xref: /freebsd/sys/kern/vfs_cache.c (revision e7dd6e9402cae324c2190a70081854c3c8a8feb9)
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/ktr.h>
50 #include <sys/lock.h>
51 #include <sys/malloc.h>
52 #include <sys/fcntl.h>
53 #include <sys/mount.h>
54 #include <sys/namei.h>
55 #include <sys/proc.h>
56 #include <sys/rwlock.h>
57 #include <sys/sdt.h>
58 #include <sys/smp.h>
59 #include <sys/syscallsubr.h>
60 #include <sys/sysctl.h>
61 #include <sys/sysproto.h>
62 #include <sys/vnode.h>
63 #ifdef KTRACE
64 #include <sys/ktrace.h>
65 #endif
66 
67 #ifdef DDB
68 #include <ddb/ddb.h>
69 #endif
70 
71 #include <vm/uma.h>
72 
73 SDT_PROVIDER_DECLARE(vfs);
74 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
75     "struct vnode *");
76 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
77     "char *");
78 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
79 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
80     "char *", "struct vnode *");
81 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
82 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
83     "struct vnode *", "char *");
84 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
85     "struct vnode *");
86 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
87     "struct vnode *", "char *");
88 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
89     "char *");
90 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *");
91 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
92 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
93 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
94     "struct vnode *");
95 SDT_PROBE_DEFINE3(vfs, namecache, zap_negative, done, "struct vnode *",
96     "char *", "int");
97 SDT_PROBE_DEFINE3(vfs, namecache, shrink_negative, done, "struct vnode *",
98     "char *", "int");
99 
100 /*
101  * This structure describes the elements in the cache of recent
102  * names looked up by namei.
103  */
104 
105 struct	namecache {
106 	LIST_ENTRY(namecache) nc_hash;	/* hash chain */
107 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
108 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
109 	struct	vnode *nc_dvp;		/* vnode of parent of name */
110 	union {
111 		struct	vnode *nu_vp;	/* vnode the name refers to */
112 		u_int	nu_neghits;	/* negative entry hits */
113 	} n_un;
114 	u_char	nc_flag;		/* flag bits */
115 	u_char	nc_nlen;		/* length of name */
116 	char	nc_name[0];		/* segment name + nul */
117 };
118 
119 /*
120  * struct namecache_ts repeats struct namecache layout up to the
121  * nc_nlen member.
122  * struct namecache_ts is used in place of struct namecache when time(s) need
123  * to be stored.  The nc_dotdottime field is used when a cache entry is mapping
124  * both a non-dotdot directory name plus dotdot for the directory's
125  * parent.
126  */
127 struct	namecache_ts {
128 	struct	timespec nc_time;	/* timespec provided by fs */
129 	struct	timespec nc_dotdottime;	/* dotdot timespec provided by fs */
130 	int	nc_ticks;		/* ticks value when entry was added */
131 	struct namecache nc_nc;
132 };
133 
134 #define	nc_vp		n_un.nu_vp
135 #define	nc_neghits	n_un.nu_neghits
136 
137 /*
138  * Flags in namecache.nc_flag
139  */
140 #define NCF_WHITE	0x01
141 #define NCF_ISDOTDOT	0x02
142 #define	NCF_TS		0x04
143 #define	NCF_DTS		0x08
144 #define	NCF_DVDROP	0x10
145 #define	NCF_NEGATIVE	0x20
146 #define	NCF_HOTNEGATIVE	0x40
147 
148 /*
149  * Name caching works as follows:
150  *
151  * Names found by directory scans are retained in a cache
152  * for future reference.  It is managed LRU, so frequently
153  * used names will hang around.  Cache is indexed by hash value
154  * obtained from (dvp, name) where dvp refers to the directory
155  * containing name.
156  *
157  * If it is a "negative" entry, (i.e. for a name that is known NOT to
158  * exist) the vnode pointer will be NULL.
159  *
160  * Upon reaching the last segment of a path, if the reference
161  * is for DELETE, or NOCACHE is set (rewrite), and the
162  * name is located in the cache, it will be dropped.
163  *
164  * These locks are used (in the order in which they can be taken):
165  * NAME		TYPE	ROLE
166  * vnodelock	mtx	vnode lists and v_cache_dd field protection
167  * bucketlock	rwlock	for access to given set of hash buckets
168  * neglist	mtx	negative entry LRU management
169  *
170  * Additionally, ncneg_shrink_lock mtx is used to have at most one thread
171  * shrinking the LRU list.
172  *
173  * It is legal to take multiple vnodelock and bucketlock locks. The locking
174  * order is lower address first. Both are recursive.
175  *
176  * "." lookups are lockless.
177  *
178  * ".." and vnode -> name lookups require vnodelock.
179  *
180  * name -> vnode lookup requires the relevant bucketlock to be held for reading.
181  *
182  * Insertions and removals of entries require involved vnodes and bucketlocks
183  * to be write-locked to prevent other threads from seeing the entry.
184  *
185  * Some lookups result in removal of the found entry (e.g. getting rid of a
186  * negative entry with the intent to create a positive one), which poses a
187  * problem when multiple threads reach the state. Similarly, two different
188  * threads can purge two different vnodes and try to remove the same name.
189  *
190  * If the already held vnode lock is lower than the second required lock, we
191  * can just take the other lock. However, in the opposite case, this could
192  * deadlock. As such, this is resolved by trylocking and if that fails unlocking
193  * the first node, locking everything in order and revalidating the state.
194  */
195 
196 /*
197  * Structures associated with name caching.
198  */
199 #define NCHHASH(hash) \
200 	(&nchashtbl[(hash) & nchash])
201 static __read_mostly LIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */
202 static u_long __read_mostly	nchash;			/* size of hash table */
203 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
204     "Size of namecache hash table");
205 static u_long __read_mostly	ncnegfactor = 12; /* ratio of negative entries */
206 SYSCTL_ULONG(_vfs, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
207     "Ratio of negative namecache entries");
208 static u_long __exclusive_cache_line	numneg;	/* number of negative entries allocated */
209 SYSCTL_ULONG(_debug, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0,
210     "Number of negative entries in namecache");
211 static u_long __exclusive_cache_line	numcache;/* number of cache entries allocated */
212 SYSCTL_ULONG(_debug, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0,
213     "Number of namecache entries");
214 static u_long __exclusive_cache_line	numcachehv;/* number of cache entries with vnodes held */
215 SYSCTL_ULONG(_debug, OID_AUTO, numcachehv, CTLFLAG_RD, &numcachehv, 0,
216     "Number of namecache entries with vnodes held");
217 u_int __read_mostly	ncsizefactor = 2;
218 SYSCTL_UINT(_vfs, OID_AUTO, ncsizefactor, CTLFLAG_RW, &ncsizefactor, 0,
219     "Size factor for namecache");
220 static u_int __read_mostly	ncpurgeminvnodes;
221 SYSCTL_UINT(_vfs, OID_AUTO, ncpurgeminvnodes, CTLFLAG_RW, &ncpurgeminvnodes, 0,
222     "Number of vnodes below which purgevfs ignores the request");
223 static u_int __read_mostly	ncneghitsrequeue = 8;
224 SYSCTL_UINT(_vfs, OID_AUTO, ncneghitsrequeue, CTLFLAG_RW, &ncneghitsrequeue, 0,
225     "Number of hits to requeue a negative entry in the LRU list");
226 
227 struct nchstats	nchstats;		/* cache effectiveness statistics */
228 
229 static struct mtx __exclusive_cache_line	ncneg_shrink_lock;
230 static int	shrink_list_turn;
231 
232 struct neglist {
233 	struct mtx		nl_lock;
234 	TAILQ_HEAD(, namecache) nl_list;
235 } __aligned(CACHE_LINE_SIZE);
236 
237 static struct neglist __read_mostly	*neglists;
238 static struct neglist ncneg_hot;
239 static u_long numhotneg;
240 
241 #define	numneglists (ncneghash + 1)
242 static u_int __read_mostly	ncneghash;
243 static inline struct neglist *
244 NCP2NEGLIST(struct namecache *ncp)
245 {
246 
247 	return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
248 }
249 
250 #define	numbucketlocks (ncbuckethash + 1)
251 static u_int __read_mostly  ncbuckethash;
252 static struct rwlock_padalign __read_mostly  *bucketlocks;
253 #define	HASH2BUCKETLOCK(hash) \
254 	((struct rwlock *)(&bucketlocks[((hash) & ncbuckethash)]))
255 
256 #define	numvnodelocks (ncvnodehash + 1)
257 static u_int __read_mostly  ncvnodehash;
258 static struct mtx __read_mostly *vnodelocks;
259 static inline struct mtx *
260 VP2VNODELOCK(struct vnode *vp)
261 {
262 
263 	return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
264 }
265 
266 /*
267  * UMA zones for the VFS cache.
268  *
269  * The small cache is used for entries with short names, which are the
270  * most common.  The large cache is used for entries which are too big to
271  * fit in the small cache.
272  */
273 static uma_zone_t __read_mostly cache_zone_small;
274 static uma_zone_t __read_mostly cache_zone_small_ts;
275 static uma_zone_t __read_mostly cache_zone_large;
276 static uma_zone_t __read_mostly cache_zone_large_ts;
277 
278 #define	CACHE_PATH_CUTOFF	35
279 
280 static struct namecache *
281 cache_alloc(int len, int ts)
282 {
283 	struct namecache_ts *ncp_ts;
284 	struct namecache *ncp;
285 
286 	if (__predict_false(ts)) {
287 		if (len <= CACHE_PATH_CUTOFF)
288 			ncp_ts = uma_zalloc(cache_zone_small_ts, M_WAITOK);
289 		else
290 			ncp_ts = uma_zalloc(cache_zone_large_ts, M_WAITOK);
291 		ncp = &ncp_ts->nc_nc;
292 	} else {
293 		if (len <= CACHE_PATH_CUTOFF)
294 			ncp = uma_zalloc(cache_zone_small, M_WAITOK);
295 		else
296 			ncp = uma_zalloc(cache_zone_large, M_WAITOK);
297 	}
298 	return (ncp);
299 }
300 
301 static void
302 cache_free(struct namecache *ncp)
303 {
304 	struct namecache_ts *ncp_ts;
305 
306 	if (ncp == NULL)
307 		return;
308 	if ((ncp->nc_flag & NCF_DVDROP) != 0)
309 		vdrop(ncp->nc_dvp);
310 	if (__predict_false(ncp->nc_flag & NCF_TS)) {
311 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
312 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
313 			uma_zfree(cache_zone_small_ts, ncp_ts);
314 		else
315 			uma_zfree(cache_zone_large_ts, ncp_ts);
316 	} else {
317 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
318 			uma_zfree(cache_zone_small, ncp);
319 		else
320 			uma_zfree(cache_zone_large, ncp);
321 	}
322 }
323 
324 static void
325 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
326 {
327 	struct namecache_ts *ncp_ts;
328 
329 	KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
330 	    (tsp == NULL && ticksp == NULL),
331 	    ("No NCF_TS"));
332 
333 	if (tsp == NULL && ticksp == NULL)
334 		return;
335 
336 	ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
337 	if (tsp != NULL)
338 		*tsp = ncp_ts->nc_time;
339 	if (ticksp != NULL)
340 		*ticksp = ncp_ts->nc_ticks;
341 }
342 
343 static int __read_mostly	doingcache = 1;	/* 1 => enable the cache */
344 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
345     "VFS namecache enabled");
346 
347 /* Export size information to userland */
348 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
349     sizeof(struct namecache), "sizeof(struct namecache)");
350 
351 /*
352  * The new name cache statistics
353  */
354 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0,
355     "Name cache statistics");
356 #define STATNODE_ULONG(name, descr)	\
357 	SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, descr);
358 #define STATNODE_COUNTER(name, descr)	\
359 	static counter_u64_t __read_mostly name; \
360 	SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, descr);
361 STATNODE_ULONG(numneg, "Number of negative cache entries");
362 STATNODE_ULONG(numcache, "Number of cache entries");
363 STATNODE_COUNTER(numcalls, "Number of cache lookups");
364 STATNODE_COUNTER(dothits, "Number of '.' hits");
365 STATNODE_COUNTER(dotdothits, "Number of '..' hits");
366 STATNODE_COUNTER(numchecks, "Number of checks in lookup");
367 STATNODE_COUNTER(nummiss, "Number of cache misses");
368 STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache");
369 STATNODE_COUNTER(numposzaps,
370     "Number of cache hits (positive) we do not want to cache");
371 STATNODE_COUNTER(numposhits, "Number of cache hits (positive)");
372 STATNODE_COUNTER(numnegzaps,
373     "Number of cache hits (negative) we do not want to cache");
374 STATNODE_COUNTER(numneghits, "Number of cache hits (negative)");
375 /* These count for kern___getcwd(), too. */
376 STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls");
377 STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
378 STATNODE_COUNTER(numfullpathfail2,
379     "Number of fullpath search errors (VOP_VPTOCNP failures)");
380 STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
381 STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls");
382 static long numneg_evicted; STATNODE_ULONG(numneg_evicted,
383     "Number of negative entries evicted when adding a new entry");
384 STATNODE_COUNTER(zap_and_exit_bucket_relock_success,
385     "Number of successful removals after relocking");
386 static long zap_and_exit_bucket_fail; STATNODE_ULONG(zap_and_exit_bucket_fail,
387     "Number of times zap_and_exit failed to lock");
388 static long zap_and_exit_bucket_fail2; STATNODE_ULONG(zap_and_exit_bucket_fail2,
389     "Number of times zap_and_exit failed to lock");
390 static long cache_lock_vnodes_cel_3_failures;
391 STATNODE_ULONG(cache_lock_vnodes_cel_3_failures,
392     "Number of times 3-way vnode locking failed");
393 STATNODE_ULONG(numhotneg, "Number of hot negative entries");
394 
395 static void cache_zap_locked(struct namecache *ncp, bool neg_locked);
396 static int vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
397     char *buf, char **retbuf, u_int buflen);
398 
399 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
400 
401 static int cache_yield;
402 SYSCTL_INT(_vfs_cache, OID_AUTO, yield, CTLFLAG_RD, &cache_yield, 0,
403     "Number of times cache called yield");
404 
405 static void __noinline
406 cache_maybe_yield(void)
407 {
408 
409 	if (should_yield()) {
410 		cache_yield++;
411 		kern_yield(PRI_USER);
412 	}
413 }
414 
415 static inline void
416 cache_assert_vlp_locked(struct mtx *vlp)
417 {
418 
419 	if (vlp != NULL)
420 		mtx_assert(vlp, MA_OWNED);
421 }
422 
423 static inline void
424 cache_assert_vnode_locked(struct vnode *vp)
425 {
426 	struct mtx *vlp;
427 
428 	vlp = VP2VNODELOCK(vp);
429 	cache_assert_vlp_locked(vlp);
430 }
431 
432 static uint32_t
433 cache_get_hash(char *name, u_char len, struct vnode *dvp)
434 {
435 	uint32_t hash;
436 
437 	hash = fnv_32_buf(name, len, FNV1_32_INIT);
438 	hash = fnv_32_buf(&dvp, sizeof(dvp), hash);
439 	return (hash);
440 }
441 
442 static inline struct rwlock *
443 NCP2BUCKETLOCK(struct namecache *ncp)
444 {
445 	uint32_t hash;
446 
447 	hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
448 	return (HASH2BUCKETLOCK(hash));
449 }
450 
451 #ifdef INVARIANTS
452 static void
453 cache_assert_bucket_locked(struct namecache *ncp, int mode)
454 {
455 	struct rwlock *blp;
456 
457 	blp = NCP2BUCKETLOCK(ncp);
458 	rw_assert(blp, mode);
459 }
460 #else
461 #define cache_assert_bucket_locked(x, y) do { } while (0)
462 #endif
463 
464 #define cache_sort_vnodes(x, y)	_cache_sort_vnodes((void **)(x), (void **)(y))
465 static void
466 _cache_sort_vnodes(void **p1, void **p2)
467 {
468 	void *tmp;
469 
470 	MPASS(*p1 != NULL || *p2 != NULL);
471 
472 	if (*p1 > *p2) {
473 		tmp = *p2;
474 		*p2 = *p1;
475 		*p1 = tmp;
476 	}
477 }
478 
479 static void
480 cache_lock_all_buckets(void)
481 {
482 	u_int i;
483 
484 	for (i = 0; i < numbucketlocks; i++)
485 		rw_wlock(&bucketlocks[i]);
486 }
487 
488 static void
489 cache_unlock_all_buckets(void)
490 {
491 	u_int i;
492 
493 	for (i = 0; i < numbucketlocks; i++)
494 		rw_wunlock(&bucketlocks[i]);
495 }
496 
497 static void
498 cache_lock_all_vnodes(void)
499 {
500 	u_int i;
501 
502 	for (i = 0; i < numvnodelocks; i++)
503 		mtx_lock(&vnodelocks[i]);
504 }
505 
506 static void
507 cache_unlock_all_vnodes(void)
508 {
509 	u_int i;
510 
511 	for (i = 0; i < numvnodelocks; i++)
512 		mtx_unlock(&vnodelocks[i]);
513 }
514 
515 static int
516 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
517 {
518 
519 	cache_sort_vnodes(&vlp1, &vlp2);
520 
521 	if (vlp1 != NULL) {
522 		if (!mtx_trylock(vlp1))
523 			return (EAGAIN);
524 	}
525 	if (!mtx_trylock(vlp2)) {
526 		if (vlp1 != NULL)
527 			mtx_unlock(vlp1);
528 		return (EAGAIN);
529 	}
530 
531 	return (0);
532 }
533 
534 static void
535 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
536 {
537 
538 	MPASS(vlp1 != NULL || vlp2 != NULL);
539 	MPASS(vlp1 <= vlp2);
540 
541 	if (vlp1 != NULL)
542 		mtx_lock(vlp1);
543 	if (vlp2 != NULL)
544 		mtx_lock(vlp2);
545 }
546 
547 static void
548 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
549 {
550 
551 	MPASS(vlp1 != NULL || vlp2 != NULL);
552 
553 	if (vlp1 != NULL)
554 		mtx_unlock(vlp1);
555 	if (vlp2 != NULL)
556 		mtx_unlock(vlp2);
557 }
558 
559 static int
560 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
561 {
562 	struct nchstats snap;
563 
564 	if (req->oldptr == NULL)
565 		return (SYSCTL_OUT(req, 0, sizeof(snap)));
566 
567 	snap = nchstats;
568 	snap.ncs_goodhits = counter_u64_fetch(numposhits);
569 	snap.ncs_neghits = counter_u64_fetch(numneghits);
570 	snap.ncs_badhits = counter_u64_fetch(numposzaps) +
571 	    counter_u64_fetch(numnegzaps);
572 	snap.ncs_miss = counter_u64_fetch(nummisszap) +
573 	    counter_u64_fetch(nummiss);
574 
575 	return (SYSCTL_OUT(req, &snap, sizeof(snap)));
576 }
577 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
578     CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
579     "VFS cache effectiveness statistics");
580 
581 #ifdef DIAGNOSTIC
582 /*
583  * Grab an atomic snapshot of the name cache hash chain lengths
584  */
585 static SYSCTL_NODE(_debug, OID_AUTO, hashstat, CTLFLAG_RW, NULL,
586     "hash table stats");
587 
588 static int
589 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
590 {
591 	struct nchashhead *ncpp;
592 	struct namecache *ncp;
593 	int i, error, n_nchash, *cntbuf;
594 
595 retry:
596 	n_nchash = nchash + 1;	/* nchash is max index, not count */
597 	if (req->oldptr == NULL)
598 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
599 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
600 	cache_lock_all_buckets();
601 	if (n_nchash != nchash + 1) {
602 		cache_unlock_all_buckets();
603 		free(cntbuf, M_TEMP);
604 		goto retry;
605 	}
606 	/* Scan hash tables counting entries */
607 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
608 		LIST_FOREACH(ncp, ncpp, nc_hash)
609 			cntbuf[i]++;
610 	cache_unlock_all_buckets();
611 	for (error = 0, i = 0; i < n_nchash; i++)
612 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
613 			break;
614 	free(cntbuf, M_TEMP);
615 	return (error);
616 }
617 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
618     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
619     "nchash chain lengths");
620 
621 static int
622 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
623 {
624 	int error;
625 	struct nchashhead *ncpp;
626 	struct namecache *ncp;
627 	int n_nchash;
628 	int count, maxlength, used, pct;
629 
630 	if (!req->oldptr)
631 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
632 
633 	cache_lock_all_buckets();
634 	n_nchash = nchash + 1;	/* nchash is max index, not count */
635 	used = 0;
636 	maxlength = 0;
637 
638 	/* Scan hash tables for applicable entries */
639 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
640 		count = 0;
641 		LIST_FOREACH(ncp, ncpp, nc_hash) {
642 			count++;
643 		}
644 		if (count)
645 			used++;
646 		if (maxlength < count)
647 			maxlength = count;
648 	}
649 	n_nchash = nchash + 1;
650 	cache_unlock_all_buckets();
651 	pct = (used * 100) / (n_nchash / 100);
652 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
653 	if (error)
654 		return (error);
655 	error = SYSCTL_OUT(req, &used, sizeof(used));
656 	if (error)
657 		return (error);
658 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
659 	if (error)
660 		return (error);
661 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
662 	if (error)
663 		return (error);
664 	return (0);
665 }
666 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
667     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
668     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
669 #endif
670 
671 /*
672  * Negative entries management
673  *
674  * A variation of LRU scheme is used. New entries are hashed into one of
675  * numneglists cold lists. Entries get promoted to the hot list on first hit.
676  * Partial LRU for the hot list is maintained by requeueing them every
677  * ncneghitsrequeue hits.
678  *
679  * The shrinker will demote hot list head and evict from the cold list in a
680  * round-robin manner.
681  */
682 static void
683 cache_negative_hit(struct namecache *ncp)
684 {
685 	struct neglist *neglist;
686 	u_int hits;
687 
688 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
689 	hits = atomic_fetchadd_int(&ncp->nc_neghits, 1);
690 	if (ncp->nc_flag & NCF_HOTNEGATIVE) {
691 		if ((hits % ncneghitsrequeue) != 0)
692 			return;
693 		mtx_lock(&ncneg_hot.nl_lock);
694 		if (ncp->nc_flag & NCF_HOTNEGATIVE) {
695 			TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
696 			TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
697 			mtx_unlock(&ncneg_hot.nl_lock);
698 			return;
699 		}
700 		/*
701 		 * The shrinker cleared the flag and removed the entry from
702 		 * the hot list. Put it back.
703 		 */
704 	} else {
705 		mtx_lock(&ncneg_hot.nl_lock);
706 	}
707 	neglist = NCP2NEGLIST(ncp);
708 	mtx_lock(&neglist->nl_lock);
709 	if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
710 		numhotneg++;
711 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
712 		TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
713 		ncp->nc_flag |= NCF_HOTNEGATIVE;
714 	}
715 	mtx_unlock(&neglist->nl_lock);
716 	mtx_unlock(&ncneg_hot.nl_lock);
717 }
718 
719 static void
720 cache_negative_insert(struct namecache *ncp, bool neg_locked)
721 {
722 	struct neglist *neglist;
723 
724 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
725 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
726 	neglist = NCP2NEGLIST(ncp);
727 	if (!neg_locked) {
728 		mtx_lock(&neglist->nl_lock);
729 	} else {
730 		mtx_assert(&neglist->nl_lock, MA_OWNED);
731 	}
732 	TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
733 	if (!neg_locked)
734 		mtx_unlock(&neglist->nl_lock);
735 	atomic_add_rel_long(&numneg, 1);
736 }
737 
738 static void
739 cache_negative_remove(struct namecache *ncp, bool neg_locked)
740 {
741 	struct neglist *neglist;
742 	bool hot_locked = false;
743 	bool list_locked = false;
744 
745 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
746 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
747 	neglist = NCP2NEGLIST(ncp);
748 	if (!neg_locked) {
749 		if (ncp->nc_flag & NCF_HOTNEGATIVE) {
750 			hot_locked = true;
751 			mtx_lock(&ncneg_hot.nl_lock);
752 			if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
753 				list_locked = true;
754 				mtx_lock(&neglist->nl_lock);
755 			}
756 		} else {
757 			list_locked = true;
758 			mtx_lock(&neglist->nl_lock);
759 		}
760 	}
761 	if (ncp->nc_flag & NCF_HOTNEGATIVE) {
762 		mtx_assert(&ncneg_hot.nl_lock, MA_OWNED);
763 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
764 		numhotneg--;
765 	} else {
766 		mtx_assert(&neglist->nl_lock, MA_OWNED);
767 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
768 	}
769 	if (list_locked)
770 		mtx_unlock(&neglist->nl_lock);
771 	if (hot_locked)
772 		mtx_unlock(&ncneg_hot.nl_lock);
773 	atomic_subtract_rel_long(&numneg, 1);
774 }
775 
776 static void
777 cache_negative_shrink_select(int start, struct namecache **ncpp,
778     struct neglist **neglistpp)
779 {
780 	struct neglist *neglist;
781 	struct namecache *ncp;
782 	int i;
783 
784 	*ncpp = ncp = NULL;
785 	neglist = NULL;
786 
787 	for (i = start; i < numneglists; i++) {
788 		neglist = &neglists[i];
789 		if (TAILQ_FIRST(&neglist->nl_list) == NULL)
790 			continue;
791 		mtx_lock(&neglist->nl_lock);
792 		ncp = TAILQ_FIRST(&neglist->nl_list);
793 		if (ncp != NULL)
794 			break;
795 		mtx_unlock(&neglist->nl_lock);
796 	}
797 
798 	*neglistpp = neglist;
799 	*ncpp = ncp;
800 }
801 
802 static void
803 cache_negative_zap_one(void)
804 {
805 	struct namecache *ncp, *ncp2;
806 	struct neglist *neglist;
807 	struct mtx *dvlp;
808 	struct rwlock *blp;
809 
810 	if (mtx_owner(&ncneg_shrink_lock) != NULL ||
811 	    !mtx_trylock(&ncneg_shrink_lock))
812 		return;
813 
814 	mtx_lock(&ncneg_hot.nl_lock);
815 	ncp = TAILQ_FIRST(&ncneg_hot.nl_list);
816 	if (ncp != NULL) {
817 		neglist = NCP2NEGLIST(ncp);
818 		mtx_lock(&neglist->nl_lock);
819 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
820 		TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
821 		ncp->nc_flag &= ~NCF_HOTNEGATIVE;
822 		numhotneg--;
823 		mtx_unlock(&neglist->nl_lock);
824 	}
825 	mtx_unlock(&ncneg_hot.nl_lock);
826 
827 	cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
828 	shrink_list_turn++;
829 	if (shrink_list_turn == numneglists)
830 		shrink_list_turn = 0;
831 	if (ncp == NULL && shrink_list_turn == 0)
832 		cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
833 	if (ncp == NULL)
834 		goto out;
835 
836 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
837 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
838 	blp = NCP2BUCKETLOCK(ncp);
839 	mtx_unlock(&neglist->nl_lock);
840 	mtx_lock(dvlp);
841 	rw_wlock(blp);
842 	mtx_lock(&neglist->nl_lock);
843 	ncp2 = TAILQ_FIRST(&neglist->nl_list);
844 	if (ncp != ncp2 || dvlp != VP2VNODELOCK(ncp2->nc_dvp) ||
845 	    blp != NCP2BUCKETLOCK(ncp2) || !(ncp2->nc_flag & NCF_NEGATIVE)) {
846 		ncp = NULL;
847 		goto out_unlock_all;
848 	}
849 	SDT_PROBE3(vfs, namecache, shrink_negative, done, ncp->nc_dvp,
850 	    ncp->nc_name, ncp->nc_neghits);
851 
852 	cache_zap_locked(ncp, true);
853 	numneg_evicted++;
854 out_unlock_all:
855 	mtx_unlock(&neglist->nl_lock);
856 	rw_wunlock(blp);
857 	mtx_unlock(dvlp);
858 out:
859 	mtx_unlock(&ncneg_shrink_lock);
860 	cache_free(ncp);
861 }
862 
863 /*
864  * cache_zap_locked():
865  *
866  *   Removes a namecache entry from cache, whether it contains an actual
867  *   pointer to a vnode or if it is just a negative cache entry.
868  */
869 static void
870 cache_zap_locked(struct namecache *ncp, bool neg_locked)
871 {
872 
873 	if (!(ncp->nc_flag & NCF_NEGATIVE))
874 		cache_assert_vnode_locked(ncp->nc_vp);
875 	cache_assert_vnode_locked(ncp->nc_dvp);
876 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
877 
878 	CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp,
879 	    (ncp->nc_flag & NCF_NEGATIVE) ? NULL : ncp->nc_vp);
880 	LIST_REMOVE(ncp, nc_hash);
881 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
882 		SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
883 		    ncp->nc_name, ncp->nc_vp);
884 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
885 		if (ncp == ncp->nc_vp->v_cache_dd)
886 			ncp->nc_vp->v_cache_dd = NULL;
887 	} else {
888 		SDT_PROBE3(vfs, namecache, zap_negative, done, ncp->nc_dvp,
889 		    ncp->nc_name, ncp->nc_neghits);
890 		cache_negative_remove(ncp, neg_locked);
891 	}
892 	if (ncp->nc_flag & NCF_ISDOTDOT) {
893 		if (ncp == ncp->nc_dvp->v_cache_dd)
894 			ncp->nc_dvp->v_cache_dd = NULL;
895 	} else {
896 		LIST_REMOVE(ncp, nc_src);
897 		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
898 			ncp->nc_flag |= NCF_DVDROP;
899 			atomic_subtract_rel_long(&numcachehv, 1);
900 		}
901 	}
902 	atomic_subtract_rel_long(&numcache, 1);
903 }
904 
905 static void
906 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
907 {
908 	struct rwlock *blp;
909 
910 	MPASS(ncp->nc_dvp == vp);
911 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
912 	cache_assert_vnode_locked(vp);
913 
914 	blp = NCP2BUCKETLOCK(ncp);
915 	rw_wlock(blp);
916 	cache_zap_locked(ncp, false);
917 	rw_wunlock(blp);
918 }
919 
920 static bool
921 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
922     struct mtx **vlpp)
923 {
924 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
925 	struct rwlock *blp;
926 
927 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
928 	cache_assert_vnode_locked(vp);
929 
930 	if (ncp->nc_flag & NCF_NEGATIVE) {
931 		if (*vlpp != NULL) {
932 			mtx_unlock(*vlpp);
933 			*vlpp = NULL;
934 		}
935 		cache_zap_negative_locked_vnode_kl(ncp, vp);
936 		return (true);
937 	}
938 
939 	pvlp = VP2VNODELOCK(vp);
940 	blp = NCP2BUCKETLOCK(ncp);
941 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
942 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
943 
944 	if (*vlpp == vlp1 || *vlpp == vlp2) {
945 		to_unlock = *vlpp;
946 		*vlpp = NULL;
947 	} else {
948 		if (*vlpp != NULL) {
949 			mtx_unlock(*vlpp);
950 			*vlpp = NULL;
951 		}
952 		cache_sort_vnodes(&vlp1, &vlp2);
953 		if (vlp1 == pvlp) {
954 			mtx_lock(vlp2);
955 			to_unlock = vlp2;
956 		} else {
957 			if (!mtx_trylock(vlp1))
958 				goto out_relock;
959 			to_unlock = vlp1;
960 		}
961 	}
962 	rw_wlock(blp);
963 	cache_zap_locked(ncp, false);
964 	rw_wunlock(blp);
965 	if (to_unlock != NULL)
966 		mtx_unlock(to_unlock);
967 	return (true);
968 
969 out_relock:
970 	mtx_unlock(vlp2);
971 	mtx_lock(vlp1);
972 	mtx_lock(vlp2);
973 	MPASS(*vlpp == NULL);
974 	*vlpp = vlp1;
975 	return (false);
976 }
977 
978 static int __noinline
979 cache_zap_locked_vnode(struct namecache *ncp, struct vnode *vp)
980 {
981 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
982 	struct rwlock *blp;
983 	int error = 0;
984 
985 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
986 	cache_assert_vnode_locked(vp);
987 
988 	pvlp = VP2VNODELOCK(vp);
989 	if (ncp->nc_flag & NCF_NEGATIVE) {
990 		cache_zap_negative_locked_vnode_kl(ncp, vp);
991 		goto out;
992 	}
993 
994 	blp = NCP2BUCKETLOCK(ncp);
995 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
996 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
997 	cache_sort_vnodes(&vlp1, &vlp2);
998 	if (vlp1 == pvlp) {
999 		mtx_lock(vlp2);
1000 		to_unlock = vlp2;
1001 	} else {
1002 		if (!mtx_trylock(vlp1)) {
1003 			error = EAGAIN;
1004 			goto out;
1005 		}
1006 		to_unlock = vlp1;
1007 	}
1008 	rw_wlock(blp);
1009 	cache_zap_locked(ncp, false);
1010 	rw_wunlock(blp);
1011 	mtx_unlock(to_unlock);
1012 out:
1013 	mtx_unlock(pvlp);
1014 	return (error);
1015 }
1016 
1017 /*
1018  * If trylocking failed we can get here. We know enough to take all needed locks
1019  * in the right order and re-lookup the entry.
1020  */
1021 static int
1022 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1023     struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1024     struct rwlock *blp)
1025 {
1026 	struct namecache *rncp;
1027 
1028 	cache_assert_bucket_locked(ncp, RA_UNLOCKED);
1029 
1030 	cache_sort_vnodes(&dvlp, &vlp);
1031 	cache_lock_vnodes(dvlp, vlp);
1032 	rw_wlock(blp);
1033 	LIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1034 		if (rncp == ncp && rncp->nc_dvp == dvp &&
1035 		    rncp->nc_nlen == cnp->cn_namelen &&
1036 		    !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen))
1037 			break;
1038 	}
1039 	if (rncp != NULL) {
1040 		cache_zap_locked(rncp, false);
1041 		rw_wunlock(blp);
1042 		cache_unlock_vnodes(dvlp, vlp);
1043 		counter_u64_add(zap_and_exit_bucket_relock_success, 1);
1044 		return (0);
1045 	}
1046 
1047 	rw_wunlock(blp);
1048 	cache_unlock_vnodes(dvlp, vlp);
1049 	return (EAGAIN);
1050 }
1051 
1052 static int __noinline
1053 cache_zap_wlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1054     uint32_t hash, struct rwlock *blp)
1055 {
1056 	struct mtx *dvlp, *vlp;
1057 	struct vnode *dvp;
1058 
1059 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1060 
1061 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1062 	vlp = NULL;
1063 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1064 		vlp = VP2VNODELOCK(ncp->nc_vp);
1065 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1066 		cache_zap_locked(ncp, false);
1067 		rw_wunlock(blp);
1068 		cache_unlock_vnodes(dvlp, vlp);
1069 		return (0);
1070 	}
1071 
1072 	dvp = ncp->nc_dvp;
1073 	rw_wunlock(blp);
1074 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1075 }
1076 
1077 static int __noinline
1078 cache_zap_rlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1079     uint32_t hash, struct rwlock *blp)
1080 {
1081 	struct mtx *dvlp, *vlp;
1082 	struct vnode *dvp;
1083 
1084 	cache_assert_bucket_locked(ncp, RA_RLOCKED);
1085 
1086 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1087 	vlp = NULL;
1088 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1089 		vlp = VP2VNODELOCK(ncp->nc_vp);
1090 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1091 		rw_runlock(blp);
1092 		rw_wlock(blp);
1093 		cache_zap_locked(ncp, false);
1094 		rw_wunlock(blp);
1095 		cache_unlock_vnodes(dvlp, vlp);
1096 		return (0);
1097 	}
1098 
1099 	dvp = ncp->nc_dvp;
1100 	rw_runlock(blp);
1101 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1102 }
1103 
1104 static int
1105 cache_zap_wlocked_bucket_kl(struct namecache *ncp, struct rwlock *blp,
1106     struct mtx **vlpp1, struct mtx **vlpp2)
1107 {
1108 	struct mtx *dvlp, *vlp;
1109 
1110 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1111 
1112 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1113 	vlp = NULL;
1114 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1115 		vlp = VP2VNODELOCK(ncp->nc_vp);
1116 	cache_sort_vnodes(&dvlp, &vlp);
1117 
1118 	if (*vlpp1 == dvlp && *vlpp2 == vlp) {
1119 		cache_zap_locked(ncp, false);
1120 		cache_unlock_vnodes(dvlp, vlp);
1121 		*vlpp1 = NULL;
1122 		*vlpp2 = NULL;
1123 		return (0);
1124 	}
1125 
1126 	if (*vlpp1 != NULL)
1127 		mtx_unlock(*vlpp1);
1128 	if (*vlpp2 != NULL)
1129 		mtx_unlock(*vlpp2);
1130 	*vlpp1 = NULL;
1131 	*vlpp2 = NULL;
1132 
1133 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1134 		cache_zap_locked(ncp, false);
1135 		cache_unlock_vnodes(dvlp, vlp);
1136 		return (0);
1137 	}
1138 
1139 	rw_wunlock(blp);
1140 	*vlpp1 = dvlp;
1141 	*vlpp2 = vlp;
1142 	if (*vlpp1 != NULL)
1143 		mtx_lock(*vlpp1);
1144 	mtx_lock(*vlpp2);
1145 	rw_wlock(blp);
1146 	return (EAGAIN);
1147 }
1148 
1149 static void
1150 cache_lookup_unlock(struct rwlock *blp, struct mtx *vlp)
1151 {
1152 
1153 	if (blp != NULL) {
1154 		rw_runlock(blp);
1155 	} else {
1156 		mtx_unlock(vlp);
1157 	}
1158 }
1159 
1160 static int __noinline
1161 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1162     struct timespec *tsp, int *ticksp)
1163 {
1164 	int ltype;
1165 
1166 	*vpp = dvp;
1167 	CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .",
1168 			dvp, cnp->cn_nameptr);
1169 	counter_u64_add(dothits, 1);
1170 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1171 	if (tsp != NULL)
1172 		timespecclear(tsp);
1173 	if (ticksp != NULL)
1174 		*ticksp = ticks;
1175 	vrefact(*vpp);
1176 	/*
1177 	 * When we lookup "." we still can be asked to lock it
1178 	 * differently...
1179 	 */
1180 	ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1181 	if (ltype != VOP_ISLOCKED(*vpp)) {
1182 		if (ltype == LK_EXCLUSIVE) {
1183 			vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1184 			if ((*vpp)->v_iflag & VI_DOOMED) {
1185 				/* forced unmount */
1186 				vrele(*vpp);
1187 				*vpp = NULL;
1188 				return (ENOENT);
1189 			}
1190 		} else
1191 			vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1192 	}
1193 	return (-1);
1194 }
1195 
1196 static __noinline int
1197 cache_lookup_nomakeentry(struct vnode *dvp, struct vnode **vpp,
1198     struct componentname *cnp, struct timespec *tsp, int *ticksp)
1199 {
1200 	struct namecache *ncp;
1201 	struct rwlock *blp;
1202 	struct mtx *dvlp, *dvlp2;
1203 	uint32_t hash;
1204 	int error;
1205 
1206 	if (cnp->cn_namelen == 2 &&
1207 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1208 		counter_u64_add(dotdothits, 1);
1209 		dvlp = VP2VNODELOCK(dvp);
1210 		dvlp2 = NULL;
1211 		mtx_lock(dvlp);
1212 retry_dotdot:
1213 		ncp = dvp->v_cache_dd;
1214 		if (ncp == NULL) {
1215 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1216 			    "..", NULL);
1217 			mtx_unlock(dvlp);
1218 			if (dvlp2 != NULL)
1219 				mtx_unlock(dvlp2);
1220 			return (0);
1221 		}
1222 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1223 			if (ncp->nc_dvp != dvp)
1224 				panic("dvp %p v_cache_dd %p\n", dvp, ncp);
1225 			if (!cache_zap_locked_vnode_kl2(ncp,
1226 			    dvp, &dvlp2))
1227 				goto retry_dotdot;
1228 			MPASS(dvp->v_cache_dd == NULL);
1229 			mtx_unlock(dvlp);
1230 			if (dvlp2 != NULL)
1231 				mtx_unlock(dvlp2);
1232 			cache_free(ncp);
1233 		} else {
1234 			dvp->v_cache_dd = NULL;
1235 			mtx_unlock(dvlp);
1236 			if (dvlp2 != NULL)
1237 				mtx_unlock(dvlp2);
1238 		}
1239 		return (0);
1240 	}
1241 
1242 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1243 	blp = HASH2BUCKETLOCK(hash);
1244 retry:
1245 	if (LIST_EMPTY(NCHHASH(hash)))
1246 		goto out_no_entry;
1247 
1248 	rw_wlock(blp);
1249 
1250 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1251 		counter_u64_add(numchecks, 1);
1252 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1253 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1254 			break;
1255 	}
1256 
1257 	/* We failed to find an entry */
1258 	if (ncp == NULL) {
1259 		rw_wunlock(blp);
1260 		goto out_no_entry;
1261 	}
1262 
1263 	error = cache_zap_wlocked_bucket(ncp, cnp, hash, blp);
1264 	if (__predict_false(error != 0)) {
1265 		zap_and_exit_bucket_fail++;
1266 		cache_maybe_yield();
1267 		goto retry;
1268 	}
1269 	counter_u64_add(numposzaps, 1);
1270 	cache_free(ncp);
1271 	return (0);
1272 out_no_entry:
1273 	SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr, NULL);
1274 	counter_u64_add(nummisszap, 1);
1275 	return (0);
1276 }
1277 
1278 /**
1279  * Lookup a name in the name cache
1280  *
1281  * # Arguments
1282  *
1283  * - dvp:	Parent directory in which to search.
1284  * - vpp:	Return argument.  Will contain desired vnode on cache hit.
1285  * - cnp:	Parameters of the name search.  The most interesting bits of
1286  *   		the cn_flags field have the following meanings:
1287  *   	- MAKEENTRY:	If clear, free an entry from the cache rather than look
1288  *   			it up.
1289  *   	- ISDOTDOT:	Must be set if and only if cn_nameptr == ".."
1290  * - tsp:	Return storage for cache timestamp.  On a successful (positive
1291  *   		or negative) lookup, tsp will be filled with any timespec that
1292  *   		was stored when this cache entry was created.  However, it will
1293  *   		be clear for "." entries.
1294  * - ticks:	Return storage for alternate cache timestamp.  On a successful
1295  *   		(positive or negative) lookup, it will contain the ticks value
1296  *   		that was current when the cache entry was created, unless cnp
1297  *   		was ".".
1298  *
1299  * # Returns
1300  *
1301  * - -1:	A positive cache hit.  vpp will contain the desired vnode.
1302  * - ENOENT:	A negative cache hit, or dvp was recycled out from under us due
1303  *		to a forced unmount.  vpp will not be modified.  If the entry
1304  *		is a whiteout, then the ISWHITEOUT flag will be set in
1305  *		cnp->cn_flags.
1306  * - 0:		A cache miss.  vpp will not be modified.
1307  *
1308  * # Locking
1309  *
1310  * On a cache hit, vpp will be returned locked and ref'd.  If we're looking up
1311  * .., dvp is unlocked.  If we're looking up . an extra ref is taken, but the
1312  * lock is not recursively acquired.
1313  */
1314 int
1315 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1316     struct timespec *tsp, int *ticksp)
1317 {
1318 	struct namecache_ts *ncp_ts;
1319 	struct namecache *ncp;
1320 	struct rwlock *blp;
1321 	struct mtx *dvlp;
1322 	uint32_t hash;
1323 	enum vgetstate vs;
1324 	int error, ltype;
1325 
1326 	if (__predict_false(!doingcache)) {
1327 		cnp->cn_flags &= ~MAKEENTRY;
1328 		return (0);
1329 	}
1330 
1331 	counter_u64_add(numcalls, 1);
1332 
1333 	if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.'))
1334 		return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1335 
1336 	if ((cnp->cn_flags & MAKEENTRY) == 0)
1337 		return (cache_lookup_nomakeentry(dvp, vpp, cnp, tsp, ticksp));
1338 
1339 retry:
1340 	blp = NULL;
1341 	dvlp = NULL;
1342 	error = 0;
1343 	if (cnp->cn_namelen == 2 &&
1344 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1345 		counter_u64_add(dotdothits, 1);
1346 		dvlp = VP2VNODELOCK(dvp);
1347 		mtx_lock(dvlp);
1348 		ncp = dvp->v_cache_dd;
1349 		if (ncp == NULL) {
1350 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1351 			    "..", NULL);
1352 			mtx_unlock(dvlp);
1353 			return (0);
1354 		}
1355 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1356 			if (ncp->nc_flag & NCF_NEGATIVE)
1357 				*vpp = NULL;
1358 			else
1359 				*vpp = ncp->nc_vp;
1360 		} else
1361 			*vpp = ncp->nc_dvp;
1362 		/* Return failure if negative entry was found. */
1363 		if (*vpp == NULL)
1364 			goto negative_success;
1365 		CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
1366 		    dvp, cnp->cn_nameptr, *vpp);
1367 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..",
1368 		    *vpp);
1369 		cache_out_ts(ncp, tsp, ticksp);
1370 		if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1371 		    NCF_DTS && tsp != NULL) {
1372 			ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1373 			*tsp = ncp_ts->nc_dotdottime;
1374 		}
1375 		goto success;
1376 	}
1377 
1378 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1379 	blp = HASH2BUCKETLOCK(hash);
1380 	rw_rlock(blp);
1381 
1382 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1383 		counter_u64_add(numchecks, 1);
1384 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1385 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1386 			break;
1387 	}
1388 
1389 	/* We failed to find an entry */
1390 	if (__predict_false(ncp == NULL)) {
1391 		rw_runlock(blp);
1392 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1393 		    NULL);
1394 		counter_u64_add(nummiss, 1);
1395 		return (0);
1396 	}
1397 
1398 	if (ncp->nc_flag & NCF_NEGATIVE)
1399 		goto negative_success;
1400 
1401 	/* We found a "positive" match, return the vnode */
1402 	counter_u64_add(numposhits, 1);
1403 	*vpp = ncp->nc_vp;
1404 	CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p",
1405 	    dvp, cnp->cn_nameptr, *vpp, ncp);
1406 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name,
1407 	    *vpp);
1408 	cache_out_ts(ncp, tsp, ticksp);
1409 success:
1410 	/*
1411 	 * On success we return a locked and ref'd vnode as per the lookup
1412 	 * protocol.
1413 	 */
1414 	MPASS(dvp != *vpp);
1415 	ltype = 0;	/* silence gcc warning */
1416 	if (cnp->cn_flags & ISDOTDOT) {
1417 		ltype = VOP_ISLOCKED(dvp);
1418 		VOP_UNLOCK(dvp, 0);
1419 	}
1420 	vs = vget_prep(*vpp);
1421 	cache_lookup_unlock(blp, dvlp);
1422 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1423 	if (cnp->cn_flags & ISDOTDOT) {
1424 		vn_lock(dvp, ltype | LK_RETRY);
1425 		if (dvp->v_iflag & VI_DOOMED) {
1426 			if (error == 0)
1427 				vput(*vpp);
1428 			*vpp = NULL;
1429 			return (ENOENT);
1430 		}
1431 	}
1432 	if (error) {
1433 		*vpp = NULL;
1434 		goto retry;
1435 	}
1436 	if ((cnp->cn_flags & ISLASTCN) &&
1437 	    (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) {
1438 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
1439 	}
1440 	return (-1);
1441 
1442 negative_success:
1443 	/* We found a negative match, and want to create it, so purge */
1444 	if (cnp->cn_nameiop == CREATE) {
1445 		counter_u64_add(numnegzaps, 1);
1446 		goto zap_and_exit;
1447 	}
1448 
1449 	counter_u64_add(numneghits, 1);
1450 	cache_negative_hit(ncp);
1451 	if (ncp->nc_flag & NCF_WHITE)
1452 		cnp->cn_flags |= ISWHITEOUT;
1453 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp,
1454 	    ncp->nc_name);
1455 	cache_out_ts(ncp, tsp, ticksp);
1456 	cache_lookup_unlock(blp, dvlp);
1457 	return (ENOENT);
1458 
1459 zap_and_exit:
1460 	if (blp != NULL)
1461 		error = cache_zap_rlocked_bucket(ncp, cnp, hash, blp);
1462 	else
1463 		error = cache_zap_locked_vnode(ncp, dvp);
1464 	if (__predict_false(error != 0)) {
1465 		zap_and_exit_bucket_fail2++;
1466 		cache_maybe_yield();
1467 		goto retry;
1468 	}
1469 	cache_free(ncp);
1470 	return (0);
1471 }
1472 
1473 struct celockstate {
1474 	struct mtx *vlp[3];
1475 	struct rwlock *blp[2];
1476 };
1477 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1478 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1479 
1480 static inline void
1481 cache_celockstate_init(struct celockstate *cel)
1482 {
1483 
1484 	bzero(cel, sizeof(*cel));
1485 }
1486 
1487 static void
1488 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1489     struct vnode *dvp)
1490 {
1491 	struct mtx *vlp1, *vlp2;
1492 
1493 	MPASS(cel->vlp[0] == NULL);
1494 	MPASS(cel->vlp[1] == NULL);
1495 	MPASS(cel->vlp[2] == NULL);
1496 
1497 	MPASS(vp != NULL || dvp != NULL);
1498 
1499 	vlp1 = VP2VNODELOCK(vp);
1500 	vlp2 = VP2VNODELOCK(dvp);
1501 	cache_sort_vnodes(&vlp1, &vlp2);
1502 
1503 	if (vlp1 != NULL) {
1504 		mtx_lock(vlp1);
1505 		cel->vlp[0] = vlp1;
1506 	}
1507 	mtx_lock(vlp2);
1508 	cel->vlp[1] = vlp2;
1509 }
1510 
1511 static void
1512 cache_unlock_vnodes_cel(struct celockstate *cel)
1513 {
1514 
1515 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
1516 
1517 	if (cel->vlp[0] != NULL)
1518 		mtx_unlock(cel->vlp[0]);
1519 	if (cel->vlp[1] != NULL)
1520 		mtx_unlock(cel->vlp[1]);
1521 	if (cel->vlp[2] != NULL)
1522 		mtx_unlock(cel->vlp[2]);
1523 }
1524 
1525 static bool
1526 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
1527 {
1528 	struct mtx *vlp;
1529 	bool ret;
1530 
1531 	cache_assert_vlp_locked(cel->vlp[0]);
1532 	cache_assert_vlp_locked(cel->vlp[1]);
1533 	MPASS(cel->vlp[2] == NULL);
1534 
1535 	MPASS(vp != NULL);
1536 	vlp = VP2VNODELOCK(vp);
1537 
1538 	ret = true;
1539 	if (vlp >= cel->vlp[1]) {
1540 		mtx_lock(vlp);
1541 	} else {
1542 		if (mtx_trylock(vlp))
1543 			goto out;
1544 		cache_lock_vnodes_cel_3_failures++;
1545 		cache_unlock_vnodes_cel(cel);
1546 		if (vlp < cel->vlp[0]) {
1547 			mtx_lock(vlp);
1548 			mtx_lock(cel->vlp[0]);
1549 			mtx_lock(cel->vlp[1]);
1550 		} else {
1551 			if (cel->vlp[0] != NULL)
1552 				mtx_lock(cel->vlp[0]);
1553 			mtx_lock(vlp);
1554 			mtx_lock(cel->vlp[1]);
1555 		}
1556 		ret = false;
1557 	}
1558 out:
1559 	cel->vlp[2] = vlp;
1560 	return (ret);
1561 }
1562 
1563 static void
1564 cache_lock_buckets_cel(struct celockstate *cel, struct rwlock *blp1,
1565     struct rwlock *blp2)
1566 {
1567 
1568 	MPASS(cel->blp[0] == NULL);
1569 	MPASS(cel->blp[1] == NULL);
1570 
1571 	cache_sort_vnodes(&blp1, &blp2);
1572 
1573 	if (blp1 != NULL) {
1574 		rw_wlock(blp1);
1575 		cel->blp[0] = blp1;
1576 	}
1577 	rw_wlock(blp2);
1578 	cel->blp[1] = blp2;
1579 }
1580 
1581 static void
1582 cache_unlock_buckets_cel(struct celockstate *cel)
1583 {
1584 
1585 	if (cel->blp[0] != NULL)
1586 		rw_wunlock(cel->blp[0]);
1587 	rw_wunlock(cel->blp[1]);
1588 }
1589 
1590 /*
1591  * Lock part of the cache affected by the insertion.
1592  *
1593  * This means vnodelocks for dvp, vp and the relevant bucketlock.
1594  * However, insertion can result in removal of an old entry. In this
1595  * case we have an additional vnode and bucketlock pair to lock. If the
1596  * entry is negative, ncelock is locked instead of the vnode.
1597  *
1598  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
1599  * preserving the locking order (smaller address first).
1600  */
1601 static void
1602 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1603     uint32_t hash)
1604 {
1605 	struct namecache *ncp;
1606 	struct rwlock *blps[2];
1607 
1608 	blps[0] = HASH2BUCKETLOCK(hash);
1609 	for (;;) {
1610 		blps[1] = NULL;
1611 		cache_lock_vnodes_cel(cel, dvp, vp);
1612 		if (vp == NULL || vp->v_type != VDIR)
1613 			break;
1614 		ncp = vp->v_cache_dd;
1615 		if (ncp == NULL)
1616 			break;
1617 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1618 			break;
1619 		MPASS(ncp->nc_dvp == vp);
1620 		blps[1] = NCP2BUCKETLOCK(ncp);
1621 		if (ncp->nc_flag & NCF_NEGATIVE)
1622 			break;
1623 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1624 			break;
1625 		/*
1626 		 * All vnodes got re-locked. Re-validate the state and if
1627 		 * nothing changed we are done. Otherwise restart.
1628 		 */
1629 		if (ncp == vp->v_cache_dd &&
1630 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1631 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1632 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1633 			break;
1634 		cache_unlock_vnodes_cel(cel);
1635 		cel->vlp[0] = NULL;
1636 		cel->vlp[1] = NULL;
1637 		cel->vlp[2] = NULL;
1638 	}
1639 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1640 }
1641 
1642 static void
1643 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1644     uint32_t hash)
1645 {
1646 	struct namecache *ncp;
1647 	struct rwlock *blps[2];
1648 
1649 	blps[0] = HASH2BUCKETLOCK(hash);
1650 	for (;;) {
1651 		blps[1] = NULL;
1652 		cache_lock_vnodes_cel(cel, dvp, vp);
1653 		ncp = dvp->v_cache_dd;
1654 		if (ncp == NULL)
1655 			break;
1656 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1657 			break;
1658 		MPASS(ncp->nc_dvp == dvp);
1659 		blps[1] = NCP2BUCKETLOCK(ncp);
1660 		if (ncp->nc_flag & NCF_NEGATIVE)
1661 			break;
1662 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1663 			break;
1664 		if (ncp == dvp->v_cache_dd &&
1665 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1666 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1667 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1668 			break;
1669 		cache_unlock_vnodes_cel(cel);
1670 		cel->vlp[0] = NULL;
1671 		cel->vlp[1] = NULL;
1672 		cel->vlp[2] = NULL;
1673 	}
1674 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1675 }
1676 
1677 static void
1678 cache_enter_unlock(struct celockstate *cel)
1679 {
1680 
1681 	cache_unlock_buckets_cel(cel);
1682 	cache_unlock_vnodes_cel(cel);
1683 }
1684 
1685 /*
1686  * Add an entry to the cache.
1687  */
1688 void
1689 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
1690     struct timespec *tsp, struct timespec *dtsp)
1691 {
1692 	struct celockstate cel;
1693 	struct namecache *ncp, *n2, *ndd;
1694 	struct namecache_ts *ncp_ts, *n2_ts;
1695 	struct nchashhead *ncpp;
1696 	struct neglist *neglist;
1697 	uint32_t hash;
1698 	int flag;
1699 	int len;
1700 	bool neg_locked, held_dvp;
1701 	u_long lnumcache;
1702 
1703 	CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr);
1704 	VNASSERT(vp == NULL || (vp->v_iflag & VI_DOOMED) == 0, vp,
1705 	    ("cache_enter: Adding a doomed vnode"));
1706 	VNASSERT(dvp == NULL || (dvp->v_iflag & VI_DOOMED) == 0, dvp,
1707 	    ("cache_enter: Doomed vnode used as src"));
1708 
1709 	if (__predict_false(!doingcache))
1710 		return;
1711 
1712 	/*
1713 	 * Avoid blowout in namecache entries.
1714 	 */
1715 	lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
1716 	if (__predict_false(lnumcache >= desiredvnodes * ncsizefactor)) {
1717 		atomic_add_long(&numcache, -1);
1718 		return;
1719 	}
1720 
1721 	cache_celockstate_init(&cel);
1722 	ndd = NULL;
1723 	ncp_ts = NULL;
1724 	flag = 0;
1725 	if (cnp->cn_nameptr[0] == '.') {
1726 		if (cnp->cn_namelen == 1)
1727 			return;
1728 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1729 			len = cnp->cn_namelen;
1730 			hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1731 			cache_enter_lock_dd(&cel, dvp, vp, hash);
1732 			/*
1733 			 * If dotdot entry already exists, just retarget it
1734 			 * to new parent vnode, otherwise continue with new
1735 			 * namecache entry allocation.
1736 			 */
1737 			if ((ncp = dvp->v_cache_dd) != NULL &&
1738 			    ncp->nc_flag & NCF_ISDOTDOT) {
1739 				KASSERT(ncp->nc_dvp == dvp,
1740 				    ("wrong isdotdot parent"));
1741 				neg_locked = false;
1742 				if (ncp->nc_flag & NCF_NEGATIVE || vp == NULL) {
1743 					neglist = NCP2NEGLIST(ncp);
1744 					mtx_lock(&ncneg_hot.nl_lock);
1745 					mtx_lock(&neglist->nl_lock);
1746 					neg_locked = true;
1747 				}
1748 				if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1749 					TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst,
1750 					    ncp, nc_dst);
1751 				} else {
1752 					cache_negative_remove(ncp, true);
1753 				}
1754 				if (vp != NULL) {
1755 					TAILQ_INSERT_HEAD(&vp->v_cache_dst,
1756 					    ncp, nc_dst);
1757 					if (ncp->nc_flag & NCF_HOTNEGATIVE)
1758 						numhotneg--;
1759 					ncp->nc_flag &= ~(NCF_NEGATIVE|NCF_HOTNEGATIVE);
1760 				} else {
1761 					if (ncp->nc_flag & NCF_HOTNEGATIVE) {
1762 						numhotneg--;
1763 						ncp->nc_flag &= ~(NCF_HOTNEGATIVE);
1764 					}
1765 					ncp->nc_flag |= NCF_NEGATIVE;
1766 					cache_negative_insert(ncp, true);
1767 				}
1768 				if (neg_locked) {
1769 					mtx_unlock(&neglist->nl_lock);
1770 					mtx_unlock(&ncneg_hot.nl_lock);
1771 				}
1772 				ncp->nc_vp = vp;
1773 				cache_enter_unlock(&cel);
1774 				return;
1775 			}
1776 			dvp->v_cache_dd = NULL;
1777 			cache_enter_unlock(&cel);
1778 			cache_celockstate_init(&cel);
1779 			SDT_PROBE3(vfs, namecache, enter, done, dvp, "..", vp);
1780 			flag = NCF_ISDOTDOT;
1781 		}
1782 	}
1783 
1784 	held_dvp = false;
1785 	if (LIST_EMPTY(&dvp->v_cache_src) && flag != NCF_ISDOTDOT) {
1786 		vhold(dvp);
1787 		atomic_add_long(&numcachehv, 1);
1788 		held_dvp = true;
1789 	}
1790 
1791 	/*
1792 	 * Calculate the hash key and setup as much of the new
1793 	 * namecache entry as possible before acquiring the lock.
1794 	 */
1795 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
1796 	ncp->nc_flag = flag;
1797 	ncp->nc_vp = vp;
1798 	if (vp == NULL)
1799 		ncp->nc_flag |= NCF_NEGATIVE;
1800 	ncp->nc_dvp = dvp;
1801 	if (tsp != NULL) {
1802 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1803 		ncp_ts->nc_time = *tsp;
1804 		ncp_ts->nc_ticks = ticks;
1805 		ncp_ts->nc_nc.nc_flag |= NCF_TS;
1806 		if (dtsp != NULL) {
1807 			ncp_ts->nc_dotdottime = *dtsp;
1808 			ncp_ts->nc_nc.nc_flag |= NCF_DTS;
1809 		}
1810 	}
1811 	len = ncp->nc_nlen = cnp->cn_namelen;
1812 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1813 	strlcpy(ncp->nc_name, cnp->cn_nameptr, len + 1);
1814 	cache_enter_lock(&cel, dvp, vp, hash);
1815 
1816 	/*
1817 	 * See if this vnode or negative entry is already in the cache
1818 	 * with this name.  This can happen with concurrent lookups of
1819 	 * the same path name.
1820 	 */
1821 	ncpp = NCHHASH(hash);
1822 	LIST_FOREACH(n2, ncpp, nc_hash) {
1823 		if (n2->nc_dvp == dvp &&
1824 		    n2->nc_nlen == cnp->cn_namelen &&
1825 		    !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
1826 			if (tsp != NULL) {
1827 				KASSERT((n2->nc_flag & NCF_TS) != 0,
1828 				    ("no NCF_TS"));
1829 				n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
1830 				n2_ts->nc_time = ncp_ts->nc_time;
1831 				n2_ts->nc_ticks = ncp_ts->nc_ticks;
1832 				if (dtsp != NULL) {
1833 					n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
1834 					if (ncp->nc_flag & NCF_NEGATIVE)
1835 						mtx_lock(&ncneg_hot.nl_lock);
1836 					n2_ts->nc_nc.nc_flag |= NCF_DTS;
1837 					if (ncp->nc_flag & NCF_NEGATIVE)
1838 						mtx_unlock(&ncneg_hot.nl_lock);
1839 				}
1840 			}
1841 			goto out_unlock_free;
1842 		}
1843 	}
1844 
1845 	if (flag == NCF_ISDOTDOT) {
1846 		/*
1847 		 * See if we are trying to add .. entry, but some other lookup
1848 		 * has populated v_cache_dd pointer already.
1849 		 */
1850 		if (dvp->v_cache_dd != NULL)
1851 			goto out_unlock_free;
1852 		KASSERT(vp == NULL || vp->v_type == VDIR,
1853 		    ("wrong vnode type %p", vp));
1854 		dvp->v_cache_dd = ncp;
1855 	}
1856 
1857 	if (vp != NULL) {
1858 		if (vp->v_type == VDIR) {
1859 			if (flag != NCF_ISDOTDOT) {
1860 				/*
1861 				 * For this case, the cache entry maps both the
1862 				 * directory name in it and the name ".." for the
1863 				 * directory's parent.
1864 				 */
1865 				if ((ndd = vp->v_cache_dd) != NULL) {
1866 					if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
1867 						cache_zap_locked(ndd, false);
1868 					else
1869 						ndd = NULL;
1870 				}
1871 				vp->v_cache_dd = ncp;
1872 			}
1873 		} else {
1874 			vp->v_cache_dd = NULL;
1875 		}
1876 	}
1877 
1878 	if (flag != NCF_ISDOTDOT) {
1879 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1880 			if (!held_dvp) {
1881 				vhold(dvp);
1882 				atomic_add_long(&numcachehv, 1);
1883 			}
1884 		} else {
1885 			if (held_dvp) {
1886 				/*
1887 				 * This will not take the interlock as someone
1888 				 * else already holds the vnode on account of
1889 				 * the namecache and we hold locks preventing
1890 				 * this from changing.
1891 				 */
1892 				vdrop(dvp);
1893 				atomic_subtract_long(&numcachehv, 1);
1894 			}
1895 		}
1896 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
1897 	}
1898 
1899 	/*
1900 	 * Insert the new namecache entry into the appropriate chain
1901 	 * within the cache entries table.
1902 	 */
1903 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
1904 
1905 	/*
1906 	 * If the entry is "negative", we place it into the
1907 	 * "negative" cache queue, otherwise, we place it into the
1908 	 * destination vnode's cache entries queue.
1909 	 */
1910 	if (vp != NULL) {
1911 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
1912 		SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
1913 		    vp);
1914 	} else {
1915 		if (cnp->cn_flags & ISWHITEOUT)
1916 			ncp->nc_flag |= NCF_WHITE;
1917 		cache_negative_insert(ncp, false);
1918 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
1919 		    ncp->nc_name);
1920 	}
1921 	cache_enter_unlock(&cel);
1922 	if (numneg * ncnegfactor > lnumcache)
1923 		cache_negative_zap_one();
1924 	cache_free(ndd);
1925 	return;
1926 out_unlock_free:
1927 	cache_enter_unlock(&cel);
1928 	cache_free(ncp);
1929 	if (held_dvp) {
1930 		vdrop(dvp);
1931 		atomic_subtract_long(&numcachehv, 1);
1932 	}
1933 	return;
1934 }
1935 
1936 static u_int
1937 cache_roundup_2(u_int val)
1938 {
1939 	u_int res;
1940 
1941 	for (res = 1; res <= val; res <<= 1)
1942 		continue;
1943 
1944 	return (res);
1945 }
1946 
1947 /*
1948  * Name cache initialization, from vfs_init() when we are booting
1949  */
1950 static void
1951 nchinit(void *dummy __unused)
1952 {
1953 	u_int i;
1954 
1955 	cache_zone_small = uma_zcreate("S VFS Cache",
1956 	    sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1,
1957 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1958 	    UMA_ZONE_ZINIT);
1959 	cache_zone_small_ts = uma_zcreate("STS VFS Cache",
1960 	    sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1,
1961 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1962 	    UMA_ZONE_ZINIT);
1963 	cache_zone_large = uma_zcreate("L VFS Cache",
1964 	    sizeof(struct namecache) + NAME_MAX + 1,
1965 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1966 	    UMA_ZONE_ZINIT);
1967 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache",
1968 	    sizeof(struct namecache_ts) + NAME_MAX + 1,
1969 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1970 	    UMA_ZONE_ZINIT);
1971 
1972 	nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash);
1973 	ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
1974 	if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
1975 		ncbuckethash = 7;
1976 	if (ncbuckethash > nchash)
1977 		ncbuckethash = nchash;
1978 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
1979 	    M_WAITOK | M_ZERO);
1980 	for (i = 0; i < numbucketlocks; i++)
1981 		rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK | RW_RECURSE);
1982 	ncvnodehash = ncbuckethash;
1983 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
1984 	    M_WAITOK | M_ZERO);
1985 	for (i = 0; i < numvnodelocks; i++)
1986 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
1987 	ncpurgeminvnodes = numbucketlocks * 2;
1988 
1989 	ncneghash = 3;
1990 	neglists = malloc(sizeof(*neglists) * numneglists, M_VFSCACHE,
1991 	    M_WAITOK | M_ZERO);
1992 	for (i = 0; i < numneglists; i++) {
1993 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
1994 		TAILQ_INIT(&neglists[i].nl_list);
1995 	}
1996 	mtx_init(&ncneg_hot.nl_lock, "ncneglh", NULL, MTX_DEF);
1997 	TAILQ_INIT(&ncneg_hot.nl_list);
1998 
1999 	mtx_init(&ncneg_shrink_lock, "ncnegs", NULL, MTX_DEF);
2000 
2001 	numcalls = counter_u64_alloc(M_WAITOK);
2002 	dothits = counter_u64_alloc(M_WAITOK);
2003 	dotdothits = counter_u64_alloc(M_WAITOK);
2004 	numchecks = counter_u64_alloc(M_WAITOK);
2005 	nummiss = counter_u64_alloc(M_WAITOK);
2006 	nummisszap = counter_u64_alloc(M_WAITOK);
2007 	numposzaps = counter_u64_alloc(M_WAITOK);
2008 	numposhits = counter_u64_alloc(M_WAITOK);
2009 	numnegzaps = counter_u64_alloc(M_WAITOK);
2010 	numneghits = counter_u64_alloc(M_WAITOK);
2011 	numfullpathcalls = counter_u64_alloc(M_WAITOK);
2012 	numfullpathfail1 = counter_u64_alloc(M_WAITOK);
2013 	numfullpathfail2 = counter_u64_alloc(M_WAITOK);
2014 	numfullpathfail4 = counter_u64_alloc(M_WAITOK);
2015 	numfullpathfound = counter_u64_alloc(M_WAITOK);
2016 	zap_and_exit_bucket_relock_success = counter_u64_alloc(M_WAITOK);
2017 }
2018 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
2019 
2020 void
2021 cache_changesize(int newmaxvnodes)
2022 {
2023 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
2024 	u_long new_nchash, old_nchash;
2025 	struct namecache *ncp;
2026 	uint32_t hash;
2027 	int i;
2028 
2029 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
2030 	if (newmaxvnodes < numbucketlocks)
2031 		newmaxvnodes = numbucketlocks;
2032 
2033 	new_nchashtbl = hashinit(newmaxvnodes, M_VFSCACHE, &new_nchash);
2034 	/* If same hash table size, nothing to do */
2035 	if (nchash == new_nchash) {
2036 		free(new_nchashtbl, M_VFSCACHE);
2037 		return;
2038 	}
2039 	/*
2040 	 * Move everything from the old hash table to the new table.
2041 	 * None of the namecache entries in the table can be removed
2042 	 * because to do so, they have to be removed from the hash table.
2043 	 */
2044 	cache_lock_all_vnodes();
2045 	cache_lock_all_buckets();
2046 	old_nchashtbl = nchashtbl;
2047 	old_nchash = nchash;
2048 	nchashtbl = new_nchashtbl;
2049 	nchash = new_nchash;
2050 	for (i = 0; i <= old_nchash; i++) {
2051 		while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) {
2052 			hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
2053 			    ncp->nc_dvp);
2054 			LIST_REMOVE(ncp, nc_hash);
2055 			LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
2056 		}
2057 	}
2058 	cache_unlock_all_buckets();
2059 	cache_unlock_all_vnodes();
2060 	free(old_nchashtbl, M_VFSCACHE);
2061 }
2062 
2063 /*
2064  * Invalidate all entries from and to a particular vnode.
2065  */
2066 void
2067 cache_purge(struct vnode *vp)
2068 {
2069 	TAILQ_HEAD(, namecache) ncps;
2070 	struct namecache *ncp, *nnp;
2071 	struct mtx *vlp, *vlp2;
2072 
2073 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
2074 	SDT_PROBE1(vfs, namecache, purge, done, vp);
2075 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2076 	    vp->v_cache_dd == NULL)
2077 		return;
2078 	TAILQ_INIT(&ncps);
2079 	vlp = VP2VNODELOCK(vp);
2080 	vlp2 = NULL;
2081 	mtx_lock(vlp);
2082 retry:
2083 	while (!LIST_EMPTY(&vp->v_cache_src)) {
2084 		ncp = LIST_FIRST(&vp->v_cache_src);
2085 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2086 			goto retry;
2087 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2088 	}
2089 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2090 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
2091 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2092 			goto retry;
2093 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2094 	}
2095 	ncp = vp->v_cache_dd;
2096 	if (ncp != NULL) {
2097 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2098 		   ("lost dotdot link"));
2099 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2100 			goto retry;
2101 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2102 	}
2103 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2104 	mtx_unlock(vlp);
2105 	if (vlp2 != NULL)
2106 		mtx_unlock(vlp2);
2107 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2108 		cache_free(ncp);
2109 	}
2110 }
2111 
2112 /*
2113  * Invalidate all negative entries for a particular directory vnode.
2114  */
2115 void
2116 cache_purge_negative(struct vnode *vp)
2117 {
2118 	TAILQ_HEAD(, namecache) ncps;
2119 	struct namecache *ncp, *nnp;
2120 	struct mtx *vlp;
2121 
2122 	CTR1(KTR_VFS, "cache_purge_negative(%p)", vp);
2123 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2124 	if (LIST_EMPTY(&vp->v_cache_src))
2125 		return;
2126 	TAILQ_INIT(&ncps);
2127 	vlp = VP2VNODELOCK(vp);
2128 	mtx_lock(vlp);
2129 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2130 		if (!(ncp->nc_flag & NCF_NEGATIVE))
2131 			continue;
2132 		cache_zap_negative_locked_vnode_kl(ncp, vp);
2133 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2134 	}
2135 	mtx_unlock(vlp);
2136 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2137 		cache_free(ncp);
2138 	}
2139 }
2140 
2141 /*
2142  * Flush all entries referencing a particular filesystem.
2143  */
2144 void
2145 cache_purgevfs(struct mount *mp, bool force)
2146 {
2147 	TAILQ_HEAD(, namecache) ncps;
2148 	struct mtx *vlp1, *vlp2;
2149 	struct rwlock *blp;
2150 	struct nchashhead *bucket;
2151 	struct namecache *ncp, *nnp;
2152 	u_long i, j, n_nchash;
2153 	int error;
2154 
2155 	/* Scan hash tables for applicable entries */
2156 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2157 	if (!force && mp->mnt_nvnodelistsize <= ncpurgeminvnodes)
2158 		return;
2159 	TAILQ_INIT(&ncps);
2160 	n_nchash = nchash + 1;
2161 	vlp1 = vlp2 = NULL;
2162 	for (i = 0; i < numbucketlocks; i++) {
2163 		blp = (struct rwlock *)&bucketlocks[i];
2164 		rw_wlock(blp);
2165 		for (j = i; j < n_nchash; j += numbucketlocks) {
2166 retry:
2167 			bucket = &nchashtbl[j];
2168 			LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) {
2169 				cache_assert_bucket_locked(ncp, RA_WLOCKED);
2170 				if (ncp->nc_dvp->v_mount != mp)
2171 					continue;
2172 				error = cache_zap_wlocked_bucket_kl(ncp, blp,
2173 				    &vlp1, &vlp2);
2174 				if (error != 0)
2175 					goto retry;
2176 				TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst);
2177 			}
2178 		}
2179 		rw_wunlock(blp);
2180 		if (vlp1 == NULL && vlp2 == NULL)
2181 			cache_maybe_yield();
2182 	}
2183 	if (vlp1 != NULL)
2184 		mtx_unlock(vlp1);
2185 	if (vlp2 != NULL)
2186 		mtx_unlock(vlp2);
2187 
2188 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2189 		cache_free(ncp);
2190 	}
2191 }
2192 
2193 /*
2194  * Perform canonical checks and cache lookup and pass on to filesystem
2195  * through the vop_cachedlookup only if needed.
2196  */
2197 
2198 int
2199 vfs_cache_lookup(struct vop_lookup_args *ap)
2200 {
2201 	struct vnode *dvp;
2202 	int error;
2203 	struct vnode **vpp = ap->a_vpp;
2204 	struct componentname *cnp = ap->a_cnp;
2205 	struct ucred *cred = cnp->cn_cred;
2206 	int flags = cnp->cn_flags;
2207 	struct thread *td = cnp->cn_thread;
2208 
2209 	*vpp = NULL;
2210 	dvp = ap->a_dvp;
2211 
2212 	if (dvp->v_type != VDIR)
2213 		return (ENOTDIR);
2214 
2215 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2216 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2217 		return (EROFS);
2218 
2219 	error = VOP_ACCESS(dvp, VEXEC, cred, td);
2220 	if (error)
2221 		return (error);
2222 
2223 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2224 	if (error == 0)
2225 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2226 	if (error == -1)
2227 		return (0);
2228 	return (error);
2229 }
2230 
2231 /*
2232  * XXX All of these sysctls would probably be more productive dead.
2233  */
2234 static int __read_mostly disablecwd;
2235 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
2236    "Disable the getcwd syscall");
2237 
2238 /* Implementation of the getcwd syscall. */
2239 int
2240 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2241 {
2242 
2243 	return (kern___getcwd(td, uap->buf, UIO_USERSPACE, uap->buflen,
2244 	    MAXPATHLEN));
2245 }
2246 
2247 int
2248 kern___getcwd(struct thread *td, char *buf, enum uio_seg bufseg, size_t buflen,
2249     size_t path_max)
2250 {
2251 	char *bp, *tmpbuf;
2252 	struct filedesc *fdp;
2253 	struct vnode *cdir, *rdir;
2254 	int error;
2255 
2256 	if (__predict_false(disablecwd))
2257 		return (ENODEV);
2258 	if (__predict_false(buflen < 2))
2259 		return (EINVAL);
2260 	if (buflen > path_max)
2261 		buflen = path_max;
2262 
2263 	tmpbuf = malloc(buflen, M_TEMP, M_WAITOK);
2264 	fdp = td->td_proc->p_fd;
2265 	FILEDESC_SLOCK(fdp);
2266 	cdir = fdp->fd_cdir;
2267 	vrefact(cdir);
2268 	rdir = fdp->fd_rdir;
2269 	vrefact(rdir);
2270 	FILEDESC_SUNLOCK(fdp);
2271 	error = vn_fullpath1(td, cdir, rdir, tmpbuf, &bp, buflen);
2272 	vrele(rdir);
2273 	vrele(cdir);
2274 
2275 	if (!error) {
2276 		if (bufseg == UIO_SYSSPACE)
2277 			bcopy(bp, buf, strlen(bp) + 1);
2278 		else
2279 			error = copyout(bp, buf, strlen(bp) + 1);
2280 #ifdef KTRACE
2281 	if (KTRPOINT(curthread, KTR_NAMEI))
2282 		ktrnamei(bp);
2283 #endif
2284 	}
2285 	free(tmpbuf, M_TEMP);
2286 	return (error);
2287 }
2288 
2289 /*
2290  * Thus begins the fullpath magic.
2291  */
2292 
2293 static int __read_mostly disablefullpath;
2294 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, &disablefullpath, 0,
2295     "Disable the vn_fullpath function");
2296 
2297 /*
2298  * Retrieve the full filesystem path that correspond to a vnode from the name
2299  * cache (if available)
2300  */
2301 int
2302 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf)
2303 {
2304 	char *buf;
2305 	struct filedesc *fdp;
2306 	struct vnode *rdir;
2307 	int error;
2308 
2309 	if (__predict_false(disablefullpath))
2310 		return (ENODEV);
2311 	if (__predict_false(vn == NULL))
2312 		return (EINVAL);
2313 
2314 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2315 	fdp = td->td_proc->p_fd;
2316 	FILEDESC_SLOCK(fdp);
2317 	rdir = fdp->fd_rdir;
2318 	vrefact(rdir);
2319 	FILEDESC_SUNLOCK(fdp);
2320 	error = vn_fullpath1(td, vn, rdir, buf, retbuf, MAXPATHLEN);
2321 	vrele(rdir);
2322 
2323 	if (!error)
2324 		*freebuf = buf;
2325 	else
2326 		free(buf, M_TEMP);
2327 	return (error);
2328 }
2329 
2330 /*
2331  * This function is similar to vn_fullpath, but it attempts to lookup the
2332  * pathname relative to the global root mount point.  This is required for the
2333  * auditing sub-system, as audited pathnames must be absolute, relative to the
2334  * global root mount point.
2335  */
2336 int
2337 vn_fullpath_global(struct thread *td, struct vnode *vn,
2338     char **retbuf, char **freebuf)
2339 {
2340 	char *buf;
2341 	int error;
2342 
2343 	if (__predict_false(disablefullpath))
2344 		return (ENODEV);
2345 	if (__predict_false(vn == NULL))
2346 		return (EINVAL);
2347 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2348 	error = vn_fullpath1(td, vn, rootvnode, buf, retbuf, MAXPATHLEN);
2349 	if (!error)
2350 		*freebuf = buf;
2351 	else
2352 		free(buf, M_TEMP);
2353 	return (error);
2354 }
2355 
2356 int
2357 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen)
2358 {
2359 	struct vnode *dvp;
2360 	struct namecache *ncp;
2361 	struct mtx *vlp;
2362 	int error;
2363 
2364 	vlp = VP2VNODELOCK(*vp);
2365 	mtx_lock(vlp);
2366 	TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) {
2367 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2368 			break;
2369 	}
2370 	if (ncp != NULL) {
2371 		if (*buflen < ncp->nc_nlen) {
2372 			mtx_unlock(vlp);
2373 			vrele(*vp);
2374 			counter_u64_add(numfullpathfail4, 1);
2375 			error = ENOMEM;
2376 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2377 			    vp, NULL);
2378 			return (error);
2379 		}
2380 		*buflen -= ncp->nc_nlen;
2381 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2382 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2383 		    ncp->nc_name, vp);
2384 		dvp = *vp;
2385 		*vp = ncp->nc_dvp;
2386 		vref(*vp);
2387 		mtx_unlock(vlp);
2388 		vrele(dvp);
2389 		return (0);
2390 	}
2391 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2392 
2393 	mtx_unlock(vlp);
2394 	vn_lock(*vp, LK_SHARED | LK_RETRY);
2395 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
2396 	vput(*vp);
2397 	if (error) {
2398 		counter_u64_add(numfullpathfail2, 1);
2399 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
2400 		return (error);
2401 	}
2402 
2403 	*vp = dvp;
2404 	if (dvp->v_iflag & VI_DOOMED) {
2405 		/* forced unmount */
2406 		vrele(dvp);
2407 		error = ENOENT;
2408 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2409 		return (error);
2410 	}
2411 	/*
2412 	 * *vp has its use count incremented still.
2413 	 */
2414 
2415 	return (0);
2416 }
2417 
2418 /*
2419  * The magic behind kern___getcwd() and vn_fullpath().
2420  */
2421 static int
2422 vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
2423     char *buf, char **retbuf, u_int buflen)
2424 {
2425 	int error, slash_prefixed;
2426 #ifdef KDTRACE_HOOKS
2427 	struct vnode *startvp = vp;
2428 #endif
2429 	struct vnode *vp1;
2430 
2431 	buflen--;
2432 	buf[buflen] = '\0';
2433 	error = 0;
2434 	slash_prefixed = 0;
2435 
2436 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
2437 	counter_u64_add(numfullpathcalls, 1);
2438 	vref(vp);
2439 	if (vp->v_type != VDIR) {
2440 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2441 		if (error)
2442 			return (error);
2443 		if (buflen == 0) {
2444 			vrele(vp);
2445 			return (ENOMEM);
2446 		}
2447 		buf[--buflen] = '/';
2448 		slash_prefixed = 1;
2449 	}
2450 	while (vp != rdir && vp != rootvnode) {
2451 		/*
2452 		 * The vp vnode must be already fully constructed,
2453 		 * since it is either found in namecache or obtained
2454 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
2455 		 * without obtaining the vnode lock.
2456 		 */
2457 		if ((vp->v_vflag & VV_ROOT) != 0) {
2458 			vn_lock(vp, LK_RETRY | LK_SHARED);
2459 
2460 			/*
2461 			 * With the vnode locked, check for races with
2462 			 * unmount, forced or not.  Note that we
2463 			 * already verified that vp is not equal to
2464 			 * the root vnode, which means that
2465 			 * mnt_vnodecovered can be NULL only for the
2466 			 * case of unmount.
2467 			 */
2468 			if ((vp->v_iflag & VI_DOOMED) != 0 ||
2469 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
2470 			    vp1->v_mountedhere != vp->v_mount) {
2471 				vput(vp);
2472 				error = ENOENT;
2473 				SDT_PROBE3(vfs, namecache, fullpath, return,
2474 				    error, vp, NULL);
2475 				break;
2476 			}
2477 
2478 			vref(vp1);
2479 			vput(vp);
2480 			vp = vp1;
2481 			continue;
2482 		}
2483 		if (vp->v_type != VDIR) {
2484 			vrele(vp);
2485 			counter_u64_add(numfullpathfail1, 1);
2486 			error = ENOTDIR;
2487 			SDT_PROBE3(vfs, namecache, fullpath, return,
2488 			    error, vp, NULL);
2489 			break;
2490 		}
2491 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2492 		if (error)
2493 			break;
2494 		if (buflen == 0) {
2495 			vrele(vp);
2496 			error = ENOMEM;
2497 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2498 			    startvp, NULL);
2499 			break;
2500 		}
2501 		buf[--buflen] = '/';
2502 		slash_prefixed = 1;
2503 	}
2504 	if (error)
2505 		return (error);
2506 	if (!slash_prefixed) {
2507 		if (buflen == 0) {
2508 			vrele(vp);
2509 			counter_u64_add(numfullpathfail4, 1);
2510 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
2511 			    startvp, NULL);
2512 			return (ENOMEM);
2513 		}
2514 		buf[--buflen] = '/';
2515 	}
2516 	counter_u64_add(numfullpathfound, 1);
2517 	vrele(vp);
2518 
2519 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, buf + buflen);
2520 	*retbuf = buf + buflen;
2521 	return (0);
2522 }
2523 
2524 struct vnode *
2525 vn_dir_dd_ino(struct vnode *vp)
2526 {
2527 	struct namecache *ncp;
2528 	struct vnode *ddvp;
2529 	struct mtx *vlp;
2530 	enum vgetstate vs;
2531 
2532 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
2533 	vlp = VP2VNODELOCK(vp);
2534 	mtx_lock(vlp);
2535 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
2536 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
2537 			continue;
2538 		ddvp = ncp->nc_dvp;
2539 		vs = vget_prep(ddvp);
2540 		mtx_unlock(vlp);
2541 		if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
2542 			return (NULL);
2543 		return (ddvp);
2544 	}
2545 	mtx_unlock(vlp);
2546 	return (NULL);
2547 }
2548 
2549 int
2550 vn_commname(struct vnode *vp, char *buf, u_int buflen)
2551 {
2552 	struct namecache *ncp;
2553 	struct mtx *vlp;
2554 	int l;
2555 
2556 	vlp = VP2VNODELOCK(vp);
2557 	mtx_lock(vlp);
2558 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
2559 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2560 			break;
2561 	if (ncp == NULL) {
2562 		mtx_unlock(vlp);
2563 		return (ENOENT);
2564 	}
2565 	l = min(ncp->nc_nlen, buflen - 1);
2566 	memcpy(buf, ncp->nc_name, l);
2567 	mtx_unlock(vlp);
2568 	buf[l] = '\0';
2569 	return (0);
2570 }
2571 
2572 /*
2573  * This function updates path string to vnode's full global path
2574  * and checks the size of the new path string against the pathlen argument.
2575  *
2576  * Requires a locked, referenced vnode.
2577  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
2578  *
2579  * If sysctl debug.disablefullpath is set, ENODEV is returned,
2580  * vnode is left locked and path remain untouched.
2581  *
2582  * If vp is a directory, the call to vn_fullpath_global() always succeeds
2583  * because it falls back to the ".." lookup if the namecache lookup fails.
2584  */
2585 int
2586 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
2587     u_int pathlen)
2588 {
2589 	struct nameidata nd;
2590 	struct vnode *vp1;
2591 	char *rpath, *fbuf;
2592 	int error;
2593 
2594 	ASSERT_VOP_ELOCKED(vp, __func__);
2595 
2596 	/* Return ENODEV if sysctl debug.disablefullpath==1 */
2597 	if (__predict_false(disablefullpath))
2598 		return (ENODEV);
2599 
2600 	/* Construct global filesystem path from vp. */
2601 	VOP_UNLOCK(vp, 0);
2602 	error = vn_fullpath_global(td, vp, &rpath, &fbuf);
2603 
2604 	if (error != 0) {
2605 		vrele(vp);
2606 		return (error);
2607 	}
2608 
2609 	if (strlen(rpath) >= pathlen) {
2610 		vrele(vp);
2611 		error = ENAMETOOLONG;
2612 		goto out;
2613 	}
2614 
2615 	/*
2616 	 * Re-lookup the vnode by path to detect a possible rename.
2617 	 * As a side effect, the vnode is relocked.
2618 	 * If vnode was renamed, return ENOENT.
2619 	 */
2620 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
2621 	    UIO_SYSSPACE, path, td);
2622 	error = namei(&nd);
2623 	if (error != 0) {
2624 		vrele(vp);
2625 		goto out;
2626 	}
2627 	NDFREE(&nd, NDF_ONLY_PNBUF);
2628 	vp1 = nd.ni_vp;
2629 	vrele(vp);
2630 	if (vp1 == vp)
2631 		strcpy(path, rpath);
2632 	else {
2633 		vput(vp1);
2634 		error = ENOENT;
2635 	}
2636 
2637 out:
2638 	free(fbuf, M_TEMP);
2639 	return (error);
2640 }
2641 
2642 #ifdef DDB
2643 static void
2644 db_print_vpath(struct vnode *vp)
2645 {
2646 
2647 	while (vp != NULL) {
2648 		db_printf("%p: ", vp);
2649 		if (vp == rootvnode) {
2650 			db_printf("/");
2651 			vp = NULL;
2652 		} else {
2653 			if (vp->v_vflag & VV_ROOT) {
2654 				db_printf("<mount point>");
2655 				vp = vp->v_mount->mnt_vnodecovered;
2656 			} else {
2657 				struct namecache *ncp;
2658 				char *ncn;
2659 				int i;
2660 
2661 				ncp = TAILQ_FIRST(&vp->v_cache_dst);
2662 				if (ncp != NULL) {
2663 					ncn = ncp->nc_name;
2664 					for (i = 0; i < ncp->nc_nlen; i++)
2665 						db_printf("%c", *ncn++);
2666 					vp = ncp->nc_dvp;
2667 				} else {
2668 					vp = NULL;
2669 				}
2670 			}
2671 		}
2672 		db_printf("\n");
2673 	}
2674 
2675 	return;
2676 }
2677 
2678 DB_SHOW_COMMAND(vpath, db_show_vpath)
2679 {
2680 	struct vnode *vp;
2681 
2682 	if (!have_addr) {
2683 		db_printf("usage: show vpath <struct vnode *>\n");
2684 		return;
2685 	}
2686 
2687 	vp = (struct vnode *)addr;
2688 	db_print_vpath(vp);
2689 }
2690 
2691 #endif
2692