xref: /freebsd/sys/kern/vfs_cache.c (revision ce6a89e27cd190313be39bb479880aeda4778436)
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 | CTLFLAG_MPSAFE, 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_DEFINE_EARLY(name);				\
352 	SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, \
353 	    descr);
354 STATNODE_ULONG(numneg, "Number of negative cache entries");
355 STATNODE_ULONG(numcache, "Number of cache entries");
356 STATNODE_COUNTER(numcachehv, "Number of namecache entries with vnodes held");
357 STATNODE_COUNTER(numcalls, "Number of cache lookups");
358 STATNODE_COUNTER(dothits, "Number of '.' hits");
359 STATNODE_COUNTER(dotdothits, "Number of '..' hits");
360 STATNODE_COUNTER(numchecks, "Number of checks in lookup");
361 STATNODE_COUNTER(nummiss, "Number of cache misses");
362 STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache");
363 STATNODE_COUNTER(numposzaps,
364     "Number of cache hits (positive) we do not want to cache");
365 STATNODE_COUNTER(numposhits, "Number of cache hits (positive)");
366 STATNODE_COUNTER(numnegzaps,
367     "Number of cache hits (negative) we do not want to cache");
368 STATNODE_COUNTER(numneghits, "Number of cache hits (negative)");
369 /* These count for vn_getcwd(), too. */
370 STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls");
371 STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
372 STATNODE_COUNTER(numfullpathfail2,
373     "Number of fullpath search errors (VOP_VPTOCNP failures)");
374 STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
375 STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls");
376 STATNODE_COUNTER(zap_and_exit_bucket_relock_success,
377     "Number of successful removals after relocking");
378 static long zap_and_exit_bucket_fail; STATNODE_ULONG(zap_and_exit_bucket_fail,
379     "Number of times zap_and_exit failed to lock");
380 static long zap_and_exit_bucket_fail2; STATNODE_ULONG(zap_and_exit_bucket_fail2,
381     "Number of times zap_and_exit failed to lock");
382 static long cache_lock_vnodes_cel_3_failures;
383 STATNODE_ULONG(cache_lock_vnodes_cel_3_failures,
384     "Number of times 3-way vnode locking failed");
385 STATNODE_ULONG(numhotneg, "Number of hot negative entries");
386 STATNODE_COUNTER(numneg_evicted,
387     "Number of negative entries evicted when adding a new entry");
388 STATNODE_COUNTER(shrinking_skipped,
389     "Number of times shrinking was already in progress");
390 
391 static void cache_zap_locked(struct namecache *ncp, bool neg_locked);
392 static int vn_fullpath_hardlink(struct thread *td, struct nameidata *ndp, char **retbuf,
393     char **freebuf, size_t *buflen);
394 static int vn_fullpath_any(struct thread *td, struct vnode *vp, struct vnode *rdir,
395     char *buf, char **retbuf, size_t *buflen);
396 static int vn_fullpath_dir(struct thread *td, struct vnode *vp, struct vnode *rdir,
397     char *buf, char **retbuf, size_t *len, bool slash_prefixed, size_t addend);
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,
586     CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
587     "hash table stats");
588 
589 static int
590 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
591 {
592 	struct nchashhead *ncpp;
593 	struct namecache *ncp;
594 	int i, error, n_nchash, *cntbuf;
595 
596 retry:
597 	n_nchash = nchash + 1;	/* nchash is max index, not count */
598 	if (req->oldptr == NULL)
599 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
600 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
601 	cache_lock_all_buckets();
602 	if (n_nchash != nchash + 1) {
603 		cache_unlock_all_buckets();
604 		free(cntbuf, M_TEMP);
605 		goto retry;
606 	}
607 	/* Scan hash tables counting entries */
608 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
609 		LIST_FOREACH(ncp, ncpp, nc_hash)
610 			cntbuf[i]++;
611 	cache_unlock_all_buckets();
612 	for (error = 0, i = 0; i < n_nchash; i++)
613 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
614 			break;
615 	free(cntbuf, M_TEMP);
616 	return (error);
617 }
618 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
619     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
620     "nchash chain lengths");
621 
622 static int
623 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
624 {
625 	int error;
626 	struct nchashhead *ncpp;
627 	struct namecache *ncp;
628 	int n_nchash;
629 	int count, maxlength, used, pct;
630 
631 	if (!req->oldptr)
632 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
633 
634 	cache_lock_all_buckets();
635 	n_nchash = nchash + 1;	/* nchash is max index, not count */
636 	used = 0;
637 	maxlength = 0;
638 
639 	/* Scan hash tables for applicable entries */
640 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
641 		count = 0;
642 		LIST_FOREACH(ncp, ncpp, nc_hash) {
643 			count++;
644 		}
645 		if (count)
646 			used++;
647 		if (maxlength < count)
648 			maxlength = count;
649 	}
650 	n_nchash = nchash + 1;
651 	cache_unlock_all_buckets();
652 	pct = (used * 100) / (n_nchash / 100);
653 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
654 	if (error)
655 		return (error);
656 	error = SYSCTL_OUT(req, &used, sizeof(used));
657 	if (error)
658 		return (error);
659 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
660 	if (error)
661 		return (error);
662 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
663 	if (error)
664 		return (error);
665 	return (0);
666 }
667 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
668     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
669     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
670 #endif
671 
672 /*
673  * Negative entries management
674  *
675  * A variation of LRU scheme is used. New entries are hashed into one of
676  * numneglists cold lists. Entries get promoted to the hot list on first hit.
677  *
678  * The shrinker will demote hot list head and evict from the cold list in a
679  * round-robin manner.
680  */
681 static void
682 cache_negative_hit(struct namecache *ncp)
683 {
684 	struct neglist *neglist;
685 
686 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
687 	if (ncp->nc_flag & NCF_HOTNEGATIVE)
688 		return;
689 	neglist = NCP2NEGLIST(ncp);
690 	mtx_lock(&ncneg_hot.nl_lock);
691 	mtx_lock(&neglist->nl_lock);
692 	if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
693 		numhotneg++;
694 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
695 		TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
696 		ncp->nc_flag |= NCF_HOTNEGATIVE;
697 	}
698 	mtx_unlock(&neglist->nl_lock);
699 	mtx_unlock(&ncneg_hot.nl_lock);
700 }
701 
702 static void
703 cache_negative_insert(struct namecache *ncp, bool neg_locked)
704 {
705 	struct neglist *neglist;
706 
707 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
708 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
709 	neglist = NCP2NEGLIST(ncp);
710 	if (!neg_locked) {
711 		mtx_lock(&neglist->nl_lock);
712 	} else {
713 		mtx_assert(&neglist->nl_lock, MA_OWNED);
714 	}
715 	TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
716 	if (!neg_locked)
717 		mtx_unlock(&neglist->nl_lock);
718 	atomic_add_rel_long(&numneg, 1);
719 }
720 
721 static void
722 cache_negative_remove(struct namecache *ncp, bool neg_locked)
723 {
724 	struct neglist *neglist;
725 	bool hot_locked = false;
726 	bool list_locked = false;
727 
728 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
729 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
730 	neglist = NCP2NEGLIST(ncp);
731 	if (!neg_locked) {
732 		if (ncp->nc_flag & NCF_HOTNEGATIVE) {
733 			hot_locked = true;
734 			mtx_lock(&ncneg_hot.nl_lock);
735 			if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
736 				list_locked = true;
737 				mtx_lock(&neglist->nl_lock);
738 			}
739 		} else {
740 			list_locked = true;
741 			mtx_lock(&neglist->nl_lock);
742 		}
743 	}
744 	if (ncp->nc_flag & NCF_HOTNEGATIVE) {
745 		mtx_assert(&ncneg_hot.nl_lock, MA_OWNED);
746 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
747 		numhotneg--;
748 	} else {
749 		mtx_assert(&neglist->nl_lock, MA_OWNED);
750 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
751 	}
752 	if (list_locked)
753 		mtx_unlock(&neglist->nl_lock);
754 	if (hot_locked)
755 		mtx_unlock(&ncneg_hot.nl_lock);
756 	atomic_subtract_rel_long(&numneg, 1);
757 }
758 
759 static void
760 cache_negative_shrink_select(int start, struct namecache **ncpp,
761     struct neglist **neglistpp)
762 {
763 	struct neglist *neglist;
764 	struct namecache *ncp;
765 	int i;
766 
767 	*ncpp = ncp = NULL;
768 	neglist = NULL;
769 
770 	for (i = start; i < numneglists; i++) {
771 		neglist = &neglists[i];
772 		if (TAILQ_FIRST(&neglist->nl_list) == NULL)
773 			continue;
774 		mtx_lock(&neglist->nl_lock);
775 		ncp = TAILQ_FIRST(&neglist->nl_list);
776 		if (ncp != NULL)
777 			break;
778 		mtx_unlock(&neglist->nl_lock);
779 	}
780 
781 	*neglistpp = neglist;
782 	*ncpp = ncp;
783 }
784 
785 static void
786 cache_negative_zap_one(void)
787 {
788 	struct namecache *ncp, *ncp2;
789 	struct neglist *neglist;
790 	struct mtx *dvlp;
791 	struct rwlock *blp;
792 
793 	if (mtx_owner(&ncneg_shrink_lock) != NULL ||
794 	    !mtx_trylock(&ncneg_shrink_lock)) {
795 		counter_u64_add(shrinking_skipped, 1);
796 		return;
797 	}
798 
799 	mtx_lock(&ncneg_hot.nl_lock);
800 	ncp = TAILQ_FIRST(&ncneg_hot.nl_list);
801 	if (ncp != NULL) {
802 		neglist = NCP2NEGLIST(ncp);
803 		mtx_lock(&neglist->nl_lock);
804 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
805 		TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
806 		ncp->nc_flag &= ~NCF_HOTNEGATIVE;
807 		numhotneg--;
808 		mtx_unlock(&neglist->nl_lock);
809 	}
810 	mtx_unlock(&ncneg_hot.nl_lock);
811 
812 	cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
813 	shrink_list_turn++;
814 	if (shrink_list_turn == numneglists)
815 		shrink_list_turn = 0;
816 	if (ncp == NULL && shrink_list_turn == 0)
817 		cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
818 	mtx_unlock(&ncneg_shrink_lock);
819 	if (ncp == NULL)
820 		return;
821 
822 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
823 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
824 	blp = NCP2BUCKETLOCK(ncp);
825 	mtx_unlock(&neglist->nl_lock);
826 	mtx_lock(dvlp);
827 	rw_wlock(blp);
828 	mtx_lock(&neglist->nl_lock);
829 	ncp2 = TAILQ_FIRST(&neglist->nl_list);
830 	if (ncp != ncp2 || dvlp != VP2VNODELOCK(ncp2->nc_dvp) ||
831 	    blp != NCP2BUCKETLOCK(ncp2) || !(ncp2->nc_flag & NCF_NEGATIVE)) {
832 		ncp = NULL;
833 	} else {
834 		SDT_PROBE2(vfs, namecache, shrink_negative, done, ncp->nc_dvp,
835 		    ncp->nc_name);
836 
837 		cache_zap_locked(ncp, true);
838 		counter_u64_add(numneg_evicted, 1);
839 	}
840 	mtx_unlock(&neglist->nl_lock);
841 	rw_wunlock(blp);
842 	mtx_unlock(dvlp);
843 	cache_free(ncp);
844 }
845 
846 /*
847  * cache_zap_locked():
848  *
849  *   Removes a namecache entry from cache, whether it contains an actual
850  *   pointer to a vnode or if it is just a negative cache entry.
851  */
852 static void
853 cache_zap_locked(struct namecache *ncp, bool neg_locked)
854 {
855 
856 	if (!(ncp->nc_flag & NCF_NEGATIVE))
857 		cache_assert_vnode_locked(ncp->nc_vp);
858 	cache_assert_vnode_locked(ncp->nc_dvp);
859 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
860 
861 	CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp,
862 	    (ncp->nc_flag & NCF_NEGATIVE) ? NULL : ncp->nc_vp);
863 	LIST_REMOVE(ncp, nc_hash);
864 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
865 		SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
866 		    ncp->nc_name, ncp->nc_vp);
867 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
868 		if (ncp == ncp->nc_vp->v_cache_dd)
869 			ncp->nc_vp->v_cache_dd = NULL;
870 	} else {
871 		SDT_PROBE2(vfs, namecache, zap_negative, done, ncp->nc_dvp,
872 		    ncp->nc_name);
873 		cache_negative_remove(ncp, neg_locked);
874 	}
875 	if (ncp->nc_flag & NCF_ISDOTDOT) {
876 		if (ncp == ncp->nc_dvp->v_cache_dd)
877 			ncp->nc_dvp->v_cache_dd = NULL;
878 	} else {
879 		LIST_REMOVE(ncp, nc_src);
880 		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
881 			ncp->nc_flag |= NCF_DVDROP;
882 			counter_u64_add(numcachehv, -1);
883 		}
884 	}
885 	atomic_subtract_rel_long(&numcache, 1);
886 }
887 
888 static void
889 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
890 {
891 	struct rwlock *blp;
892 
893 	MPASS(ncp->nc_dvp == vp);
894 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
895 	cache_assert_vnode_locked(vp);
896 
897 	blp = NCP2BUCKETLOCK(ncp);
898 	rw_wlock(blp);
899 	cache_zap_locked(ncp, false);
900 	rw_wunlock(blp);
901 }
902 
903 static bool
904 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
905     struct mtx **vlpp)
906 {
907 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
908 	struct rwlock *blp;
909 
910 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
911 	cache_assert_vnode_locked(vp);
912 
913 	if (ncp->nc_flag & NCF_NEGATIVE) {
914 		if (*vlpp != NULL) {
915 			mtx_unlock(*vlpp);
916 			*vlpp = NULL;
917 		}
918 		cache_zap_negative_locked_vnode_kl(ncp, vp);
919 		return (true);
920 	}
921 
922 	pvlp = VP2VNODELOCK(vp);
923 	blp = NCP2BUCKETLOCK(ncp);
924 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
925 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
926 
927 	if (*vlpp == vlp1 || *vlpp == vlp2) {
928 		to_unlock = *vlpp;
929 		*vlpp = NULL;
930 	} else {
931 		if (*vlpp != NULL) {
932 			mtx_unlock(*vlpp);
933 			*vlpp = NULL;
934 		}
935 		cache_sort_vnodes(&vlp1, &vlp2);
936 		if (vlp1 == pvlp) {
937 			mtx_lock(vlp2);
938 			to_unlock = vlp2;
939 		} else {
940 			if (!mtx_trylock(vlp1))
941 				goto out_relock;
942 			to_unlock = vlp1;
943 		}
944 	}
945 	rw_wlock(blp);
946 	cache_zap_locked(ncp, false);
947 	rw_wunlock(blp);
948 	if (to_unlock != NULL)
949 		mtx_unlock(to_unlock);
950 	return (true);
951 
952 out_relock:
953 	mtx_unlock(vlp2);
954 	mtx_lock(vlp1);
955 	mtx_lock(vlp2);
956 	MPASS(*vlpp == NULL);
957 	*vlpp = vlp1;
958 	return (false);
959 }
960 
961 static int __noinline
962 cache_zap_locked_vnode(struct namecache *ncp, struct vnode *vp)
963 {
964 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
965 	struct rwlock *blp;
966 	int error = 0;
967 
968 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
969 	cache_assert_vnode_locked(vp);
970 
971 	pvlp = VP2VNODELOCK(vp);
972 	if (ncp->nc_flag & NCF_NEGATIVE) {
973 		cache_zap_negative_locked_vnode_kl(ncp, vp);
974 		goto out;
975 	}
976 
977 	blp = NCP2BUCKETLOCK(ncp);
978 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
979 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
980 	cache_sort_vnodes(&vlp1, &vlp2);
981 	if (vlp1 == pvlp) {
982 		mtx_lock(vlp2);
983 		to_unlock = vlp2;
984 	} else {
985 		if (!mtx_trylock(vlp1)) {
986 			error = EAGAIN;
987 			goto out;
988 		}
989 		to_unlock = vlp1;
990 	}
991 	rw_wlock(blp);
992 	cache_zap_locked(ncp, false);
993 	rw_wunlock(blp);
994 	mtx_unlock(to_unlock);
995 out:
996 	mtx_unlock(pvlp);
997 	return (error);
998 }
999 
1000 /*
1001  * If trylocking failed we can get here. We know enough to take all needed locks
1002  * in the right order and re-lookup the entry.
1003  */
1004 static int
1005 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1006     struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1007     struct rwlock *blp)
1008 {
1009 	struct namecache *rncp;
1010 
1011 	cache_assert_bucket_locked(ncp, RA_UNLOCKED);
1012 
1013 	cache_sort_vnodes(&dvlp, &vlp);
1014 	cache_lock_vnodes(dvlp, vlp);
1015 	rw_wlock(blp);
1016 	LIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1017 		if (rncp == ncp && rncp->nc_dvp == dvp &&
1018 		    rncp->nc_nlen == cnp->cn_namelen &&
1019 		    !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen))
1020 			break;
1021 	}
1022 	if (rncp != NULL) {
1023 		cache_zap_locked(rncp, false);
1024 		rw_wunlock(blp);
1025 		cache_unlock_vnodes(dvlp, vlp);
1026 		counter_u64_add(zap_and_exit_bucket_relock_success, 1);
1027 		return (0);
1028 	}
1029 
1030 	rw_wunlock(blp);
1031 	cache_unlock_vnodes(dvlp, vlp);
1032 	return (EAGAIN);
1033 }
1034 
1035 static int __noinline
1036 cache_zap_wlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1037     uint32_t hash, struct rwlock *blp)
1038 {
1039 	struct mtx *dvlp, *vlp;
1040 	struct vnode *dvp;
1041 
1042 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1043 
1044 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1045 	vlp = NULL;
1046 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1047 		vlp = VP2VNODELOCK(ncp->nc_vp);
1048 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1049 		cache_zap_locked(ncp, false);
1050 		rw_wunlock(blp);
1051 		cache_unlock_vnodes(dvlp, vlp);
1052 		return (0);
1053 	}
1054 
1055 	dvp = ncp->nc_dvp;
1056 	rw_wunlock(blp);
1057 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1058 }
1059 
1060 static int __noinline
1061 cache_zap_rlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1062     uint32_t hash, struct rwlock *blp)
1063 {
1064 	struct mtx *dvlp, *vlp;
1065 	struct vnode *dvp;
1066 
1067 	cache_assert_bucket_locked(ncp, RA_RLOCKED);
1068 
1069 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1070 	vlp = NULL;
1071 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1072 		vlp = VP2VNODELOCK(ncp->nc_vp);
1073 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1074 		rw_runlock(blp);
1075 		rw_wlock(blp);
1076 		cache_zap_locked(ncp, false);
1077 		rw_wunlock(blp);
1078 		cache_unlock_vnodes(dvlp, vlp);
1079 		return (0);
1080 	}
1081 
1082 	dvp = ncp->nc_dvp;
1083 	rw_runlock(blp);
1084 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1085 }
1086 
1087 static int
1088 cache_zap_wlocked_bucket_kl(struct namecache *ncp, struct rwlock *blp,
1089     struct mtx **vlpp1, struct mtx **vlpp2)
1090 {
1091 	struct mtx *dvlp, *vlp;
1092 
1093 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1094 
1095 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1096 	vlp = NULL;
1097 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1098 		vlp = VP2VNODELOCK(ncp->nc_vp);
1099 	cache_sort_vnodes(&dvlp, &vlp);
1100 
1101 	if (*vlpp1 == dvlp && *vlpp2 == vlp) {
1102 		cache_zap_locked(ncp, false);
1103 		cache_unlock_vnodes(dvlp, vlp);
1104 		*vlpp1 = NULL;
1105 		*vlpp2 = NULL;
1106 		return (0);
1107 	}
1108 
1109 	if (*vlpp1 != NULL)
1110 		mtx_unlock(*vlpp1);
1111 	if (*vlpp2 != NULL)
1112 		mtx_unlock(*vlpp2);
1113 	*vlpp1 = NULL;
1114 	*vlpp2 = NULL;
1115 
1116 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1117 		cache_zap_locked(ncp, false);
1118 		cache_unlock_vnodes(dvlp, vlp);
1119 		return (0);
1120 	}
1121 
1122 	rw_wunlock(blp);
1123 	*vlpp1 = dvlp;
1124 	*vlpp2 = vlp;
1125 	if (*vlpp1 != NULL)
1126 		mtx_lock(*vlpp1);
1127 	mtx_lock(*vlpp2);
1128 	rw_wlock(blp);
1129 	return (EAGAIN);
1130 }
1131 
1132 static void
1133 cache_lookup_unlock(struct rwlock *blp, struct mtx *vlp)
1134 {
1135 
1136 	if (blp != NULL) {
1137 		rw_runlock(blp);
1138 	} else {
1139 		mtx_unlock(vlp);
1140 	}
1141 }
1142 
1143 static int __noinline
1144 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1145     struct timespec *tsp, int *ticksp)
1146 {
1147 	int ltype;
1148 
1149 	*vpp = dvp;
1150 	CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .",
1151 			dvp, cnp->cn_nameptr);
1152 	counter_u64_add(dothits, 1);
1153 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1154 	if (tsp != NULL)
1155 		timespecclear(tsp);
1156 	if (ticksp != NULL)
1157 		*ticksp = ticks;
1158 	vrefact(*vpp);
1159 	/*
1160 	 * When we lookup "." we still can be asked to lock it
1161 	 * differently...
1162 	 */
1163 	ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1164 	if (ltype != VOP_ISLOCKED(*vpp)) {
1165 		if (ltype == LK_EXCLUSIVE) {
1166 			vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1167 			if (VN_IS_DOOMED((*vpp))) {
1168 				/* forced unmount */
1169 				vrele(*vpp);
1170 				*vpp = NULL;
1171 				return (ENOENT);
1172 			}
1173 		} else
1174 			vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1175 	}
1176 	return (-1);
1177 }
1178 
1179 static __noinline int
1180 cache_lookup_nomakeentry(struct vnode *dvp, struct vnode **vpp,
1181     struct componentname *cnp, struct timespec *tsp, int *ticksp)
1182 {
1183 	struct namecache *ncp;
1184 	struct rwlock *blp;
1185 	struct mtx *dvlp, *dvlp2;
1186 	uint32_t hash;
1187 	int error;
1188 
1189 	if (cnp->cn_namelen == 2 &&
1190 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1191 		counter_u64_add(dotdothits, 1);
1192 		dvlp = VP2VNODELOCK(dvp);
1193 		dvlp2 = NULL;
1194 		mtx_lock(dvlp);
1195 retry_dotdot:
1196 		ncp = dvp->v_cache_dd;
1197 		if (ncp == NULL) {
1198 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1199 			    "..", NULL);
1200 			mtx_unlock(dvlp);
1201 			if (dvlp2 != NULL)
1202 				mtx_unlock(dvlp2);
1203 			return (0);
1204 		}
1205 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1206 			if (ncp->nc_dvp != dvp)
1207 				panic("dvp %p v_cache_dd %p\n", dvp, ncp);
1208 			if (!cache_zap_locked_vnode_kl2(ncp,
1209 			    dvp, &dvlp2))
1210 				goto retry_dotdot;
1211 			MPASS(dvp->v_cache_dd == NULL);
1212 			mtx_unlock(dvlp);
1213 			if (dvlp2 != NULL)
1214 				mtx_unlock(dvlp2);
1215 			cache_free(ncp);
1216 		} else {
1217 			dvp->v_cache_dd = NULL;
1218 			mtx_unlock(dvlp);
1219 			if (dvlp2 != NULL)
1220 				mtx_unlock(dvlp2);
1221 		}
1222 		return (0);
1223 	}
1224 
1225 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1226 	blp = HASH2BUCKETLOCK(hash);
1227 retry:
1228 	if (LIST_EMPTY(NCHHASH(hash)))
1229 		goto out_no_entry;
1230 
1231 	rw_wlock(blp);
1232 
1233 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1234 		counter_u64_add(numchecks, 1);
1235 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1236 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1237 			break;
1238 	}
1239 
1240 	/* We failed to find an entry */
1241 	if (ncp == NULL) {
1242 		rw_wunlock(blp);
1243 		goto out_no_entry;
1244 	}
1245 
1246 	error = cache_zap_wlocked_bucket(ncp, cnp, hash, blp);
1247 	if (__predict_false(error != 0)) {
1248 		zap_and_exit_bucket_fail++;
1249 		cache_maybe_yield();
1250 		goto retry;
1251 	}
1252 	counter_u64_add(numposzaps, 1);
1253 	cache_free(ncp);
1254 	return (0);
1255 out_no_entry:
1256 	SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr, NULL);
1257 	counter_u64_add(nummisszap, 1);
1258 	return (0);
1259 }
1260 
1261 /**
1262  * Lookup a name in the name cache
1263  *
1264  * # Arguments
1265  *
1266  * - dvp:	Parent directory in which to search.
1267  * - vpp:	Return argument.  Will contain desired vnode on cache hit.
1268  * - cnp:	Parameters of the name search.  The most interesting bits of
1269  *   		the cn_flags field have the following meanings:
1270  *   	- MAKEENTRY:	If clear, free an entry from the cache rather than look
1271  *   			it up.
1272  *   	- ISDOTDOT:	Must be set if and only if cn_nameptr == ".."
1273  * - tsp:	Return storage for cache timestamp.  On a successful (positive
1274  *   		or negative) lookup, tsp will be filled with any timespec that
1275  *   		was stored when this cache entry was created.  However, it will
1276  *   		be clear for "." entries.
1277  * - ticks:	Return storage for alternate cache timestamp.  On a successful
1278  *   		(positive or negative) lookup, it will contain the ticks value
1279  *   		that was current when the cache entry was created, unless cnp
1280  *   		was ".".
1281  *
1282  * # Returns
1283  *
1284  * - -1:	A positive cache hit.  vpp will contain the desired vnode.
1285  * - ENOENT:	A negative cache hit, or dvp was recycled out from under us due
1286  *		to a forced unmount.  vpp will not be modified.  If the entry
1287  *		is a whiteout, then the ISWHITEOUT flag will be set in
1288  *		cnp->cn_flags.
1289  * - 0:		A cache miss.  vpp will not be modified.
1290  *
1291  * # Locking
1292  *
1293  * On a cache hit, vpp will be returned locked and ref'd.  If we're looking up
1294  * .., dvp is unlocked.  If we're looking up . an extra ref is taken, but the
1295  * lock is not recursively acquired.
1296  */
1297 int
1298 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1299     struct timespec *tsp, int *ticksp)
1300 {
1301 	struct namecache_ts *ncp_ts;
1302 	struct namecache *ncp;
1303 	struct rwlock *blp;
1304 	struct mtx *dvlp;
1305 	uint32_t hash;
1306 	enum vgetstate vs;
1307 	int error, ltype;
1308 
1309 #ifdef DEBUG_CACHE
1310 	if (__predict_false(!doingcache)) {
1311 		cnp->cn_flags &= ~MAKEENTRY;
1312 		return (0);
1313 	}
1314 #endif
1315 
1316 	counter_u64_add(numcalls, 1);
1317 
1318 	if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.'))
1319 		return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1320 
1321 	if ((cnp->cn_flags & MAKEENTRY) == 0)
1322 		return (cache_lookup_nomakeentry(dvp, vpp, cnp, tsp, ticksp));
1323 
1324 retry:
1325 	blp = NULL;
1326 	dvlp = NULL;
1327 	error = 0;
1328 	if (cnp->cn_namelen == 2 &&
1329 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1330 		counter_u64_add(dotdothits, 1);
1331 		dvlp = VP2VNODELOCK(dvp);
1332 		mtx_lock(dvlp);
1333 		ncp = dvp->v_cache_dd;
1334 		if (ncp == NULL) {
1335 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1336 			    "..", NULL);
1337 			mtx_unlock(dvlp);
1338 			return (0);
1339 		}
1340 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1341 			if (ncp->nc_flag & NCF_NEGATIVE)
1342 				*vpp = NULL;
1343 			else
1344 				*vpp = ncp->nc_vp;
1345 		} else
1346 			*vpp = ncp->nc_dvp;
1347 		/* Return failure if negative entry was found. */
1348 		if (*vpp == NULL)
1349 			goto negative_success;
1350 		CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
1351 		    dvp, cnp->cn_nameptr, *vpp);
1352 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..",
1353 		    *vpp);
1354 		cache_out_ts(ncp, tsp, ticksp);
1355 		if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1356 		    NCF_DTS && tsp != NULL) {
1357 			ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1358 			*tsp = ncp_ts->nc_dotdottime;
1359 		}
1360 		goto success;
1361 	}
1362 
1363 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1364 	blp = HASH2BUCKETLOCK(hash);
1365 	rw_rlock(blp);
1366 
1367 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1368 		counter_u64_add(numchecks, 1);
1369 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1370 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1371 			break;
1372 	}
1373 
1374 	/* We failed to find an entry */
1375 	if (__predict_false(ncp == NULL)) {
1376 		rw_runlock(blp);
1377 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1378 		    NULL);
1379 		counter_u64_add(nummiss, 1);
1380 		return (0);
1381 	}
1382 
1383 	if (ncp->nc_flag & NCF_NEGATIVE)
1384 		goto negative_success;
1385 
1386 	/* We found a "positive" match, return the vnode */
1387 	counter_u64_add(numposhits, 1);
1388 	*vpp = ncp->nc_vp;
1389 	CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p",
1390 	    dvp, cnp->cn_nameptr, *vpp, ncp);
1391 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name,
1392 	    *vpp);
1393 	cache_out_ts(ncp, tsp, ticksp);
1394 success:
1395 	/*
1396 	 * On success we return a locked and ref'd vnode as per the lookup
1397 	 * protocol.
1398 	 */
1399 	MPASS(dvp != *vpp);
1400 	ltype = 0;	/* silence gcc warning */
1401 	if (cnp->cn_flags & ISDOTDOT) {
1402 		ltype = VOP_ISLOCKED(dvp);
1403 		VOP_UNLOCK(dvp);
1404 	}
1405 	vs = vget_prep(*vpp);
1406 	cache_lookup_unlock(blp, dvlp);
1407 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1408 	if (cnp->cn_flags & ISDOTDOT) {
1409 		vn_lock(dvp, ltype | LK_RETRY);
1410 		if (VN_IS_DOOMED(dvp)) {
1411 			if (error == 0)
1412 				vput(*vpp);
1413 			*vpp = NULL;
1414 			return (ENOENT);
1415 		}
1416 	}
1417 	if (error) {
1418 		*vpp = NULL;
1419 		goto retry;
1420 	}
1421 	if ((cnp->cn_flags & ISLASTCN) &&
1422 	    (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) {
1423 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
1424 	}
1425 	return (-1);
1426 
1427 negative_success:
1428 	/* We found a negative match, and want to create it, so purge */
1429 	if (cnp->cn_nameiop == CREATE) {
1430 		counter_u64_add(numnegzaps, 1);
1431 		goto zap_and_exit;
1432 	}
1433 
1434 	counter_u64_add(numneghits, 1);
1435 	cache_negative_hit(ncp);
1436 	if (ncp->nc_flag & NCF_WHITE)
1437 		cnp->cn_flags |= ISWHITEOUT;
1438 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp,
1439 	    ncp->nc_name);
1440 	cache_out_ts(ncp, tsp, ticksp);
1441 	cache_lookup_unlock(blp, dvlp);
1442 	return (ENOENT);
1443 
1444 zap_and_exit:
1445 	if (blp != NULL)
1446 		error = cache_zap_rlocked_bucket(ncp, cnp, hash, blp);
1447 	else
1448 		error = cache_zap_locked_vnode(ncp, dvp);
1449 	if (__predict_false(error != 0)) {
1450 		zap_and_exit_bucket_fail2++;
1451 		cache_maybe_yield();
1452 		goto retry;
1453 	}
1454 	cache_free(ncp);
1455 	return (0);
1456 }
1457 
1458 struct celockstate {
1459 	struct mtx *vlp[3];
1460 	struct rwlock *blp[2];
1461 };
1462 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1463 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1464 
1465 static inline void
1466 cache_celockstate_init(struct celockstate *cel)
1467 {
1468 
1469 	bzero(cel, sizeof(*cel));
1470 }
1471 
1472 static void
1473 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1474     struct vnode *dvp)
1475 {
1476 	struct mtx *vlp1, *vlp2;
1477 
1478 	MPASS(cel->vlp[0] == NULL);
1479 	MPASS(cel->vlp[1] == NULL);
1480 	MPASS(cel->vlp[2] == NULL);
1481 
1482 	MPASS(vp != NULL || dvp != NULL);
1483 
1484 	vlp1 = VP2VNODELOCK(vp);
1485 	vlp2 = VP2VNODELOCK(dvp);
1486 	cache_sort_vnodes(&vlp1, &vlp2);
1487 
1488 	if (vlp1 != NULL) {
1489 		mtx_lock(vlp1);
1490 		cel->vlp[0] = vlp1;
1491 	}
1492 	mtx_lock(vlp2);
1493 	cel->vlp[1] = vlp2;
1494 }
1495 
1496 static void
1497 cache_unlock_vnodes_cel(struct celockstate *cel)
1498 {
1499 
1500 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
1501 
1502 	if (cel->vlp[0] != NULL)
1503 		mtx_unlock(cel->vlp[0]);
1504 	if (cel->vlp[1] != NULL)
1505 		mtx_unlock(cel->vlp[1]);
1506 	if (cel->vlp[2] != NULL)
1507 		mtx_unlock(cel->vlp[2]);
1508 }
1509 
1510 static bool
1511 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
1512 {
1513 	struct mtx *vlp;
1514 	bool ret;
1515 
1516 	cache_assert_vlp_locked(cel->vlp[0]);
1517 	cache_assert_vlp_locked(cel->vlp[1]);
1518 	MPASS(cel->vlp[2] == NULL);
1519 
1520 	MPASS(vp != NULL);
1521 	vlp = VP2VNODELOCK(vp);
1522 
1523 	ret = true;
1524 	if (vlp >= cel->vlp[1]) {
1525 		mtx_lock(vlp);
1526 	} else {
1527 		if (mtx_trylock(vlp))
1528 			goto out;
1529 		cache_lock_vnodes_cel_3_failures++;
1530 		cache_unlock_vnodes_cel(cel);
1531 		if (vlp < cel->vlp[0]) {
1532 			mtx_lock(vlp);
1533 			mtx_lock(cel->vlp[0]);
1534 			mtx_lock(cel->vlp[1]);
1535 		} else {
1536 			if (cel->vlp[0] != NULL)
1537 				mtx_lock(cel->vlp[0]);
1538 			mtx_lock(vlp);
1539 			mtx_lock(cel->vlp[1]);
1540 		}
1541 		ret = false;
1542 	}
1543 out:
1544 	cel->vlp[2] = vlp;
1545 	return (ret);
1546 }
1547 
1548 static void
1549 cache_lock_buckets_cel(struct celockstate *cel, struct rwlock *blp1,
1550     struct rwlock *blp2)
1551 {
1552 
1553 	MPASS(cel->blp[0] == NULL);
1554 	MPASS(cel->blp[1] == NULL);
1555 
1556 	cache_sort_vnodes(&blp1, &blp2);
1557 
1558 	if (blp1 != NULL) {
1559 		rw_wlock(blp1);
1560 		cel->blp[0] = blp1;
1561 	}
1562 	rw_wlock(blp2);
1563 	cel->blp[1] = blp2;
1564 }
1565 
1566 static void
1567 cache_unlock_buckets_cel(struct celockstate *cel)
1568 {
1569 
1570 	if (cel->blp[0] != NULL)
1571 		rw_wunlock(cel->blp[0]);
1572 	rw_wunlock(cel->blp[1]);
1573 }
1574 
1575 /*
1576  * Lock part of the cache affected by the insertion.
1577  *
1578  * This means vnodelocks for dvp, vp and the relevant bucketlock.
1579  * However, insertion can result in removal of an old entry. In this
1580  * case we have an additional vnode and bucketlock pair to lock. If the
1581  * entry is negative, ncelock is locked instead of the vnode.
1582  *
1583  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
1584  * preserving the locking order (smaller address first).
1585  */
1586 static void
1587 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1588     uint32_t hash)
1589 {
1590 	struct namecache *ncp;
1591 	struct rwlock *blps[2];
1592 
1593 	blps[0] = HASH2BUCKETLOCK(hash);
1594 	for (;;) {
1595 		blps[1] = NULL;
1596 		cache_lock_vnodes_cel(cel, dvp, vp);
1597 		if (vp == NULL || vp->v_type != VDIR)
1598 			break;
1599 		ncp = vp->v_cache_dd;
1600 		if (ncp == NULL)
1601 			break;
1602 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1603 			break;
1604 		MPASS(ncp->nc_dvp == vp);
1605 		blps[1] = NCP2BUCKETLOCK(ncp);
1606 		if (ncp->nc_flag & NCF_NEGATIVE)
1607 			break;
1608 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1609 			break;
1610 		/*
1611 		 * All vnodes got re-locked. Re-validate the state and if
1612 		 * nothing changed we are done. Otherwise restart.
1613 		 */
1614 		if (ncp == vp->v_cache_dd &&
1615 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1616 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1617 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1618 			break;
1619 		cache_unlock_vnodes_cel(cel);
1620 		cel->vlp[0] = NULL;
1621 		cel->vlp[1] = NULL;
1622 		cel->vlp[2] = NULL;
1623 	}
1624 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1625 }
1626 
1627 static void
1628 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1629     uint32_t hash)
1630 {
1631 	struct namecache *ncp;
1632 	struct rwlock *blps[2];
1633 
1634 	blps[0] = HASH2BUCKETLOCK(hash);
1635 	for (;;) {
1636 		blps[1] = NULL;
1637 		cache_lock_vnodes_cel(cel, dvp, vp);
1638 		ncp = dvp->v_cache_dd;
1639 		if (ncp == NULL)
1640 			break;
1641 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1642 			break;
1643 		MPASS(ncp->nc_dvp == dvp);
1644 		blps[1] = NCP2BUCKETLOCK(ncp);
1645 		if (ncp->nc_flag & NCF_NEGATIVE)
1646 			break;
1647 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1648 			break;
1649 		if (ncp == dvp->v_cache_dd &&
1650 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1651 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1652 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1653 			break;
1654 		cache_unlock_vnodes_cel(cel);
1655 		cel->vlp[0] = NULL;
1656 		cel->vlp[1] = NULL;
1657 		cel->vlp[2] = NULL;
1658 	}
1659 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1660 }
1661 
1662 static void
1663 cache_enter_unlock(struct celockstate *cel)
1664 {
1665 
1666 	cache_unlock_buckets_cel(cel);
1667 	cache_unlock_vnodes_cel(cel);
1668 }
1669 
1670 static void __noinline
1671 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp,
1672     struct componentname *cnp)
1673 {
1674 	struct celockstate cel;
1675 	struct namecache *ncp;
1676 	uint32_t hash;
1677 	int len;
1678 
1679 	if (dvp->v_cache_dd == NULL)
1680 		return;
1681 	len = cnp->cn_namelen;
1682 	cache_celockstate_init(&cel);
1683 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1684 	cache_enter_lock_dd(&cel, dvp, vp, hash);
1685 	ncp = dvp->v_cache_dd;
1686 	if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) {
1687 		KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent"));
1688 		cache_zap_locked(ncp, false);
1689 	} else {
1690 		ncp = NULL;
1691 	}
1692 	dvp->v_cache_dd = NULL;
1693 	cache_enter_unlock(&cel);
1694 	cache_free(ncp);
1695 }
1696 
1697 /*
1698  * Add an entry to the cache.
1699  */
1700 void
1701 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
1702     struct timespec *tsp, struct timespec *dtsp)
1703 {
1704 	struct celockstate cel;
1705 	struct namecache *ncp, *n2, *ndd;
1706 	struct namecache_ts *ncp_ts, *n2_ts;
1707 	struct nchashhead *ncpp;
1708 	uint32_t hash;
1709 	int flag;
1710 	int len;
1711 	u_long lnumcache;
1712 
1713 	CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr);
1714 	VNASSERT(vp == NULL || !VN_IS_DOOMED(vp), vp,
1715 	    ("cache_enter: Adding a doomed vnode"));
1716 	VNASSERT(dvp == NULL || !VN_IS_DOOMED(dvp), dvp,
1717 	    ("cache_enter: Doomed vnode used as src"));
1718 
1719 #ifdef DEBUG_CACHE
1720 	if (__predict_false(!doingcache))
1721 		return;
1722 #endif
1723 
1724 	flag = 0;
1725 	if (__predict_false(cnp->cn_nameptr[0] == '.')) {
1726 		if (cnp->cn_namelen == 1)
1727 			return;
1728 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1729 			cache_enter_dotdot_prep(dvp, vp, cnp);
1730 			flag = NCF_ISDOTDOT;
1731 		}
1732 	}
1733 
1734 	/*
1735 	 * Avoid blowout in namecache entries.
1736 	 */
1737 	lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
1738 	if (__predict_false(lnumcache >= ncsize)) {
1739 		atomic_add_long(&numcache, -1);
1740 		return;
1741 	}
1742 
1743 	cache_celockstate_init(&cel);
1744 	ndd = NULL;
1745 	ncp_ts = NULL;
1746 
1747 	/*
1748 	 * Calculate the hash key and setup as much of the new
1749 	 * namecache entry as possible before acquiring the lock.
1750 	 */
1751 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
1752 	ncp->nc_flag = flag;
1753 	ncp->nc_vp = vp;
1754 	if (vp == NULL)
1755 		ncp->nc_flag |= NCF_NEGATIVE;
1756 	ncp->nc_dvp = dvp;
1757 	if (tsp != NULL) {
1758 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1759 		ncp_ts->nc_time = *tsp;
1760 		ncp_ts->nc_ticks = ticks;
1761 		ncp_ts->nc_nc.nc_flag |= NCF_TS;
1762 		if (dtsp != NULL) {
1763 			ncp_ts->nc_dotdottime = *dtsp;
1764 			ncp_ts->nc_nc.nc_flag |= NCF_DTS;
1765 		}
1766 	}
1767 	len = ncp->nc_nlen = cnp->cn_namelen;
1768 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1769 	strlcpy(ncp->nc_name, cnp->cn_nameptr, len + 1);
1770 	cache_enter_lock(&cel, dvp, vp, hash);
1771 
1772 	/*
1773 	 * See if this vnode or negative entry is already in the cache
1774 	 * with this name.  This can happen with concurrent lookups of
1775 	 * the same path name.
1776 	 */
1777 	ncpp = NCHHASH(hash);
1778 	LIST_FOREACH(n2, ncpp, nc_hash) {
1779 		if (n2->nc_dvp == dvp &&
1780 		    n2->nc_nlen == cnp->cn_namelen &&
1781 		    !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
1782 			if (tsp != NULL) {
1783 				KASSERT((n2->nc_flag & NCF_TS) != 0,
1784 				    ("no NCF_TS"));
1785 				n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
1786 				n2_ts->nc_time = ncp_ts->nc_time;
1787 				n2_ts->nc_ticks = ncp_ts->nc_ticks;
1788 				if (dtsp != NULL) {
1789 					n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
1790 					if (ncp->nc_flag & NCF_NEGATIVE)
1791 						mtx_lock(&ncneg_hot.nl_lock);
1792 					n2_ts->nc_nc.nc_flag |= NCF_DTS;
1793 					if (ncp->nc_flag & NCF_NEGATIVE)
1794 						mtx_unlock(&ncneg_hot.nl_lock);
1795 				}
1796 			}
1797 			goto out_unlock_free;
1798 		}
1799 	}
1800 
1801 	if (flag == NCF_ISDOTDOT) {
1802 		/*
1803 		 * See if we are trying to add .. entry, but some other lookup
1804 		 * has populated v_cache_dd pointer already.
1805 		 */
1806 		if (dvp->v_cache_dd != NULL)
1807 			goto out_unlock_free;
1808 		KASSERT(vp == NULL || vp->v_type == VDIR,
1809 		    ("wrong vnode type %p", vp));
1810 		dvp->v_cache_dd = ncp;
1811 	}
1812 
1813 	if (vp != NULL) {
1814 		if (vp->v_type == VDIR) {
1815 			if (flag != NCF_ISDOTDOT) {
1816 				/*
1817 				 * For this case, the cache entry maps both the
1818 				 * directory name in it and the name ".." for the
1819 				 * directory's parent.
1820 				 */
1821 				if ((ndd = vp->v_cache_dd) != NULL) {
1822 					if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
1823 						cache_zap_locked(ndd, false);
1824 					else
1825 						ndd = NULL;
1826 				}
1827 				vp->v_cache_dd = ncp;
1828 			}
1829 		} else {
1830 			vp->v_cache_dd = NULL;
1831 		}
1832 	}
1833 
1834 	if (flag != NCF_ISDOTDOT) {
1835 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1836 			vhold(dvp);
1837 			counter_u64_add(numcachehv, 1);
1838 		}
1839 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
1840 	}
1841 
1842 	/*
1843 	 * Insert the new namecache entry into the appropriate chain
1844 	 * within the cache entries table.
1845 	 */
1846 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
1847 
1848 	/*
1849 	 * If the entry is "negative", we place it into the
1850 	 * "negative" cache queue, otherwise, we place it into the
1851 	 * destination vnode's cache entries queue.
1852 	 */
1853 	if (vp != NULL) {
1854 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
1855 		SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
1856 		    vp);
1857 	} else {
1858 		if (cnp->cn_flags & ISWHITEOUT)
1859 			ncp->nc_flag |= NCF_WHITE;
1860 		cache_negative_insert(ncp, false);
1861 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
1862 		    ncp->nc_name);
1863 	}
1864 	cache_enter_unlock(&cel);
1865 	if (numneg * ncnegfactor > lnumcache)
1866 		cache_negative_zap_one();
1867 	cache_free(ndd);
1868 	return;
1869 out_unlock_free:
1870 	cache_enter_unlock(&cel);
1871 	cache_free(ncp);
1872 	return;
1873 }
1874 
1875 static u_int
1876 cache_roundup_2(u_int val)
1877 {
1878 	u_int res;
1879 
1880 	for (res = 1; res <= val; res <<= 1)
1881 		continue;
1882 
1883 	return (res);
1884 }
1885 
1886 /*
1887  * Name cache initialization, from vfs_init() when we are booting
1888  */
1889 static void
1890 nchinit(void *dummy __unused)
1891 {
1892 	u_int i;
1893 
1894 	cache_zone_small = uma_zcreate("S VFS Cache",
1895 	    sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1,
1896 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1897 	    UMA_ZONE_ZINIT);
1898 	cache_zone_small_ts = uma_zcreate("STS VFS Cache",
1899 	    sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1,
1900 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1901 	    UMA_ZONE_ZINIT);
1902 	cache_zone_large = uma_zcreate("L VFS Cache",
1903 	    sizeof(struct namecache) + NAME_MAX + 1,
1904 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1905 	    UMA_ZONE_ZINIT);
1906 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache",
1907 	    sizeof(struct namecache_ts) + NAME_MAX + 1,
1908 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1909 	    UMA_ZONE_ZINIT);
1910 
1911 	ncsize = desiredvnodes * ncsizefactor;
1912 	nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash);
1913 	ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
1914 	if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
1915 		ncbuckethash = 7;
1916 	if (ncbuckethash > nchash)
1917 		ncbuckethash = nchash;
1918 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
1919 	    M_WAITOK | M_ZERO);
1920 	for (i = 0; i < numbucketlocks; i++)
1921 		rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK | RW_RECURSE);
1922 	ncvnodehash = ncbuckethash;
1923 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
1924 	    M_WAITOK | M_ZERO);
1925 	for (i = 0; i < numvnodelocks; i++)
1926 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
1927 	ncpurgeminvnodes = numbucketlocks * 2;
1928 
1929 	ncneghash = 3;
1930 	neglists = malloc(sizeof(*neglists) * numneglists, M_VFSCACHE,
1931 	    M_WAITOK | M_ZERO);
1932 	for (i = 0; i < numneglists; i++) {
1933 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
1934 		TAILQ_INIT(&neglists[i].nl_list);
1935 	}
1936 	mtx_init(&ncneg_hot.nl_lock, "ncneglh", NULL, MTX_DEF);
1937 	TAILQ_INIT(&ncneg_hot.nl_list);
1938 
1939 	mtx_init(&ncneg_shrink_lock, "ncnegs", NULL, MTX_DEF);
1940 }
1941 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
1942 
1943 void
1944 cache_changesize(u_long newmaxvnodes)
1945 {
1946 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
1947 	u_long new_nchash, old_nchash;
1948 	struct namecache *ncp;
1949 	uint32_t hash;
1950 	u_long newncsize;
1951 	int i;
1952 
1953 	newncsize = newmaxvnodes * ncsizefactor;
1954 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
1955 	if (newmaxvnodes < numbucketlocks)
1956 		newmaxvnodes = numbucketlocks;
1957 
1958 	new_nchashtbl = hashinit(newmaxvnodes, M_VFSCACHE, &new_nchash);
1959 	/* If same hash table size, nothing to do */
1960 	if (nchash == new_nchash) {
1961 		free(new_nchashtbl, M_VFSCACHE);
1962 		return;
1963 	}
1964 	/*
1965 	 * Move everything from the old hash table to the new table.
1966 	 * None of the namecache entries in the table can be removed
1967 	 * because to do so, they have to be removed from the hash table.
1968 	 */
1969 	cache_lock_all_vnodes();
1970 	cache_lock_all_buckets();
1971 	old_nchashtbl = nchashtbl;
1972 	old_nchash = nchash;
1973 	nchashtbl = new_nchashtbl;
1974 	nchash = new_nchash;
1975 	for (i = 0; i <= old_nchash; i++) {
1976 		while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) {
1977 			hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
1978 			    ncp->nc_dvp);
1979 			LIST_REMOVE(ncp, nc_hash);
1980 			LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
1981 		}
1982 	}
1983 	ncsize = newncsize;
1984 	cache_unlock_all_buckets();
1985 	cache_unlock_all_vnodes();
1986 	free(old_nchashtbl, M_VFSCACHE);
1987 }
1988 
1989 /*
1990  * Invalidate all entries from and to a particular vnode.
1991  */
1992 void
1993 cache_purge(struct vnode *vp)
1994 {
1995 	TAILQ_HEAD(, namecache) ncps;
1996 	struct namecache *ncp, *nnp;
1997 	struct mtx *vlp, *vlp2;
1998 
1999 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
2000 	SDT_PROBE1(vfs, namecache, purge, done, vp);
2001 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2002 	    vp->v_cache_dd == NULL)
2003 		return;
2004 	TAILQ_INIT(&ncps);
2005 	vlp = VP2VNODELOCK(vp);
2006 	vlp2 = NULL;
2007 	mtx_lock(vlp);
2008 retry:
2009 	while (!LIST_EMPTY(&vp->v_cache_src)) {
2010 		ncp = LIST_FIRST(&vp->v_cache_src);
2011 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2012 			goto retry;
2013 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2014 	}
2015 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2016 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
2017 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2018 			goto retry;
2019 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2020 	}
2021 	ncp = vp->v_cache_dd;
2022 	if (ncp != NULL) {
2023 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2024 		   ("lost dotdot link"));
2025 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2026 			goto retry;
2027 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2028 	}
2029 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2030 	mtx_unlock(vlp);
2031 	if (vlp2 != NULL)
2032 		mtx_unlock(vlp2);
2033 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2034 		cache_free(ncp);
2035 	}
2036 }
2037 
2038 /*
2039  * Invalidate all negative entries for a particular directory vnode.
2040  */
2041 void
2042 cache_purge_negative(struct vnode *vp)
2043 {
2044 	TAILQ_HEAD(, namecache) ncps;
2045 	struct namecache *ncp, *nnp;
2046 	struct mtx *vlp;
2047 
2048 	CTR1(KTR_VFS, "cache_purge_negative(%p)", vp);
2049 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2050 	if (LIST_EMPTY(&vp->v_cache_src))
2051 		return;
2052 	TAILQ_INIT(&ncps);
2053 	vlp = VP2VNODELOCK(vp);
2054 	mtx_lock(vlp);
2055 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2056 		if (!(ncp->nc_flag & NCF_NEGATIVE))
2057 			continue;
2058 		cache_zap_negative_locked_vnode_kl(ncp, vp);
2059 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2060 	}
2061 	mtx_unlock(vlp);
2062 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2063 		cache_free(ncp);
2064 	}
2065 }
2066 
2067 /*
2068  * Flush all entries referencing a particular filesystem.
2069  */
2070 void
2071 cache_purgevfs(struct mount *mp, bool force)
2072 {
2073 	TAILQ_HEAD(, namecache) ncps;
2074 	struct mtx *vlp1, *vlp2;
2075 	struct rwlock *blp;
2076 	struct nchashhead *bucket;
2077 	struct namecache *ncp, *nnp;
2078 	u_long i, j, n_nchash;
2079 	int error;
2080 
2081 	/* Scan hash tables for applicable entries */
2082 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2083 	if (!force && mp->mnt_nvnodelistsize <= ncpurgeminvnodes)
2084 		return;
2085 	TAILQ_INIT(&ncps);
2086 	n_nchash = nchash + 1;
2087 	vlp1 = vlp2 = NULL;
2088 	for (i = 0; i < numbucketlocks; i++) {
2089 		blp = (struct rwlock *)&bucketlocks[i];
2090 		rw_wlock(blp);
2091 		for (j = i; j < n_nchash; j += numbucketlocks) {
2092 retry:
2093 			bucket = &nchashtbl[j];
2094 			LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) {
2095 				cache_assert_bucket_locked(ncp, RA_WLOCKED);
2096 				if (ncp->nc_dvp->v_mount != mp)
2097 					continue;
2098 				error = cache_zap_wlocked_bucket_kl(ncp, blp,
2099 				    &vlp1, &vlp2);
2100 				if (error != 0)
2101 					goto retry;
2102 				TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst);
2103 			}
2104 		}
2105 		rw_wunlock(blp);
2106 		if (vlp1 == NULL && vlp2 == NULL)
2107 			cache_maybe_yield();
2108 	}
2109 	if (vlp1 != NULL)
2110 		mtx_unlock(vlp1);
2111 	if (vlp2 != NULL)
2112 		mtx_unlock(vlp2);
2113 
2114 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2115 		cache_free(ncp);
2116 	}
2117 }
2118 
2119 /*
2120  * Perform canonical checks and cache lookup and pass on to filesystem
2121  * through the vop_cachedlookup only if needed.
2122  */
2123 
2124 int
2125 vfs_cache_lookup(struct vop_lookup_args *ap)
2126 {
2127 	struct vnode *dvp;
2128 	int error;
2129 	struct vnode **vpp = ap->a_vpp;
2130 	struct componentname *cnp = ap->a_cnp;
2131 	int flags = cnp->cn_flags;
2132 
2133 	*vpp = NULL;
2134 	dvp = ap->a_dvp;
2135 
2136 	if (dvp->v_type != VDIR)
2137 		return (ENOTDIR);
2138 
2139 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2140 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2141 		return (EROFS);
2142 
2143 	error = vn_dir_check_exec(dvp, cnp);
2144 	if (error != 0)
2145 		return (error);
2146 
2147 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2148 	if (error == 0)
2149 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2150 	if (error == -1)
2151 		return (0);
2152 	return (error);
2153 }
2154 
2155 /* Implementation of the getcwd syscall. */
2156 int
2157 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2158 {
2159 	char *buf, *retbuf;
2160 	size_t buflen;
2161 	int error;
2162 
2163 	buflen = uap->buflen;
2164 	if (__predict_false(buflen < 2))
2165 		return (EINVAL);
2166 	if (buflen > MAXPATHLEN)
2167 		buflen = MAXPATHLEN;
2168 
2169 	buf = malloc(buflen, M_TEMP, M_WAITOK);
2170 	error = vn_getcwd(td, buf, &retbuf, &buflen);
2171 	if (error == 0)
2172 		error = copyout(retbuf, uap->buf, buflen);
2173 	free(buf, M_TEMP);
2174 	return (error);
2175 }
2176 
2177 int
2178 vn_getcwd(struct thread *td, char *buf, char **retbuf, size_t *buflen)
2179 {
2180 	struct pwd *pwd;
2181 	int error;
2182 
2183 	pwd = pwd_hold(td);
2184 	error = vn_fullpath_any(td, pwd->pwd_cdir, pwd->pwd_rdir, buf, retbuf, buflen);
2185 	pwd_drop(pwd);
2186 
2187 #ifdef KTRACE
2188 	if (KTRPOINT(curthread, KTR_NAMEI) && error == 0)
2189 		ktrnamei(*retbuf);
2190 #endif
2191 	return (error);
2192 }
2193 
2194 static int
2195 kern___realpathat(struct thread *td, int fd, const char *path, char *buf,
2196     size_t size, int flags, enum uio_seg pathseg)
2197 {
2198 	struct nameidata nd;
2199 	char *retbuf, *freebuf;
2200 	int error;
2201 
2202 	if (flags != 0)
2203 		return (EINVAL);
2204 	NDINIT_ATRIGHTS(&nd, LOOKUP, FOLLOW | SAVENAME | WANTPARENT | AUDITVNODE1,
2205 	    pathseg, path, fd, &cap_fstat_rights, td);
2206 	if ((error = namei(&nd)) != 0)
2207 		return (error);
2208 	error = vn_fullpath_hardlink(td, &nd, &retbuf, &freebuf, &size);
2209 	if (error == 0) {
2210 		error = copyout(retbuf, buf, size);
2211 		free(freebuf, M_TEMP);
2212 	}
2213 	NDFREE(&nd, 0);
2214 	return (error);
2215 }
2216 
2217 int
2218 sys___realpathat(struct thread *td, struct __realpathat_args *uap)
2219 {
2220 
2221 	return (kern___realpathat(td, uap->fd, uap->path, uap->buf, uap->size,
2222 	    uap->flags, UIO_USERSPACE));
2223 }
2224 
2225 /*
2226  * Retrieve the full filesystem path that correspond to a vnode from the name
2227  * cache (if available)
2228  */
2229 int
2230 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf)
2231 {
2232 	struct pwd *pwd;
2233 	char *buf;
2234 	size_t buflen;
2235 	int error;
2236 
2237 	if (__predict_false(vn == NULL))
2238 		return (EINVAL);
2239 
2240 	buflen = MAXPATHLEN;
2241 	buf = malloc(buflen, M_TEMP, M_WAITOK);
2242 	pwd = pwd_hold(td);
2243 	error = vn_fullpath_any(td, vn, pwd->pwd_rdir, buf, retbuf, &buflen);
2244 	pwd_drop(pwd);
2245 
2246 	if (!error)
2247 		*freebuf = buf;
2248 	else
2249 		free(buf, M_TEMP);
2250 	return (error);
2251 }
2252 
2253 /*
2254  * This function is similar to vn_fullpath, but it attempts to lookup the
2255  * pathname relative to the global root mount point.  This is required for the
2256  * auditing sub-system, as audited pathnames must be absolute, relative to the
2257  * global root mount point.
2258  */
2259 int
2260 vn_fullpath_global(struct thread *td, struct vnode *vn,
2261     char **retbuf, char **freebuf)
2262 {
2263 	char *buf;
2264 	size_t buflen;
2265 	int error;
2266 
2267 	if (__predict_false(vn == NULL))
2268 		return (EINVAL);
2269 	buflen = MAXPATHLEN;
2270 	buf = malloc(buflen, M_TEMP, M_WAITOK);
2271 	error = vn_fullpath_any(td, vn, rootvnode, buf, retbuf, &buflen);
2272 	if (!error)
2273 		*freebuf = buf;
2274 	else
2275 		free(buf, M_TEMP);
2276 	return (error);
2277 }
2278 
2279 int
2280 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, size_t *buflen)
2281 {
2282 	struct vnode *dvp;
2283 	struct namecache *ncp;
2284 	struct mtx *vlp;
2285 	int error;
2286 
2287 	vlp = VP2VNODELOCK(*vp);
2288 	mtx_lock(vlp);
2289 	TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) {
2290 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2291 			break;
2292 	}
2293 	if (ncp != NULL) {
2294 		if (*buflen < ncp->nc_nlen) {
2295 			mtx_unlock(vlp);
2296 			vrele(*vp);
2297 			counter_u64_add(numfullpathfail4, 1);
2298 			error = ENOMEM;
2299 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2300 			    vp, NULL);
2301 			return (error);
2302 		}
2303 		*buflen -= ncp->nc_nlen;
2304 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2305 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2306 		    ncp->nc_name, vp);
2307 		dvp = *vp;
2308 		*vp = ncp->nc_dvp;
2309 		vref(*vp);
2310 		mtx_unlock(vlp);
2311 		vrele(dvp);
2312 		return (0);
2313 	}
2314 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2315 
2316 	mtx_unlock(vlp);
2317 	vn_lock(*vp, LK_SHARED | LK_RETRY);
2318 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
2319 	vput(*vp);
2320 	if (error) {
2321 		counter_u64_add(numfullpathfail2, 1);
2322 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
2323 		return (error);
2324 	}
2325 
2326 	*vp = dvp;
2327 	if (VN_IS_DOOMED(dvp)) {
2328 		/* forced unmount */
2329 		vrele(dvp);
2330 		error = ENOENT;
2331 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2332 		return (error);
2333 	}
2334 	/*
2335 	 * *vp has its use count incremented still.
2336 	 */
2337 
2338 	return (0);
2339 }
2340 
2341 /*
2342  * Resolve a directory to a pathname.
2343  *
2344  * The name of the directory can always be found in the namecache or fetched
2345  * from the filesystem. There is also guaranteed to be only one parent, meaning
2346  * we can just follow vnodes up until we find the root.
2347  *
2348  * The vnode must be referenced.
2349  */
2350 static int
2351 vn_fullpath_dir(struct thread *td, struct vnode *vp, struct vnode *rdir,
2352     char *buf, char **retbuf, size_t *len, bool slash_prefixed, size_t addend)
2353 {
2354 #ifdef KDTRACE_HOOKS
2355 	struct vnode *startvp = vp;
2356 #endif
2357 	struct vnode *vp1;
2358 	size_t buflen;
2359 	int error;
2360 
2361 	VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp);
2362 	VNPASS(vp->v_usecount > 0, vp);
2363 
2364 	buflen = *len;
2365 
2366 	if (!slash_prefixed) {
2367 		MPASS(*len >= 2);
2368 		buflen--;
2369 		buf[buflen] = '\0';
2370 	}
2371 
2372 	error = 0;
2373 
2374 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
2375 	counter_u64_add(numfullpathcalls, 1);
2376 	while (vp != rdir && vp != rootvnode) {
2377 		/*
2378 		 * The vp vnode must be already fully constructed,
2379 		 * since it is either found in namecache or obtained
2380 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
2381 		 * without obtaining the vnode lock.
2382 		 */
2383 		if ((vp->v_vflag & VV_ROOT) != 0) {
2384 			vn_lock(vp, LK_RETRY | LK_SHARED);
2385 
2386 			/*
2387 			 * With the vnode locked, check for races with
2388 			 * unmount, forced or not.  Note that we
2389 			 * already verified that vp is not equal to
2390 			 * the root vnode, which means that
2391 			 * mnt_vnodecovered can be NULL only for the
2392 			 * case of unmount.
2393 			 */
2394 			if (VN_IS_DOOMED(vp) ||
2395 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
2396 			    vp1->v_mountedhere != vp->v_mount) {
2397 				vput(vp);
2398 				error = ENOENT;
2399 				SDT_PROBE3(vfs, namecache, fullpath, return,
2400 				    error, vp, NULL);
2401 				break;
2402 			}
2403 
2404 			vref(vp1);
2405 			vput(vp);
2406 			vp = vp1;
2407 			continue;
2408 		}
2409 		if (vp->v_type != VDIR) {
2410 			vrele(vp);
2411 			counter_u64_add(numfullpathfail1, 1);
2412 			error = ENOTDIR;
2413 			SDT_PROBE3(vfs, namecache, fullpath, return,
2414 			    error, vp, NULL);
2415 			break;
2416 		}
2417 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2418 		if (error)
2419 			break;
2420 		if (buflen == 0) {
2421 			vrele(vp);
2422 			error = ENOMEM;
2423 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2424 			    startvp, NULL);
2425 			break;
2426 		}
2427 		buf[--buflen] = '/';
2428 		slash_prefixed = true;
2429 	}
2430 	if (error)
2431 		return (error);
2432 	if (!slash_prefixed) {
2433 		if (buflen == 0) {
2434 			vrele(vp);
2435 			counter_u64_add(numfullpathfail4, 1);
2436 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
2437 			    startvp, NULL);
2438 			return (ENOMEM);
2439 		}
2440 		buf[--buflen] = '/';
2441 	}
2442 	counter_u64_add(numfullpathfound, 1);
2443 	vrele(vp);
2444 
2445 	*retbuf = buf + buflen;
2446 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, *retbuf);
2447 	*len -= buflen;
2448 	*len += addend;
2449 	return (0);
2450 }
2451 
2452 /*
2453  * Resolve an arbitrary vnode to a pathname.
2454  *
2455  * Note 2 caveats:
2456  * - hardlinks are not tracked, thus if the vnode is not a directory this can
2457  *   resolve to a different path than the one used to find it
2458  * - namecache is not mandatory, meaning names are not guaranteed to be added
2459  *   (in which case resolving fails)
2460  */
2461 static int
2462 vn_fullpath_any(struct thread *td, struct vnode *vp, struct vnode *rdir,
2463     char *buf, char **retbuf, size_t *buflen)
2464 {
2465 	size_t orig_buflen;
2466 	bool slash_prefixed;
2467 	int error;
2468 
2469 	if (*buflen < 2)
2470 		return (EINVAL);
2471 
2472 	orig_buflen = *buflen;
2473 
2474 	vref(vp);
2475 	slash_prefixed = false;
2476 	if (vp->v_type != VDIR) {
2477 		*buflen -= 1;
2478 		buf[*buflen] = '\0';
2479 		error = vn_vptocnp(&vp, td->td_ucred, buf, buflen);
2480 		if (error)
2481 			return (error);
2482 		if (*buflen == 0) {
2483 			vrele(vp);
2484 			return (ENOMEM);
2485 		}
2486 		*buflen -= 1;
2487 		buf[*buflen] = '/';
2488 		slash_prefixed = true;
2489 	}
2490 
2491 	return (vn_fullpath_dir(td, vp, rdir, buf, retbuf, buflen, slash_prefixed,
2492 	    orig_buflen - *buflen));
2493 }
2494 
2495 /*
2496  * Resolve an arbitrary vnode to a pathname (taking care of hardlinks).
2497  *
2498  * Since the namecache does not track handlings, the caller is expected to first
2499  * look up the target vnode with SAVENAME | WANTPARENT flags passed to namei.
2500  *
2501  * Then we have 2 cases:
2502  * - if the found vnode is a directory, the path can be constructed just by
2503  *   fullowing names up the chain
2504  * - otherwise we populate the buffer with the saved name and start resolving
2505  *   from the parent
2506  */
2507 static int
2508 vn_fullpath_hardlink(struct thread *td, struct nameidata *ndp, char **retbuf,
2509     char **freebuf, size_t *buflen)
2510 {
2511 	char *buf, *tmpbuf;
2512 	struct pwd *pwd;
2513 	struct componentname *cnp;
2514 	struct vnode *vp;
2515 	size_t addend;
2516 	int error;
2517 	bool slash_prefixed;
2518 
2519 	if (*buflen < 2)
2520 		return (EINVAL);
2521 	if (*buflen > MAXPATHLEN)
2522 		*buflen = MAXPATHLEN;
2523 
2524 	slash_prefixed = false;
2525 
2526 	buf = malloc(*buflen, M_TEMP, M_WAITOK);
2527 	pwd = pwd_hold(td);
2528 
2529 	addend = 0;
2530 	vp = ndp->ni_vp;
2531 	if (vp->v_type != VDIR) {
2532 		cnp = &ndp->ni_cnd;
2533 		addend = cnp->cn_namelen + 2;
2534 		if (*buflen < addend) {
2535 			error = ENOMEM;
2536 			goto out_bad;
2537 		}
2538 		*buflen -= addend;
2539 		tmpbuf = buf + *buflen;
2540 		tmpbuf[0] = '/';
2541 		memcpy(&tmpbuf[1], cnp->cn_nameptr, cnp->cn_namelen);
2542 		tmpbuf[addend - 1] = '\0';
2543 		slash_prefixed = true;
2544 		vp = ndp->ni_dvp;
2545 	}
2546 
2547 	vref(vp);
2548 	error = vn_fullpath_dir(td, vp, pwd->pwd_rdir, buf, retbuf, buflen,
2549 	    slash_prefixed, addend);
2550 	if (error != 0)
2551 		goto out_bad;
2552 
2553 	pwd_drop(pwd);
2554 	*freebuf = buf;
2555 
2556 	return (0);
2557 out_bad:
2558 	pwd_drop(pwd);
2559 	free(buf, M_TEMP);
2560 	return (error);
2561 }
2562 
2563 struct vnode *
2564 vn_dir_dd_ino(struct vnode *vp)
2565 {
2566 	struct namecache *ncp;
2567 	struct vnode *ddvp;
2568 	struct mtx *vlp;
2569 	enum vgetstate vs;
2570 
2571 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
2572 	vlp = VP2VNODELOCK(vp);
2573 	mtx_lock(vlp);
2574 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
2575 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
2576 			continue;
2577 		ddvp = ncp->nc_dvp;
2578 		vs = vget_prep(ddvp);
2579 		mtx_unlock(vlp);
2580 		if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
2581 			return (NULL);
2582 		return (ddvp);
2583 	}
2584 	mtx_unlock(vlp);
2585 	return (NULL);
2586 }
2587 
2588 int
2589 vn_commname(struct vnode *vp, char *buf, u_int buflen)
2590 {
2591 	struct namecache *ncp;
2592 	struct mtx *vlp;
2593 	int l;
2594 
2595 	vlp = VP2VNODELOCK(vp);
2596 	mtx_lock(vlp);
2597 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
2598 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2599 			break;
2600 	if (ncp == NULL) {
2601 		mtx_unlock(vlp);
2602 		return (ENOENT);
2603 	}
2604 	l = min(ncp->nc_nlen, buflen - 1);
2605 	memcpy(buf, ncp->nc_name, l);
2606 	mtx_unlock(vlp);
2607 	buf[l] = '\0';
2608 	return (0);
2609 }
2610 
2611 /*
2612  * This function updates path string to vnode's full global path
2613  * and checks the size of the new path string against the pathlen argument.
2614  *
2615  * Requires a locked, referenced vnode.
2616  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
2617  *
2618  * If vp is a directory, the call to vn_fullpath_global() always succeeds
2619  * because it falls back to the ".." lookup if the namecache lookup fails.
2620  */
2621 int
2622 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
2623     u_int pathlen)
2624 {
2625 	struct nameidata nd;
2626 	struct vnode *vp1;
2627 	char *rpath, *fbuf;
2628 	int error;
2629 
2630 	ASSERT_VOP_ELOCKED(vp, __func__);
2631 
2632 	/* Construct global filesystem path from vp. */
2633 	VOP_UNLOCK(vp);
2634 	error = vn_fullpath_global(td, vp, &rpath, &fbuf);
2635 
2636 	if (error != 0) {
2637 		vrele(vp);
2638 		return (error);
2639 	}
2640 
2641 	if (strlen(rpath) >= pathlen) {
2642 		vrele(vp);
2643 		error = ENAMETOOLONG;
2644 		goto out;
2645 	}
2646 
2647 	/*
2648 	 * Re-lookup the vnode by path to detect a possible rename.
2649 	 * As a side effect, the vnode is relocked.
2650 	 * If vnode was renamed, return ENOENT.
2651 	 */
2652 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
2653 	    UIO_SYSSPACE, path, td);
2654 	error = namei(&nd);
2655 	if (error != 0) {
2656 		vrele(vp);
2657 		goto out;
2658 	}
2659 	NDFREE(&nd, NDF_ONLY_PNBUF);
2660 	vp1 = nd.ni_vp;
2661 	vrele(vp);
2662 	if (vp1 == vp)
2663 		strcpy(path, rpath);
2664 	else {
2665 		vput(vp1);
2666 		error = ENOENT;
2667 	}
2668 
2669 out:
2670 	free(fbuf, M_TEMP);
2671 	return (error);
2672 }
2673 
2674 #ifdef DDB
2675 static void
2676 db_print_vpath(struct vnode *vp)
2677 {
2678 
2679 	while (vp != NULL) {
2680 		db_printf("%p: ", vp);
2681 		if (vp == rootvnode) {
2682 			db_printf("/");
2683 			vp = NULL;
2684 		} else {
2685 			if (vp->v_vflag & VV_ROOT) {
2686 				db_printf("<mount point>");
2687 				vp = vp->v_mount->mnt_vnodecovered;
2688 			} else {
2689 				struct namecache *ncp;
2690 				char *ncn;
2691 				int i;
2692 
2693 				ncp = TAILQ_FIRST(&vp->v_cache_dst);
2694 				if (ncp != NULL) {
2695 					ncn = ncp->nc_name;
2696 					for (i = 0; i < ncp->nc_nlen; i++)
2697 						db_printf("%c", *ncn++);
2698 					vp = ncp->nc_dvp;
2699 				} else {
2700 					vp = NULL;
2701 				}
2702 			}
2703 		}
2704 		db_printf("\n");
2705 	}
2706 
2707 	return;
2708 }
2709 
2710 DB_SHOW_COMMAND(vpath, db_show_vpath)
2711 {
2712 	struct vnode *vp;
2713 
2714 	if (!have_addr) {
2715 		db_printf("usage: show vpath <struct vnode *>\n");
2716 		return;
2717 	}
2718 
2719 	vp = (struct vnode *)addr;
2720 	db_print_vpath(vp);
2721 }
2722 
2723 #endif
2724