xref: /freebsd/sys/kern/vfs_cache.c (revision 52c2bb75163559a6e2866ad374a7de67a4ea1273)
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 (dvp, name) where dvp 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 	LIST_REMOVE(ncp, nc_hash);
855 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
856 		SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
857 		    ncp->nc_name, ncp->nc_vp);
858 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
859 		if (ncp == ncp->nc_vp->v_cache_dd)
860 			ncp->nc_vp->v_cache_dd = NULL;
861 	} else {
862 		SDT_PROBE3(vfs, namecache, zap_negative, done, ncp->nc_dvp,
863 		    ncp->nc_name, ncp->nc_neghits);
864 		cache_negative_remove(ncp, neg_locked);
865 	}
866 	if (ncp->nc_flag & NCF_ISDOTDOT) {
867 		if (ncp == ncp->nc_dvp->v_cache_dd)
868 			ncp->nc_dvp->v_cache_dd = NULL;
869 	} else {
870 		LIST_REMOVE(ncp, nc_src);
871 		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
872 			ncp->nc_flag |= NCF_DVDROP;
873 			atomic_subtract_rel_long(&numcachehv, 1);
874 		}
875 	}
876 	atomic_subtract_rel_long(&numcache, 1);
877 }
878 
879 static void
880 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
881 {
882 	struct rwlock *blp;
883 
884 	MPASS(ncp->nc_dvp == vp);
885 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
886 	cache_assert_vnode_locked(vp);
887 
888 	blp = NCP2BUCKETLOCK(ncp);
889 	rw_wlock(blp);
890 	cache_zap_locked(ncp, false);
891 	rw_wunlock(blp);
892 }
893 
894 static bool
895 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
896     struct mtx **vlpp)
897 {
898 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
899 	struct rwlock *blp;
900 
901 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
902 	cache_assert_vnode_locked(vp);
903 
904 	if (ncp->nc_flag & NCF_NEGATIVE) {
905 		if (*vlpp != NULL) {
906 			mtx_unlock(*vlpp);
907 			*vlpp = NULL;
908 		}
909 		cache_zap_negative_locked_vnode_kl(ncp, vp);
910 		return (true);
911 	}
912 
913 	pvlp = VP2VNODELOCK(vp);
914 	blp = NCP2BUCKETLOCK(ncp);
915 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
916 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
917 
918 	if (*vlpp == vlp1 || *vlpp == vlp2) {
919 		to_unlock = *vlpp;
920 		*vlpp = NULL;
921 	} else {
922 		if (*vlpp != NULL) {
923 			mtx_unlock(*vlpp);
924 			*vlpp = NULL;
925 		}
926 		cache_sort(&vlp1, &vlp2);
927 		if (vlp1 == pvlp) {
928 			mtx_lock(vlp2);
929 			to_unlock = vlp2;
930 		} else {
931 			if (!mtx_trylock(vlp1))
932 				goto out_relock;
933 			to_unlock = vlp1;
934 		}
935 	}
936 	rw_wlock(blp);
937 	cache_zap_locked(ncp, false);
938 	rw_wunlock(blp);
939 	if (to_unlock != NULL)
940 		mtx_unlock(to_unlock);
941 	return (true);
942 
943 out_relock:
944 	mtx_unlock(vlp2);
945 	mtx_lock(vlp1);
946 	mtx_lock(vlp2);
947 	MPASS(*vlpp == NULL);
948 	*vlpp = vlp1;
949 	return (false);
950 }
951 
952 static int
953 cache_zap_locked_vnode(struct namecache *ncp, struct vnode *vp)
954 {
955 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
956 	struct rwlock *blp;
957 	int error = 0;
958 
959 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
960 	cache_assert_vnode_locked(vp);
961 
962 	pvlp = VP2VNODELOCK(vp);
963 	if (ncp->nc_flag & NCF_NEGATIVE) {
964 		cache_zap_negative_locked_vnode_kl(ncp, vp);
965 		goto out;
966 	}
967 
968 	blp = NCP2BUCKETLOCK(ncp);
969 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
970 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
971 	cache_sort(&vlp1, &vlp2);
972 	if (vlp1 == pvlp) {
973 		mtx_lock(vlp2);
974 		to_unlock = vlp2;
975 	} else {
976 		if (!mtx_trylock(vlp1)) {
977 			error = EAGAIN;
978 			goto out;
979 		}
980 		to_unlock = vlp1;
981 	}
982 	rw_wlock(blp);
983 	cache_zap_locked(ncp, false);
984 	rw_wunlock(blp);
985 	mtx_unlock(to_unlock);
986 out:
987 	mtx_unlock(pvlp);
988 	return (error);
989 }
990 
991 static int
992 cache_zap_wlocked_bucket(struct namecache *ncp, struct rwlock *blp)
993 {
994 	struct mtx *dvlp, *vlp;
995 
996 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
997 
998 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
999 	vlp = NULL;
1000 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1001 		vlp = VP2VNODELOCK(ncp->nc_vp);
1002 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1003 		cache_zap_locked(ncp, false);
1004 		rw_wunlock(blp);
1005 		cache_unlock_vnodes(dvlp, vlp);
1006 		return (0);
1007 	}
1008 
1009 	rw_wunlock(blp);
1010 	return (EAGAIN);
1011 }
1012 
1013 static int
1014 cache_zap_rlocked_bucket(struct namecache *ncp, struct rwlock *blp)
1015 {
1016 	struct mtx *dvlp, *vlp;
1017 
1018 	cache_assert_bucket_locked(ncp, RA_RLOCKED);
1019 
1020 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1021 	vlp = NULL;
1022 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1023 		vlp = VP2VNODELOCK(ncp->nc_vp);
1024 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1025 		rw_runlock(blp);
1026 		rw_wlock(blp);
1027 		cache_zap_locked(ncp, false);
1028 		rw_wunlock(blp);
1029 		cache_unlock_vnodes(dvlp, vlp);
1030 		return (0);
1031 	}
1032 
1033 	rw_runlock(blp);
1034 	return (EAGAIN);
1035 }
1036 
1037 static int
1038 cache_zap_wlocked_bucket_kl(struct namecache *ncp, struct rwlock *blp,
1039     struct mtx **vlpp1, struct mtx **vlpp2)
1040 {
1041 	struct mtx *dvlp, *vlp;
1042 
1043 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1044 
1045 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1046 	vlp = NULL;
1047 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1048 		vlp = VP2VNODELOCK(ncp->nc_vp);
1049 	cache_sort(&dvlp, &vlp);
1050 
1051 	if (*vlpp1 == dvlp && *vlpp2 == vlp) {
1052 		cache_zap_locked(ncp, false);
1053 		cache_unlock_vnodes(dvlp, vlp);
1054 		*vlpp1 = NULL;
1055 		*vlpp2 = NULL;
1056 		return (0);
1057 	}
1058 
1059 	if (*vlpp1 != NULL)
1060 		mtx_unlock(*vlpp1);
1061 	if (*vlpp2 != NULL)
1062 		mtx_unlock(*vlpp2);
1063 	*vlpp1 = NULL;
1064 	*vlpp2 = NULL;
1065 
1066 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1067 		cache_zap_locked(ncp, false);
1068 		cache_unlock_vnodes(dvlp, vlp);
1069 		return (0);
1070 	}
1071 
1072 	rw_wunlock(blp);
1073 	*vlpp1 = dvlp;
1074 	*vlpp2 = vlp;
1075 	if (*vlpp1 != NULL)
1076 		mtx_lock(*vlpp1);
1077 	mtx_lock(*vlpp2);
1078 	rw_wlock(blp);
1079 	return (EAGAIN);
1080 }
1081 
1082 static void
1083 cache_lookup_unlock(struct rwlock *blp, struct mtx *vlp)
1084 {
1085 
1086 	if (blp != NULL) {
1087 		rw_runlock(blp);
1088 	} else {
1089 		mtx_unlock(vlp);
1090 	}
1091 }
1092 
1093 static int __noinline
1094 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1095     struct timespec *tsp, int *ticksp)
1096 {
1097 	int ltype;
1098 
1099 	*vpp = dvp;
1100 	CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .",
1101 			dvp, cnp->cn_nameptr);
1102 	counter_u64_add(dothits, 1);
1103 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1104 	if (tsp != NULL)
1105 		timespecclear(tsp);
1106 	if (ticksp != NULL)
1107 		*ticksp = ticks;
1108 	vrefact(*vpp);
1109 	/*
1110 	 * When we lookup "." we still can be asked to lock it
1111 	 * differently...
1112 	 */
1113 	ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1114 	if (ltype != VOP_ISLOCKED(*vpp)) {
1115 		if (ltype == LK_EXCLUSIVE) {
1116 			vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1117 			if ((*vpp)->v_iflag & VI_DOOMED) {
1118 				/* forced unmount */
1119 				vrele(*vpp);
1120 				*vpp = NULL;
1121 				return (ENOENT);
1122 			}
1123 		} else
1124 			vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1125 	}
1126 	return (-1);
1127 }
1128 
1129 static __noinline int
1130 cache_lookup_nomakeentry(struct vnode *dvp, struct vnode **vpp,
1131     struct componentname *cnp, struct timespec *tsp, int *ticksp)
1132 {
1133 	struct namecache *ncp;
1134 	struct rwlock *blp;
1135 	struct mtx *dvlp, *dvlp2;
1136 	uint32_t hash;
1137 	int error;
1138 
1139 	if (cnp->cn_namelen == 2 &&
1140 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1141 		counter_u64_add(dotdothits, 1);
1142 		dvlp = VP2VNODELOCK(dvp);
1143 		dvlp2 = NULL;
1144 		mtx_lock(dvlp);
1145 retry_dotdot:
1146 		ncp = dvp->v_cache_dd;
1147 		if (ncp == NULL) {
1148 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1149 			    "..", NULL);
1150 			mtx_unlock(dvlp);
1151 			if (dvlp2 != NULL)
1152 				mtx_unlock(dvlp2);
1153 			return (0);
1154 		}
1155 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1156 			if (ncp->nc_dvp != dvp)
1157 				panic("dvp %p v_cache_dd %p\n", dvp, ncp);
1158 			if (!cache_zap_locked_vnode_kl2(ncp,
1159 			    dvp, &dvlp2))
1160 				goto retry_dotdot;
1161 			MPASS(dvp->v_cache_dd == NULL);
1162 			mtx_unlock(dvlp);
1163 			if (dvlp2 != NULL)
1164 				mtx_unlock(dvlp2);
1165 			cache_free(ncp);
1166 		} else {
1167 			dvp->v_cache_dd = NULL;
1168 			mtx_unlock(dvlp);
1169 			if (dvlp2 != NULL)
1170 				mtx_unlock(dvlp2);
1171 		}
1172 		return (0);
1173 	}
1174 
1175 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1176 	blp = HASH2BUCKETLOCK(hash);
1177 retry:
1178 	if (LIST_EMPTY(NCHHASH(hash)))
1179 		goto out_no_entry;
1180 
1181 	rw_wlock(blp);
1182 
1183 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1184 		counter_u64_add(numchecks, 1);
1185 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1186 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1187 			break;
1188 	}
1189 
1190 	/* We failed to find an entry */
1191 	if (ncp == NULL) {
1192 		rw_wunlock(blp);
1193 		goto out_no_entry;
1194 	}
1195 
1196 	counter_u64_add(numposzaps, 1);
1197 
1198 	error = cache_zap_wlocked_bucket(ncp, blp);
1199 	if (error != 0) {
1200 		zap_and_exit_bucket_fail++;
1201 		cache_maybe_yield();
1202 		goto retry;
1203 	}
1204 	cache_free(ncp);
1205 	return (0);
1206 out_no_entry:
1207 	SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr, NULL);
1208 	counter_u64_add(nummisszap, 1);
1209 	return (0);
1210 }
1211 
1212 /**
1213  * Lookup a name in the name cache
1214  *
1215  * # Arguments
1216  *
1217  * - dvp:	Parent directory in which to search.
1218  * - vpp:	Return argument.  Will contain desired vnode on cache hit.
1219  * - cnp:	Parameters of the name search.  The most interesting bits of
1220  *   		the cn_flags field have the following meanings:
1221  *   	- MAKEENTRY:	If clear, free an entry from the cache rather than look
1222  *   			it up.
1223  *   	- ISDOTDOT:	Must be set if and only if cn_nameptr == ".."
1224  * - tsp:	Return storage for cache timestamp.  On a successful (positive
1225  *   		or negative) lookup, tsp will be filled with any timespec that
1226  *   		was stored when this cache entry was created.  However, it will
1227  *   		be clear for "." entries.
1228  * - ticks:	Return storage for alternate cache timestamp.  On a successful
1229  *   		(positive or negative) lookup, it will contain the ticks value
1230  *   		that was current when the cache entry was created, unless cnp
1231  *   		was ".".
1232  *
1233  * # Returns
1234  *
1235  * - -1:	A positive cache hit.  vpp will contain the desired vnode.
1236  * - ENOENT:	A negative cache hit, or dvp was recycled out from under us due
1237  *		to a forced unmount.  vpp will not be modified.  If the entry
1238  *		is a whiteout, then the ISWHITEOUT flag will be set in
1239  *		cnp->cn_flags.
1240  * - 0:		A cache miss.  vpp will not be modified.
1241  *
1242  * # Locking
1243  *
1244  * On a cache hit, vpp will be returned locked and ref'd.  If we're looking up
1245  * .., dvp is unlocked.  If we're looking up . an extra ref is taken, but the
1246  * lock is not recursively acquired.
1247  */
1248 int
1249 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1250     struct timespec *tsp, int *ticksp)
1251 {
1252 	struct namecache_ts *ncp_ts;
1253 	struct namecache *ncp;
1254 	struct rwlock *blp;
1255 	struct mtx *dvlp;
1256 	uint32_t hash;
1257 	int error, ltype;
1258 
1259 	if (__predict_false(!doingcache)) {
1260 		cnp->cn_flags &= ~MAKEENTRY;
1261 		return (0);
1262 	}
1263 
1264 	counter_u64_add(numcalls, 1);
1265 
1266 	if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.'))
1267 		return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1268 
1269 	if ((cnp->cn_flags & MAKEENTRY) == 0)
1270 		return (cache_lookup_nomakeentry(dvp, vpp, cnp, tsp, ticksp));
1271 
1272 retry:
1273 	blp = NULL;
1274 	dvlp = NULL;
1275 	error = 0;
1276 	if (cnp->cn_namelen == 2 &&
1277 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1278 		counter_u64_add(dotdothits, 1);
1279 		dvlp = VP2VNODELOCK(dvp);
1280 		mtx_lock(dvlp);
1281 		ncp = dvp->v_cache_dd;
1282 		if (ncp == NULL) {
1283 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1284 			    "..", NULL);
1285 			mtx_unlock(dvlp);
1286 			return (0);
1287 		}
1288 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1289 			if (ncp->nc_flag & NCF_NEGATIVE)
1290 				*vpp = NULL;
1291 			else
1292 				*vpp = ncp->nc_vp;
1293 		} else
1294 			*vpp = ncp->nc_dvp;
1295 		/* Return failure if negative entry was found. */
1296 		if (*vpp == NULL)
1297 			goto negative_success;
1298 		CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
1299 		    dvp, cnp->cn_nameptr, *vpp);
1300 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..",
1301 		    *vpp);
1302 		cache_out_ts(ncp, tsp, ticksp);
1303 		if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1304 		    NCF_DTS && tsp != NULL) {
1305 			ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1306 			*tsp = ncp_ts->nc_dotdottime;
1307 		}
1308 		goto success;
1309 	}
1310 
1311 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1312 	blp = HASH2BUCKETLOCK(hash);
1313 	rw_rlock(blp);
1314 
1315 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1316 		counter_u64_add(numchecks, 1);
1317 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1318 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1319 			break;
1320 	}
1321 
1322 	/* We failed to find an entry */
1323 	if (ncp == NULL) {
1324 		rw_runlock(blp);
1325 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1326 		    NULL);
1327 		counter_u64_add(nummiss, 1);
1328 		return (0);
1329 	}
1330 
1331 	/* We found a "positive" match, return the vnode */
1332 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1333 		counter_u64_add(numposhits, 1);
1334 		*vpp = ncp->nc_vp;
1335 		CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p",
1336 		    dvp, cnp->cn_nameptr, *vpp, ncp);
1337 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name,
1338 		    *vpp);
1339 		cache_out_ts(ncp, tsp, ticksp);
1340 		goto success;
1341 	}
1342 
1343 negative_success:
1344 	/* We found a negative match, and want to create it, so purge */
1345 	if (cnp->cn_nameiop == CREATE) {
1346 		counter_u64_add(numnegzaps, 1);
1347 		goto zap_and_exit;
1348 	}
1349 
1350 	counter_u64_add(numneghits, 1);
1351 	cache_negative_hit(ncp);
1352 	if (ncp->nc_flag & NCF_WHITE)
1353 		cnp->cn_flags |= ISWHITEOUT;
1354 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp,
1355 	    ncp->nc_name);
1356 	cache_out_ts(ncp, tsp, ticksp);
1357 	cache_lookup_unlock(blp, dvlp);
1358 	return (ENOENT);
1359 
1360 success:
1361 	/*
1362 	 * On success we return a locked and ref'd vnode as per the lookup
1363 	 * protocol.
1364 	 */
1365 	MPASS(dvp != *vpp);
1366 	ltype = 0;	/* silence gcc warning */
1367 	if (cnp->cn_flags & ISDOTDOT) {
1368 		ltype = VOP_ISLOCKED(dvp);
1369 		VOP_UNLOCK(dvp, 0);
1370 	}
1371 	vhold(*vpp);
1372 	cache_lookup_unlock(blp, dvlp);
1373 	error = vget(*vpp, cnp->cn_lkflags | LK_VNHELD, cnp->cn_thread);
1374 	if (cnp->cn_flags & ISDOTDOT) {
1375 		vn_lock(dvp, ltype | LK_RETRY);
1376 		if (dvp->v_iflag & VI_DOOMED) {
1377 			if (error == 0)
1378 				vput(*vpp);
1379 			*vpp = NULL;
1380 			return (ENOENT);
1381 		}
1382 	}
1383 	if (error) {
1384 		*vpp = NULL;
1385 		goto retry;
1386 	}
1387 	if ((cnp->cn_flags & ISLASTCN) &&
1388 	    (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) {
1389 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
1390 	}
1391 	return (-1);
1392 
1393 zap_and_exit:
1394 	if (blp != NULL)
1395 		error = cache_zap_rlocked_bucket(ncp, blp);
1396 	else
1397 		error = cache_zap_locked_vnode(ncp, dvp);
1398 	if (error != 0) {
1399 		zap_and_exit_bucket_fail++;
1400 		cache_maybe_yield();
1401 		goto retry;
1402 	}
1403 	cache_free(ncp);
1404 	return (0);
1405 }
1406 
1407 struct celockstate {
1408 	struct mtx *vlp[3];
1409 	struct rwlock *blp[2];
1410 };
1411 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1412 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1413 
1414 static inline void
1415 cache_celockstate_init(struct celockstate *cel)
1416 {
1417 
1418 	bzero(cel, sizeof(*cel));
1419 }
1420 
1421 static void
1422 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1423     struct vnode *dvp)
1424 {
1425 	struct mtx *vlp1, *vlp2;
1426 
1427 	MPASS(cel->vlp[0] == NULL);
1428 	MPASS(cel->vlp[1] == NULL);
1429 	MPASS(cel->vlp[2] == NULL);
1430 
1431 	MPASS(vp != NULL || dvp != NULL);
1432 
1433 	vlp1 = VP2VNODELOCK(vp);
1434 	vlp2 = VP2VNODELOCK(dvp);
1435 	cache_sort(&vlp1, &vlp2);
1436 
1437 	if (vlp1 != NULL) {
1438 		mtx_lock(vlp1);
1439 		cel->vlp[0] = vlp1;
1440 	}
1441 	mtx_lock(vlp2);
1442 	cel->vlp[1] = vlp2;
1443 }
1444 
1445 static void
1446 cache_unlock_vnodes_cel(struct celockstate *cel)
1447 {
1448 
1449 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
1450 
1451 	if (cel->vlp[0] != NULL)
1452 		mtx_unlock(cel->vlp[0]);
1453 	if (cel->vlp[1] != NULL)
1454 		mtx_unlock(cel->vlp[1]);
1455 	if (cel->vlp[2] != NULL)
1456 		mtx_unlock(cel->vlp[2]);
1457 }
1458 
1459 static bool
1460 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
1461 {
1462 	struct mtx *vlp;
1463 	bool ret;
1464 
1465 	cache_assert_vlp_locked(cel->vlp[0]);
1466 	cache_assert_vlp_locked(cel->vlp[1]);
1467 	MPASS(cel->vlp[2] == NULL);
1468 
1469 	MPASS(vp != NULL);
1470 	vlp = VP2VNODELOCK(vp);
1471 
1472 	ret = true;
1473 	if (vlp >= cel->vlp[1]) {
1474 		mtx_lock(vlp);
1475 	} else {
1476 		if (mtx_trylock(vlp))
1477 			goto out;
1478 		cache_lock_vnodes_cel_3_failures++;
1479 		cache_unlock_vnodes_cel(cel);
1480 		if (vlp < cel->vlp[0]) {
1481 			mtx_lock(vlp);
1482 			mtx_lock(cel->vlp[0]);
1483 			mtx_lock(cel->vlp[1]);
1484 		} else {
1485 			if (cel->vlp[0] != NULL)
1486 				mtx_lock(cel->vlp[0]);
1487 			mtx_lock(vlp);
1488 			mtx_lock(cel->vlp[1]);
1489 		}
1490 		ret = false;
1491 	}
1492 out:
1493 	cel->vlp[2] = vlp;
1494 	return (ret);
1495 }
1496 
1497 static void
1498 cache_lock_buckets_cel(struct celockstate *cel, struct rwlock *blp1,
1499     struct rwlock *blp2)
1500 {
1501 
1502 	MPASS(cel->blp[0] == NULL);
1503 	MPASS(cel->blp[1] == NULL);
1504 
1505 	cache_sort(&blp1, &blp2);
1506 
1507 	if (blp1 != NULL) {
1508 		rw_wlock(blp1);
1509 		cel->blp[0] = blp1;
1510 	}
1511 	rw_wlock(blp2);
1512 	cel->blp[1] = blp2;
1513 }
1514 
1515 static void
1516 cache_unlock_buckets_cel(struct celockstate *cel)
1517 {
1518 
1519 	if (cel->blp[0] != NULL)
1520 		rw_wunlock(cel->blp[0]);
1521 	rw_wunlock(cel->blp[1]);
1522 }
1523 
1524 /*
1525  * Lock part of the cache affected by the insertion.
1526  *
1527  * This means vnodelocks for dvp, vp and the relevant bucketlock.
1528  * However, insertion can result in removal of an old entry. In this
1529  * case we have an additional vnode and bucketlock pair to lock. If the
1530  * entry is negative, ncelock is locked instead of the vnode.
1531  *
1532  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
1533  * preserving the locking order (smaller address first).
1534  */
1535 static void
1536 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1537     uint32_t hash)
1538 {
1539 	struct namecache *ncp;
1540 	struct rwlock *blps[2];
1541 
1542 	blps[0] = HASH2BUCKETLOCK(hash);
1543 	for (;;) {
1544 		blps[1] = NULL;
1545 		cache_lock_vnodes_cel(cel, dvp, vp);
1546 		if (vp == NULL || vp->v_type != VDIR)
1547 			break;
1548 		ncp = vp->v_cache_dd;
1549 		if (ncp == NULL)
1550 			break;
1551 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1552 			break;
1553 		MPASS(ncp->nc_dvp == vp);
1554 		blps[1] = NCP2BUCKETLOCK(ncp);
1555 		if (ncp->nc_flag & NCF_NEGATIVE)
1556 			break;
1557 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1558 			break;
1559 		/*
1560 		 * All vnodes got re-locked. Re-validate the state and if
1561 		 * nothing changed we are done. Otherwise restart.
1562 		 */
1563 		if (ncp == vp->v_cache_dd &&
1564 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1565 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1566 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1567 			break;
1568 		cache_unlock_vnodes_cel(cel);
1569 		cel->vlp[0] = NULL;
1570 		cel->vlp[1] = NULL;
1571 		cel->vlp[2] = NULL;
1572 	}
1573 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1574 }
1575 
1576 static void
1577 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1578     uint32_t hash)
1579 {
1580 	struct namecache *ncp;
1581 	struct rwlock *blps[2];
1582 
1583 	blps[0] = HASH2BUCKETLOCK(hash);
1584 	for (;;) {
1585 		blps[1] = NULL;
1586 		cache_lock_vnodes_cel(cel, dvp, vp);
1587 		ncp = dvp->v_cache_dd;
1588 		if (ncp == NULL)
1589 			break;
1590 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1591 			break;
1592 		MPASS(ncp->nc_dvp == dvp);
1593 		blps[1] = NCP2BUCKETLOCK(ncp);
1594 		if (ncp->nc_flag & NCF_NEGATIVE)
1595 			break;
1596 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1597 			break;
1598 		if (ncp == dvp->v_cache_dd &&
1599 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1600 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1601 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1602 			break;
1603 		cache_unlock_vnodes_cel(cel);
1604 		cel->vlp[0] = NULL;
1605 		cel->vlp[1] = NULL;
1606 		cel->vlp[2] = NULL;
1607 	}
1608 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1609 }
1610 
1611 static void
1612 cache_enter_unlock(struct celockstate *cel)
1613 {
1614 
1615 	cache_unlock_buckets_cel(cel);
1616 	cache_unlock_vnodes_cel(cel);
1617 }
1618 
1619 /*
1620  * Add an entry to the cache.
1621  */
1622 void
1623 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
1624     struct timespec *tsp, struct timespec *dtsp)
1625 {
1626 	struct celockstate cel;
1627 	struct namecache *ncp, *n2, *ndd;
1628 	struct namecache_ts *ncp_ts, *n2_ts;
1629 	struct nchashhead *ncpp;
1630 	struct neglist *neglist;
1631 	uint32_t hash;
1632 	int flag;
1633 	int len;
1634 	bool neg_locked;
1635 	u_long lnumcache;
1636 
1637 	CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr);
1638 	VNASSERT(vp == NULL || (vp->v_iflag & VI_DOOMED) == 0, vp,
1639 	    ("cache_enter: Adding a doomed vnode"));
1640 	VNASSERT(dvp == NULL || (dvp->v_iflag & VI_DOOMED) == 0, dvp,
1641 	    ("cache_enter: Doomed vnode used as src"));
1642 
1643 	if (__predict_false(!doingcache))
1644 		return;
1645 
1646 	/*
1647 	 * Avoid blowout in namecache entries.
1648 	 */
1649 	lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
1650 	if (__predict_false(lnumcache >= desiredvnodes * ncsizefactor)) {
1651 		atomic_add_long(&numcache, -1);
1652 		return;
1653 	}
1654 
1655 	cache_celockstate_init(&cel);
1656 	ndd = NULL;
1657 	ncp_ts = NULL;
1658 	flag = 0;
1659 	if (cnp->cn_nameptr[0] == '.') {
1660 		if (cnp->cn_namelen == 1)
1661 			return;
1662 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1663 			len = cnp->cn_namelen;
1664 			hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1665 			cache_enter_lock_dd(&cel, dvp, vp, hash);
1666 			/*
1667 			 * If dotdot entry already exists, just retarget it
1668 			 * to new parent vnode, otherwise continue with new
1669 			 * namecache entry allocation.
1670 			 */
1671 			if ((ncp = dvp->v_cache_dd) != NULL &&
1672 			    ncp->nc_flag & NCF_ISDOTDOT) {
1673 				KASSERT(ncp->nc_dvp == dvp,
1674 				    ("wrong isdotdot parent"));
1675 				neg_locked = false;
1676 				if (ncp->nc_flag & NCF_NEGATIVE || vp == NULL) {
1677 					neglist = NCP2NEGLIST(ncp);
1678 					mtx_lock(&ncneg_hot.nl_lock);
1679 					mtx_lock(&neglist->nl_lock);
1680 					neg_locked = true;
1681 				}
1682 				if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1683 					TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst,
1684 					    ncp, nc_dst);
1685 				} else {
1686 					cache_negative_remove(ncp, true);
1687 				}
1688 				if (vp != NULL) {
1689 					TAILQ_INSERT_HEAD(&vp->v_cache_dst,
1690 					    ncp, nc_dst);
1691 					ncp->nc_flag &= ~(NCF_NEGATIVE|NCF_HOTNEGATIVE);
1692 				} else {
1693 					ncp->nc_flag &= ~(NCF_HOTNEGATIVE);
1694 					ncp->nc_flag |= NCF_NEGATIVE;
1695 					cache_negative_insert(ncp, true);
1696 				}
1697 				if (neg_locked) {
1698 					mtx_unlock(&neglist->nl_lock);
1699 					mtx_unlock(&ncneg_hot.nl_lock);
1700 				}
1701 				ncp->nc_vp = vp;
1702 				cache_enter_unlock(&cel);
1703 				return;
1704 			}
1705 			dvp->v_cache_dd = NULL;
1706 			cache_enter_unlock(&cel);
1707 			cache_celockstate_init(&cel);
1708 			SDT_PROBE3(vfs, namecache, enter, done, dvp, "..", vp);
1709 			flag = NCF_ISDOTDOT;
1710 		}
1711 	}
1712 
1713 	/*
1714 	 * Calculate the hash key and setup as much of the new
1715 	 * namecache entry as possible before acquiring the lock.
1716 	 */
1717 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
1718 	ncp->nc_flag = flag;
1719 	ncp->nc_vp = vp;
1720 	if (vp == NULL)
1721 		ncp->nc_flag |= NCF_NEGATIVE;
1722 	ncp->nc_dvp = dvp;
1723 	if (tsp != NULL) {
1724 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1725 		ncp_ts->nc_time = *tsp;
1726 		ncp_ts->nc_ticks = ticks;
1727 		ncp_ts->nc_nc.nc_flag |= NCF_TS;
1728 		if (dtsp != NULL) {
1729 			ncp_ts->nc_dotdottime = *dtsp;
1730 			ncp_ts->nc_nc.nc_flag |= NCF_DTS;
1731 		}
1732 	}
1733 	len = ncp->nc_nlen = cnp->cn_namelen;
1734 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1735 	strlcpy(ncp->nc_name, cnp->cn_nameptr, len + 1);
1736 	cache_enter_lock(&cel, dvp, vp, hash);
1737 
1738 	/*
1739 	 * See if this vnode or negative entry is already in the cache
1740 	 * with this name.  This can happen with concurrent lookups of
1741 	 * the same path name.
1742 	 */
1743 	ncpp = NCHHASH(hash);
1744 	LIST_FOREACH(n2, ncpp, nc_hash) {
1745 		if (n2->nc_dvp == dvp &&
1746 		    n2->nc_nlen == cnp->cn_namelen &&
1747 		    !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
1748 			if (tsp != NULL) {
1749 				KASSERT((n2->nc_flag & NCF_TS) != 0,
1750 				    ("no NCF_TS"));
1751 				n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
1752 				n2_ts->nc_time = ncp_ts->nc_time;
1753 				n2_ts->nc_ticks = ncp_ts->nc_ticks;
1754 				if (dtsp != NULL) {
1755 					n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
1756 					if (ncp->nc_flag & NCF_NEGATIVE)
1757 						mtx_lock(&ncneg_hot.nl_lock);
1758 					n2_ts->nc_nc.nc_flag |= NCF_DTS;
1759 					if (ncp->nc_flag & NCF_NEGATIVE)
1760 						mtx_unlock(&ncneg_hot.nl_lock);
1761 				}
1762 			}
1763 			goto out_unlock_free;
1764 		}
1765 	}
1766 
1767 	if (flag == NCF_ISDOTDOT) {
1768 		/*
1769 		 * See if we are trying to add .. entry, but some other lookup
1770 		 * has populated v_cache_dd pointer already.
1771 		 */
1772 		if (dvp->v_cache_dd != NULL)
1773 			goto out_unlock_free;
1774 		KASSERT(vp == NULL || vp->v_type == VDIR,
1775 		    ("wrong vnode type %p", vp));
1776 		dvp->v_cache_dd = ncp;
1777 	}
1778 
1779 	if (vp != NULL) {
1780 		if (vp->v_type == VDIR) {
1781 			if (flag != NCF_ISDOTDOT) {
1782 				/*
1783 				 * For this case, the cache entry maps both the
1784 				 * directory name in it and the name ".." for the
1785 				 * directory's parent.
1786 				 */
1787 				if ((ndd = vp->v_cache_dd) != NULL) {
1788 					if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
1789 						cache_zap_locked(ndd, false);
1790 					else
1791 						ndd = NULL;
1792 				}
1793 				vp->v_cache_dd = ncp;
1794 			}
1795 		} else {
1796 			vp->v_cache_dd = NULL;
1797 		}
1798 	}
1799 
1800 	if (flag != NCF_ISDOTDOT) {
1801 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1802 			vhold(dvp);
1803 			atomic_add_rel_long(&numcachehv, 1);
1804 		}
1805 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
1806 	}
1807 
1808 	/*
1809 	 * Insert the new namecache entry into the appropriate chain
1810 	 * within the cache entries table.
1811 	 */
1812 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
1813 
1814 	/*
1815 	 * If the entry is "negative", we place it into the
1816 	 * "negative" cache queue, otherwise, we place it into the
1817 	 * destination vnode's cache entries queue.
1818 	 */
1819 	if (vp != NULL) {
1820 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
1821 		SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
1822 		    vp);
1823 	} else {
1824 		if (cnp->cn_flags & ISWHITEOUT)
1825 			ncp->nc_flag |= NCF_WHITE;
1826 		cache_negative_insert(ncp, false);
1827 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
1828 		    ncp->nc_name);
1829 	}
1830 	cache_enter_unlock(&cel);
1831 	if (numneg * ncnegfactor > lnumcache)
1832 		cache_negative_zap_one();
1833 	cache_free(ndd);
1834 	return;
1835 out_unlock_free:
1836 	cache_enter_unlock(&cel);
1837 	cache_free(ncp);
1838 	return;
1839 }
1840 
1841 static u_int
1842 cache_roundup_2(u_int val)
1843 {
1844 	u_int res;
1845 
1846 	for (res = 1; res <= val; res <<= 1)
1847 		continue;
1848 
1849 	return (res);
1850 }
1851 
1852 /*
1853  * Name cache initialization, from vfs_init() when we are booting
1854  */
1855 static void
1856 nchinit(void *dummy __unused)
1857 {
1858 	u_int i;
1859 
1860 	cache_zone_small = uma_zcreate("S VFS Cache",
1861 	    sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1,
1862 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1863 	    UMA_ZONE_ZINIT);
1864 	cache_zone_small_ts = uma_zcreate("STS VFS Cache",
1865 	    sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1,
1866 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1867 	    UMA_ZONE_ZINIT);
1868 	cache_zone_large = uma_zcreate("L VFS Cache",
1869 	    sizeof(struct namecache) + NAME_MAX + 1,
1870 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1871 	    UMA_ZONE_ZINIT);
1872 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache",
1873 	    sizeof(struct namecache_ts) + NAME_MAX + 1,
1874 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1875 	    UMA_ZONE_ZINIT);
1876 
1877 	nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash);
1878 	ncbuckethash = cache_roundup_2(mp_ncpus * 64) - 1;
1879 	if (ncbuckethash > nchash)
1880 		ncbuckethash = nchash;
1881 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
1882 	    M_WAITOK | M_ZERO);
1883 	for (i = 0; i < numbucketlocks; i++)
1884 		rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK | RW_RECURSE);
1885 	ncvnodehash = cache_roundup_2(mp_ncpus * 64) - 1;
1886 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
1887 	    M_WAITOK | M_ZERO);
1888 	for (i = 0; i < numvnodelocks; i++)
1889 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
1890 	ncpurgeminvnodes = numbucketlocks;
1891 
1892 	ncneghash = 3;
1893 	neglists = malloc(sizeof(*neglists) * numneglists, M_VFSCACHE,
1894 	    M_WAITOK | M_ZERO);
1895 	for (i = 0; i < numneglists; i++) {
1896 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
1897 		TAILQ_INIT(&neglists[i].nl_list);
1898 	}
1899 	mtx_init(&ncneg_hot.nl_lock, "ncneglh", NULL, MTX_DEF);
1900 	TAILQ_INIT(&ncneg_hot.nl_list);
1901 
1902 	mtx_init(&ncneg_shrink_lock, "ncnegs", NULL, MTX_DEF);
1903 
1904 	numcalls = counter_u64_alloc(M_WAITOK);
1905 	dothits = counter_u64_alloc(M_WAITOK);
1906 	dotdothits = counter_u64_alloc(M_WAITOK);
1907 	numchecks = counter_u64_alloc(M_WAITOK);
1908 	nummiss = counter_u64_alloc(M_WAITOK);
1909 	nummisszap = counter_u64_alloc(M_WAITOK);
1910 	numposzaps = counter_u64_alloc(M_WAITOK);
1911 	numposhits = counter_u64_alloc(M_WAITOK);
1912 	numnegzaps = counter_u64_alloc(M_WAITOK);
1913 	numneghits = counter_u64_alloc(M_WAITOK);
1914 	numfullpathcalls = counter_u64_alloc(M_WAITOK);
1915 	numfullpathfail1 = counter_u64_alloc(M_WAITOK);
1916 	numfullpathfail2 = counter_u64_alloc(M_WAITOK);
1917 	numfullpathfail4 = counter_u64_alloc(M_WAITOK);
1918 	numfullpathfound = counter_u64_alloc(M_WAITOK);
1919 }
1920 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
1921 
1922 void
1923 cache_changesize(int newmaxvnodes)
1924 {
1925 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
1926 	u_long new_nchash, old_nchash;
1927 	struct namecache *ncp;
1928 	uint32_t hash;
1929 	int i;
1930 
1931 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
1932 	if (newmaxvnodes < numbucketlocks)
1933 		newmaxvnodes = numbucketlocks;
1934 
1935 	new_nchashtbl = hashinit(newmaxvnodes, M_VFSCACHE, &new_nchash);
1936 	/* If same hash table size, nothing to do */
1937 	if (nchash == new_nchash) {
1938 		free(new_nchashtbl, M_VFSCACHE);
1939 		return;
1940 	}
1941 	/*
1942 	 * Move everything from the old hash table to the new table.
1943 	 * None of the namecache entries in the table can be removed
1944 	 * because to do so, they have to be removed from the hash table.
1945 	 */
1946 	cache_lock_all_vnodes();
1947 	cache_lock_all_buckets();
1948 	old_nchashtbl = nchashtbl;
1949 	old_nchash = nchash;
1950 	nchashtbl = new_nchashtbl;
1951 	nchash = new_nchash;
1952 	for (i = 0; i <= old_nchash; i++) {
1953 		while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) {
1954 			hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
1955 			    ncp->nc_dvp);
1956 			LIST_REMOVE(ncp, nc_hash);
1957 			LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
1958 		}
1959 	}
1960 	cache_unlock_all_buckets();
1961 	cache_unlock_all_vnodes();
1962 	free(old_nchashtbl, M_VFSCACHE);
1963 }
1964 
1965 /*
1966  * Invalidate all entries to a particular vnode.
1967  */
1968 void
1969 cache_purge(struct vnode *vp)
1970 {
1971 	TAILQ_HEAD(, namecache) ncps;
1972 	struct namecache *ncp, *nnp;
1973 	struct mtx *vlp, *vlp2;
1974 
1975 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
1976 	SDT_PROBE1(vfs, namecache, purge, done, vp);
1977 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
1978 	    vp->v_cache_dd == NULL)
1979 		return;
1980 	TAILQ_INIT(&ncps);
1981 	vlp = VP2VNODELOCK(vp);
1982 	vlp2 = NULL;
1983 	mtx_lock(vlp);
1984 retry:
1985 	while (!LIST_EMPTY(&vp->v_cache_src)) {
1986 		ncp = LIST_FIRST(&vp->v_cache_src);
1987 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
1988 			goto retry;
1989 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1990 	}
1991 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
1992 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
1993 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
1994 			goto retry;
1995 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1996 	}
1997 	ncp = vp->v_cache_dd;
1998 	if (ncp != NULL) {
1999 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2000 		   ("lost dotdot link"));
2001 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2002 			goto retry;
2003 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2004 	}
2005 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2006 	mtx_unlock(vlp);
2007 	if (vlp2 != NULL)
2008 		mtx_unlock(vlp2);
2009 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2010 		cache_free(ncp);
2011 	}
2012 }
2013 
2014 /*
2015  * Invalidate all negative entries for a particular directory vnode.
2016  */
2017 void
2018 cache_purge_negative(struct vnode *vp)
2019 {
2020 	TAILQ_HEAD(, namecache) ncps;
2021 	struct namecache *ncp, *nnp;
2022 	struct mtx *vlp;
2023 
2024 	CTR1(KTR_VFS, "cache_purge_negative(%p)", vp);
2025 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2026 	if (LIST_EMPTY(&vp->v_cache_src))
2027 		return;
2028 	TAILQ_INIT(&ncps);
2029 	vlp = VP2VNODELOCK(vp);
2030 	mtx_lock(vlp);
2031 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2032 		if (!(ncp->nc_flag & NCF_NEGATIVE))
2033 			continue;
2034 		cache_zap_negative_locked_vnode_kl(ncp, vp);
2035 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2036 	}
2037 	mtx_unlock(vlp);
2038 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2039 		cache_free(ncp);
2040 	}
2041 }
2042 
2043 /*
2044  * Flush all entries referencing a particular filesystem.
2045  */
2046 void
2047 cache_purgevfs(struct mount *mp, bool force)
2048 {
2049 	TAILQ_HEAD(, namecache) ncps;
2050 	struct mtx *vlp1, *vlp2;
2051 	struct rwlock *blp;
2052 	struct nchashhead *bucket;
2053 	struct namecache *ncp, *nnp;
2054 	u_long i, j, n_nchash;
2055 	int error;
2056 
2057 	/* Scan hash tables for applicable entries */
2058 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2059 	if (!force && mp->mnt_nvnodelistsize <= ncpurgeminvnodes)
2060 		return;
2061 	TAILQ_INIT(&ncps);
2062 	n_nchash = nchash + 1;
2063 	vlp1 = vlp2 = NULL;
2064 	for (i = 0; i < numbucketlocks; i++) {
2065 		blp = (struct rwlock *)&bucketlocks[i];
2066 		rw_wlock(blp);
2067 		for (j = i; j < n_nchash; j += numbucketlocks) {
2068 retry:
2069 			bucket = &nchashtbl[j];
2070 			LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) {
2071 				cache_assert_bucket_locked(ncp, RA_WLOCKED);
2072 				if (ncp->nc_dvp->v_mount != mp)
2073 					continue;
2074 				error = cache_zap_wlocked_bucket_kl(ncp, blp,
2075 				    &vlp1, &vlp2);
2076 				if (error != 0)
2077 					goto retry;
2078 				TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst);
2079 			}
2080 		}
2081 		rw_wunlock(blp);
2082 		if (vlp1 == NULL && vlp2 == NULL)
2083 			cache_maybe_yield();
2084 	}
2085 	if (vlp1 != NULL)
2086 		mtx_unlock(vlp1);
2087 	if (vlp2 != NULL)
2088 		mtx_unlock(vlp2);
2089 
2090 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2091 		cache_free(ncp);
2092 	}
2093 }
2094 
2095 /*
2096  * Perform canonical checks and cache lookup and pass on to filesystem
2097  * through the vop_cachedlookup only if needed.
2098  */
2099 
2100 int
2101 vfs_cache_lookup(struct vop_lookup_args *ap)
2102 {
2103 	struct vnode *dvp;
2104 	int error;
2105 	struct vnode **vpp = ap->a_vpp;
2106 	struct componentname *cnp = ap->a_cnp;
2107 	struct ucred *cred = cnp->cn_cred;
2108 	int flags = cnp->cn_flags;
2109 	struct thread *td = cnp->cn_thread;
2110 
2111 	*vpp = NULL;
2112 	dvp = ap->a_dvp;
2113 
2114 	if (dvp->v_type != VDIR)
2115 		return (ENOTDIR);
2116 
2117 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2118 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2119 		return (EROFS);
2120 
2121 	error = VOP_ACCESS(dvp, VEXEC, cred, td);
2122 	if (error)
2123 		return (error);
2124 
2125 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2126 	if (error == 0)
2127 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2128 	if (error == -1)
2129 		return (0);
2130 	return (error);
2131 }
2132 
2133 /*
2134  * XXX All of these sysctls would probably be more productive dead.
2135  */
2136 static int __read_mostly disablecwd;
2137 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
2138    "Disable the getcwd syscall");
2139 
2140 /* Implementation of the getcwd syscall. */
2141 int
2142 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2143 {
2144 
2145 	return (kern___getcwd(td, uap->buf, UIO_USERSPACE, uap->buflen,
2146 	    MAXPATHLEN));
2147 }
2148 
2149 int
2150 kern___getcwd(struct thread *td, char *buf, enum uio_seg bufseg, size_t buflen,
2151     size_t path_max)
2152 {
2153 	char *bp, *tmpbuf;
2154 	struct filedesc *fdp;
2155 	struct vnode *cdir, *rdir;
2156 	int error;
2157 
2158 	if (__predict_false(disablecwd))
2159 		return (ENODEV);
2160 	if (__predict_false(buflen < 2))
2161 		return (EINVAL);
2162 	if (buflen > path_max)
2163 		buflen = path_max;
2164 
2165 	tmpbuf = malloc(buflen, M_TEMP, M_WAITOK);
2166 	fdp = td->td_proc->p_fd;
2167 	FILEDESC_SLOCK(fdp);
2168 	cdir = fdp->fd_cdir;
2169 	vrefact(cdir);
2170 	rdir = fdp->fd_rdir;
2171 	vrefact(rdir);
2172 	FILEDESC_SUNLOCK(fdp);
2173 	error = vn_fullpath1(td, cdir, rdir, tmpbuf, &bp, buflen);
2174 	vrele(rdir);
2175 	vrele(cdir);
2176 
2177 	if (!error) {
2178 		if (bufseg == UIO_SYSSPACE)
2179 			bcopy(bp, buf, strlen(bp) + 1);
2180 		else
2181 			error = copyout(bp, buf, strlen(bp) + 1);
2182 #ifdef KTRACE
2183 	if (KTRPOINT(curthread, KTR_NAMEI))
2184 		ktrnamei(bp);
2185 #endif
2186 	}
2187 	free(tmpbuf, M_TEMP);
2188 	return (error);
2189 }
2190 
2191 /*
2192  * Thus begins the fullpath magic.
2193  */
2194 
2195 static int __read_mostly disablefullpath;
2196 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, &disablefullpath, 0,
2197     "Disable the vn_fullpath function");
2198 
2199 /*
2200  * Retrieve the full filesystem path that correspond to a vnode from the name
2201  * cache (if available)
2202  */
2203 int
2204 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf)
2205 {
2206 	char *buf;
2207 	struct filedesc *fdp;
2208 	struct vnode *rdir;
2209 	int error;
2210 
2211 	if (__predict_false(disablefullpath))
2212 		return (ENODEV);
2213 	if (__predict_false(vn == NULL))
2214 		return (EINVAL);
2215 
2216 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2217 	fdp = td->td_proc->p_fd;
2218 	FILEDESC_SLOCK(fdp);
2219 	rdir = fdp->fd_rdir;
2220 	vrefact(rdir);
2221 	FILEDESC_SUNLOCK(fdp);
2222 	error = vn_fullpath1(td, vn, rdir, buf, retbuf, MAXPATHLEN);
2223 	vrele(rdir);
2224 
2225 	if (!error)
2226 		*freebuf = buf;
2227 	else
2228 		free(buf, M_TEMP);
2229 	return (error);
2230 }
2231 
2232 /*
2233  * This function is similar to vn_fullpath, but it attempts to lookup the
2234  * pathname relative to the global root mount point.  This is required for the
2235  * auditing sub-system, as audited pathnames must be absolute, relative to the
2236  * global root mount point.
2237  */
2238 int
2239 vn_fullpath_global(struct thread *td, struct vnode *vn,
2240     char **retbuf, char **freebuf)
2241 {
2242 	char *buf;
2243 	int error;
2244 
2245 	if (__predict_false(disablefullpath))
2246 		return (ENODEV);
2247 	if (__predict_false(vn == NULL))
2248 		return (EINVAL);
2249 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2250 	error = vn_fullpath1(td, vn, rootvnode, buf, retbuf, MAXPATHLEN);
2251 	if (!error)
2252 		*freebuf = buf;
2253 	else
2254 		free(buf, M_TEMP);
2255 	return (error);
2256 }
2257 
2258 int
2259 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen)
2260 {
2261 	struct vnode *dvp;
2262 	struct namecache *ncp;
2263 	struct mtx *vlp;
2264 	int error;
2265 
2266 	vlp = VP2VNODELOCK(*vp);
2267 	mtx_lock(vlp);
2268 	TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) {
2269 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2270 			break;
2271 	}
2272 	if (ncp != NULL) {
2273 		if (*buflen < ncp->nc_nlen) {
2274 			mtx_unlock(vlp);
2275 			vrele(*vp);
2276 			counter_u64_add(numfullpathfail4, 1);
2277 			error = ENOMEM;
2278 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2279 			    vp, NULL);
2280 			return (error);
2281 		}
2282 		*buflen -= ncp->nc_nlen;
2283 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2284 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2285 		    ncp->nc_name, vp);
2286 		dvp = *vp;
2287 		*vp = ncp->nc_dvp;
2288 		vref(*vp);
2289 		mtx_unlock(vlp);
2290 		vrele(dvp);
2291 		return (0);
2292 	}
2293 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2294 
2295 	mtx_unlock(vlp);
2296 	vn_lock(*vp, LK_SHARED | LK_RETRY);
2297 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
2298 	vput(*vp);
2299 	if (error) {
2300 		counter_u64_add(numfullpathfail2, 1);
2301 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
2302 		return (error);
2303 	}
2304 
2305 	*vp = dvp;
2306 	if (dvp->v_iflag & VI_DOOMED) {
2307 		/* forced unmount */
2308 		vrele(dvp);
2309 		error = ENOENT;
2310 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2311 		return (error);
2312 	}
2313 	/*
2314 	 * *vp has its use count incremented still.
2315 	 */
2316 
2317 	return (0);
2318 }
2319 
2320 /*
2321  * The magic behind kern___getcwd() and vn_fullpath().
2322  */
2323 static int
2324 vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
2325     char *buf, char **retbuf, u_int buflen)
2326 {
2327 	int error, slash_prefixed;
2328 #ifdef KDTRACE_HOOKS
2329 	struct vnode *startvp = vp;
2330 #endif
2331 	struct vnode *vp1;
2332 
2333 	buflen--;
2334 	buf[buflen] = '\0';
2335 	error = 0;
2336 	slash_prefixed = 0;
2337 
2338 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
2339 	counter_u64_add(numfullpathcalls, 1);
2340 	vref(vp);
2341 	if (vp->v_type != VDIR) {
2342 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2343 		if (error)
2344 			return (error);
2345 		if (buflen == 0) {
2346 			vrele(vp);
2347 			return (ENOMEM);
2348 		}
2349 		buf[--buflen] = '/';
2350 		slash_prefixed = 1;
2351 	}
2352 	while (vp != rdir && vp != rootvnode) {
2353 		/*
2354 		 * The vp vnode must be already fully constructed,
2355 		 * since it is either found in namecache or obtained
2356 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
2357 		 * without obtaining the vnode lock.
2358 		 */
2359 		if ((vp->v_vflag & VV_ROOT) != 0) {
2360 			vn_lock(vp, LK_RETRY | LK_SHARED);
2361 
2362 			/*
2363 			 * With the vnode locked, check for races with
2364 			 * unmount, forced or not.  Note that we
2365 			 * already verified that vp is not equal to
2366 			 * the root vnode, which means that
2367 			 * mnt_vnodecovered can be NULL only for the
2368 			 * case of unmount.
2369 			 */
2370 			if ((vp->v_iflag & VI_DOOMED) != 0 ||
2371 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
2372 			    vp1->v_mountedhere != vp->v_mount) {
2373 				vput(vp);
2374 				error = ENOENT;
2375 				SDT_PROBE3(vfs, namecache, fullpath, return,
2376 				    error, vp, NULL);
2377 				break;
2378 			}
2379 
2380 			vref(vp1);
2381 			vput(vp);
2382 			vp = vp1;
2383 			continue;
2384 		}
2385 		if (vp->v_type != VDIR) {
2386 			vrele(vp);
2387 			counter_u64_add(numfullpathfail1, 1);
2388 			error = ENOTDIR;
2389 			SDT_PROBE3(vfs, namecache, fullpath, return,
2390 			    error, vp, NULL);
2391 			break;
2392 		}
2393 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2394 		if (error)
2395 			break;
2396 		if (buflen == 0) {
2397 			vrele(vp);
2398 			error = ENOMEM;
2399 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2400 			    startvp, NULL);
2401 			break;
2402 		}
2403 		buf[--buflen] = '/';
2404 		slash_prefixed = 1;
2405 	}
2406 	if (error)
2407 		return (error);
2408 	if (!slash_prefixed) {
2409 		if (buflen == 0) {
2410 			vrele(vp);
2411 			counter_u64_add(numfullpathfail4, 1);
2412 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
2413 			    startvp, NULL);
2414 			return (ENOMEM);
2415 		}
2416 		buf[--buflen] = '/';
2417 	}
2418 	counter_u64_add(numfullpathfound, 1);
2419 	vrele(vp);
2420 
2421 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, buf + buflen);
2422 	*retbuf = buf + buflen;
2423 	return (0);
2424 }
2425 
2426 struct vnode *
2427 vn_dir_dd_ino(struct vnode *vp)
2428 {
2429 	struct namecache *ncp;
2430 	struct vnode *ddvp;
2431 	struct mtx *vlp;
2432 
2433 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
2434 	vlp = VP2VNODELOCK(vp);
2435 	mtx_lock(vlp);
2436 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
2437 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
2438 			continue;
2439 		ddvp = ncp->nc_dvp;
2440 		vhold(ddvp);
2441 		mtx_unlock(vlp);
2442 		if (vget(ddvp, LK_SHARED | LK_NOWAIT | LK_VNHELD, curthread))
2443 			return (NULL);
2444 		return (ddvp);
2445 	}
2446 	mtx_unlock(vlp);
2447 	return (NULL);
2448 }
2449 
2450 int
2451 vn_commname(struct vnode *vp, char *buf, u_int buflen)
2452 {
2453 	struct namecache *ncp;
2454 	struct mtx *vlp;
2455 	int l;
2456 
2457 	vlp = VP2VNODELOCK(vp);
2458 	mtx_lock(vlp);
2459 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
2460 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2461 			break;
2462 	if (ncp == NULL) {
2463 		mtx_unlock(vlp);
2464 		return (ENOENT);
2465 	}
2466 	l = min(ncp->nc_nlen, buflen - 1);
2467 	memcpy(buf, ncp->nc_name, l);
2468 	mtx_unlock(vlp);
2469 	buf[l] = '\0';
2470 	return (0);
2471 }
2472 
2473 /*
2474  * This function updates path string to vnode's full global path
2475  * and checks the size of the new path string against the pathlen argument.
2476  *
2477  * Requires a locked, referenced vnode.
2478  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
2479  *
2480  * If sysctl debug.disablefullpath is set, ENODEV is returned,
2481  * vnode is left locked and path remain untouched.
2482  *
2483  * If vp is a directory, the call to vn_fullpath_global() always succeeds
2484  * because it falls back to the ".." lookup if the namecache lookup fails.
2485  */
2486 int
2487 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
2488     u_int pathlen)
2489 {
2490 	struct nameidata nd;
2491 	struct vnode *vp1;
2492 	char *rpath, *fbuf;
2493 	int error;
2494 
2495 	ASSERT_VOP_ELOCKED(vp, __func__);
2496 
2497 	/* Return ENODEV if sysctl debug.disablefullpath==1 */
2498 	if (__predict_false(disablefullpath))
2499 		return (ENODEV);
2500 
2501 	/* Construct global filesystem path from vp. */
2502 	VOP_UNLOCK(vp, 0);
2503 	error = vn_fullpath_global(td, vp, &rpath, &fbuf);
2504 
2505 	if (error != 0) {
2506 		vrele(vp);
2507 		return (error);
2508 	}
2509 
2510 	if (strlen(rpath) >= pathlen) {
2511 		vrele(vp);
2512 		error = ENAMETOOLONG;
2513 		goto out;
2514 	}
2515 
2516 	/*
2517 	 * Re-lookup the vnode by path to detect a possible rename.
2518 	 * As a side effect, the vnode is relocked.
2519 	 * If vnode was renamed, return ENOENT.
2520 	 */
2521 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
2522 	    UIO_SYSSPACE, path, td);
2523 	error = namei(&nd);
2524 	if (error != 0) {
2525 		vrele(vp);
2526 		goto out;
2527 	}
2528 	NDFREE(&nd, NDF_ONLY_PNBUF);
2529 	vp1 = nd.ni_vp;
2530 	vrele(vp);
2531 	if (vp1 == vp)
2532 		strcpy(path, rpath);
2533 	else {
2534 		vput(vp1);
2535 		error = ENOENT;
2536 	}
2537 
2538 out:
2539 	free(fbuf, M_TEMP);
2540 	return (error);
2541 }
2542 
2543 #ifdef DDB
2544 static void
2545 db_print_vpath(struct vnode *vp)
2546 {
2547 
2548 	while (vp != NULL) {
2549 		db_printf("%p: ", vp);
2550 		if (vp == rootvnode) {
2551 			db_printf("/");
2552 			vp = NULL;
2553 		} else {
2554 			if (vp->v_vflag & VV_ROOT) {
2555 				db_printf("<mount point>");
2556 				vp = vp->v_mount->mnt_vnodecovered;
2557 			} else {
2558 				struct namecache *ncp;
2559 				char *ncn;
2560 				int i;
2561 
2562 				ncp = TAILQ_FIRST(&vp->v_cache_dst);
2563 				if (ncp != NULL) {
2564 					ncn = ncp->nc_name;
2565 					for (i = 0; i < ncp->nc_nlen; i++)
2566 						db_printf("%c", *ncn++);
2567 					vp = ncp->nc_dvp;
2568 				} else {
2569 					vp = NULL;
2570 				}
2571 			}
2572 		}
2573 		db_printf("\n");
2574 	}
2575 
2576 	return;
2577 }
2578 
2579 DB_SHOW_COMMAND(vpath, db_show_vpath)
2580 {
2581 	struct vnode *vp;
2582 
2583 	if (!have_addr) {
2584 		db_printf("usage: show vpath <struct vnode *>\n");
2585 		return;
2586 	}
2587 
2588 	vp = (struct vnode *)addr;
2589 	db_print_vpath(vp);
2590 }
2591 
2592 #endif
2593