xref: /freebsd/sys/kern/vfs_cache.c (revision ca987d4641cdcd7f27e153db17c5bf064934faf5)
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 = 12; /* 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_wlocked_bucket(struct namecache *ncp, struct rwlock *blp)
988 {
989 	struct mtx *dvlp, *vlp;
990 
991 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
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 		cache_zap_locked(ncp, false);
999 		rw_wunlock(blp);
1000 		cache_unlock_vnodes(dvlp, vlp);
1001 		return (0);
1002 	}
1003 
1004 	rw_wunlock(blp);
1005 	return (EAGAIN);
1006 }
1007 
1008 static int
1009 cache_zap_rlocked_bucket(struct namecache *ncp, struct rwlock *blp)
1010 {
1011 	struct mtx *dvlp, *vlp;
1012 
1013 	cache_assert_bucket_locked(ncp, RA_RLOCKED);
1014 
1015 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1016 	vlp = NULL;
1017 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1018 		vlp = VP2VNODELOCK(ncp->nc_vp);
1019 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1020 		rw_runlock(blp);
1021 		rw_wlock(blp);
1022 		cache_zap_locked(ncp, false);
1023 		rw_wunlock(blp);
1024 		cache_unlock_vnodes(dvlp, vlp);
1025 		return (0);
1026 	}
1027 
1028 	rw_runlock(blp);
1029 	return (EAGAIN);
1030 }
1031 
1032 static int
1033 cache_zap_wlocked_bucket_kl(struct namecache *ncp, struct rwlock *blp,
1034     struct mtx **vlpp1, struct mtx **vlpp2)
1035 {
1036 	struct mtx *dvlp, *vlp;
1037 
1038 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1039 
1040 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1041 	vlp = NULL;
1042 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1043 		vlp = VP2VNODELOCK(ncp->nc_vp);
1044 	cache_sort(&dvlp, &vlp);
1045 
1046 	if (*vlpp1 == dvlp && *vlpp2 == vlp) {
1047 		cache_zap_locked(ncp, false);
1048 		cache_unlock_vnodes(dvlp, vlp);
1049 		*vlpp1 = NULL;
1050 		*vlpp2 = NULL;
1051 		return (0);
1052 	}
1053 
1054 	if (*vlpp1 != NULL)
1055 		mtx_unlock(*vlpp1);
1056 	if (*vlpp2 != NULL)
1057 		mtx_unlock(*vlpp2);
1058 	*vlpp1 = NULL;
1059 	*vlpp2 = NULL;
1060 
1061 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1062 		cache_zap_locked(ncp, false);
1063 		cache_unlock_vnodes(dvlp, vlp);
1064 		return (0);
1065 	}
1066 
1067 	rw_wunlock(blp);
1068 	*vlpp1 = dvlp;
1069 	*vlpp2 = vlp;
1070 	if (*vlpp1 != NULL)
1071 		mtx_lock(*vlpp1);
1072 	mtx_lock(*vlpp2);
1073 	rw_wlock(blp);
1074 	return (EAGAIN);
1075 }
1076 
1077 static void
1078 cache_lookup_unlock(struct rwlock *blp, struct mtx *vlp)
1079 {
1080 
1081 	if (blp != NULL) {
1082 		rw_runlock(blp);
1083 	} else {
1084 		mtx_unlock(vlp);
1085 	}
1086 }
1087 
1088 static int __noinline
1089 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1090     struct timespec *tsp, int *ticksp)
1091 {
1092 	int ltype;
1093 
1094 	*vpp = dvp;
1095 	CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .",
1096 			dvp, cnp->cn_nameptr);
1097 	counter_u64_add(dothits, 1);
1098 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1099 	if (tsp != NULL)
1100 		timespecclear(tsp);
1101 	if (ticksp != NULL)
1102 		*ticksp = ticks;
1103 	vrefact(*vpp);
1104 	/*
1105 	 * When we lookup "." we still can be asked to lock it
1106 	 * differently...
1107 	 */
1108 	ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1109 	if (ltype != VOP_ISLOCKED(*vpp)) {
1110 		if (ltype == LK_EXCLUSIVE) {
1111 			vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1112 			if ((*vpp)->v_iflag & VI_DOOMED) {
1113 				/* forced unmount */
1114 				vrele(*vpp);
1115 				*vpp = NULL;
1116 				return (ENOENT);
1117 			}
1118 		} else
1119 			vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1120 	}
1121 	return (-1);
1122 }
1123 
1124 /*
1125  * Lookup an entry in the cache
1126  *
1127  * Lookup is called with dvp pointing to the directory to search,
1128  * cnp pointing to the name of the entry being sought. If the lookup
1129  * succeeds, the vnode is returned in *vpp, and a status of -1 is
1130  * returned. If the lookup determines that the name does not exist
1131  * (negative caching), a status of ENOENT is returned. If the lookup
1132  * fails, a status of zero is returned.  If the directory vnode is
1133  * recycled out from under us due to a forced unmount, a status of
1134  * ENOENT is returned.
1135  *
1136  * vpp is locked and ref'd on return.  If we're looking up DOTDOT, dvp is
1137  * unlocked.  If we're looking up . an extra ref is taken, but the lock is
1138  * not recursively acquired.
1139  */
1140 
1141 static __noinline int
1142 cache_lookup_nomakeentry(struct vnode *dvp, struct vnode **vpp,
1143     struct componentname *cnp, struct timespec *tsp, int *ticksp)
1144 {
1145 	struct namecache *ncp;
1146 	struct rwlock *blp;
1147 	struct mtx *dvlp, *dvlp2;
1148 	uint32_t hash;
1149 	int error;
1150 
1151 	if (cnp->cn_namelen == 2 &&
1152 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1153 		counter_u64_add(dotdothits, 1);
1154 		dvlp = VP2VNODELOCK(dvp);
1155 		dvlp2 = NULL;
1156 		mtx_lock(dvlp);
1157 retry_dotdot:
1158 		ncp = dvp->v_cache_dd;
1159 		if (ncp == NULL) {
1160 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1161 			    "..", NULL);
1162 			mtx_unlock(dvlp);
1163 			return (0);
1164 		}
1165 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1166 			if (ncp->nc_dvp != dvp)
1167 				panic("dvp %p v_cache_dd %p\n", dvp, ncp);
1168 			if (!cache_zap_locked_vnode_kl2(ncp,
1169 			    dvp, &dvlp2))
1170 				goto retry_dotdot;
1171 			MPASS(dvp->v_cache_dd == NULL);
1172 			mtx_unlock(dvlp);
1173 			if (dvlp2 != NULL)
1174 				mtx_unlock(dvlp2);
1175 			cache_free(ncp);
1176 		} else {
1177 			dvp->v_cache_dd = NULL;
1178 			mtx_unlock(dvlp);
1179 			if (dvlp2 != NULL)
1180 				mtx_unlock(dvlp2);
1181 		}
1182 		return (0);
1183 	}
1184 
1185 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1186 	blp = HASH2BUCKETLOCK(hash);
1187 retry:
1188 	if (LIST_EMPTY(NCHHASH(hash)))
1189 		goto out_no_entry;
1190 
1191 	rw_wlock(blp);
1192 
1193 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1194 		counter_u64_add(numchecks, 1);
1195 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1196 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1197 			break;
1198 	}
1199 
1200 	/* We failed to find an entry */
1201 	if (ncp == NULL) {
1202 		rw_wunlock(blp);
1203 		goto out_no_entry;
1204 	}
1205 
1206 	counter_u64_add(numposzaps, 1);
1207 
1208 	error = cache_zap_wlocked_bucket(ncp, blp);
1209 	if (error != 0) {
1210 		zap_and_exit_bucket_fail++;
1211 		cache_maybe_yield();
1212 		goto retry;
1213 	}
1214 	cache_free(ncp);
1215 	return (0);
1216 out_no_entry:
1217 	SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr, NULL);
1218 	counter_u64_add(nummisszap, 1);
1219 	return (0);
1220 }
1221 
1222 int
1223 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1224     struct timespec *tsp, int *ticksp)
1225 {
1226 	struct namecache_ts *ncp_ts;
1227 	struct namecache *ncp;
1228 	struct rwlock *blp;
1229 	struct mtx *dvlp, *dvlp2;
1230 	uint32_t hash;
1231 	int error, ltype;
1232 
1233 	if (__predict_false(!doingcache)) {
1234 		cnp->cn_flags &= ~MAKEENTRY;
1235 		return (0);
1236 	}
1237 
1238 	counter_u64_add(numcalls, 1);
1239 
1240 	if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.'))
1241 		return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1242 
1243 	if ((cnp->cn_flags & MAKEENTRY) == 0)
1244 		return (cache_lookup_nomakeentry(dvp, vpp, cnp, tsp, ticksp));
1245 
1246 retry:
1247 	blp = NULL;
1248 	error = 0;
1249 	if (cnp->cn_namelen == 2 &&
1250 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1251 		counter_u64_add(dotdothits, 1);
1252 		dvlp = VP2VNODELOCK(dvp);
1253 		dvlp2 = NULL;
1254 		mtx_lock(dvlp);
1255 		ncp = dvp->v_cache_dd;
1256 		if (ncp == NULL) {
1257 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1258 			    "..", NULL);
1259 			mtx_unlock(dvlp);
1260 			return (0);
1261 		}
1262 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1263 			if (ncp->nc_flag & NCF_NEGATIVE)
1264 				*vpp = NULL;
1265 			else
1266 				*vpp = ncp->nc_vp;
1267 		} else
1268 			*vpp = ncp->nc_dvp;
1269 		/* Return failure if negative entry was found. */
1270 		if (*vpp == NULL)
1271 			goto negative_success;
1272 		CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
1273 		    dvp, cnp->cn_nameptr, *vpp);
1274 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..",
1275 		    *vpp);
1276 		cache_out_ts(ncp, tsp, ticksp);
1277 		if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1278 		    NCF_DTS && tsp != NULL) {
1279 			ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1280 			*tsp = ncp_ts->nc_dotdottime;
1281 		}
1282 		goto success;
1283 	}
1284 
1285 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1286 	blp = HASH2BUCKETLOCK(hash);
1287 	rw_rlock(blp);
1288 
1289 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1290 		counter_u64_add(numchecks, 1);
1291 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1292 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1293 			break;
1294 	}
1295 
1296 	/* We failed to find an entry */
1297 	if (ncp == NULL) {
1298 		rw_runlock(blp);
1299 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1300 		    NULL);
1301 		counter_u64_add(nummiss, 1);
1302 		return (0);
1303 	}
1304 
1305 	/* We found a "positive" match, return the vnode */
1306 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1307 		counter_u64_add(numposhits, 1);
1308 		*vpp = ncp->nc_vp;
1309 		CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p",
1310 		    dvp, cnp->cn_nameptr, *vpp, ncp);
1311 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name,
1312 		    *vpp);
1313 		cache_out_ts(ncp, tsp, ticksp);
1314 		goto success;
1315 	}
1316 
1317 negative_success:
1318 	/* We found a negative match, and want to create it, so purge */
1319 	if (cnp->cn_nameiop == CREATE) {
1320 		counter_u64_add(numnegzaps, 1);
1321 		goto zap_and_exit;
1322 	}
1323 
1324 	counter_u64_add(numneghits, 1);
1325 	cache_negative_hit(ncp);
1326 	if (ncp->nc_flag & NCF_WHITE)
1327 		cnp->cn_flags |= ISWHITEOUT;
1328 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp,
1329 	    ncp->nc_name);
1330 	cache_out_ts(ncp, tsp, ticksp);
1331 	cache_lookup_unlock(blp, dvlp);
1332 	return (ENOENT);
1333 
1334 success:
1335 	/*
1336 	 * On success we return a locked and ref'd vnode as per the lookup
1337 	 * protocol.
1338 	 */
1339 	MPASS(dvp != *vpp);
1340 	ltype = 0;	/* silence gcc warning */
1341 	if (cnp->cn_flags & ISDOTDOT) {
1342 		ltype = VOP_ISLOCKED(dvp);
1343 		VOP_UNLOCK(dvp, 0);
1344 	}
1345 	vhold(*vpp);
1346 	cache_lookup_unlock(blp, dvlp);
1347 	error = vget(*vpp, cnp->cn_lkflags | LK_VNHELD, cnp->cn_thread);
1348 	if (cnp->cn_flags & ISDOTDOT) {
1349 		vn_lock(dvp, ltype | LK_RETRY);
1350 		if (dvp->v_iflag & VI_DOOMED) {
1351 			if (error == 0)
1352 				vput(*vpp);
1353 			*vpp = NULL;
1354 			return (ENOENT);
1355 		}
1356 	}
1357 	if (error) {
1358 		*vpp = NULL;
1359 		goto retry;
1360 	}
1361 	if ((cnp->cn_flags & ISLASTCN) &&
1362 	    (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) {
1363 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
1364 	}
1365 	return (-1);
1366 
1367 zap_and_exit:
1368 	if (blp != NULL)
1369 		error = cache_zap_rlocked_bucket(ncp, blp);
1370 	else
1371 		error = cache_zap_locked_vnode(ncp, dvp);
1372 	if (error != 0) {
1373 		zap_and_exit_bucket_fail++;
1374 		cache_maybe_yield();
1375 		goto retry;
1376 	}
1377 	cache_free(ncp);
1378 	return (0);
1379 }
1380 
1381 struct celockstate {
1382 	struct mtx *vlp[3];
1383 	struct rwlock *blp[2];
1384 };
1385 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1386 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1387 
1388 static inline void
1389 cache_celockstate_init(struct celockstate *cel)
1390 {
1391 
1392 	bzero(cel, sizeof(*cel));
1393 }
1394 
1395 static void
1396 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1397     struct vnode *dvp)
1398 {
1399 	struct mtx *vlp1, *vlp2;
1400 
1401 	MPASS(cel->vlp[0] == NULL);
1402 	MPASS(cel->vlp[1] == NULL);
1403 	MPASS(cel->vlp[2] == NULL);
1404 
1405 	MPASS(vp != NULL || dvp != NULL);
1406 
1407 	vlp1 = VP2VNODELOCK(vp);
1408 	vlp2 = VP2VNODELOCK(dvp);
1409 	cache_sort(&vlp1, &vlp2);
1410 
1411 	if (vlp1 != NULL) {
1412 		mtx_lock(vlp1);
1413 		cel->vlp[0] = vlp1;
1414 	}
1415 	mtx_lock(vlp2);
1416 	cel->vlp[1] = vlp2;
1417 }
1418 
1419 static void
1420 cache_unlock_vnodes_cel(struct celockstate *cel)
1421 {
1422 
1423 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
1424 
1425 	if (cel->vlp[0] != NULL)
1426 		mtx_unlock(cel->vlp[0]);
1427 	if (cel->vlp[1] != NULL)
1428 		mtx_unlock(cel->vlp[1]);
1429 	if (cel->vlp[2] != NULL)
1430 		mtx_unlock(cel->vlp[2]);
1431 }
1432 
1433 static bool
1434 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
1435 {
1436 	struct mtx *vlp;
1437 	bool ret;
1438 
1439 	cache_assert_vlp_locked(cel->vlp[0]);
1440 	cache_assert_vlp_locked(cel->vlp[1]);
1441 	MPASS(cel->vlp[2] == NULL);
1442 
1443 	MPASS(vp != NULL);
1444 	vlp = VP2VNODELOCK(vp);
1445 
1446 	ret = true;
1447 	if (vlp >= cel->vlp[1]) {
1448 		mtx_lock(vlp);
1449 	} else {
1450 		if (mtx_trylock(vlp))
1451 			goto out;
1452 		cache_lock_vnodes_cel_3_failures++;
1453 		cache_unlock_vnodes_cel(cel);
1454 		if (vlp < cel->vlp[0]) {
1455 			mtx_lock(vlp);
1456 			mtx_lock(cel->vlp[0]);
1457 			mtx_lock(cel->vlp[1]);
1458 		} else {
1459 			if (cel->vlp[0] != NULL)
1460 				mtx_lock(cel->vlp[0]);
1461 			mtx_lock(vlp);
1462 			mtx_lock(cel->vlp[1]);
1463 		}
1464 		ret = false;
1465 	}
1466 out:
1467 	cel->vlp[2] = vlp;
1468 	return (ret);
1469 }
1470 
1471 static void
1472 cache_lock_buckets_cel(struct celockstate *cel, struct rwlock *blp1,
1473     struct rwlock *blp2)
1474 {
1475 
1476 	MPASS(cel->blp[0] == NULL);
1477 	MPASS(cel->blp[1] == NULL);
1478 
1479 	cache_sort(&blp1, &blp2);
1480 
1481 	if (blp1 != NULL) {
1482 		rw_wlock(blp1);
1483 		cel->blp[0] = blp1;
1484 	}
1485 	rw_wlock(blp2);
1486 	cel->blp[1] = blp2;
1487 }
1488 
1489 static void
1490 cache_unlock_buckets_cel(struct celockstate *cel)
1491 {
1492 
1493 	if (cel->blp[0] != NULL)
1494 		rw_wunlock(cel->blp[0]);
1495 	rw_wunlock(cel->blp[1]);
1496 }
1497 
1498 /*
1499  * Lock part of the cache affected by the insertion.
1500  *
1501  * This means vnodelocks for dvp, vp and the relevant bucketlock.
1502  * However, insertion can result in removal of an old entry. In this
1503  * case we have an additional vnode and bucketlock pair to lock. If the
1504  * entry is negative, ncelock is locked instead of the vnode.
1505  *
1506  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
1507  * preserving the locking order (smaller address first).
1508  */
1509 static void
1510 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1511     uint32_t hash)
1512 {
1513 	struct namecache *ncp;
1514 	struct rwlock *blps[2];
1515 
1516 	blps[0] = HASH2BUCKETLOCK(hash);
1517 	for (;;) {
1518 		blps[1] = NULL;
1519 		cache_lock_vnodes_cel(cel, dvp, vp);
1520 		if (vp == NULL || vp->v_type != VDIR)
1521 			break;
1522 		ncp = vp->v_cache_dd;
1523 		if (ncp == NULL)
1524 			break;
1525 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1526 			break;
1527 		MPASS(ncp->nc_dvp == vp);
1528 		blps[1] = NCP2BUCKETLOCK(ncp);
1529 		if (ncp->nc_flag & NCF_NEGATIVE)
1530 			break;
1531 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1532 			break;
1533 		/*
1534 		 * All vnodes got re-locked. Re-validate the state and if
1535 		 * nothing changed we are done. Otherwise restart.
1536 		 */
1537 		if (ncp == vp->v_cache_dd &&
1538 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1539 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1540 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1541 			break;
1542 		cache_unlock_vnodes_cel(cel);
1543 		cel->vlp[0] = NULL;
1544 		cel->vlp[1] = NULL;
1545 		cel->vlp[2] = NULL;
1546 	}
1547 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1548 }
1549 
1550 static void
1551 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1552     uint32_t hash)
1553 {
1554 	struct namecache *ncp;
1555 	struct rwlock *blps[2];
1556 
1557 	blps[0] = HASH2BUCKETLOCK(hash);
1558 	for (;;) {
1559 		blps[1] = NULL;
1560 		cache_lock_vnodes_cel(cel, dvp, vp);
1561 		ncp = dvp->v_cache_dd;
1562 		if (ncp == NULL)
1563 			break;
1564 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1565 			break;
1566 		MPASS(ncp->nc_dvp == dvp);
1567 		blps[1] = NCP2BUCKETLOCK(ncp);
1568 		if (ncp->nc_flag & NCF_NEGATIVE)
1569 			break;
1570 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1571 			break;
1572 		if (ncp == dvp->v_cache_dd &&
1573 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1574 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1575 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1576 			break;
1577 		cache_unlock_vnodes_cel(cel);
1578 		cel->vlp[0] = NULL;
1579 		cel->vlp[1] = NULL;
1580 		cel->vlp[2] = NULL;
1581 	}
1582 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1583 }
1584 
1585 static void
1586 cache_enter_unlock(struct celockstate *cel)
1587 {
1588 
1589 	cache_unlock_buckets_cel(cel);
1590 	cache_unlock_vnodes_cel(cel);
1591 }
1592 
1593 /*
1594  * Add an entry to the cache.
1595  */
1596 void
1597 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
1598     struct timespec *tsp, struct timespec *dtsp)
1599 {
1600 	struct celockstate cel;
1601 	struct namecache *ncp, *n2, *ndd;
1602 	struct namecache_ts *ncp_ts, *n2_ts;
1603 	struct nchashhead *ncpp;
1604 	struct neglist *neglist;
1605 	uint32_t hash;
1606 	int flag;
1607 	int len;
1608 	bool neg_locked;
1609 	int lnumcache;
1610 
1611 	CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr);
1612 	VNASSERT(vp == NULL || (vp->v_iflag & VI_DOOMED) == 0, vp,
1613 	    ("cache_enter: Adding a doomed vnode"));
1614 	VNASSERT(dvp == NULL || (dvp->v_iflag & VI_DOOMED) == 0, dvp,
1615 	    ("cache_enter: Doomed vnode used as src"));
1616 
1617 	if (__predict_false(!doingcache))
1618 		return;
1619 
1620 	/*
1621 	 * Avoid blowout in namecache entries.
1622 	 */
1623 	if (__predict_false(numcache >= desiredvnodes * ncsizefactor))
1624 		return;
1625 
1626 	cache_celockstate_init(&cel);
1627 	ndd = NULL;
1628 	flag = 0;
1629 	if (cnp->cn_nameptr[0] == '.') {
1630 		if (cnp->cn_namelen == 1)
1631 			return;
1632 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1633 			len = cnp->cn_namelen;
1634 			hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1635 			cache_enter_lock_dd(&cel, dvp, vp, hash);
1636 			/*
1637 			 * If dotdot entry already exists, just retarget it
1638 			 * to new parent vnode, otherwise continue with new
1639 			 * namecache entry allocation.
1640 			 */
1641 			if ((ncp = dvp->v_cache_dd) != NULL &&
1642 			    ncp->nc_flag & NCF_ISDOTDOT) {
1643 				KASSERT(ncp->nc_dvp == dvp,
1644 				    ("wrong isdotdot parent"));
1645 				neg_locked = false;
1646 				if (ncp->nc_flag & NCF_NEGATIVE || vp == NULL) {
1647 					neglist = NCP2NEGLIST(ncp);
1648 					mtx_lock(&ncneg_hot.nl_lock);
1649 					mtx_lock(&neglist->nl_lock);
1650 					neg_locked = true;
1651 				}
1652 				if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1653 					TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst,
1654 					    ncp, nc_dst);
1655 				} else {
1656 					cache_negative_remove(ncp, true);
1657 				}
1658 				if (vp != NULL) {
1659 					TAILQ_INSERT_HEAD(&vp->v_cache_dst,
1660 					    ncp, nc_dst);
1661 					ncp->nc_flag &= ~(NCF_NEGATIVE|NCF_HOTNEGATIVE);
1662 				} else {
1663 					ncp->nc_flag &= ~(NCF_HOTNEGATIVE);
1664 					ncp->nc_flag |= NCF_NEGATIVE;
1665 					cache_negative_insert(ncp, true);
1666 				}
1667 				if (neg_locked) {
1668 					mtx_unlock(&neglist->nl_lock);
1669 					mtx_unlock(&ncneg_hot.nl_lock);
1670 				}
1671 				ncp->nc_vp = vp;
1672 				cache_enter_unlock(&cel);
1673 				return;
1674 			}
1675 			dvp->v_cache_dd = NULL;
1676 			cache_enter_unlock(&cel);
1677 			cache_celockstate_init(&cel);
1678 			SDT_PROBE3(vfs, namecache, enter, done, dvp, "..", vp);
1679 			flag = NCF_ISDOTDOT;
1680 		}
1681 	}
1682 
1683 	/*
1684 	 * Calculate the hash key and setup as much of the new
1685 	 * namecache entry as possible before acquiring the lock.
1686 	 */
1687 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
1688 	ncp->nc_flag = flag;
1689 	ncp->nc_vp = vp;
1690 	if (vp == NULL)
1691 		ncp->nc_flag |= NCF_NEGATIVE;
1692 	ncp->nc_dvp = dvp;
1693 	if (tsp != NULL) {
1694 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1695 		ncp_ts->nc_time = *tsp;
1696 		ncp_ts->nc_ticks = ticks;
1697 		ncp_ts->nc_nc.nc_flag |= NCF_TS;
1698 		if (dtsp != NULL) {
1699 			ncp_ts->nc_dotdottime = *dtsp;
1700 			ncp_ts->nc_nc.nc_flag |= NCF_DTS;
1701 		}
1702 	}
1703 	len = ncp->nc_nlen = cnp->cn_namelen;
1704 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1705 	strlcpy(ncp->nc_name, cnp->cn_nameptr, len + 1);
1706 	cache_enter_lock(&cel, dvp, vp, hash);
1707 
1708 	/*
1709 	 * See if this vnode or negative entry is already in the cache
1710 	 * with this name.  This can happen with concurrent lookups of
1711 	 * the same path name.
1712 	 */
1713 	ncpp = NCHHASH(hash);
1714 	LIST_FOREACH(n2, ncpp, nc_hash) {
1715 		if (n2->nc_dvp == dvp &&
1716 		    n2->nc_nlen == cnp->cn_namelen &&
1717 		    !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
1718 			if (tsp != NULL) {
1719 				KASSERT((n2->nc_flag & NCF_TS) != 0,
1720 				    ("no NCF_TS"));
1721 				n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
1722 				n2_ts->nc_time = ncp_ts->nc_time;
1723 				n2_ts->nc_ticks = ncp_ts->nc_ticks;
1724 				if (dtsp != NULL) {
1725 					n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
1726 					if (ncp->nc_flag & NCF_NEGATIVE)
1727 						mtx_lock(&ncneg_hot.nl_lock);
1728 					n2_ts->nc_nc.nc_flag |= NCF_DTS;
1729 					if (ncp->nc_flag & NCF_NEGATIVE)
1730 						mtx_unlock(&ncneg_hot.nl_lock);
1731 				}
1732 			}
1733 			goto out_unlock_free;
1734 		}
1735 	}
1736 
1737 	if (flag == NCF_ISDOTDOT) {
1738 		/*
1739 		 * See if we are trying to add .. entry, but some other lookup
1740 		 * has populated v_cache_dd pointer already.
1741 		 */
1742 		if (dvp->v_cache_dd != NULL)
1743 			goto out_unlock_free;
1744 		KASSERT(vp == NULL || vp->v_type == VDIR,
1745 		    ("wrong vnode type %p", vp));
1746 		dvp->v_cache_dd = ncp;
1747 	}
1748 
1749 	if (vp != NULL) {
1750 		if (vp->v_type == VDIR) {
1751 			if (flag != NCF_ISDOTDOT) {
1752 				/*
1753 				 * For this case, the cache entry maps both the
1754 				 * directory name in it and the name ".." for the
1755 				 * directory's parent.
1756 				 */
1757 				if ((ndd = vp->v_cache_dd) != NULL) {
1758 					if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
1759 						cache_zap_locked(ndd, false);
1760 					else
1761 						ndd = NULL;
1762 				}
1763 				vp->v_cache_dd = ncp;
1764 			}
1765 		} else {
1766 			vp->v_cache_dd = NULL;
1767 		}
1768 	}
1769 
1770 	if (flag != NCF_ISDOTDOT) {
1771 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1772 			vhold(dvp);
1773 			atomic_add_rel_long(&numcachehv, 1);
1774 		}
1775 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
1776 	}
1777 
1778 	/*
1779 	 * Insert the new namecache entry into the appropriate chain
1780 	 * within the cache entries table.
1781 	 */
1782 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
1783 
1784 	/*
1785 	 * If the entry is "negative", we place it into the
1786 	 * "negative" cache queue, otherwise, we place it into the
1787 	 * destination vnode's cache entries queue.
1788 	 */
1789 	if (vp != NULL) {
1790 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
1791 		SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
1792 		    vp);
1793 	} else {
1794 		if (cnp->cn_flags & ISWHITEOUT)
1795 			ncp->nc_flag |= NCF_WHITE;
1796 		cache_negative_insert(ncp, false);
1797 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
1798 		    ncp->nc_name);
1799 	}
1800 	cache_enter_unlock(&cel);
1801 	lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
1802 	if (numneg * ncnegfactor > lnumcache)
1803 		cache_negative_zap_one();
1804 	cache_free(ndd);
1805 	return;
1806 out_unlock_free:
1807 	cache_enter_unlock(&cel);
1808 	cache_free(ncp);
1809 	return;
1810 }
1811 
1812 static u_int
1813 cache_roundup_2(u_int val)
1814 {
1815 	u_int res;
1816 
1817 	for (res = 1; res <= val; res <<= 1)
1818 		continue;
1819 
1820 	return (res);
1821 }
1822 
1823 /*
1824  * Name cache initialization, from vfs_init() when we are booting
1825  */
1826 static void
1827 nchinit(void *dummy __unused)
1828 {
1829 	u_int i;
1830 
1831 	cache_zone_small = uma_zcreate("S VFS Cache",
1832 	    sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1,
1833 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1834 	    UMA_ZONE_ZINIT);
1835 	cache_zone_small_ts = uma_zcreate("STS VFS Cache",
1836 	    sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1,
1837 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1838 	    UMA_ZONE_ZINIT);
1839 	cache_zone_large = uma_zcreate("L VFS Cache",
1840 	    sizeof(struct namecache) + NAME_MAX + 1,
1841 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1842 	    UMA_ZONE_ZINIT);
1843 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache",
1844 	    sizeof(struct namecache_ts) + NAME_MAX + 1,
1845 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1846 	    UMA_ZONE_ZINIT);
1847 
1848 	nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash);
1849 	ncbuckethash = cache_roundup_2(mp_ncpus * 64) - 1;
1850 	if (ncbuckethash > nchash)
1851 		ncbuckethash = nchash;
1852 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
1853 	    M_WAITOK | M_ZERO);
1854 	for (i = 0; i < numbucketlocks; i++)
1855 		rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK | RW_RECURSE);
1856 	ncvnodehash = cache_roundup_2(mp_ncpus * 64) - 1;
1857 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
1858 	    M_WAITOK | M_ZERO);
1859 	for (i = 0; i < numvnodelocks; i++)
1860 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
1861 	ncpurgeminvnodes = numbucketlocks;
1862 
1863 	ncneghash = 3;
1864 	neglists = malloc(sizeof(*neglists) * numneglists, M_VFSCACHE,
1865 	    M_WAITOK | M_ZERO);
1866 	for (i = 0; i < numneglists; i++) {
1867 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
1868 		TAILQ_INIT(&neglists[i].nl_list);
1869 	}
1870 	mtx_init(&ncneg_hot.nl_lock, "ncneglh", NULL, MTX_DEF);
1871 	TAILQ_INIT(&ncneg_hot.nl_list);
1872 
1873 	mtx_init(&ncneg_shrink_lock, "ncnegs", NULL, MTX_DEF);
1874 
1875 	numcalls = counter_u64_alloc(M_WAITOK);
1876 	dothits = counter_u64_alloc(M_WAITOK);
1877 	dotdothits = counter_u64_alloc(M_WAITOK);
1878 	numchecks = counter_u64_alloc(M_WAITOK);
1879 	nummiss = counter_u64_alloc(M_WAITOK);
1880 	nummisszap = counter_u64_alloc(M_WAITOK);
1881 	numposzaps = counter_u64_alloc(M_WAITOK);
1882 	numposhits = counter_u64_alloc(M_WAITOK);
1883 	numnegzaps = counter_u64_alloc(M_WAITOK);
1884 	numneghits = counter_u64_alloc(M_WAITOK);
1885 	numfullpathcalls = counter_u64_alloc(M_WAITOK);
1886 	numfullpathfail1 = counter_u64_alloc(M_WAITOK);
1887 	numfullpathfail2 = counter_u64_alloc(M_WAITOK);
1888 	numfullpathfail4 = counter_u64_alloc(M_WAITOK);
1889 	numfullpathfound = counter_u64_alloc(M_WAITOK);
1890 }
1891 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
1892 
1893 void
1894 cache_changesize(int newmaxvnodes)
1895 {
1896 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
1897 	u_long new_nchash, old_nchash;
1898 	struct namecache *ncp;
1899 	uint32_t hash;
1900 	int i;
1901 
1902 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
1903 	if (newmaxvnodes < numbucketlocks)
1904 		newmaxvnodes = numbucketlocks;
1905 
1906 	new_nchashtbl = hashinit(newmaxvnodes, M_VFSCACHE, &new_nchash);
1907 	/* If same hash table size, nothing to do */
1908 	if (nchash == new_nchash) {
1909 		free(new_nchashtbl, M_VFSCACHE);
1910 		return;
1911 	}
1912 	/*
1913 	 * Move everything from the old hash table to the new table.
1914 	 * None of the namecache entries in the table can be removed
1915 	 * because to do so, they have to be removed from the hash table.
1916 	 */
1917 	cache_lock_all_vnodes();
1918 	cache_lock_all_buckets();
1919 	old_nchashtbl = nchashtbl;
1920 	old_nchash = nchash;
1921 	nchashtbl = new_nchashtbl;
1922 	nchash = new_nchash;
1923 	for (i = 0; i <= old_nchash; i++) {
1924 		while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) {
1925 			hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
1926 			    ncp->nc_dvp);
1927 			LIST_REMOVE(ncp, nc_hash);
1928 			LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
1929 		}
1930 	}
1931 	cache_unlock_all_buckets();
1932 	cache_unlock_all_vnodes();
1933 	free(old_nchashtbl, M_VFSCACHE);
1934 }
1935 
1936 /*
1937  * Invalidate all entries to a particular vnode.
1938  */
1939 void
1940 cache_purge(struct vnode *vp)
1941 {
1942 	TAILQ_HEAD(, namecache) ncps;
1943 	struct namecache *ncp, *nnp;
1944 	struct mtx *vlp, *vlp2;
1945 
1946 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
1947 	SDT_PROBE1(vfs, namecache, purge, done, vp);
1948 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
1949 	    vp->v_cache_dd == NULL)
1950 		return;
1951 	TAILQ_INIT(&ncps);
1952 	vlp = VP2VNODELOCK(vp);
1953 	vlp2 = NULL;
1954 	mtx_lock(vlp);
1955 retry:
1956 	while (!LIST_EMPTY(&vp->v_cache_src)) {
1957 		ncp = LIST_FIRST(&vp->v_cache_src);
1958 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
1959 			goto retry;
1960 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1961 	}
1962 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
1963 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
1964 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
1965 			goto retry;
1966 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1967 	}
1968 	ncp = vp->v_cache_dd;
1969 	if (ncp != NULL) {
1970 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
1971 		   ("lost dotdot link"));
1972 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
1973 			goto retry;
1974 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1975 	}
1976 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
1977 	mtx_unlock(vlp);
1978 	if (vlp2 != NULL)
1979 		mtx_unlock(vlp2);
1980 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
1981 		cache_free(ncp);
1982 	}
1983 }
1984 
1985 /*
1986  * Invalidate all negative entries for a particular directory vnode.
1987  */
1988 void
1989 cache_purge_negative(struct vnode *vp)
1990 {
1991 	TAILQ_HEAD(, namecache) ncps;
1992 	struct namecache *ncp, *nnp;
1993 	struct mtx *vlp;
1994 
1995 	CTR1(KTR_VFS, "cache_purge_negative(%p)", vp);
1996 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
1997 	if (LIST_EMPTY(&vp->v_cache_src))
1998 		return;
1999 	TAILQ_INIT(&ncps);
2000 	vlp = VP2VNODELOCK(vp);
2001 	mtx_lock(vlp);
2002 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2003 		if (!(ncp->nc_flag & NCF_NEGATIVE))
2004 			continue;
2005 		cache_zap_negative_locked_vnode_kl(ncp, vp);
2006 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2007 	}
2008 	mtx_unlock(vlp);
2009 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2010 		cache_free(ncp);
2011 	}
2012 }
2013 
2014 /*
2015  * Flush all entries referencing a particular filesystem.
2016  */
2017 void
2018 cache_purgevfs(struct mount *mp, bool force)
2019 {
2020 	TAILQ_HEAD(, namecache) ncps;
2021 	struct mtx *vlp1, *vlp2;
2022 	struct rwlock *blp;
2023 	struct nchashhead *bucket;
2024 	struct namecache *ncp, *nnp;
2025 	u_long i, j, n_nchash;
2026 	int error;
2027 
2028 	/* Scan hash tables for applicable entries */
2029 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2030 	if (!force && mp->mnt_nvnodelistsize <= ncpurgeminvnodes)
2031 		return;
2032 	TAILQ_INIT(&ncps);
2033 	n_nchash = nchash + 1;
2034 	vlp1 = vlp2 = NULL;
2035 	for (i = 0; i < numbucketlocks; i++) {
2036 		blp = (struct rwlock *)&bucketlocks[i];
2037 		rw_wlock(blp);
2038 		for (j = i; j < n_nchash; j += numbucketlocks) {
2039 retry:
2040 			bucket = &nchashtbl[j];
2041 			LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) {
2042 				cache_assert_bucket_locked(ncp, RA_WLOCKED);
2043 				if (ncp->nc_dvp->v_mount != mp)
2044 					continue;
2045 				error = cache_zap_wlocked_bucket_kl(ncp, blp,
2046 				    &vlp1, &vlp2);
2047 				if (error != 0)
2048 					goto retry;
2049 				TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst);
2050 			}
2051 		}
2052 		rw_wunlock(blp);
2053 		if (vlp1 == NULL && vlp2 == NULL)
2054 			cache_maybe_yield();
2055 	}
2056 	if (vlp1 != NULL)
2057 		mtx_unlock(vlp1);
2058 	if (vlp2 != NULL)
2059 		mtx_unlock(vlp2);
2060 
2061 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2062 		cache_free(ncp);
2063 	}
2064 }
2065 
2066 /*
2067  * Perform canonical checks and cache lookup and pass on to filesystem
2068  * through the vop_cachedlookup only if needed.
2069  */
2070 
2071 int
2072 vfs_cache_lookup(struct vop_lookup_args *ap)
2073 {
2074 	struct vnode *dvp;
2075 	int error;
2076 	struct vnode **vpp = ap->a_vpp;
2077 	struct componentname *cnp = ap->a_cnp;
2078 	struct ucred *cred = cnp->cn_cred;
2079 	int flags = cnp->cn_flags;
2080 	struct thread *td = cnp->cn_thread;
2081 
2082 	*vpp = NULL;
2083 	dvp = ap->a_dvp;
2084 
2085 	if (dvp->v_type != VDIR)
2086 		return (ENOTDIR);
2087 
2088 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2089 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2090 		return (EROFS);
2091 
2092 	error = VOP_ACCESS(dvp, VEXEC, cred, td);
2093 	if (error)
2094 		return (error);
2095 
2096 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2097 	if (error == 0)
2098 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2099 	if (error == -1)
2100 		return (0);
2101 	return (error);
2102 }
2103 
2104 /*
2105  * XXX All of these sysctls would probably be more productive dead.
2106  */
2107 static int __read_mostly disablecwd;
2108 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
2109    "Disable the getcwd syscall");
2110 
2111 /* Implementation of the getcwd syscall. */
2112 int
2113 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2114 {
2115 
2116 	return (kern___getcwd(td, uap->buf, UIO_USERSPACE, uap->buflen,
2117 	    MAXPATHLEN));
2118 }
2119 
2120 int
2121 kern___getcwd(struct thread *td, char *buf, enum uio_seg bufseg, size_t buflen,
2122     size_t path_max)
2123 {
2124 	char *bp, *tmpbuf;
2125 	struct filedesc *fdp;
2126 	struct vnode *cdir, *rdir;
2127 	int error;
2128 
2129 	if (__predict_false(disablecwd))
2130 		return (ENODEV);
2131 	if (__predict_false(buflen < 2))
2132 		return (EINVAL);
2133 	if (buflen > path_max)
2134 		buflen = path_max;
2135 
2136 	tmpbuf = malloc(buflen, M_TEMP, M_WAITOK);
2137 	fdp = td->td_proc->p_fd;
2138 	FILEDESC_SLOCK(fdp);
2139 	cdir = fdp->fd_cdir;
2140 	vrefact(cdir);
2141 	rdir = fdp->fd_rdir;
2142 	vrefact(rdir);
2143 	FILEDESC_SUNLOCK(fdp);
2144 	error = vn_fullpath1(td, cdir, rdir, tmpbuf, &bp, buflen);
2145 	vrele(rdir);
2146 	vrele(cdir);
2147 
2148 	if (!error) {
2149 		if (bufseg == UIO_SYSSPACE)
2150 			bcopy(bp, buf, strlen(bp) + 1);
2151 		else
2152 			error = copyout(bp, buf, strlen(bp) + 1);
2153 #ifdef KTRACE
2154 	if (KTRPOINT(curthread, KTR_NAMEI))
2155 		ktrnamei(bp);
2156 #endif
2157 	}
2158 	free(tmpbuf, M_TEMP);
2159 	return (error);
2160 }
2161 
2162 /*
2163  * Thus begins the fullpath magic.
2164  */
2165 
2166 static int __read_mostly disablefullpath;
2167 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, &disablefullpath, 0,
2168     "Disable the vn_fullpath function");
2169 
2170 /*
2171  * Retrieve the full filesystem path that correspond to a vnode from the name
2172  * cache (if available)
2173  */
2174 int
2175 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf)
2176 {
2177 	char *buf;
2178 	struct filedesc *fdp;
2179 	struct vnode *rdir;
2180 	int error;
2181 
2182 	if (__predict_false(disablefullpath))
2183 		return (ENODEV);
2184 	if (__predict_false(vn == NULL))
2185 		return (EINVAL);
2186 
2187 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2188 	fdp = td->td_proc->p_fd;
2189 	FILEDESC_SLOCK(fdp);
2190 	rdir = fdp->fd_rdir;
2191 	vrefact(rdir);
2192 	FILEDESC_SUNLOCK(fdp);
2193 	error = vn_fullpath1(td, vn, rdir, buf, retbuf, MAXPATHLEN);
2194 	vrele(rdir);
2195 
2196 	if (!error)
2197 		*freebuf = buf;
2198 	else
2199 		free(buf, M_TEMP);
2200 	return (error);
2201 }
2202 
2203 /*
2204  * This function is similar to vn_fullpath, but it attempts to lookup the
2205  * pathname relative to the global root mount point.  This is required for the
2206  * auditing sub-system, as audited pathnames must be absolute, relative to the
2207  * global root mount point.
2208  */
2209 int
2210 vn_fullpath_global(struct thread *td, struct vnode *vn,
2211     char **retbuf, char **freebuf)
2212 {
2213 	char *buf;
2214 	int error;
2215 
2216 	if (__predict_false(disablefullpath))
2217 		return (ENODEV);
2218 	if (__predict_false(vn == NULL))
2219 		return (EINVAL);
2220 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2221 	error = vn_fullpath1(td, vn, rootvnode, buf, retbuf, MAXPATHLEN);
2222 	if (!error)
2223 		*freebuf = buf;
2224 	else
2225 		free(buf, M_TEMP);
2226 	return (error);
2227 }
2228 
2229 int
2230 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen)
2231 {
2232 	struct vnode *dvp;
2233 	struct namecache *ncp;
2234 	struct mtx *vlp;
2235 	int error;
2236 
2237 	vlp = VP2VNODELOCK(*vp);
2238 	mtx_lock(vlp);
2239 	TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) {
2240 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2241 			break;
2242 	}
2243 	if (ncp != NULL) {
2244 		if (*buflen < ncp->nc_nlen) {
2245 			mtx_unlock(vlp);
2246 			vrele(*vp);
2247 			counter_u64_add(numfullpathfail4, 1);
2248 			error = ENOMEM;
2249 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2250 			    vp, NULL);
2251 			return (error);
2252 		}
2253 		*buflen -= ncp->nc_nlen;
2254 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2255 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2256 		    ncp->nc_name, vp);
2257 		dvp = *vp;
2258 		*vp = ncp->nc_dvp;
2259 		vref(*vp);
2260 		mtx_unlock(vlp);
2261 		vrele(dvp);
2262 		return (0);
2263 	}
2264 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2265 
2266 	mtx_unlock(vlp);
2267 	vn_lock(*vp, LK_SHARED | LK_RETRY);
2268 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
2269 	vput(*vp);
2270 	if (error) {
2271 		counter_u64_add(numfullpathfail2, 1);
2272 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
2273 		return (error);
2274 	}
2275 
2276 	*vp = dvp;
2277 	if (dvp->v_iflag & VI_DOOMED) {
2278 		/* forced unmount */
2279 		vrele(dvp);
2280 		error = ENOENT;
2281 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2282 		return (error);
2283 	}
2284 	/*
2285 	 * *vp has its use count incremented still.
2286 	 */
2287 
2288 	return (0);
2289 }
2290 
2291 /*
2292  * The magic behind kern___getcwd() and vn_fullpath().
2293  */
2294 static int
2295 vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
2296     char *buf, char **retbuf, u_int buflen)
2297 {
2298 	int error, slash_prefixed;
2299 #ifdef KDTRACE_HOOKS
2300 	struct vnode *startvp = vp;
2301 #endif
2302 	struct vnode *vp1;
2303 
2304 	buflen--;
2305 	buf[buflen] = '\0';
2306 	error = 0;
2307 	slash_prefixed = 0;
2308 
2309 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
2310 	counter_u64_add(numfullpathcalls, 1);
2311 	vref(vp);
2312 	if (vp->v_type != VDIR) {
2313 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2314 		if (error)
2315 			return (error);
2316 		if (buflen == 0) {
2317 			vrele(vp);
2318 			return (ENOMEM);
2319 		}
2320 		buf[--buflen] = '/';
2321 		slash_prefixed = 1;
2322 	}
2323 	while (vp != rdir && vp != rootvnode) {
2324 		/*
2325 		 * The vp vnode must be already fully constructed,
2326 		 * since it is either found in namecache or obtained
2327 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
2328 		 * without obtaining the vnode lock.
2329 		 */
2330 		if ((vp->v_vflag & VV_ROOT) != 0) {
2331 			vn_lock(vp, LK_RETRY | LK_SHARED);
2332 
2333 			/*
2334 			 * With the vnode locked, check for races with
2335 			 * unmount, forced or not.  Note that we
2336 			 * already verified that vp is not equal to
2337 			 * the root vnode, which means that
2338 			 * mnt_vnodecovered can be NULL only for the
2339 			 * case of unmount.
2340 			 */
2341 			if ((vp->v_iflag & VI_DOOMED) != 0 ||
2342 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
2343 			    vp1->v_mountedhere != vp->v_mount) {
2344 				vput(vp);
2345 				error = ENOENT;
2346 				SDT_PROBE3(vfs, namecache, fullpath, return,
2347 				    error, vp, NULL);
2348 				break;
2349 			}
2350 
2351 			vref(vp1);
2352 			vput(vp);
2353 			vp = vp1;
2354 			continue;
2355 		}
2356 		if (vp->v_type != VDIR) {
2357 			vrele(vp);
2358 			counter_u64_add(numfullpathfail1, 1);
2359 			error = ENOTDIR;
2360 			SDT_PROBE3(vfs, namecache, fullpath, return,
2361 			    error, vp, NULL);
2362 			break;
2363 		}
2364 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2365 		if (error)
2366 			break;
2367 		if (buflen == 0) {
2368 			vrele(vp);
2369 			error = ENOMEM;
2370 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2371 			    startvp, NULL);
2372 			break;
2373 		}
2374 		buf[--buflen] = '/';
2375 		slash_prefixed = 1;
2376 	}
2377 	if (error)
2378 		return (error);
2379 	if (!slash_prefixed) {
2380 		if (buflen == 0) {
2381 			vrele(vp);
2382 			counter_u64_add(numfullpathfail4, 1);
2383 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
2384 			    startvp, NULL);
2385 			return (ENOMEM);
2386 		}
2387 		buf[--buflen] = '/';
2388 	}
2389 	counter_u64_add(numfullpathfound, 1);
2390 	vrele(vp);
2391 
2392 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, buf + buflen);
2393 	*retbuf = buf + buflen;
2394 	return (0);
2395 }
2396 
2397 struct vnode *
2398 vn_dir_dd_ino(struct vnode *vp)
2399 {
2400 	struct namecache *ncp;
2401 	struct vnode *ddvp;
2402 	struct mtx *vlp;
2403 
2404 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
2405 	vlp = VP2VNODELOCK(vp);
2406 	mtx_lock(vlp);
2407 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
2408 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
2409 			continue;
2410 		ddvp = ncp->nc_dvp;
2411 		vhold(ddvp);
2412 		mtx_unlock(vlp);
2413 		if (vget(ddvp, LK_SHARED | LK_NOWAIT | LK_VNHELD, curthread))
2414 			return (NULL);
2415 		return (ddvp);
2416 	}
2417 	mtx_unlock(vlp);
2418 	return (NULL);
2419 }
2420 
2421 int
2422 vn_commname(struct vnode *vp, char *buf, u_int buflen)
2423 {
2424 	struct namecache *ncp;
2425 	struct mtx *vlp;
2426 	int l;
2427 
2428 	vlp = VP2VNODELOCK(vp);
2429 	mtx_lock(vlp);
2430 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
2431 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2432 			break;
2433 	if (ncp == NULL) {
2434 		mtx_unlock(vlp);
2435 		return (ENOENT);
2436 	}
2437 	l = min(ncp->nc_nlen, buflen - 1);
2438 	memcpy(buf, ncp->nc_name, l);
2439 	mtx_unlock(vlp);
2440 	buf[l] = '\0';
2441 	return (0);
2442 }
2443 
2444 /* ABI compat shims for old kernel modules. */
2445 #undef cache_enter
2446 
2447 void	cache_enter(struct vnode *dvp, struct vnode *vp,
2448 	    struct componentname *cnp);
2449 
2450 void
2451 cache_enter(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
2452 {
2453 
2454 	cache_enter_time(dvp, vp, cnp, NULL, NULL);
2455 }
2456 
2457 /*
2458  * This function updates path string to vnode's full global path
2459  * and checks the size of the new path string against the pathlen argument.
2460  *
2461  * Requires a locked, referenced vnode.
2462  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
2463  *
2464  * If sysctl debug.disablefullpath is set, ENODEV is returned,
2465  * vnode is left locked and path remain untouched.
2466  *
2467  * If vp is a directory, the call to vn_fullpath_global() always succeeds
2468  * because it falls back to the ".." lookup if the namecache lookup fails.
2469  */
2470 int
2471 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
2472     u_int pathlen)
2473 {
2474 	struct nameidata nd;
2475 	struct vnode *vp1;
2476 	char *rpath, *fbuf;
2477 	int error;
2478 
2479 	ASSERT_VOP_ELOCKED(vp, __func__);
2480 
2481 	/* Return ENODEV if sysctl debug.disablefullpath==1 */
2482 	if (__predict_false(disablefullpath))
2483 		return (ENODEV);
2484 
2485 	/* Construct global filesystem path from vp. */
2486 	VOP_UNLOCK(vp, 0);
2487 	error = vn_fullpath_global(td, vp, &rpath, &fbuf);
2488 
2489 	if (error != 0) {
2490 		vrele(vp);
2491 		return (error);
2492 	}
2493 
2494 	if (strlen(rpath) >= pathlen) {
2495 		vrele(vp);
2496 		error = ENAMETOOLONG;
2497 		goto out;
2498 	}
2499 
2500 	/*
2501 	 * Re-lookup the vnode by path to detect a possible rename.
2502 	 * As a side effect, the vnode is relocked.
2503 	 * If vnode was renamed, return ENOENT.
2504 	 */
2505 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
2506 	    UIO_SYSSPACE, path, td);
2507 	error = namei(&nd);
2508 	if (error != 0) {
2509 		vrele(vp);
2510 		goto out;
2511 	}
2512 	NDFREE(&nd, NDF_ONLY_PNBUF);
2513 	vp1 = nd.ni_vp;
2514 	vrele(vp);
2515 	if (vp1 == vp)
2516 		strcpy(path, rpath);
2517 	else {
2518 		vput(vp1);
2519 		error = ENOENT;
2520 	}
2521 
2522 out:
2523 	free(fbuf, M_TEMP);
2524 	return (error);
2525 }
2526