xref: /freebsd/sys/kern/vfs_cache.c (revision 6c925b9c81036a86db387f75a32b423420eadf6c)
1 /*-
2  * Copyright (c) 1989, 1993, 1995
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * Poul-Henning Kamp of the FreeBSD Project.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	@(#)vfs_cache.c	8.5 (Berkeley) 3/22/95
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include "opt_ktrace.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/counter.h>
43 #include <sys/filedesc.h>
44 #include <sys/fnv_hash.h>
45 #include <sys/kernel.h>
46 #include <sys/lock.h>
47 #include <sys/malloc.h>
48 #include <sys/fcntl.h>
49 #include <sys/mount.h>
50 #include <sys/namei.h>
51 #include <sys/proc.h>
52 #include <sys/rwlock.h>
53 #include <sys/sdt.h>
54 #include <sys/smp.h>
55 #include <sys/syscallsubr.h>
56 #include <sys/sysctl.h>
57 #include <sys/sysproto.h>
58 #include <sys/vnode.h>
59 #ifdef KTRACE
60 #include <sys/ktrace.h>
61 #endif
62 
63 #include <vm/uma.h>
64 
65 SDT_PROVIDER_DECLARE(vfs);
66 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
67     "struct vnode *");
68 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
69     "char *");
70 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
71 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
72     "char *", "struct vnode *");
73 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
74 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
75     "struct vnode *", "char *");
76 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
77     "struct vnode *");
78 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
79     "struct vnode *", "char *");
80 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
81     "char *");
82 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *");
83 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
84 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
85 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
86     "struct vnode *");
87 SDT_PROBE_DEFINE3(vfs, namecache, zap_negative, done, "struct vnode *",
88     "char *", "int");
89 SDT_PROBE_DEFINE3(vfs, namecache, shrink_negative, done, "struct vnode *",
90     "char *", "int");
91 
92 /*
93  * This structure describes the elements in the cache of recent
94  * names looked up by namei.
95  */
96 
97 struct	namecache {
98 	LIST_ENTRY(namecache) nc_hash;	/* hash chain */
99 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
100 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
101 	struct	vnode *nc_dvp;		/* vnode of parent of name */
102 	union {
103 		struct	vnode *nu_vp;	/* vnode the name refers to */
104 		u_int	nu_neghits;	/* negative entry hits */
105 	} n_un;
106 	u_char	nc_flag;		/* flag bits */
107 	u_char	nc_nlen;		/* length of name */
108 	char	nc_name[0];		/* segment name + nul */
109 };
110 
111 /*
112  * struct namecache_ts repeats struct namecache layout up to the
113  * nc_nlen member.
114  * struct namecache_ts is used in place of struct namecache when time(s) need
115  * to be stored.  The nc_dotdottime field is used when a cache entry is mapping
116  * both a non-dotdot directory name plus dotdot for the directory's
117  * parent.
118  */
119 struct	namecache_ts {
120 	LIST_ENTRY(namecache) nc_hash;	/* hash chain */
121 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
122 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
123 	struct	vnode *nc_dvp;		/* vnode of parent of name */
124 	union {
125 		struct	vnode *nu_vp;	/* vnode the name refers to */
126 		u_int	nu_neghits;	/* negative entry hits */
127 	} n_un;
128 	u_char	nc_flag;		/* flag bits */
129 	u_char	nc_nlen;		/* length of name */
130 	struct	timespec nc_time;	/* timespec provided by fs */
131 	struct	timespec nc_dotdottime;	/* dotdot timespec provided by fs */
132 	int	nc_ticks;		/* ticks value when entry was added */
133 	char	nc_name[0];		/* segment name + nul */
134 };
135 
136 #define	nc_vp		n_un.nu_vp
137 #define	nc_neghits	n_un.nu_neghits
138 
139 /*
140  * Flags in namecache.nc_flag
141  */
142 #define NCF_WHITE	0x01
143 #define NCF_ISDOTDOT	0x02
144 #define	NCF_TS		0x04
145 #define	NCF_DTS		0x08
146 #define	NCF_DVDROP	0x10
147 #define	NCF_NEGATIVE	0x20
148 #define	NCF_HOTNEGATIVE	0x40
149 
150 /*
151  * Name caching works as follows:
152  *
153  * Names found by directory scans are retained in a cache
154  * for future reference.  It is managed LRU, so frequently
155  * used names will hang around.  Cache is indexed by hash value
156  * obtained from (vp, name) where vp refers to the directory
157  * containing name.
158  *
159  * If it is a "negative" entry, (i.e. for a name that is known NOT to
160  * exist) the vnode pointer will be NULL.
161  *
162  * Upon reaching the last segment of a path, if the reference
163  * is for DELETE, or NOCACHE is set (rewrite), and the
164  * name is located in the cache, it will be dropped.
165  *
166  * These locks are used (in the order in which they can be taken):
167  * NAME		TYPE	ROLE
168  * vnodelock	mtx	vnode lists and v_cache_dd field protection
169  * bucketlock	rwlock	for access to given set of hash buckets
170  * neglist	mtx	negative entry LRU management
171  *
172  * Additionally, ncneg_shrink_lock mtx is used to have at most one thread
173  * shrinking the LRU list.
174  *
175  * It is legal to take multiple vnodelock and bucketlock locks. The locking
176  * order is lower address first. Both are recursive.
177  *
178  * "." lookups are lockless.
179  *
180  * ".." and vnode -> name lookups require vnodelock.
181  *
182  * name -> vnode lookup requires the relevant bucketlock to be held for reading.
183  *
184  * Insertions and removals of entries require involved vnodes and bucketlocks
185  * to be write-locked to prevent other threads from seeing the entry.
186  *
187  * Some lookups result in removal of the found entry (e.g. getting rid of a
188  * negative entry with the intent to create a positive one), which poses a
189  * problem when multiple threads reach the state. Similarly, two different
190  * threads can purge two different vnodes and try to remove the same name.
191  *
192  * If the already held vnode lock is lower than the second required lock, we
193  * can just take the other lock. However, in the opposite case, this could
194  * deadlock. As such, this is resolved by trylocking and if that fails unlocking
195  * the first node, locking everything in order and revalidating the state.
196  */
197 
198 /*
199  * Structures associated with name caching.
200  */
201 #define NCHHASH(hash) \
202 	(&nchashtbl[(hash) & nchash])
203 static LIST_HEAD(nchashhead, namecache) *nchashtbl;	/* Hash Table */
204 static u_long	nchash;			/* size of hash table */
205 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
206     "Size of namecache hash table");
207 static u_long	ncnegfactor = 16;	/* ratio of negative entries */
208 SYSCTL_ULONG(_vfs, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
209     "Ratio of negative namecache entries");
210 static u_long	numneg;			/* number of negative entries allocated */
211 SYSCTL_ULONG(_debug, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0,
212     "Number of negative entries in namecache");
213 static u_long	numcache;		/* number of cache entries allocated */
214 SYSCTL_ULONG(_debug, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0,
215     "Number of namecache entries");
216 static u_long	numcachehv;		/* number of cache entries with vnodes held */
217 SYSCTL_ULONG(_debug, OID_AUTO, numcachehv, CTLFLAG_RD, &numcachehv, 0,
218     "Number of namecache entries with vnodes held");
219 u_int	ncsizefactor = 2;
220 SYSCTL_UINT(_vfs, OID_AUTO, ncsizefactor, CTLFLAG_RW, &ncsizefactor, 0,
221     "Size factor for namecache");
222 static u_int	ncpurgeminvnodes;
223 SYSCTL_UINT(_vfs, OID_AUTO, ncpurgeminvnodes, CTLFLAG_RW, &ncpurgeminvnodes, 0,
224     "Number of vnodes below which purgevfs ignores the request");
225 static u_int	ncneghitsrequeue = 8;
226 SYSCTL_UINT(_vfs, OID_AUTO, ncneghitsrequeue, CTLFLAG_RW, &ncneghitsrequeue, 0,
227     "Number of hits to requeue a negative entry in the LRU list");
228 
229 struct nchstats	nchstats;		/* cache effectiveness statistics */
230 
231 static struct mtx       ncneg_shrink_lock;
232 MTX_SYSINIT(vfscache_shrink_neg, &ncneg_shrink_lock, "Name Cache shrink neg",
233     MTX_DEF);
234 
235 struct neglist {
236 	struct mtx		nl_lock;
237 	TAILQ_HEAD(, namecache) nl_list;
238 } __aligned(CACHE_LINE_SIZE);
239 
240 static struct neglist *neglists;
241 static struct neglist ncneg_hot;
242 
243 static int	shrink_list_turn;
244 
245 static u_int	numneglists;
246 static inline struct neglist *
247 NCP2NEGLIST(struct namecache *ncp)
248 {
249 
250 	return (&neglists[(((uintptr_t)(ncp) >> 8) % numneglists)]);
251 }
252 
253 static u_int   numbucketlocks;
254 static struct rwlock_padalign  *bucketlocks;
255 #define	HASH2BUCKETLOCK(hash) \
256 	((struct rwlock *)(&bucketlocks[((hash) % numbucketlocks)]))
257 
258 static u_int   numvnodelocks;
259 static struct mtx *vnodelocks;
260 static inline struct mtx *
261 VP2VNODELOCK(struct vnode *vp)
262 {
263 	struct mtx *vlp;
264 
265 	if (vp == NULL)
266 		return (NULL);
267 	vlp = &vnodelocks[(((uintptr_t)(vp) >> 8) % numvnodelocks)];
268 	return (vlp);
269 }
270 
271 /*
272  * UMA zones for the VFS cache.
273  *
274  * The small cache is used for entries with short names, which are the
275  * most common.  The large cache is used for entries which are too big to
276  * fit in the small cache.
277  */
278 static uma_zone_t cache_zone_small;
279 static uma_zone_t cache_zone_small_ts;
280 static uma_zone_t cache_zone_large;
281 static uma_zone_t cache_zone_large_ts;
282 
283 #define	CACHE_PATH_CUTOFF	35
284 
285 static struct namecache *
286 cache_alloc(int len, int ts)
287 {
288 
289 	if (len > CACHE_PATH_CUTOFF) {
290 		if (ts)
291 			return (uma_zalloc(cache_zone_large_ts, M_WAITOK));
292 		else
293 			return (uma_zalloc(cache_zone_large, M_WAITOK));
294 	}
295 	if (ts)
296 		return (uma_zalloc(cache_zone_small_ts, M_WAITOK));
297 	else
298 		return (uma_zalloc(cache_zone_small, M_WAITOK));
299 }
300 
301 static void
302 cache_free(struct namecache *ncp)
303 {
304 	int ts;
305 
306 	if (ncp == NULL)
307 		return;
308 	ts = ncp->nc_flag & NCF_TS;
309 	if ((ncp->nc_flag & NCF_DVDROP) != 0)
310 		vdrop(ncp->nc_dvp);
311 	if (ncp->nc_nlen <= CACHE_PATH_CUTOFF) {
312 		if (ts)
313 			uma_zfree(cache_zone_small_ts, ncp);
314 		else
315 			uma_zfree(cache_zone_small, ncp);
316 	} else if (ts)
317 		uma_zfree(cache_zone_large_ts, ncp);
318 	else
319 		uma_zfree(cache_zone_large, ncp);
320 }
321 
322 static char *
323 nc_get_name(struct namecache *ncp)
324 {
325 	struct namecache_ts *ncp_ts;
326 
327 	if ((ncp->nc_flag & NCF_TS) == 0)
328 		return (ncp->nc_name);
329 	ncp_ts = (struct namecache_ts *)ncp;
330 	return (ncp_ts->nc_name);
331 }
332 
333 static void
334 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
335 {
336 
337 	KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
338 	    (tsp == NULL && ticksp == NULL),
339 	    ("No NCF_TS"));
340 
341 	if (tsp != NULL)
342 		*tsp = ((struct namecache_ts *)ncp)->nc_time;
343 	if (ticksp != NULL)
344 		*ticksp = ((struct namecache_ts *)ncp)->nc_ticks;
345 }
346 
347 static int	doingcache = 1;		/* 1 => enable the cache */
348 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
349     "VFS namecache enabled");
350 
351 /* Export size information to userland */
352 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
353     sizeof(struct namecache), "sizeof(struct namecache)");
354 
355 /*
356  * The new name cache statistics
357  */
358 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0,
359     "Name cache statistics");
360 #define STATNODE_ULONG(name, descr)	\
361 	SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, descr);
362 #define STATNODE_COUNTER(name, descr)	\
363 	static counter_u64_t name;	\
364 	SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, descr);
365 STATNODE_ULONG(numneg, "Number of negative cache entries");
366 STATNODE_ULONG(numcache, "Number of cache entries");
367 STATNODE_COUNTER(numcalls, "Number of cache lookups");
368 STATNODE_COUNTER(dothits, "Number of '.' hits");
369 STATNODE_COUNTER(dotdothits, "Number of '..' hits");
370 STATNODE_COUNTER(numchecks, "Number of checks in lookup");
371 STATNODE_COUNTER(nummiss, "Number of cache misses");
372 STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache");
373 STATNODE_COUNTER(numposzaps,
374     "Number of cache hits (positive) we do not want to cache");
375 STATNODE_COUNTER(numposhits, "Number of cache hits (positive)");
376 STATNODE_COUNTER(numnegzaps,
377     "Number of cache hits (negative) we do not want to cache");
378 STATNODE_COUNTER(numneghits, "Number of cache hits (negative)");
379 /* These count for kern___getcwd(), too. */
380 STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls");
381 STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
382 STATNODE_COUNTER(numfullpathfail2,
383     "Number of fullpath search errors (VOP_VPTOCNP failures)");
384 STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
385 STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls");
386 static long zap_and_exit_bucket_fail; STATNODE_ULONG(zap_and_exit_bucket_fail,
387     "Number of times zap_and_exit failed to lock");
388 static long cache_lock_vnodes_cel_3_failures;
389 STATNODE_ULONG(cache_lock_vnodes_cel_3_failures,
390     "Number of times 3-way vnode locking failed");
391 
392 static void cache_zap_locked(struct namecache *ncp, bool neg_locked);
393 static int vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
394     char *buf, char **retbuf, u_int buflen);
395 
396 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
397 
398 static int cache_yield;
399 SYSCTL_INT(_vfs_cache, OID_AUTO, yield, CTLFLAG_RD, &cache_yield, 0,
400     "Number of times cache called yield");
401 
402 static void
403 cache_maybe_yield(void)
404 {
405 
406 	if (should_yield()) {
407 		cache_yield++;
408 		kern_yield(PRI_USER);
409 	}
410 }
411 
412 static inline void
413 cache_assert_vlp_locked(struct mtx *vlp)
414 {
415 
416 	if (vlp != NULL)
417 		mtx_assert(vlp, MA_OWNED);
418 }
419 
420 static inline void
421 cache_assert_vnode_locked(struct vnode *vp)
422 {
423 	struct mtx *vlp;
424 
425 	vlp = VP2VNODELOCK(vp);
426 	cache_assert_vlp_locked(vlp);
427 }
428 
429 static uint32_t
430 cache_get_hash(char *name, u_char len, struct vnode *dvp)
431 {
432 	uint32_t hash;
433 
434 	hash = fnv_32_buf(name, len, FNV1_32_INIT);
435 	hash = fnv_32_buf(&dvp, sizeof(dvp), hash);
436 	return (hash);
437 }
438 
439 static inline struct rwlock *
440 NCP2BUCKETLOCK(struct namecache *ncp)
441 {
442 	uint32_t hash;
443 
444 	hash = cache_get_hash(nc_get_name(ncp), ncp->nc_nlen, ncp->nc_dvp);
445 	return (HASH2BUCKETLOCK(hash));
446 }
447 
448 #ifdef INVARIANTS
449 static void
450 cache_assert_bucket_locked(struct namecache *ncp, int mode)
451 {
452 	struct rwlock *blp;
453 
454 	blp = NCP2BUCKETLOCK(ncp);
455 	rw_assert(blp, mode);
456 }
457 #else
458 #define cache_assert_bucket_locked(x, y) do { } while (0)
459 #endif
460 
461 #define cache_sort(x, y)	_cache_sort((void **)(x), (void **)(y))
462 static void
463 _cache_sort(void **p1, void **p2)
464 {
465 	void *tmp;
466 
467 	if (*p1 > *p2) {
468 		tmp = *p2;
469 		*p2 = *p1;
470 		*p1 = tmp;
471 	}
472 }
473 
474 static void
475 cache_lock_all_buckets(void)
476 {
477 	u_int i;
478 
479 	for (i = 0; i < numbucketlocks; i++)
480 		rw_wlock(&bucketlocks[i]);
481 }
482 
483 static void
484 cache_unlock_all_buckets(void)
485 {
486 	u_int i;
487 
488 	for (i = 0; i < numbucketlocks; i++)
489 		rw_wunlock(&bucketlocks[i]);
490 }
491 
492 static void
493 cache_lock_all_vnodes(void)
494 {
495 	u_int i;
496 
497 	for (i = 0; i < numvnodelocks; i++)
498 		mtx_lock(&vnodelocks[i]);
499 }
500 
501 static void
502 cache_unlock_all_vnodes(void)
503 {
504 	u_int i;
505 
506 	for (i = 0; i < numvnodelocks; i++)
507 		mtx_unlock(&vnodelocks[i]);
508 }
509 
510 static int
511 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
512 {
513 
514 	cache_sort(&vlp1, &vlp2);
515 	MPASS(vlp2 != NULL);
516 
517 	if (vlp1 != NULL) {
518 		if (!mtx_trylock(vlp1))
519 			return (EAGAIN);
520 	}
521 	if (!mtx_trylock(vlp2)) {
522 		if (vlp1 != NULL)
523 			mtx_unlock(vlp1);
524 		return (EAGAIN);
525 	}
526 
527 	return (0);
528 }
529 
530 static void
531 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
532 {
533 
534 	MPASS(vlp1 != NULL || vlp2 != NULL);
535 
536 	if (vlp1 != NULL)
537 		mtx_unlock(vlp1);
538 	if (vlp2 != NULL)
539 		mtx_unlock(vlp2);
540 }
541 
542 static int
543 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
544 {
545 	struct nchstats snap;
546 
547 	if (req->oldptr == NULL)
548 		return (SYSCTL_OUT(req, 0, sizeof(snap)));
549 
550 	snap = nchstats;
551 	snap.ncs_goodhits = counter_u64_fetch(numposhits);
552 	snap.ncs_neghits = counter_u64_fetch(numneghits);
553 	snap.ncs_badhits = counter_u64_fetch(numposzaps) +
554 	    counter_u64_fetch(numnegzaps);
555 	snap.ncs_miss = counter_u64_fetch(nummisszap) +
556 	    counter_u64_fetch(nummiss);
557 
558 	return (SYSCTL_OUT(req, &snap, sizeof(snap)));
559 }
560 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
561     CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
562     "VFS cache effectiveness statistics");
563 
564 #ifdef DIAGNOSTIC
565 /*
566  * Grab an atomic snapshot of the name cache hash chain lengths
567  */
568 static SYSCTL_NODE(_debug, OID_AUTO, hashstat, CTLFLAG_RW, NULL,
569     "hash table stats");
570 
571 static int
572 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
573 {
574 	struct nchashhead *ncpp;
575 	struct namecache *ncp;
576 	int i, error, n_nchash, *cntbuf;
577 
578 retry:
579 	n_nchash = nchash + 1;	/* nchash is max index, not count */
580 	if (req->oldptr == NULL)
581 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
582 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
583 	cache_lock_all_buckets();
584 	if (n_nchash != nchash + 1) {
585 		cache_unlock_all_buckets();
586 		free(cntbuf, M_TEMP);
587 		goto retry;
588 	}
589 	/* Scan hash tables counting entries */
590 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
591 		LIST_FOREACH(ncp, ncpp, nc_hash)
592 			cntbuf[i]++;
593 	cache_unlock_all_buckets();
594 	for (error = 0, i = 0; i < n_nchash; i++)
595 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
596 			break;
597 	free(cntbuf, M_TEMP);
598 	return (error);
599 }
600 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
601     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
602     "nchash chain lengths");
603 
604 static int
605 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
606 {
607 	int error;
608 	struct nchashhead *ncpp;
609 	struct namecache *ncp;
610 	int n_nchash;
611 	int count, maxlength, used, pct;
612 
613 	if (!req->oldptr)
614 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
615 
616 	cache_lock_all_buckets();
617 	n_nchash = nchash + 1;	/* nchash is max index, not count */
618 	used = 0;
619 	maxlength = 0;
620 
621 	/* Scan hash tables for applicable entries */
622 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
623 		count = 0;
624 		LIST_FOREACH(ncp, ncpp, nc_hash) {
625 			count++;
626 		}
627 		if (count)
628 			used++;
629 		if (maxlength < count)
630 			maxlength = count;
631 	}
632 	n_nchash = nchash + 1;
633 	cache_unlock_all_buckets();
634 	pct = (used * 100) / (n_nchash / 100);
635 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
636 	if (error)
637 		return (error);
638 	error = SYSCTL_OUT(req, &used, sizeof(used));
639 	if (error)
640 		return (error);
641 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
642 	if (error)
643 		return (error);
644 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
645 	if (error)
646 		return (error);
647 	return (0);
648 }
649 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
650     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
651     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
652 #endif
653 
654 /*
655  * Negative entries management
656  *
657  * A variation of LRU scheme is used. New entries are hashed into one of
658  * numneglists cold lists. Entries get promoted to the hot list on first hit.
659  * Partial LRU for the hot list is maintained by requeueing them every
660  * ncneghitsrequeue hits.
661  *
662  * The shrinker will demote hot list head and evict from the cold list in a
663  * round-robin manner.
664  */
665 static void
666 cache_negative_hit(struct namecache *ncp)
667 {
668 	struct neglist *neglist;
669 	u_int hits;
670 
671 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
672 	hits = atomic_fetchadd_int(&ncp->nc_neghits, 1);
673 	if (ncp->nc_flag & NCF_HOTNEGATIVE) {
674 		if ((hits % ncneghitsrequeue) != 0)
675 			return;
676 		mtx_lock(&ncneg_hot.nl_lock);
677 		if (ncp->nc_flag & NCF_HOTNEGATIVE) {
678 			TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
679 			TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
680 			mtx_unlock(&ncneg_hot.nl_lock);
681 			return;
682 		}
683 		/*
684 		 * The shrinker cleared the flag and removed the entry from
685 		 * the hot list. Put it back.
686 		 */
687 	} else {
688 		mtx_lock(&ncneg_hot.nl_lock);
689 	}
690 	neglist = NCP2NEGLIST(ncp);
691 	mtx_lock(&neglist->nl_lock);
692 	if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
693 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
694 		TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
695 		ncp->nc_flag |= NCF_HOTNEGATIVE;
696 	}
697 	mtx_unlock(&neglist->nl_lock);
698 	mtx_unlock(&ncneg_hot.nl_lock);
699 }
700 
701 static void
702 cache_negative_insert(struct namecache *ncp, bool neg_locked)
703 {
704 	struct neglist *neglist;
705 
706 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
707 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
708 	neglist = NCP2NEGLIST(ncp);
709 	if (!neg_locked) {
710 		mtx_lock(&neglist->nl_lock);
711 	} else {
712 		mtx_assert(&neglist->nl_lock, MA_OWNED);
713 	}
714 	TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
715 	if (!neg_locked)
716 		mtx_unlock(&neglist->nl_lock);
717 	atomic_add_rel_long(&numneg, 1);
718 }
719 
720 static void
721 cache_negative_remove(struct namecache *ncp, bool neg_locked)
722 {
723 	struct neglist *neglist;
724 	bool hot_locked = false;
725 	bool list_locked = false;
726 
727 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
728 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
729 	neglist = NCP2NEGLIST(ncp);
730 	if (!neg_locked) {
731 		if (ncp->nc_flag & NCF_HOTNEGATIVE) {
732 			hot_locked = true;
733 			mtx_lock(&ncneg_hot.nl_lock);
734 			if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
735 				list_locked = true;
736 				mtx_lock(&neglist->nl_lock);
737 			}
738 		} else {
739 			list_locked = true;
740 			mtx_lock(&neglist->nl_lock);
741 		}
742 	} else {
743 		mtx_assert(&neglist->nl_lock, MA_OWNED);
744 		mtx_assert(&ncneg_hot.nl_lock, MA_OWNED);
745 	}
746 	if (ncp->nc_flag & NCF_HOTNEGATIVE) {
747 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
748 	} else {
749 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
750 	}
751 	if (list_locked)
752 		mtx_unlock(&neglist->nl_lock);
753 	if (hot_locked)
754 		mtx_unlock(&ncneg_hot.nl_lock);
755 	atomic_subtract_rel_long(&numneg, 1);
756 }
757 
758 static void
759 cache_negative_shrink_select(int start, struct namecache **ncpp,
760     struct neglist **neglistpp)
761 {
762 	struct neglist *neglist;
763 	struct namecache *ncp;
764 	int i;
765 
766 	*ncpp = ncp = NULL;
767 
768 	for (i = start; i < numneglists; i++) {
769 		neglist = &neglists[i];
770 		if (TAILQ_FIRST(&neglist->nl_list) == NULL)
771 			continue;
772 		mtx_lock(&neglist->nl_lock);
773 		ncp = TAILQ_FIRST(&neglist->nl_list);
774 		if (ncp != NULL)
775 			break;
776 		mtx_unlock(&neglist->nl_lock);
777 	}
778 
779 	*neglistpp = neglist;
780 	*ncpp = ncp;
781 }
782 
783 static void
784 cache_negative_zap_one(void)
785 {
786 	struct namecache *ncp, *ncp2;
787 	struct neglist *neglist;
788 	struct mtx *dvlp;
789 	struct rwlock *blp;
790 
791 	if (!mtx_trylock(&ncneg_shrink_lock))
792 		return;
793 
794 	mtx_lock(&ncneg_hot.nl_lock);
795 	ncp = TAILQ_FIRST(&ncneg_hot.nl_list);
796 	if (ncp != NULL) {
797 		neglist = NCP2NEGLIST(ncp);
798 		mtx_lock(&neglist->nl_lock);
799 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
800 		TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
801 		ncp->nc_flag &= ~NCF_HOTNEGATIVE;
802 		mtx_unlock(&neglist->nl_lock);
803 	}
804 
805 	cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
806 	shrink_list_turn++;
807 	if (shrink_list_turn == numneglists)
808 		shrink_list_turn = 0;
809 	if (ncp == NULL && shrink_list_turn == 0)
810 		cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
811 	if (ncp == NULL) {
812 		mtx_unlock(&ncneg_hot.nl_lock);
813 		goto out;
814 	}
815 
816 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
817 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
818 	blp = NCP2BUCKETLOCK(ncp);
819 	mtx_unlock(&neglist->nl_lock);
820 	mtx_unlock(&ncneg_hot.nl_lock);
821 	mtx_lock(dvlp);
822 	rw_wlock(blp);
823 	mtx_lock(&ncneg_hot.nl_lock);
824 	mtx_lock(&neglist->nl_lock);
825 	ncp2 = TAILQ_FIRST(&neglist->nl_list);
826 	if (ncp != ncp2 || dvlp != VP2VNODELOCK(ncp2->nc_dvp) ||
827 	    blp != NCP2BUCKETLOCK(ncp2) || !(ncp2->nc_flag & NCF_NEGATIVE)) {
828 		ncp = NULL;
829 		goto out_unlock_all;
830 	}
831 	SDT_PROBE3(vfs, namecache, shrink_negative, done, ncp->nc_dvp,
832 	    nc_get_name(ncp), ncp->nc_neghits);
833 
834 	cache_zap_locked(ncp, true);
835 out_unlock_all:
836 	mtx_unlock(&neglist->nl_lock);
837 	mtx_unlock(&ncneg_hot.nl_lock);
838 	rw_wunlock(blp);
839 	mtx_unlock(dvlp);
840 out:
841 	mtx_unlock(&ncneg_shrink_lock);
842 	cache_free(ncp);
843 }
844 
845 /*
846  * cache_zap_locked():
847  *
848  *   Removes a namecache entry from cache, whether it contains an actual
849  *   pointer to a vnode or if it is just a negative cache entry.
850  */
851 static void
852 cache_zap_locked(struct namecache *ncp, bool neg_locked)
853 {
854 
855 	if (!(ncp->nc_flag & NCF_NEGATIVE))
856 		cache_assert_vnode_locked(ncp->nc_vp);
857 	cache_assert_vnode_locked(ncp->nc_dvp);
858 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
859 
860 	CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp,
861 	    (ncp->nc_flag & NCF_NEGATIVE) ? NULL : ncp->nc_vp);
862 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
863 		SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
864 		    nc_get_name(ncp), ncp->nc_vp);
865 	} else {
866 		SDT_PROBE3(vfs, namecache, zap_negative, done, ncp->nc_dvp,
867 		    nc_get_name(ncp), ncp->nc_neghits);
868 	}
869 	LIST_REMOVE(ncp, nc_hash);
870 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
871 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
872 		if (ncp == ncp->nc_vp->v_cache_dd)
873 			ncp->nc_vp->v_cache_dd = NULL;
874 	} else {
875 		cache_negative_remove(ncp, neg_locked);
876 	}
877 	if (ncp->nc_flag & NCF_ISDOTDOT) {
878 		if (ncp == ncp->nc_dvp->v_cache_dd)
879 			ncp->nc_dvp->v_cache_dd = NULL;
880 	} else {
881 		LIST_REMOVE(ncp, nc_src);
882 		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
883 			ncp->nc_flag |= NCF_DVDROP;
884 			atomic_subtract_rel_long(&numcachehv, 1);
885 		}
886 	}
887 	atomic_subtract_rel_long(&numcache, 1);
888 }
889 
890 static void
891 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
892 {
893 	struct rwlock *blp;
894 
895 	MPASS(ncp->nc_dvp == vp);
896 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
897 	cache_assert_vnode_locked(vp);
898 
899 	blp = NCP2BUCKETLOCK(ncp);
900 	rw_wlock(blp);
901 	cache_zap_locked(ncp, false);
902 	rw_wunlock(blp);
903 }
904 
905 static bool
906 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
907     struct mtx **vlpp)
908 {
909 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
910 	struct rwlock *blp;
911 
912 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
913 	cache_assert_vnode_locked(vp);
914 
915 	if (ncp->nc_flag & NCF_NEGATIVE) {
916 		if (*vlpp != NULL) {
917 			mtx_unlock(*vlpp);
918 			*vlpp = NULL;
919 		}
920 		cache_zap_negative_locked_vnode_kl(ncp, vp);
921 		return (true);
922 	}
923 
924 	pvlp = VP2VNODELOCK(vp);
925 	blp = NCP2BUCKETLOCK(ncp);
926 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
927 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
928 
929 	if (*vlpp == vlp1 || *vlpp == vlp2) {
930 		to_unlock = *vlpp;
931 		*vlpp = NULL;
932 	} else {
933 		if (*vlpp != NULL) {
934 			mtx_unlock(*vlpp);
935 			*vlpp = NULL;
936 		}
937 		cache_sort(&vlp1, &vlp2);
938 		if (vlp1 == pvlp) {
939 			mtx_lock(vlp2);
940 			to_unlock = vlp2;
941 		} else {
942 			if (!mtx_trylock(vlp1))
943 				goto out_relock;
944 			to_unlock = vlp1;
945 		}
946 	}
947 	rw_wlock(blp);
948 	cache_zap_locked(ncp, false);
949 	rw_wunlock(blp);
950 	if (to_unlock != NULL)
951 		mtx_unlock(to_unlock);
952 	return (true);
953 
954 out_relock:
955 	mtx_unlock(vlp2);
956 	mtx_lock(vlp1);
957 	mtx_lock(vlp2);
958 	MPASS(*vlpp == NULL);
959 	*vlpp = vlp1;
960 	return (false);
961 }
962 
963 static int
964 cache_zap_locked_vnode(struct namecache *ncp, struct vnode *vp)
965 {
966 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
967 	struct rwlock *blp;
968 	int error = 0;
969 
970 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
971 	cache_assert_vnode_locked(vp);
972 
973 	pvlp = VP2VNODELOCK(vp);
974 	if (ncp->nc_flag & NCF_NEGATIVE) {
975 		cache_zap_negative_locked_vnode_kl(ncp, vp);
976 		goto out;
977 	}
978 
979 	blp = NCP2BUCKETLOCK(ncp);
980 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
981 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
982 	cache_sort(&vlp1, &vlp2);
983 	if (vlp1 == pvlp) {
984 		mtx_lock(vlp2);
985 		to_unlock = vlp2;
986 	} else {
987 		if (!mtx_trylock(vlp1)) {
988 			error = EAGAIN;
989 			goto out;
990 		}
991 		to_unlock = vlp1;
992 	}
993 	rw_wlock(blp);
994 	cache_zap_locked(ncp, false);
995 	rw_wunlock(blp);
996 	mtx_unlock(to_unlock);
997 out:
998 	mtx_unlock(pvlp);
999 	return (error);
1000 }
1001 
1002 static int
1003 cache_zap_rlocked_bucket(struct namecache *ncp, struct rwlock *blp)
1004 {
1005 	struct mtx *dvlp, *vlp;
1006 
1007 	cache_assert_bucket_locked(ncp, RA_RLOCKED);
1008 
1009 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1010 	vlp = NULL;
1011 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1012 		vlp = VP2VNODELOCK(ncp->nc_vp);
1013 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1014 		rw_runlock(blp);
1015 		rw_wlock(blp);
1016 		cache_zap_locked(ncp, false);
1017 		rw_wunlock(blp);
1018 		cache_unlock_vnodes(dvlp, vlp);
1019 		return (0);
1020 	}
1021 
1022 	rw_runlock(blp);
1023 	return (EAGAIN);
1024 }
1025 
1026 static int
1027 cache_zap_wlocked_bucket_kl(struct namecache *ncp, struct rwlock *blp,
1028     struct mtx **vlpp1, struct mtx **vlpp2)
1029 {
1030 	struct mtx *dvlp, *vlp;
1031 
1032 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1033 
1034 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1035 	vlp = NULL;
1036 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1037 		vlp = VP2VNODELOCK(ncp->nc_vp);
1038 	cache_sort(&dvlp, &vlp);
1039 
1040 	if (*vlpp1 == dvlp && *vlpp2 == vlp) {
1041 		cache_zap_locked(ncp, false);
1042 		cache_unlock_vnodes(dvlp, vlp);
1043 		*vlpp1 = NULL;
1044 		*vlpp2 = NULL;
1045 		return (0);
1046 	}
1047 
1048 	if (*vlpp1 != NULL)
1049 		mtx_unlock(*vlpp1);
1050 	if (*vlpp2 != NULL)
1051 		mtx_unlock(*vlpp2);
1052 	*vlpp1 = NULL;
1053 	*vlpp2 = NULL;
1054 
1055 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1056 		cache_zap_locked(ncp, false);
1057 		cache_unlock_vnodes(dvlp, vlp);
1058 		return (0);
1059 	}
1060 
1061 	rw_wunlock(blp);
1062 	*vlpp1 = dvlp;
1063 	*vlpp2 = vlp;
1064 	if (*vlpp1 != NULL)
1065 		mtx_lock(*vlpp1);
1066 	mtx_lock(*vlpp2);
1067 	rw_wlock(blp);
1068 	return (EAGAIN);
1069 }
1070 
1071 static void
1072 cache_lookup_unlock(struct rwlock *blp, struct mtx *vlp)
1073 {
1074 
1075 	if (blp != NULL) {
1076 		rw_runlock(blp);
1077 		mtx_assert(vlp, MA_NOTOWNED);
1078 	} else {
1079 		mtx_unlock(vlp);
1080 	}
1081 }
1082 
1083 /*
1084  * Lookup an entry in the cache
1085  *
1086  * Lookup is called with dvp pointing to the directory to search,
1087  * cnp pointing to the name of the entry being sought. If the lookup
1088  * succeeds, the vnode is returned in *vpp, and a status of -1 is
1089  * returned. If the lookup determines that the name does not exist
1090  * (negative caching), a status of ENOENT is returned. If the lookup
1091  * fails, a status of zero is returned.  If the directory vnode is
1092  * recycled out from under us due to a forced unmount, a status of
1093  * ENOENT is returned.
1094  *
1095  * vpp is locked and ref'd on return.  If we're looking up DOTDOT, dvp is
1096  * unlocked.  If we're looking up . an extra ref is taken, but the lock is
1097  * not recursively acquired.
1098  */
1099 
1100 int
1101 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1102     struct timespec *tsp, int *ticksp)
1103 {
1104 	struct namecache *ncp;
1105 	struct rwlock *blp;
1106 	struct mtx *dvlp, *dvlp2;
1107 	uint32_t hash;
1108 	int error, ltype;
1109 
1110 	if (!doingcache) {
1111 		cnp->cn_flags &= ~MAKEENTRY;
1112 		return (0);
1113 	}
1114 retry:
1115 	blp = NULL;
1116 	dvlp = VP2VNODELOCK(dvp);
1117 	error = 0;
1118 	counter_u64_add(numcalls, 1);
1119 
1120 	if (cnp->cn_nameptr[0] == '.') {
1121 		if (cnp->cn_namelen == 1) {
1122 			*vpp = dvp;
1123 			CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .",
1124 			    dvp, cnp->cn_nameptr);
1125 			counter_u64_add(dothits, 1);
1126 			SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1127 			if (tsp != NULL)
1128 				timespecclear(tsp);
1129 			if (ticksp != NULL)
1130 				*ticksp = ticks;
1131 			VREF(*vpp);
1132 			/*
1133 			 * When we lookup "." we still can be asked to lock it
1134 			 * differently...
1135 			 */
1136 			ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1137 			if (ltype != VOP_ISLOCKED(*vpp)) {
1138 				if (ltype == LK_EXCLUSIVE) {
1139 					vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1140 					if ((*vpp)->v_iflag & VI_DOOMED) {
1141 						/* forced unmount */
1142 						vrele(*vpp);
1143 						*vpp = NULL;
1144 						return (ENOENT);
1145 					}
1146 				} else
1147 					vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1148 			}
1149 			return (-1);
1150 		}
1151 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1152 			counter_u64_add(dotdothits, 1);
1153 			dvlp2 = NULL;
1154 			mtx_lock(dvlp);
1155 retry_dotdot:
1156 			ncp = dvp->v_cache_dd;
1157 			if (ncp == NULL) {
1158 				SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1159 				    "..", NULL);
1160 				mtx_unlock(dvlp);
1161 				return (0);
1162 			}
1163 			if ((cnp->cn_flags & MAKEENTRY) == 0) {
1164 				if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1165 					if (ncp->nc_dvp != dvp)
1166 						panic("dvp %p v_cache_dd %p\n", dvp, ncp);
1167 					if (!cache_zap_locked_vnode_kl2(ncp,
1168 					    dvp, &dvlp2))
1169 						goto retry_dotdot;
1170 					MPASS(dvp->v_cache_dd == NULL);
1171 					mtx_unlock(dvlp);
1172 					if (dvlp2 != NULL)
1173 						mtx_unlock(dvlp2);
1174 					cache_free(ncp);
1175 				} else {
1176 					dvp->v_cache_dd = NULL;
1177 					mtx_unlock(dvlp);
1178 					if (dvlp2 != NULL)
1179 						mtx_unlock(dvlp2);
1180 				}
1181 				return (0);
1182 			}
1183 			if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1184 				if (ncp->nc_flag & NCF_NEGATIVE)
1185 					*vpp = NULL;
1186 				else
1187 					*vpp = ncp->nc_vp;
1188 			} else
1189 				*vpp = ncp->nc_dvp;
1190 			/* Return failure if negative entry was found. */
1191 			if (*vpp == NULL)
1192 				goto negative_success;
1193 			CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
1194 			    dvp, cnp->cn_nameptr, *vpp);
1195 			SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..",
1196 			    *vpp);
1197 			cache_out_ts(ncp, tsp, ticksp);
1198 			if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1199 			    NCF_DTS && tsp != NULL)
1200 				*tsp = ((struct namecache_ts *)ncp)->
1201 				    nc_dotdottime;
1202 			goto success;
1203 		}
1204 	}
1205 
1206 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1207 	blp = HASH2BUCKETLOCK(hash);
1208 	rw_rlock(blp);
1209 
1210 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1211 		counter_u64_add(numchecks, 1);
1212 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1213 		    !bcmp(nc_get_name(ncp), cnp->cn_nameptr, ncp->nc_nlen))
1214 			break;
1215 	}
1216 
1217 	/* We failed to find an entry */
1218 	if (ncp == NULL) {
1219 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1220 		    NULL);
1221 		if ((cnp->cn_flags & MAKEENTRY) == 0) {
1222 			counter_u64_add(nummisszap, 1);
1223 		} else {
1224 			counter_u64_add(nummiss, 1);
1225 		}
1226 		goto unlock;
1227 	}
1228 
1229 	/* We don't want to have an entry, so dump it */
1230 	if ((cnp->cn_flags & MAKEENTRY) == 0) {
1231 		counter_u64_add(numposzaps, 1);
1232 		goto zap_and_exit;
1233 	}
1234 
1235 	/* We found a "positive" match, return the vnode */
1236 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1237 		counter_u64_add(numposhits, 1);
1238 		*vpp = ncp->nc_vp;
1239 		CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p",
1240 		    dvp, cnp->cn_nameptr, *vpp, ncp);
1241 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, nc_get_name(ncp),
1242 		    *vpp);
1243 		cache_out_ts(ncp, tsp, ticksp);
1244 		goto success;
1245 	}
1246 
1247 negative_success:
1248 	/* We found a negative match, and want to create it, so purge */
1249 	if (cnp->cn_nameiop == CREATE) {
1250 		counter_u64_add(numnegzaps, 1);
1251 		goto zap_and_exit;
1252 	}
1253 
1254 	counter_u64_add(numneghits, 1);
1255 	cache_negative_hit(ncp);
1256 	if (ncp->nc_flag & NCF_WHITE)
1257 		cnp->cn_flags |= ISWHITEOUT;
1258 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp,
1259 	    nc_get_name(ncp));
1260 	cache_out_ts(ncp, tsp, ticksp);
1261 	cache_lookup_unlock(blp, dvlp);
1262 	return (ENOENT);
1263 
1264 success:
1265 	/*
1266 	 * On success we return a locked and ref'd vnode as per the lookup
1267 	 * protocol.
1268 	 */
1269 	MPASS(dvp != *vpp);
1270 	ltype = 0;	/* silence gcc warning */
1271 	if (cnp->cn_flags & ISDOTDOT) {
1272 		ltype = VOP_ISLOCKED(dvp);
1273 		VOP_UNLOCK(dvp, 0);
1274 	}
1275 	vhold(*vpp);
1276 	cache_lookup_unlock(blp, dvlp);
1277 	error = vget(*vpp, cnp->cn_lkflags | LK_VNHELD, cnp->cn_thread);
1278 	if (cnp->cn_flags & ISDOTDOT) {
1279 		vn_lock(dvp, ltype | LK_RETRY);
1280 		if (dvp->v_iflag & VI_DOOMED) {
1281 			if (error == 0)
1282 				vput(*vpp);
1283 			*vpp = NULL;
1284 			return (ENOENT);
1285 		}
1286 	}
1287 	if (error) {
1288 		*vpp = NULL;
1289 		goto retry;
1290 	}
1291 	if ((cnp->cn_flags & ISLASTCN) &&
1292 	    (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) {
1293 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
1294 	}
1295 	return (-1);
1296 
1297 unlock:
1298 	cache_lookup_unlock(blp, dvlp);
1299 	return (0);
1300 
1301 zap_and_exit:
1302 	if (blp != NULL)
1303 		error = cache_zap_rlocked_bucket(ncp, blp);
1304 	else
1305 		error = cache_zap_locked_vnode(ncp, dvp);
1306 	if (error != 0) {
1307 		zap_and_exit_bucket_fail++;
1308 		cache_maybe_yield();
1309 		goto retry;
1310 	}
1311 	cache_free(ncp);
1312 	return (0);
1313 }
1314 
1315 struct celockstate {
1316 	struct mtx *vlp[3];
1317 	struct rwlock *blp[2];
1318 };
1319 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1320 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1321 
1322 static inline void
1323 cache_celockstate_init(struct celockstate *cel)
1324 {
1325 
1326 	bzero(cel, sizeof(*cel));
1327 }
1328 
1329 static void
1330 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1331     struct vnode *dvp)
1332 {
1333 	struct mtx *vlp1, *vlp2;
1334 
1335 	MPASS(cel->vlp[0] == NULL);
1336 	MPASS(cel->vlp[1] == NULL);
1337 	MPASS(cel->vlp[2] == NULL);
1338 
1339 	MPASS(vp != NULL || dvp != NULL);
1340 
1341 	vlp1 = VP2VNODELOCK(vp);
1342 	vlp2 = VP2VNODELOCK(dvp);
1343 	cache_sort(&vlp1, &vlp2);
1344 
1345 	if (vlp1 != NULL) {
1346 		mtx_lock(vlp1);
1347 		cel->vlp[0] = vlp1;
1348 	}
1349 	mtx_lock(vlp2);
1350 	cel->vlp[1] = vlp2;
1351 }
1352 
1353 static void
1354 cache_unlock_vnodes_cel(struct celockstate *cel)
1355 {
1356 
1357 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
1358 
1359 	if (cel->vlp[0] != NULL)
1360 		mtx_unlock(cel->vlp[0]);
1361 	if (cel->vlp[1] != NULL)
1362 		mtx_unlock(cel->vlp[1]);
1363 	if (cel->vlp[2] != NULL)
1364 		mtx_unlock(cel->vlp[2]);
1365 }
1366 
1367 static bool
1368 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
1369 {
1370 	struct mtx *vlp;
1371 	bool ret;
1372 
1373 	cache_assert_vlp_locked(cel->vlp[0]);
1374 	cache_assert_vlp_locked(cel->vlp[1]);
1375 	MPASS(cel->vlp[2] == NULL);
1376 
1377 	vlp = VP2VNODELOCK(vp);
1378 	MPASS(vlp != NULL);
1379 
1380 	ret = true;
1381 	if (vlp >= cel->vlp[1]) {
1382 		mtx_lock(vlp);
1383 	} else {
1384 		if (mtx_trylock(vlp))
1385 			goto out;
1386 		cache_lock_vnodes_cel_3_failures++;
1387 		cache_unlock_vnodes_cel(cel);
1388 		if (vlp < cel->vlp[0]) {
1389 			mtx_lock(vlp);
1390 			mtx_lock(cel->vlp[0]);
1391 			mtx_lock(cel->vlp[1]);
1392 		} else {
1393 			if (cel->vlp[0] != NULL)
1394 				mtx_lock(cel->vlp[0]);
1395 			mtx_lock(vlp);
1396 			mtx_lock(cel->vlp[1]);
1397 		}
1398 		ret = false;
1399 	}
1400 out:
1401 	cel->vlp[2] = vlp;
1402 	return (ret);
1403 }
1404 
1405 static void
1406 cache_lock_buckets_cel(struct celockstate *cel, struct rwlock *blp1,
1407     struct rwlock *blp2)
1408 {
1409 
1410 	MPASS(cel->blp[0] == NULL);
1411 	MPASS(cel->blp[1] == NULL);
1412 
1413 	cache_sort(&blp1, &blp2);
1414 
1415 	if (blp1 != NULL) {
1416 		rw_wlock(blp1);
1417 		cel->blp[0] = blp1;
1418 	}
1419 	rw_wlock(blp2);
1420 	cel->blp[1] = blp2;
1421 }
1422 
1423 static void
1424 cache_unlock_buckets_cel(struct celockstate *cel)
1425 {
1426 
1427 	if (cel->blp[0] != NULL)
1428 		rw_wunlock(cel->blp[0]);
1429 	rw_wunlock(cel->blp[1]);
1430 }
1431 
1432 /*
1433  * Lock part of the cache affected by the insertion.
1434  *
1435  * This means vnodelocks for dvp, vp and the relevant bucketlock.
1436  * However, insertion can result in removal of an old entry. In this
1437  * case we have an additional vnode and bucketlock pair to lock. If the
1438  * entry is negative, ncelock is locked instead of the vnode.
1439  *
1440  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
1441  * preserving the locking order (smaller address first).
1442  */
1443 static void
1444 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1445     uint32_t hash)
1446 {
1447 	struct namecache *ncp;
1448 	struct rwlock *blps[2];
1449 
1450 	blps[0] = HASH2BUCKETLOCK(hash);
1451 	for (;;) {
1452 		blps[1] = NULL;
1453 		cache_lock_vnodes_cel(cel, dvp, vp);
1454 		if (vp == NULL || vp->v_type != VDIR)
1455 			break;
1456 		ncp = vp->v_cache_dd;
1457 		if (ncp == NULL)
1458 			break;
1459 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1460 			break;
1461 		MPASS(ncp->nc_dvp == vp);
1462 		blps[1] = NCP2BUCKETLOCK(ncp);
1463 		if (ncp->nc_flag & NCF_NEGATIVE)
1464 			break;
1465 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1466 			break;
1467 		/*
1468 		 * All vnodes got re-locked. Re-validate the state and if
1469 		 * nothing changed we are done. Otherwise restart.
1470 		 */
1471 		if (ncp == vp->v_cache_dd &&
1472 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1473 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1474 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1475 			break;
1476 		cache_unlock_vnodes_cel(cel);
1477 		cel->vlp[0] = NULL;
1478 		cel->vlp[1] = NULL;
1479 		cel->vlp[2] = NULL;
1480 	}
1481 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1482 }
1483 
1484 static void
1485 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1486     uint32_t hash)
1487 {
1488 	struct namecache *ncp;
1489 	struct rwlock *blps[2];
1490 
1491 	blps[0] = HASH2BUCKETLOCK(hash);
1492 	for (;;) {
1493 		blps[1] = NULL;
1494 		cache_lock_vnodes_cel(cel, dvp, vp);
1495 		ncp = dvp->v_cache_dd;
1496 		if (ncp == NULL)
1497 			break;
1498 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1499 			break;
1500 		MPASS(ncp->nc_dvp == dvp);
1501 		blps[1] = NCP2BUCKETLOCK(ncp);
1502 		if (ncp->nc_flag & NCF_NEGATIVE)
1503 			break;
1504 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1505 			break;
1506 		if (ncp == dvp->v_cache_dd &&
1507 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1508 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1509 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1510 			break;
1511 		cache_unlock_vnodes_cel(cel);
1512 		cel->vlp[0] = NULL;
1513 		cel->vlp[1] = NULL;
1514 		cel->vlp[2] = NULL;
1515 	}
1516 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1517 }
1518 
1519 static void
1520 cache_enter_unlock(struct celockstate *cel)
1521 {
1522 
1523 	cache_unlock_buckets_cel(cel);
1524 	cache_unlock_vnodes_cel(cel);
1525 }
1526 
1527 /*
1528  * Add an entry to the cache.
1529  */
1530 void
1531 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
1532     struct timespec *tsp, struct timespec *dtsp)
1533 {
1534 	struct celockstate cel;
1535 	struct namecache *ncp, *n2, *ndd;
1536 	struct namecache_ts *n3;
1537 	struct nchashhead *ncpp;
1538 	struct neglist *neglist;
1539 	uint32_t hash;
1540 	int flag;
1541 	int len;
1542 	bool neg_locked;
1543 
1544 	CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr);
1545 	VNASSERT(vp == NULL || (vp->v_iflag & VI_DOOMED) == 0, vp,
1546 	    ("cache_enter: Adding a doomed vnode"));
1547 	VNASSERT(dvp == NULL || (dvp->v_iflag & VI_DOOMED) == 0, dvp,
1548 	    ("cache_enter: Doomed vnode used as src"));
1549 
1550 	if (!doingcache)
1551 		return;
1552 
1553 	/*
1554 	 * Avoid blowout in namecache entries.
1555 	 */
1556 	if (numcache >= desiredvnodes * ncsizefactor)
1557 		return;
1558 
1559 	cache_celockstate_init(&cel);
1560 	ndd = NULL;
1561 	flag = 0;
1562 	if (cnp->cn_nameptr[0] == '.') {
1563 		if (cnp->cn_namelen == 1)
1564 			return;
1565 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1566 			len = cnp->cn_namelen;
1567 			hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1568 			cache_enter_lock_dd(&cel, dvp, vp, hash);
1569 			/*
1570 			 * If dotdot entry already exists, just retarget it
1571 			 * to new parent vnode, otherwise continue with new
1572 			 * namecache entry allocation.
1573 			 */
1574 			if ((ncp = dvp->v_cache_dd) != NULL &&
1575 			    ncp->nc_flag & NCF_ISDOTDOT) {
1576 				KASSERT(ncp->nc_dvp == dvp,
1577 				    ("wrong isdotdot parent"));
1578 				neg_locked = false;
1579 				if (ncp->nc_flag & NCF_NEGATIVE || vp == NULL) {
1580 					neglist = NCP2NEGLIST(ncp);
1581 					mtx_lock(&ncneg_hot.nl_lock);
1582 					mtx_lock(&neglist->nl_lock);
1583 					neg_locked = true;
1584 				}
1585 				if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1586 					TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst,
1587 					    ncp, nc_dst);
1588 				} else {
1589 					cache_negative_remove(ncp, true);
1590 				}
1591 				if (vp != NULL) {
1592 					TAILQ_INSERT_HEAD(&vp->v_cache_dst,
1593 					    ncp, nc_dst);
1594 					ncp->nc_flag &= ~(NCF_NEGATIVE|NCF_HOTNEGATIVE);
1595 				} else {
1596 					ncp->nc_flag &= ~(NCF_HOTNEGATIVE);
1597 					ncp->nc_flag |= NCF_NEGATIVE;
1598 					cache_negative_insert(ncp, true);
1599 				}
1600 				if (neg_locked) {
1601 					mtx_unlock(&neglist->nl_lock);
1602 					mtx_unlock(&ncneg_hot.nl_lock);
1603 				}
1604 				ncp->nc_vp = vp;
1605 				cache_enter_unlock(&cel);
1606 				return;
1607 			}
1608 			dvp->v_cache_dd = NULL;
1609 			cache_enter_unlock(&cel);
1610 			cache_celockstate_init(&cel);
1611 			SDT_PROBE3(vfs, namecache, enter, done, dvp, "..", vp);
1612 			flag = NCF_ISDOTDOT;
1613 		}
1614 	}
1615 
1616 	/*
1617 	 * Calculate the hash key and setup as much of the new
1618 	 * namecache entry as possible before acquiring the lock.
1619 	 */
1620 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
1621 	ncp->nc_flag = flag;
1622 	ncp->nc_vp = vp;
1623 	if (vp == NULL)
1624 		ncp->nc_flag |= NCF_NEGATIVE;
1625 	ncp->nc_dvp = dvp;
1626 	if (tsp != NULL) {
1627 		n3 = (struct namecache_ts *)ncp;
1628 		n3->nc_time = *tsp;
1629 		n3->nc_ticks = ticks;
1630 		n3->nc_flag |= NCF_TS;
1631 		if (dtsp != NULL) {
1632 			n3->nc_dotdottime = *dtsp;
1633 			n3->nc_flag |= NCF_DTS;
1634 		}
1635 	}
1636 	len = ncp->nc_nlen = cnp->cn_namelen;
1637 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1638 	strlcpy(nc_get_name(ncp), cnp->cn_nameptr, len + 1);
1639 	cache_enter_lock(&cel, dvp, vp, hash);
1640 
1641 	/*
1642 	 * See if this vnode or negative entry is already in the cache
1643 	 * with this name.  This can happen with concurrent lookups of
1644 	 * the same path name.
1645 	 */
1646 	ncpp = NCHHASH(hash);
1647 	LIST_FOREACH(n2, ncpp, nc_hash) {
1648 		if (n2->nc_dvp == dvp &&
1649 		    n2->nc_nlen == cnp->cn_namelen &&
1650 		    !bcmp(nc_get_name(n2), cnp->cn_nameptr, n2->nc_nlen)) {
1651 			if (tsp != NULL) {
1652 				KASSERT((n2->nc_flag & NCF_TS) != 0,
1653 				    ("no NCF_TS"));
1654 				n3 = (struct namecache_ts *)n2;
1655 				n3->nc_time =
1656 				    ((struct namecache_ts *)ncp)->nc_time;
1657 				n3->nc_ticks =
1658 				    ((struct namecache_ts *)ncp)->nc_ticks;
1659 				if (dtsp != NULL) {
1660 					n3->nc_dotdottime =
1661 					    ((struct namecache_ts *)ncp)->
1662 					    nc_dotdottime;
1663 					if (ncp->nc_flag & NCF_NEGATIVE)
1664 						mtx_lock(&ncneg_hot.nl_lock);
1665 					n3->nc_flag |= NCF_DTS;
1666 					if (ncp->nc_flag & NCF_NEGATIVE)
1667 						mtx_unlock(&ncneg_hot.nl_lock);
1668 				}
1669 			}
1670 			goto out_unlock_free;
1671 		}
1672 	}
1673 
1674 	if (flag == NCF_ISDOTDOT) {
1675 		/*
1676 		 * See if we are trying to add .. entry, but some other lookup
1677 		 * has populated v_cache_dd pointer already.
1678 		 */
1679 		if (dvp->v_cache_dd != NULL)
1680 			goto out_unlock_free;
1681 		KASSERT(vp == NULL || vp->v_type == VDIR,
1682 		    ("wrong vnode type %p", vp));
1683 		dvp->v_cache_dd = ncp;
1684 	}
1685 
1686 	atomic_add_rel_long(&numcache, 1);
1687 	if (vp != NULL) {
1688 		if (vp->v_type == VDIR) {
1689 			if (flag != NCF_ISDOTDOT) {
1690 				/*
1691 				 * For this case, the cache entry maps both the
1692 				 * directory name in it and the name ".." for the
1693 				 * directory's parent.
1694 				 */
1695 				if ((ndd = vp->v_cache_dd) != NULL) {
1696 					if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
1697 						cache_zap_locked(ndd, false);
1698 					else
1699 						ndd = NULL;
1700 				}
1701 				vp->v_cache_dd = ncp;
1702 			}
1703 		} else {
1704 			vp->v_cache_dd = NULL;
1705 		}
1706 	}
1707 
1708 	if (flag != NCF_ISDOTDOT) {
1709 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1710 			vhold(dvp);
1711 			atomic_add_rel_long(&numcachehv, 1);
1712 		}
1713 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
1714 	}
1715 
1716 	/*
1717 	 * Insert the new namecache entry into the appropriate chain
1718 	 * within the cache entries table.
1719 	 */
1720 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
1721 
1722 	/*
1723 	 * If the entry is "negative", we place it into the
1724 	 * "negative" cache queue, otherwise, we place it into the
1725 	 * destination vnode's cache entries queue.
1726 	 */
1727 	if (vp != NULL) {
1728 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
1729 		SDT_PROBE3(vfs, namecache, enter, done, dvp, nc_get_name(ncp),
1730 		    vp);
1731 	} else {
1732 		if (cnp->cn_flags & ISWHITEOUT)
1733 			ncp->nc_flag |= NCF_WHITE;
1734 		cache_negative_insert(ncp, false);
1735 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
1736 		    nc_get_name(ncp));
1737 	}
1738 	cache_enter_unlock(&cel);
1739 	if (numneg * ncnegfactor > numcache)
1740 		cache_negative_zap_one();
1741 	cache_free(ndd);
1742 	return;
1743 out_unlock_free:
1744 	cache_enter_unlock(&cel);
1745 	cache_free(ncp);
1746 	return;
1747 }
1748 
1749 static u_int
1750 cache_roundup_2(u_int val)
1751 {
1752 	u_int res;
1753 
1754 	for (res = 1; res <= val; res <<= 1)
1755 		continue;
1756 
1757 	return (res);
1758 }
1759 
1760 /*
1761  * Name cache initialization, from vfs_init() when we are booting
1762  */
1763 static void
1764 nchinit(void *dummy __unused)
1765 {
1766 	u_int i;
1767 
1768 	cache_zone_small = uma_zcreate("S VFS Cache",
1769 	    sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1,
1770 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
1771 	cache_zone_small_ts = uma_zcreate("STS VFS Cache",
1772 	    sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1,
1773 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
1774 	cache_zone_large = uma_zcreate("L VFS Cache",
1775 	    sizeof(struct namecache) + NAME_MAX + 1,
1776 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
1777 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache",
1778 	    sizeof(struct namecache_ts) + NAME_MAX + 1,
1779 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
1780 
1781 	nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash);
1782 	numbucketlocks = cache_roundup_2(mp_ncpus * 64);
1783 	if (numbucketlocks > nchash + 1)
1784 		numbucketlocks = nchash + 1;
1785 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
1786 	    M_WAITOK | M_ZERO);
1787 	for (i = 0; i < numbucketlocks; i++)
1788 		rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK | RW_RECURSE);
1789 	numvnodelocks = cache_roundup_2(mp_ncpus * 64);
1790 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
1791 	    M_WAITOK | M_ZERO);
1792 	for (i = 0; i < numvnodelocks; i++)
1793 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
1794 	ncpurgeminvnodes = numbucketlocks;
1795 
1796 	numneglists = 4;
1797 	neglists = malloc(sizeof(*neglists) * numneglists, M_VFSCACHE,
1798 	    M_WAITOK | M_ZERO);
1799 	for (i = 0; i < numneglists; i++) {
1800 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
1801 		TAILQ_INIT(&neglists[i].nl_list);
1802 	}
1803 	mtx_init(&ncneg_hot.nl_lock, "ncneglh", NULL, MTX_DEF);
1804 	TAILQ_INIT(&ncneg_hot.nl_list);
1805 
1806 	numcalls = counter_u64_alloc(M_WAITOK);
1807 	dothits = counter_u64_alloc(M_WAITOK);
1808 	dotdothits = counter_u64_alloc(M_WAITOK);
1809 	numchecks = counter_u64_alloc(M_WAITOK);
1810 	nummiss = counter_u64_alloc(M_WAITOK);
1811 	nummisszap = counter_u64_alloc(M_WAITOK);
1812 	numposzaps = counter_u64_alloc(M_WAITOK);
1813 	numposhits = counter_u64_alloc(M_WAITOK);
1814 	numnegzaps = counter_u64_alloc(M_WAITOK);
1815 	numneghits = counter_u64_alloc(M_WAITOK);
1816 	numfullpathcalls = counter_u64_alloc(M_WAITOK);
1817 	numfullpathfail1 = counter_u64_alloc(M_WAITOK);
1818 	numfullpathfail2 = counter_u64_alloc(M_WAITOK);
1819 	numfullpathfail4 = counter_u64_alloc(M_WAITOK);
1820 	numfullpathfound = counter_u64_alloc(M_WAITOK);
1821 }
1822 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
1823 
1824 void
1825 cache_changesize(int newmaxvnodes)
1826 {
1827 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
1828 	u_long new_nchash, old_nchash;
1829 	struct namecache *ncp;
1830 	uint32_t hash;
1831 	int i;
1832 
1833 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
1834 	if (newmaxvnodes < numbucketlocks)
1835 		newmaxvnodes = numbucketlocks;
1836 
1837 	new_nchashtbl = hashinit(newmaxvnodes, M_VFSCACHE, &new_nchash);
1838 	/* If same hash table size, nothing to do */
1839 	if (nchash == new_nchash) {
1840 		free(new_nchashtbl, M_VFSCACHE);
1841 		return;
1842 	}
1843 	/*
1844 	 * Move everything from the old hash table to the new table.
1845 	 * None of the namecache entries in the table can be removed
1846 	 * because to do so, they have to be removed from the hash table.
1847 	 */
1848 	cache_lock_all_vnodes();
1849 	cache_lock_all_buckets();
1850 	old_nchashtbl = nchashtbl;
1851 	old_nchash = nchash;
1852 	nchashtbl = new_nchashtbl;
1853 	nchash = new_nchash;
1854 	for (i = 0; i <= old_nchash; i++) {
1855 		while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) {
1856 			hash = cache_get_hash(nc_get_name(ncp), ncp->nc_nlen,
1857 			    ncp->nc_dvp);
1858 			LIST_REMOVE(ncp, nc_hash);
1859 			LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
1860 		}
1861 	}
1862 	cache_unlock_all_buckets();
1863 	cache_unlock_all_vnodes();
1864 	free(old_nchashtbl, M_VFSCACHE);
1865 }
1866 
1867 /*
1868  * Invalidate all entries to a particular vnode.
1869  */
1870 void
1871 cache_purge(struct vnode *vp)
1872 {
1873 	TAILQ_HEAD(, namecache) ncps;
1874 	struct namecache *ncp, *nnp;
1875 	struct mtx *vlp, *vlp2;
1876 
1877 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
1878 	SDT_PROBE1(vfs, namecache, purge, done, vp);
1879 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
1880 	    vp->v_cache_dd == NULL)
1881 		return;
1882 	TAILQ_INIT(&ncps);
1883 	vlp = VP2VNODELOCK(vp);
1884 	vlp2 = NULL;
1885 	mtx_lock(vlp);
1886 retry:
1887 	while (!LIST_EMPTY(&vp->v_cache_src)) {
1888 		ncp = LIST_FIRST(&vp->v_cache_src);
1889 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
1890 			goto retry;
1891 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1892 	}
1893 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
1894 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
1895 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
1896 			goto retry;
1897 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1898 	}
1899 	ncp = vp->v_cache_dd;
1900 	if (ncp != NULL) {
1901 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
1902 		   ("lost dotdot link"));
1903 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
1904 			goto retry;
1905 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1906 	}
1907 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
1908 	mtx_unlock(vlp);
1909 	if (vlp2 != NULL)
1910 		mtx_unlock(vlp2);
1911 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
1912 		cache_free(ncp);
1913 	}
1914 }
1915 
1916 /*
1917  * Invalidate all negative entries for a particular directory vnode.
1918  */
1919 void
1920 cache_purge_negative(struct vnode *vp)
1921 {
1922 	TAILQ_HEAD(, namecache) ncps;
1923 	struct namecache *ncp, *nnp;
1924 	struct mtx *vlp;
1925 
1926 	CTR1(KTR_VFS, "cache_purge_negative(%p)", vp);
1927 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
1928 	TAILQ_INIT(&ncps);
1929 	vlp = VP2VNODELOCK(vp);
1930 	mtx_lock(vlp);
1931 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
1932 		if (!(ncp->nc_flag & NCF_NEGATIVE))
1933 			continue;
1934 		cache_zap_negative_locked_vnode_kl(ncp, vp);
1935 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1936 	}
1937 	mtx_unlock(vlp);
1938 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
1939 		cache_free(ncp);
1940 	}
1941 }
1942 
1943 /*
1944  * Flush all entries referencing a particular filesystem.
1945  */
1946 void
1947 cache_purgevfs(struct mount *mp, bool force)
1948 {
1949 	TAILQ_HEAD(, namecache) ncps;
1950 	struct mtx *vlp1, *vlp2;
1951 	struct rwlock *blp;
1952 	struct nchashhead *bucket;
1953 	struct namecache *ncp, *nnp;
1954 	u_long i, j, n_nchash;
1955 	int error;
1956 
1957 	/* Scan hash tables for applicable entries */
1958 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
1959 	if (!force && mp->mnt_nvnodelistsize <= ncpurgeminvnodes)
1960 		return;
1961 	TAILQ_INIT(&ncps);
1962 	n_nchash = nchash + 1;
1963 	vlp1 = vlp2 = NULL;
1964 	for (i = 0; i < numbucketlocks; i++) {
1965 		blp = (struct rwlock *)&bucketlocks[i];
1966 		rw_wlock(blp);
1967 		for (j = i; j < n_nchash; j += numbucketlocks) {
1968 retry:
1969 			bucket = &nchashtbl[j];
1970 			LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) {
1971 				cache_assert_bucket_locked(ncp, RA_WLOCKED);
1972 				if (ncp->nc_dvp->v_mount != mp)
1973 					continue;
1974 				error = cache_zap_wlocked_bucket_kl(ncp, blp,
1975 				    &vlp1, &vlp2);
1976 				if (error != 0)
1977 					goto retry;
1978 				TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst);
1979 			}
1980 		}
1981 		rw_wunlock(blp);
1982 		if (vlp1 == NULL && vlp2 == NULL)
1983 			cache_maybe_yield();
1984 	}
1985 	if (vlp1 != NULL)
1986 		mtx_unlock(vlp1);
1987 	if (vlp2 != NULL)
1988 		mtx_unlock(vlp2);
1989 
1990 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
1991 		cache_free(ncp);
1992 	}
1993 }
1994 
1995 /*
1996  * Perform canonical checks and cache lookup and pass on to filesystem
1997  * through the vop_cachedlookup only if needed.
1998  */
1999 
2000 int
2001 vfs_cache_lookup(struct vop_lookup_args *ap)
2002 {
2003 	struct vnode *dvp;
2004 	int error;
2005 	struct vnode **vpp = ap->a_vpp;
2006 	struct componentname *cnp = ap->a_cnp;
2007 	struct ucred *cred = cnp->cn_cred;
2008 	int flags = cnp->cn_flags;
2009 	struct thread *td = cnp->cn_thread;
2010 
2011 	*vpp = NULL;
2012 	dvp = ap->a_dvp;
2013 
2014 	if (dvp->v_type != VDIR)
2015 		return (ENOTDIR);
2016 
2017 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2018 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2019 		return (EROFS);
2020 
2021 	error = VOP_ACCESS(dvp, VEXEC, cred, td);
2022 	if (error)
2023 		return (error);
2024 
2025 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2026 	if (error == 0)
2027 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2028 	if (error == -1)
2029 		return (0);
2030 	return (error);
2031 }
2032 
2033 /*
2034  * XXX All of these sysctls would probably be more productive dead.
2035  */
2036 static int disablecwd;
2037 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
2038    "Disable the getcwd syscall");
2039 
2040 /* Implementation of the getcwd syscall. */
2041 int
2042 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2043 {
2044 
2045 	return (kern___getcwd(td, uap->buf, UIO_USERSPACE, uap->buflen,
2046 	    MAXPATHLEN));
2047 }
2048 
2049 int
2050 kern___getcwd(struct thread *td, char *buf, enum uio_seg bufseg, u_int buflen,
2051     u_int path_max)
2052 {
2053 	char *bp, *tmpbuf;
2054 	struct filedesc *fdp;
2055 	struct vnode *cdir, *rdir;
2056 	int error;
2057 
2058 	if (disablecwd)
2059 		return (ENODEV);
2060 	if (buflen < 2)
2061 		return (EINVAL);
2062 	if (buflen > path_max)
2063 		buflen = path_max;
2064 
2065 	tmpbuf = malloc(buflen, M_TEMP, M_WAITOK);
2066 	fdp = td->td_proc->p_fd;
2067 	FILEDESC_SLOCK(fdp);
2068 	cdir = fdp->fd_cdir;
2069 	vrefact(cdir);
2070 	rdir = fdp->fd_rdir;
2071 	vrefact(rdir);
2072 	FILEDESC_SUNLOCK(fdp);
2073 	error = vn_fullpath1(td, cdir, rdir, tmpbuf, &bp, buflen);
2074 	vrele(rdir);
2075 	vrele(cdir);
2076 
2077 	if (!error) {
2078 		if (bufseg == UIO_SYSSPACE)
2079 			bcopy(bp, buf, strlen(bp) + 1);
2080 		else
2081 			error = copyout(bp, buf, strlen(bp) + 1);
2082 #ifdef KTRACE
2083 	if (KTRPOINT(curthread, KTR_NAMEI))
2084 		ktrnamei(bp);
2085 #endif
2086 	}
2087 	free(tmpbuf, M_TEMP);
2088 	return (error);
2089 }
2090 
2091 /*
2092  * Thus begins the fullpath magic.
2093  */
2094 
2095 static int disablefullpath;
2096 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, &disablefullpath, 0,
2097     "Disable the vn_fullpath function");
2098 
2099 /*
2100  * Retrieve the full filesystem path that correspond to a vnode from the name
2101  * cache (if available)
2102  */
2103 int
2104 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf)
2105 {
2106 	char *buf;
2107 	struct filedesc *fdp;
2108 	struct vnode *rdir;
2109 	int error;
2110 
2111 	if (disablefullpath)
2112 		return (ENODEV);
2113 	if (vn == NULL)
2114 		return (EINVAL);
2115 
2116 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2117 	fdp = td->td_proc->p_fd;
2118 	FILEDESC_SLOCK(fdp);
2119 	rdir = fdp->fd_rdir;
2120 	VREF(rdir);
2121 	FILEDESC_SUNLOCK(fdp);
2122 	error = vn_fullpath1(td, vn, rdir, buf, retbuf, MAXPATHLEN);
2123 	vrele(rdir);
2124 
2125 	if (!error)
2126 		*freebuf = buf;
2127 	else
2128 		free(buf, M_TEMP);
2129 	return (error);
2130 }
2131 
2132 /*
2133  * This function is similar to vn_fullpath, but it attempts to lookup the
2134  * pathname relative to the global root mount point.  This is required for the
2135  * auditing sub-system, as audited pathnames must be absolute, relative to the
2136  * global root mount point.
2137  */
2138 int
2139 vn_fullpath_global(struct thread *td, struct vnode *vn,
2140     char **retbuf, char **freebuf)
2141 {
2142 	char *buf;
2143 	int error;
2144 
2145 	if (disablefullpath)
2146 		return (ENODEV);
2147 	if (vn == NULL)
2148 		return (EINVAL);
2149 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2150 	error = vn_fullpath1(td, vn, rootvnode, buf, retbuf, MAXPATHLEN);
2151 	if (!error)
2152 		*freebuf = buf;
2153 	else
2154 		free(buf, M_TEMP);
2155 	return (error);
2156 }
2157 
2158 int
2159 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen)
2160 {
2161 	struct vnode *dvp;
2162 	struct namecache *ncp;
2163 	struct mtx *vlp;
2164 	int error;
2165 
2166 	vlp = VP2VNODELOCK(*vp);
2167 	mtx_lock(vlp);
2168 	TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) {
2169 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2170 			break;
2171 	}
2172 	if (ncp != NULL) {
2173 		if (*buflen < ncp->nc_nlen) {
2174 			mtx_unlock(vlp);
2175 			vrele(*vp);
2176 			counter_u64_add(numfullpathfail4, 1);
2177 			error = ENOMEM;
2178 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2179 			    vp, NULL);
2180 			return (error);
2181 		}
2182 		*buflen -= ncp->nc_nlen;
2183 		memcpy(buf + *buflen, nc_get_name(ncp), ncp->nc_nlen);
2184 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2185 		    nc_get_name(ncp), vp);
2186 		dvp = *vp;
2187 		*vp = ncp->nc_dvp;
2188 		vref(*vp);
2189 		mtx_unlock(vlp);
2190 		vrele(dvp);
2191 		return (0);
2192 	}
2193 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2194 
2195 	mtx_unlock(vlp);
2196 	vn_lock(*vp, LK_SHARED | LK_RETRY);
2197 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
2198 	vput(*vp);
2199 	if (error) {
2200 		counter_u64_add(numfullpathfail2, 1);
2201 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
2202 		return (error);
2203 	}
2204 
2205 	*vp = dvp;
2206 	if (dvp->v_iflag & VI_DOOMED) {
2207 		/* forced unmount */
2208 		vrele(dvp);
2209 		error = ENOENT;
2210 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2211 		return (error);
2212 	}
2213 	/*
2214 	 * *vp has its use count incremented still.
2215 	 */
2216 
2217 	return (0);
2218 }
2219 
2220 /*
2221  * The magic behind kern___getcwd() and vn_fullpath().
2222  */
2223 static int
2224 vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
2225     char *buf, char **retbuf, u_int buflen)
2226 {
2227 	int error, slash_prefixed;
2228 #ifdef KDTRACE_HOOKS
2229 	struct vnode *startvp = vp;
2230 #endif
2231 	struct vnode *vp1;
2232 
2233 	buflen--;
2234 	buf[buflen] = '\0';
2235 	error = 0;
2236 	slash_prefixed = 0;
2237 
2238 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
2239 	counter_u64_add(numfullpathcalls, 1);
2240 	vref(vp);
2241 	if (vp->v_type != VDIR) {
2242 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2243 		if (error)
2244 			return (error);
2245 		if (buflen == 0) {
2246 			vrele(vp);
2247 			return (ENOMEM);
2248 		}
2249 		buf[--buflen] = '/';
2250 		slash_prefixed = 1;
2251 	}
2252 	while (vp != rdir && vp != rootvnode) {
2253 		/*
2254 		 * The vp vnode must be already fully constructed,
2255 		 * since it is either found in namecache or obtained
2256 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
2257 		 * without obtaining the vnode lock.
2258 		 */
2259 		if ((vp->v_vflag & VV_ROOT) != 0) {
2260 			vn_lock(vp, LK_RETRY | LK_SHARED);
2261 
2262 			/*
2263 			 * With the vnode locked, check for races with
2264 			 * unmount, forced or not.  Note that we
2265 			 * already verified that vp is not equal to
2266 			 * the root vnode, which means that
2267 			 * mnt_vnodecovered can be NULL only for the
2268 			 * case of unmount.
2269 			 */
2270 			if ((vp->v_iflag & VI_DOOMED) != 0 ||
2271 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
2272 			    vp1->v_mountedhere != vp->v_mount) {
2273 				vput(vp);
2274 				error = ENOENT;
2275 				SDT_PROBE3(vfs, namecache, fullpath, return,
2276 				    error, vp, NULL);
2277 				break;
2278 			}
2279 
2280 			vref(vp1);
2281 			vput(vp);
2282 			vp = vp1;
2283 			continue;
2284 		}
2285 		if (vp->v_type != VDIR) {
2286 			vrele(vp);
2287 			counter_u64_add(numfullpathfail1, 1);
2288 			error = ENOTDIR;
2289 			SDT_PROBE3(vfs, namecache, fullpath, return,
2290 			    error, vp, NULL);
2291 			break;
2292 		}
2293 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2294 		if (error)
2295 			break;
2296 		if (buflen == 0) {
2297 			vrele(vp);
2298 			error = ENOMEM;
2299 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2300 			    startvp, NULL);
2301 			break;
2302 		}
2303 		buf[--buflen] = '/';
2304 		slash_prefixed = 1;
2305 	}
2306 	if (error)
2307 		return (error);
2308 	if (!slash_prefixed) {
2309 		if (buflen == 0) {
2310 			vrele(vp);
2311 			counter_u64_add(numfullpathfail4, 1);
2312 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
2313 			    startvp, NULL);
2314 			return (ENOMEM);
2315 		}
2316 		buf[--buflen] = '/';
2317 	}
2318 	counter_u64_add(numfullpathfound, 1);
2319 	vrele(vp);
2320 
2321 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, buf + buflen);
2322 	*retbuf = buf + buflen;
2323 	return (0);
2324 }
2325 
2326 struct vnode *
2327 vn_dir_dd_ino(struct vnode *vp)
2328 {
2329 	struct namecache *ncp;
2330 	struct vnode *ddvp;
2331 	struct mtx *vlp;
2332 
2333 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
2334 	vlp = VP2VNODELOCK(vp);
2335 	mtx_lock(vlp);
2336 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
2337 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
2338 			continue;
2339 		ddvp = ncp->nc_dvp;
2340 		vhold(ddvp);
2341 		mtx_unlock(vlp);
2342 		if (vget(ddvp, LK_SHARED | LK_NOWAIT | LK_VNHELD, curthread))
2343 			return (NULL);
2344 		return (ddvp);
2345 	}
2346 	mtx_unlock(vlp);
2347 	return (NULL);
2348 }
2349 
2350 int
2351 vn_commname(struct vnode *vp, char *buf, u_int buflen)
2352 {
2353 	struct namecache *ncp;
2354 	struct mtx *vlp;
2355 	int l;
2356 
2357 	vlp = VP2VNODELOCK(vp);
2358 	mtx_lock(vlp);
2359 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
2360 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2361 			break;
2362 	if (ncp == NULL) {
2363 		mtx_unlock(vlp);
2364 		return (ENOENT);
2365 	}
2366 	l = min(ncp->nc_nlen, buflen - 1);
2367 	memcpy(buf, nc_get_name(ncp), l);
2368 	mtx_unlock(vlp);
2369 	buf[l] = '\0';
2370 	return (0);
2371 }
2372 
2373 /* ABI compat shims for old kernel modules. */
2374 #undef cache_enter
2375 
2376 void	cache_enter(struct vnode *dvp, struct vnode *vp,
2377 	    struct componentname *cnp);
2378 
2379 void
2380 cache_enter(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
2381 {
2382 
2383 	cache_enter_time(dvp, vp, cnp, NULL, NULL);
2384 }
2385 
2386 /*
2387  * This function updates path string to vnode's full global path
2388  * and checks the size of the new path string against the pathlen argument.
2389  *
2390  * Requires a locked, referenced vnode.
2391  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
2392  *
2393  * If sysctl debug.disablefullpath is set, ENODEV is returned,
2394  * vnode is left locked and path remain untouched.
2395  *
2396  * If vp is a directory, the call to vn_fullpath_global() always succeeds
2397  * because it falls back to the ".." lookup if the namecache lookup fails.
2398  */
2399 int
2400 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
2401     u_int pathlen)
2402 {
2403 	struct nameidata nd;
2404 	struct vnode *vp1;
2405 	char *rpath, *fbuf;
2406 	int error;
2407 
2408 	ASSERT_VOP_ELOCKED(vp, __func__);
2409 
2410 	/* Return ENODEV if sysctl debug.disablefullpath==1 */
2411 	if (disablefullpath)
2412 		return (ENODEV);
2413 
2414 	/* Construct global filesystem path from vp. */
2415 	VOP_UNLOCK(vp, 0);
2416 	error = vn_fullpath_global(td, vp, &rpath, &fbuf);
2417 
2418 	if (error != 0) {
2419 		vrele(vp);
2420 		return (error);
2421 	}
2422 
2423 	if (strlen(rpath) >= pathlen) {
2424 		vrele(vp);
2425 		error = ENAMETOOLONG;
2426 		goto out;
2427 	}
2428 
2429 	/*
2430 	 * Re-lookup the vnode by path to detect a possible rename.
2431 	 * As a side effect, the vnode is relocked.
2432 	 * If vnode was renamed, return ENOENT.
2433 	 */
2434 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
2435 	    UIO_SYSSPACE, path, td);
2436 	error = namei(&nd);
2437 	if (error != 0) {
2438 		vrele(vp);
2439 		goto out;
2440 	}
2441 	NDFREE(&nd, NDF_ONLY_PNBUF);
2442 	vp1 = nd.ni_vp;
2443 	vrele(vp);
2444 	if (vp1 == vp)
2445 		strcpy(path, rpath);
2446 	else {
2447 		vput(vp1);
2448 		error = ENOENT;
2449 	}
2450 
2451 out:
2452 	free(fbuf, M_TEMP);
2453 	return (error);
2454 }
2455