xref: /freebsd/sys/kern/vfs_cache.c (revision 473957941922d17be72089e385e2e2a995fd0e1c)
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/capsicum.h>
46 #include <sys/counter.h>
47 #include <sys/filedesc.h>
48 #include <sys/fnv_hash.h>
49 #include <sys/kernel.h>
50 #include <sys/ktr.h>
51 #include <sys/lock.h>
52 #include <sys/malloc.h>
53 #include <sys/fcntl.h>
54 #include <sys/mount.h>
55 #include <sys/namei.h>
56 #include <sys/proc.h>
57 #include <sys/rwlock.h>
58 #include <sys/sdt.h>
59 #include <sys/smp.h>
60 #include <sys/syscallsubr.h>
61 #include <sys/sysctl.h>
62 #include <sys/sysproto.h>
63 #include <sys/vnode.h>
64 #ifdef KTRACE
65 #include <sys/ktrace.h>
66 #endif
67 
68 #ifdef DDB
69 #include <ddb/ddb.h>
70 #endif
71 
72 #include <vm/uma.h>
73 
74 SDT_PROVIDER_DECLARE(vfs);
75 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
76     "struct vnode *");
77 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
78     "char *");
79 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
80 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
81     "char *", "struct vnode *");
82 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
83 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
84     "struct vnode *", "char *");
85 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
86     "struct vnode *");
87 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
88     "struct vnode *", "char *");
89 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
90     "char *");
91 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *");
92 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
93 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
94 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
95     "struct vnode *");
96 SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *",
97     "char *");
98 SDT_PROBE_DEFINE2(vfs, namecache, shrink_negative, done, "struct vnode *",
99     "char *");
100 
101 /*
102  * This structure describes the elements in the cache of recent
103  * names looked up by namei.
104  */
105 
106 struct	namecache {
107 	LIST_ENTRY(namecache) nc_hash;	/* hash chain */
108 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
109 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
110 	struct	vnode *nc_dvp;		/* vnode of parent of name */
111 	union {
112 		struct	vnode *nu_vp;	/* vnode the name refers to */
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 
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 = 5; /* 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 static u_long __exclusive_cache_line	numcache;/* number of cache entries allocated */
209 u_int ncsizefactor = 2;
210 SYSCTL_UINT(_vfs, OID_AUTO, ncsizefactor, CTLFLAG_RW, &ncsizefactor, 0,
211     "Size factor for namecache");
212 static u_int __read_mostly	ncpurgeminvnodes;
213 SYSCTL_UINT(_vfs, OID_AUTO, ncpurgeminvnodes, CTLFLAG_RW, &ncpurgeminvnodes, 0,
214     "Number of vnodes below which purgevfs ignores the request");
215 static u_int __read_mostly	ncsize; /* the size as computed on creation or resizing */
216 
217 struct nchstats	nchstats;		/* cache effectiveness statistics */
218 
219 static struct mtx __exclusive_cache_line	ncneg_shrink_lock;
220 static int	shrink_list_turn;
221 
222 struct neglist {
223 	struct mtx		nl_lock;
224 	TAILQ_HEAD(, namecache) nl_list;
225 } __aligned(CACHE_LINE_SIZE);
226 
227 static struct neglist __read_mostly	*neglists;
228 static struct neglist ncneg_hot;
229 static u_long numhotneg;
230 
231 #define	numneglists (ncneghash + 1)
232 static u_int __read_mostly	ncneghash;
233 static inline struct neglist *
234 NCP2NEGLIST(struct namecache *ncp)
235 {
236 
237 	return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
238 }
239 
240 #define	numbucketlocks (ncbuckethash + 1)
241 static u_int __read_mostly  ncbuckethash;
242 static struct rwlock_padalign __read_mostly  *bucketlocks;
243 #define	HASH2BUCKETLOCK(hash) \
244 	((struct rwlock *)(&bucketlocks[((hash) & ncbuckethash)]))
245 
246 #define	numvnodelocks (ncvnodehash + 1)
247 static u_int __read_mostly  ncvnodehash;
248 static struct mtx __read_mostly *vnodelocks;
249 static inline struct mtx *
250 VP2VNODELOCK(struct vnode *vp)
251 {
252 
253 	return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
254 }
255 
256 /*
257  * UMA zones for the VFS cache.
258  *
259  * The small cache is used for entries with short names, which are the
260  * most common.  The large cache is used for entries which are too big to
261  * fit in the small cache.
262  */
263 static uma_zone_t __read_mostly cache_zone_small;
264 static uma_zone_t __read_mostly cache_zone_small_ts;
265 static uma_zone_t __read_mostly cache_zone_large;
266 static uma_zone_t __read_mostly cache_zone_large_ts;
267 
268 #define	CACHE_PATH_CUTOFF	35
269 
270 static struct namecache *
271 cache_alloc(int len, int ts)
272 {
273 	struct namecache_ts *ncp_ts;
274 	struct namecache *ncp;
275 
276 	if (__predict_false(ts)) {
277 		if (len <= CACHE_PATH_CUTOFF)
278 			ncp_ts = uma_zalloc(cache_zone_small_ts, M_WAITOK);
279 		else
280 			ncp_ts = uma_zalloc(cache_zone_large_ts, M_WAITOK);
281 		ncp = &ncp_ts->nc_nc;
282 	} else {
283 		if (len <= CACHE_PATH_CUTOFF)
284 			ncp = uma_zalloc(cache_zone_small, M_WAITOK);
285 		else
286 			ncp = uma_zalloc(cache_zone_large, M_WAITOK);
287 	}
288 	return (ncp);
289 }
290 
291 static void
292 cache_free(struct namecache *ncp)
293 {
294 	struct namecache_ts *ncp_ts;
295 
296 	if (ncp == NULL)
297 		return;
298 	if ((ncp->nc_flag & NCF_DVDROP) != 0)
299 		vdrop(ncp->nc_dvp);
300 	if (__predict_false(ncp->nc_flag & NCF_TS)) {
301 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
302 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
303 			uma_zfree(cache_zone_small_ts, ncp_ts);
304 		else
305 			uma_zfree(cache_zone_large_ts, ncp_ts);
306 	} else {
307 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
308 			uma_zfree(cache_zone_small, ncp);
309 		else
310 			uma_zfree(cache_zone_large, ncp);
311 	}
312 }
313 
314 static void
315 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
316 {
317 	struct namecache_ts *ncp_ts;
318 
319 	KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
320 	    (tsp == NULL && ticksp == NULL),
321 	    ("No NCF_TS"));
322 
323 	if (tsp == NULL && ticksp == NULL)
324 		return;
325 
326 	ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
327 	if (tsp != NULL)
328 		*tsp = ncp_ts->nc_time;
329 	if (ticksp != NULL)
330 		*ticksp = ncp_ts->nc_ticks;
331 }
332 
333 #ifdef DEBUG_CACHE
334 static int __read_mostly	doingcache = 1;	/* 1 => enable the cache */
335 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
336     "VFS namecache enabled");
337 #endif
338 
339 /* Export size information to userland */
340 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
341     sizeof(struct namecache), "sizeof(struct namecache)");
342 
343 /*
344  * The new name cache statistics
345  */
346 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0,
347     "Name cache statistics");
348 #define STATNODE_ULONG(name, descr)	\
349 	SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, descr);
350 #define STATNODE_COUNTER(name, descr)	\
351 	static counter_u64_t __read_mostly name; \
352 	SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, descr);
353 STATNODE_ULONG(numneg, "Number of negative cache entries");
354 STATNODE_ULONG(numcache, "Number of cache entries");
355 STATNODE_COUNTER(numcachehv, "Number of namecache entries with vnodes held");
356 STATNODE_COUNTER(numcalls, "Number of cache lookups");
357 STATNODE_COUNTER(dothits, "Number of '.' hits");
358 STATNODE_COUNTER(dotdothits, "Number of '..' hits");
359 STATNODE_COUNTER(numchecks, "Number of checks in lookup");
360 STATNODE_COUNTER(nummiss, "Number of cache misses");
361 STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache");
362 STATNODE_COUNTER(numposzaps,
363     "Number of cache hits (positive) we do not want to cache");
364 STATNODE_COUNTER(numposhits, "Number of cache hits (positive)");
365 STATNODE_COUNTER(numnegzaps,
366     "Number of cache hits (negative) we do not want to cache");
367 STATNODE_COUNTER(numneghits, "Number of cache hits (negative)");
368 /* These count for vn_getcwd(), too. */
369 STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls");
370 STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
371 STATNODE_COUNTER(numfullpathfail2,
372     "Number of fullpath search errors (VOP_VPTOCNP failures)");
373 STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
374 STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls");
375 STATNODE_COUNTER(zap_and_exit_bucket_relock_success,
376     "Number of successful removals after relocking");
377 static long zap_and_exit_bucket_fail; STATNODE_ULONG(zap_and_exit_bucket_fail,
378     "Number of times zap_and_exit failed to lock");
379 static long zap_and_exit_bucket_fail2; STATNODE_ULONG(zap_and_exit_bucket_fail2,
380     "Number of times zap_and_exit failed to lock");
381 static long cache_lock_vnodes_cel_3_failures;
382 STATNODE_ULONG(cache_lock_vnodes_cel_3_failures,
383     "Number of times 3-way vnode locking failed");
384 STATNODE_ULONG(numhotneg, "Number of hot negative entries");
385 STATNODE_COUNTER(numneg_evicted,
386     "Number of negative entries evicted when adding a new entry");
387 STATNODE_COUNTER(shrinking_skipped,
388     "Number of times shrinking was already in progress");
389 
390 static void cache_zap_locked(struct namecache *ncp, bool neg_locked);
391 static int vn_fullpath_hardlink(struct thread *td, struct nameidata *ndp, char **retbuf,
392     char **freebuf, size_t *buflen);
393 static int vn_fullpath_any(struct thread *td, struct vnode *vp, struct vnode *rdir,
394     char *buf, char **retbuf, size_t *buflen);
395 static int vn_fullpath_dir(struct thread *td, struct vnode *vp, struct vnode *rdir,
396     char *buf, char **retbuf, size_t *len, bool slash_prefixed, size_t addend);
397 
398 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
399 
400 static int cache_yield;
401 SYSCTL_INT(_vfs_cache, OID_AUTO, yield, CTLFLAG_RD, &cache_yield, 0,
402     "Number of times cache called yield");
403 
404 static void __noinline
405 cache_maybe_yield(void)
406 {
407 
408 	if (should_yield()) {
409 		cache_yield++;
410 		kern_yield(PRI_USER);
411 	}
412 }
413 
414 static inline void
415 cache_assert_vlp_locked(struct mtx *vlp)
416 {
417 
418 	if (vlp != NULL)
419 		mtx_assert(vlp, MA_OWNED);
420 }
421 
422 static inline void
423 cache_assert_vnode_locked(struct vnode *vp)
424 {
425 	struct mtx *vlp;
426 
427 	vlp = VP2VNODELOCK(vp);
428 	cache_assert_vlp_locked(vlp);
429 }
430 
431 static uint32_t
432 cache_get_hash(char *name, u_char len, struct vnode *dvp)
433 {
434 	uint32_t hash;
435 
436 	hash = fnv_32_buf(name, len, FNV1_32_INIT);
437 	hash = fnv_32_buf(&dvp, sizeof(dvp), hash);
438 	return (hash);
439 }
440 
441 static inline struct rwlock *
442 NCP2BUCKETLOCK(struct namecache *ncp)
443 {
444 	uint32_t hash;
445 
446 	hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
447 	return (HASH2BUCKETLOCK(hash));
448 }
449 
450 #ifdef INVARIANTS
451 static void
452 cache_assert_bucket_locked(struct namecache *ncp, int mode)
453 {
454 	struct rwlock *blp;
455 
456 	blp = NCP2BUCKETLOCK(ncp);
457 	rw_assert(blp, mode);
458 }
459 #else
460 #define cache_assert_bucket_locked(x, y) do { } while (0)
461 #endif
462 
463 #define cache_sort_vnodes(x, y)	_cache_sort_vnodes((void **)(x), (void **)(y))
464 static void
465 _cache_sort_vnodes(void **p1, void **p2)
466 {
467 	void *tmp;
468 
469 	MPASS(*p1 != NULL || *p2 != NULL);
470 
471 	if (*p1 > *p2) {
472 		tmp = *p2;
473 		*p2 = *p1;
474 		*p1 = tmp;
475 	}
476 }
477 
478 static void
479 cache_lock_all_buckets(void)
480 {
481 	u_int i;
482 
483 	for (i = 0; i < numbucketlocks; i++)
484 		rw_wlock(&bucketlocks[i]);
485 }
486 
487 static void
488 cache_unlock_all_buckets(void)
489 {
490 	u_int i;
491 
492 	for (i = 0; i < numbucketlocks; i++)
493 		rw_wunlock(&bucketlocks[i]);
494 }
495 
496 static void
497 cache_lock_all_vnodes(void)
498 {
499 	u_int i;
500 
501 	for (i = 0; i < numvnodelocks; i++)
502 		mtx_lock(&vnodelocks[i]);
503 }
504 
505 static void
506 cache_unlock_all_vnodes(void)
507 {
508 	u_int i;
509 
510 	for (i = 0; i < numvnodelocks; i++)
511 		mtx_unlock(&vnodelocks[i]);
512 }
513 
514 static int
515 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
516 {
517 
518 	cache_sort_vnodes(&vlp1, &vlp2);
519 
520 	if (vlp1 != NULL) {
521 		if (!mtx_trylock(vlp1))
522 			return (EAGAIN);
523 	}
524 	if (!mtx_trylock(vlp2)) {
525 		if (vlp1 != NULL)
526 			mtx_unlock(vlp1);
527 		return (EAGAIN);
528 	}
529 
530 	return (0);
531 }
532 
533 static void
534 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
535 {
536 
537 	MPASS(vlp1 != NULL || vlp2 != NULL);
538 	MPASS(vlp1 <= vlp2);
539 
540 	if (vlp1 != NULL)
541 		mtx_lock(vlp1);
542 	if (vlp2 != NULL)
543 		mtx_lock(vlp2);
544 }
545 
546 static void
547 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
548 {
549 
550 	MPASS(vlp1 != NULL || vlp2 != NULL);
551 
552 	if (vlp1 != NULL)
553 		mtx_unlock(vlp1);
554 	if (vlp2 != NULL)
555 		mtx_unlock(vlp2);
556 }
557 
558 static int
559 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
560 {
561 	struct nchstats snap;
562 
563 	if (req->oldptr == NULL)
564 		return (SYSCTL_OUT(req, 0, sizeof(snap)));
565 
566 	snap = nchstats;
567 	snap.ncs_goodhits = counter_u64_fetch(numposhits);
568 	snap.ncs_neghits = counter_u64_fetch(numneghits);
569 	snap.ncs_badhits = counter_u64_fetch(numposzaps) +
570 	    counter_u64_fetch(numnegzaps);
571 	snap.ncs_miss = counter_u64_fetch(nummisszap) +
572 	    counter_u64_fetch(nummiss);
573 
574 	return (SYSCTL_OUT(req, &snap, sizeof(snap)));
575 }
576 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
577     CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
578     "VFS cache effectiveness statistics");
579 
580 #ifdef DIAGNOSTIC
581 /*
582  * Grab an atomic snapshot of the name cache hash chain lengths
583  */
584 static SYSCTL_NODE(_debug, OID_AUTO, hashstat, CTLFLAG_RW, NULL,
585     "hash table stats");
586 
587 static int
588 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
589 {
590 	struct nchashhead *ncpp;
591 	struct namecache *ncp;
592 	int i, error, n_nchash, *cntbuf;
593 
594 retry:
595 	n_nchash = nchash + 1;	/* nchash is max index, not count */
596 	if (req->oldptr == NULL)
597 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
598 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
599 	cache_lock_all_buckets();
600 	if (n_nchash != nchash + 1) {
601 		cache_unlock_all_buckets();
602 		free(cntbuf, M_TEMP);
603 		goto retry;
604 	}
605 	/* Scan hash tables counting entries */
606 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
607 		LIST_FOREACH(ncp, ncpp, nc_hash)
608 			cntbuf[i]++;
609 	cache_unlock_all_buckets();
610 	for (error = 0, i = 0; i < n_nchash; i++)
611 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
612 			break;
613 	free(cntbuf, M_TEMP);
614 	return (error);
615 }
616 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
617     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
618     "nchash chain lengths");
619 
620 static int
621 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
622 {
623 	int error;
624 	struct nchashhead *ncpp;
625 	struct namecache *ncp;
626 	int n_nchash;
627 	int count, maxlength, used, pct;
628 
629 	if (!req->oldptr)
630 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
631 
632 	cache_lock_all_buckets();
633 	n_nchash = nchash + 1;	/* nchash is max index, not count */
634 	used = 0;
635 	maxlength = 0;
636 
637 	/* Scan hash tables for applicable entries */
638 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
639 		count = 0;
640 		LIST_FOREACH(ncp, ncpp, nc_hash) {
641 			count++;
642 		}
643 		if (count)
644 			used++;
645 		if (maxlength < count)
646 			maxlength = count;
647 	}
648 	n_nchash = nchash + 1;
649 	cache_unlock_all_buckets();
650 	pct = (used * 100) / (n_nchash / 100);
651 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
652 	if (error)
653 		return (error);
654 	error = SYSCTL_OUT(req, &used, sizeof(used));
655 	if (error)
656 		return (error);
657 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
658 	if (error)
659 		return (error);
660 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
661 	if (error)
662 		return (error);
663 	return (0);
664 }
665 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
666     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
667     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
668 #endif
669 
670 /*
671  * Negative entries management
672  *
673  * A variation of LRU scheme is used. New entries are hashed into one of
674  * numneglists cold lists. Entries get promoted to the hot list on first hit.
675  *
676  * The shrinker will demote hot list head and evict from the cold list in a
677  * round-robin manner.
678  */
679 static void
680 cache_negative_hit(struct namecache *ncp)
681 {
682 	struct neglist *neglist;
683 
684 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
685 	if (ncp->nc_flag & NCF_HOTNEGATIVE)
686 		return;
687 	neglist = NCP2NEGLIST(ncp);
688 	mtx_lock(&ncneg_hot.nl_lock);
689 	mtx_lock(&neglist->nl_lock);
690 	if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
691 		numhotneg++;
692 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
693 		TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
694 		ncp->nc_flag |= NCF_HOTNEGATIVE;
695 	}
696 	mtx_unlock(&neglist->nl_lock);
697 	mtx_unlock(&ncneg_hot.nl_lock);
698 }
699 
700 static void
701 cache_negative_insert(struct namecache *ncp, bool neg_locked)
702 {
703 	struct neglist *neglist;
704 
705 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
706 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
707 	neglist = NCP2NEGLIST(ncp);
708 	if (!neg_locked) {
709 		mtx_lock(&neglist->nl_lock);
710 	} else {
711 		mtx_assert(&neglist->nl_lock, MA_OWNED);
712 	}
713 	TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
714 	if (!neg_locked)
715 		mtx_unlock(&neglist->nl_lock);
716 	atomic_add_rel_long(&numneg, 1);
717 }
718 
719 static void
720 cache_negative_remove(struct namecache *ncp, bool neg_locked)
721 {
722 	struct neglist *neglist;
723 	bool hot_locked = false;
724 	bool list_locked = false;
725 
726 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
727 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
728 	neglist = NCP2NEGLIST(ncp);
729 	if (!neg_locked) {
730 		if (ncp->nc_flag & NCF_HOTNEGATIVE) {
731 			hot_locked = true;
732 			mtx_lock(&ncneg_hot.nl_lock);
733 			if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
734 				list_locked = true;
735 				mtx_lock(&neglist->nl_lock);
736 			}
737 		} else {
738 			list_locked = true;
739 			mtx_lock(&neglist->nl_lock);
740 		}
741 	}
742 	if (ncp->nc_flag & NCF_HOTNEGATIVE) {
743 		mtx_assert(&ncneg_hot.nl_lock, MA_OWNED);
744 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
745 		numhotneg--;
746 	} else {
747 		mtx_assert(&neglist->nl_lock, MA_OWNED);
748 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
749 	}
750 	if (list_locked)
751 		mtx_unlock(&neglist->nl_lock);
752 	if (hot_locked)
753 		mtx_unlock(&ncneg_hot.nl_lock);
754 	atomic_subtract_rel_long(&numneg, 1);
755 }
756 
757 static void
758 cache_negative_shrink_select(int start, struct namecache **ncpp,
759     struct neglist **neglistpp)
760 {
761 	struct neglist *neglist;
762 	struct namecache *ncp;
763 	int i;
764 
765 	*ncpp = ncp = NULL;
766 	neglist = NULL;
767 
768 	for (i = start; i < numneglists; i++) {
769 		neglist = &neglists[i];
770 		if (TAILQ_FIRST(&neglist->nl_list) == NULL)
771 			continue;
772 		mtx_lock(&neglist->nl_lock);
773 		ncp = TAILQ_FIRST(&neglist->nl_list);
774 		if (ncp != NULL)
775 			break;
776 		mtx_unlock(&neglist->nl_lock);
777 	}
778 
779 	*neglistpp = neglist;
780 	*ncpp = ncp;
781 }
782 
783 static void
784 cache_negative_zap_one(void)
785 {
786 	struct namecache *ncp, *ncp2;
787 	struct neglist *neglist;
788 	struct mtx *dvlp;
789 	struct rwlock *blp;
790 
791 	if (mtx_owner(&ncneg_shrink_lock) != NULL ||
792 	    !mtx_trylock(&ncneg_shrink_lock)) {
793 		counter_u64_add(shrinking_skipped, 1);
794 		return;
795 	}
796 
797 	mtx_lock(&ncneg_hot.nl_lock);
798 	ncp = TAILQ_FIRST(&ncneg_hot.nl_list);
799 	if (ncp != NULL) {
800 		neglist = NCP2NEGLIST(ncp);
801 		mtx_lock(&neglist->nl_lock);
802 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
803 		TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
804 		ncp->nc_flag &= ~NCF_HOTNEGATIVE;
805 		numhotneg--;
806 		mtx_unlock(&neglist->nl_lock);
807 	}
808 	mtx_unlock(&ncneg_hot.nl_lock);
809 
810 	cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
811 	shrink_list_turn++;
812 	if (shrink_list_turn == numneglists)
813 		shrink_list_turn = 0;
814 	if (ncp == NULL && shrink_list_turn == 0)
815 		cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
816 	mtx_unlock(&ncneg_shrink_lock);
817 	if (ncp == NULL)
818 		return;
819 
820 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
821 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
822 	blp = NCP2BUCKETLOCK(ncp);
823 	mtx_unlock(&neglist->nl_lock);
824 	mtx_lock(dvlp);
825 	rw_wlock(blp);
826 	mtx_lock(&neglist->nl_lock);
827 	ncp2 = TAILQ_FIRST(&neglist->nl_list);
828 	if (ncp != ncp2 || dvlp != VP2VNODELOCK(ncp2->nc_dvp) ||
829 	    blp != NCP2BUCKETLOCK(ncp2) || !(ncp2->nc_flag & NCF_NEGATIVE)) {
830 		ncp = NULL;
831 	} else {
832 		SDT_PROBE2(vfs, namecache, shrink_negative, done, ncp->nc_dvp,
833 		    ncp->nc_name);
834 
835 		cache_zap_locked(ncp, true);
836 		counter_u64_add(numneg_evicted, 1);
837 	}
838 	mtx_unlock(&neglist->nl_lock);
839 	rw_wunlock(blp);
840 	mtx_unlock(dvlp);
841 	cache_free(ncp);
842 }
843 
844 /*
845  * cache_zap_locked():
846  *
847  *   Removes a namecache entry from cache, whether it contains an actual
848  *   pointer to a vnode or if it is just a negative cache entry.
849  */
850 static void
851 cache_zap_locked(struct namecache *ncp, bool neg_locked)
852 {
853 
854 	if (!(ncp->nc_flag & NCF_NEGATIVE))
855 		cache_assert_vnode_locked(ncp->nc_vp);
856 	cache_assert_vnode_locked(ncp->nc_dvp);
857 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
858 
859 	CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp,
860 	    (ncp->nc_flag & NCF_NEGATIVE) ? NULL : ncp->nc_vp);
861 	LIST_REMOVE(ncp, nc_hash);
862 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
863 		SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
864 		    ncp->nc_name, ncp->nc_vp);
865 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
866 		if (ncp == ncp->nc_vp->v_cache_dd)
867 			ncp->nc_vp->v_cache_dd = NULL;
868 	} else {
869 		SDT_PROBE2(vfs, namecache, zap_negative, done, ncp->nc_dvp,
870 		    ncp->nc_name);
871 		cache_negative_remove(ncp, neg_locked);
872 	}
873 	if (ncp->nc_flag & NCF_ISDOTDOT) {
874 		if (ncp == ncp->nc_dvp->v_cache_dd)
875 			ncp->nc_dvp->v_cache_dd = NULL;
876 	} else {
877 		LIST_REMOVE(ncp, nc_src);
878 		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
879 			ncp->nc_flag |= NCF_DVDROP;
880 			counter_u64_add(numcachehv, -1);
881 		}
882 	}
883 	atomic_subtract_rel_long(&numcache, 1);
884 }
885 
886 static void
887 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
888 {
889 	struct rwlock *blp;
890 
891 	MPASS(ncp->nc_dvp == vp);
892 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
893 	cache_assert_vnode_locked(vp);
894 
895 	blp = NCP2BUCKETLOCK(ncp);
896 	rw_wlock(blp);
897 	cache_zap_locked(ncp, false);
898 	rw_wunlock(blp);
899 }
900 
901 static bool
902 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
903     struct mtx **vlpp)
904 {
905 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
906 	struct rwlock *blp;
907 
908 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
909 	cache_assert_vnode_locked(vp);
910 
911 	if (ncp->nc_flag & NCF_NEGATIVE) {
912 		if (*vlpp != NULL) {
913 			mtx_unlock(*vlpp);
914 			*vlpp = NULL;
915 		}
916 		cache_zap_negative_locked_vnode_kl(ncp, vp);
917 		return (true);
918 	}
919 
920 	pvlp = VP2VNODELOCK(vp);
921 	blp = NCP2BUCKETLOCK(ncp);
922 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
923 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
924 
925 	if (*vlpp == vlp1 || *vlpp == vlp2) {
926 		to_unlock = *vlpp;
927 		*vlpp = NULL;
928 	} else {
929 		if (*vlpp != NULL) {
930 			mtx_unlock(*vlpp);
931 			*vlpp = NULL;
932 		}
933 		cache_sort_vnodes(&vlp1, &vlp2);
934 		if (vlp1 == pvlp) {
935 			mtx_lock(vlp2);
936 			to_unlock = vlp2;
937 		} else {
938 			if (!mtx_trylock(vlp1))
939 				goto out_relock;
940 			to_unlock = vlp1;
941 		}
942 	}
943 	rw_wlock(blp);
944 	cache_zap_locked(ncp, false);
945 	rw_wunlock(blp);
946 	if (to_unlock != NULL)
947 		mtx_unlock(to_unlock);
948 	return (true);
949 
950 out_relock:
951 	mtx_unlock(vlp2);
952 	mtx_lock(vlp1);
953 	mtx_lock(vlp2);
954 	MPASS(*vlpp == NULL);
955 	*vlpp = vlp1;
956 	return (false);
957 }
958 
959 static int __noinline
960 cache_zap_locked_vnode(struct namecache *ncp, struct vnode *vp)
961 {
962 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
963 	struct rwlock *blp;
964 	int error = 0;
965 
966 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
967 	cache_assert_vnode_locked(vp);
968 
969 	pvlp = VP2VNODELOCK(vp);
970 	if (ncp->nc_flag & NCF_NEGATIVE) {
971 		cache_zap_negative_locked_vnode_kl(ncp, vp);
972 		goto out;
973 	}
974 
975 	blp = NCP2BUCKETLOCK(ncp);
976 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
977 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
978 	cache_sort_vnodes(&vlp1, &vlp2);
979 	if (vlp1 == pvlp) {
980 		mtx_lock(vlp2);
981 		to_unlock = vlp2;
982 	} else {
983 		if (!mtx_trylock(vlp1)) {
984 			error = EAGAIN;
985 			goto out;
986 		}
987 		to_unlock = vlp1;
988 	}
989 	rw_wlock(blp);
990 	cache_zap_locked(ncp, false);
991 	rw_wunlock(blp);
992 	mtx_unlock(to_unlock);
993 out:
994 	mtx_unlock(pvlp);
995 	return (error);
996 }
997 
998 /*
999  * If trylocking failed we can get here. We know enough to take all needed locks
1000  * in the right order and re-lookup the entry.
1001  */
1002 static int
1003 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1004     struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1005     struct rwlock *blp)
1006 {
1007 	struct namecache *rncp;
1008 
1009 	cache_assert_bucket_locked(ncp, RA_UNLOCKED);
1010 
1011 	cache_sort_vnodes(&dvlp, &vlp);
1012 	cache_lock_vnodes(dvlp, vlp);
1013 	rw_wlock(blp);
1014 	LIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1015 		if (rncp == ncp && rncp->nc_dvp == dvp &&
1016 		    rncp->nc_nlen == cnp->cn_namelen &&
1017 		    !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen))
1018 			break;
1019 	}
1020 	if (rncp != NULL) {
1021 		cache_zap_locked(rncp, false);
1022 		rw_wunlock(blp);
1023 		cache_unlock_vnodes(dvlp, vlp);
1024 		counter_u64_add(zap_and_exit_bucket_relock_success, 1);
1025 		return (0);
1026 	}
1027 
1028 	rw_wunlock(blp);
1029 	cache_unlock_vnodes(dvlp, vlp);
1030 	return (EAGAIN);
1031 }
1032 
1033 static int __noinline
1034 cache_zap_wlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1035     uint32_t hash, struct rwlock *blp)
1036 {
1037 	struct mtx *dvlp, *vlp;
1038 	struct vnode *dvp;
1039 
1040 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1041 
1042 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1043 	vlp = NULL;
1044 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1045 		vlp = VP2VNODELOCK(ncp->nc_vp);
1046 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1047 		cache_zap_locked(ncp, false);
1048 		rw_wunlock(blp);
1049 		cache_unlock_vnodes(dvlp, vlp);
1050 		return (0);
1051 	}
1052 
1053 	dvp = ncp->nc_dvp;
1054 	rw_wunlock(blp);
1055 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1056 }
1057 
1058 static int __noinline
1059 cache_zap_rlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1060     uint32_t hash, struct rwlock *blp)
1061 {
1062 	struct mtx *dvlp, *vlp;
1063 	struct vnode *dvp;
1064 
1065 	cache_assert_bucket_locked(ncp, RA_RLOCKED);
1066 
1067 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1068 	vlp = NULL;
1069 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1070 		vlp = VP2VNODELOCK(ncp->nc_vp);
1071 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1072 		rw_runlock(blp);
1073 		rw_wlock(blp);
1074 		cache_zap_locked(ncp, false);
1075 		rw_wunlock(blp);
1076 		cache_unlock_vnodes(dvlp, vlp);
1077 		return (0);
1078 	}
1079 
1080 	dvp = ncp->nc_dvp;
1081 	rw_runlock(blp);
1082 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1083 }
1084 
1085 static int
1086 cache_zap_wlocked_bucket_kl(struct namecache *ncp, struct rwlock *blp,
1087     struct mtx **vlpp1, struct mtx **vlpp2)
1088 {
1089 	struct mtx *dvlp, *vlp;
1090 
1091 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1092 
1093 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1094 	vlp = NULL;
1095 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1096 		vlp = VP2VNODELOCK(ncp->nc_vp);
1097 	cache_sort_vnodes(&dvlp, &vlp);
1098 
1099 	if (*vlpp1 == dvlp && *vlpp2 == vlp) {
1100 		cache_zap_locked(ncp, false);
1101 		cache_unlock_vnodes(dvlp, vlp);
1102 		*vlpp1 = NULL;
1103 		*vlpp2 = NULL;
1104 		return (0);
1105 	}
1106 
1107 	if (*vlpp1 != NULL)
1108 		mtx_unlock(*vlpp1);
1109 	if (*vlpp2 != NULL)
1110 		mtx_unlock(*vlpp2);
1111 	*vlpp1 = NULL;
1112 	*vlpp2 = NULL;
1113 
1114 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1115 		cache_zap_locked(ncp, false);
1116 		cache_unlock_vnodes(dvlp, vlp);
1117 		return (0);
1118 	}
1119 
1120 	rw_wunlock(blp);
1121 	*vlpp1 = dvlp;
1122 	*vlpp2 = vlp;
1123 	if (*vlpp1 != NULL)
1124 		mtx_lock(*vlpp1);
1125 	mtx_lock(*vlpp2);
1126 	rw_wlock(blp);
1127 	return (EAGAIN);
1128 }
1129 
1130 static void
1131 cache_lookup_unlock(struct rwlock *blp, struct mtx *vlp)
1132 {
1133 
1134 	if (blp != NULL) {
1135 		rw_runlock(blp);
1136 	} else {
1137 		mtx_unlock(vlp);
1138 	}
1139 }
1140 
1141 static int __noinline
1142 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1143     struct timespec *tsp, int *ticksp)
1144 {
1145 	int ltype;
1146 
1147 	*vpp = dvp;
1148 	CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .",
1149 			dvp, cnp->cn_nameptr);
1150 	counter_u64_add(dothits, 1);
1151 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1152 	if (tsp != NULL)
1153 		timespecclear(tsp);
1154 	if (ticksp != NULL)
1155 		*ticksp = ticks;
1156 	vrefact(*vpp);
1157 	/*
1158 	 * When we lookup "." we still can be asked to lock it
1159 	 * differently...
1160 	 */
1161 	ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1162 	if (ltype != VOP_ISLOCKED(*vpp)) {
1163 		if (ltype == LK_EXCLUSIVE) {
1164 			vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1165 			if (VN_IS_DOOMED((*vpp))) {
1166 				/* forced unmount */
1167 				vrele(*vpp);
1168 				*vpp = NULL;
1169 				return (ENOENT);
1170 			}
1171 		} else
1172 			vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1173 	}
1174 	return (-1);
1175 }
1176 
1177 static __noinline int
1178 cache_lookup_nomakeentry(struct vnode *dvp, struct vnode **vpp,
1179     struct componentname *cnp, struct timespec *tsp, int *ticksp)
1180 {
1181 	struct namecache *ncp;
1182 	struct rwlock *blp;
1183 	struct mtx *dvlp, *dvlp2;
1184 	uint32_t hash;
1185 	int error;
1186 
1187 	if (cnp->cn_namelen == 2 &&
1188 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1189 		counter_u64_add(dotdothits, 1);
1190 		dvlp = VP2VNODELOCK(dvp);
1191 		dvlp2 = NULL;
1192 		mtx_lock(dvlp);
1193 retry_dotdot:
1194 		ncp = dvp->v_cache_dd;
1195 		if (ncp == NULL) {
1196 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1197 			    "..", NULL);
1198 			mtx_unlock(dvlp);
1199 			if (dvlp2 != NULL)
1200 				mtx_unlock(dvlp2);
1201 			return (0);
1202 		}
1203 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1204 			if (ncp->nc_dvp != dvp)
1205 				panic("dvp %p v_cache_dd %p\n", dvp, ncp);
1206 			if (!cache_zap_locked_vnode_kl2(ncp,
1207 			    dvp, &dvlp2))
1208 				goto retry_dotdot;
1209 			MPASS(dvp->v_cache_dd == NULL);
1210 			mtx_unlock(dvlp);
1211 			if (dvlp2 != NULL)
1212 				mtx_unlock(dvlp2);
1213 			cache_free(ncp);
1214 		} else {
1215 			dvp->v_cache_dd = NULL;
1216 			mtx_unlock(dvlp);
1217 			if (dvlp2 != NULL)
1218 				mtx_unlock(dvlp2);
1219 		}
1220 		return (0);
1221 	}
1222 
1223 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1224 	blp = HASH2BUCKETLOCK(hash);
1225 retry:
1226 	if (LIST_EMPTY(NCHHASH(hash)))
1227 		goto out_no_entry;
1228 
1229 	rw_wlock(blp);
1230 
1231 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1232 		counter_u64_add(numchecks, 1);
1233 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1234 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1235 			break;
1236 	}
1237 
1238 	/* We failed to find an entry */
1239 	if (ncp == NULL) {
1240 		rw_wunlock(blp);
1241 		goto out_no_entry;
1242 	}
1243 
1244 	error = cache_zap_wlocked_bucket(ncp, cnp, hash, blp);
1245 	if (__predict_false(error != 0)) {
1246 		zap_and_exit_bucket_fail++;
1247 		cache_maybe_yield();
1248 		goto retry;
1249 	}
1250 	counter_u64_add(numposzaps, 1);
1251 	cache_free(ncp);
1252 	return (0);
1253 out_no_entry:
1254 	SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr, NULL);
1255 	counter_u64_add(nummisszap, 1);
1256 	return (0);
1257 }
1258 
1259 /**
1260  * Lookup a name in the name cache
1261  *
1262  * # Arguments
1263  *
1264  * - dvp:	Parent directory in which to search.
1265  * - vpp:	Return argument.  Will contain desired vnode on cache hit.
1266  * - cnp:	Parameters of the name search.  The most interesting bits of
1267  *   		the cn_flags field have the following meanings:
1268  *   	- MAKEENTRY:	If clear, free an entry from the cache rather than look
1269  *   			it up.
1270  *   	- ISDOTDOT:	Must be set if and only if cn_nameptr == ".."
1271  * - tsp:	Return storage for cache timestamp.  On a successful (positive
1272  *   		or negative) lookup, tsp will be filled with any timespec that
1273  *   		was stored when this cache entry was created.  However, it will
1274  *   		be clear for "." entries.
1275  * - ticks:	Return storage for alternate cache timestamp.  On a successful
1276  *   		(positive or negative) lookup, it will contain the ticks value
1277  *   		that was current when the cache entry was created, unless cnp
1278  *   		was ".".
1279  *
1280  * # Returns
1281  *
1282  * - -1:	A positive cache hit.  vpp will contain the desired vnode.
1283  * - ENOENT:	A negative cache hit, or dvp was recycled out from under us due
1284  *		to a forced unmount.  vpp will not be modified.  If the entry
1285  *		is a whiteout, then the ISWHITEOUT flag will be set in
1286  *		cnp->cn_flags.
1287  * - 0:		A cache miss.  vpp will not be modified.
1288  *
1289  * # Locking
1290  *
1291  * On a cache hit, vpp will be returned locked and ref'd.  If we're looking up
1292  * .., dvp is unlocked.  If we're looking up . an extra ref is taken, but the
1293  * lock is not recursively acquired.
1294  */
1295 int
1296 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1297     struct timespec *tsp, int *ticksp)
1298 {
1299 	struct namecache_ts *ncp_ts;
1300 	struct namecache *ncp;
1301 	struct rwlock *blp;
1302 	struct mtx *dvlp;
1303 	uint32_t hash;
1304 	enum vgetstate vs;
1305 	int error, ltype;
1306 
1307 #ifdef DEBUG_CACHE
1308 	if (__predict_false(!doingcache)) {
1309 		cnp->cn_flags &= ~MAKEENTRY;
1310 		return (0);
1311 	}
1312 #endif
1313 
1314 	counter_u64_add(numcalls, 1);
1315 
1316 	if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.'))
1317 		return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1318 
1319 	if ((cnp->cn_flags & MAKEENTRY) == 0)
1320 		return (cache_lookup_nomakeentry(dvp, vpp, cnp, tsp, ticksp));
1321 
1322 retry:
1323 	blp = NULL;
1324 	dvlp = NULL;
1325 	error = 0;
1326 	if (cnp->cn_namelen == 2 &&
1327 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1328 		counter_u64_add(dotdothits, 1);
1329 		dvlp = VP2VNODELOCK(dvp);
1330 		mtx_lock(dvlp);
1331 		ncp = dvp->v_cache_dd;
1332 		if (ncp == NULL) {
1333 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1334 			    "..", NULL);
1335 			mtx_unlock(dvlp);
1336 			return (0);
1337 		}
1338 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1339 			if (ncp->nc_flag & NCF_NEGATIVE)
1340 				*vpp = NULL;
1341 			else
1342 				*vpp = ncp->nc_vp;
1343 		} else
1344 			*vpp = ncp->nc_dvp;
1345 		/* Return failure if negative entry was found. */
1346 		if (*vpp == NULL)
1347 			goto negative_success;
1348 		CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
1349 		    dvp, cnp->cn_nameptr, *vpp);
1350 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..",
1351 		    *vpp);
1352 		cache_out_ts(ncp, tsp, ticksp);
1353 		if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1354 		    NCF_DTS && tsp != NULL) {
1355 			ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1356 			*tsp = ncp_ts->nc_dotdottime;
1357 		}
1358 		goto success;
1359 	}
1360 
1361 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1362 	blp = HASH2BUCKETLOCK(hash);
1363 	rw_rlock(blp);
1364 
1365 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1366 		counter_u64_add(numchecks, 1);
1367 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1368 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1369 			break;
1370 	}
1371 
1372 	/* We failed to find an entry */
1373 	if (__predict_false(ncp == NULL)) {
1374 		rw_runlock(blp);
1375 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1376 		    NULL);
1377 		counter_u64_add(nummiss, 1);
1378 		return (0);
1379 	}
1380 
1381 	if (ncp->nc_flag & NCF_NEGATIVE)
1382 		goto negative_success;
1383 
1384 	/* We found a "positive" match, return the vnode */
1385 	counter_u64_add(numposhits, 1);
1386 	*vpp = ncp->nc_vp;
1387 	CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p",
1388 	    dvp, cnp->cn_nameptr, *vpp, ncp);
1389 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name,
1390 	    *vpp);
1391 	cache_out_ts(ncp, tsp, ticksp);
1392 success:
1393 	/*
1394 	 * On success we return a locked and ref'd vnode as per the lookup
1395 	 * protocol.
1396 	 */
1397 	MPASS(dvp != *vpp);
1398 	ltype = 0;	/* silence gcc warning */
1399 	if (cnp->cn_flags & ISDOTDOT) {
1400 		ltype = VOP_ISLOCKED(dvp);
1401 		VOP_UNLOCK(dvp);
1402 	}
1403 	vs = vget_prep(*vpp);
1404 	cache_lookup_unlock(blp, dvlp);
1405 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1406 	if (cnp->cn_flags & ISDOTDOT) {
1407 		vn_lock(dvp, ltype | LK_RETRY);
1408 		if (VN_IS_DOOMED(dvp)) {
1409 			if (error == 0)
1410 				vput(*vpp);
1411 			*vpp = NULL;
1412 			return (ENOENT);
1413 		}
1414 	}
1415 	if (error) {
1416 		*vpp = NULL;
1417 		goto retry;
1418 	}
1419 	if ((cnp->cn_flags & ISLASTCN) &&
1420 	    (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) {
1421 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
1422 	}
1423 	return (-1);
1424 
1425 negative_success:
1426 	/* We found a negative match, and want to create it, so purge */
1427 	if (cnp->cn_nameiop == CREATE) {
1428 		counter_u64_add(numnegzaps, 1);
1429 		goto zap_and_exit;
1430 	}
1431 
1432 	counter_u64_add(numneghits, 1);
1433 	cache_negative_hit(ncp);
1434 	if (ncp->nc_flag & NCF_WHITE)
1435 		cnp->cn_flags |= ISWHITEOUT;
1436 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp,
1437 	    ncp->nc_name);
1438 	cache_out_ts(ncp, tsp, ticksp);
1439 	cache_lookup_unlock(blp, dvlp);
1440 	return (ENOENT);
1441 
1442 zap_and_exit:
1443 	if (blp != NULL)
1444 		error = cache_zap_rlocked_bucket(ncp, cnp, hash, blp);
1445 	else
1446 		error = cache_zap_locked_vnode(ncp, dvp);
1447 	if (__predict_false(error != 0)) {
1448 		zap_and_exit_bucket_fail2++;
1449 		cache_maybe_yield();
1450 		goto retry;
1451 	}
1452 	cache_free(ncp);
1453 	return (0);
1454 }
1455 
1456 struct celockstate {
1457 	struct mtx *vlp[3];
1458 	struct rwlock *blp[2];
1459 };
1460 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1461 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1462 
1463 static inline void
1464 cache_celockstate_init(struct celockstate *cel)
1465 {
1466 
1467 	bzero(cel, sizeof(*cel));
1468 }
1469 
1470 static void
1471 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1472     struct vnode *dvp)
1473 {
1474 	struct mtx *vlp1, *vlp2;
1475 
1476 	MPASS(cel->vlp[0] == NULL);
1477 	MPASS(cel->vlp[1] == NULL);
1478 	MPASS(cel->vlp[2] == NULL);
1479 
1480 	MPASS(vp != NULL || dvp != NULL);
1481 
1482 	vlp1 = VP2VNODELOCK(vp);
1483 	vlp2 = VP2VNODELOCK(dvp);
1484 	cache_sort_vnodes(&vlp1, &vlp2);
1485 
1486 	if (vlp1 != NULL) {
1487 		mtx_lock(vlp1);
1488 		cel->vlp[0] = vlp1;
1489 	}
1490 	mtx_lock(vlp2);
1491 	cel->vlp[1] = vlp2;
1492 }
1493 
1494 static void
1495 cache_unlock_vnodes_cel(struct celockstate *cel)
1496 {
1497 
1498 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
1499 
1500 	if (cel->vlp[0] != NULL)
1501 		mtx_unlock(cel->vlp[0]);
1502 	if (cel->vlp[1] != NULL)
1503 		mtx_unlock(cel->vlp[1]);
1504 	if (cel->vlp[2] != NULL)
1505 		mtx_unlock(cel->vlp[2]);
1506 }
1507 
1508 static bool
1509 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
1510 {
1511 	struct mtx *vlp;
1512 	bool ret;
1513 
1514 	cache_assert_vlp_locked(cel->vlp[0]);
1515 	cache_assert_vlp_locked(cel->vlp[1]);
1516 	MPASS(cel->vlp[2] == NULL);
1517 
1518 	MPASS(vp != NULL);
1519 	vlp = VP2VNODELOCK(vp);
1520 
1521 	ret = true;
1522 	if (vlp >= cel->vlp[1]) {
1523 		mtx_lock(vlp);
1524 	} else {
1525 		if (mtx_trylock(vlp))
1526 			goto out;
1527 		cache_lock_vnodes_cel_3_failures++;
1528 		cache_unlock_vnodes_cel(cel);
1529 		if (vlp < cel->vlp[0]) {
1530 			mtx_lock(vlp);
1531 			mtx_lock(cel->vlp[0]);
1532 			mtx_lock(cel->vlp[1]);
1533 		} else {
1534 			if (cel->vlp[0] != NULL)
1535 				mtx_lock(cel->vlp[0]);
1536 			mtx_lock(vlp);
1537 			mtx_lock(cel->vlp[1]);
1538 		}
1539 		ret = false;
1540 	}
1541 out:
1542 	cel->vlp[2] = vlp;
1543 	return (ret);
1544 }
1545 
1546 static void
1547 cache_lock_buckets_cel(struct celockstate *cel, struct rwlock *blp1,
1548     struct rwlock *blp2)
1549 {
1550 
1551 	MPASS(cel->blp[0] == NULL);
1552 	MPASS(cel->blp[1] == NULL);
1553 
1554 	cache_sort_vnodes(&blp1, &blp2);
1555 
1556 	if (blp1 != NULL) {
1557 		rw_wlock(blp1);
1558 		cel->blp[0] = blp1;
1559 	}
1560 	rw_wlock(blp2);
1561 	cel->blp[1] = blp2;
1562 }
1563 
1564 static void
1565 cache_unlock_buckets_cel(struct celockstate *cel)
1566 {
1567 
1568 	if (cel->blp[0] != NULL)
1569 		rw_wunlock(cel->blp[0]);
1570 	rw_wunlock(cel->blp[1]);
1571 }
1572 
1573 /*
1574  * Lock part of the cache affected by the insertion.
1575  *
1576  * This means vnodelocks for dvp, vp and the relevant bucketlock.
1577  * However, insertion can result in removal of an old entry. In this
1578  * case we have an additional vnode and bucketlock pair to lock. If the
1579  * entry is negative, ncelock is locked instead of the vnode.
1580  *
1581  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
1582  * preserving the locking order (smaller address first).
1583  */
1584 static void
1585 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1586     uint32_t hash)
1587 {
1588 	struct namecache *ncp;
1589 	struct rwlock *blps[2];
1590 
1591 	blps[0] = HASH2BUCKETLOCK(hash);
1592 	for (;;) {
1593 		blps[1] = NULL;
1594 		cache_lock_vnodes_cel(cel, dvp, vp);
1595 		if (vp == NULL || vp->v_type != VDIR)
1596 			break;
1597 		ncp = vp->v_cache_dd;
1598 		if (ncp == NULL)
1599 			break;
1600 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1601 			break;
1602 		MPASS(ncp->nc_dvp == vp);
1603 		blps[1] = NCP2BUCKETLOCK(ncp);
1604 		if (ncp->nc_flag & NCF_NEGATIVE)
1605 			break;
1606 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1607 			break;
1608 		/*
1609 		 * All vnodes got re-locked. Re-validate the state and if
1610 		 * nothing changed we are done. Otherwise restart.
1611 		 */
1612 		if (ncp == vp->v_cache_dd &&
1613 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1614 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1615 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1616 			break;
1617 		cache_unlock_vnodes_cel(cel);
1618 		cel->vlp[0] = NULL;
1619 		cel->vlp[1] = NULL;
1620 		cel->vlp[2] = NULL;
1621 	}
1622 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1623 }
1624 
1625 static void
1626 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1627     uint32_t hash)
1628 {
1629 	struct namecache *ncp;
1630 	struct rwlock *blps[2];
1631 
1632 	blps[0] = HASH2BUCKETLOCK(hash);
1633 	for (;;) {
1634 		blps[1] = NULL;
1635 		cache_lock_vnodes_cel(cel, dvp, vp);
1636 		ncp = dvp->v_cache_dd;
1637 		if (ncp == NULL)
1638 			break;
1639 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1640 			break;
1641 		MPASS(ncp->nc_dvp == dvp);
1642 		blps[1] = NCP2BUCKETLOCK(ncp);
1643 		if (ncp->nc_flag & NCF_NEGATIVE)
1644 			break;
1645 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1646 			break;
1647 		if (ncp == dvp->v_cache_dd &&
1648 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1649 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1650 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1651 			break;
1652 		cache_unlock_vnodes_cel(cel);
1653 		cel->vlp[0] = NULL;
1654 		cel->vlp[1] = NULL;
1655 		cel->vlp[2] = NULL;
1656 	}
1657 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1658 }
1659 
1660 static void
1661 cache_enter_unlock(struct celockstate *cel)
1662 {
1663 
1664 	cache_unlock_buckets_cel(cel);
1665 	cache_unlock_vnodes_cel(cel);
1666 }
1667 
1668 static void __noinline
1669 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp,
1670     struct componentname *cnp)
1671 {
1672 	struct celockstate cel;
1673 	struct namecache *ncp;
1674 	uint32_t hash;
1675 	int len;
1676 
1677 	if (dvp->v_cache_dd == NULL)
1678 		return;
1679 	len = cnp->cn_namelen;
1680 	cache_celockstate_init(&cel);
1681 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1682 	cache_enter_lock_dd(&cel, dvp, vp, hash);
1683 	ncp = dvp->v_cache_dd;
1684 	if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) {
1685 		KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent"));
1686 		cache_zap_locked(ncp, false);
1687 	} else {
1688 		ncp = NULL;
1689 	}
1690 	dvp->v_cache_dd = NULL;
1691 	cache_enter_unlock(&cel);
1692 	cache_free(ncp);
1693 }
1694 
1695 /*
1696  * Add an entry to the cache.
1697  */
1698 void
1699 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
1700     struct timespec *tsp, struct timespec *dtsp)
1701 {
1702 	struct celockstate cel;
1703 	struct namecache *ncp, *n2, *ndd;
1704 	struct namecache_ts *ncp_ts, *n2_ts;
1705 	struct nchashhead *ncpp;
1706 	uint32_t hash;
1707 	int flag;
1708 	int len;
1709 	u_long lnumcache;
1710 
1711 	CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr);
1712 	VNASSERT(vp == NULL || !VN_IS_DOOMED(vp), vp,
1713 	    ("cache_enter: Adding a doomed vnode"));
1714 	VNASSERT(dvp == NULL || !VN_IS_DOOMED(dvp), dvp,
1715 	    ("cache_enter: Doomed vnode used as src"));
1716 
1717 #ifdef DEBUG_CACHE
1718 	if (__predict_false(!doingcache))
1719 		return;
1720 #endif
1721 
1722 	flag = 0;
1723 	if (__predict_false(cnp->cn_nameptr[0] == '.')) {
1724 		if (cnp->cn_namelen == 1)
1725 			return;
1726 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1727 			cache_enter_dotdot_prep(dvp, vp, cnp);
1728 			flag = NCF_ISDOTDOT;
1729 		}
1730 	}
1731 
1732 	/*
1733 	 * Avoid blowout in namecache entries.
1734 	 */
1735 	lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
1736 	if (__predict_false(lnumcache >= ncsize)) {
1737 		atomic_add_long(&numcache, -1);
1738 		return;
1739 	}
1740 
1741 	cache_celockstate_init(&cel);
1742 	ndd = NULL;
1743 	ncp_ts = NULL;
1744 
1745 	/*
1746 	 * Calculate the hash key and setup as much of the new
1747 	 * namecache entry as possible before acquiring the lock.
1748 	 */
1749 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
1750 	ncp->nc_flag = flag;
1751 	ncp->nc_vp = vp;
1752 	if (vp == NULL)
1753 		ncp->nc_flag |= NCF_NEGATIVE;
1754 	ncp->nc_dvp = dvp;
1755 	if (tsp != NULL) {
1756 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1757 		ncp_ts->nc_time = *tsp;
1758 		ncp_ts->nc_ticks = ticks;
1759 		ncp_ts->nc_nc.nc_flag |= NCF_TS;
1760 		if (dtsp != NULL) {
1761 			ncp_ts->nc_dotdottime = *dtsp;
1762 			ncp_ts->nc_nc.nc_flag |= NCF_DTS;
1763 		}
1764 	}
1765 	len = ncp->nc_nlen = cnp->cn_namelen;
1766 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1767 	strlcpy(ncp->nc_name, cnp->cn_nameptr, len + 1);
1768 	cache_enter_lock(&cel, dvp, vp, hash);
1769 
1770 	/*
1771 	 * See if this vnode or negative entry is already in the cache
1772 	 * with this name.  This can happen with concurrent lookups of
1773 	 * the same path name.
1774 	 */
1775 	ncpp = NCHHASH(hash);
1776 	LIST_FOREACH(n2, ncpp, nc_hash) {
1777 		if (n2->nc_dvp == dvp &&
1778 		    n2->nc_nlen == cnp->cn_namelen &&
1779 		    !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
1780 			if (tsp != NULL) {
1781 				KASSERT((n2->nc_flag & NCF_TS) != 0,
1782 				    ("no NCF_TS"));
1783 				n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
1784 				n2_ts->nc_time = ncp_ts->nc_time;
1785 				n2_ts->nc_ticks = ncp_ts->nc_ticks;
1786 				if (dtsp != NULL) {
1787 					n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
1788 					if (ncp->nc_flag & NCF_NEGATIVE)
1789 						mtx_lock(&ncneg_hot.nl_lock);
1790 					n2_ts->nc_nc.nc_flag |= NCF_DTS;
1791 					if (ncp->nc_flag & NCF_NEGATIVE)
1792 						mtx_unlock(&ncneg_hot.nl_lock);
1793 				}
1794 			}
1795 			goto out_unlock_free;
1796 		}
1797 	}
1798 
1799 	if (flag == NCF_ISDOTDOT) {
1800 		/*
1801 		 * See if we are trying to add .. entry, but some other lookup
1802 		 * has populated v_cache_dd pointer already.
1803 		 */
1804 		if (dvp->v_cache_dd != NULL)
1805 			goto out_unlock_free;
1806 		KASSERT(vp == NULL || vp->v_type == VDIR,
1807 		    ("wrong vnode type %p", vp));
1808 		dvp->v_cache_dd = ncp;
1809 	}
1810 
1811 	if (vp != NULL) {
1812 		if (vp->v_type == VDIR) {
1813 			if (flag != NCF_ISDOTDOT) {
1814 				/*
1815 				 * For this case, the cache entry maps both the
1816 				 * directory name in it and the name ".." for the
1817 				 * directory's parent.
1818 				 */
1819 				if ((ndd = vp->v_cache_dd) != NULL) {
1820 					if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
1821 						cache_zap_locked(ndd, false);
1822 					else
1823 						ndd = NULL;
1824 				}
1825 				vp->v_cache_dd = ncp;
1826 			}
1827 		} else {
1828 			vp->v_cache_dd = NULL;
1829 		}
1830 	}
1831 
1832 	if (flag != NCF_ISDOTDOT) {
1833 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1834 			vhold(dvp);
1835 			counter_u64_add(numcachehv, 1);
1836 		}
1837 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
1838 	}
1839 
1840 	/*
1841 	 * Insert the new namecache entry into the appropriate chain
1842 	 * within the cache entries table.
1843 	 */
1844 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
1845 
1846 	/*
1847 	 * If the entry is "negative", we place it into the
1848 	 * "negative" cache queue, otherwise, we place it into the
1849 	 * destination vnode's cache entries queue.
1850 	 */
1851 	if (vp != NULL) {
1852 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
1853 		SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
1854 		    vp);
1855 	} else {
1856 		if (cnp->cn_flags & ISWHITEOUT)
1857 			ncp->nc_flag |= NCF_WHITE;
1858 		cache_negative_insert(ncp, false);
1859 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
1860 		    ncp->nc_name);
1861 	}
1862 	cache_enter_unlock(&cel);
1863 	if (numneg * ncnegfactor > lnumcache)
1864 		cache_negative_zap_one();
1865 	cache_free(ndd);
1866 	return;
1867 out_unlock_free:
1868 	cache_enter_unlock(&cel);
1869 	cache_free(ncp);
1870 	return;
1871 }
1872 
1873 static u_int
1874 cache_roundup_2(u_int val)
1875 {
1876 	u_int res;
1877 
1878 	for (res = 1; res <= val; res <<= 1)
1879 		continue;
1880 
1881 	return (res);
1882 }
1883 
1884 /*
1885  * Name cache initialization, from vfs_init() when we are booting
1886  */
1887 static void
1888 nchinit(void *dummy __unused)
1889 {
1890 	u_int i;
1891 
1892 	cache_zone_small = uma_zcreate("S VFS Cache",
1893 	    sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1,
1894 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1895 	    UMA_ZONE_ZINIT);
1896 	cache_zone_small_ts = uma_zcreate("STS VFS Cache",
1897 	    sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1,
1898 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1899 	    UMA_ZONE_ZINIT);
1900 	cache_zone_large = uma_zcreate("L VFS Cache",
1901 	    sizeof(struct namecache) + NAME_MAX + 1,
1902 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1903 	    UMA_ZONE_ZINIT);
1904 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache",
1905 	    sizeof(struct namecache_ts) + NAME_MAX + 1,
1906 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1907 	    UMA_ZONE_ZINIT);
1908 
1909 	ncsize = desiredvnodes * ncsizefactor;
1910 	nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash);
1911 	ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
1912 	if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
1913 		ncbuckethash = 7;
1914 	if (ncbuckethash > nchash)
1915 		ncbuckethash = nchash;
1916 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
1917 	    M_WAITOK | M_ZERO);
1918 	for (i = 0; i < numbucketlocks; i++)
1919 		rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK | RW_RECURSE);
1920 	ncvnodehash = ncbuckethash;
1921 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
1922 	    M_WAITOK | M_ZERO);
1923 	for (i = 0; i < numvnodelocks; i++)
1924 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
1925 	ncpurgeminvnodes = numbucketlocks * 2;
1926 
1927 	ncneghash = 3;
1928 	neglists = malloc(sizeof(*neglists) * numneglists, M_VFSCACHE,
1929 	    M_WAITOK | M_ZERO);
1930 	for (i = 0; i < numneglists; i++) {
1931 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
1932 		TAILQ_INIT(&neglists[i].nl_list);
1933 	}
1934 	mtx_init(&ncneg_hot.nl_lock, "ncneglh", NULL, MTX_DEF);
1935 	TAILQ_INIT(&ncneg_hot.nl_list);
1936 
1937 	mtx_init(&ncneg_shrink_lock, "ncnegs", NULL, MTX_DEF);
1938 
1939 	numcachehv = counter_u64_alloc(M_WAITOK);
1940 	numcalls = counter_u64_alloc(M_WAITOK);
1941 	dothits = counter_u64_alloc(M_WAITOK);
1942 	dotdothits = counter_u64_alloc(M_WAITOK);
1943 	numchecks = counter_u64_alloc(M_WAITOK);
1944 	nummiss = counter_u64_alloc(M_WAITOK);
1945 	nummisszap = counter_u64_alloc(M_WAITOK);
1946 	numposzaps = counter_u64_alloc(M_WAITOK);
1947 	numposhits = counter_u64_alloc(M_WAITOK);
1948 	numnegzaps = counter_u64_alloc(M_WAITOK);
1949 	numneghits = counter_u64_alloc(M_WAITOK);
1950 	numfullpathcalls = counter_u64_alloc(M_WAITOK);
1951 	numfullpathfail1 = counter_u64_alloc(M_WAITOK);
1952 	numfullpathfail2 = counter_u64_alloc(M_WAITOK);
1953 	numfullpathfail4 = counter_u64_alloc(M_WAITOK);
1954 	numfullpathfound = counter_u64_alloc(M_WAITOK);
1955 	zap_and_exit_bucket_relock_success = counter_u64_alloc(M_WAITOK);
1956 	numneg_evicted = counter_u64_alloc(M_WAITOK);
1957 	shrinking_skipped = counter_u64_alloc(M_WAITOK);
1958 }
1959 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
1960 
1961 void
1962 cache_changesize(u_long newmaxvnodes)
1963 {
1964 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
1965 	u_long new_nchash, old_nchash;
1966 	struct namecache *ncp;
1967 	uint32_t hash;
1968 	u_long newncsize;
1969 	int i;
1970 
1971 	newncsize = newmaxvnodes * ncsizefactor;
1972 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
1973 	if (newmaxvnodes < numbucketlocks)
1974 		newmaxvnodes = numbucketlocks;
1975 
1976 	new_nchashtbl = hashinit(newmaxvnodes, M_VFSCACHE, &new_nchash);
1977 	/* If same hash table size, nothing to do */
1978 	if (nchash == new_nchash) {
1979 		free(new_nchashtbl, M_VFSCACHE);
1980 		return;
1981 	}
1982 	/*
1983 	 * Move everything from the old hash table to the new table.
1984 	 * None of the namecache entries in the table can be removed
1985 	 * because to do so, they have to be removed from the hash table.
1986 	 */
1987 	cache_lock_all_vnodes();
1988 	cache_lock_all_buckets();
1989 	old_nchashtbl = nchashtbl;
1990 	old_nchash = nchash;
1991 	nchashtbl = new_nchashtbl;
1992 	nchash = new_nchash;
1993 	for (i = 0; i <= old_nchash; i++) {
1994 		while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) {
1995 			hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
1996 			    ncp->nc_dvp);
1997 			LIST_REMOVE(ncp, nc_hash);
1998 			LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
1999 		}
2000 	}
2001 	ncsize = newncsize;
2002 	cache_unlock_all_buckets();
2003 	cache_unlock_all_vnodes();
2004 	free(old_nchashtbl, M_VFSCACHE);
2005 }
2006 
2007 /*
2008  * Invalidate all entries from and to a particular vnode.
2009  */
2010 void
2011 cache_purge(struct vnode *vp)
2012 {
2013 	TAILQ_HEAD(, namecache) ncps;
2014 	struct namecache *ncp, *nnp;
2015 	struct mtx *vlp, *vlp2;
2016 
2017 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
2018 	SDT_PROBE1(vfs, namecache, purge, done, vp);
2019 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2020 	    vp->v_cache_dd == NULL)
2021 		return;
2022 	TAILQ_INIT(&ncps);
2023 	vlp = VP2VNODELOCK(vp);
2024 	vlp2 = NULL;
2025 	mtx_lock(vlp);
2026 retry:
2027 	while (!LIST_EMPTY(&vp->v_cache_src)) {
2028 		ncp = LIST_FIRST(&vp->v_cache_src);
2029 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2030 			goto retry;
2031 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2032 	}
2033 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2034 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
2035 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2036 			goto retry;
2037 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2038 	}
2039 	ncp = vp->v_cache_dd;
2040 	if (ncp != NULL) {
2041 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2042 		   ("lost dotdot link"));
2043 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2044 			goto retry;
2045 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2046 	}
2047 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2048 	mtx_unlock(vlp);
2049 	if (vlp2 != NULL)
2050 		mtx_unlock(vlp2);
2051 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2052 		cache_free(ncp);
2053 	}
2054 }
2055 
2056 /*
2057  * Invalidate all negative entries for a particular directory vnode.
2058  */
2059 void
2060 cache_purge_negative(struct vnode *vp)
2061 {
2062 	TAILQ_HEAD(, namecache) ncps;
2063 	struct namecache *ncp, *nnp;
2064 	struct mtx *vlp;
2065 
2066 	CTR1(KTR_VFS, "cache_purge_negative(%p)", vp);
2067 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2068 	if (LIST_EMPTY(&vp->v_cache_src))
2069 		return;
2070 	TAILQ_INIT(&ncps);
2071 	vlp = VP2VNODELOCK(vp);
2072 	mtx_lock(vlp);
2073 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2074 		if (!(ncp->nc_flag & NCF_NEGATIVE))
2075 			continue;
2076 		cache_zap_negative_locked_vnode_kl(ncp, vp);
2077 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2078 	}
2079 	mtx_unlock(vlp);
2080 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2081 		cache_free(ncp);
2082 	}
2083 }
2084 
2085 /*
2086  * Flush all entries referencing a particular filesystem.
2087  */
2088 void
2089 cache_purgevfs(struct mount *mp, bool force)
2090 {
2091 	TAILQ_HEAD(, namecache) ncps;
2092 	struct mtx *vlp1, *vlp2;
2093 	struct rwlock *blp;
2094 	struct nchashhead *bucket;
2095 	struct namecache *ncp, *nnp;
2096 	u_long i, j, n_nchash;
2097 	int error;
2098 
2099 	/* Scan hash tables for applicable entries */
2100 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2101 	if (!force && mp->mnt_nvnodelistsize <= ncpurgeminvnodes)
2102 		return;
2103 	TAILQ_INIT(&ncps);
2104 	n_nchash = nchash + 1;
2105 	vlp1 = vlp2 = NULL;
2106 	for (i = 0; i < numbucketlocks; i++) {
2107 		blp = (struct rwlock *)&bucketlocks[i];
2108 		rw_wlock(blp);
2109 		for (j = i; j < n_nchash; j += numbucketlocks) {
2110 retry:
2111 			bucket = &nchashtbl[j];
2112 			LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) {
2113 				cache_assert_bucket_locked(ncp, RA_WLOCKED);
2114 				if (ncp->nc_dvp->v_mount != mp)
2115 					continue;
2116 				error = cache_zap_wlocked_bucket_kl(ncp, blp,
2117 				    &vlp1, &vlp2);
2118 				if (error != 0)
2119 					goto retry;
2120 				TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst);
2121 			}
2122 		}
2123 		rw_wunlock(blp);
2124 		if (vlp1 == NULL && vlp2 == NULL)
2125 			cache_maybe_yield();
2126 	}
2127 	if (vlp1 != NULL)
2128 		mtx_unlock(vlp1);
2129 	if (vlp2 != NULL)
2130 		mtx_unlock(vlp2);
2131 
2132 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2133 		cache_free(ncp);
2134 	}
2135 }
2136 
2137 /*
2138  * Perform canonical checks and cache lookup and pass on to filesystem
2139  * through the vop_cachedlookup only if needed.
2140  */
2141 
2142 int
2143 vfs_cache_lookup(struct vop_lookup_args *ap)
2144 {
2145 	struct vnode *dvp;
2146 	int error;
2147 	struct vnode **vpp = ap->a_vpp;
2148 	struct componentname *cnp = ap->a_cnp;
2149 	int flags = cnp->cn_flags;
2150 
2151 	*vpp = NULL;
2152 	dvp = ap->a_dvp;
2153 
2154 	if (dvp->v_type != VDIR)
2155 		return (ENOTDIR);
2156 
2157 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2158 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2159 		return (EROFS);
2160 
2161 	error = vn_dir_check_exec(dvp, cnp);
2162 	if (error != 0)
2163 		return (error);
2164 
2165 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2166 	if (error == 0)
2167 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2168 	if (error == -1)
2169 		return (0);
2170 	return (error);
2171 }
2172 
2173 /* Implementation of the getcwd syscall. */
2174 int
2175 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2176 {
2177 	char *buf, *retbuf;
2178 	size_t buflen;
2179 	int error;
2180 
2181 	buflen = uap->buflen;
2182 	if (__predict_false(buflen < 2))
2183 		return (EINVAL);
2184 	if (buflen > MAXPATHLEN)
2185 		buflen = MAXPATHLEN;
2186 
2187 	buf = malloc(buflen, M_TEMP, M_WAITOK);
2188 	error = vn_getcwd(td, buf, &retbuf, &buflen);
2189 	if (error == 0)
2190 		error = copyout(retbuf, uap->buf, buflen);
2191 	free(buf, M_TEMP);
2192 	return (error);
2193 }
2194 
2195 int
2196 vn_getcwd(struct thread *td, char *buf, char **retbuf, size_t *buflen)
2197 {
2198 	struct filedesc *fdp;
2199 	struct vnode *cdir, *rdir;
2200 	int error;
2201 
2202 	fdp = td->td_proc->p_fd;
2203 	FILEDESC_SLOCK(fdp);
2204 	cdir = fdp->fd_cdir;
2205 	vrefact(cdir);
2206 	rdir = fdp->fd_rdir;
2207 	vrefact(rdir);
2208 	FILEDESC_SUNLOCK(fdp);
2209 	error = vn_fullpath_any(td, cdir, rdir, buf, retbuf, buflen);
2210 	vrele(rdir);
2211 	vrele(cdir);
2212 
2213 #ifdef KTRACE
2214 	if (KTRPOINT(curthread, KTR_NAMEI) && error == 0)
2215 		ktrnamei(*retbuf);
2216 #endif
2217 	return (error);
2218 }
2219 
2220 static int
2221 kern___realpathat(struct thread *td, int fd, const char *path, char *buf,
2222     size_t size, int flags, enum uio_seg pathseg)
2223 {
2224 	struct nameidata nd;
2225 	char *retbuf, *freebuf;
2226 	int error;
2227 
2228 	if (flags != 0)
2229 		return (EINVAL);
2230 	NDINIT_ATRIGHTS(&nd, LOOKUP, FOLLOW | SAVENAME | WANTPARENT | AUDITVNODE1,
2231 	    pathseg, path, fd, &cap_fstat_rights, td);
2232 	if ((error = namei(&nd)) != 0)
2233 		return (error);
2234 	error = vn_fullpath_hardlink(td, &nd, &retbuf, &freebuf, &size);
2235 	if (error == 0) {
2236 		error = copyout(retbuf, buf, size);
2237 		free(freebuf, M_TEMP);
2238 	}
2239 	NDFREE(&nd, 0);
2240 	return (error);
2241 }
2242 
2243 int
2244 sys___realpathat(struct thread *td, struct __realpathat_args *uap)
2245 {
2246 
2247 	return (kern___realpathat(td, uap->fd, uap->path, uap->buf, uap->size,
2248 	    uap->flags, UIO_USERSPACE));
2249 }
2250 
2251 /*
2252  * Retrieve the full filesystem path that correspond to a vnode from the name
2253  * cache (if available)
2254  */
2255 int
2256 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf)
2257 {
2258 	char *buf;
2259 	struct filedesc *fdp;
2260 	struct vnode *rdir;
2261 	size_t buflen;
2262 	int error;
2263 
2264 	if (__predict_false(vn == NULL))
2265 		return (EINVAL);
2266 
2267 	buflen = MAXPATHLEN;
2268 	buf = malloc(buflen, M_TEMP, M_WAITOK);
2269 	fdp = td->td_proc->p_fd;
2270 	FILEDESC_SLOCK(fdp);
2271 	rdir = fdp->fd_rdir;
2272 	vrefact(rdir);
2273 	FILEDESC_SUNLOCK(fdp);
2274 	error = vn_fullpath_any(td, vn, rdir, buf, retbuf, &buflen);
2275 	vrele(rdir);
2276 
2277 	if (!error)
2278 		*freebuf = buf;
2279 	else
2280 		free(buf, M_TEMP);
2281 	return (error);
2282 }
2283 
2284 /*
2285  * This function is similar to vn_fullpath, but it attempts to lookup the
2286  * pathname relative to the global root mount point.  This is required for the
2287  * auditing sub-system, as audited pathnames must be absolute, relative to the
2288  * global root mount point.
2289  */
2290 int
2291 vn_fullpath_global(struct thread *td, struct vnode *vn,
2292     char **retbuf, char **freebuf)
2293 {
2294 	char *buf;
2295 	size_t buflen;
2296 	int error;
2297 
2298 	if (__predict_false(vn == NULL))
2299 		return (EINVAL);
2300 	buflen = MAXPATHLEN;
2301 	buf = malloc(buflen, M_TEMP, M_WAITOK);
2302 	error = vn_fullpath_any(td, vn, rootvnode, buf, retbuf, &buflen);
2303 	if (!error)
2304 		*freebuf = buf;
2305 	else
2306 		free(buf, M_TEMP);
2307 	return (error);
2308 }
2309 
2310 int
2311 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, size_t *buflen)
2312 {
2313 	struct vnode *dvp;
2314 	struct namecache *ncp;
2315 	struct mtx *vlp;
2316 	int error;
2317 
2318 	vlp = VP2VNODELOCK(*vp);
2319 	mtx_lock(vlp);
2320 	TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) {
2321 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2322 			break;
2323 	}
2324 	if (ncp != NULL) {
2325 		if (*buflen < ncp->nc_nlen) {
2326 			mtx_unlock(vlp);
2327 			vrele(*vp);
2328 			counter_u64_add(numfullpathfail4, 1);
2329 			error = ENOMEM;
2330 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2331 			    vp, NULL);
2332 			return (error);
2333 		}
2334 		*buflen -= ncp->nc_nlen;
2335 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2336 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2337 		    ncp->nc_name, vp);
2338 		dvp = *vp;
2339 		*vp = ncp->nc_dvp;
2340 		vref(*vp);
2341 		mtx_unlock(vlp);
2342 		vrele(dvp);
2343 		return (0);
2344 	}
2345 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2346 
2347 	mtx_unlock(vlp);
2348 	vn_lock(*vp, LK_SHARED | LK_RETRY);
2349 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
2350 	vput(*vp);
2351 	if (error) {
2352 		counter_u64_add(numfullpathfail2, 1);
2353 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
2354 		return (error);
2355 	}
2356 
2357 	*vp = dvp;
2358 	if (VN_IS_DOOMED(dvp)) {
2359 		/* forced unmount */
2360 		vrele(dvp);
2361 		error = ENOENT;
2362 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2363 		return (error);
2364 	}
2365 	/*
2366 	 * *vp has its use count incremented still.
2367 	 */
2368 
2369 	return (0);
2370 }
2371 
2372 /*
2373  * Resolve a directory to a pathname.
2374  *
2375  * The name of the directory can always be found in the namecache or fetched
2376  * from the filesystem. There is also guaranteed to be only one parent, meaning
2377  * we can just follow vnodes up until we find the root.
2378  *
2379  * The vnode must be referenced.
2380  */
2381 static int
2382 vn_fullpath_dir(struct thread *td, struct vnode *vp, struct vnode *rdir,
2383     char *buf, char **retbuf, size_t *len, bool slash_prefixed, size_t addend)
2384 {
2385 #ifdef KDTRACE_HOOKS
2386 	struct vnode *startvp = vp;
2387 #endif
2388 	struct vnode *vp1;
2389 	size_t buflen;
2390 	int error;
2391 
2392 	VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp);
2393 	VNPASS(vp->v_usecount > 0, vp);
2394 
2395 	buflen = *len;
2396 
2397 	if (!slash_prefixed) {
2398 		MPASS(*len >= 2);
2399 		buflen--;
2400 		buf[buflen] = '\0';
2401 	}
2402 
2403 	error = 0;
2404 
2405 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
2406 	counter_u64_add(numfullpathcalls, 1);
2407 	while (vp != rdir && vp != rootvnode) {
2408 		/*
2409 		 * The vp vnode must be already fully constructed,
2410 		 * since it is either found in namecache or obtained
2411 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
2412 		 * without obtaining the vnode lock.
2413 		 */
2414 		if ((vp->v_vflag & VV_ROOT) != 0) {
2415 			vn_lock(vp, LK_RETRY | LK_SHARED);
2416 
2417 			/*
2418 			 * With the vnode locked, check for races with
2419 			 * unmount, forced or not.  Note that we
2420 			 * already verified that vp is not equal to
2421 			 * the root vnode, which means that
2422 			 * mnt_vnodecovered can be NULL only for the
2423 			 * case of unmount.
2424 			 */
2425 			if (VN_IS_DOOMED(vp) ||
2426 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
2427 			    vp1->v_mountedhere != vp->v_mount) {
2428 				vput(vp);
2429 				error = ENOENT;
2430 				SDT_PROBE3(vfs, namecache, fullpath, return,
2431 				    error, vp, NULL);
2432 				break;
2433 			}
2434 
2435 			vref(vp1);
2436 			vput(vp);
2437 			vp = vp1;
2438 			continue;
2439 		}
2440 		if (vp->v_type != VDIR) {
2441 			vrele(vp);
2442 			counter_u64_add(numfullpathfail1, 1);
2443 			error = ENOTDIR;
2444 			SDT_PROBE3(vfs, namecache, fullpath, return,
2445 			    error, vp, NULL);
2446 			break;
2447 		}
2448 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2449 		if (error)
2450 			break;
2451 		if (buflen == 0) {
2452 			vrele(vp);
2453 			error = ENOMEM;
2454 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2455 			    startvp, NULL);
2456 			break;
2457 		}
2458 		buf[--buflen] = '/';
2459 		slash_prefixed = true;
2460 	}
2461 	if (error)
2462 		return (error);
2463 	if (!slash_prefixed) {
2464 		if (buflen == 0) {
2465 			vrele(vp);
2466 			counter_u64_add(numfullpathfail4, 1);
2467 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
2468 			    startvp, NULL);
2469 			return (ENOMEM);
2470 		}
2471 		buf[--buflen] = '/';
2472 	}
2473 	counter_u64_add(numfullpathfound, 1);
2474 	vrele(vp);
2475 
2476 	*retbuf = buf + buflen;
2477 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, *retbuf);
2478 	*len -= buflen;
2479 	*len += addend;
2480 	return (0);
2481 }
2482 
2483 /*
2484  * Resolve an arbitrary vnode to a pathname.
2485  *
2486  * Note 2 caveats:
2487  * - hardlinks are not tracked, thus if the vnode is not a directory this can
2488  *   resolve to a different path than the one used to find it
2489  * - namecache is not mandatory, meaning names are not guaranteed to be added
2490  *   (in which case resolving fails)
2491  */
2492 static int
2493 vn_fullpath_any(struct thread *td, struct vnode *vp, struct vnode *rdir,
2494     char *buf, char **retbuf, size_t *buflen)
2495 {
2496 	size_t orig_buflen;
2497 	bool slash_prefixed;
2498 	int error;
2499 
2500 	if (*buflen < 2)
2501 		return (EINVAL);
2502 
2503 	orig_buflen = *buflen;
2504 
2505 	vref(vp);
2506 	slash_prefixed = false;
2507 	if (vp->v_type != VDIR) {
2508 		*buflen -= 1;
2509 		buf[*buflen] = '\0';
2510 		error = vn_vptocnp(&vp, td->td_ucred, buf, buflen);
2511 		if (error)
2512 			return (error);
2513 		if (*buflen == 0) {
2514 			vrele(vp);
2515 			return (ENOMEM);
2516 		}
2517 		*buflen -= 1;
2518 		buf[*buflen] = '/';
2519 		slash_prefixed = true;
2520 	}
2521 
2522 	return (vn_fullpath_dir(td, vp, rdir, buf, retbuf, buflen, slash_prefixed,
2523 	    orig_buflen - *buflen));
2524 }
2525 
2526 /*
2527  * Resolve an arbitrary vnode to a pathname (taking care of hardlinks).
2528  *
2529  * Since the namecache does not track handlings, the caller is expected to first
2530  * look up the target vnode with SAVENAME | WANTPARENT flags passed to namei.
2531  *
2532  * Then we have 2 cases:
2533  * - if the found vnode is a directory, the path can be constructed just by
2534  *   fullowing names up the chain
2535  * - otherwise we populate the buffer with the saved name and start resolving
2536  *   from the parent
2537  */
2538 static int
2539 vn_fullpath_hardlink(struct thread *td, struct nameidata *ndp, char **retbuf,
2540     char **freebuf, size_t *buflen)
2541 {
2542 	char *buf, *tmpbuf;
2543 	struct filedesc *fdp;
2544 	struct vnode *rdir;
2545 	struct componentname *cnp;
2546 	struct vnode *vp;
2547 	size_t addend;
2548 	int error;
2549 	bool slash_prefixed;
2550 
2551 	if (*buflen < 2)
2552 		return (EINVAL);
2553 	if (*buflen > MAXPATHLEN)
2554 		*buflen = MAXPATHLEN;
2555 
2556 	slash_prefixed = false;
2557 
2558 	buf = malloc(*buflen, M_TEMP, M_WAITOK);
2559 	fdp = td->td_proc->p_fd;
2560 	FILEDESC_SLOCK(fdp);
2561 	rdir = fdp->fd_rdir;
2562 	vrefact(rdir);
2563 	FILEDESC_SUNLOCK(fdp);
2564 
2565 	addend = 0;
2566 	vp = ndp->ni_vp;
2567 	if (vp->v_type != VDIR) {
2568 		cnp = &ndp->ni_cnd;
2569 		addend = cnp->cn_namelen + 2;
2570 		if (*buflen < addend) {
2571 			error = ENOMEM;
2572 			goto out_bad;
2573 		}
2574 		*buflen -= addend;
2575 		tmpbuf = buf + *buflen;
2576 		tmpbuf[0] = '/';
2577 		memcpy(&tmpbuf[1], cnp->cn_nameptr, cnp->cn_namelen);
2578 		tmpbuf[addend - 1] = '\0';
2579 		slash_prefixed = true;
2580 		vp = ndp->ni_dvp;
2581 	}
2582 
2583 	vref(vp);
2584 	error = vn_fullpath_dir(td, vp, rdir, buf, retbuf, buflen, slash_prefixed, addend);
2585 	if (error != 0)
2586 		goto out_bad;
2587 
2588 	vrele(rdir);
2589 	*freebuf = buf;
2590 
2591 	return (0);
2592 out_bad:
2593 	vrele(rdir);
2594 	free(buf, M_TEMP);
2595 	return (error);
2596 }
2597 
2598 struct vnode *
2599 vn_dir_dd_ino(struct vnode *vp)
2600 {
2601 	struct namecache *ncp;
2602 	struct vnode *ddvp;
2603 	struct mtx *vlp;
2604 	enum vgetstate vs;
2605 
2606 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
2607 	vlp = VP2VNODELOCK(vp);
2608 	mtx_lock(vlp);
2609 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
2610 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
2611 			continue;
2612 		ddvp = ncp->nc_dvp;
2613 		vs = vget_prep(ddvp);
2614 		mtx_unlock(vlp);
2615 		if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
2616 			return (NULL);
2617 		return (ddvp);
2618 	}
2619 	mtx_unlock(vlp);
2620 	return (NULL);
2621 }
2622 
2623 int
2624 vn_commname(struct vnode *vp, char *buf, u_int buflen)
2625 {
2626 	struct namecache *ncp;
2627 	struct mtx *vlp;
2628 	int l;
2629 
2630 	vlp = VP2VNODELOCK(vp);
2631 	mtx_lock(vlp);
2632 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
2633 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2634 			break;
2635 	if (ncp == NULL) {
2636 		mtx_unlock(vlp);
2637 		return (ENOENT);
2638 	}
2639 	l = min(ncp->nc_nlen, buflen - 1);
2640 	memcpy(buf, ncp->nc_name, l);
2641 	mtx_unlock(vlp);
2642 	buf[l] = '\0';
2643 	return (0);
2644 }
2645 
2646 /*
2647  * This function updates path string to vnode's full global path
2648  * and checks the size of the new path string against the pathlen argument.
2649  *
2650  * Requires a locked, referenced vnode.
2651  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
2652  *
2653  * If vp is a directory, the call to vn_fullpath_global() always succeeds
2654  * because it falls back to the ".." lookup if the namecache lookup fails.
2655  */
2656 int
2657 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
2658     u_int pathlen)
2659 {
2660 	struct nameidata nd;
2661 	struct vnode *vp1;
2662 	char *rpath, *fbuf;
2663 	int error;
2664 
2665 	ASSERT_VOP_ELOCKED(vp, __func__);
2666 
2667 	/* Construct global filesystem path from vp. */
2668 	VOP_UNLOCK(vp);
2669 	error = vn_fullpath_global(td, vp, &rpath, &fbuf);
2670 
2671 	if (error != 0) {
2672 		vrele(vp);
2673 		return (error);
2674 	}
2675 
2676 	if (strlen(rpath) >= pathlen) {
2677 		vrele(vp);
2678 		error = ENAMETOOLONG;
2679 		goto out;
2680 	}
2681 
2682 	/*
2683 	 * Re-lookup the vnode by path to detect a possible rename.
2684 	 * As a side effect, the vnode is relocked.
2685 	 * If vnode was renamed, return ENOENT.
2686 	 */
2687 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
2688 	    UIO_SYSSPACE, path, td);
2689 	error = namei(&nd);
2690 	if (error != 0) {
2691 		vrele(vp);
2692 		goto out;
2693 	}
2694 	NDFREE(&nd, NDF_ONLY_PNBUF);
2695 	vp1 = nd.ni_vp;
2696 	vrele(vp);
2697 	if (vp1 == vp)
2698 		strcpy(path, rpath);
2699 	else {
2700 		vput(vp1);
2701 		error = ENOENT;
2702 	}
2703 
2704 out:
2705 	free(fbuf, M_TEMP);
2706 	return (error);
2707 }
2708 
2709 #ifdef DDB
2710 static void
2711 db_print_vpath(struct vnode *vp)
2712 {
2713 
2714 	while (vp != NULL) {
2715 		db_printf("%p: ", vp);
2716 		if (vp == rootvnode) {
2717 			db_printf("/");
2718 			vp = NULL;
2719 		} else {
2720 			if (vp->v_vflag & VV_ROOT) {
2721 				db_printf("<mount point>");
2722 				vp = vp->v_mount->mnt_vnodecovered;
2723 			} else {
2724 				struct namecache *ncp;
2725 				char *ncn;
2726 				int i;
2727 
2728 				ncp = TAILQ_FIRST(&vp->v_cache_dst);
2729 				if (ncp != NULL) {
2730 					ncn = ncp->nc_name;
2731 					for (i = 0; i < ncp->nc_nlen; i++)
2732 						db_printf("%c", *ncn++);
2733 					vp = ncp->nc_dvp;
2734 				} else {
2735 					vp = NULL;
2736 				}
2737 			}
2738 		}
2739 		db_printf("\n");
2740 	}
2741 
2742 	return;
2743 }
2744 
2745 DB_SHOW_COMMAND(vpath, db_show_vpath)
2746 {
2747 	struct vnode *vp;
2748 
2749 	if (!have_addr) {
2750 		db_printf("usage: show vpath <struct vnode *>\n");
2751 		return;
2752 	}
2753 
2754 	vp = (struct vnode *)addr;
2755 	db_print_vpath(vp);
2756 }
2757 
2758 #endif
2759