uint64_t __db_obj = (dn)->dn_object; \
578 	if (__db_obj == DMU_META_DNODE_OBJECT) \
579 		(void) strcpy(__db_buf, "mdn"); \
580 	else \
581 		(void) snprintf(__db_buf, sizeof (__db_buf), "%lld", \
582 		    (u_longlong_t)__db_obj);\
583 	dprintf_ds((dn)->dn_objset->os_dsl_dataset, "obj=%s " fmt, \
584 	    __db_buf, __VA_ARGS__); \
585 	} \
586 _NOTE(CONSTCOND) } while (0)
587 
588 #define	DNODE_VERIFY(dn)		dnode_verify(dn)
589 #define	FREE_VERIFY(db, start, end, tx)	free_verify(db, start, end, tx)
590 
591 #else
592 
593 #define	dprintf_dnode(db, fmt, ...)
594 #define	DNODE_VERIFY(dn)
595 #define	FREE_VERIFY(db, start, end, tx)
596 
597 #endif
598 
599 #ifdef	__cplusplus
600 }
601 #endif
602 
603 #endif	/* _SYS_DNODE_H */
604