xref: /freebsd/sys/kern/vfs_cache.c (revision 8bcb0991864975618c09697b1aca10683346d9f0)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993, 1995
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * Poul-Henning Kamp of the FreeBSD Project.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)vfs_cache.c	8.5 (Berkeley) 3/22/95
35  */
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include "opt_ddb.h"
41 #include "opt_ktrace.h"
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/counter.h>
46 #include <sys/filedesc.h>
47 #include <sys/fnv_hash.h>
48 #include <sys/kernel.h>
49 #include <sys/ktr.h>
50 #include <sys/lock.h>
51 #include <sys/malloc.h>
52 #include <sys/fcntl.h>
53 #include <sys/mount.h>
54 #include <sys/namei.h>
55 #include <sys/proc.h>
56 #include <sys/rwlock.h>
57 #include <sys/sdt.h>
58 #include <sys/smp.h>
59 #include <sys/syscallsubr.h>
60 #include <sys/sysctl.h>
61 #include <sys/sysproto.h>
62 #include <sys/vnode.h>
63 #ifdef KTRACE
64 #include <sys/ktrace.h>
65 #endif
66 
67 #ifdef DDB
68 #include <ddb/ddb.h>
69 #endif
70 
71 #include <vm/uma.h>
72 
73 SDT_PROVIDER_DECLARE(vfs);
74 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
75     "struct vnode *");
76 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
77     "char *");
78 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
79 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
80     "char *", "struct vnode *");
81 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
82 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
83     "struct vnode *", "char *");
84 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
85     "struct vnode *");
86 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
87     "struct vnode *", "char *");
88 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
89     "char *");
90 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *");
91 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
92 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
93 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
94     "struct vnode *");
95 SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *",
96     "char *");
97 SDT_PROBE_DEFINE2(vfs, namecache, shrink_negative, done, "struct vnode *",
98     "char *");
99 
100 /*
101  * This structure describes the elements in the cache of recent
102  * names looked up by namei.
103  */
104 
105 struct	namecache {
106 	LIST_ENTRY(namecache) nc_hash;	/* hash chain */
107 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
108 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
109 	struct	vnode *nc_dvp;		/* vnode of parent of name */
110 	union {
111 		struct	vnode *nu_vp;	/* vnode the name refers to */
112 	} n_un;
113 	u_char	nc_flag;		/* flag bits */
114 	u_char	nc_nlen;		/* length of name */
115 	char	nc_name[0];		/* segment name + nul */
116 };
117 
118 /*
119  * struct namecache_ts repeats struct namecache layout up to the
120  * nc_nlen member.
121  * struct namecache_ts is used in place of struct namecache when time(s) need
122  * to be stored.  The nc_dotdottime field is used when a cache entry is mapping
123  * both a non-dotdot directory name plus dotdot for the directory's
124  * parent.
125  */
126 struct	namecache_ts {
127 	struct	timespec nc_time;	/* timespec provided by fs */
128 	struct	timespec nc_dotdottime;	/* dotdot timespec provided by fs */
129 	int	nc_ticks;		/* ticks value when entry was added */
130 	struct namecache nc_nc;
131 };
132 
133 #define	nc_vp		n_un.nu_vp
134 
135 /*
136  * Flags in namecache.nc_flag
137  */
138 #define NCF_WHITE	0x01
139 #define NCF_ISDOTDOT	0x02
140 #define	NCF_TS		0x04
141 #define	NCF_DTS		0x08
142 #define	NCF_DVDROP	0x10
143 #define	NCF_NEGATIVE	0x20
144 #define	NCF_HOTNEGATIVE	0x40
145 
146 /*
147  * Name caching works as follows:
148  *
149  * Names found by directory scans are retained in a cache
150  * for future reference.  It is managed LRU, so frequently
151  * used names will hang around.  Cache is indexed by hash value
152  * obtained from (dvp, name) where dvp refers to the directory
153  * containing name.
154  *
155  * If it is a "negative" entry, (i.e. for a name that is known NOT to
156  * exist) the vnode pointer will be NULL.
157  *
158  * Upon reaching the last segment of a path, if the reference
159  * is for DELETE, or NOCACHE is set (rewrite), and the
160  * name is located in the cache, it will be dropped.
161  *
162  * These locks are used (in the order in which they can be taken):
163  * NAME		TYPE	ROLE
164  * vnodelock	mtx	vnode lists and v_cache_dd field protection
165  * bucketlock	rwlock	for access to given set of hash buckets
166  * neglist	mtx	negative entry LRU management
167  *
168  * Additionally, ncneg_shrink_lock mtx is used to have at most one thread
169  * shrinking the LRU list.
170  *
171  * It is legal to take multiple vnodelock and bucketlock locks. The locking
172  * order is lower address first. Both are recursive.
173  *
174  * "." lookups are lockless.
175  *
176  * ".." and vnode -> name lookups require vnodelock.
177  *
178  * name -> vnode lookup requires the relevant bucketlock to be held for reading.
179  *
180  * Insertions and removals of entries require involved vnodes and bucketlocks
181  * to be write-locked to prevent other threads from seeing the entry.
182  *
183  * Some lookups result in removal of the found entry (e.g. getting rid of a
184  * negative entry with the intent to create a positive one), which poses a
185  * problem when multiple threads reach the state. Similarly, two different
186  * threads can purge two different vnodes and try to remove the same name.
187  *
188  * If the already held vnode lock is lower than the second required lock, we
189  * can just take the other lock. However, in the opposite case, this could
190  * deadlock. As such, this is resolved by trylocking and if that fails unlocking
191  * the first node, locking everything in order and revalidating the state.
192  */
193 
194 /*
195  * Structures associated with name caching.
196  */
197 #define NCHHASH(hash) \
198 	(&nchashtbl[(hash) & nchash])
199 static __read_mostly LIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */
200 static u_long __read_mostly	nchash;			/* size of hash table */
201 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
202     "Size of namecache hash table");
203 static u_long __read_mostly	ncnegfactor = 5; /* ratio of negative entries */
204 SYSCTL_ULONG(_vfs, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
205     "Ratio of negative namecache entries");
206 static u_long __exclusive_cache_line	numneg;	/* number of negative entries allocated */
207 static u_long __exclusive_cache_line	numcache;/* number of cache entries allocated */
208 u_int ncsizefactor = 2;
209 SYSCTL_UINT(_vfs, OID_AUTO, ncsizefactor, CTLFLAG_RW, &ncsizefactor, 0,
210     "Size factor for namecache");
211 static u_int __read_mostly	ncpurgeminvnodes;
212 SYSCTL_UINT(_vfs, OID_AUTO, ncpurgeminvnodes, CTLFLAG_RW, &ncpurgeminvnodes, 0,
213     "Number of vnodes below which purgevfs ignores the request");
214 static u_int __read_mostly	ncsize; /* the size as computed on creation or resizing */
215 
216 struct nchstats	nchstats;		/* cache effectiveness statistics */
217 
218 static struct mtx __exclusive_cache_line	ncneg_shrink_lock;
219 static int	shrink_list_turn;
220 
221 struct neglist {
222 	struct mtx		nl_lock;
223 	TAILQ_HEAD(, namecache) nl_list;
224 } __aligned(CACHE_LINE_SIZE);
225 
226 static struct neglist __read_mostly	*neglists;
227 static struct neglist ncneg_hot;
228 static u_long numhotneg;
229 
230 #define	numneglists (ncneghash + 1)
231 static u_int __read_mostly	ncneghash;
232 static inline struct neglist *
233 NCP2NEGLIST(struct namecache *ncp)
234 {
235 
236 	return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
237 }
238 
239 #define	numbucketlocks (ncbuckethash + 1)
240 static u_int __read_mostly  ncbuckethash;
241 static struct rwlock_padalign __read_mostly  *bucketlocks;
242 #define	HASH2BUCKETLOCK(hash) \
243 	((struct rwlock *)(&bucketlocks[((hash) & ncbuckethash)]))
244 
245 #define	numvnodelocks (ncvnodehash + 1)
246 static u_int __read_mostly  ncvnodehash;
247 static struct mtx __read_mostly *vnodelocks;
248 static inline struct mtx *
249 VP2VNODELOCK(struct vnode *vp)
250 {
251 
252 	return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
253 }
254 
255 /*
256  * UMA zones for the VFS cache.
257  *
258  * The small cache is used for entries with short names, which are the
259  * most common.  The large cache is used for entries which are too big to
260  * fit in the small cache.
261  */
262 static uma_zone_t __read_mostly cache_zone_small;
263 static uma_zone_t __read_mostly cache_zone_small_ts;
264 static uma_zone_t __read_mostly cache_zone_large;
265 static uma_zone_t __read_mostly cache_zone_large_ts;
266 
267 #define	CACHE_PATH_CUTOFF	35
268 
269 static struct namecache *
270 cache_alloc(int len, int ts)
271 {
272 	struct namecache_ts *ncp_ts;
273 	struct namecache *ncp;
274 
275 	if (__predict_false(ts)) {
276 		if (len <= CACHE_PATH_CUTOFF)
277 			ncp_ts = uma_zalloc(cache_zone_small_ts, M_WAITOK);
278 		else
279 			ncp_ts = uma_zalloc(cache_zone_large_ts, M_WAITOK);
280 		ncp = &ncp_ts->nc_nc;
281 	} else {
282 		if (len <= CACHE_PATH_CUTOFF)
283 			ncp = uma_zalloc(cache_zone_small, M_WAITOK);
284 		else
285 			ncp = uma_zalloc(cache_zone_large, M_WAITOK);
286 	}
287 	return (ncp);
288 }
289 
290 static void
291 cache_free(struct namecache *ncp)
292 {
293 	struct namecache_ts *ncp_ts;
294 
295 	if (ncp == NULL)
296 		return;
297 	if ((ncp->nc_flag & NCF_DVDROP) != 0)
298 		vdrop(ncp->nc_dvp);
299 	if (__predict_false(ncp->nc_flag & NCF_TS)) {
300 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
301 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
302 			uma_zfree(cache_zone_small_ts, ncp_ts);
303 		else
304 			uma_zfree(cache_zone_large_ts, ncp_ts);
305 	} else {
306 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
307 			uma_zfree(cache_zone_small, ncp);
308 		else
309 			uma_zfree(cache_zone_large, ncp);
310 	}
311 }
312 
313 static void
314 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
315 {
316 	struct namecache_ts *ncp_ts;
317 
318 	KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
319 	    (tsp == NULL && ticksp == NULL),
320 	    ("No NCF_TS"));
321 
322 	if (tsp == NULL && ticksp == NULL)
323 		return;
324 
325 	ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
326 	if (tsp != NULL)
327 		*tsp = ncp_ts->nc_time;
328 	if (ticksp != NULL)
329 		*ticksp = ncp_ts->nc_ticks;
330 }
331 
332 #ifdef DEBUG_CACHE
333 static int __read_mostly	doingcache = 1;	/* 1 => enable the cache */
334 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
335     "VFS namecache enabled");
336 #endif
337 
338 /* Export size information to userland */
339 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
340     sizeof(struct namecache), "sizeof(struct namecache)");
341 
342 /*
343  * The new name cache statistics
344  */
345 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0,
346     "Name cache statistics");
347 #define STATNODE_ULONG(name, descr)	\
348 	SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, descr);
349 #define STATNODE_COUNTER(name, descr)	\
350 	static counter_u64_t __read_mostly name; \
351 	SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, descr);
352 STATNODE_ULONG(numneg, "Number of negative cache entries");
353 STATNODE_ULONG(numcache, "Number of cache entries");
354 STATNODE_COUNTER(numcachehv, "Number of namecache entries with vnodes held");
355 STATNODE_COUNTER(numcalls, "Number of cache lookups");
356 STATNODE_COUNTER(dothits, "Number of '.' hits");
357 STATNODE_COUNTER(dotdothits, "Number of '..' hits");
358 STATNODE_COUNTER(numchecks, "Number of checks in lookup");
359 STATNODE_COUNTER(nummiss, "Number of cache misses");
360 STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache");
361 STATNODE_COUNTER(numposzaps,
362     "Number of cache hits (positive) we do not want to cache");
363 STATNODE_COUNTER(numposhits, "Number of cache hits (positive)");
364 STATNODE_COUNTER(numnegzaps,
365     "Number of cache hits (negative) we do not want to cache");
366 STATNODE_COUNTER(numneghits, "Number of cache hits (negative)");
367 /* These count for kern___getcwd(), too. */
368 STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls");
369 STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
370 STATNODE_COUNTER(numfullpathfail2,
371     "Number of fullpath search errors (VOP_VPTOCNP failures)");
372 STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
373 STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls");
374 STATNODE_COUNTER(zap_and_exit_bucket_relock_success,
375     "Number of successful removals after relocking");
376 static long zap_and_exit_bucket_fail; STATNODE_ULONG(zap_and_exit_bucket_fail,
377     "Number of times zap_and_exit failed to lock");
378 static long zap_and_exit_bucket_fail2; STATNODE_ULONG(zap_and_exit_bucket_fail2,
379     "Number of times zap_and_exit failed to lock");
380 static long cache_lock_vnodes_cel_3_failures;
381 STATNODE_ULONG(cache_lock_vnodes_cel_3_failures,
382     "Number of times 3-way vnode locking failed");
383 STATNODE_ULONG(numhotneg, "Number of hot negative entries");
384 STATNODE_COUNTER(numneg_evicted,
385     "Number of negative entries evicted when adding a new entry");
386 STATNODE_COUNTER(shrinking_skipped,
387     "Number of times shrinking was already in progress");
388 
389 static void cache_zap_locked(struct namecache *ncp, bool neg_locked);
390 static int vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
391     char *buf, char **retbuf, u_int buflen);
392 
393 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
394 
395 static int cache_yield;
396 SYSCTL_INT(_vfs_cache, OID_AUTO, yield, CTLFLAG_RD, &cache_yield, 0,
397     "Number of times cache called yield");
398 
399 static void __noinline
400 cache_maybe_yield(void)
401 {
402 
403 	if (should_yield()) {
404 		cache_yield++;
405 		kern_yield(PRI_USER);
406 	}
407 }
408 
409 static inline void
410 cache_assert_vlp_locked(struct mtx *vlp)
411 {
412 
413 	if (vlp != NULL)
414 		mtx_assert(vlp, MA_OWNED);
415 }
416 
417 static inline void
418 cache_assert_vnode_locked(struct vnode *vp)
419 {
420 	struct mtx *vlp;
421 
422 	vlp = VP2VNODELOCK(vp);
423 	cache_assert_vlp_locked(vlp);
424 }
425 
426 static uint32_t
427 cache_get_hash(char *name, u_char len, struct vnode *dvp)
428 {
429 	uint32_t hash;
430 
431 	hash = fnv_32_buf(name, len, FNV1_32_INIT);
432 	hash = fnv_32_buf(&dvp, sizeof(dvp), hash);
433 	return (hash);
434 }
435 
436 static inline struct rwlock *
437 NCP2BUCKETLOCK(struct namecache *ncp)
438 {
439 	uint32_t hash;
440 
441 	hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
442 	return (HASH2BUCKETLOCK(hash));
443 }
444 
445 #ifdef INVARIANTS
446 static void
447 cache_assert_bucket_locked(struct namecache *ncp, int mode)
448 {
449 	struct rwlock *blp;
450 
451 	blp = NCP2BUCKETLOCK(ncp);
452 	rw_assert(blp, mode);
453 }
454 #else
455 #define cache_assert_bucket_locked(x, y) do { } while (0)
456 #endif
457 
458 #define cache_sort_vnodes(x, y)	_cache_sort_vnodes((void **)(x), (void **)(y))
459 static void
460 _cache_sort_vnodes(void **p1, void **p2)
461 {
462 	void *tmp;
463 
464 	MPASS(*p1 != NULL || *p2 != NULL);
465 
466 	if (*p1 > *p2) {
467 		tmp = *p2;
468 		*p2 = *p1;
469 		*p1 = tmp;
470 	}
471 }
472 
473 static void
474 cache_lock_all_buckets(void)
475 {
476 	u_int i;
477 
478 	for (i = 0; i < numbucketlocks; i++)
479 		rw_wlock(&bucketlocks[i]);
480 }
481 
482 static void
483 cache_unlock_all_buckets(void)
484 {
485 	u_int i;
486 
487 	for (i = 0; i < numbucketlocks; i++)
488 		rw_wunlock(&bucketlocks[i]);
489 }
490 
491 static void
492 cache_lock_all_vnodes(void)
493 {
494 	u_int i;
495 
496 	for (i = 0; i < numvnodelocks; i++)
497 		mtx_lock(&vnodelocks[i]);
498 }
499 
500 static void
501 cache_unlock_all_vnodes(void)
502 {
503 	u_int i;
504 
505 	for (i = 0; i < numvnodelocks; i++)
506 		mtx_unlock(&vnodelocks[i]);
507 }
508 
509 static int
510 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
511 {
512 
513 	cache_sort_vnodes(&vlp1, &vlp2);
514 
515 	if (vlp1 != NULL) {
516 		if (!mtx_trylock(vlp1))
517 			return (EAGAIN);
518 	}
519 	if (!mtx_trylock(vlp2)) {
520 		if (vlp1 != NULL)
521 			mtx_unlock(vlp1);
522 		return (EAGAIN);
523 	}
524 
525 	return (0);
526 }
527 
528 static void
529 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
530 {
531 
532 	MPASS(vlp1 != NULL || vlp2 != NULL);
533 	MPASS(vlp1 <= vlp2);
534 
535 	if (vlp1 != NULL)
536 		mtx_lock(vlp1);
537 	if (vlp2 != NULL)
538 		mtx_lock(vlp2);
539 }
540 
541 static void
542 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
543 {
544 
545 	MPASS(vlp1 != NULL || vlp2 != NULL);
546 
547 	if (vlp1 != NULL)
548 		mtx_unlock(vlp1);
549 	if (vlp2 != NULL)
550 		mtx_unlock(vlp2);
551 }
552 
553 static int
554 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
555 {
556 	struct nchstats snap;
557 
558 	if (req->oldptr == NULL)
559 		return (SYSCTL_OUT(req, 0, sizeof(snap)));
560 
561 	snap = nchstats;
562 	snap.ncs_goodhits = counter_u64_fetch(numposhits);
563 	snap.ncs_neghits = counter_u64_fetch(numneghits);
564 	snap.ncs_badhits = counter_u64_fetch(numposzaps) +
565 	    counter_u64_fetch(numnegzaps);
566 	snap.ncs_miss = counter_u64_fetch(nummisszap) +
567 	    counter_u64_fetch(nummiss);
568 
569 	return (SYSCTL_OUT(req, &snap, sizeof(snap)));
570 }
571 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
572     CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
573     "VFS cache effectiveness statistics");
574 
575 #ifdef DIAGNOSTIC
576 /*
577  * Grab an atomic snapshot of the name cache hash chain lengths
578  */
579 static SYSCTL_NODE(_debug, OID_AUTO, hashstat, CTLFLAG_RW, NULL,
580     "hash table stats");
581 
582 static int
583 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
584 {
585 	struct nchashhead *ncpp;
586 	struct namecache *ncp;
587 	int i, error, n_nchash, *cntbuf;
588 
589 retry:
590 	n_nchash = nchash + 1;	/* nchash is max index, not count */
591 	if (req->oldptr == NULL)
592 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
593 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
594 	cache_lock_all_buckets();
595 	if (n_nchash != nchash + 1) {
596 		cache_unlock_all_buckets();
597 		free(cntbuf, M_TEMP);
598 		goto retry;
599 	}
600 	/* Scan hash tables counting entries */
601 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
602 		LIST_FOREACH(ncp, ncpp, nc_hash)
603 			cntbuf[i]++;
604 	cache_unlock_all_buckets();
605 	for (error = 0, i = 0; i < n_nchash; i++)
606 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
607 			break;
608 	free(cntbuf, M_TEMP);
609 	return (error);
610 }
611 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
612     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
613     "nchash chain lengths");
614 
615 static int
616 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
617 {
618 	int error;
619 	struct nchashhead *ncpp;
620 	struct namecache *ncp;
621 	int n_nchash;
622 	int count, maxlength, used, pct;
623 
624 	if (!req->oldptr)
625 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
626 
627 	cache_lock_all_buckets();
628 	n_nchash = nchash + 1;	/* nchash is max index, not count */
629 	used = 0;
630 	maxlength = 0;
631 
632 	/* Scan hash tables for applicable entries */
633 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
634 		count = 0;
635 		LIST_FOREACH(ncp, ncpp, nc_hash) {
636 			count++;
637 		}
638 		if (count)
639 			used++;
640 		if (maxlength < count)
641 			maxlength = count;
642 	}
643 	n_nchash = nchash + 1;
644 	cache_unlock_all_buckets();
645 	pct = (used * 100) / (n_nchash / 100);
646 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
647 	if (error)
648 		return (error);
649 	error = SYSCTL_OUT(req, &used, sizeof(used));
650 	if (error)
651 		return (error);
652 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
653 	if (error)
654 		return (error);
655 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
656 	if (error)
657 		return (error);
658 	return (0);
659 }
660 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
661     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
662     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
663 #endif
664 
665 /*
666  * Negative entries management
667  *
668  * A variation of LRU scheme is used. New entries are hashed into one of
669  * numneglists cold lists. Entries get promoted to the hot list on first hit.
670  *
671  * The shrinker will demote hot list head and evict from the cold list in a
672  * round-robin manner.
673  */
674 static void
675 cache_negative_hit(struct namecache *ncp)
676 {
677 	struct neglist *neglist;
678 
679 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
680 	if (ncp->nc_flag & NCF_HOTNEGATIVE)
681 		return;
682 	neglist = NCP2NEGLIST(ncp);
683 	mtx_lock(&ncneg_hot.nl_lock);
684 	mtx_lock(&neglist->nl_lock);
685 	if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
686 		numhotneg++;
687 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
688 		TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
689 		ncp->nc_flag |= NCF_HOTNEGATIVE;
690 	}
691 	mtx_unlock(&neglist->nl_lock);
692 	mtx_unlock(&ncneg_hot.nl_lock);
693 }
694 
695 static void
696 cache_negative_insert(struct namecache *ncp, bool neg_locked)
697 {
698 	struct neglist *neglist;
699 
700 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
701 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
702 	neglist = NCP2NEGLIST(ncp);
703 	if (!neg_locked) {
704 		mtx_lock(&neglist->nl_lock);
705 	} else {
706 		mtx_assert(&neglist->nl_lock, MA_OWNED);
707 	}
708 	TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
709 	if (!neg_locked)
710 		mtx_unlock(&neglist->nl_lock);
711 	atomic_add_rel_long(&numneg, 1);
712 }
713 
714 static void
715 cache_negative_remove(struct namecache *ncp, bool neg_locked)
716 {
717 	struct neglist *neglist;
718 	bool hot_locked = false;
719 	bool list_locked = false;
720 
721 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
722 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
723 	neglist = NCP2NEGLIST(ncp);
724 	if (!neg_locked) {
725 		if (ncp->nc_flag & NCF_HOTNEGATIVE) {
726 			hot_locked = true;
727 			mtx_lock(&ncneg_hot.nl_lock);
728 			if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
729 				list_locked = true;
730 				mtx_lock(&neglist->nl_lock);
731 			}
732 		} else {
733 			list_locked = true;
734 			mtx_lock(&neglist->nl_lock);
735 		}
736 	}
737 	if (ncp->nc_flag & NCF_HOTNEGATIVE) {
738 		mtx_assert(&ncneg_hot.nl_lock, MA_OWNED);
739 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
740 		numhotneg--;
741 	} else {
742 		mtx_assert(&neglist->nl_lock, MA_OWNED);
743 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
744 	}
745 	if (list_locked)
746 		mtx_unlock(&neglist->nl_lock);
747 	if (hot_locked)
748 		mtx_unlock(&ncneg_hot.nl_lock);
749 	atomic_subtract_rel_long(&numneg, 1);
750 }
751 
752 static void
753 cache_negative_shrink_select(int start, struct namecache **ncpp,
754     struct neglist **neglistpp)
755 {
756 	struct neglist *neglist;
757 	struct namecache *ncp;
758 	int i;
759 
760 	*ncpp = ncp = NULL;
761 	neglist = NULL;
762 
763 	for (i = start; i < numneglists; i++) {
764 		neglist = &neglists[i];
765 		if (TAILQ_FIRST(&neglist->nl_list) == NULL)
766 			continue;
767 		mtx_lock(&neglist->nl_lock);
768 		ncp = TAILQ_FIRST(&neglist->nl_list);
769 		if (ncp != NULL)
770 			break;
771 		mtx_unlock(&neglist->nl_lock);
772 	}
773 
774 	*neglistpp = neglist;
775 	*ncpp = ncp;
776 }
777 
778 static void
779 cache_negative_zap_one(void)
780 {
781 	struct namecache *ncp, *ncp2;
782 	struct neglist *neglist;
783 	struct mtx *dvlp;
784 	struct rwlock *blp;
785 
786 	if (mtx_owner(&ncneg_shrink_lock) != NULL ||
787 	    !mtx_trylock(&ncneg_shrink_lock)) {
788 		counter_u64_add(shrinking_skipped, 1);
789 		return;
790 	}
791 
792 	mtx_lock(&ncneg_hot.nl_lock);
793 	ncp = TAILQ_FIRST(&ncneg_hot.nl_list);
794 	if (ncp != NULL) {
795 		neglist = NCP2NEGLIST(ncp);
796 		mtx_lock(&neglist->nl_lock);
797 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
798 		TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
799 		ncp->nc_flag &= ~NCF_HOTNEGATIVE;
800 		numhotneg--;
801 		mtx_unlock(&neglist->nl_lock);
802 	}
803 	mtx_unlock(&ncneg_hot.nl_lock);
804 
805 	cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
806 	shrink_list_turn++;
807 	if (shrink_list_turn == numneglists)
808 		shrink_list_turn = 0;
809 	if (ncp == NULL && shrink_list_turn == 0)
810 		cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
811 	mtx_unlock(&ncneg_shrink_lock);
812 	if (ncp == NULL)
813 		return;
814 
815 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
816 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
817 	blp = NCP2BUCKETLOCK(ncp);
818 	mtx_unlock(&neglist->nl_lock);
819 	mtx_lock(dvlp);
820 	rw_wlock(blp);
821 	mtx_lock(&neglist->nl_lock);
822 	ncp2 = TAILQ_FIRST(&neglist->nl_list);
823 	if (ncp != ncp2 || dvlp != VP2VNODELOCK(ncp2->nc_dvp) ||
824 	    blp != NCP2BUCKETLOCK(ncp2) || !(ncp2->nc_flag & NCF_NEGATIVE)) {
825 		ncp = NULL;
826 	} else {
827 		SDT_PROBE2(vfs, namecache, shrink_negative, done, ncp->nc_dvp,
828 		    ncp->nc_name);
829 
830 		cache_zap_locked(ncp, true);
831 		counter_u64_add(numneg_evicted, 1);
832 	}
833 	mtx_unlock(&neglist->nl_lock);
834 	rw_wunlock(blp);
835 	mtx_unlock(dvlp);
836 	cache_free(ncp);
837 }
838 
839 /*
840  * cache_zap_locked():
841  *
842  *   Removes a namecache entry from cache, whether it contains an actual
843  *   pointer to a vnode or if it is just a negative cache entry.
844  */
845 static void
846 cache_zap_locked(struct namecache *ncp, bool neg_locked)
847 {
848 
849 	if (!(ncp->nc_flag & NCF_NEGATIVE))
850 		cache_assert_vnode_locked(ncp->nc_vp);
851 	cache_assert_vnode_locked(ncp->nc_dvp);
852 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
853 
854 	CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp,
855 	    (ncp->nc_flag & NCF_NEGATIVE) ? NULL : ncp->nc_vp);
856 	LIST_REMOVE(ncp, nc_hash);
857 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
858 		SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
859 		    ncp->nc_name, ncp->nc_vp);
860 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
861 		if (ncp == ncp->nc_vp->v_cache_dd)
862 			ncp->nc_vp->v_cache_dd = NULL;
863 	} else {
864 		SDT_PROBE2(vfs, namecache, zap_negative, done, ncp->nc_dvp,
865 		    ncp->nc_name);
866 		cache_negative_remove(ncp, neg_locked);
867 	}
868 	if (ncp->nc_flag & NCF_ISDOTDOT) {
869 		if (ncp == ncp->nc_dvp->v_cache_dd)
870 			ncp->nc_dvp->v_cache_dd = NULL;
871 	} else {
872 		LIST_REMOVE(ncp, nc_src);
873 		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
874 			ncp->nc_flag |= NCF_DVDROP;
875 			counter_u64_add(numcachehv, -1);
876 		}
877 	}
878 	atomic_subtract_rel_long(&numcache, 1);
879 }
880 
881 static void
882 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
883 {
884 	struct rwlock *blp;
885 
886 	MPASS(ncp->nc_dvp == vp);
887 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
888 	cache_assert_vnode_locked(vp);
889 
890 	blp = NCP2BUCKETLOCK(ncp);
891 	rw_wlock(blp);
892 	cache_zap_locked(ncp, false);
893 	rw_wunlock(blp);
894 }
895 
896 static bool
897 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
898     struct mtx **vlpp)
899 {
900 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
901 	struct rwlock *blp;
902 
903 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
904 	cache_assert_vnode_locked(vp);
905 
906 	if (ncp->nc_flag & NCF_NEGATIVE) {
907 		if (*vlpp != NULL) {
908 			mtx_unlock(*vlpp);
909 			*vlpp = NULL;
910 		}
911 		cache_zap_negative_locked_vnode_kl(ncp, vp);
912 		return (true);
913 	}
914 
915 	pvlp = VP2VNODELOCK(vp);
916 	blp = NCP2BUCKETLOCK(ncp);
917 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
918 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
919 
920 	if (*vlpp == vlp1 || *vlpp == vlp2) {
921 		to_unlock = *vlpp;
922 		*vlpp = NULL;
923 	} else {
924 		if (*vlpp != NULL) {
925 			mtx_unlock(*vlpp);
926 			*vlpp = NULL;
927 		}
928 		cache_sort_vnodes(&vlp1, &vlp2);
929 		if (vlp1 == pvlp) {
930 			mtx_lock(vlp2);
931 			to_unlock = vlp2;
932 		} else {
933 			if (!mtx_trylock(vlp1))
934 				goto out_relock;
935 			to_unlock = vlp1;
936 		}
937 	}
938 	rw_wlock(blp);
939 	cache_zap_locked(ncp, false);
940 	rw_wunlock(blp);
941 	if (to_unlock != NULL)
942 		mtx_unlock(to_unlock);
943 	return (true);
944 
945 out_relock:
946 	mtx_unlock(vlp2);
947 	mtx_lock(vlp1);
948 	mtx_lock(vlp2);
949 	MPASS(*vlpp == NULL);
950 	*vlpp = vlp1;
951 	return (false);
952 }
953 
954 static int __noinline
955 cache_zap_locked_vnode(struct namecache *ncp, struct vnode *vp)
956 {
957 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
958 	struct rwlock *blp;
959 	int error = 0;
960 
961 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
962 	cache_assert_vnode_locked(vp);
963 
964 	pvlp = VP2VNODELOCK(vp);
965 	if (ncp->nc_flag & NCF_NEGATIVE) {
966 		cache_zap_negative_locked_vnode_kl(ncp, vp);
967 		goto out;
968 	}
969 
970 	blp = NCP2BUCKETLOCK(ncp);
971 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
972 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
973 	cache_sort_vnodes(&vlp1, &vlp2);
974 	if (vlp1 == pvlp) {
975 		mtx_lock(vlp2);
976 		to_unlock = vlp2;
977 	} else {
978 		if (!mtx_trylock(vlp1)) {
979 			error = EAGAIN;
980 			goto out;
981 		}
982 		to_unlock = vlp1;
983 	}
984 	rw_wlock(blp);
985 	cache_zap_locked(ncp, false);
986 	rw_wunlock(blp);
987 	mtx_unlock(to_unlock);
988 out:
989 	mtx_unlock(pvlp);
990 	return (error);
991 }
992 
993 /*
994  * If trylocking failed we can get here. We know enough to take all needed locks
995  * in the right order and re-lookup the entry.
996  */
997 static int
998 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
999     struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1000     struct rwlock *blp)
1001 {
1002 	struct namecache *rncp;
1003 
1004 	cache_assert_bucket_locked(ncp, RA_UNLOCKED);
1005 
1006 	cache_sort_vnodes(&dvlp, &vlp);
1007 	cache_lock_vnodes(dvlp, vlp);
1008 	rw_wlock(blp);
1009 	LIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1010 		if (rncp == ncp && rncp->nc_dvp == dvp &&
1011 		    rncp->nc_nlen == cnp->cn_namelen &&
1012 		    !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen))
1013 			break;
1014 	}
1015 	if (rncp != NULL) {
1016 		cache_zap_locked(rncp, false);
1017 		rw_wunlock(blp);
1018 		cache_unlock_vnodes(dvlp, vlp);
1019 		counter_u64_add(zap_and_exit_bucket_relock_success, 1);
1020 		return (0);
1021 	}
1022 
1023 	rw_wunlock(blp);
1024 	cache_unlock_vnodes(dvlp, vlp);
1025 	return (EAGAIN);
1026 }
1027 
1028 static int __noinline
1029 cache_zap_wlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1030     uint32_t hash, struct rwlock *blp)
1031 {
1032 	struct mtx *dvlp, *vlp;
1033 	struct vnode *dvp;
1034 
1035 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1036 
1037 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1038 	vlp = NULL;
1039 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1040 		vlp = VP2VNODELOCK(ncp->nc_vp);
1041 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1042 		cache_zap_locked(ncp, false);
1043 		rw_wunlock(blp);
1044 		cache_unlock_vnodes(dvlp, vlp);
1045 		return (0);
1046 	}
1047 
1048 	dvp = ncp->nc_dvp;
1049 	rw_wunlock(blp);
1050 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1051 }
1052 
1053 static int __noinline
1054 cache_zap_rlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1055     uint32_t hash, struct rwlock *blp)
1056 {
1057 	struct mtx *dvlp, *vlp;
1058 	struct vnode *dvp;
1059 
1060 	cache_assert_bucket_locked(ncp, RA_RLOCKED);
1061 
1062 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1063 	vlp = NULL;
1064 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1065 		vlp = VP2VNODELOCK(ncp->nc_vp);
1066 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1067 		rw_runlock(blp);
1068 		rw_wlock(blp);
1069 		cache_zap_locked(ncp, false);
1070 		rw_wunlock(blp);
1071 		cache_unlock_vnodes(dvlp, vlp);
1072 		return (0);
1073 	}
1074 
1075 	dvp = ncp->nc_dvp;
1076 	rw_runlock(blp);
1077 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1078 }
1079 
1080 static int
1081 cache_zap_wlocked_bucket_kl(struct namecache *ncp, struct rwlock *blp,
1082     struct mtx **vlpp1, struct mtx **vlpp2)
1083 {
1084 	struct mtx *dvlp, *vlp;
1085 
1086 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1087 
1088 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1089 	vlp = NULL;
1090 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1091 		vlp = VP2VNODELOCK(ncp->nc_vp);
1092 	cache_sort_vnodes(&dvlp, &vlp);
1093 
1094 	if (*vlpp1 == dvlp && *vlpp2 == vlp) {
1095 		cache_zap_locked(ncp, false);
1096 		cache_unlock_vnodes(dvlp, vlp);
1097 		*vlpp1 = NULL;
1098 		*vlpp2 = NULL;
1099 		return (0);
1100 	}
1101 
1102 	if (*vlpp1 != NULL)
1103 		mtx_unlock(*vlpp1);
1104 	if (*vlpp2 != NULL)
1105 		mtx_unlock(*vlpp2);
1106 	*vlpp1 = NULL;
1107 	*vlpp2 = NULL;
1108 
1109 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1110 		cache_zap_locked(ncp, false);
1111 		cache_unlock_vnodes(dvlp, vlp);
1112 		return (0);
1113 	}
1114 
1115 	rw_wunlock(blp);
1116 	*vlpp1 = dvlp;
1117 	*vlpp2 = vlp;
1118 	if (*vlpp1 != NULL)
1119 		mtx_lock(*vlpp1);
1120 	mtx_lock(*vlpp2);
1121 	rw_wlock(blp);
1122 	return (EAGAIN);
1123 }
1124 
1125 static void
1126 cache_lookup_unlock(struct rwlock *blp, struct mtx *vlp)
1127 {
1128 
1129 	if (blp != NULL) {
1130 		rw_runlock(blp);
1131 	} else {
1132 		mtx_unlock(vlp);
1133 	}
1134 }
1135 
1136 static int __noinline
1137 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1138     struct timespec *tsp, int *ticksp)
1139 {
1140 	int ltype;
1141 
1142 	*vpp = dvp;
1143 	CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .",
1144 			dvp, cnp->cn_nameptr);
1145 	counter_u64_add(dothits, 1);
1146 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1147 	if (tsp != NULL)
1148 		timespecclear(tsp);
1149 	if (ticksp != NULL)
1150 		*ticksp = ticks;
1151 	vrefact(*vpp);
1152 	/*
1153 	 * When we lookup "." we still can be asked to lock it
1154 	 * differently...
1155 	 */
1156 	ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1157 	if (ltype != VOP_ISLOCKED(*vpp)) {
1158 		if (ltype == LK_EXCLUSIVE) {
1159 			vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1160 			if (VN_IS_DOOMED((*vpp))) {
1161 				/* forced unmount */
1162 				vrele(*vpp);
1163 				*vpp = NULL;
1164 				return (ENOENT);
1165 			}
1166 		} else
1167 			vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1168 	}
1169 	return (-1);
1170 }
1171 
1172 static __noinline int
1173 cache_lookup_nomakeentry(struct vnode *dvp, struct vnode **vpp,
1174     struct componentname *cnp, struct timespec *tsp, int *ticksp)
1175 {
1176 	struct namecache *ncp;
1177 	struct rwlock *blp;
1178 	struct mtx *dvlp, *dvlp2;
1179 	uint32_t hash;
1180 	int error;
1181 
1182 	if (cnp->cn_namelen == 2 &&
1183 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1184 		counter_u64_add(dotdothits, 1);
1185 		dvlp = VP2VNODELOCK(dvp);
1186 		dvlp2 = NULL;
1187 		mtx_lock(dvlp);
1188 retry_dotdot:
1189 		ncp = dvp->v_cache_dd;
1190 		if (ncp == NULL) {
1191 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1192 			    "..", NULL);
1193 			mtx_unlock(dvlp);
1194 			if (dvlp2 != NULL)
1195 				mtx_unlock(dvlp2);
1196 			return (0);
1197 		}
1198 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1199 			if (ncp->nc_dvp != dvp)
1200 				panic("dvp %p v_cache_dd %p\n", dvp, ncp);
1201 			if (!cache_zap_locked_vnode_kl2(ncp,
1202 			    dvp, &dvlp2))
1203 				goto retry_dotdot;
1204 			MPASS(dvp->v_cache_dd == NULL);
1205 			mtx_unlock(dvlp);
1206 			if (dvlp2 != NULL)
1207 				mtx_unlock(dvlp2);
1208 			cache_free(ncp);
1209 		} else {
1210 			dvp->v_cache_dd = NULL;
1211 			mtx_unlock(dvlp);
1212 			if (dvlp2 != NULL)
1213 				mtx_unlock(dvlp2);
1214 		}
1215 		return (0);
1216 	}
1217 
1218 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1219 	blp = HASH2BUCKETLOCK(hash);
1220 retry:
1221 	if (LIST_EMPTY(NCHHASH(hash)))
1222 		goto out_no_entry;
1223 
1224 	rw_wlock(blp);
1225 
1226 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1227 		counter_u64_add(numchecks, 1);
1228 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1229 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1230 			break;
1231 	}
1232 
1233 	/* We failed to find an entry */
1234 	if (ncp == NULL) {
1235 		rw_wunlock(blp);
1236 		goto out_no_entry;
1237 	}
1238 
1239 	error = cache_zap_wlocked_bucket(ncp, cnp, hash, blp);
1240 	if (__predict_false(error != 0)) {
1241 		zap_and_exit_bucket_fail++;
1242 		cache_maybe_yield();
1243 		goto retry;
1244 	}
1245 	counter_u64_add(numposzaps, 1);
1246 	cache_free(ncp);
1247 	return (0);
1248 out_no_entry:
1249 	SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr, NULL);
1250 	counter_u64_add(nummisszap, 1);
1251 	return (0);
1252 }
1253 
1254 /**
1255  * Lookup a name in the name cache
1256  *
1257  * # Arguments
1258  *
1259  * - dvp:	Parent directory in which to search.
1260  * - vpp:	Return argument.  Will contain desired vnode on cache hit.
1261  * - cnp:	Parameters of the name search.  The most interesting bits of
1262  *   		the cn_flags field have the following meanings:
1263  *   	- MAKEENTRY:	If clear, free an entry from the cache rather than look
1264  *   			it up.
1265  *   	- ISDOTDOT:	Must be set if and only if cn_nameptr == ".."
1266  * - tsp:	Return storage for cache timestamp.  On a successful (positive
1267  *   		or negative) lookup, tsp will be filled with any timespec that
1268  *   		was stored when this cache entry was created.  However, it will
1269  *   		be clear for "." entries.
1270  * - ticks:	Return storage for alternate cache timestamp.  On a successful
1271  *   		(positive or negative) lookup, it will contain the ticks value
1272  *   		that was current when the cache entry was created, unless cnp
1273  *   		was ".".
1274  *
1275  * # Returns
1276  *
1277  * - -1:	A positive cache hit.  vpp will contain the desired vnode.
1278  * - ENOENT:	A negative cache hit, or dvp was recycled out from under us due
1279  *		to a forced unmount.  vpp will not be modified.  If the entry
1280  *		is a whiteout, then the ISWHITEOUT flag will be set in
1281  *		cnp->cn_flags.
1282  * - 0:		A cache miss.  vpp will not be modified.
1283  *
1284  * # Locking
1285  *
1286  * On a cache hit, vpp will be returned locked and ref'd.  If we're looking up
1287  * .., dvp is unlocked.  If we're looking up . an extra ref is taken, but the
1288  * lock is not recursively acquired.
1289  */
1290 int
1291 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1292     struct timespec *tsp, int *ticksp)
1293 {
1294 	struct namecache_ts *ncp_ts;
1295 	struct namecache *ncp;
1296 	struct rwlock *blp;
1297 	struct mtx *dvlp;
1298 	uint32_t hash;
1299 	enum vgetstate vs;
1300 	int error, ltype;
1301 
1302 #ifdef DEBUG_CACHE
1303 	if (__predict_false(!doingcache)) {
1304 		cnp->cn_flags &= ~MAKEENTRY;
1305 		return (0);
1306 	}
1307 #endif
1308 
1309 	counter_u64_add(numcalls, 1);
1310 
1311 	if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.'))
1312 		return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1313 
1314 	if ((cnp->cn_flags & MAKEENTRY) == 0)
1315 		return (cache_lookup_nomakeentry(dvp, vpp, cnp, tsp, ticksp));
1316 
1317 retry:
1318 	blp = NULL;
1319 	dvlp = NULL;
1320 	error = 0;
1321 	if (cnp->cn_namelen == 2 &&
1322 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1323 		counter_u64_add(dotdothits, 1);
1324 		dvlp = VP2VNODELOCK(dvp);
1325 		mtx_lock(dvlp);
1326 		ncp = dvp->v_cache_dd;
1327 		if (ncp == NULL) {
1328 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1329 			    "..", NULL);
1330 			mtx_unlock(dvlp);
1331 			return (0);
1332 		}
1333 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1334 			if (ncp->nc_flag & NCF_NEGATIVE)
1335 				*vpp = NULL;
1336 			else
1337 				*vpp = ncp->nc_vp;
1338 		} else
1339 			*vpp = ncp->nc_dvp;
1340 		/* Return failure if negative entry was found. */
1341 		if (*vpp == NULL)
1342 			goto negative_success;
1343 		CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
1344 		    dvp, cnp->cn_nameptr, *vpp);
1345 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..",
1346 		    *vpp);
1347 		cache_out_ts(ncp, tsp, ticksp);
1348 		if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1349 		    NCF_DTS && tsp != NULL) {
1350 			ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1351 			*tsp = ncp_ts->nc_dotdottime;
1352 		}
1353 		goto success;
1354 	}
1355 
1356 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1357 	blp = HASH2BUCKETLOCK(hash);
1358 	rw_rlock(blp);
1359 
1360 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1361 		counter_u64_add(numchecks, 1);
1362 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1363 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1364 			break;
1365 	}
1366 
1367 	/* We failed to find an entry */
1368 	if (__predict_false(ncp == NULL)) {
1369 		rw_runlock(blp);
1370 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1371 		    NULL);
1372 		counter_u64_add(nummiss, 1);
1373 		return (0);
1374 	}
1375 
1376 	if (ncp->nc_flag & NCF_NEGATIVE)
1377 		goto negative_success;
1378 
1379 	/* We found a "positive" match, return the vnode */
1380 	counter_u64_add(numposhits, 1);
1381 	*vpp = ncp->nc_vp;
1382 	CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p",
1383 	    dvp, cnp->cn_nameptr, *vpp, ncp);
1384 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name,
1385 	    *vpp);
1386 	cache_out_ts(ncp, tsp, ticksp);
1387 success:
1388 	/*
1389 	 * On success we return a locked and ref'd vnode as per the lookup
1390 	 * protocol.
1391 	 */
1392 	MPASS(dvp != *vpp);
1393 	ltype = 0;	/* silence gcc warning */
1394 	if (cnp->cn_flags & ISDOTDOT) {
1395 		ltype = VOP_ISLOCKED(dvp);
1396 		VOP_UNLOCK(dvp);
1397 	}
1398 	vs = vget_prep(*vpp);
1399 	cache_lookup_unlock(blp, dvlp);
1400 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1401 	if (cnp->cn_flags & ISDOTDOT) {
1402 		vn_lock(dvp, ltype | LK_RETRY);
1403 		if (VN_IS_DOOMED(dvp)) {
1404 			if (error == 0)
1405 				vput(*vpp);
1406 			*vpp = NULL;
1407 			return (ENOENT);
1408 		}
1409 	}
1410 	if (error) {
1411 		*vpp = NULL;
1412 		goto retry;
1413 	}
1414 	if ((cnp->cn_flags & ISLASTCN) &&
1415 	    (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) {
1416 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
1417 	}
1418 	return (-1);
1419 
1420 negative_success:
1421 	/* We found a negative match, and want to create it, so purge */
1422 	if (cnp->cn_nameiop == CREATE) {
1423 		counter_u64_add(numnegzaps, 1);
1424 		goto zap_and_exit;
1425 	}
1426 
1427 	counter_u64_add(numneghits, 1);
1428 	cache_negative_hit(ncp);
1429 	if (ncp->nc_flag & NCF_WHITE)
1430 		cnp->cn_flags |= ISWHITEOUT;
1431 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp,
1432 	    ncp->nc_name);
1433 	cache_out_ts(ncp, tsp, ticksp);
1434 	cache_lookup_unlock(blp, dvlp);
1435 	return (ENOENT);
1436 
1437 zap_and_exit:
1438 	if (blp != NULL)
1439 		error = cache_zap_rlocked_bucket(ncp, cnp, hash, blp);
1440 	else
1441 		error = cache_zap_locked_vnode(ncp, dvp);
1442 	if (__predict_false(error != 0)) {
1443 		zap_and_exit_bucket_fail2++;
1444 		cache_maybe_yield();
1445 		goto retry;
1446 	}
1447 	cache_free(ncp);
1448 	return (0);
1449 }
1450 
1451 struct celockstate {
1452 	struct mtx *vlp[3];
1453 	struct rwlock *blp[2];
1454 };
1455 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1456 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1457 
1458 static inline void
1459 cache_celockstate_init(struct celockstate *cel)
1460 {
1461 
1462 	bzero(cel, sizeof(*cel));
1463 }
1464 
1465 static void
1466 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1467     struct vnode *dvp)
1468 {
1469 	struct mtx *vlp1, *vlp2;
1470 
1471 	MPASS(cel->vlp[0] == NULL);
1472 	MPASS(cel->vlp[1] == NULL);
1473 	MPASS(cel->vlp[2] == NULL);
1474 
1475 	MPASS(vp != NULL || dvp != NULL);
1476 
1477 	vlp1 = VP2VNODELOCK(vp);
1478 	vlp2 = VP2VNODELOCK(dvp);
1479 	cache_sort_vnodes(&vlp1, &vlp2);
1480 
1481 	if (vlp1 != NULL) {
1482 		mtx_lock(vlp1);
1483 		cel->vlp[0] = vlp1;
1484 	}
1485 	mtx_lock(vlp2);
1486 	cel->vlp[1] = vlp2;
1487 }
1488 
1489 static void
1490 cache_unlock_vnodes_cel(struct celockstate *cel)
1491 {
1492 
1493 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
1494 
1495 	if (cel->vlp[0] != NULL)
1496 		mtx_unlock(cel->vlp[0]);
1497 	if (cel->vlp[1] != NULL)
1498 		mtx_unlock(cel->vlp[1]);
1499 	if (cel->vlp[2] != NULL)
1500 		mtx_unlock(cel->vlp[2]);
1501 }
1502 
1503 static bool
1504 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
1505 {
1506 	struct mtx *vlp;
1507 	bool ret;
1508 
1509 	cache_assert_vlp_locked(cel->vlp[0]);
1510 	cache_assert_vlp_locked(cel->vlp[1]);
1511 	MPASS(cel->vlp[2] == NULL);
1512 
1513 	MPASS(vp != NULL);
1514 	vlp = VP2VNODELOCK(vp);
1515 
1516 	ret = true;
1517 	if (vlp >= cel->vlp[1]) {
1518 		mtx_lock(vlp);
1519 	} else {
1520 		if (mtx_trylock(vlp))
1521 			goto out;
1522 		cache_lock_vnodes_cel_3_failures++;
1523 		cache_unlock_vnodes_cel(cel);
1524 		if (vlp < cel->vlp[0]) {
1525 			mtx_lock(vlp);
1526 			mtx_lock(cel->vlp[0]);
1527 			mtx_lock(cel->vlp[1]);
1528 		} else {
1529 			if (cel->vlp[0] != NULL)
1530 				mtx_lock(cel->vlp[0]);
1531 			mtx_lock(vlp);
1532 			mtx_lock(cel->vlp[1]);
1533 		}
1534 		ret = false;
1535 	}
1536 out:
1537 	cel->vlp[2] = vlp;
1538 	return (ret);
1539 }
1540 
1541 static void
1542 cache_lock_buckets_cel(struct celockstate *cel, struct rwlock *blp1,
1543     struct rwlock *blp2)
1544 {
1545 
1546 	MPASS(cel->blp[0] == NULL);
1547 	MPASS(cel->blp[1] == NULL);
1548 
1549 	cache_sort_vnodes(&blp1, &blp2);
1550 
1551 	if (blp1 != NULL) {
1552 		rw_wlock(blp1);
1553 		cel->blp[0] = blp1;
1554 	}
1555 	rw_wlock(blp2);
1556 	cel->blp[1] = blp2;
1557 }
1558 
1559 static void
1560 cache_unlock_buckets_cel(struct celockstate *cel)
1561 {
1562 
1563 	if (cel->blp[0] != NULL)
1564 		rw_wunlock(cel->blp[0]);
1565 	rw_wunlock(cel->blp[1]);
1566 }
1567 
1568 /*
1569  * Lock part of the cache affected by the insertion.
1570  *
1571  * This means vnodelocks for dvp, vp and the relevant bucketlock.
1572  * However, insertion can result in removal of an old entry. In this
1573  * case we have an additional vnode and bucketlock pair to lock. If the
1574  * entry is negative, ncelock is locked instead of the vnode.
1575  *
1576  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
1577  * preserving the locking order (smaller address first).
1578  */
1579 static void
1580 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1581     uint32_t hash)
1582 {
1583 	struct namecache *ncp;
1584 	struct rwlock *blps[2];
1585 
1586 	blps[0] = HASH2BUCKETLOCK(hash);
1587 	for (;;) {
1588 		blps[1] = NULL;
1589 		cache_lock_vnodes_cel(cel, dvp, vp);
1590 		if (vp == NULL || vp->v_type != VDIR)
1591 			break;
1592 		ncp = vp->v_cache_dd;
1593 		if (ncp == NULL)
1594 			break;
1595 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1596 			break;
1597 		MPASS(ncp->nc_dvp == vp);
1598 		blps[1] = NCP2BUCKETLOCK(ncp);
1599 		if (ncp->nc_flag & NCF_NEGATIVE)
1600 			break;
1601 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1602 			break;
1603 		/*
1604 		 * All vnodes got re-locked. Re-validate the state and if
1605 		 * nothing changed we are done. Otherwise restart.
1606 		 */
1607 		if (ncp == vp->v_cache_dd &&
1608 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1609 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1610 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1611 			break;
1612 		cache_unlock_vnodes_cel(cel);
1613 		cel->vlp[0] = NULL;
1614 		cel->vlp[1] = NULL;
1615 		cel->vlp[2] = NULL;
1616 	}
1617 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1618 }
1619 
1620 static void
1621 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1622     uint32_t hash)
1623 {
1624 	struct namecache *ncp;
1625 	struct rwlock *blps[2];
1626 
1627 	blps[0] = HASH2BUCKETLOCK(hash);
1628 	for (;;) {
1629 		blps[1] = NULL;
1630 		cache_lock_vnodes_cel(cel, dvp, vp);
1631 		ncp = dvp->v_cache_dd;
1632 		if (ncp == NULL)
1633 			break;
1634 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1635 			break;
1636 		MPASS(ncp->nc_dvp == dvp);
1637 		blps[1] = NCP2BUCKETLOCK(ncp);
1638 		if (ncp->nc_flag & NCF_NEGATIVE)
1639 			break;
1640 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1641 			break;
1642 		if (ncp == dvp->v_cache_dd &&
1643 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1644 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1645 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1646 			break;
1647 		cache_unlock_vnodes_cel(cel);
1648 		cel->vlp[0] = NULL;
1649 		cel->vlp[1] = NULL;
1650 		cel->vlp[2] = NULL;
1651 	}
1652 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1653 }
1654 
1655 static void
1656 cache_enter_unlock(struct celockstate *cel)
1657 {
1658 
1659 	cache_unlock_buckets_cel(cel);
1660 	cache_unlock_vnodes_cel(cel);
1661 }
1662 
1663 static void __noinline
1664 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp,
1665     struct componentname *cnp)
1666 {
1667 	struct celockstate cel;
1668 	struct namecache *ncp;
1669 	uint32_t hash;
1670 	int len;
1671 
1672 	if (dvp->v_cache_dd == NULL)
1673 		return;
1674 	len = cnp->cn_namelen;
1675 	cache_celockstate_init(&cel);
1676 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1677 	cache_enter_lock_dd(&cel, dvp, vp, hash);
1678 	ncp = dvp->v_cache_dd;
1679 	if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) {
1680 		KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent"));
1681 		cache_zap_locked(ncp, false);
1682 	} else {
1683 		ncp = NULL;
1684 	}
1685 	dvp->v_cache_dd = NULL;
1686 	cache_enter_unlock(&cel);
1687 	cache_free(ncp);
1688 }
1689 
1690 /*
1691  * Add an entry to the cache.
1692  */
1693 void
1694 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
1695     struct timespec *tsp, struct timespec *dtsp)
1696 {
1697 	struct celockstate cel;
1698 	struct namecache *ncp, *n2, *ndd;
1699 	struct namecache_ts *ncp_ts, *n2_ts;
1700 	struct nchashhead *ncpp;
1701 	uint32_t hash;
1702 	int flag;
1703 	int len;
1704 	u_long lnumcache;
1705 
1706 	CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr);
1707 	VNASSERT(vp == NULL || !VN_IS_DOOMED(vp), vp,
1708 	    ("cache_enter: Adding a doomed vnode"));
1709 	VNASSERT(dvp == NULL || !VN_IS_DOOMED(dvp), dvp,
1710 	    ("cache_enter: Doomed vnode used as src"));
1711 
1712 #ifdef DEBUG_CACHE
1713 	if (__predict_false(!doingcache))
1714 		return;
1715 #endif
1716 
1717 	flag = 0;
1718 	if (__predict_false(cnp->cn_nameptr[0] == '.')) {
1719 		if (cnp->cn_namelen == 1)
1720 			return;
1721 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1722 			cache_enter_dotdot_prep(dvp, vp, cnp);
1723 			flag = NCF_ISDOTDOT;
1724 		}
1725 	}
1726 
1727 	/*
1728 	 * Avoid blowout in namecache entries.
1729 	 */
1730 	lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
1731 	if (__predict_false(lnumcache >= ncsize)) {
1732 		atomic_add_long(&numcache, -1);
1733 		return;
1734 	}
1735 
1736 	cache_celockstate_init(&cel);
1737 	ndd = NULL;
1738 	ncp_ts = NULL;
1739 
1740 	/*
1741 	 * Calculate the hash key and setup as much of the new
1742 	 * namecache entry as possible before acquiring the lock.
1743 	 */
1744 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
1745 	ncp->nc_flag = flag;
1746 	ncp->nc_vp = vp;
1747 	if (vp == NULL)
1748 		ncp->nc_flag |= NCF_NEGATIVE;
1749 	ncp->nc_dvp = dvp;
1750 	if (tsp != NULL) {
1751 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1752 		ncp_ts->nc_time = *tsp;
1753 		ncp_ts->nc_ticks = ticks;
1754 		ncp_ts->nc_nc.nc_flag |= NCF_TS;
1755 		if (dtsp != NULL) {
1756 			ncp_ts->nc_dotdottime = *dtsp;
1757 			ncp_ts->nc_nc.nc_flag |= NCF_DTS;
1758 		}
1759 	}
1760 	len = ncp->nc_nlen = cnp->cn_namelen;
1761 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1762 	strlcpy(ncp->nc_name, cnp->cn_nameptr, len + 1);
1763 	cache_enter_lock(&cel, dvp, vp, hash);
1764 
1765 	/*
1766 	 * See if this vnode or negative entry is already in the cache
1767 	 * with this name.  This can happen with concurrent lookups of
1768 	 * the same path name.
1769 	 */
1770 	ncpp = NCHHASH(hash);
1771 	LIST_FOREACH(n2, ncpp, nc_hash) {
1772 		if (n2->nc_dvp == dvp &&
1773 		    n2->nc_nlen == cnp->cn_namelen &&
1774 		    !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
1775 			if (tsp != NULL) {
1776 				KASSERT((n2->nc_flag & NCF_TS) != 0,
1777 				    ("no NCF_TS"));
1778 				n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
1779 				n2_ts->nc_time = ncp_ts->nc_time;
1780 				n2_ts->nc_ticks = ncp_ts->nc_ticks;
1781 				if (dtsp != NULL) {
1782 					n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
1783 					if (ncp->nc_flag & NCF_NEGATIVE)
1784 						mtx_lock(&ncneg_hot.nl_lock);
1785 					n2_ts->nc_nc.nc_flag |= NCF_DTS;
1786 					if (ncp->nc_flag & NCF_NEGATIVE)
1787 						mtx_unlock(&ncneg_hot.nl_lock);
1788 				}
1789 			}
1790 			goto out_unlock_free;
1791 		}
1792 	}
1793 
1794 	if (flag == NCF_ISDOTDOT) {
1795 		/*
1796 		 * See if we are trying to add .. entry, but some other lookup
1797 		 * has populated v_cache_dd pointer already.
1798 		 */
1799 		if (dvp->v_cache_dd != NULL)
1800 			goto out_unlock_free;
1801 		KASSERT(vp == NULL || vp->v_type == VDIR,
1802 		    ("wrong vnode type %p", vp));
1803 		dvp->v_cache_dd = ncp;
1804 	}
1805 
1806 	if (vp != NULL) {
1807 		if (vp->v_type == VDIR) {
1808 			if (flag != NCF_ISDOTDOT) {
1809 				/*
1810 				 * For this case, the cache entry maps both the
1811 				 * directory name in it and the name ".." for the
1812 				 * directory's parent.
1813 				 */
1814 				if ((ndd = vp->v_cache_dd) != NULL) {
1815 					if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
1816 						cache_zap_locked(ndd, false);
1817 					else
1818 						ndd = NULL;
1819 				}
1820 				vp->v_cache_dd = ncp;
1821 			}
1822 		} else {
1823 			vp->v_cache_dd = NULL;
1824 		}
1825 	}
1826 
1827 	if (flag != NCF_ISDOTDOT) {
1828 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1829 			vhold(dvp);
1830 			counter_u64_add(numcachehv, 1);
1831 		}
1832 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
1833 	}
1834 
1835 	/*
1836 	 * Insert the new namecache entry into the appropriate chain
1837 	 * within the cache entries table.
1838 	 */
1839 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
1840 
1841 	/*
1842 	 * If the entry is "negative", we place it into the
1843 	 * "negative" cache queue, otherwise, we place it into the
1844 	 * destination vnode's cache entries queue.
1845 	 */
1846 	if (vp != NULL) {
1847 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
1848 		SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
1849 		    vp);
1850 	} else {
1851 		if (cnp->cn_flags & ISWHITEOUT)
1852 			ncp->nc_flag |= NCF_WHITE;
1853 		cache_negative_insert(ncp, false);
1854 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
1855 		    ncp->nc_name);
1856 	}
1857 	cache_enter_unlock(&cel);
1858 	if (numneg * ncnegfactor > lnumcache)
1859 		cache_negative_zap_one();
1860 	cache_free(ndd);
1861 	return;
1862 out_unlock_free:
1863 	cache_enter_unlock(&cel);
1864 	cache_free(ncp);
1865 	return;
1866 }
1867 
1868 static u_int
1869 cache_roundup_2(u_int val)
1870 {
1871 	u_int res;
1872 
1873 	for (res = 1; res <= val; res <<= 1)
1874 		continue;
1875 
1876 	return (res);
1877 }
1878 
1879 /*
1880  * Name cache initialization, from vfs_init() when we are booting
1881  */
1882 static void
1883 nchinit(void *dummy __unused)
1884 {
1885 	u_int i;
1886 
1887 	cache_zone_small = uma_zcreate("S VFS Cache",
1888 	    sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1,
1889 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1890 	    UMA_ZONE_ZINIT);
1891 	cache_zone_small_ts = uma_zcreate("STS VFS Cache",
1892 	    sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1,
1893 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1894 	    UMA_ZONE_ZINIT);
1895 	cache_zone_large = uma_zcreate("L VFS Cache",
1896 	    sizeof(struct namecache) + NAME_MAX + 1,
1897 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1898 	    UMA_ZONE_ZINIT);
1899 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache",
1900 	    sizeof(struct namecache_ts) + NAME_MAX + 1,
1901 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1902 	    UMA_ZONE_ZINIT);
1903 
1904 	ncsize = desiredvnodes * ncsizefactor;
1905 	nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash);
1906 	ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
1907 	if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
1908 		ncbuckethash = 7;
1909 	if (ncbuckethash > nchash)
1910 		ncbuckethash = nchash;
1911 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
1912 	    M_WAITOK | M_ZERO);
1913 	for (i = 0; i < numbucketlocks; i++)
1914 		rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK | RW_RECURSE);
1915 	ncvnodehash = ncbuckethash;
1916 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
1917 	    M_WAITOK | M_ZERO);
1918 	for (i = 0; i < numvnodelocks; i++)
1919 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
1920 	ncpurgeminvnodes = numbucketlocks * 2;
1921 
1922 	ncneghash = 3;
1923 	neglists = malloc(sizeof(*neglists) * numneglists, M_VFSCACHE,
1924 	    M_WAITOK | M_ZERO);
1925 	for (i = 0; i < numneglists; i++) {
1926 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
1927 		TAILQ_INIT(&neglists[i].nl_list);
1928 	}
1929 	mtx_init(&ncneg_hot.nl_lock, "ncneglh", NULL, MTX_DEF);
1930 	TAILQ_INIT(&ncneg_hot.nl_list);
1931 
1932 	mtx_init(&ncneg_shrink_lock, "ncnegs", NULL, MTX_DEF);
1933 
1934 	numcachehv = counter_u64_alloc(M_WAITOK);
1935 	numcalls = counter_u64_alloc(M_WAITOK);
1936 	dothits = counter_u64_alloc(M_WAITOK);
1937 	dotdothits = counter_u64_alloc(M_WAITOK);
1938 	numchecks = counter_u64_alloc(M_WAITOK);
1939 	nummiss = counter_u64_alloc(M_WAITOK);
1940 	nummisszap = counter_u64_alloc(M_WAITOK);
1941 	numposzaps = counter_u64_alloc(M_WAITOK);
1942 	numposhits = counter_u64_alloc(M_WAITOK);
1943 	numnegzaps = counter_u64_alloc(M_WAITOK);
1944 	numneghits = counter_u64_alloc(M_WAITOK);
1945 	numfullpathcalls = counter_u64_alloc(M_WAITOK);
1946 	numfullpathfail1 = counter_u64_alloc(M_WAITOK);
1947 	numfullpathfail2 = counter_u64_alloc(M_WAITOK);
1948 	numfullpathfail4 = counter_u64_alloc(M_WAITOK);
1949 	numfullpathfound = counter_u64_alloc(M_WAITOK);
1950 	zap_and_exit_bucket_relock_success = counter_u64_alloc(M_WAITOK);
1951 	numneg_evicted = counter_u64_alloc(M_WAITOK);
1952 	shrinking_skipped = counter_u64_alloc(M_WAITOK);
1953 }
1954 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
1955 
1956 void
1957 cache_changesize(u_long newmaxvnodes)
1958 {
1959 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
1960 	u_long new_nchash, old_nchash;
1961 	struct namecache *ncp;
1962 	uint32_t hash;
1963 	u_long newncsize;
1964 	int i;
1965 
1966 	newncsize = newmaxvnodes * ncsizefactor;
1967 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
1968 	if (newmaxvnodes < numbucketlocks)
1969 		newmaxvnodes = numbucketlocks;
1970 
1971 	new_nchashtbl = hashinit(newmaxvnodes, M_VFSCACHE, &new_nchash);
1972 	/* If same hash table size, nothing to do */
1973 	if (nchash == new_nchash) {
1974 		free(new_nchashtbl, M_VFSCACHE);
1975 		return;
1976 	}
1977 	/*
1978 	 * Move everything from the old hash table to the new table.
1979 	 * None of the namecache entries in the table can be removed
1980 	 * because to do so, they have to be removed from the hash table.
1981 	 */
1982 	cache_lock_all_vnodes();
1983 	cache_lock_all_buckets();
1984 	old_nchashtbl = nchashtbl;
1985 	old_nchash = nchash;
1986 	nchashtbl = new_nchashtbl;
1987 	nchash = new_nchash;
1988 	for (i = 0; i <= old_nchash; i++) {
1989 		while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) {
1990 			hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
1991 			    ncp->nc_dvp);
1992 			LIST_REMOVE(ncp, nc_hash);
1993 			LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
1994 		}
1995 	}
1996 	ncsize = newncsize;
1997 	cache_unlock_all_buckets();
1998 	cache_unlock_all_vnodes();
1999 	free(old_nchashtbl, M_VFSCACHE);
2000 }
2001 
2002 /*
2003  * Invalidate all entries from and to a particular vnode.
2004  */
2005 void
2006 cache_purge(struct vnode *vp)
2007 {
2008 	TAILQ_HEAD(, namecache) ncps;
2009 	struct namecache *ncp, *nnp;
2010 	struct mtx *vlp, *vlp2;
2011 
2012 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
2013 	SDT_PROBE1(vfs, namecache, purge, done, vp);
2014 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2015 	    vp->v_cache_dd == NULL)
2016 		return;
2017 	TAILQ_INIT(&ncps);
2018 	vlp = VP2VNODELOCK(vp);
2019 	vlp2 = NULL;
2020 	mtx_lock(vlp);
2021 retry:
2022 	while (!LIST_EMPTY(&vp->v_cache_src)) {
2023 		ncp = LIST_FIRST(&vp->v_cache_src);
2024 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2025 			goto retry;
2026 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2027 	}
2028 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2029 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
2030 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2031 			goto retry;
2032 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2033 	}
2034 	ncp = vp->v_cache_dd;
2035 	if (ncp != NULL) {
2036 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2037 		   ("lost dotdot link"));
2038 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2039 			goto retry;
2040 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2041 	}
2042 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2043 	mtx_unlock(vlp);
2044 	if (vlp2 != NULL)
2045 		mtx_unlock(vlp2);
2046 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2047 		cache_free(ncp);
2048 	}
2049 }
2050 
2051 /*
2052  * Invalidate all negative entries for a particular directory vnode.
2053  */
2054 void
2055 cache_purge_negative(struct vnode *vp)
2056 {
2057 	TAILQ_HEAD(, namecache) ncps;
2058 	struct namecache *ncp, *nnp;
2059 	struct mtx *vlp;
2060 
2061 	CTR1(KTR_VFS, "cache_purge_negative(%p)", vp);
2062 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2063 	if (LIST_EMPTY(&vp->v_cache_src))
2064 		return;
2065 	TAILQ_INIT(&ncps);
2066 	vlp = VP2VNODELOCK(vp);
2067 	mtx_lock(vlp);
2068 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2069 		if (!(ncp->nc_flag & NCF_NEGATIVE))
2070 			continue;
2071 		cache_zap_negative_locked_vnode_kl(ncp, vp);
2072 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2073 	}
2074 	mtx_unlock(vlp);
2075 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2076 		cache_free(ncp);
2077 	}
2078 }
2079 
2080 /*
2081  * Flush all entries referencing a particular filesystem.
2082  */
2083 void
2084 cache_purgevfs(struct mount *mp, bool force)
2085 {
2086 	TAILQ_HEAD(, namecache) ncps;
2087 	struct mtx *vlp1, *vlp2;
2088 	struct rwlock *blp;
2089 	struct nchashhead *bucket;
2090 	struct namecache *ncp, *nnp;
2091 	u_long i, j, n_nchash;
2092 	int error;
2093 
2094 	/* Scan hash tables for applicable entries */
2095 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2096 	if (!force && mp->mnt_nvnodelistsize <= ncpurgeminvnodes)
2097 		return;
2098 	TAILQ_INIT(&ncps);
2099 	n_nchash = nchash + 1;
2100 	vlp1 = vlp2 = NULL;
2101 	for (i = 0; i < numbucketlocks; i++) {
2102 		blp = (struct rwlock *)&bucketlocks[i];
2103 		rw_wlock(blp);
2104 		for (j = i; j < n_nchash; j += numbucketlocks) {
2105 retry:
2106 			bucket = &nchashtbl[j];
2107 			LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) {
2108 				cache_assert_bucket_locked(ncp, RA_WLOCKED);
2109 				if (ncp->nc_dvp->v_mount != mp)
2110 					continue;
2111 				error = cache_zap_wlocked_bucket_kl(ncp, blp,
2112 				    &vlp1, &vlp2);
2113 				if (error != 0)
2114 					goto retry;
2115 				TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst);
2116 			}
2117 		}
2118 		rw_wunlock(blp);
2119 		if (vlp1 == NULL && vlp2 == NULL)
2120 			cache_maybe_yield();
2121 	}
2122 	if (vlp1 != NULL)
2123 		mtx_unlock(vlp1);
2124 	if (vlp2 != NULL)
2125 		mtx_unlock(vlp2);
2126 
2127 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2128 		cache_free(ncp);
2129 	}
2130 }
2131 
2132 /*
2133  * Perform canonical checks and cache lookup and pass on to filesystem
2134  * through the vop_cachedlookup only if needed.
2135  */
2136 
2137 int
2138 vfs_cache_lookup(struct vop_lookup_args *ap)
2139 {
2140 	struct vnode *dvp;
2141 	int error;
2142 	struct vnode **vpp = ap->a_vpp;
2143 	struct componentname *cnp = ap->a_cnp;
2144 	struct ucred *cred = cnp->cn_cred;
2145 	int flags = cnp->cn_flags;
2146 	struct thread *td = cnp->cn_thread;
2147 
2148 	*vpp = NULL;
2149 	dvp = ap->a_dvp;
2150 
2151 	if (dvp->v_type != VDIR)
2152 		return (ENOTDIR);
2153 
2154 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2155 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2156 		return (EROFS);
2157 
2158 	error = VOP_ACCESS(dvp, VEXEC, cred, td);
2159 	if (error)
2160 		return (error);
2161 
2162 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2163 	if (error == 0)
2164 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2165 	if (error == -1)
2166 		return (0);
2167 	return (error);
2168 }
2169 
2170 /*
2171  * XXX All of these sysctls would probably be more productive dead.
2172  */
2173 static int __read_mostly disablecwd;
2174 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
2175    "Disable the getcwd syscall");
2176 
2177 /* Implementation of the getcwd syscall. */
2178 int
2179 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2180 {
2181 
2182 	return (kern___getcwd(td, uap->buf, UIO_USERSPACE, uap->buflen,
2183 	    MAXPATHLEN));
2184 }
2185 
2186 int
2187 kern___getcwd(struct thread *td, char *buf, enum uio_seg bufseg, size_t buflen,
2188     size_t path_max)
2189 {
2190 	char *bp, *tmpbuf;
2191 	struct filedesc *fdp;
2192 	struct vnode *cdir, *rdir;
2193 	int error;
2194 
2195 	if (__predict_false(disablecwd))
2196 		return (ENODEV);
2197 	if (__predict_false(buflen < 2))
2198 		return (EINVAL);
2199 	if (buflen > path_max)
2200 		buflen = path_max;
2201 
2202 	tmpbuf = malloc(buflen, M_TEMP, M_WAITOK);
2203 	fdp = td->td_proc->p_fd;
2204 	FILEDESC_SLOCK(fdp);
2205 	cdir = fdp->fd_cdir;
2206 	vrefact(cdir);
2207 	rdir = fdp->fd_rdir;
2208 	vrefact(rdir);
2209 	FILEDESC_SUNLOCK(fdp);
2210 	error = vn_fullpath1(td, cdir, rdir, tmpbuf, &bp, buflen);
2211 	vrele(rdir);
2212 	vrele(cdir);
2213 
2214 	if (!error) {
2215 		if (bufseg == UIO_SYSSPACE)
2216 			bcopy(bp, buf, strlen(bp) + 1);
2217 		else
2218 			error = copyout(bp, buf, strlen(bp) + 1);
2219 #ifdef KTRACE
2220 	if (KTRPOINT(curthread, KTR_NAMEI))
2221 		ktrnamei(bp);
2222 #endif
2223 	}
2224 	free(tmpbuf, M_TEMP);
2225 	return (error);
2226 }
2227 
2228 /*
2229  * Thus begins the fullpath magic.
2230  */
2231 
2232 static int __read_mostly disablefullpath;
2233 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, &disablefullpath, 0,
2234     "Disable the vn_fullpath function");
2235 
2236 /*
2237  * Retrieve the full filesystem path that correspond to a vnode from the name
2238  * cache (if available)
2239  */
2240 int
2241 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf)
2242 {
2243 	char *buf;
2244 	struct filedesc *fdp;
2245 	struct vnode *rdir;
2246 	int error;
2247 
2248 	if (__predict_false(disablefullpath))
2249 		return (ENODEV);
2250 	if (__predict_false(vn == NULL))
2251 		return (EINVAL);
2252 
2253 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2254 	fdp = td->td_proc->p_fd;
2255 	FILEDESC_SLOCK(fdp);
2256 	rdir = fdp->fd_rdir;
2257 	vrefact(rdir);
2258 	FILEDESC_SUNLOCK(fdp);
2259 	error = vn_fullpath1(td, vn, rdir, buf, retbuf, MAXPATHLEN);
2260 	vrele(rdir);
2261 
2262 	if (!error)
2263 		*freebuf = buf;
2264 	else
2265 		free(buf, M_TEMP);
2266 	return (error);
2267 }
2268 
2269 /*
2270  * This function is similar to vn_fullpath, but it attempts to lookup the
2271  * pathname relative to the global root mount point.  This is required for the
2272  * auditing sub-system, as audited pathnames must be absolute, relative to the
2273  * global root mount point.
2274  */
2275 int
2276 vn_fullpath_global(struct thread *td, struct vnode *vn,
2277     char **retbuf, char **freebuf)
2278 {
2279 	char *buf;
2280 	int error;
2281 
2282 	if (__predict_false(disablefullpath))
2283 		return (ENODEV);
2284 	if (__predict_false(vn == NULL))
2285 		return (EINVAL);
2286 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2287 	error = vn_fullpath1(td, vn, rootvnode, buf, retbuf, MAXPATHLEN);
2288 	if (!error)
2289 		*freebuf = buf;
2290 	else
2291 		free(buf, M_TEMP);
2292 	return (error);
2293 }
2294 
2295 int
2296 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen)
2297 {
2298 	struct vnode *dvp;
2299 	struct namecache *ncp;
2300 	struct mtx *vlp;
2301 	int error;
2302 
2303 	vlp = VP2VNODELOCK(*vp);
2304 	mtx_lock(vlp);
2305 	TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) {
2306 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2307 			break;
2308 	}
2309 	if (ncp != NULL) {
2310 		if (*buflen < ncp->nc_nlen) {
2311 			mtx_unlock(vlp);
2312 			vrele(*vp);
2313 			counter_u64_add(numfullpathfail4, 1);
2314 			error = ENOMEM;
2315 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2316 			    vp, NULL);
2317 			return (error);
2318 		}
2319 		*buflen -= ncp->nc_nlen;
2320 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2321 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2322 		    ncp->nc_name, vp);
2323 		dvp = *vp;
2324 		*vp = ncp->nc_dvp;
2325 		vref(*vp);
2326 		mtx_unlock(vlp);
2327 		vrele(dvp);
2328 		return (0);
2329 	}
2330 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2331 
2332 	mtx_unlock(vlp);
2333 	vn_lock(*vp, LK_SHARED | LK_RETRY);
2334 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
2335 	vput(*vp);
2336 	if (error) {
2337 		counter_u64_add(numfullpathfail2, 1);
2338 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
2339 		return (error);
2340 	}
2341 
2342 	*vp = dvp;
2343 	if (VN_IS_DOOMED(dvp)) {
2344 		/* forced unmount */
2345 		vrele(dvp);
2346 		error = ENOENT;
2347 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2348 		return (error);
2349 	}
2350 	/*
2351 	 * *vp has its use count incremented still.
2352 	 */
2353 
2354 	return (0);
2355 }
2356 
2357 /*
2358  * The magic behind kern___getcwd() and vn_fullpath().
2359  */
2360 static int
2361 vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
2362     char *buf, char **retbuf, u_int buflen)
2363 {
2364 	int error, slash_prefixed;
2365 #ifdef KDTRACE_HOOKS
2366 	struct vnode *startvp = vp;
2367 #endif
2368 	struct vnode *vp1;
2369 
2370 	buflen--;
2371 	buf[buflen] = '\0';
2372 	error = 0;
2373 	slash_prefixed = 0;
2374 
2375 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
2376 	counter_u64_add(numfullpathcalls, 1);
2377 	vref(vp);
2378 	if (vp->v_type != VDIR) {
2379 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2380 		if (error)
2381 			return (error);
2382 		if (buflen == 0) {
2383 			vrele(vp);
2384 			return (ENOMEM);
2385 		}
2386 		buf[--buflen] = '/';
2387 		slash_prefixed = 1;
2388 	}
2389 	while (vp != rdir && vp != rootvnode) {
2390 		/*
2391 		 * The vp vnode must be already fully constructed,
2392 		 * since it is either found in namecache or obtained
2393 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
2394 		 * without obtaining the vnode lock.
2395 		 */
2396 		if ((vp->v_vflag & VV_ROOT) != 0) {
2397 			vn_lock(vp, LK_RETRY | LK_SHARED);
2398 
2399 			/*
2400 			 * With the vnode locked, check for races with
2401 			 * unmount, forced or not.  Note that we
2402 			 * already verified that vp is not equal to
2403 			 * the root vnode, which means that
2404 			 * mnt_vnodecovered can be NULL only for the
2405 			 * case of unmount.
2406 			 */
2407 			if (VN_IS_DOOMED(vp) ||
2408 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
2409 			    vp1->v_mountedhere != vp->v_mount) {
2410 				vput(vp);
2411 				error = ENOENT;
2412 				SDT_PROBE3(vfs, namecache, fullpath, return,
2413 				    error, vp, NULL);
2414 				break;
2415 			}
2416 
2417 			vref(vp1);
2418 			vput(vp);
2419 			vp = vp1;
2420 			continue;
2421 		}
2422 		if (vp->v_type != VDIR) {
2423 			vrele(vp);
2424 			counter_u64_add(numfullpathfail1, 1);
2425 			error = ENOTDIR;
2426 			SDT_PROBE3(vfs, namecache, fullpath, return,
2427 			    error, vp, NULL);
2428 			break;
2429 		}
2430 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2431 		if (error)
2432 			break;
2433 		if (buflen == 0) {
2434 			vrele(vp);
2435 			error = ENOMEM;
2436 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2437 			    startvp, NULL);
2438 			break;
2439 		}
2440 		buf[--buflen] = '/';
2441 		slash_prefixed = 1;
2442 	}
2443 	if (error)
2444 		return (error);
2445 	if (!slash_prefixed) {
2446 		if (buflen == 0) {
2447 			vrele(vp);
2448 			counter_u64_add(numfullpathfail4, 1);
2449 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
2450 			    startvp, NULL);
2451 			return (ENOMEM);
2452 		}
2453 		buf[--buflen] = '/';
2454 	}
2455 	counter_u64_add(numfullpathfound, 1);
2456 	vrele(vp);
2457 
2458 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, buf + buflen);
2459 	*retbuf = buf + buflen;
2460 	return (0);
2461 }
2462 
2463 struct vnode *
2464 vn_dir_dd_ino(struct vnode *vp)
2465 {
2466 	struct namecache *ncp;
2467 	struct vnode *ddvp;
2468 	struct mtx *vlp;
2469 	enum vgetstate vs;
2470 
2471 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
2472 	vlp = VP2VNODELOCK(vp);
2473 	mtx_lock(vlp);
2474 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
2475 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
2476 			continue;
2477 		ddvp = ncp->nc_dvp;
2478 		vs = vget_prep(ddvp);
2479 		mtx_unlock(vlp);
2480 		if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
2481 			return (NULL);
2482 		return (ddvp);
2483 	}
2484 	mtx_unlock(vlp);
2485 	return (NULL);
2486 }
2487 
2488 int
2489 vn_commname(struct vnode *vp, char *buf, u_int buflen)
2490 {
2491 	struct namecache *ncp;
2492 	struct mtx *vlp;
2493 	int l;
2494 
2495 	vlp = VP2VNODELOCK(vp);
2496 	mtx_lock(vlp);
2497 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
2498 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2499 			break;
2500 	if (ncp == NULL) {
2501 		mtx_unlock(vlp);
2502 		return (ENOENT);
2503 	}
2504 	l = min(ncp->nc_nlen, buflen - 1);
2505 	memcpy(buf, ncp->nc_name, l);
2506 	mtx_unlock(vlp);
2507 	buf[l] = '\0';
2508 	return (0);
2509 }
2510 
2511 /*
2512  * This function updates path string to vnode's full global path
2513  * and checks the size of the new path string against the pathlen argument.
2514  *
2515  * Requires a locked, referenced vnode.
2516  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
2517  *
2518  * If sysctl debug.disablefullpath is set, ENODEV is returned,
2519  * vnode is left locked and path remain untouched.
2520  *
2521  * If vp is a directory, the call to vn_fullpath_global() always succeeds
2522  * because it falls back to the ".." lookup if the namecache lookup fails.
2523  */
2524 int
2525 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
2526     u_int pathlen)
2527 {
2528 	struct nameidata nd;
2529 	struct vnode *vp1;
2530 	char *rpath, *fbuf;
2531 	int error;
2532 
2533 	ASSERT_VOP_ELOCKED(vp, __func__);
2534 
2535 	/* Return ENODEV if sysctl debug.disablefullpath==1 */
2536 	if (__predict_false(disablefullpath))
2537 		return (ENODEV);
2538 
2539 	/* Construct global filesystem path from vp. */
2540 	VOP_UNLOCK(vp);
2541 	error = vn_fullpath_global(td, vp, &rpath, &fbuf);
2542 
2543 	if (error != 0) {
2544 		vrele(vp);
2545 		return (error);
2546 	}
2547 
2548 	if (strlen(rpath) >= pathlen) {
2549 		vrele(vp);
2550 		error = ENAMETOOLONG;
2551 		goto out;
2552 	}
2553 
2554 	/*
2555 	 * Re-lookup the vnode by path to detect a possible rename.
2556 	 * As a side effect, the vnode is relocked.
2557 	 * If vnode was renamed, return ENOENT.
2558 	 */
2559 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
2560 	    UIO_SYSSPACE, path, td);
2561 	error = namei(&nd);
2562 	if (error != 0) {
2563 		vrele(vp);
2564 		goto out;
2565 	}
2566 	NDFREE(&nd, NDF_ONLY_PNBUF);
2567 	vp1 = nd.ni_vp;
2568 	vrele(vp);
2569 	if (vp1 == vp)
2570 		strcpy(path, rpath);
2571 	else {
2572 		vput(vp1);
2573 		error = ENOENT;
2574 	}
2575 
2576 out:
2577 	free(fbuf, M_TEMP);
2578 	return (error);
2579 }
2580 
2581 #ifdef DDB
2582 static void
2583 db_print_vpath(struct vnode *vp)
2584 {
2585 
2586 	while (vp != NULL) {
2587 		db_printf("%p: ", vp);
2588 		if (vp == rootvnode) {
2589 			db_printf("/");
2590 			vp = NULL;
2591 		} else {
2592 			if (vp->v_vflag & VV_ROOT) {
2593 				db_printf("<mount point>");
2594 				vp = vp->v_mount->mnt_vnodecovered;
2595 			} else {
2596 				struct namecache *ncp;
2597 				char *ncn;
2598 				int i;
2599 
2600 				ncp = TAILQ_FIRST(&vp->v_cache_dst);
2601 				if (ncp != NULL) {
2602 					ncn = ncp->nc_name;
2603 					for (i = 0; i < ncp->nc_nlen; i++)
2604 						db_printf("%c", *ncn++);
2605 					vp = ncp->nc_dvp;
2606 				} else {
2607 					vp = NULL;
2608 				}
2609 			}
2610 		}
2611 		db_printf("\n");
2612 	}
2613 
2614 	return;
2615 }
2616 
2617 DB_SHOW_COMMAND(vpath, db_show_vpath)
2618 {
2619 	struct vnode *vp;
2620 
2621 	if (!have_addr) {
2622 		db_printf("usage: show vpath <struct vnode *>\n");
2623 		return;
2624 	}
2625 
2626 	vp = (struct vnode *)addr;
2627 	db_print_vpath(vp);
2628 }
2629 
2630 #endif
2631