xref: /freebsd/sys/kern/vfs_cache.c (revision ea825d02749f382c3f7e17f28247f20a48733eab)
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 	struct	timespec nc_time;	/* timespec provided by fs */
121 	struct	timespec nc_dotdottime;	/* dotdot timespec provided by fs */
122 	int	nc_ticks;		/* ticks value when entry was added */
123 	struct namecache nc_nc;
124 };
125 
126 #define	nc_vp		n_un.nu_vp
127 #define	nc_neghits	n_un.nu_neghits
128 
129 /*
130  * Flags in namecache.nc_flag
131  */
132 #define NCF_WHITE	0x01
133 #define NCF_ISDOTDOT	0x02
134 #define	NCF_TS		0x04
135 #define	NCF_DTS		0x08
136 #define	NCF_DVDROP	0x10
137 #define	NCF_NEGATIVE	0x20
138 #define	NCF_HOTNEGATIVE	0x40
139 
140 /*
141  * Name caching works as follows:
142  *
143  * Names found by directory scans are retained in a cache
144  * for future reference.  It is managed LRU, so frequently
145  * used names will hang around.  Cache is indexed by hash value
146  * obtained from (vp, name) where vp refers to the directory
147  * containing name.
148  *
149  * If it is a "negative" entry, (i.e. for a name that is known NOT to
150  * exist) the vnode pointer will be NULL.
151  *
152  * Upon reaching the last segment of a path, if the reference
153  * is for DELETE, or NOCACHE is set (rewrite), and the
154  * name is located in the cache, it will be dropped.
155  *
156  * These locks are used (in the order in which they can be taken):
157  * NAME		TYPE	ROLE
158  * vnodelock	mtx	vnode lists and v_cache_dd field protection
159  * bucketlock	rwlock	for access to given set of hash buckets
160  * neglist	mtx	negative entry LRU management
161  *
162  * Additionally, ncneg_shrink_lock mtx is used to have at most one thread
163  * shrinking the LRU list.
164  *
165  * It is legal to take multiple vnodelock and bucketlock locks. The locking
166  * order is lower address first. Both are recursive.
167  *
168  * "." lookups are lockless.
169  *
170  * ".." and vnode -> name lookups require vnodelock.
171  *
172  * name -> vnode lookup requires the relevant bucketlock to be held for reading.
173  *
174  * Insertions and removals of entries require involved vnodes and bucketlocks
175  * to be write-locked to prevent other threads from seeing the entry.
176  *
177  * Some lookups result in removal of the found entry (e.g. getting rid of a
178  * negative entry with the intent to create a positive one), which poses a
179  * problem when multiple threads reach the state. Similarly, two different
180  * threads can purge two different vnodes and try to remove the same name.
181  *
182  * If the already held vnode lock is lower than the second required lock, we
183  * can just take the other lock. However, in the opposite case, this could
184  * deadlock. As such, this is resolved by trylocking and if that fails unlocking
185  * the first node, locking everything in order and revalidating the state.
186  */
187 
188 /*
189  * Structures associated with name caching.
190  */
191 #define NCHHASH(hash) \
192 	(&nchashtbl[(hash) & nchash])
193 static __read_mostly LIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */
194 static u_long __read_mostly	nchash;			/* size of hash table */
195 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
196     "Size of namecache hash table");
197 static u_long __read_mostly	ncnegfactor = 16; /* ratio of negative entries */
198 SYSCTL_ULONG(_vfs, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
199     "Ratio of negative namecache entries");
200 static u_long __exclusive_cache_line	numneg;	/* number of negative entries allocated */
201 SYSCTL_ULONG(_debug, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0,
202     "Number of negative entries in namecache");
203 static u_long __exclusive_cache_line	numcache;/* number of cache entries allocated */
204 SYSCTL_ULONG(_debug, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0,
205     "Number of namecache entries");
206 static u_long __exclusive_cache_line	numcachehv;/* number of cache entries with vnodes held */
207 SYSCTL_ULONG(_debug, OID_AUTO, numcachehv, CTLFLAG_RD, &numcachehv, 0,
208     "Number of namecache entries with vnodes held");
209 u_int __read_mostly	ncsizefactor = 2;
210 SYSCTL_UINT(_vfs, OID_AUTO, ncsizefactor, CTLFLAG_RW, &ncsizefactor, 0,
211     "Size factor for namecache");
212 static u_int __read_mostly	ncpurgeminvnodes;
213 SYSCTL_UINT(_vfs, OID_AUTO, ncpurgeminvnodes, CTLFLAG_RW, &ncpurgeminvnodes, 0,
214     "Number of vnodes below which purgevfs ignores the request");
215 static u_int __read_mostly	ncneghitsrequeue = 8;
216 SYSCTL_UINT(_vfs, OID_AUTO, ncneghitsrequeue, CTLFLAG_RW, &ncneghitsrequeue, 0,
217     "Number of hits to requeue a negative entry in the LRU list");
218 
219 struct nchstats	nchstats;		/* cache effectiveness statistics */
220 
221 static struct mtx       ncneg_shrink_lock;
222 static int	shrink_list_turn;
223 
224 struct neglist {
225 	struct mtx		nl_lock;
226 	TAILQ_HEAD(, namecache) nl_list;
227 } __aligned(CACHE_LINE_SIZE);
228 
229 static struct neglist __read_mostly	*neglists;
230 static struct neglist ncneg_hot;
231 
232 #define	numneglists (ncneghash + 1)
233 static u_int __read_mostly	ncneghash;
234 static inline struct neglist *
235 NCP2NEGLIST(struct namecache *ncp)
236 {
237 
238 	return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
239 }
240 
241 #define	numbucketlocks (ncbuckethash + 1)
242 static u_int __read_mostly  ncbuckethash;
243 static struct rwlock_padalign __read_mostly  *bucketlocks;
244 #define	HASH2BUCKETLOCK(hash) \
245 	((struct rwlock *)(&bucketlocks[((hash) & ncbuckethash)]))
246 
247 #define	numvnodelocks (ncvnodehash + 1)
248 static u_int __read_mostly  ncvnodehash;
249 static struct mtx __read_mostly *vnodelocks;
250 static inline struct mtx *
251 VP2VNODELOCK(struct vnode *vp)
252 {
253 
254 	return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
255 }
256 
257 /*
258  * UMA zones for the VFS cache.
259  *
260  * The small cache is used for entries with short names, which are the
261  * most common.  The large cache is used for entries which are too big to
262  * fit in the small cache.
263  */
264 static uma_zone_t __read_mostly cache_zone_small;
265 static uma_zone_t __read_mostly cache_zone_small_ts;
266 static uma_zone_t __read_mostly cache_zone_large;
267 static uma_zone_t __read_mostly cache_zone_large_ts;
268 
269 #define	CACHE_PATH_CUTOFF	35
270 
271 static struct namecache *
272 cache_alloc(int len, int ts)
273 {
274 	struct namecache_ts *ncp_ts;
275 	struct namecache *ncp;
276 
277 	if (__predict_false(ts)) {
278 		if (len <= CACHE_PATH_CUTOFF)
279 			ncp_ts = uma_zalloc(cache_zone_small_ts, M_WAITOK);
280 		else
281 			ncp_ts = uma_zalloc(cache_zone_large_ts, M_WAITOK);
282 		ncp = &ncp_ts->nc_nc;
283 	} else {
284 		if (len <= CACHE_PATH_CUTOFF)
285 			ncp = uma_zalloc(cache_zone_small, M_WAITOK);
286 		else
287 			ncp = uma_zalloc(cache_zone_large, M_WAITOK);
288 	}
289 	return (ncp);
290 }
291 
292 static void
293 cache_free(struct namecache *ncp)
294 {
295 	struct namecache_ts *ncp_ts;
296 
297 	if (ncp == NULL)
298 		return;
299 	if ((ncp->nc_flag & NCF_DVDROP) != 0)
300 		vdrop(ncp->nc_dvp);
301 	if (__predict_false(ncp->nc_flag & NCF_TS)) {
302 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
303 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
304 			uma_zfree(cache_zone_small_ts, ncp_ts);
305 		else
306 			uma_zfree(cache_zone_large_ts, ncp_ts);
307 	} else {
308 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
309 			uma_zfree(cache_zone_small, ncp);
310 		else
311 			uma_zfree(cache_zone_large, ncp);
312 	}
313 }
314 
315 static void
316 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
317 {
318 	struct namecache_ts *ncp_ts;
319 
320 	KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
321 	    (tsp == NULL && ticksp == NULL),
322 	    ("No NCF_TS"));
323 
324 	if (tsp == NULL && ticksp == NULL)
325 		return;
326 
327 	ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
328 	if (tsp != NULL)
329 		*tsp = ncp_ts->nc_time;
330 	if (ticksp != NULL)
331 		*ticksp = ncp_ts->nc_ticks;
332 }
333 
334 static int __read_mostly	doingcache = 1;	/* 1 => enable the cache */
335 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
336     "VFS namecache enabled");
337 
338 /* Export size information to userland */
339 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
340     sizeof(struct namecache), "sizeof(struct namecache)");
341 
342 /*
343  * The new name cache statistics
344  */
345 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0,
346     "Name cache statistics");
347 #define STATNODE_ULONG(name, descr)	\
348 	SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, descr);
349 #define STATNODE_COUNTER(name, descr)	\
350 	static counter_u64_t __read_mostly name; \
351 	SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, descr);
352 STATNODE_ULONG(numneg, "Number of negative cache entries");
353 STATNODE_ULONG(numcache, "Number of cache entries");
354 STATNODE_COUNTER(numcalls, "Number of cache lookups");
355 STATNODE_COUNTER(dothits, "Number of '.' hits");
356 STATNODE_COUNTER(dotdothits, "Number of '..' hits");
357 STATNODE_COUNTER(numchecks, "Number of checks in lookup");
358 STATNODE_COUNTER(nummiss, "Number of cache misses");
359 STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache");
360 STATNODE_COUNTER(numposzaps,
361     "Number of cache hits (positive) we do not want to cache");
362 STATNODE_COUNTER(numposhits, "Number of cache hits (positive)");
363 STATNODE_COUNTER(numnegzaps,
364     "Number of cache hits (negative) we do not want to cache");
365 STATNODE_COUNTER(numneghits, "Number of cache hits (negative)");
366 /* These count for kern___getcwd(), too. */
367 STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls");
368 STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
369 STATNODE_COUNTER(numfullpathfail2,
370     "Number of fullpath search errors (VOP_VPTOCNP failures)");
371 STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
372 STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls");
373 static long zap_and_exit_bucket_fail; STATNODE_ULONG(zap_and_exit_bucket_fail,
374     "Number of times zap_and_exit failed to lock");
375 static long cache_lock_vnodes_cel_3_failures;
376 STATNODE_ULONG(cache_lock_vnodes_cel_3_failures,
377     "Number of times 3-way vnode locking failed");
378 
379 static void cache_zap_locked(struct namecache *ncp, bool neg_locked);
380 static int vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
381     char *buf, char **retbuf, u_int buflen);
382 
383 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
384 
385 static int cache_yield;
386 SYSCTL_INT(_vfs_cache, OID_AUTO, yield, CTLFLAG_RD, &cache_yield, 0,
387     "Number of times cache called yield");
388 
389 static void
390 cache_maybe_yield(void)
391 {
392 
393 	if (should_yield()) {
394 		cache_yield++;
395 		kern_yield(PRI_USER);
396 	}
397 }
398 
399 static inline void
400 cache_assert_vlp_locked(struct mtx *vlp)
401 {
402 
403 	if (vlp != NULL)
404 		mtx_assert(vlp, MA_OWNED);
405 }
406 
407 static inline void
408 cache_assert_vnode_locked(struct vnode *vp)
409 {
410 	struct mtx *vlp;
411 
412 	vlp = VP2VNODELOCK(vp);
413 	cache_assert_vlp_locked(vlp);
414 }
415 
416 static uint32_t
417 cache_get_hash(char *name, u_char len, struct vnode *dvp)
418 {
419 	uint32_t hash;
420 
421 	hash = fnv_32_buf(name, len, FNV1_32_INIT);
422 	hash = fnv_32_buf(&dvp, sizeof(dvp), hash);
423 	return (hash);
424 }
425 
426 static inline struct rwlock *
427 NCP2BUCKETLOCK(struct namecache *ncp)
428 {
429 	uint32_t hash;
430 
431 	hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
432 	return (HASH2BUCKETLOCK(hash));
433 }
434 
435 #ifdef INVARIANTS
436 static void
437 cache_assert_bucket_locked(struct namecache *ncp, int mode)
438 {
439 	struct rwlock *blp;
440 
441 	blp = NCP2BUCKETLOCK(ncp);
442 	rw_assert(blp, mode);
443 }
444 #else
445 #define cache_assert_bucket_locked(x, y) do { } while (0)
446 #endif
447 
448 #define cache_sort(x, y)	_cache_sort((void **)(x), (void **)(y))
449 static void
450 _cache_sort(void **p1, void **p2)
451 {
452 	void *tmp;
453 
454 	if (*p1 > *p2) {
455 		tmp = *p2;
456 		*p2 = *p1;
457 		*p1 = tmp;
458 	}
459 }
460 
461 static void
462 cache_lock_all_buckets(void)
463 {
464 	u_int i;
465 
466 	for (i = 0; i < numbucketlocks; i++)
467 		rw_wlock(&bucketlocks[i]);
468 }
469 
470 static void
471 cache_unlock_all_buckets(void)
472 {
473 	u_int i;
474 
475 	for (i = 0; i < numbucketlocks; i++)
476 		rw_wunlock(&bucketlocks[i]);
477 }
478 
479 static void
480 cache_lock_all_vnodes(void)
481 {
482 	u_int i;
483 
484 	for (i = 0; i < numvnodelocks; i++)
485 		mtx_lock(&vnodelocks[i]);
486 }
487 
488 static void
489 cache_unlock_all_vnodes(void)
490 {
491 	u_int i;
492 
493 	for (i = 0; i < numvnodelocks; i++)
494 		mtx_unlock(&vnodelocks[i]);
495 }
496 
497 static int
498 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
499 {
500 
501 	cache_sort(&vlp1, &vlp2);
502 	MPASS(vlp2 != NULL);
503 
504 	if (vlp1 != NULL) {
505 		if (!mtx_trylock(vlp1))
506 			return (EAGAIN);
507 	}
508 	if (!mtx_trylock(vlp2)) {
509 		if (vlp1 != NULL)
510 			mtx_unlock(vlp1);
511 		return (EAGAIN);
512 	}
513 
514 	return (0);
515 }
516 
517 static void
518 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
519 {
520 
521 	MPASS(vlp1 != NULL || vlp2 != NULL);
522 
523 	if (vlp1 != NULL)
524 		mtx_unlock(vlp1);
525 	if (vlp2 != NULL)
526 		mtx_unlock(vlp2);
527 }
528 
529 static int
530 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
531 {
532 	struct nchstats snap;
533 
534 	if (req->oldptr == NULL)
535 		return (SYSCTL_OUT(req, 0, sizeof(snap)));
536 
537 	snap = nchstats;
538 	snap.ncs_goodhits = counter_u64_fetch(numposhits);
539 	snap.ncs_neghits = counter_u64_fetch(numneghits);
540 	snap.ncs_badhits = counter_u64_fetch(numposzaps) +
541 	    counter_u64_fetch(numnegzaps);
542 	snap.ncs_miss = counter_u64_fetch(nummisszap) +
543 	    counter_u64_fetch(nummiss);
544 
545 	return (SYSCTL_OUT(req, &snap, sizeof(snap)));
546 }
547 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
548     CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
549     "VFS cache effectiveness statistics");
550 
551 #ifdef DIAGNOSTIC
552 /*
553  * Grab an atomic snapshot of the name cache hash chain lengths
554  */
555 static SYSCTL_NODE(_debug, OID_AUTO, hashstat, CTLFLAG_RW, NULL,
556     "hash table stats");
557 
558 static int
559 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
560 {
561 	struct nchashhead *ncpp;
562 	struct namecache *ncp;
563 	int i, error, n_nchash, *cntbuf;
564 
565 retry:
566 	n_nchash = nchash + 1;	/* nchash is max index, not count */
567 	if (req->oldptr == NULL)
568 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
569 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
570 	cache_lock_all_buckets();
571 	if (n_nchash != nchash + 1) {
572 		cache_unlock_all_buckets();
573 		free(cntbuf, M_TEMP);
574 		goto retry;
575 	}
576 	/* Scan hash tables counting entries */
577 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
578 		LIST_FOREACH(ncp, ncpp, nc_hash)
579 			cntbuf[i]++;
580 	cache_unlock_all_buckets();
581 	for (error = 0, i = 0; i < n_nchash; i++)
582 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
583 			break;
584 	free(cntbuf, M_TEMP);
585 	return (error);
586 }
587 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
588     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
589     "nchash chain lengths");
590 
591 static int
592 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
593 {
594 	int error;
595 	struct nchashhead *ncpp;
596 	struct namecache *ncp;
597 	int n_nchash;
598 	int count, maxlength, used, pct;
599 
600 	if (!req->oldptr)
601 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
602 
603 	cache_lock_all_buckets();
604 	n_nchash = nchash + 1;	/* nchash is max index, not count */
605 	used = 0;
606 	maxlength = 0;
607 
608 	/* Scan hash tables for applicable entries */
609 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
610 		count = 0;
611 		LIST_FOREACH(ncp, ncpp, nc_hash) {
612 			count++;
613 		}
614 		if (count)
615 			used++;
616 		if (maxlength < count)
617 			maxlength = count;
618 	}
619 	n_nchash = nchash + 1;
620 	cache_unlock_all_buckets();
621 	pct = (used * 100) / (n_nchash / 100);
622 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
623 	if (error)
624 		return (error);
625 	error = SYSCTL_OUT(req, &used, sizeof(used));
626 	if (error)
627 		return (error);
628 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
629 	if (error)
630 		return (error);
631 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
632 	if (error)
633 		return (error);
634 	return (0);
635 }
636 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
637     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
638     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
639 #endif
640 
641 /*
642  * Negative entries management
643  *
644  * A variation of LRU scheme is used. New entries are hashed into one of
645  * numneglists cold lists. Entries get promoted to the hot list on first hit.
646  * Partial LRU for the hot list is maintained by requeueing them every
647  * ncneghitsrequeue hits.
648  *
649  * The shrinker will demote hot list head and evict from the cold list in a
650  * round-robin manner.
651  */
652 static void
653 cache_negative_hit(struct namecache *ncp)
654 {
655 	struct neglist *neglist;
656 	u_int hits;
657 
658 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
659 	hits = atomic_fetchadd_int(&ncp->nc_neghits, 1);
660 	if (ncp->nc_flag & NCF_HOTNEGATIVE) {
661 		if ((hits % ncneghitsrequeue) != 0)
662 			return;
663 		mtx_lock(&ncneg_hot.nl_lock);
664 		if (ncp->nc_flag & NCF_HOTNEGATIVE) {
665 			TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
666 			TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
667 			mtx_unlock(&ncneg_hot.nl_lock);
668 			return;
669 		}
670 		/*
671 		 * The shrinker cleared the flag and removed the entry from
672 		 * the hot list. Put it back.
673 		 */
674 	} else {
675 		mtx_lock(&ncneg_hot.nl_lock);
676 	}
677 	neglist = NCP2NEGLIST(ncp);
678 	mtx_lock(&neglist->nl_lock);
679 	if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
680 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
681 		TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
682 		ncp->nc_flag |= NCF_HOTNEGATIVE;
683 	}
684 	mtx_unlock(&neglist->nl_lock);
685 	mtx_unlock(&ncneg_hot.nl_lock);
686 }
687 
688 static void
689 cache_negative_insert(struct namecache *ncp, bool neg_locked)
690 {
691 	struct neglist *neglist;
692 
693 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
694 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
695 	neglist = NCP2NEGLIST(ncp);
696 	if (!neg_locked) {
697 		mtx_lock(&neglist->nl_lock);
698 	} else {
699 		mtx_assert(&neglist->nl_lock, MA_OWNED);
700 	}
701 	TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
702 	if (!neg_locked)
703 		mtx_unlock(&neglist->nl_lock);
704 	atomic_add_rel_long(&numneg, 1);
705 }
706 
707 static void
708 cache_negative_remove(struct namecache *ncp, bool neg_locked)
709 {
710 	struct neglist *neglist;
711 	bool hot_locked = false;
712 	bool list_locked = false;
713 
714 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
715 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
716 	neglist = NCP2NEGLIST(ncp);
717 	if (!neg_locked) {
718 		if (ncp->nc_flag & NCF_HOTNEGATIVE) {
719 			hot_locked = true;
720 			mtx_lock(&ncneg_hot.nl_lock);
721 			if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
722 				list_locked = true;
723 				mtx_lock(&neglist->nl_lock);
724 			}
725 		} else {
726 			list_locked = true;
727 			mtx_lock(&neglist->nl_lock);
728 		}
729 	}
730 	if (ncp->nc_flag & NCF_HOTNEGATIVE) {
731 		mtx_assert(&ncneg_hot.nl_lock, MA_OWNED);
732 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
733 	} else {
734 		mtx_assert(&neglist->nl_lock, MA_OWNED);
735 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
736 	}
737 	if (list_locked)
738 		mtx_unlock(&neglist->nl_lock);
739 	if (hot_locked)
740 		mtx_unlock(&ncneg_hot.nl_lock);
741 	atomic_subtract_rel_long(&numneg, 1);
742 }
743 
744 static void
745 cache_negative_shrink_select(int start, struct namecache **ncpp,
746     struct neglist **neglistpp)
747 {
748 	struct neglist *neglist;
749 	struct namecache *ncp;
750 	int i;
751 
752 	*ncpp = ncp = NULL;
753 
754 	for (i = start; i < numneglists; i++) {
755 		neglist = &neglists[i];
756 		if (TAILQ_FIRST(&neglist->nl_list) == NULL)
757 			continue;
758 		mtx_lock(&neglist->nl_lock);
759 		ncp = TAILQ_FIRST(&neglist->nl_list);
760 		if (ncp != NULL)
761 			break;
762 		mtx_unlock(&neglist->nl_lock);
763 	}
764 
765 	*neglistpp = neglist;
766 	*ncpp = ncp;
767 }
768 
769 static void
770 cache_negative_zap_one(void)
771 {
772 	struct namecache *ncp, *ncp2;
773 	struct neglist *neglist;
774 	struct mtx *dvlp;
775 	struct rwlock *blp;
776 
777 	if (!mtx_trylock(&ncneg_shrink_lock))
778 		return;
779 
780 	mtx_lock(&ncneg_hot.nl_lock);
781 	ncp = TAILQ_FIRST(&ncneg_hot.nl_list);
782 	if (ncp != NULL) {
783 		neglist = NCP2NEGLIST(ncp);
784 		mtx_lock(&neglist->nl_lock);
785 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
786 		TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
787 		ncp->nc_flag &= ~NCF_HOTNEGATIVE;
788 		mtx_unlock(&neglist->nl_lock);
789 	}
790 
791 	cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
792 	shrink_list_turn++;
793 	if (shrink_list_turn == numneglists)
794 		shrink_list_turn = 0;
795 	if (ncp == NULL && shrink_list_turn == 0)
796 		cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
797 	if (ncp == NULL) {
798 		mtx_unlock(&ncneg_hot.nl_lock);
799 		goto out;
800 	}
801 
802 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
803 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
804 	blp = NCP2BUCKETLOCK(ncp);
805 	mtx_unlock(&neglist->nl_lock);
806 	mtx_unlock(&ncneg_hot.nl_lock);
807 	mtx_lock(dvlp);
808 	rw_wlock(blp);
809 	mtx_lock(&neglist->nl_lock);
810 	ncp2 = TAILQ_FIRST(&neglist->nl_list);
811 	if (ncp != ncp2 || dvlp != VP2VNODELOCK(ncp2->nc_dvp) ||
812 	    blp != NCP2BUCKETLOCK(ncp2) || !(ncp2->nc_flag & NCF_NEGATIVE)) {
813 		ncp = NULL;
814 		goto out_unlock_all;
815 	}
816 	SDT_PROBE3(vfs, namecache, shrink_negative, done, ncp->nc_dvp,
817 	    ncp->nc_name, ncp->nc_neghits);
818 
819 	cache_zap_locked(ncp, true);
820 out_unlock_all:
821 	mtx_unlock(&neglist->nl_lock);
822 	rw_wunlock(blp);
823 	mtx_unlock(dvlp);
824 out:
825 	mtx_unlock(&ncneg_shrink_lock);
826 	cache_free(ncp);
827 }
828 
829 /*
830  * cache_zap_locked():
831  *
832  *   Removes a namecache entry from cache, whether it contains an actual
833  *   pointer to a vnode or if it is just a negative cache entry.
834  */
835 static void
836 cache_zap_locked(struct namecache *ncp, bool neg_locked)
837 {
838 
839 	if (!(ncp->nc_flag & NCF_NEGATIVE))
840 		cache_assert_vnode_locked(ncp->nc_vp);
841 	cache_assert_vnode_locked(ncp->nc_dvp);
842 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
843 
844 	CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp,
845 	    (ncp->nc_flag & NCF_NEGATIVE) ? NULL : ncp->nc_vp);
846 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
847 		SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
848 		    ncp->nc_name, ncp->nc_vp);
849 	} else {
850 		SDT_PROBE3(vfs, namecache, zap_negative, done, ncp->nc_dvp,
851 		    ncp->nc_name, ncp->nc_neghits);
852 	}
853 	LIST_REMOVE(ncp, nc_hash);
854 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
855 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
856 		if (ncp == ncp->nc_vp->v_cache_dd)
857 			ncp->nc_vp->v_cache_dd = NULL;
858 	} else {
859 		cache_negative_remove(ncp, neg_locked);
860 	}
861 	if (ncp->nc_flag & NCF_ISDOTDOT) {
862 		if (ncp == ncp->nc_dvp->v_cache_dd)
863 			ncp->nc_dvp->v_cache_dd = NULL;
864 	} else {
865 		LIST_REMOVE(ncp, nc_src);
866 		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
867 			ncp->nc_flag |= NCF_DVDROP;
868 			atomic_subtract_rel_long(&numcachehv, 1);
869 		}
870 	}
871 	atomic_subtract_rel_long(&numcache, 1);
872 }
873 
874 static void
875 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
876 {
877 	struct rwlock *blp;
878 
879 	MPASS(ncp->nc_dvp == vp);
880 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
881 	cache_assert_vnode_locked(vp);
882 
883 	blp = NCP2BUCKETLOCK(ncp);
884 	rw_wlock(blp);
885 	cache_zap_locked(ncp, false);
886 	rw_wunlock(blp);
887 }
888 
889 static bool
890 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
891     struct mtx **vlpp)
892 {
893 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
894 	struct rwlock *blp;
895 
896 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
897 	cache_assert_vnode_locked(vp);
898 
899 	if (ncp->nc_flag & NCF_NEGATIVE) {
900 		if (*vlpp != NULL) {
901 			mtx_unlock(*vlpp);
902 			*vlpp = NULL;
903 		}
904 		cache_zap_negative_locked_vnode_kl(ncp, vp);
905 		return (true);
906 	}
907 
908 	pvlp = VP2VNODELOCK(vp);
909 	blp = NCP2BUCKETLOCK(ncp);
910 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
911 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
912 
913 	if (*vlpp == vlp1 || *vlpp == vlp2) {
914 		to_unlock = *vlpp;
915 		*vlpp = NULL;
916 	} else {
917 		if (*vlpp != NULL) {
918 			mtx_unlock(*vlpp);
919 			*vlpp = NULL;
920 		}
921 		cache_sort(&vlp1, &vlp2);
922 		if (vlp1 == pvlp) {
923 			mtx_lock(vlp2);
924 			to_unlock = vlp2;
925 		} else {
926 			if (!mtx_trylock(vlp1))
927 				goto out_relock;
928 			to_unlock = vlp1;
929 		}
930 	}
931 	rw_wlock(blp);
932 	cache_zap_locked(ncp, false);
933 	rw_wunlock(blp);
934 	if (to_unlock != NULL)
935 		mtx_unlock(to_unlock);
936 	return (true);
937 
938 out_relock:
939 	mtx_unlock(vlp2);
940 	mtx_lock(vlp1);
941 	mtx_lock(vlp2);
942 	MPASS(*vlpp == NULL);
943 	*vlpp = vlp1;
944 	return (false);
945 }
946 
947 static int
948 cache_zap_locked_vnode(struct namecache *ncp, struct vnode *vp)
949 {
950 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
951 	struct rwlock *blp;
952 	int error = 0;
953 
954 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
955 	cache_assert_vnode_locked(vp);
956 
957 	pvlp = VP2VNODELOCK(vp);
958 	if (ncp->nc_flag & NCF_NEGATIVE) {
959 		cache_zap_negative_locked_vnode_kl(ncp, vp);
960 		goto out;
961 	}
962 
963 	blp = NCP2BUCKETLOCK(ncp);
964 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
965 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
966 	cache_sort(&vlp1, &vlp2);
967 	if (vlp1 == pvlp) {
968 		mtx_lock(vlp2);
969 		to_unlock = vlp2;
970 	} else {
971 		if (!mtx_trylock(vlp1)) {
972 			error = EAGAIN;
973 			goto out;
974 		}
975 		to_unlock = vlp1;
976 	}
977 	rw_wlock(blp);
978 	cache_zap_locked(ncp, false);
979 	rw_wunlock(blp);
980 	mtx_unlock(to_unlock);
981 out:
982 	mtx_unlock(pvlp);
983 	return (error);
984 }
985 
986 static int
987 cache_zap_rlocked_bucket(struct namecache *ncp, struct rwlock *blp)
988 {
989 	struct mtx *dvlp, *vlp;
990 
991 	cache_assert_bucket_locked(ncp, RA_RLOCKED);
992 
993 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
994 	vlp = NULL;
995 	if (!(ncp->nc_flag & NCF_NEGATIVE))
996 		vlp = VP2VNODELOCK(ncp->nc_vp);
997 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
998 		rw_runlock(blp);
999 		rw_wlock(blp);
1000 		cache_zap_locked(ncp, false);
1001 		rw_wunlock(blp);
1002 		cache_unlock_vnodes(dvlp, vlp);
1003 		return (0);
1004 	}
1005 
1006 	rw_runlock(blp);
1007 	return (EAGAIN);
1008 }
1009 
1010 static int
1011 cache_zap_wlocked_bucket_kl(struct namecache *ncp, struct rwlock *blp,
1012     struct mtx **vlpp1, struct mtx **vlpp2)
1013 {
1014 	struct mtx *dvlp, *vlp;
1015 
1016 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1017 
1018 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1019 	vlp = NULL;
1020 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1021 		vlp = VP2VNODELOCK(ncp->nc_vp);
1022 	cache_sort(&dvlp, &vlp);
1023 
1024 	if (*vlpp1 == dvlp && *vlpp2 == vlp) {
1025 		cache_zap_locked(ncp, false);
1026 		cache_unlock_vnodes(dvlp, vlp);
1027 		*vlpp1 = NULL;
1028 		*vlpp2 = NULL;
1029 		return (0);
1030 	}
1031 
1032 	if (*vlpp1 != NULL)
1033 		mtx_unlock(*vlpp1);
1034 	if (*vlpp2 != NULL)
1035 		mtx_unlock(*vlpp2);
1036 	*vlpp1 = NULL;
1037 	*vlpp2 = NULL;
1038 
1039 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1040 		cache_zap_locked(ncp, false);
1041 		cache_unlock_vnodes(dvlp, vlp);
1042 		return (0);
1043 	}
1044 
1045 	rw_wunlock(blp);
1046 	*vlpp1 = dvlp;
1047 	*vlpp2 = vlp;
1048 	if (*vlpp1 != NULL)
1049 		mtx_lock(*vlpp1);
1050 	mtx_lock(*vlpp2);
1051 	rw_wlock(blp);
1052 	return (EAGAIN);
1053 }
1054 
1055 static void
1056 cache_lookup_unlock(struct rwlock *blp, struct mtx *vlp)
1057 {
1058 
1059 	if (blp != NULL) {
1060 		rw_runlock(blp);
1061 	} else {
1062 		mtx_unlock(vlp);
1063 	}
1064 }
1065 
1066 static int __noinline
1067 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1068     struct timespec *tsp, int *ticksp)
1069 {
1070 	int ltype;
1071 
1072 	*vpp = dvp;
1073 	CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .",
1074 			dvp, cnp->cn_nameptr);
1075 	counter_u64_add(dothits, 1);
1076 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1077 	if (tsp != NULL)
1078 		timespecclear(tsp);
1079 	if (ticksp != NULL)
1080 		*ticksp = ticks;
1081 	vrefact(*vpp);
1082 	/*
1083 	 * When we lookup "." we still can be asked to lock it
1084 	 * differently...
1085 	 */
1086 	ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1087 	if (ltype != VOP_ISLOCKED(*vpp)) {
1088 		if (ltype == LK_EXCLUSIVE) {
1089 			vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1090 			if ((*vpp)->v_iflag & VI_DOOMED) {
1091 				/* forced unmount */
1092 				vrele(*vpp);
1093 				*vpp = NULL;
1094 				return (ENOENT);
1095 			}
1096 		} else
1097 			vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1098 	}
1099 	return (-1);
1100 }
1101 
1102 /*
1103  * Lookup an entry in the cache
1104  *
1105  * Lookup is called with dvp pointing to the directory to search,
1106  * cnp pointing to the name of the entry being sought. If the lookup
1107  * succeeds, the vnode is returned in *vpp, and a status of -1 is
1108  * returned. If the lookup determines that the name does not exist
1109  * (negative caching), a status of ENOENT is returned. If the lookup
1110  * fails, a status of zero is returned.  If the directory vnode is
1111  * recycled out from under us due to a forced unmount, a status of
1112  * ENOENT is returned.
1113  *
1114  * vpp is locked and ref'd on return.  If we're looking up DOTDOT, dvp is
1115  * unlocked.  If we're looking up . an extra ref is taken, but the lock is
1116  * not recursively acquired.
1117  */
1118 
1119 static __noinline int
1120 cache_lookup_nomakeentry(struct vnode *dvp, struct vnode **vpp,
1121     struct componentname *cnp, struct timespec *tsp, int *ticksp)
1122 {
1123 	struct namecache *ncp;
1124 	struct rwlock *blp;
1125 	struct mtx *dvlp, *dvlp2;
1126 	uint32_t hash;
1127 	int error;
1128 
1129 	if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1130 		counter_u64_add(dotdothits, 1);
1131 		dvlp = VP2VNODELOCK(dvp);
1132 		dvlp2 = NULL;
1133 		mtx_lock(dvlp);
1134 retry_dotdot:
1135 		ncp = dvp->v_cache_dd;
1136 		if (ncp == NULL) {
1137 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1138 			    "..", NULL);
1139 			mtx_unlock(dvlp);
1140 			return (0);
1141 		}
1142 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1143 			if (ncp->nc_dvp != dvp)
1144 				panic("dvp %p v_cache_dd %p\n", dvp, ncp);
1145 			if (!cache_zap_locked_vnode_kl2(ncp,
1146 			    dvp, &dvlp2))
1147 				goto retry_dotdot;
1148 			MPASS(dvp->v_cache_dd == NULL);
1149 			mtx_unlock(dvlp);
1150 			if (dvlp2 != NULL)
1151 				mtx_unlock(dvlp2);
1152 			cache_free(ncp);
1153 		} else {
1154 			dvp->v_cache_dd = NULL;
1155 			mtx_unlock(dvlp);
1156 			if (dvlp2 != NULL)
1157 				mtx_unlock(dvlp2);
1158 		}
1159 		return (0);
1160 	}
1161 
1162 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1163 	blp = HASH2BUCKETLOCK(hash);
1164 retry:
1165 	rw_rlock(blp);
1166 
1167 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1168 		counter_u64_add(numchecks, 1);
1169 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1170 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1171 			break;
1172 	}
1173 
1174 	/* We failed to find an entry */
1175 	if (ncp == NULL) {
1176 		rw_runlock(blp);
1177 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1178 		    NULL);
1179 		counter_u64_add(nummisszap, 1);
1180 		return (0);
1181 	}
1182 
1183 	counter_u64_add(numposzaps, 1);
1184 
1185 	error = cache_zap_rlocked_bucket(ncp, blp);
1186 	if (error != 0) {
1187 		zap_and_exit_bucket_fail++;
1188 		cache_maybe_yield();
1189 		goto retry;
1190 	}
1191 	cache_free(ncp);
1192 	return (0);
1193 }
1194 
1195 int
1196 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1197     struct timespec *tsp, int *ticksp)
1198 {
1199 	struct namecache_ts *ncp_ts;
1200 	struct namecache *ncp;
1201 	struct rwlock *blp;
1202 	struct mtx *dvlp, *dvlp2;
1203 	uint32_t hash;
1204 	int error, ltype;
1205 
1206 	if (__predict_false(!doingcache)) {
1207 		cnp->cn_flags &= ~MAKEENTRY;
1208 		return (0);
1209 	}
1210 
1211 	counter_u64_add(numcalls, 1);
1212 
1213 	if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.'))
1214 		return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1215 
1216 	if ((cnp->cn_flags & MAKEENTRY) == 0)
1217 		return (cache_lookup_nomakeentry(dvp, vpp, cnp, tsp, ticksp));
1218 
1219 retry:
1220 	blp = NULL;
1221 	error = 0;
1222 	if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1223 		counter_u64_add(dotdothits, 1);
1224 		dvlp = VP2VNODELOCK(dvp);
1225 		dvlp2 = NULL;
1226 		mtx_lock(dvlp);
1227 		ncp = dvp->v_cache_dd;
1228 		if (ncp == NULL) {
1229 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1230 			    "..", NULL);
1231 			mtx_unlock(dvlp);
1232 			return (0);
1233 		}
1234 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1235 			if (ncp->nc_flag & NCF_NEGATIVE)
1236 				*vpp = NULL;
1237 			else
1238 				*vpp = ncp->nc_vp;
1239 		} else
1240 			*vpp = ncp->nc_dvp;
1241 		/* Return failure if negative entry was found. */
1242 		if (*vpp == NULL)
1243 			goto negative_success;
1244 		CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
1245 		    dvp, cnp->cn_nameptr, *vpp);
1246 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..",
1247 		    *vpp);
1248 		cache_out_ts(ncp, tsp, ticksp);
1249 		if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1250 		    NCF_DTS && tsp != NULL) {
1251 			ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1252 			*tsp = ncp_ts->nc_dotdottime;
1253 		}
1254 		goto success;
1255 	}
1256 
1257 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1258 	blp = HASH2BUCKETLOCK(hash);
1259 	rw_rlock(blp);
1260 
1261 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1262 		counter_u64_add(numchecks, 1);
1263 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1264 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1265 			break;
1266 	}
1267 
1268 	/* We failed to find an entry */
1269 	if (ncp == NULL) {
1270 		rw_runlock(blp);
1271 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1272 		    NULL);
1273 		counter_u64_add(nummiss, 1);
1274 		return (0);
1275 	}
1276 
1277 	/* We found a "positive" match, return the vnode */
1278 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1279 		counter_u64_add(numposhits, 1);
1280 		*vpp = ncp->nc_vp;
1281 		CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p",
1282 		    dvp, cnp->cn_nameptr, *vpp, ncp);
1283 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name,
1284 		    *vpp);
1285 		cache_out_ts(ncp, tsp, ticksp);
1286 		goto success;
1287 	}
1288 
1289 negative_success:
1290 	/* We found a negative match, and want to create it, so purge */
1291 	if (cnp->cn_nameiop == CREATE) {
1292 		counter_u64_add(numnegzaps, 1);
1293 		goto zap_and_exit;
1294 	}
1295 
1296 	counter_u64_add(numneghits, 1);
1297 	cache_negative_hit(ncp);
1298 	if (ncp->nc_flag & NCF_WHITE)
1299 		cnp->cn_flags |= ISWHITEOUT;
1300 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp,
1301 	    ncp->nc_name);
1302 	cache_out_ts(ncp, tsp, ticksp);
1303 	cache_lookup_unlock(blp, dvlp);
1304 	return (ENOENT);
1305 
1306 success:
1307 	/*
1308 	 * On success we return a locked and ref'd vnode as per the lookup
1309 	 * protocol.
1310 	 */
1311 	MPASS(dvp != *vpp);
1312 	ltype = 0;	/* silence gcc warning */
1313 	if (cnp->cn_flags & ISDOTDOT) {
1314 		ltype = VOP_ISLOCKED(dvp);
1315 		VOP_UNLOCK(dvp, 0);
1316 	}
1317 	vhold(*vpp);
1318 	cache_lookup_unlock(blp, dvlp);
1319 	error = vget(*vpp, cnp->cn_lkflags | LK_VNHELD, cnp->cn_thread);
1320 	if (cnp->cn_flags & ISDOTDOT) {
1321 		vn_lock(dvp, ltype | LK_RETRY);
1322 		if (dvp->v_iflag & VI_DOOMED) {
1323 			if (error == 0)
1324 				vput(*vpp);
1325 			*vpp = NULL;
1326 			return (ENOENT);
1327 		}
1328 	}
1329 	if (error) {
1330 		*vpp = NULL;
1331 		goto retry;
1332 	}
1333 	if ((cnp->cn_flags & ISLASTCN) &&
1334 	    (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) {
1335 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
1336 	}
1337 	return (-1);
1338 
1339 zap_and_exit:
1340 	if (blp != NULL)
1341 		error = cache_zap_rlocked_bucket(ncp, blp);
1342 	else
1343 		error = cache_zap_locked_vnode(ncp, dvp);
1344 	if (error != 0) {
1345 		zap_and_exit_bucket_fail++;
1346 		cache_maybe_yield();
1347 		goto retry;
1348 	}
1349 	cache_free(ncp);
1350 	return (0);
1351 }
1352 
1353 struct celockstate {
1354 	struct mtx *vlp[3];
1355 	struct rwlock *blp[2];
1356 };
1357 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1358 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1359 
1360 static inline void
1361 cache_celockstate_init(struct celockstate *cel)
1362 {
1363 
1364 	bzero(cel, sizeof(*cel));
1365 }
1366 
1367 static void
1368 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1369     struct vnode *dvp)
1370 {
1371 	struct mtx *vlp1, *vlp2;
1372 
1373 	MPASS(cel->vlp[0] == NULL);
1374 	MPASS(cel->vlp[1] == NULL);
1375 	MPASS(cel->vlp[2] == NULL);
1376 
1377 	MPASS(vp != NULL || dvp != NULL);
1378 
1379 	vlp1 = VP2VNODELOCK(vp);
1380 	vlp2 = VP2VNODELOCK(dvp);
1381 	cache_sort(&vlp1, &vlp2);
1382 
1383 	if (vlp1 != NULL) {
1384 		mtx_lock(vlp1);
1385 		cel->vlp[0] = vlp1;
1386 	}
1387 	mtx_lock(vlp2);
1388 	cel->vlp[1] = vlp2;
1389 }
1390 
1391 static void
1392 cache_unlock_vnodes_cel(struct celockstate *cel)
1393 {
1394 
1395 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
1396 
1397 	if (cel->vlp[0] != NULL)
1398 		mtx_unlock(cel->vlp[0]);
1399 	if (cel->vlp[1] != NULL)
1400 		mtx_unlock(cel->vlp[1]);
1401 	if (cel->vlp[2] != NULL)
1402 		mtx_unlock(cel->vlp[2]);
1403 }
1404 
1405 static bool
1406 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
1407 {
1408 	struct mtx *vlp;
1409 	bool ret;
1410 
1411 	cache_assert_vlp_locked(cel->vlp[0]);
1412 	cache_assert_vlp_locked(cel->vlp[1]);
1413 	MPASS(cel->vlp[2] == NULL);
1414 
1415 	MPASS(vp != NULL);
1416 	vlp = VP2VNODELOCK(vp);
1417 
1418 	ret = true;
1419 	if (vlp >= cel->vlp[1]) {
1420 		mtx_lock(vlp);
1421 	} else {
1422 		if (mtx_trylock(vlp))
1423 			goto out;
1424 		cache_lock_vnodes_cel_3_failures++;
1425 		cache_unlock_vnodes_cel(cel);
1426 		if (vlp < cel->vlp[0]) {
1427 			mtx_lock(vlp);
1428 			mtx_lock(cel->vlp[0]);
1429 			mtx_lock(cel->vlp[1]);
1430 		} else {
1431 			if (cel->vlp[0] != NULL)
1432 				mtx_lock(cel->vlp[0]);
1433 			mtx_lock(vlp);
1434 			mtx_lock(cel->vlp[1]);
1435 		}
1436 		ret = false;
1437 	}
1438 out:
1439 	cel->vlp[2] = vlp;
1440 	return (ret);
1441 }
1442 
1443 static void
1444 cache_lock_buckets_cel(struct celockstate *cel, struct rwlock *blp1,
1445     struct rwlock *blp2)
1446 {
1447 
1448 	MPASS(cel->blp[0] == NULL);
1449 	MPASS(cel->blp[1] == NULL);
1450 
1451 	cache_sort(&blp1, &blp2);
1452 
1453 	if (blp1 != NULL) {
1454 		rw_wlock(blp1);
1455 		cel->blp[0] = blp1;
1456 	}
1457 	rw_wlock(blp2);
1458 	cel->blp[1] = blp2;
1459 }
1460 
1461 static void
1462 cache_unlock_buckets_cel(struct celockstate *cel)
1463 {
1464 
1465 	if (cel->blp[0] != NULL)
1466 		rw_wunlock(cel->blp[0]);
1467 	rw_wunlock(cel->blp[1]);
1468 }
1469 
1470 /*
1471  * Lock part of the cache affected by the insertion.
1472  *
1473  * This means vnodelocks for dvp, vp and the relevant bucketlock.
1474  * However, insertion can result in removal of an old entry. In this
1475  * case we have an additional vnode and bucketlock pair to lock. If the
1476  * entry is negative, ncelock is locked instead of the vnode.
1477  *
1478  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
1479  * preserving the locking order (smaller address first).
1480  */
1481 static void
1482 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1483     uint32_t hash)
1484 {
1485 	struct namecache *ncp;
1486 	struct rwlock *blps[2];
1487 
1488 	blps[0] = HASH2BUCKETLOCK(hash);
1489 	for (;;) {
1490 		blps[1] = NULL;
1491 		cache_lock_vnodes_cel(cel, dvp, vp);
1492 		if (vp == NULL || vp->v_type != VDIR)
1493 			break;
1494 		ncp = vp->v_cache_dd;
1495 		if (ncp == NULL)
1496 			break;
1497 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1498 			break;
1499 		MPASS(ncp->nc_dvp == vp);
1500 		blps[1] = NCP2BUCKETLOCK(ncp);
1501 		if (ncp->nc_flag & NCF_NEGATIVE)
1502 			break;
1503 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1504 			break;
1505 		/*
1506 		 * All vnodes got re-locked. Re-validate the state and if
1507 		 * nothing changed we are done. Otherwise restart.
1508 		 */
1509 		if (ncp == vp->v_cache_dd &&
1510 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1511 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1512 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1513 			break;
1514 		cache_unlock_vnodes_cel(cel);
1515 		cel->vlp[0] = NULL;
1516 		cel->vlp[1] = NULL;
1517 		cel->vlp[2] = NULL;
1518 	}
1519 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1520 }
1521 
1522 static void
1523 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1524     uint32_t hash)
1525 {
1526 	struct namecache *ncp;
1527 	struct rwlock *blps[2];
1528 
1529 	blps[0] = HASH2BUCKETLOCK(hash);
1530 	for (;;) {
1531 		blps[1] = NULL;
1532 		cache_lock_vnodes_cel(cel, dvp, vp);
1533 		ncp = dvp->v_cache_dd;
1534 		if (ncp == NULL)
1535 			break;
1536 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1537 			break;
1538 		MPASS(ncp->nc_dvp == dvp);
1539 		blps[1] = NCP2BUCKETLOCK(ncp);
1540 		if (ncp->nc_flag & NCF_NEGATIVE)
1541 			break;
1542 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1543 			break;
1544 		if (ncp == dvp->v_cache_dd &&
1545 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1546 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1547 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1548 			break;
1549 		cache_unlock_vnodes_cel(cel);
1550 		cel->vlp[0] = NULL;
1551 		cel->vlp[1] = NULL;
1552 		cel->vlp[2] = NULL;
1553 	}
1554 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1555 }
1556 
1557 static void
1558 cache_enter_unlock(struct celockstate *cel)
1559 {
1560 
1561 	cache_unlock_buckets_cel(cel);
1562 	cache_unlock_vnodes_cel(cel);
1563 }
1564 
1565 /*
1566  * Add an entry to the cache.
1567  */
1568 void
1569 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
1570     struct timespec *tsp, struct timespec *dtsp)
1571 {
1572 	struct celockstate cel;
1573 	struct namecache *ncp, *n2, *ndd;
1574 	struct namecache_ts *ncp_ts, *n2_ts;
1575 	struct nchashhead *ncpp;
1576 	struct neglist *neglist;
1577 	uint32_t hash;
1578 	int flag;
1579 	int len;
1580 	bool neg_locked;
1581 
1582 	CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr);
1583 	VNASSERT(vp == NULL || (vp->v_iflag & VI_DOOMED) == 0, vp,
1584 	    ("cache_enter: Adding a doomed vnode"));
1585 	VNASSERT(dvp == NULL || (dvp->v_iflag & VI_DOOMED) == 0, dvp,
1586 	    ("cache_enter: Doomed vnode used as src"));
1587 
1588 	if (__predict_false(!doingcache))
1589 		return;
1590 
1591 	/*
1592 	 * Avoid blowout in namecache entries.
1593 	 */
1594 	if (__predict_false(numcache >= desiredvnodes * ncsizefactor))
1595 		return;
1596 
1597 	cache_celockstate_init(&cel);
1598 	ndd = NULL;
1599 	flag = 0;
1600 	if (cnp->cn_nameptr[0] == '.') {
1601 		if (cnp->cn_namelen == 1)
1602 			return;
1603 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1604 			len = cnp->cn_namelen;
1605 			hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1606 			cache_enter_lock_dd(&cel, dvp, vp, hash);
1607 			/*
1608 			 * If dotdot entry already exists, just retarget it
1609 			 * to new parent vnode, otherwise continue with new
1610 			 * namecache entry allocation.
1611 			 */
1612 			if ((ncp = dvp->v_cache_dd) != NULL &&
1613 			    ncp->nc_flag & NCF_ISDOTDOT) {
1614 				KASSERT(ncp->nc_dvp == dvp,
1615 				    ("wrong isdotdot parent"));
1616 				neg_locked = false;
1617 				if (ncp->nc_flag & NCF_NEGATIVE || vp == NULL) {
1618 					neglist = NCP2NEGLIST(ncp);
1619 					mtx_lock(&ncneg_hot.nl_lock);
1620 					mtx_lock(&neglist->nl_lock);
1621 					neg_locked = true;
1622 				}
1623 				if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1624 					TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst,
1625 					    ncp, nc_dst);
1626 				} else {
1627 					cache_negative_remove(ncp, true);
1628 				}
1629 				if (vp != NULL) {
1630 					TAILQ_INSERT_HEAD(&vp->v_cache_dst,
1631 					    ncp, nc_dst);
1632 					ncp->nc_flag &= ~(NCF_NEGATIVE|NCF_HOTNEGATIVE);
1633 				} else {
1634 					ncp->nc_flag &= ~(NCF_HOTNEGATIVE);
1635 					ncp->nc_flag |= NCF_NEGATIVE;
1636 					cache_negative_insert(ncp, true);
1637 				}
1638 				if (neg_locked) {
1639 					mtx_unlock(&neglist->nl_lock);
1640 					mtx_unlock(&ncneg_hot.nl_lock);
1641 				}
1642 				ncp->nc_vp = vp;
1643 				cache_enter_unlock(&cel);
1644 				return;
1645 			}
1646 			dvp->v_cache_dd = NULL;
1647 			cache_enter_unlock(&cel);
1648 			cache_celockstate_init(&cel);
1649 			SDT_PROBE3(vfs, namecache, enter, done, dvp, "..", vp);
1650 			flag = NCF_ISDOTDOT;
1651 		}
1652 	}
1653 
1654 	/*
1655 	 * Calculate the hash key and setup as much of the new
1656 	 * namecache entry as possible before acquiring the lock.
1657 	 */
1658 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
1659 	ncp->nc_flag = flag;
1660 	ncp->nc_vp = vp;
1661 	if (vp == NULL)
1662 		ncp->nc_flag |= NCF_NEGATIVE;
1663 	ncp->nc_dvp = dvp;
1664 	if (tsp != NULL) {
1665 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1666 		ncp_ts->nc_time = *tsp;
1667 		ncp_ts->nc_ticks = ticks;
1668 		ncp_ts->nc_nc.nc_flag |= NCF_TS;
1669 		if (dtsp != NULL) {
1670 			ncp_ts->nc_dotdottime = *dtsp;
1671 			ncp_ts->nc_nc.nc_flag |= NCF_DTS;
1672 		}
1673 	}
1674 	len = ncp->nc_nlen = cnp->cn_namelen;
1675 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1676 	strlcpy(ncp->nc_name, cnp->cn_nameptr, len + 1);
1677 	cache_enter_lock(&cel, dvp, vp, hash);
1678 
1679 	/*
1680 	 * See if this vnode or negative entry is already in the cache
1681 	 * with this name.  This can happen with concurrent lookups of
1682 	 * the same path name.
1683 	 */
1684 	ncpp = NCHHASH(hash);
1685 	LIST_FOREACH(n2, ncpp, nc_hash) {
1686 		if (n2->nc_dvp == dvp &&
1687 		    n2->nc_nlen == cnp->cn_namelen &&
1688 		    !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
1689 			if (tsp != NULL) {
1690 				KASSERT((n2->nc_flag & NCF_TS) != 0,
1691 				    ("no NCF_TS"));
1692 				n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
1693 				n2_ts->nc_time = ncp_ts->nc_time;
1694 				n2_ts->nc_ticks = ncp_ts->nc_ticks;
1695 				if (dtsp != NULL) {
1696 					n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
1697 					if (ncp->nc_flag & NCF_NEGATIVE)
1698 						mtx_lock(&ncneg_hot.nl_lock);
1699 					n2_ts->nc_nc.nc_flag |= NCF_DTS;
1700 					if (ncp->nc_flag & NCF_NEGATIVE)
1701 						mtx_unlock(&ncneg_hot.nl_lock);
1702 				}
1703 			}
1704 			goto out_unlock_free;
1705 		}
1706 	}
1707 
1708 	if (flag == NCF_ISDOTDOT) {
1709 		/*
1710 		 * See if we are trying to add .. entry, but some other lookup
1711 		 * has populated v_cache_dd pointer already.
1712 		 */
1713 		if (dvp->v_cache_dd != NULL)
1714 			goto out_unlock_free;
1715 		KASSERT(vp == NULL || vp->v_type == VDIR,
1716 		    ("wrong vnode type %p", vp));
1717 		dvp->v_cache_dd = ncp;
1718 	}
1719 
1720 	atomic_add_rel_long(&numcache, 1);
1721 	if (vp != NULL) {
1722 		if (vp->v_type == VDIR) {
1723 			if (flag != NCF_ISDOTDOT) {
1724 				/*
1725 				 * For this case, the cache entry maps both the
1726 				 * directory name in it and the name ".." for the
1727 				 * directory's parent.
1728 				 */
1729 				if ((ndd = vp->v_cache_dd) != NULL) {
1730 					if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
1731 						cache_zap_locked(ndd, false);
1732 					else
1733 						ndd = NULL;
1734 				}
1735 				vp->v_cache_dd = ncp;
1736 			}
1737 		} else {
1738 			vp->v_cache_dd = NULL;
1739 		}
1740 	}
1741 
1742 	if (flag != NCF_ISDOTDOT) {
1743 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1744 			vhold(dvp);
1745 			atomic_add_rel_long(&numcachehv, 1);
1746 		}
1747 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
1748 	}
1749 
1750 	/*
1751 	 * Insert the new namecache entry into the appropriate chain
1752 	 * within the cache entries table.
1753 	 */
1754 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
1755 
1756 	/*
1757 	 * If the entry is "negative", we place it into the
1758 	 * "negative" cache queue, otherwise, we place it into the
1759 	 * destination vnode's cache entries queue.
1760 	 */
1761 	if (vp != NULL) {
1762 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
1763 		SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
1764 		    vp);
1765 	} else {
1766 		if (cnp->cn_flags & ISWHITEOUT)
1767 			ncp->nc_flag |= NCF_WHITE;
1768 		cache_negative_insert(ncp, false);
1769 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
1770 		    ncp->nc_name);
1771 	}
1772 	cache_enter_unlock(&cel);
1773 	if (numneg * ncnegfactor > numcache)
1774 		cache_negative_zap_one();
1775 	cache_free(ndd);
1776 	return;
1777 out_unlock_free:
1778 	cache_enter_unlock(&cel);
1779 	cache_free(ncp);
1780 	return;
1781 }
1782 
1783 static u_int
1784 cache_roundup_2(u_int val)
1785 {
1786 	u_int res;
1787 
1788 	for (res = 1; res <= val; res <<= 1)
1789 		continue;
1790 
1791 	return (res);
1792 }
1793 
1794 /*
1795  * Name cache initialization, from vfs_init() when we are booting
1796  */
1797 static void
1798 nchinit(void *dummy __unused)
1799 {
1800 	u_int i;
1801 
1802 	cache_zone_small = uma_zcreate("S VFS Cache",
1803 	    sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1,
1804 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1805 	    UMA_ZONE_ZINIT);
1806 	cache_zone_small_ts = uma_zcreate("STS VFS Cache",
1807 	    sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1,
1808 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1809 	    UMA_ZONE_ZINIT);
1810 	cache_zone_large = uma_zcreate("L VFS Cache",
1811 	    sizeof(struct namecache) + NAME_MAX + 1,
1812 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1813 	    UMA_ZONE_ZINIT);
1814 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache",
1815 	    sizeof(struct namecache_ts) + NAME_MAX + 1,
1816 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1817 	    UMA_ZONE_ZINIT);
1818 
1819 	nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash);
1820 	ncbuckethash = cache_roundup_2(mp_ncpus * 64) - 1;
1821 	if (ncbuckethash > nchash)
1822 		ncbuckethash = nchash;
1823 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
1824 	    M_WAITOK | M_ZERO);
1825 	for (i = 0; i < numbucketlocks; i++)
1826 		rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK | RW_RECURSE);
1827 	ncvnodehash = cache_roundup_2(mp_ncpus * 64) - 1;
1828 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
1829 	    M_WAITOK | M_ZERO);
1830 	for (i = 0; i < numvnodelocks; i++)
1831 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
1832 	ncpurgeminvnodes = numbucketlocks;
1833 
1834 	ncneghash = 3;
1835 	neglists = malloc(sizeof(*neglists) * numneglists, M_VFSCACHE,
1836 	    M_WAITOK | M_ZERO);
1837 	for (i = 0; i < numneglists; i++) {
1838 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
1839 		TAILQ_INIT(&neglists[i].nl_list);
1840 	}
1841 	mtx_init(&ncneg_hot.nl_lock, "ncneglh", NULL, MTX_DEF);
1842 	TAILQ_INIT(&ncneg_hot.nl_list);
1843 
1844 	mtx_init(&ncneg_shrink_lock, "ncnegs", NULL, MTX_DEF);
1845 
1846 	numcalls = counter_u64_alloc(M_WAITOK);
1847 	dothits = counter_u64_alloc(M_WAITOK);
1848 	dotdothits = counter_u64_alloc(M_WAITOK);
1849 	numchecks = counter_u64_alloc(M_WAITOK);
1850 	nummiss = counter_u64_alloc(M_WAITOK);
1851 	nummisszap = counter_u64_alloc(M_WAITOK);
1852 	numposzaps = counter_u64_alloc(M_WAITOK);
1853 	numposhits = counter_u64_alloc(M_WAITOK);
1854 	numnegzaps = counter_u64_alloc(M_WAITOK);
1855 	numneghits = counter_u64_alloc(M_WAITOK);
1856 	numfullpathcalls = counter_u64_alloc(M_WAITOK);
1857 	numfullpathfail1 = counter_u64_alloc(M_WAITOK);
1858 	numfullpathfail2 = counter_u64_alloc(M_WAITOK);
1859 	numfullpathfail4 = counter_u64_alloc(M_WAITOK);
1860 	numfullpathfound = counter_u64_alloc(M_WAITOK);
1861 }
1862 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
1863 
1864 void
1865 cache_changesize(int newmaxvnodes)
1866 {
1867 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
1868 	u_long new_nchash, old_nchash;
1869 	struct namecache *ncp;
1870 	uint32_t hash;
1871 	int i;
1872 
1873 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
1874 	if (newmaxvnodes < numbucketlocks)
1875 		newmaxvnodes = numbucketlocks;
1876 
1877 	new_nchashtbl = hashinit(newmaxvnodes, M_VFSCACHE, &new_nchash);
1878 	/* If same hash table size, nothing to do */
1879 	if (nchash == new_nchash) {
1880 		free(new_nchashtbl, M_VFSCACHE);
1881 		return;
1882 	}
1883 	/*
1884 	 * Move everything from the old hash table to the new table.
1885 	 * None of the namecache entries in the table can be removed
1886 	 * because to do so, they have to be removed from the hash table.
1887 	 */
1888 	cache_lock_all_vnodes();
1889 	cache_lock_all_buckets();
1890 	old_nchashtbl = nchashtbl;
1891 	old_nchash = nchash;
1892 	nchashtbl = new_nchashtbl;
1893 	nchash = new_nchash;
1894 	for (i = 0; i <= old_nchash; i++) {
1895 		while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) {
1896 			hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
1897 			    ncp->nc_dvp);
1898 			LIST_REMOVE(ncp, nc_hash);
1899 			LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
1900 		}
1901 	}
1902 	cache_unlock_all_buckets();
1903 	cache_unlock_all_vnodes();
1904 	free(old_nchashtbl, M_VFSCACHE);
1905 }
1906 
1907 /*
1908  * Invalidate all entries to a particular vnode.
1909  */
1910 void
1911 cache_purge(struct vnode *vp)
1912 {
1913 	TAILQ_HEAD(, namecache) ncps;
1914 	struct namecache *ncp, *nnp;
1915 	struct mtx *vlp, *vlp2;
1916 
1917 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
1918 	SDT_PROBE1(vfs, namecache, purge, done, vp);
1919 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
1920 	    vp->v_cache_dd == NULL)
1921 		return;
1922 	TAILQ_INIT(&ncps);
1923 	vlp = VP2VNODELOCK(vp);
1924 	vlp2 = NULL;
1925 	mtx_lock(vlp);
1926 retry:
1927 	while (!LIST_EMPTY(&vp->v_cache_src)) {
1928 		ncp = LIST_FIRST(&vp->v_cache_src);
1929 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
1930 			goto retry;
1931 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1932 	}
1933 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
1934 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
1935 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
1936 			goto retry;
1937 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1938 	}
1939 	ncp = vp->v_cache_dd;
1940 	if (ncp != NULL) {
1941 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
1942 		   ("lost dotdot link"));
1943 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
1944 			goto retry;
1945 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1946 	}
1947 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
1948 	mtx_unlock(vlp);
1949 	if (vlp2 != NULL)
1950 		mtx_unlock(vlp2);
1951 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
1952 		cache_free(ncp);
1953 	}
1954 }
1955 
1956 /*
1957  * Invalidate all negative entries for a particular directory vnode.
1958  */
1959 void
1960 cache_purge_negative(struct vnode *vp)
1961 {
1962 	TAILQ_HEAD(, namecache) ncps;
1963 	struct namecache *ncp, *nnp;
1964 	struct mtx *vlp;
1965 
1966 	CTR1(KTR_VFS, "cache_purge_negative(%p)", vp);
1967 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
1968 	TAILQ_INIT(&ncps);
1969 	vlp = VP2VNODELOCK(vp);
1970 	mtx_lock(vlp);
1971 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
1972 		if (!(ncp->nc_flag & NCF_NEGATIVE))
1973 			continue;
1974 		cache_zap_negative_locked_vnode_kl(ncp, vp);
1975 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1976 	}
1977 	mtx_unlock(vlp);
1978 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
1979 		cache_free(ncp);
1980 	}
1981 }
1982 
1983 /*
1984  * Flush all entries referencing a particular filesystem.
1985  */
1986 void
1987 cache_purgevfs(struct mount *mp, bool force)
1988 {
1989 	TAILQ_HEAD(, namecache) ncps;
1990 	struct mtx *vlp1, *vlp2;
1991 	struct rwlock *blp;
1992 	struct nchashhead *bucket;
1993 	struct namecache *ncp, *nnp;
1994 	u_long i, j, n_nchash;
1995 	int error;
1996 
1997 	/* Scan hash tables for applicable entries */
1998 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
1999 	if (!force && mp->mnt_nvnodelistsize <= ncpurgeminvnodes)
2000 		return;
2001 	TAILQ_INIT(&ncps);
2002 	n_nchash = nchash + 1;
2003 	vlp1 = vlp2 = NULL;
2004 	for (i = 0; i < numbucketlocks; i++) {
2005 		blp = (struct rwlock *)&bucketlocks[i];
2006 		rw_wlock(blp);
2007 		for (j = i; j < n_nchash; j += numbucketlocks) {
2008 retry:
2009 			bucket = &nchashtbl[j];
2010 			LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) {
2011 				cache_assert_bucket_locked(ncp, RA_WLOCKED);
2012 				if (ncp->nc_dvp->v_mount != mp)
2013 					continue;
2014 				error = cache_zap_wlocked_bucket_kl(ncp, blp,
2015 				    &vlp1, &vlp2);
2016 				if (error != 0)
2017 					goto retry;
2018 				TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst);
2019 			}
2020 		}
2021 		rw_wunlock(blp);
2022 		if (vlp1 == NULL && vlp2 == NULL)
2023 			cache_maybe_yield();
2024 	}
2025 	if (vlp1 != NULL)
2026 		mtx_unlock(vlp1);
2027 	if (vlp2 != NULL)
2028 		mtx_unlock(vlp2);
2029 
2030 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2031 		cache_free(ncp);
2032 	}
2033 }
2034 
2035 /*
2036  * Perform canonical checks and cache lookup and pass on to filesystem
2037  * through the vop_cachedlookup only if needed.
2038  */
2039 
2040 int
2041 vfs_cache_lookup(struct vop_lookup_args *ap)
2042 {
2043 	struct vnode *dvp;
2044 	int error;
2045 	struct vnode **vpp = ap->a_vpp;
2046 	struct componentname *cnp = ap->a_cnp;
2047 	struct ucred *cred = cnp->cn_cred;
2048 	int flags = cnp->cn_flags;
2049 	struct thread *td = cnp->cn_thread;
2050 
2051 	*vpp = NULL;
2052 	dvp = ap->a_dvp;
2053 
2054 	if (dvp->v_type != VDIR)
2055 		return (ENOTDIR);
2056 
2057 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2058 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2059 		return (EROFS);
2060 
2061 	error = VOP_ACCESS(dvp, VEXEC, cred, td);
2062 	if (error)
2063 		return (error);
2064 
2065 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2066 	if (error == 0)
2067 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2068 	if (error == -1)
2069 		return (0);
2070 	return (error);
2071 }
2072 
2073 /*
2074  * XXX All of these sysctls would probably be more productive dead.
2075  */
2076 static int __read_mostly disablecwd;
2077 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
2078    "Disable the getcwd syscall");
2079 
2080 /* Implementation of the getcwd syscall. */
2081 int
2082 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2083 {
2084 
2085 	return (kern___getcwd(td, uap->buf, UIO_USERSPACE, uap->buflen,
2086 	    MAXPATHLEN));
2087 }
2088 
2089 int
2090 kern___getcwd(struct thread *td, char *buf, enum uio_seg bufseg, size_t buflen,
2091     size_t path_max)
2092 {
2093 	char *bp, *tmpbuf;
2094 	struct filedesc *fdp;
2095 	struct vnode *cdir, *rdir;
2096 	int error;
2097 
2098 	if (__predict_false(disablecwd))
2099 		return (ENODEV);
2100 	if (__predict_false(buflen < 2))
2101 		return (EINVAL);
2102 	if (buflen > path_max)
2103 		buflen = path_max;
2104 
2105 	tmpbuf = malloc(buflen, M_TEMP, M_WAITOK);
2106 	fdp = td->td_proc->p_fd;
2107 	FILEDESC_SLOCK(fdp);
2108 	cdir = fdp->fd_cdir;
2109 	vrefact(cdir);
2110 	rdir = fdp->fd_rdir;
2111 	vrefact(rdir);
2112 	FILEDESC_SUNLOCK(fdp);
2113 	error = vn_fullpath1(td, cdir, rdir, tmpbuf, &bp, buflen);
2114 	vrele(rdir);
2115 	vrele(cdir);
2116 
2117 	if (!error) {
2118 		if (bufseg == UIO_SYSSPACE)
2119 			bcopy(bp, buf, strlen(bp) + 1);
2120 		else
2121 			error = copyout(bp, buf, strlen(bp) + 1);
2122 #ifdef KTRACE
2123 	if (KTRPOINT(curthread, KTR_NAMEI))
2124 		ktrnamei(bp);
2125 #endif
2126 	}
2127 	free(tmpbuf, M_TEMP);
2128 	return (error);
2129 }
2130 
2131 /*
2132  * Thus begins the fullpath magic.
2133  */
2134 
2135 static int __read_mostly disablefullpath;
2136 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, &disablefullpath, 0,
2137     "Disable the vn_fullpath function");
2138 
2139 /*
2140  * Retrieve the full filesystem path that correspond to a vnode from the name
2141  * cache (if available)
2142  */
2143 int
2144 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf)
2145 {
2146 	char *buf;
2147 	struct filedesc *fdp;
2148 	struct vnode *rdir;
2149 	int error;
2150 
2151 	if (__predict_false(disablefullpath))
2152 		return (ENODEV);
2153 	if (__predict_false(vn == NULL))
2154 		return (EINVAL);
2155 
2156 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2157 	fdp = td->td_proc->p_fd;
2158 	FILEDESC_SLOCK(fdp);
2159 	rdir = fdp->fd_rdir;
2160 	vrefact(rdir);
2161 	FILEDESC_SUNLOCK(fdp);
2162 	error = vn_fullpath1(td, vn, rdir, buf, retbuf, MAXPATHLEN);
2163 	vrele(rdir);
2164 
2165 	if (!error)
2166 		*freebuf = buf;
2167 	else
2168 		free(buf, M_TEMP);
2169 	return (error);
2170 }
2171 
2172 /*
2173  * This function is similar to vn_fullpath, but it attempts to lookup the
2174  * pathname relative to the global root mount point.  This is required for the
2175  * auditing sub-system, as audited pathnames must be absolute, relative to the
2176  * global root mount point.
2177  */
2178 int
2179 vn_fullpath_global(struct thread *td, struct vnode *vn,
2180     char **retbuf, char **freebuf)
2181 {
2182 	char *buf;
2183 	int error;
2184 
2185 	if (__predict_false(disablefullpath))
2186 		return (ENODEV);
2187 	if (__predict_false(vn == NULL))
2188 		return (EINVAL);
2189 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2190 	error = vn_fullpath1(td, vn, rootvnode, buf, retbuf, MAXPATHLEN);
2191 	if (!error)
2192 		*freebuf = buf;
2193 	else
2194 		free(buf, M_TEMP);
2195 	return (error);
2196 }
2197 
2198 int
2199 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen)
2200 {
2201 	struct vnode *dvp;
2202 	struct namecache *ncp;
2203 	struct mtx *vlp;
2204 	int error;
2205 
2206 	vlp = VP2VNODELOCK(*vp);
2207 	mtx_lock(vlp);
2208 	TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) {
2209 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2210 			break;
2211 	}
2212 	if (ncp != NULL) {
2213 		if (*buflen < ncp->nc_nlen) {
2214 			mtx_unlock(vlp);
2215 			vrele(*vp);
2216 			counter_u64_add(numfullpathfail4, 1);
2217 			error = ENOMEM;
2218 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2219 			    vp, NULL);
2220 			return (error);
2221 		}
2222 		*buflen -= ncp->nc_nlen;
2223 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2224 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2225 		    ncp->nc_name, vp);
2226 		dvp = *vp;
2227 		*vp = ncp->nc_dvp;
2228 		vref(*vp);
2229 		mtx_unlock(vlp);
2230 		vrele(dvp);
2231 		return (0);
2232 	}
2233 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2234 
2235 	mtx_unlock(vlp);
2236 	vn_lock(*vp, LK_SHARED | LK_RETRY);
2237 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
2238 	vput(*vp);
2239 	if (error) {
2240 		counter_u64_add(numfullpathfail2, 1);
2241 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
2242 		return (error);
2243 	}
2244 
2245 	*vp = dvp;
2246 	if (dvp->v_iflag & VI_DOOMED) {
2247 		/* forced unmount */
2248 		vrele(dvp);
2249 		error = ENOENT;
2250 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2251 		return (error);
2252 	}
2253 	/*
2254 	 * *vp has its use count incremented still.
2255 	 */
2256 
2257 	return (0);
2258 }
2259 
2260 /*
2261  * The magic behind kern___getcwd() and vn_fullpath().
2262  */
2263 static int
2264 vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
2265     char *buf, char **retbuf, u_int buflen)
2266 {
2267 	int error, slash_prefixed;
2268 #ifdef KDTRACE_HOOKS
2269 	struct vnode *startvp = vp;
2270 #endif
2271 	struct vnode *vp1;
2272 
2273 	buflen--;
2274 	buf[buflen] = '\0';
2275 	error = 0;
2276 	slash_prefixed = 0;
2277 
2278 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
2279 	counter_u64_add(numfullpathcalls, 1);
2280 	vref(vp);
2281 	if (vp->v_type != VDIR) {
2282 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2283 		if (error)
2284 			return (error);
2285 		if (buflen == 0) {
2286 			vrele(vp);
2287 			return (ENOMEM);
2288 		}
2289 		buf[--buflen] = '/';
2290 		slash_prefixed = 1;
2291 	}
2292 	while (vp != rdir && vp != rootvnode) {
2293 		/*
2294 		 * The vp vnode must be already fully constructed,
2295 		 * since it is either found in namecache or obtained
2296 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
2297 		 * without obtaining the vnode lock.
2298 		 */
2299 		if ((vp->v_vflag & VV_ROOT) != 0) {
2300 			vn_lock(vp, LK_RETRY | LK_SHARED);
2301 
2302 			/*
2303 			 * With the vnode locked, check for races with
2304 			 * unmount, forced or not.  Note that we
2305 			 * already verified that vp is not equal to
2306 			 * the root vnode, which means that
2307 			 * mnt_vnodecovered can be NULL only for the
2308 			 * case of unmount.
2309 			 */
2310 			if ((vp->v_iflag & VI_DOOMED) != 0 ||
2311 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
2312 			    vp1->v_mountedhere != vp->v_mount) {
2313 				vput(vp);
2314 				error = ENOENT;
2315 				SDT_PROBE3(vfs, namecache, fullpath, return,
2316 				    error, vp, NULL);
2317 				break;
2318 			}
2319 
2320 			vref(vp1);
2321 			vput(vp);
2322 			vp = vp1;
2323 			continue;
2324 		}
2325 		if (vp->v_type != VDIR) {
2326 			vrele(vp);
2327 			counter_u64_add(numfullpathfail1, 1);
2328 			error = ENOTDIR;
2329 			SDT_PROBE3(vfs, namecache, fullpath, return,
2330 			    error, vp, NULL);
2331 			break;
2332 		}
2333 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2334 		if (error)
2335 			break;
2336 		if (buflen == 0) {
2337 			vrele(vp);
2338 			error = ENOMEM;
2339 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2340 			    startvp, NULL);
2341 			break;
2342 		}
2343 		buf[--buflen] = '/';
2344 		slash_prefixed = 1;
2345 	}
2346 	if (error)
2347 		return (error);
2348 	if (!slash_prefixed) {
2349 		if (buflen == 0) {
2350 			vrele(vp);
2351 			counter_u64_add(numfullpathfail4, 1);
2352 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
2353 			    startvp, NULL);
2354 			return (ENOMEM);
2355 		}
2356 		buf[--buflen] = '/';
2357 	}
2358 	counter_u64_add(numfullpathfound, 1);
2359 	vrele(vp);
2360 
2361 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, buf + buflen);
2362 	*retbuf = buf + buflen;
2363 	return (0);
2364 }
2365 
2366 struct vnode *
2367 vn_dir_dd_ino(struct vnode *vp)
2368 {
2369 	struct namecache *ncp;
2370 	struct vnode *ddvp;
2371 	struct mtx *vlp;
2372 
2373 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
2374 	vlp = VP2VNODELOCK(vp);
2375 	mtx_lock(vlp);
2376 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
2377 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
2378 			continue;
2379 		ddvp = ncp->nc_dvp;
2380 		vhold(ddvp);
2381 		mtx_unlock(vlp);
2382 		if (vget(ddvp, LK_SHARED | LK_NOWAIT | LK_VNHELD, curthread))
2383 			return (NULL);
2384 		return (ddvp);
2385 	}
2386 	mtx_unlock(vlp);
2387 	return (NULL);
2388 }
2389 
2390 int
2391 vn_commname(struct vnode *vp, char *buf, u_int buflen)
2392 {
2393 	struct namecache *ncp;
2394 	struct mtx *vlp;
2395 	int l;
2396 
2397 	vlp = VP2VNODELOCK(vp);
2398 	mtx_lock(vlp);
2399 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
2400 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2401 			break;
2402 	if (ncp == NULL) {
2403 		mtx_unlock(vlp);
2404 		return (ENOENT);
2405 	}
2406 	l = min(ncp->nc_nlen, buflen - 1);
2407 	memcpy(buf, ncp->nc_name, l);
2408 	mtx_unlock(vlp);
2409 	buf[l] = '\0';
2410 	return (0);
2411 }
2412 
2413 /* ABI compat shims for old kernel modules. */
2414 #undef cache_enter
2415 
2416 void	cache_enter(struct vnode *dvp, struct vnode *vp,
2417 	    struct componentname *cnp);
2418 
2419 void
2420 cache_enter(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
2421 {
2422 
2423 	cache_enter_time(dvp, vp, cnp, NULL, NULL);
2424 }
2425 
2426 /*
2427  * This function updates path string to vnode's full global path
2428  * and checks the size of the new path string against the pathlen argument.
2429  *
2430  * Requires a locked, referenced vnode.
2431  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
2432  *
2433  * If sysctl debug.disablefullpath is set, ENODEV is returned,
2434  * vnode is left locked and path remain untouched.
2435  *
2436  * If vp is a directory, the call to vn_fullpath_global() always succeeds
2437  * because it falls back to the ".." lookup if the namecache lookup fails.
2438  */
2439 int
2440 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
2441     u_int pathlen)
2442 {
2443 	struct nameidata nd;
2444 	struct vnode *vp1;
2445 	char *rpath, *fbuf;
2446 	int error;
2447 
2448 	ASSERT_VOP_ELOCKED(vp, __func__);
2449 
2450 	/* Return ENODEV if sysctl debug.disablefullpath==1 */
2451 	if (__predict_false(disablefullpath))
2452 		return (ENODEV);
2453 
2454 	/* Construct global filesystem path from vp. */
2455 	VOP_UNLOCK(vp, 0);
2456 	error = vn_fullpath_global(td, vp, &rpath, &fbuf);
2457 
2458 	if (error != 0) {
2459 		vrele(vp);
2460 		return (error);
2461 	}
2462 
2463 	if (strlen(rpath) >= pathlen) {
2464 		vrele(vp);
2465 		error = ENAMETOOLONG;
2466 		goto out;
2467 	}
2468 
2469 	/*
2470 	 * Re-lookup the vnode by path to detect a possible rename.
2471 	 * As a side effect, the vnode is relocked.
2472 	 * If vnode was renamed, return ENOENT.
2473 	 */
2474 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
2475 	    UIO_SYSSPACE, path, td);
2476 	error = namei(&nd);
2477 	if (error != 0) {
2478 		vrele(vp);
2479 		goto out;
2480 	}
2481 	NDFREE(&nd, NDF_ONLY_PNBUF);
2482 	vp1 = nd.ni_vp;
2483 	vrele(vp);
2484 	if (vp1 == vp)
2485 		strcpy(path, rpath);
2486 	else {
2487 		vput(vp1);
2488 		error = ENOENT;
2489 	}
2490 
2491 out:
2492 	free(fbuf, M_TEMP);
2493 	return (error);
2494 }
2495