xref: /freebsd/sys/kern/vfs_cache.c (revision 037479ff5ee18977b1c48e1e59770aad2f200a5a)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993, 1995
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * Poul-Henning Kamp of the FreeBSD Project.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)vfs_cache.c	8.5 (Berkeley) 3/22/95
35  */
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include "opt_ddb.h"
41 #include "opt_ktrace.h"
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/counter.h>
46 #include <sys/filedesc.h>
47 #include <sys/fnv_hash.h>
48 #include <sys/kernel.h>
49 #include <sys/lock.h>
50 #include <sys/malloc.h>
51 #include <sys/fcntl.h>
52 #include <sys/mount.h>
53 #include <sys/namei.h>
54 #include <sys/proc.h>
55 #include <sys/rwlock.h>
56 #include <sys/sdt.h>
57 #include <sys/smp.h>
58 #include <sys/syscallsubr.h>
59 #include <sys/sysctl.h>
60 #include <sys/sysproto.h>
61 #include <sys/vnode.h>
62 #ifdef KTRACE
63 #include <sys/ktrace.h>
64 #endif
65 
66 #ifdef DDB
67 #include <ddb/ddb.h>
68 #endif
69 
70 #include <vm/uma.h>
71 
72 SDT_PROVIDER_DECLARE(vfs);
73 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
74     "struct vnode *");
75 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
76     "char *");
77 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
78 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
79     "char *", "struct vnode *");
80 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
81 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
82     "struct vnode *", "char *");
83 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
84     "struct vnode *");
85 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
86     "struct vnode *", "char *");
87 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
88     "char *");
89 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *");
90 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
91 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
92 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
93     "struct vnode *");
94 SDT_PROBE_DEFINE3(vfs, namecache, zap_negative, done, "struct vnode *",
95     "char *", "int");
96 SDT_PROBE_DEFINE3(vfs, namecache, shrink_negative, done, "struct vnode *",
97     "char *", "int");
98 
99 /*
100  * This structure describes the elements in the cache of recent
101  * names looked up by namei.
102  */
103 
104 struct	namecache {
105 	LIST_ENTRY(namecache) nc_hash;	/* hash chain */
106 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
107 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
108 	struct	vnode *nc_dvp;		/* vnode of parent of name */
109 	union {
110 		struct	vnode *nu_vp;	/* vnode the name refers to */
111 		u_int	nu_neghits;	/* negative entry hits */
112 	} n_un;
113 	u_char	nc_flag;		/* flag bits */
114 	u_char	nc_nlen;		/* length of name */
115 	char	nc_name[0];		/* segment name + nul */
116 };
117 
118 /*
119  * struct namecache_ts repeats struct namecache layout up to the
120  * nc_nlen member.
121  * struct namecache_ts is used in place of struct namecache when time(s) need
122  * to be stored.  The nc_dotdottime field is used when a cache entry is mapping
123  * both a non-dotdot directory name plus dotdot for the directory's
124  * parent.
125  */
126 struct	namecache_ts {
127 	struct	timespec nc_time;	/* timespec provided by fs */
128 	struct	timespec nc_dotdottime;	/* dotdot timespec provided by fs */
129 	int	nc_ticks;		/* ticks value when entry was added */
130 	struct namecache nc_nc;
131 };
132 
133 #define	nc_vp		n_un.nu_vp
134 #define	nc_neghits	n_un.nu_neghits
135 
136 /*
137  * Flags in namecache.nc_flag
138  */
139 #define NCF_WHITE	0x01
140 #define NCF_ISDOTDOT	0x02
141 #define	NCF_TS		0x04
142 #define	NCF_DTS		0x08
143 #define	NCF_DVDROP	0x10
144 #define	NCF_NEGATIVE	0x20
145 #define	NCF_HOTNEGATIVE	0x40
146 
147 /*
148  * Name caching works as follows:
149  *
150  * Names found by directory scans are retained in a cache
151  * for future reference.  It is managed LRU, so frequently
152  * used names will hang around.  Cache is indexed by hash value
153  * obtained from (vp, name) where vp refers to the directory
154  * containing name.
155  *
156  * If it is a "negative" entry, (i.e. for a name that is known NOT to
157  * exist) the vnode pointer will be NULL.
158  *
159  * Upon reaching the last segment of a path, if the reference
160  * is for DELETE, or NOCACHE is set (rewrite), and the
161  * name is located in the cache, it will be dropped.
162  *
163  * These locks are used (in the order in which they can be taken):
164  * NAME		TYPE	ROLE
165  * vnodelock	mtx	vnode lists and v_cache_dd field protection
166  * bucketlock	rwlock	for access to given set of hash buckets
167  * neglist	mtx	negative entry LRU management
168  *
169  * Additionally, ncneg_shrink_lock mtx is used to have at most one thread
170  * shrinking the LRU list.
171  *
172  * It is legal to take multiple vnodelock and bucketlock locks. The locking
173  * order is lower address first. Both are recursive.
174  *
175  * "." lookups are lockless.
176  *
177  * ".." and vnode -> name lookups require vnodelock.
178  *
179  * name -> vnode lookup requires the relevant bucketlock to be held for reading.
180  *
181  * Insertions and removals of entries require involved vnodes and bucketlocks
182  * to be write-locked to prevent other threads from seeing the entry.
183  *
184  * Some lookups result in removal of the found entry (e.g. getting rid of a
185  * negative entry with the intent to create a positive one), which poses a
186  * problem when multiple threads reach the state. Similarly, two different
187  * threads can purge two different vnodes and try to remove the same name.
188  *
189  * If the already held vnode lock is lower than the second required lock, we
190  * can just take the other lock. However, in the opposite case, this could
191  * deadlock. As such, this is resolved by trylocking and if that fails unlocking
192  * the first node, locking everything in order and revalidating the state.
193  */
194 
195 /*
196  * Structures associated with name caching.
197  */
198 #define NCHHASH(hash) \
199 	(&nchashtbl[(hash) & nchash])
200 static __read_mostly LIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */
201 static u_long __read_mostly	nchash;			/* size of hash table */
202 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
203     "Size of namecache hash table");
204 static u_long __read_mostly	ncnegfactor = 12; /* ratio of negative entries */
205 SYSCTL_ULONG(_vfs, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
206     "Ratio of negative namecache entries");
207 static u_long __exclusive_cache_line	numneg;	/* number of negative entries allocated */
208 SYSCTL_ULONG(_debug, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0,
209     "Number of negative entries in namecache");
210 static u_long __exclusive_cache_line	numcache;/* number of cache entries allocated */
211 SYSCTL_ULONG(_debug, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0,
212     "Number of namecache entries");
213 static u_long __exclusive_cache_line	numcachehv;/* number of cache entries with vnodes held */
214 SYSCTL_ULONG(_debug, OID_AUTO, numcachehv, CTLFLAG_RD, &numcachehv, 0,
215     "Number of namecache entries with vnodes held");
216 u_int __read_mostly	ncsizefactor = 2;
217 SYSCTL_UINT(_vfs, OID_AUTO, ncsizefactor, CTLFLAG_RW, &ncsizefactor, 0,
218     "Size factor for namecache");
219 static u_int __read_mostly	ncpurgeminvnodes;
220 SYSCTL_UINT(_vfs, OID_AUTO, ncpurgeminvnodes, CTLFLAG_RW, &ncpurgeminvnodes, 0,
221     "Number of vnodes below which purgevfs ignores the request");
222 static u_int __read_mostly	ncneghitsrequeue = 8;
223 SYSCTL_UINT(_vfs, OID_AUTO, ncneghitsrequeue, CTLFLAG_RW, &ncneghitsrequeue, 0,
224     "Number of hits to requeue a negative entry in the LRU list");
225 
226 struct nchstats	nchstats;		/* cache effectiveness statistics */
227 
228 static struct mtx       ncneg_shrink_lock;
229 static int	shrink_list_turn;
230 
231 struct neglist {
232 	struct mtx		nl_lock;
233 	TAILQ_HEAD(, namecache) nl_list;
234 } __aligned(CACHE_LINE_SIZE);
235 
236 static struct neglist __read_mostly	*neglists;
237 static struct neglist ncneg_hot;
238 
239 #define	numneglists (ncneghash + 1)
240 static u_int __read_mostly	ncneghash;
241 static inline struct neglist *
242 NCP2NEGLIST(struct namecache *ncp)
243 {
244 
245 	return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
246 }
247 
248 #define	numbucketlocks (ncbuckethash + 1)
249 static u_int __read_mostly  ncbuckethash;
250 static struct rwlock_padalign __read_mostly  *bucketlocks;
251 #define	HASH2BUCKETLOCK(hash) \
252 	((struct rwlock *)(&bucketlocks[((hash) & ncbuckethash)]))
253 
254 #define	numvnodelocks (ncvnodehash + 1)
255 static u_int __read_mostly  ncvnodehash;
256 static struct mtx __read_mostly *vnodelocks;
257 static inline struct mtx *
258 VP2VNODELOCK(struct vnode *vp)
259 {
260 
261 	return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
262 }
263 
264 /*
265  * UMA zones for the VFS cache.
266  *
267  * The small cache is used for entries with short names, which are the
268  * most common.  The large cache is used for entries which are too big to
269  * fit in the small cache.
270  */
271 static uma_zone_t __read_mostly cache_zone_small;
272 static uma_zone_t __read_mostly cache_zone_small_ts;
273 static uma_zone_t __read_mostly cache_zone_large;
274 static uma_zone_t __read_mostly cache_zone_large_ts;
275 
276 #define	CACHE_PATH_CUTOFF	35
277 
278 static struct namecache *
279 cache_alloc(int len, int ts)
280 {
281 	struct namecache_ts *ncp_ts;
282 	struct namecache *ncp;
283 
284 	if (__predict_false(ts)) {
285 		if (len <= CACHE_PATH_CUTOFF)
286 			ncp_ts = uma_zalloc(cache_zone_small_ts, M_WAITOK);
287 		else
288 			ncp_ts = uma_zalloc(cache_zone_large_ts, M_WAITOK);
289 		ncp = &ncp_ts->nc_nc;
290 	} else {
291 		if (len <= CACHE_PATH_CUTOFF)
292 			ncp = uma_zalloc(cache_zone_small, M_WAITOK);
293 		else
294 			ncp = uma_zalloc(cache_zone_large, M_WAITOK);
295 	}
296 	return (ncp);
297 }
298 
299 static void
300 cache_free(struct namecache *ncp)
301 {
302 	struct namecache_ts *ncp_ts;
303 
304 	if (ncp == NULL)
305 		return;
306 	if ((ncp->nc_flag & NCF_DVDROP) != 0)
307 		vdrop(ncp->nc_dvp);
308 	if (__predict_false(ncp->nc_flag & NCF_TS)) {
309 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
310 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
311 			uma_zfree(cache_zone_small_ts, ncp_ts);
312 		else
313 			uma_zfree(cache_zone_large_ts, ncp_ts);
314 	} else {
315 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
316 			uma_zfree(cache_zone_small, ncp);
317 		else
318 			uma_zfree(cache_zone_large, ncp);
319 	}
320 }
321 
322 static void
323 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
324 {
325 	struct namecache_ts *ncp_ts;
326 
327 	KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
328 	    (tsp == NULL && ticksp == NULL),
329 	    ("No NCF_TS"));
330 
331 	if (tsp == NULL && ticksp == NULL)
332 		return;
333 
334 	ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
335 	if (tsp != NULL)
336 		*tsp = ncp_ts->nc_time;
337 	if (ticksp != NULL)
338 		*ticksp = ncp_ts->nc_ticks;
339 }
340 
341 static int __read_mostly	doingcache = 1;	/* 1 => enable the cache */
342 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
343     "VFS namecache enabled");
344 
345 /* Export size information to userland */
346 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
347     sizeof(struct namecache), "sizeof(struct namecache)");
348 
349 /*
350  * The new name cache statistics
351  */
352 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0,
353     "Name cache statistics");
354 #define STATNODE_ULONG(name, descr)	\
355 	SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, descr);
356 #define STATNODE_COUNTER(name, descr)	\
357 	static counter_u64_t __read_mostly name; \
358 	SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, descr);
359 STATNODE_ULONG(numneg, "Number of negative cache entries");
360 STATNODE_ULONG(numcache, "Number of cache entries");
361 STATNODE_COUNTER(numcalls, "Number of cache lookups");
362 STATNODE_COUNTER(dothits, "Number of '.' hits");
363 STATNODE_COUNTER(dotdothits, "Number of '..' hits");
364 STATNODE_COUNTER(numchecks, "Number of checks in lookup");
365 STATNODE_COUNTER(nummiss, "Number of cache misses");
366 STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache");
367 STATNODE_COUNTER(numposzaps,
368     "Number of cache hits (positive) we do not want to cache");
369 STATNODE_COUNTER(numposhits, "Number of cache hits (positive)");
370 STATNODE_COUNTER(numnegzaps,
371     "Number of cache hits (negative) we do not want to cache");
372 STATNODE_COUNTER(numneghits, "Number of cache hits (negative)");
373 /* These count for kern___getcwd(), too. */
374 STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls");
375 STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
376 STATNODE_COUNTER(numfullpathfail2,
377     "Number of fullpath search errors (VOP_VPTOCNP failures)");
378 STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
379 STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls");
380 static long zap_and_exit_bucket_fail; STATNODE_ULONG(zap_and_exit_bucket_fail,
381     "Number of times zap_and_exit failed to lock");
382 static long cache_lock_vnodes_cel_3_failures;
383 STATNODE_ULONG(cache_lock_vnodes_cel_3_failures,
384     "Number of times 3-way vnode locking failed");
385 
386 static void cache_zap_locked(struct namecache *ncp, bool neg_locked);
387 static int vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
388     char *buf, char **retbuf, u_int buflen);
389 
390 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
391 
392 static int cache_yield;
393 SYSCTL_INT(_vfs_cache, OID_AUTO, yield, CTLFLAG_RD, &cache_yield, 0,
394     "Number of times cache called yield");
395 
396 static void
397 cache_maybe_yield(void)
398 {
399 
400 	if (should_yield()) {
401 		cache_yield++;
402 		kern_yield(PRI_USER);
403 	}
404 }
405 
406 static inline void
407 cache_assert_vlp_locked(struct mtx *vlp)
408 {
409 
410 	if (vlp != NULL)
411 		mtx_assert(vlp, MA_OWNED);
412 }
413 
414 static inline void
415 cache_assert_vnode_locked(struct vnode *vp)
416 {
417 	struct mtx *vlp;
418 
419 	vlp = VP2VNODELOCK(vp);
420 	cache_assert_vlp_locked(vlp);
421 }
422 
423 static uint32_t
424 cache_get_hash(char *name, u_char len, struct vnode *dvp)
425 {
426 	uint32_t hash;
427 
428 	hash = fnv_32_buf(name, len, FNV1_32_INIT);
429 	hash = fnv_32_buf(&dvp, sizeof(dvp), hash);
430 	return (hash);
431 }
432 
433 static inline struct rwlock *
434 NCP2BUCKETLOCK(struct namecache *ncp)
435 {
436 	uint32_t hash;
437 
438 	hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
439 	return (HASH2BUCKETLOCK(hash));
440 }
441 
442 #ifdef INVARIANTS
443 static void
444 cache_assert_bucket_locked(struct namecache *ncp, int mode)
445 {
446 	struct rwlock *blp;
447 
448 	blp = NCP2BUCKETLOCK(ncp);
449 	rw_assert(blp, mode);
450 }
451 #else
452 #define cache_assert_bucket_locked(x, y) do { } while (0)
453 #endif
454 
455 #define cache_sort(x, y)	_cache_sort((void **)(x), (void **)(y))
456 static void
457 _cache_sort(void **p1, void **p2)
458 {
459 	void *tmp;
460 
461 	if (*p1 > *p2) {
462 		tmp = *p2;
463 		*p2 = *p1;
464 		*p1 = tmp;
465 	}
466 }
467 
468 static void
469 cache_lock_all_buckets(void)
470 {
471 	u_int i;
472 
473 	for (i = 0; i < numbucketlocks; i++)
474 		rw_wlock(&bucketlocks[i]);
475 }
476 
477 static void
478 cache_unlock_all_buckets(void)
479 {
480 	u_int i;
481 
482 	for (i = 0; i < numbucketlocks; i++)
483 		rw_wunlock(&bucketlocks[i]);
484 }
485 
486 static void
487 cache_lock_all_vnodes(void)
488 {
489 	u_int i;
490 
491 	for (i = 0; i < numvnodelocks; i++)
492 		mtx_lock(&vnodelocks[i]);
493 }
494 
495 static void
496 cache_unlock_all_vnodes(void)
497 {
498 	u_int i;
499 
500 	for (i = 0; i < numvnodelocks; i++)
501 		mtx_unlock(&vnodelocks[i]);
502 }
503 
504 static int
505 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
506 {
507 
508 	cache_sort(&vlp1, &vlp2);
509 	MPASS(vlp2 != NULL);
510 
511 	if (vlp1 != NULL) {
512 		if (!mtx_trylock(vlp1))
513 			return (EAGAIN);
514 	}
515 	if (!mtx_trylock(vlp2)) {
516 		if (vlp1 != NULL)
517 			mtx_unlock(vlp1);
518 		return (EAGAIN);
519 	}
520 
521 	return (0);
522 }
523 
524 static void
525 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
526 {
527 
528 	MPASS(vlp1 != NULL || vlp2 != NULL);
529 
530 	if (vlp1 != NULL)
531 		mtx_unlock(vlp1);
532 	if (vlp2 != NULL)
533 		mtx_unlock(vlp2);
534 }
535 
536 static int
537 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
538 {
539 	struct nchstats snap;
540 
541 	if (req->oldptr == NULL)
542 		return (SYSCTL_OUT(req, 0, sizeof(snap)));
543 
544 	snap = nchstats;
545 	snap.ncs_goodhits = counter_u64_fetch(numposhits);
546 	snap.ncs_neghits = counter_u64_fetch(numneghits);
547 	snap.ncs_badhits = counter_u64_fetch(numposzaps) +
548 	    counter_u64_fetch(numnegzaps);
549 	snap.ncs_miss = counter_u64_fetch(nummisszap) +
550 	    counter_u64_fetch(nummiss);
551 
552 	return (SYSCTL_OUT(req, &snap, sizeof(snap)));
553 }
554 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
555     CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
556     "VFS cache effectiveness statistics");
557 
558 #ifdef DIAGNOSTIC
559 /*
560  * Grab an atomic snapshot of the name cache hash chain lengths
561  */
562 static SYSCTL_NODE(_debug, OID_AUTO, hashstat, CTLFLAG_RW, NULL,
563     "hash table stats");
564 
565 static int
566 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
567 {
568 	struct nchashhead *ncpp;
569 	struct namecache *ncp;
570 	int i, error, n_nchash, *cntbuf;
571 
572 retry:
573 	n_nchash = nchash + 1;	/* nchash is max index, not count */
574 	if (req->oldptr == NULL)
575 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
576 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
577 	cache_lock_all_buckets();
578 	if (n_nchash != nchash + 1) {
579 		cache_unlock_all_buckets();
580 		free(cntbuf, M_TEMP);
581 		goto retry;
582 	}
583 	/* Scan hash tables counting entries */
584 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
585 		LIST_FOREACH(ncp, ncpp, nc_hash)
586 			cntbuf[i]++;
587 	cache_unlock_all_buckets();
588 	for (error = 0, i = 0; i < n_nchash; i++)
589 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
590 			break;
591 	free(cntbuf, M_TEMP);
592 	return (error);
593 }
594 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
595     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
596     "nchash chain lengths");
597 
598 static int
599 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
600 {
601 	int error;
602 	struct nchashhead *ncpp;
603 	struct namecache *ncp;
604 	int n_nchash;
605 	int count, maxlength, used, pct;
606 
607 	if (!req->oldptr)
608 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
609 
610 	cache_lock_all_buckets();
611 	n_nchash = nchash + 1;	/* nchash is max index, not count */
612 	used = 0;
613 	maxlength = 0;
614 
615 	/* Scan hash tables for applicable entries */
616 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
617 		count = 0;
618 		LIST_FOREACH(ncp, ncpp, nc_hash) {
619 			count++;
620 		}
621 		if (count)
622 			used++;
623 		if (maxlength < count)
624 			maxlength = count;
625 	}
626 	n_nchash = nchash + 1;
627 	cache_unlock_all_buckets();
628 	pct = (used * 100) / (n_nchash / 100);
629 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
630 	if (error)
631 		return (error);
632 	error = SYSCTL_OUT(req, &used, sizeof(used));
633 	if (error)
634 		return (error);
635 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
636 	if (error)
637 		return (error);
638 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
639 	if (error)
640 		return (error);
641 	return (0);
642 }
643 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
644     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
645     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
646 #endif
647 
648 /*
649  * Negative entries management
650  *
651  * A variation of LRU scheme is used. New entries are hashed into one of
652  * numneglists cold lists. Entries get promoted to the hot list on first hit.
653  * Partial LRU for the hot list is maintained by requeueing them every
654  * ncneghitsrequeue hits.
655  *
656  * The shrinker will demote hot list head and evict from the cold list in a
657  * round-robin manner.
658  */
659 static void
660 cache_negative_hit(struct namecache *ncp)
661 {
662 	struct neglist *neglist;
663 	u_int hits;
664 
665 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
666 	hits = atomic_fetchadd_int(&ncp->nc_neghits, 1);
667 	if (ncp->nc_flag & NCF_HOTNEGATIVE) {
668 		if ((hits % ncneghitsrequeue) != 0)
669 			return;
670 		mtx_lock(&ncneg_hot.nl_lock);
671 		if (ncp->nc_flag & NCF_HOTNEGATIVE) {
672 			TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
673 			TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
674 			mtx_unlock(&ncneg_hot.nl_lock);
675 			return;
676 		}
677 		/*
678 		 * The shrinker cleared the flag and removed the entry from
679 		 * the hot list. Put it back.
680 		 */
681 	} else {
682 		mtx_lock(&ncneg_hot.nl_lock);
683 	}
684 	neglist = NCP2NEGLIST(ncp);
685 	mtx_lock(&neglist->nl_lock);
686 	if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
687 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
688 		TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
689 		ncp->nc_flag |= NCF_HOTNEGATIVE;
690 	}
691 	mtx_unlock(&neglist->nl_lock);
692 	mtx_unlock(&ncneg_hot.nl_lock);
693 }
694 
695 static void
696 cache_negative_insert(struct namecache *ncp, bool neg_locked)
697 {
698 	struct neglist *neglist;
699 
700 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
701 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
702 	neglist = NCP2NEGLIST(ncp);
703 	if (!neg_locked) {
704 		mtx_lock(&neglist->nl_lock);
705 	} else {
706 		mtx_assert(&neglist->nl_lock, MA_OWNED);
707 	}
708 	TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
709 	if (!neg_locked)
710 		mtx_unlock(&neglist->nl_lock);
711 	atomic_add_rel_long(&numneg, 1);
712 }
713 
714 static void
715 cache_negative_remove(struct namecache *ncp, bool neg_locked)
716 {
717 	struct neglist *neglist;
718 	bool hot_locked = false;
719 	bool list_locked = false;
720 
721 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
722 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
723 	neglist = NCP2NEGLIST(ncp);
724 	if (!neg_locked) {
725 		if (ncp->nc_flag & NCF_HOTNEGATIVE) {
726 			hot_locked = true;
727 			mtx_lock(&ncneg_hot.nl_lock);
728 			if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
729 				list_locked = true;
730 				mtx_lock(&neglist->nl_lock);
731 			}
732 		} else {
733 			list_locked = true;
734 			mtx_lock(&neglist->nl_lock);
735 		}
736 	}
737 	if (ncp->nc_flag & NCF_HOTNEGATIVE) {
738 		mtx_assert(&ncneg_hot.nl_lock, MA_OWNED);
739 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
740 	} else {
741 		mtx_assert(&neglist->nl_lock, MA_OWNED);
742 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
743 	}
744 	if (list_locked)
745 		mtx_unlock(&neglist->nl_lock);
746 	if (hot_locked)
747 		mtx_unlock(&ncneg_hot.nl_lock);
748 	atomic_subtract_rel_long(&numneg, 1);
749 }
750 
751 static void
752 cache_negative_shrink_select(int start, struct namecache **ncpp,
753     struct neglist **neglistpp)
754 {
755 	struct neglist *neglist;
756 	struct namecache *ncp;
757 	int i;
758 
759 	*ncpp = ncp = NULL;
760 	neglist = NULL;
761 
762 	for (i = start; i < numneglists; i++) {
763 		neglist = &neglists[i];
764 		if (TAILQ_FIRST(&neglist->nl_list) == NULL)
765 			continue;
766 		mtx_lock(&neglist->nl_lock);
767 		ncp = TAILQ_FIRST(&neglist->nl_list);
768 		if (ncp != NULL)
769 			break;
770 		mtx_unlock(&neglist->nl_lock);
771 	}
772 
773 	*neglistpp = neglist;
774 	*ncpp = ncp;
775 }
776 
777 static void
778 cache_negative_zap_one(void)
779 {
780 	struct namecache *ncp, *ncp2;
781 	struct neglist *neglist;
782 	struct mtx *dvlp;
783 	struct rwlock *blp;
784 
785 	if (!mtx_trylock(&ncneg_shrink_lock))
786 		return;
787 
788 	mtx_lock(&ncneg_hot.nl_lock);
789 	ncp = TAILQ_FIRST(&ncneg_hot.nl_list);
790 	if (ncp != NULL) {
791 		neglist = NCP2NEGLIST(ncp);
792 		mtx_lock(&neglist->nl_lock);
793 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
794 		TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
795 		ncp->nc_flag &= ~NCF_HOTNEGATIVE;
796 		mtx_unlock(&neglist->nl_lock);
797 	}
798 
799 	cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
800 	shrink_list_turn++;
801 	if (shrink_list_turn == numneglists)
802 		shrink_list_turn = 0;
803 	if (ncp == NULL && shrink_list_turn == 0)
804 		cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
805 	if (ncp == NULL) {
806 		mtx_unlock(&ncneg_hot.nl_lock);
807 		goto out;
808 	}
809 
810 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
811 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
812 	blp = NCP2BUCKETLOCK(ncp);
813 	mtx_unlock(&neglist->nl_lock);
814 	mtx_unlock(&ncneg_hot.nl_lock);
815 	mtx_lock(dvlp);
816 	rw_wlock(blp);
817 	mtx_lock(&neglist->nl_lock);
818 	ncp2 = TAILQ_FIRST(&neglist->nl_list);
819 	if (ncp != ncp2 || dvlp != VP2VNODELOCK(ncp2->nc_dvp) ||
820 	    blp != NCP2BUCKETLOCK(ncp2) || !(ncp2->nc_flag & NCF_NEGATIVE)) {
821 		ncp = NULL;
822 		goto out_unlock_all;
823 	}
824 	SDT_PROBE3(vfs, namecache, shrink_negative, done, ncp->nc_dvp,
825 	    ncp->nc_name, ncp->nc_neghits);
826 
827 	cache_zap_locked(ncp, true);
828 out_unlock_all:
829 	mtx_unlock(&neglist->nl_lock);
830 	rw_wunlock(blp);
831 	mtx_unlock(dvlp);
832 out:
833 	mtx_unlock(&ncneg_shrink_lock);
834 	cache_free(ncp);
835 }
836 
837 /*
838  * cache_zap_locked():
839  *
840  *   Removes a namecache entry from cache, whether it contains an actual
841  *   pointer to a vnode or if it is just a negative cache entry.
842  */
843 static void
844 cache_zap_locked(struct namecache *ncp, bool neg_locked)
845 {
846 
847 	if (!(ncp->nc_flag & NCF_NEGATIVE))
848 		cache_assert_vnode_locked(ncp->nc_vp);
849 	cache_assert_vnode_locked(ncp->nc_dvp);
850 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
851 
852 	CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp,
853 	    (ncp->nc_flag & NCF_NEGATIVE) ? NULL : ncp->nc_vp);
854 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
855 		SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
856 		    ncp->nc_name, ncp->nc_vp);
857 	} else {
858 		SDT_PROBE3(vfs, namecache, zap_negative, done, ncp->nc_dvp,
859 		    ncp->nc_name, ncp->nc_neghits);
860 	}
861 	LIST_REMOVE(ncp, nc_hash);
862 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
863 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
864 		if (ncp == ncp->nc_vp->v_cache_dd)
865 			ncp->nc_vp->v_cache_dd = NULL;
866 	} else {
867 		cache_negative_remove(ncp, neg_locked);
868 	}
869 	if (ncp->nc_flag & NCF_ISDOTDOT) {
870 		if (ncp == ncp->nc_dvp->v_cache_dd)
871 			ncp->nc_dvp->v_cache_dd = NULL;
872 	} else {
873 		LIST_REMOVE(ncp, nc_src);
874 		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
875 			ncp->nc_flag |= NCF_DVDROP;
876 			atomic_subtract_rel_long(&numcachehv, 1);
877 		}
878 	}
879 	atomic_subtract_rel_long(&numcache, 1);
880 }
881 
882 static void
883 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
884 {
885 	struct rwlock *blp;
886 
887 	MPASS(ncp->nc_dvp == vp);
888 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
889 	cache_assert_vnode_locked(vp);
890 
891 	blp = NCP2BUCKETLOCK(ncp);
892 	rw_wlock(blp);
893 	cache_zap_locked(ncp, false);
894 	rw_wunlock(blp);
895 }
896 
897 static bool
898 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
899     struct mtx **vlpp)
900 {
901 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
902 	struct rwlock *blp;
903 
904 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
905 	cache_assert_vnode_locked(vp);
906 
907 	if (ncp->nc_flag & NCF_NEGATIVE) {
908 		if (*vlpp != NULL) {
909 			mtx_unlock(*vlpp);
910 			*vlpp = NULL;
911 		}
912 		cache_zap_negative_locked_vnode_kl(ncp, vp);
913 		return (true);
914 	}
915 
916 	pvlp = VP2VNODELOCK(vp);
917 	blp = NCP2BUCKETLOCK(ncp);
918 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
919 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
920 
921 	if (*vlpp == vlp1 || *vlpp == vlp2) {
922 		to_unlock = *vlpp;
923 		*vlpp = NULL;
924 	} else {
925 		if (*vlpp != NULL) {
926 			mtx_unlock(*vlpp);
927 			*vlpp = NULL;
928 		}
929 		cache_sort(&vlp1, &vlp2);
930 		if (vlp1 == pvlp) {
931 			mtx_lock(vlp2);
932 			to_unlock = vlp2;
933 		} else {
934 			if (!mtx_trylock(vlp1))
935 				goto out_relock;
936 			to_unlock = vlp1;
937 		}
938 	}
939 	rw_wlock(blp);
940 	cache_zap_locked(ncp, false);
941 	rw_wunlock(blp);
942 	if (to_unlock != NULL)
943 		mtx_unlock(to_unlock);
944 	return (true);
945 
946 out_relock:
947 	mtx_unlock(vlp2);
948 	mtx_lock(vlp1);
949 	mtx_lock(vlp2);
950 	MPASS(*vlpp == NULL);
951 	*vlpp = vlp1;
952 	return (false);
953 }
954 
955 static int
956 cache_zap_locked_vnode(struct namecache *ncp, struct vnode *vp)
957 {
958 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
959 	struct rwlock *blp;
960 	int error = 0;
961 
962 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
963 	cache_assert_vnode_locked(vp);
964 
965 	pvlp = VP2VNODELOCK(vp);
966 	if (ncp->nc_flag & NCF_NEGATIVE) {
967 		cache_zap_negative_locked_vnode_kl(ncp, vp);
968 		goto out;
969 	}
970 
971 	blp = NCP2BUCKETLOCK(ncp);
972 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
973 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
974 	cache_sort(&vlp1, &vlp2);
975 	if (vlp1 == pvlp) {
976 		mtx_lock(vlp2);
977 		to_unlock = vlp2;
978 	} else {
979 		if (!mtx_trylock(vlp1)) {
980 			error = EAGAIN;
981 			goto out;
982 		}
983 		to_unlock = vlp1;
984 	}
985 	rw_wlock(blp);
986 	cache_zap_locked(ncp, false);
987 	rw_wunlock(blp);
988 	mtx_unlock(to_unlock);
989 out:
990 	mtx_unlock(pvlp);
991 	return (error);
992 }
993 
994 static int
995 cache_zap_wlocked_bucket(struct namecache *ncp, struct rwlock *blp)
996 {
997 	struct mtx *dvlp, *vlp;
998 
999 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1000 
1001 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1002 	vlp = NULL;
1003 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1004 		vlp = VP2VNODELOCK(ncp->nc_vp);
1005 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1006 		cache_zap_locked(ncp, false);
1007 		rw_wunlock(blp);
1008 		cache_unlock_vnodes(dvlp, vlp);
1009 		return (0);
1010 	}
1011 
1012 	rw_wunlock(blp);
1013 	return (EAGAIN);
1014 }
1015 
1016 static int
1017 cache_zap_rlocked_bucket(struct namecache *ncp, struct rwlock *blp)
1018 {
1019 	struct mtx *dvlp, *vlp;
1020 
1021 	cache_assert_bucket_locked(ncp, RA_RLOCKED);
1022 
1023 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1024 	vlp = NULL;
1025 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1026 		vlp = VP2VNODELOCK(ncp->nc_vp);
1027 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1028 		rw_runlock(blp);
1029 		rw_wlock(blp);
1030 		cache_zap_locked(ncp, false);
1031 		rw_wunlock(blp);
1032 		cache_unlock_vnodes(dvlp, vlp);
1033 		return (0);
1034 	}
1035 
1036 	rw_runlock(blp);
1037 	return (EAGAIN);
1038 }
1039 
1040 static int
1041 cache_zap_wlocked_bucket_kl(struct namecache *ncp, struct rwlock *blp,
1042     struct mtx **vlpp1, struct mtx **vlpp2)
1043 {
1044 	struct mtx *dvlp, *vlp;
1045 
1046 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1047 
1048 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1049 	vlp = NULL;
1050 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1051 		vlp = VP2VNODELOCK(ncp->nc_vp);
1052 	cache_sort(&dvlp, &vlp);
1053 
1054 	if (*vlpp1 == dvlp && *vlpp2 == vlp) {
1055 		cache_zap_locked(ncp, false);
1056 		cache_unlock_vnodes(dvlp, vlp);
1057 		*vlpp1 = NULL;
1058 		*vlpp2 = NULL;
1059 		return (0);
1060 	}
1061 
1062 	if (*vlpp1 != NULL)
1063 		mtx_unlock(*vlpp1);
1064 	if (*vlpp2 != NULL)
1065 		mtx_unlock(*vlpp2);
1066 	*vlpp1 = NULL;
1067 	*vlpp2 = NULL;
1068 
1069 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1070 		cache_zap_locked(ncp, false);
1071 		cache_unlock_vnodes(dvlp, vlp);
1072 		return (0);
1073 	}
1074 
1075 	rw_wunlock(blp);
1076 	*vlpp1 = dvlp;
1077 	*vlpp2 = vlp;
1078 	if (*vlpp1 != NULL)
1079 		mtx_lock(*vlpp1);
1080 	mtx_lock(*vlpp2);
1081 	rw_wlock(blp);
1082 	return (EAGAIN);
1083 }
1084 
1085 static void
1086 cache_lookup_unlock(struct rwlock *blp, struct mtx *vlp)
1087 {
1088 
1089 	if (blp != NULL) {
1090 		rw_runlock(blp);
1091 	} else {
1092 		mtx_unlock(vlp);
1093 	}
1094 }
1095 
1096 static int __noinline
1097 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1098     struct timespec *tsp, int *ticksp)
1099 {
1100 	int ltype;
1101 
1102 	*vpp = dvp;
1103 	CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .",
1104 			dvp, cnp->cn_nameptr);
1105 	counter_u64_add(dothits, 1);
1106 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1107 	if (tsp != NULL)
1108 		timespecclear(tsp);
1109 	if (ticksp != NULL)
1110 		*ticksp = ticks;
1111 	vrefact(*vpp);
1112 	/*
1113 	 * When we lookup "." we still can be asked to lock it
1114 	 * differently...
1115 	 */
1116 	ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1117 	if (ltype != VOP_ISLOCKED(*vpp)) {
1118 		if (ltype == LK_EXCLUSIVE) {
1119 			vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1120 			if ((*vpp)->v_iflag & VI_DOOMED) {
1121 				/* forced unmount */
1122 				vrele(*vpp);
1123 				*vpp = NULL;
1124 				return (ENOENT);
1125 			}
1126 		} else
1127 			vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1128 	}
1129 	return (-1);
1130 }
1131 
1132 /*
1133  * Lookup an entry in the cache
1134  *
1135  * Lookup is called with dvp pointing to the directory to search,
1136  * cnp pointing to the name of the entry being sought. If the lookup
1137  * succeeds, the vnode is returned in *vpp, and a status of -1 is
1138  * returned. If the lookup determines that the name does not exist
1139  * (negative caching), a status of ENOENT is returned. If the lookup
1140  * fails, a status of zero is returned.  If the directory vnode is
1141  * recycled out from under us due to a forced unmount, a status of
1142  * ENOENT is returned.
1143  *
1144  * vpp is locked and ref'd on return.  If we're looking up DOTDOT, dvp is
1145  * unlocked.  If we're looking up . an extra ref is taken, but the lock is
1146  * not recursively acquired.
1147  */
1148 
1149 static __noinline int
1150 cache_lookup_nomakeentry(struct vnode *dvp, struct vnode **vpp,
1151     struct componentname *cnp, struct timespec *tsp, int *ticksp)
1152 {
1153 	struct namecache *ncp;
1154 	struct rwlock *blp;
1155 	struct mtx *dvlp, *dvlp2;
1156 	uint32_t hash;
1157 	int error;
1158 
1159 	if (cnp->cn_namelen == 2 &&
1160 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1161 		counter_u64_add(dotdothits, 1);
1162 		dvlp = VP2VNODELOCK(dvp);
1163 		dvlp2 = NULL;
1164 		mtx_lock(dvlp);
1165 retry_dotdot:
1166 		ncp = dvp->v_cache_dd;
1167 		if (ncp == NULL) {
1168 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1169 			    "..", NULL);
1170 			mtx_unlock(dvlp);
1171 			if (dvlp2 != NULL)
1172 				mtx_unlock(dvlp2);
1173 			return (0);
1174 		}
1175 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1176 			if (ncp->nc_dvp != dvp)
1177 				panic("dvp %p v_cache_dd %p\n", dvp, ncp);
1178 			if (!cache_zap_locked_vnode_kl2(ncp,
1179 			    dvp, &dvlp2))
1180 				goto retry_dotdot;
1181 			MPASS(dvp->v_cache_dd == NULL);
1182 			mtx_unlock(dvlp);
1183 			if (dvlp2 != NULL)
1184 				mtx_unlock(dvlp2);
1185 			cache_free(ncp);
1186 		} else {
1187 			dvp->v_cache_dd = NULL;
1188 			mtx_unlock(dvlp);
1189 			if (dvlp2 != NULL)
1190 				mtx_unlock(dvlp2);
1191 		}
1192 		return (0);
1193 	}
1194 
1195 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1196 	blp = HASH2BUCKETLOCK(hash);
1197 retry:
1198 	if (LIST_EMPTY(NCHHASH(hash)))
1199 		goto out_no_entry;
1200 
1201 	rw_wlock(blp);
1202 
1203 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1204 		counter_u64_add(numchecks, 1);
1205 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1206 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1207 			break;
1208 	}
1209 
1210 	/* We failed to find an entry */
1211 	if (ncp == NULL) {
1212 		rw_wunlock(blp);
1213 		goto out_no_entry;
1214 	}
1215 
1216 	counter_u64_add(numposzaps, 1);
1217 
1218 	error = cache_zap_wlocked_bucket(ncp, blp);
1219 	if (error != 0) {
1220 		zap_and_exit_bucket_fail++;
1221 		cache_maybe_yield();
1222 		goto retry;
1223 	}
1224 	cache_free(ncp);
1225 	return (0);
1226 out_no_entry:
1227 	SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr, NULL);
1228 	counter_u64_add(nummisszap, 1);
1229 	return (0);
1230 }
1231 
1232 int
1233 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1234     struct timespec *tsp, int *ticksp)
1235 {
1236 	struct namecache_ts *ncp_ts;
1237 	struct namecache *ncp;
1238 	struct rwlock *blp;
1239 	struct mtx *dvlp;
1240 	uint32_t hash;
1241 	int error, ltype;
1242 
1243 	if (__predict_false(!doingcache)) {
1244 		cnp->cn_flags &= ~MAKEENTRY;
1245 		return (0);
1246 	}
1247 
1248 	counter_u64_add(numcalls, 1);
1249 
1250 	if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.'))
1251 		return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1252 
1253 	if ((cnp->cn_flags & MAKEENTRY) == 0)
1254 		return (cache_lookup_nomakeentry(dvp, vpp, cnp, tsp, ticksp));
1255 
1256 retry:
1257 	blp = NULL;
1258 	dvlp = NULL;
1259 	error = 0;
1260 	if (cnp->cn_namelen == 2 &&
1261 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1262 		counter_u64_add(dotdothits, 1);
1263 		dvlp = VP2VNODELOCK(dvp);
1264 		mtx_lock(dvlp);
1265 		ncp = dvp->v_cache_dd;
1266 		if (ncp == NULL) {
1267 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1268 			    "..", NULL);
1269 			mtx_unlock(dvlp);
1270 			return (0);
1271 		}
1272 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1273 			if (ncp->nc_flag & NCF_NEGATIVE)
1274 				*vpp = NULL;
1275 			else
1276 				*vpp = ncp->nc_vp;
1277 		} else
1278 			*vpp = ncp->nc_dvp;
1279 		/* Return failure if negative entry was found. */
1280 		if (*vpp == NULL)
1281 			goto negative_success;
1282 		CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
1283 		    dvp, cnp->cn_nameptr, *vpp);
1284 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..",
1285 		    *vpp);
1286 		cache_out_ts(ncp, tsp, ticksp);
1287 		if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1288 		    NCF_DTS && tsp != NULL) {
1289 			ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1290 			*tsp = ncp_ts->nc_dotdottime;
1291 		}
1292 		goto success;
1293 	}
1294 
1295 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1296 	blp = HASH2BUCKETLOCK(hash);
1297 	rw_rlock(blp);
1298 
1299 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1300 		counter_u64_add(numchecks, 1);
1301 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1302 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1303 			break;
1304 	}
1305 
1306 	/* We failed to find an entry */
1307 	if (ncp == NULL) {
1308 		rw_runlock(blp);
1309 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1310 		    NULL);
1311 		counter_u64_add(nummiss, 1);
1312 		return (0);
1313 	}
1314 
1315 	/* We found a "positive" match, return the vnode */
1316 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1317 		counter_u64_add(numposhits, 1);
1318 		*vpp = ncp->nc_vp;
1319 		CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p",
1320 		    dvp, cnp->cn_nameptr, *vpp, ncp);
1321 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name,
1322 		    *vpp);
1323 		cache_out_ts(ncp, tsp, ticksp);
1324 		goto success;
1325 	}
1326 
1327 negative_success:
1328 	/* We found a negative match, and want to create it, so purge */
1329 	if (cnp->cn_nameiop == CREATE) {
1330 		counter_u64_add(numnegzaps, 1);
1331 		goto zap_and_exit;
1332 	}
1333 
1334 	counter_u64_add(numneghits, 1);
1335 	cache_negative_hit(ncp);
1336 	if (ncp->nc_flag & NCF_WHITE)
1337 		cnp->cn_flags |= ISWHITEOUT;
1338 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp,
1339 	    ncp->nc_name);
1340 	cache_out_ts(ncp, tsp, ticksp);
1341 	cache_lookup_unlock(blp, dvlp);
1342 	return (ENOENT);
1343 
1344 success:
1345 	/*
1346 	 * On success we return a locked and ref'd vnode as per the lookup
1347 	 * protocol.
1348 	 */
1349 	MPASS(dvp != *vpp);
1350 	ltype = 0;	/* silence gcc warning */
1351 	if (cnp->cn_flags & ISDOTDOT) {
1352 		ltype = VOP_ISLOCKED(dvp);
1353 		VOP_UNLOCK(dvp, 0);
1354 	}
1355 	vhold(*vpp);
1356 	cache_lookup_unlock(blp, dvlp);
1357 	error = vget(*vpp, cnp->cn_lkflags | LK_VNHELD, cnp->cn_thread);
1358 	if (cnp->cn_flags & ISDOTDOT) {
1359 		vn_lock(dvp, ltype | LK_RETRY);
1360 		if (dvp->v_iflag & VI_DOOMED) {
1361 			if (error == 0)
1362 				vput(*vpp);
1363 			*vpp = NULL;
1364 			return (ENOENT);
1365 		}
1366 	}
1367 	if (error) {
1368 		*vpp = NULL;
1369 		goto retry;
1370 	}
1371 	if ((cnp->cn_flags & ISLASTCN) &&
1372 	    (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) {
1373 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
1374 	}
1375 	return (-1);
1376 
1377 zap_and_exit:
1378 	if (blp != NULL)
1379 		error = cache_zap_rlocked_bucket(ncp, blp);
1380 	else
1381 		error = cache_zap_locked_vnode(ncp, dvp);
1382 	if (error != 0) {
1383 		zap_and_exit_bucket_fail++;
1384 		cache_maybe_yield();
1385 		goto retry;
1386 	}
1387 	cache_free(ncp);
1388 	return (0);
1389 }
1390 
1391 struct celockstate {
1392 	struct mtx *vlp[3];
1393 	struct rwlock *blp[2];
1394 };
1395 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1396 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1397 
1398 static inline void
1399 cache_celockstate_init(struct celockstate *cel)
1400 {
1401 
1402 	bzero(cel, sizeof(*cel));
1403 }
1404 
1405 static void
1406 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1407     struct vnode *dvp)
1408 {
1409 	struct mtx *vlp1, *vlp2;
1410 
1411 	MPASS(cel->vlp[0] == NULL);
1412 	MPASS(cel->vlp[1] == NULL);
1413 	MPASS(cel->vlp[2] == NULL);
1414 
1415 	MPASS(vp != NULL || dvp != NULL);
1416 
1417 	vlp1 = VP2VNODELOCK(vp);
1418 	vlp2 = VP2VNODELOCK(dvp);
1419 	cache_sort(&vlp1, &vlp2);
1420 
1421 	if (vlp1 != NULL) {
1422 		mtx_lock(vlp1);
1423 		cel->vlp[0] = vlp1;
1424 	}
1425 	mtx_lock(vlp2);
1426 	cel->vlp[1] = vlp2;
1427 }
1428 
1429 static void
1430 cache_unlock_vnodes_cel(struct celockstate *cel)
1431 {
1432 
1433 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
1434 
1435 	if (cel->vlp[0] != NULL)
1436 		mtx_unlock(cel->vlp[0]);
1437 	if (cel->vlp[1] != NULL)
1438 		mtx_unlock(cel->vlp[1]);
1439 	if (cel->vlp[2] != NULL)
1440 		mtx_unlock(cel->vlp[2]);
1441 }
1442 
1443 static bool
1444 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
1445 {
1446 	struct mtx *vlp;
1447 	bool ret;
1448 
1449 	cache_assert_vlp_locked(cel->vlp[0]);
1450 	cache_assert_vlp_locked(cel->vlp[1]);
1451 	MPASS(cel->vlp[2] == NULL);
1452 
1453 	MPASS(vp != NULL);
1454 	vlp = VP2VNODELOCK(vp);
1455 
1456 	ret = true;
1457 	if (vlp >= cel->vlp[1]) {
1458 		mtx_lock(vlp);
1459 	} else {
1460 		if (mtx_trylock(vlp))
1461 			goto out;
1462 		cache_lock_vnodes_cel_3_failures++;
1463 		cache_unlock_vnodes_cel(cel);
1464 		if (vlp < cel->vlp[0]) {
1465 			mtx_lock(vlp);
1466 			mtx_lock(cel->vlp[0]);
1467 			mtx_lock(cel->vlp[1]);
1468 		} else {
1469 			if (cel->vlp[0] != NULL)
1470 				mtx_lock(cel->vlp[0]);
1471 			mtx_lock(vlp);
1472 			mtx_lock(cel->vlp[1]);
1473 		}
1474 		ret = false;
1475 	}
1476 out:
1477 	cel->vlp[2] = vlp;
1478 	return (ret);
1479 }
1480 
1481 static void
1482 cache_lock_buckets_cel(struct celockstate *cel, struct rwlock *blp1,
1483     struct rwlock *blp2)
1484 {
1485 
1486 	MPASS(cel->blp[0] == NULL);
1487 	MPASS(cel->blp[1] == NULL);
1488 
1489 	cache_sort(&blp1, &blp2);
1490 
1491 	if (blp1 != NULL) {
1492 		rw_wlock(blp1);
1493 		cel->blp[0] = blp1;
1494 	}
1495 	rw_wlock(blp2);
1496 	cel->blp[1] = blp2;
1497 }
1498 
1499 static void
1500 cache_unlock_buckets_cel(struct celockstate *cel)
1501 {
1502 
1503 	if (cel->blp[0] != NULL)
1504 		rw_wunlock(cel->blp[0]);
1505 	rw_wunlock(cel->blp[1]);
1506 }
1507 
1508 /*
1509  * Lock part of the cache affected by the insertion.
1510  *
1511  * This means vnodelocks for dvp, vp and the relevant bucketlock.
1512  * However, insertion can result in removal of an old entry. In this
1513  * case we have an additional vnode and bucketlock pair to lock. If the
1514  * entry is negative, ncelock is locked instead of the vnode.
1515  *
1516  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
1517  * preserving the locking order (smaller address first).
1518  */
1519 static void
1520 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1521     uint32_t hash)
1522 {
1523 	struct namecache *ncp;
1524 	struct rwlock *blps[2];
1525 
1526 	blps[0] = HASH2BUCKETLOCK(hash);
1527 	for (;;) {
1528 		blps[1] = NULL;
1529 		cache_lock_vnodes_cel(cel, dvp, vp);
1530 		if (vp == NULL || vp->v_type != VDIR)
1531 			break;
1532 		ncp = vp->v_cache_dd;
1533 		if (ncp == NULL)
1534 			break;
1535 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1536 			break;
1537 		MPASS(ncp->nc_dvp == vp);
1538 		blps[1] = NCP2BUCKETLOCK(ncp);
1539 		if (ncp->nc_flag & NCF_NEGATIVE)
1540 			break;
1541 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1542 			break;
1543 		/*
1544 		 * All vnodes got re-locked. Re-validate the state and if
1545 		 * nothing changed we are done. Otherwise restart.
1546 		 */
1547 		if (ncp == vp->v_cache_dd &&
1548 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1549 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1550 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1551 			break;
1552 		cache_unlock_vnodes_cel(cel);
1553 		cel->vlp[0] = NULL;
1554 		cel->vlp[1] = NULL;
1555 		cel->vlp[2] = NULL;
1556 	}
1557 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1558 }
1559 
1560 static void
1561 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1562     uint32_t hash)
1563 {
1564 	struct namecache *ncp;
1565 	struct rwlock *blps[2];
1566 
1567 	blps[0] = HASH2BUCKETLOCK(hash);
1568 	for (;;) {
1569 		blps[1] = NULL;
1570 		cache_lock_vnodes_cel(cel, dvp, vp);
1571 		ncp = dvp->v_cache_dd;
1572 		if (ncp == NULL)
1573 			break;
1574 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1575 			break;
1576 		MPASS(ncp->nc_dvp == dvp);
1577 		blps[1] = NCP2BUCKETLOCK(ncp);
1578 		if (ncp->nc_flag & NCF_NEGATIVE)
1579 			break;
1580 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1581 			break;
1582 		if (ncp == dvp->v_cache_dd &&
1583 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1584 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1585 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1586 			break;
1587 		cache_unlock_vnodes_cel(cel);
1588 		cel->vlp[0] = NULL;
1589 		cel->vlp[1] = NULL;
1590 		cel->vlp[2] = NULL;
1591 	}
1592 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1593 }
1594 
1595 static void
1596 cache_enter_unlock(struct celockstate *cel)
1597 {
1598 
1599 	cache_unlock_buckets_cel(cel);
1600 	cache_unlock_vnodes_cel(cel);
1601 }
1602 
1603 /*
1604  * Add an entry to the cache.
1605  */
1606 void
1607 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
1608     struct timespec *tsp, struct timespec *dtsp)
1609 {
1610 	struct celockstate cel;
1611 	struct namecache *ncp, *n2, *ndd;
1612 	struct namecache_ts *ncp_ts, *n2_ts;
1613 	struct nchashhead *ncpp;
1614 	struct neglist *neglist;
1615 	uint32_t hash;
1616 	int flag;
1617 	int len;
1618 	bool neg_locked;
1619 	int lnumcache;
1620 
1621 	CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr);
1622 	VNASSERT(vp == NULL || (vp->v_iflag & VI_DOOMED) == 0, vp,
1623 	    ("cache_enter: Adding a doomed vnode"));
1624 	VNASSERT(dvp == NULL || (dvp->v_iflag & VI_DOOMED) == 0, dvp,
1625 	    ("cache_enter: Doomed vnode used as src"));
1626 
1627 	if (__predict_false(!doingcache))
1628 		return;
1629 
1630 	/*
1631 	 * Avoid blowout in namecache entries.
1632 	 */
1633 	if (__predict_false(numcache >= desiredvnodes * ncsizefactor))
1634 		return;
1635 
1636 	cache_celockstate_init(&cel);
1637 	ndd = NULL;
1638 	ncp_ts = NULL;
1639 	flag = 0;
1640 	if (cnp->cn_nameptr[0] == '.') {
1641 		if (cnp->cn_namelen == 1)
1642 			return;
1643 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1644 			len = cnp->cn_namelen;
1645 			hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1646 			cache_enter_lock_dd(&cel, dvp, vp, hash);
1647 			/*
1648 			 * If dotdot entry already exists, just retarget it
1649 			 * to new parent vnode, otherwise continue with new
1650 			 * namecache entry allocation.
1651 			 */
1652 			if ((ncp = dvp->v_cache_dd) != NULL &&
1653 			    ncp->nc_flag & NCF_ISDOTDOT) {
1654 				KASSERT(ncp->nc_dvp == dvp,
1655 				    ("wrong isdotdot parent"));
1656 				neg_locked = false;
1657 				if (ncp->nc_flag & NCF_NEGATIVE || vp == NULL) {
1658 					neglist = NCP2NEGLIST(ncp);
1659 					mtx_lock(&ncneg_hot.nl_lock);
1660 					mtx_lock(&neglist->nl_lock);
1661 					neg_locked = true;
1662 				}
1663 				if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1664 					TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst,
1665 					    ncp, nc_dst);
1666 				} else {
1667 					cache_negative_remove(ncp, true);
1668 				}
1669 				if (vp != NULL) {
1670 					TAILQ_INSERT_HEAD(&vp->v_cache_dst,
1671 					    ncp, nc_dst);
1672 					ncp->nc_flag &= ~(NCF_NEGATIVE|NCF_HOTNEGATIVE);
1673 				} else {
1674 					ncp->nc_flag &= ~(NCF_HOTNEGATIVE);
1675 					ncp->nc_flag |= NCF_NEGATIVE;
1676 					cache_negative_insert(ncp, true);
1677 				}
1678 				if (neg_locked) {
1679 					mtx_unlock(&neglist->nl_lock);
1680 					mtx_unlock(&ncneg_hot.nl_lock);
1681 				}
1682 				ncp->nc_vp = vp;
1683 				cache_enter_unlock(&cel);
1684 				return;
1685 			}
1686 			dvp->v_cache_dd = NULL;
1687 			cache_enter_unlock(&cel);
1688 			cache_celockstate_init(&cel);
1689 			SDT_PROBE3(vfs, namecache, enter, done, dvp, "..", vp);
1690 			flag = NCF_ISDOTDOT;
1691 		}
1692 	}
1693 
1694 	/*
1695 	 * Calculate the hash key and setup as much of the new
1696 	 * namecache entry as possible before acquiring the lock.
1697 	 */
1698 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
1699 	ncp->nc_flag = flag;
1700 	ncp->nc_vp = vp;
1701 	if (vp == NULL)
1702 		ncp->nc_flag |= NCF_NEGATIVE;
1703 	ncp->nc_dvp = dvp;
1704 	if (tsp != NULL) {
1705 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1706 		ncp_ts->nc_time = *tsp;
1707 		ncp_ts->nc_ticks = ticks;
1708 		ncp_ts->nc_nc.nc_flag |= NCF_TS;
1709 		if (dtsp != NULL) {
1710 			ncp_ts->nc_dotdottime = *dtsp;
1711 			ncp_ts->nc_nc.nc_flag |= NCF_DTS;
1712 		}
1713 	}
1714 	len = ncp->nc_nlen = cnp->cn_namelen;
1715 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1716 	strlcpy(ncp->nc_name, cnp->cn_nameptr, len + 1);
1717 	cache_enter_lock(&cel, dvp, vp, hash);
1718 
1719 	/*
1720 	 * See if this vnode or negative entry is already in the cache
1721 	 * with this name.  This can happen with concurrent lookups of
1722 	 * the same path name.
1723 	 */
1724 	ncpp = NCHHASH(hash);
1725 	LIST_FOREACH(n2, ncpp, nc_hash) {
1726 		if (n2->nc_dvp == dvp &&
1727 		    n2->nc_nlen == cnp->cn_namelen &&
1728 		    !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
1729 			if (tsp != NULL) {
1730 				KASSERT((n2->nc_flag & NCF_TS) != 0,
1731 				    ("no NCF_TS"));
1732 				n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
1733 				n2_ts->nc_time = ncp_ts->nc_time;
1734 				n2_ts->nc_ticks = ncp_ts->nc_ticks;
1735 				if (dtsp != NULL) {
1736 					n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
1737 					if (ncp->nc_flag & NCF_NEGATIVE)
1738 						mtx_lock(&ncneg_hot.nl_lock);
1739 					n2_ts->nc_nc.nc_flag |= NCF_DTS;
1740 					if (ncp->nc_flag & NCF_NEGATIVE)
1741 						mtx_unlock(&ncneg_hot.nl_lock);
1742 				}
1743 			}
1744 			goto out_unlock_free;
1745 		}
1746 	}
1747 
1748 	if (flag == NCF_ISDOTDOT) {
1749 		/*
1750 		 * See if we are trying to add .. entry, but some other lookup
1751 		 * has populated v_cache_dd pointer already.
1752 		 */
1753 		if (dvp->v_cache_dd != NULL)
1754 			goto out_unlock_free;
1755 		KASSERT(vp == NULL || vp->v_type == VDIR,
1756 		    ("wrong vnode type %p", vp));
1757 		dvp->v_cache_dd = ncp;
1758 	}
1759 
1760 	if (vp != NULL) {
1761 		if (vp->v_type == VDIR) {
1762 			if (flag != NCF_ISDOTDOT) {
1763 				/*
1764 				 * For this case, the cache entry maps both the
1765 				 * directory name in it and the name ".." for the
1766 				 * directory's parent.
1767 				 */
1768 				if ((ndd = vp->v_cache_dd) != NULL) {
1769 					if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
1770 						cache_zap_locked(ndd, false);
1771 					else
1772 						ndd = NULL;
1773 				}
1774 				vp->v_cache_dd = ncp;
1775 			}
1776 		} else {
1777 			vp->v_cache_dd = NULL;
1778 		}
1779 	}
1780 
1781 	if (flag != NCF_ISDOTDOT) {
1782 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1783 			vhold(dvp);
1784 			atomic_add_rel_long(&numcachehv, 1);
1785 		}
1786 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
1787 	}
1788 
1789 	/*
1790 	 * Insert the new namecache entry into the appropriate chain
1791 	 * within the cache entries table.
1792 	 */
1793 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
1794 
1795 	/*
1796 	 * If the entry is "negative", we place it into the
1797 	 * "negative" cache queue, otherwise, we place it into the
1798 	 * destination vnode's cache entries queue.
1799 	 */
1800 	if (vp != NULL) {
1801 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
1802 		SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
1803 		    vp);
1804 	} else {
1805 		if (cnp->cn_flags & ISWHITEOUT)
1806 			ncp->nc_flag |= NCF_WHITE;
1807 		cache_negative_insert(ncp, false);
1808 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
1809 		    ncp->nc_name);
1810 	}
1811 	cache_enter_unlock(&cel);
1812 	lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
1813 	if (numneg * ncnegfactor > lnumcache)
1814 		cache_negative_zap_one();
1815 	cache_free(ndd);
1816 	return;
1817 out_unlock_free:
1818 	cache_enter_unlock(&cel);
1819 	cache_free(ncp);
1820 	return;
1821 }
1822 
1823 static u_int
1824 cache_roundup_2(u_int val)
1825 {
1826 	u_int res;
1827 
1828 	for (res = 1; res <= val; res <<= 1)
1829 		continue;
1830 
1831 	return (res);
1832 }
1833 
1834 /*
1835  * Name cache initialization, from vfs_init() when we are booting
1836  */
1837 static void
1838 nchinit(void *dummy __unused)
1839 {
1840 	u_int i;
1841 
1842 	cache_zone_small = uma_zcreate("S VFS Cache",
1843 	    sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1,
1844 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1845 	    UMA_ZONE_ZINIT);
1846 	cache_zone_small_ts = uma_zcreate("STS VFS Cache",
1847 	    sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1,
1848 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1849 	    UMA_ZONE_ZINIT);
1850 	cache_zone_large = uma_zcreate("L VFS Cache",
1851 	    sizeof(struct namecache) + NAME_MAX + 1,
1852 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1853 	    UMA_ZONE_ZINIT);
1854 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache",
1855 	    sizeof(struct namecache_ts) + NAME_MAX + 1,
1856 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1857 	    UMA_ZONE_ZINIT);
1858 
1859 	nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash);
1860 	ncbuckethash = cache_roundup_2(mp_ncpus * 64) - 1;
1861 	if (ncbuckethash > nchash)
1862 		ncbuckethash = nchash;
1863 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
1864 	    M_WAITOK | M_ZERO);
1865 	for (i = 0; i < numbucketlocks; i++)
1866 		rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK | RW_RECURSE);
1867 	ncvnodehash = cache_roundup_2(mp_ncpus * 64) - 1;
1868 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
1869 	    M_WAITOK | M_ZERO);
1870 	for (i = 0; i < numvnodelocks; i++)
1871 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
1872 	ncpurgeminvnodes = numbucketlocks;
1873 
1874 	ncneghash = 3;
1875 	neglists = malloc(sizeof(*neglists) * numneglists, M_VFSCACHE,
1876 	    M_WAITOK | M_ZERO);
1877 	for (i = 0; i < numneglists; i++) {
1878 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
1879 		TAILQ_INIT(&neglists[i].nl_list);
1880 	}
1881 	mtx_init(&ncneg_hot.nl_lock, "ncneglh", NULL, MTX_DEF);
1882 	TAILQ_INIT(&ncneg_hot.nl_list);
1883 
1884 	mtx_init(&ncneg_shrink_lock, "ncnegs", NULL, MTX_DEF);
1885 
1886 	numcalls = counter_u64_alloc(M_WAITOK);
1887 	dothits = counter_u64_alloc(M_WAITOK);
1888 	dotdothits = counter_u64_alloc(M_WAITOK);
1889 	numchecks = counter_u64_alloc(M_WAITOK);
1890 	nummiss = counter_u64_alloc(M_WAITOK);
1891 	nummisszap = counter_u64_alloc(M_WAITOK);
1892 	numposzaps = counter_u64_alloc(M_WAITOK);
1893 	numposhits = counter_u64_alloc(M_WAITOK);
1894 	numnegzaps = counter_u64_alloc(M_WAITOK);
1895 	numneghits = counter_u64_alloc(M_WAITOK);
1896 	numfullpathcalls = counter_u64_alloc(M_WAITOK);
1897 	numfullpathfail1 = counter_u64_alloc(M_WAITOK);
1898 	numfullpathfail2 = counter_u64_alloc(M_WAITOK);
1899 	numfullpathfail4 = counter_u64_alloc(M_WAITOK);
1900 	numfullpathfound = counter_u64_alloc(M_WAITOK);
1901 }
1902 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
1903 
1904 void
1905 cache_changesize(int newmaxvnodes)
1906 {
1907 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
1908 	u_long new_nchash, old_nchash;
1909 	struct namecache *ncp;
1910 	uint32_t hash;
1911 	int i;
1912 
1913 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
1914 	if (newmaxvnodes < numbucketlocks)
1915 		newmaxvnodes = numbucketlocks;
1916 
1917 	new_nchashtbl = hashinit(newmaxvnodes, M_VFSCACHE, &new_nchash);
1918 	/* If same hash table size, nothing to do */
1919 	if (nchash == new_nchash) {
1920 		free(new_nchashtbl, M_VFSCACHE);
1921 		return;
1922 	}
1923 	/*
1924 	 * Move everything from the old hash table to the new table.
1925 	 * None of the namecache entries in the table can be removed
1926 	 * because to do so, they have to be removed from the hash table.
1927 	 */
1928 	cache_lock_all_vnodes();
1929 	cache_lock_all_buckets();
1930 	old_nchashtbl = nchashtbl;
1931 	old_nchash = nchash;
1932 	nchashtbl = new_nchashtbl;
1933 	nchash = new_nchash;
1934 	for (i = 0; i <= old_nchash; i++) {
1935 		while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) {
1936 			hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
1937 			    ncp->nc_dvp);
1938 			LIST_REMOVE(ncp, nc_hash);
1939 			LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
1940 		}
1941 	}
1942 	cache_unlock_all_buckets();
1943 	cache_unlock_all_vnodes();
1944 	free(old_nchashtbl, M_VFSCACHE);
1945 }
1946 
1947 /*
1948  * Invalidate all entries to a particular vnode.
1949  */
1950 void
1951 cache_purge(struct vnode *vp)
1952 {
1953 	TAILQ_HEAD(, namecache) ncps;
1954 	struct namecache *ncp, *nnp;
1955 	struct mtx *vlp, *vlp2;
1956 
1957 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
1958 	SDT_PROBE1(vfs, namecache, purge, done, vp);
1959 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
1960 	    vp->v_cache_dd == NULL)
1961 		return;
1962 	TAILQ_INIT(&ncps);
1963 	vlp = VP2VNODELOCK(vp);
1964 	vlp2 = NULL;
1965 	mtx_lock(vlp);
1966 retry:
1967 	while (!LIST_EMPTY(&vp->v_cache_src)) {
1968 		ncp = LIST_FIRST(&vp->v_cache_src);
1969 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
1970 			goto retry;
1971 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1972 	}
1973 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
1974 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
1975 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
1976 			goto retry;
1977 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1978 	}
1979 	ncp = vp->v_cache_dd;
1980 	if (ncp != NULL) {
1981 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
1982 		   ("lost dotdot link"));
1983 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
1984 			goto retry;
1985 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1986 	}
1987 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
1988 	mtx_unlock(vlp);
1989 	if (vlp2 != NULL)
1990 		mtx_unlock(vlp2);
1991 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
1992 		cache_free(ncp);
1993 	}
1994 }
1995 
1996 /*
1997  * Invalidate all negative entries for a particular directory vnode.
1998  */
1999 void
2000 cache_purge_negative(struct vnode *vp)
2001 {
2002 	TAILQ_HEAD(, namecache) ncps;
2003 	struct namecache *ncp, *nnp;
2004 	struct mtx *vlp;
2005 
2006 	CTR1(KTR_VFS, "cache_purge_negative(%p)", vp);
2007 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2008 	if (LIST_EMPTY(&vp->v_cache_src))
2009 		return;
2010 	TAILQ_INIT(&ncps);
2011 	vlp = VP2VNODELOCK(vp);
2012 	mtx_lock(vlp);
2013 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2014 		if (!(ncp->nc_flag & NCF_NEGATIVE))
2015 			continue;
2016 		cache_zap_negative_locked_vnode_kl(ncp, vp);
2017 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2018 	}
2019 	mtx_unlock(vlp);
2020 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2021 		cache_free(ncp);
2022 	}
2023 }
2024 
2025 /*
2026  * Flush all entries referencing a particular filesystem.
2027  */
2028 void
2029 cache_purgevfs(struct mount *mp, bool force)
2030 {
2031 	TAILQ_HEAD(, namecache) ncps;
2032 	struct mtx *vlp1, *vlp2;
2033 	struct rwlock *blp;
2034 	struct nchashhead *bucket;
2035 	struct namecache *ncp, *nnp;
2036 	u_long i, j, n_nchash;
2037 	int error;
2038 
2039 	/* Scan hash tables for applicable entries */
2040 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2041 	if (!force && mp->mnt_nvnodelistsize <= ncpurgeminvnodes)
2042 		return;
2043 	TAILQ_INIT(&ncps);
2044 	n_nchash = nchash + 1;
2045 	vlp1 = vlp2 = NULL;
2046 	for (i = 0; i < numbucketlocks; i++) {
2047 		blp = (struct rwlock *)&bucketlocks[i];
2048 		rw_wlock(blp);
2049 		for (j = i; j < n_nchash; j += numbucketlocks) {
2050 retry:
2051 			bucket = &nchashtbl[j];
2052 			LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) {
2053 				cache_assert_bucket_locked(ncp, RA_WLOCKED);
2054 				if (ncp->nc_dvp->v_mount != mp)
2055 					continue;
2056 				error = cache_zap_wlocked_bucket_kl(ncp, blp,
2057 				    &vlp1, &vlp2);
2058 				if (error != 0)
2059 					goto retry;
2060 				TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst);
2061 			}
2062 		}
2063 		rw_wunlock(blp);
2064 		if (vlp1 == NULL && vlp2 == NULL)
2065 			cache_maybe_yield();
2066 	}
2067 	if (vlp1 != NULL)
2068 		mtx_unlock(vlp1);
2069 	if (vlp2 != NULL)
2070 		mtx_unlock(vlp2);
2071 
2072 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2073 		cache_free(ncp);
2074 	}
2075 }
2076 
2077 /*
2078  * Perform canonical checks and cache lookup and pass on to filesystem
2079  * through the vop_cachedlookup only if needed.
2080  */
2081 
2082 int
2083 vfs_cache_lookup(struct vop_lookup_args *ap)
2084 {
2085 	struct vnode *dvp;
2086 	int error;
2087 	struct vnode **vpp = ap->a_vpp;
2088 	struct componentname *cnp = ap->a_cnp;
2089 	struct ucred *cred = cnp->cn_cred;
2090 	int flags = cnp->cn_flags;
2091 	struct thread *td = cnp->cn_thread;
2092 
2093 	*vpp = NULL;
2094 	dvp = ap->a_dvp;
2095 
2096 	if (dvp->v_type != VDIR)
2097 		return (ENOTDIR);
2098 
2099 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2100 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2101 		return (EROFS);
2102 
2103 	error = VOP_ACCESS(dvp, VEXEC, cred, td);
2104 	if (error)
2105 		return (error);
2106 
2107 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2108 	if (error == 0)
2109 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2110 	if (error == -1)
2111 		return (0);
2112 	return (error);
2113 }
2114 
2115 /*
2116  * XXX All of these sysctls would probably be more productive dead.
2117  */
2118 static int __read_mostly disablecwd;
2119 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
2120    "Disable the getcwd syscall");
2121 
2122 /* Implementation of the getcwd syscall. */
2123 int
2124 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2125 {
2126 
2127 	return (kern___getcwd(td, uap->buf, UIO_USERSPACE, uap->buflen,
2128 	    MAXPATHLEN));
2129 }
2130 
2131 int
2132 kern___getcwd(struct thread *td, char *buf, enum uio_seg bufseg, size_t buflen,
2133     size_t path_max)
2134 {
2135 	char *bp, *tmpbuf;
2136 	struct filedesc *fdp;
2137 	struct vnode *cdir, *rdir;
2138 	int error;
2139 
2140 	if (__predict_false(disablecwd))
2141 		return (ENODEV);
2142 	if (__predict_false(buflen < 2))
2143 		return (EINVAL);
2144 	if (buflen > path_max)
2145 		buflen = path_max;
2146 
2147 	tmpbuf = malloc(buflen, M_TEMP, M_WAITOK);
2148 	fdp = td->td_proc->p_fd;
2149 	FILEDESC_SLOCK(fdp);
2150 	cdir = fdp->fd_cdir;
2151 	vrefact(cdir);
2152 	rdir = fdp->fd_rdir;
2153 	vrefact(rdir);
2154 	FILEDESC_SUNLOCK(fdp);
2155 	error = vn_fullpath1(td, cdir, rdir, tmpbuf, &bp, buflen);
2156 	vrele(rdir);
2157 	vrele(cdir);
2158 
2159 	if (!error) {
2160 		if (bufseg == UIO_SYSSPACE)
2161 			bcopy(bp, buf, strlen(bp) + 1);
2162 		else
2163 			error = copyout(bp, buf, strlen(bp) + 1);
2164 #ifdef KTRACE
2165 	if (KTRPOINT(curthread, KTR_NAMEI))
2166 		ktrnamei(bp);
2167 #endif
2168 	}
2169 	free(tmpbuf, M_TEMP);
2170 	return (error);
2171 }
2172 
2173 /*
2174  * Thus begins the fullpath magic.
2175  */
2176 
2177 static int __read_mostly disablefullpath;
2178 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, &disablefullpath, 0,
2179     "Disable the vn_fullpath function");
2180 
2181 /*
2182  * Retrieve the full filesystem path that correspond to a vnode from the name
2183  * cache (if available)
2184  */
2185 int
2186 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf)
2187 {
2188 	char *buf;
2189 	struct filedesc *fdp;
2190 	struct vnode *rdir;
2191 	int error;
2192 
2193 	if (__predict_false(disablefullpath))
2194 		return (ENODEV);
2195 	if (__predict_false(vn == NULL))
2196 		return (EINVAL);
2197 
2198 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2199 	fdp = td->td_proc->p_fd;
2200 	FILEDESC_SLOCK(fdp);
2201 	rdir = fdp->fd_rdir;
2202 	vrefact(rdir);
2203 	FILEDESC_SUNLOCK(fdp);
2204 	error = vn_fullpath1(td, vn, rdir, buf, retbuf, MAXPATHLEN);
2205 	vrele(rdir);
2206 
2207 	if (!error)
2208 		*freebuf = buf;
2209 	else
2210 		free(buf, M_TEMP);
2211 	return (error);
2212 }
2213 
2214 /*
2215  * This function is similar to vn_fullpath, but it attempts to lookup the
2216  * pathname relative to the global root mount point.  This is required for the
2217  * auditing sub-system, as audited pathnames must be absolute, relative to the
2218  * global root mount point.
2219  */
2220 int
2221 vn_fullpath_global(struct thread *td, struct vnode *vn,
2222     char **retbuf, char **freebuf)
2223 {
2224 	char *buf;
2225 	int error;
2226 
2227 	if (__predict_false(disablefullpath))
2228 		return (ENODEV);
2229 	if (__predict_false(vn == NULL))
2230 		return (EINVAL);
2231 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2232 	error = vn_fullpath1(td, vn, rootvnode, buf, retbuf, MAXPATHLEN);
2233 	if (!error)
2234 		*freebuf = buf;
2235 	else
2236 		free(buf, M_TEMP);
2237 	return (error);
2238 }
2239 
2240 int
2241 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen)
2242 {
2243 	struct vnode *dvp;
2244 	struct namecache *ncp;
2245 	struct mtx *vlp;
2246 	int error;
2247 
2248 	vlp = VP2VNODELOCK(*vp);
2249 	mtx_lock(vlp);
2250 	TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) {
2251 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2252 			break;
2253 	}
2254 	if (ncp != NULL) {
2255 		if (*buflen < ncp->nc_nlen) {
2256 			mtx_unlock(vlp);
2257 			vrele(*vp);
2258 			counter_u64_add(numfullpathfail4, 1);
2259 			error = ENOMEM;
2260 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2261 			    vp, NULL);
2262 			return (error);
2263 		}
2264 		*buflen -= ncp->nc_nlen;
2265 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2266 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2267 		    ncp->nc_name, vp);
2268 		dvp = *vp;
2269 		*vp = ncp->nc_dvp;
2270 		vref(*vp);
2271 		mtx_unlock(vlp);
2272 		vrele(dvp);
2273 		return (0);
2274 	}
2275 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2276 
2277 	mtx_unlock(vlp);
2278 	vn_lock(*vp, LK_SHARED | LK_RETRY);
2279 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
2280 	vput(*vp);
2281 	if (error) {
2282 		counter_u64_add(numfullpathfail2, 1);
2283 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
2284 		return (error);
2285 	}
2286 
2287 	*vp = dvp;
2288 	if (dvp->v_iflag & VI_DOOMED) {
2289 		/* forced unmount */
2290 		vrele(dvp);
2291 		error = ENOENT;
2292 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2293 		return (error);
2294 	}
2295 	/*
2296 	 * *vp has its use count incremented still.
2297 	 */
2298 
2299 	return (0);
2300 }
2301 
2302 /*
2303  * The magic behind kern___getcwd() and vn_fullpath().
2304  */
2305 static int
2306 vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
2307     char *buf, char **retbuf, u_int buflen)
2308 {
2309 	int error, slash_prefixed;
2310 #ifdef KDTRACE_HOOKS
2311 	struct vnode *startvp = vp;
2312 #endif
2313 	struct vnode *vp1;
2314 
2315 	buflen--;
2316 	buf[buflen] = '\0';
2317 	error = 0;
2318 	slash_prefixed = 0;
2319 
2320 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
2321 	counter_u64_add(numfullpathcalls, 1);
2322 	vref(vp);
2323 	if (vp->v_type != VDIR) {
2324 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2325 		if (error)
2326 			return (error);
2327 		if (buflen == 0) {
2328 			vrele(vp);
2329 			return (ENOMEM);
2330 		}
2331 		buf[--buflen] = '/';
2332 		slash_prefixed = 1;
2333 	}
2334 	while (vp != rdir && vp != rootvnode) {
2335 		/*
2336 		 * The vp vnode must be already fully constructed,
2337 		 * since it is either found in namecache or obtained
2338 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
2339 		 * without obtaining the vnode lock.
2340 		 */
2341 		if ((vp->v_vflag & VV_ROOT) != 0) {
2342 			vn_lock(vp, LK_RETRY | LK_SHARED);
2343 
2344 			/*
2345 			 * With the vnode locked, check for races with
2346 			 * unmount, forced or not.  Note that we
2347 			 * already verified that vp is not equal to
2348 			 * the root vnode, which means that
2349 			 * mnt_vnodecovered can be NULL only for the
2350 			 * case of unmount.
2351 			 */
2352 			if ((vp->v_iflag & VI_DOOMED) != 0 ||
2353 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
2354 			    vp1->v_mountedhere != vp->v_mount) {
2355 				vput(vp);
2356 				error = ENOENT;
2357 				SDT_PROBE3(vfs, namecache, fullpath, return,
2358 				    error, vp, NULL);
2359 				break;
2360 			}
2361 
2362 			vref(vp1);
2363 			vput(vp);
2364 			vp = vp1;
2365 			continue;
2366 		}
2367 		if (vp->v_type != VDIR) {
2368 			vrele(vp);
2369 			counter_u64_add(numfullpathfail1, 1);
2370 			error = ENOTDIR;
2371 			SDT_PROBE3(vfs, namecache, fullpath, return,
2372 			    error, vp, NULL);
2373 			break;
2374 		}
2375 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2376 		if (error)
2377 			break;
2378 		if (buflen == 0) {
2379 			vrele(vp);
2380 			error = ENOMEM;
2381 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2382 			    startvp, NULL);
2383 			break;
2384 		}
2385 		buf[--buflen] = '/';
2386 		slash_prefixed = 1;
2387 	}
2388 	if (error)
2389 		return (error);
2390 	if (!slash_prefixed) {
2391 		if (buflen == 0) {
2392 			vrele(vp);
2393 			counter_u64_add(numfullpathfail4, 1);
2394 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
2395 			    startvp, NULL);
2396 			return (ENOMEM);
2397 		}
2398 		buf[--buflen] = '/';
2399 	}
2400 	counter_u64_add(numfullpathfound, 1);
2401 	vrele(vp);
2402 
2403 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, buf + buflen);
2404 	*retbuf = buf + buflen;
2405 	return (0);
2406 }
2407 
2408 struct vnode *
2409 vn_dir_dd_ino(struct vnode *vp)
2410 {
2411 	struct namecache *ncp;
2412 	struct vnode *ddvp;
2413 	struct mtx *vlp;
2414 
2415 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
2416 	vlp = VP2VNODELOCK(vp);
2417 	mtx_lock(vlp);
2418 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
2419 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
2420 			continue;
2421 		ddvp = ncp->nc_dvp;
2422 		vhold(ddvp);
2423 		mtx_unlock(vlp);
2424 		if (vget(ddvp, LK_SHARED | LK_NOWAIT | LK_VNHELD, curthread))
2425 			return (NULL);
2426 		return (ddvp);
2427 	}
2428 	mtx_unlock(vlp);
2429 	return (NULL);
2430 }
2431 
2432 int
2433 vn_commname(struct vnode *vp, char *buf, u_int buflen)
2434 {
2435 	struct namecache *ncp;
2436 	struct mtx *vlp;
2437 	int l;
2438 
2439 	vlp = VP2VNODELOCK(vp);
2440 	mtx_lock(vlp);
2441 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
2442 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2443 			break;
2444 	if (ncp == NULL) {
2445 		mtx_unlock(vlp);
2446 		return (ENOENT);
2447 	}
2448 	l = min(ncp->nc_nlen, buflen - 1);
2449 	memcpy(buf, ncp->nc_name, l);
2450 	mtx_unlock(vlp);
2451 	buf[l] = '\0';
2452 	return (0);
2453 }
2454 
2455 /*
2456  * This function updates path string to vnode's full global path
2457  * and checks the size of the new path string against the pathlen argument.
2458  *
2459  * Requires a locked, referenced vnode.
2460  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
2461  *
2462  * If sysctl debug.disablefullpath is set, ENODEV is returned,
2463  * vnode is left locked and path remain untouched.
2464  *
2465  * If vp is a directory, the call to vn_fullpath_global() always succeeds
2466  * because it falls back to the ".." lookup if the namecache lookup fails.
2467  */
2468 int
2469 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
2470     u_int pathlen)
2471 {
2472 	struct nameidata nd;
2473 	struct vnode *vp1;
2474 	char *rpath, *fbuf;
2475 	int error;
2476 
2477 	ASSERT_VOP_ELOCKED(vp, __func__);
2478 
2479 	/* Return ENODEV if sysctl debug.disablefullpath==1 */
2480 	if (__predict_false(disablefullpath))
2481 		return (ENODEV);
2482 
2483 	/* Construct global filesystem path from vp. */
2484 	VOP_UNLOCK(vp, 0);
2485 	error = vn_fullpath_global(td, vp, &rpath, &fbuf);
2486 
2487 	if (error != 0) {
2488 		vrele(vp);
2489 		return (error);
2490 	}
2491 
2492 	if (strlen(rpath) >= pathlen) {
2493 		vrele(vp);
2494 		error = ENAMETOOLONG;
2495 		goto out;
2496 	}
2497 
2498 	/*
2499 	 * Re-lookup the vnode by path to detect a possible rename.
2500 	 * As a side effect, the vnode is relocked.
2501 	 * If vnode was renamed, return ENOENT.
2502 	 */
2503 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
2504 	    UIO_SYSSPACE, path, td);
2505 	error = namei(&nd);
2506 	if (error != 0) {
2507 		vrele(vp);
2508 		goto out;
2509 	}
2510 	NDFREE(&nd, NDF_ONLY_PNBUF);
2511 	vp1 = nd.ni_vp;
2512 	vrele(vp);
2513 	if (vp1 == vp)
2514 		strcpy(path, rpath);
2515 	else {
2516 		vput(vp1);
2517 		error = ENOENT;
2518 	}
2519 
2520 out:
2521 	free(fbuf, M_TEMP);
2522 	return (error);
2523 }
2524 
2525 #ifdef DDB
2526 static void
2527 db_print_vpath(struct vnode *vp)
2528 {
2529 
2530 	while (vp != NULL) {
2531 		db_printf("%p: ", vp);
2532 		if (vp == rootvnode) {
2533 			db_printf("/");
2534 			vp = NULL;
2535 		} else {
2536 			if (vp->v_vflag & VV_ROOT) {
2537 				db_printf("<mount point>");
2538 				vp = vp->v_mount->mnt_vnodecovered;
2539 			} else {
2540 				struct namecache *ncp;
2541 				char *ncn;
2542 				int i;
2543 
2544 				ncp = TAILQ_FIRST(&vp->v_cache_dst);
2545 				if (ncp != NULL) {
2546 					ncn = ncp->nc_name;
2547 					for (i = 0; i < ncp->nc_nlen; i++)
2548 						db_printf("%c", *ncn++);
2549 					vp = ncp->nc_dvp;
2550 				} else {
2551 					vp = NULL;
2552 				}
2553 			}
2554 		}
2555 		db_printf("\n");
2556 	}
2557 
2558 	return;
2559 }
2560 
2561 DB_SHOW_COMMAND(vpath, db_show_vpath)
2562 {
2563 	struct vnode *vp;
2564 
2565 	if (!have_addr) {
2566 		db_printf("usage: show vpath <struct vnode *>\n");
2567 		return;
2568 	}
2569 
2570 	vp = (struct vnode *)addr;
2571 	db_print_vpath(vp);
2572 }
2573 
2574 #endif
2575