xref: /freebsd/sys/kern/vfs_cache.c (revision 3c67996ca9a6231d5f737a32efe618b2ead7acc1)
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/smr.h>
60 #include <sys/smp.h>
61 #include <sys/syscallsubr.h>
62 #include <sys/sysctl.h>
63 #include <sys/sysproto.h>
64 #include <sys/vnode.h>
65 #include <ck_queue.h>
66 #ifdef KTRACE
67 #include <sys/ktrace.h>
68 #endif
69 
70 #ifdef DDB
71 #include <ddb/ddb.h>
72 #endif
73 
74 #include <vm/uma.h>
75 
76 SDT_PROVIDER_DECLARE(vfs);
77 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
78     "struct vnode *");
79 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
80     "char *");
81 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
82 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
83     "char *", "struct vnode *");
84 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
85 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
86     "struct vnode *", "char *");
87 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
88     "struct vnode *");
89 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
90     "struct vnode *", "char *");
91 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
92     "char *");
93 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *");
94 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
95 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
96 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
97     "struct vnode *");
98 SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *",
99     "char *");
100 SDT_PROBE_DEFINE2(vfs, namecache, shrink_negative, done, "struct vnode *",
101     "char *");
102 
103 /*
104  * This structure describes the elements in the cache of recent
105  * names looked up by namei.
106  */
107 
108 struct	namecache {
109 	CK_LIST_ENTRY(namecache) nc_hash;/* hash chain */
110 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
111 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
112 	struct	vnode *nc_dvp;		/* vnode of parent of name */
113 	union {
114 		struct	vnode *nu_vp;	/* vnode the name refers to */
115 	} n_un;
116 	u_char	nc_flag;		/* flag bits */
117 	u_char	nc_nlen;		/* length of name */
118 	char	nc_name[0];		/* segment name + nul */
119 };
120 
121 /*
122  * struct namecache_ts repeats struct namecache layout up to the
123  * nc_nlen member.
124  * struct namecache_ts is used in place of struct namecache when time(s) need
125  * to be stored.  The nc_dotdottime field is used when a cache entry is mapping
126  * both a non-dotdot directory name plus dotdot for the directory's
127  * parent.
128  */
129 struct	namecache_ts {
130 	struct	timespec nc_time;	/* timespec provided by fs */
131 	struct	timespec nc_dotdottime;	/* dotdot timespec provided by fs */
132 	int	nc_ticks;		/* ticks value when entry was added */
133 	struct namecache nc_nc;
134 };
135 
136 #define	nc_vp		n_un.nu_vp
137 
138 /*
139  * Flags in namecache.nc_flag
140  */
141 #define NCF_WHITE	0x01
142 #define NCF_ISDOTDOT	0x02
143 #define	NCF_TS		0x04
144 #define	NCF_DTS		0x08
145 #define	NCF_DVDROP	0x10
146 #define	NCF_NEGATIVE	0x20
147 #define	NCF_HOTNEGATIVE	0x40
148 #define NCF_INVALID	0x80
149 
150 /*
151  * Mark an entry as invalid.
152  *
153  * This is called before it starts getting deconstructed.
154  */
155 static void
156 cache_ncp_invalidate(struct namecache *ncp)
157 {
158 
159 	KASSERT((ncp->nc_flag & NCF_INVALID) == 0,
160 	    ("%s: entry %p already invalid", __func__, ncp));
161 	ncp->nc_flag |= NCF_INVALID;
162 	atomic_thread_fence_rel();
163 }
164 
165 /*
166  * Verify validity of an entry.
167  *
168  * All places which elide locks are supposed to call this after they are
169  * done with reading from an entry.
170  */
171 static bool
172 cache_ncp_invalid(struct namecache *ncp)
173 {
174 
175 	atomic_thread_fence_acq();
176 	return ((ncp->nc_flag & NCF_INVALID) != 0);
177 }
178 
179 /*
180  * Name caching works as follows:
181  *
182  * Names found by directory scans are retained in a cache
183  * for future reference.  It is managed LRU, so frequently
184  * used names will hang around.  Cache is indexed by hash value
185  * obtained from (dvp, name) where dvp refers to the directory
186  * containing name.
187  *
188  * If it is a "negative" entry, (i.e. for a name that is known NOT to
189  * exist) the vnode pointer will be NULL.
190  *
191  * Upon reaching the last segment of a path, if the reference
192  * is for DELETE, or NOCACHE is set (rewrite), and the
193  * name is located in the cache, it will be dropped.
194  *
195  * These locks are used (in the order in which they can be taken):
196  * NAME		TYPE	ROLE
197  * vnodelock	mtx	vnode lists and v_cache_dd field protection
198  * bucketlock	rwlock	for access to given set of hash buckets
199  * neglist	mtx	negative entry LRU management
200  *
201  * Additionally, ncneg_shrink_lock mtx is used to have at most one thread
202  * shrinking the LRU list.
203  *
204  * It is legal to take multiple vnodelock and bucketlock locks. The locking
205  * order is lower address first. Both are recursive.
206  *
207  * "." lookups are lockless.
208  *
209  * ".." and vnode -> name lookups require vnodelock.
210  *
211  * name -> vnode lookup requires the relevant bucketlock to be held for reading.
212  *
213  * Insertions and removals of entries require involved vnodes and bucketlocks
214  * to be write-locked to prevent other threads from seeing the entry.
215  *
216  * Some lookups result in removal of the found entry (e.g. getting rid of a
217  * negative entry with the intent to create a positive one), which poses a
218  * problem when multiple threads reach the state. Similarly, two different
219  * threads can purge two different vnodes and try to remove the same name.
220  *
221  * If the already held vnode lock is lower than the second required lock, we
222  * can just take the other lock. However, in the opposite case, this could
223  * deadlock. As such, this is resolved by trylocking and if that fails unlocking
224  * the first node, locking everything in order and revalidating the state.
225  */
226 
227 VFS_SMR_DECLARE;
228 
229 /*
230  * Structures associated with name caching.
231  */
232 #define NCHHASH(hash) \
233 	(&nchashtbl[(hash) & nchash])
234 static __read_mostly CK_LIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */
235 static u_long __read_mostly	nchash;			/* size of hash table */
236 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
237     "Size of namecache hash table");
238 static u_long __read_mostly	ncnegfactor = 5; /* ratio of negative entries */
239 SYSCTL_ULONG(_vfs, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
240     "Ratio of negative namecache entries");
241 static u_long __exclusive_cache_line	numneg;	/* number of negative entries allocated */
242 static u_long __exclusive_cache_line	numcache;/* number of cache entries allocated */
243 u_int ncsizefactor = 2;
244 SYSCTL_UINT(_vfs, OID_AUTO, ncsizefactor, CTLFLAG_RW, &ncsizefactor, 0,
245     "Size factor for namecache");
246 static u_int __read_mostly	ncpurgeminvnodes;
247 SYSCTL_UINT(_vfs, OID_AUTO, ncpurgeminvnodes, CTLFLAG_RW, &ncpurgeminvnodes, 0,
248     "Number of vnodes below which purgevfs ignores the request");
249 static u_int __read_mostly	ncsize; /* the size as computed on creation or resizing */
250 
251 struct nchstats	nchstats;		/* cache effectiveness statistics */
252 
253 static struct mtx __exclusive_cache_line	ncneg_shrink_lock;
254 static int	shrink_list_turn;
255 
256 struct neglist {
257 	struct mtx		nl_lock;
258 	TAILQ_HEAD(, namecache) nl_list;
259 } __aligned(CACHE_LINE_SIZE);
260 
261 static struct neglist __read_mostly	*neglists;
262 static struct neglist ncneg_hot;
263 static u_long numhotneg;
264 
265 #define	numneglists (ncneghash + 1)
266 static u_int __read_mostly	ncneghash;
267 static inline struct neglist *
268 NCP2NEGLIST(struct namecache *ncp)
269 {
270 
271 	return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
272 }
273 
274 #define	numbucketlocks (ncbuckethash + 1)
275 static u_int __read_mostly  ncbuckethash;
276 static struct rwlock_padalign __read_mostly  *bucketlocks;
277 #define	HASH2BUCKETLOCK(hash) \
278 	((struct rwlock *)(&bucketlocks[((hash) & ncbuckethash)]))
279 
280 #define	numvnodelocks (ncvnodehash + 1)
281 static u_int __read_mostly  ncvnodehash;
282 static struct mtx __read_mostly *vnodelocks;
283 static inline struct mtx *
284 VP2VNODELOCK(struct vnode *vp)
285 {
286 
287 	return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
288 }
289 
290 /*
291  * UMA zones for the VFS cache.
292  *
293  * The small cache is used for entries with short names, which are the
294  * most common.  The large cache is used for entries which are too big to
295  * fit in the small cache.
296  */
297 static uma_zone_t __read_mostly cache_zone_small;
298 static uma_zone_t __read_mostly cache_zone_small_ts;
299 static uma_zone_t __read_mostly cache_zone_large;
300 static uma_zone_t __read_mostly cache_zone_large_ts;
301 
302 #define	CACHE_PATH_CUTOFF	35
303 
304 static struct namecache *
305 cache_alloc(int len, int ts)
306 {
307 	struct namecache_ts *ncp_ts;
308 	struct namecache *ncp;
309 
310 	if (__predict_false(ts)) {
311 		if (len <= CACHE_PATH_CUTOFF)
312 			ncp_ts = uma_zalloc_smr(cache_zone_small_ts, M_WAITOK);
313 		else
314 			ncp_ts = uma_zalloc_smr(cache_zone_large_ts, M_WAITOK);
315 		ncp = &ncp_ts->nc_nc;
316 	} else {
317 		if (len <= CACHE_PATH_CUTOFF)
318 			ncp = uma_zalloc_smr(cache_zone_small, M_WAITOK);
319 		else
320 			ncp = uma_zalloc_smr(cache_zone_large, M_WAITOK);
321 	}
322 	return (ncp);
323 }
324 
325 static void
326 cache_free(struct namecache *ncp)
327 {
328 	struct namecache_ts *ncp_ts;
329 
330 	if (ncp == NULL)
331 		return;
332 	if ((ncp->nc_flag & NCF_DVDROP) != 0)
333 		vdrop(ncp->nc_dvp);
334 	if (__predict_false(ncp->nc_flag & NCF_TS)) {
335 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
336 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
337 			uma_zfree_smr(cache_zone_small_ts, ncp_ts);
338 		else
339 			uma_zfree_smr(cache_zone_large_ts, ncp_ts);
340 	} else {
341 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
342 			uma_zfree_smr(cache_zone_small, ncp);
343 		else
344 			uma_zfree_smr(cache_zone_large, ncp);
345 	}
346 }
347 
348 static void
349 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
350 {
351 	struct namecache_ts *ncp_ts;
352 
353 	KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
354 	    (tsp == NULL && ticksp == NULL),
355 	    ("No NCF_TS"));
356 
357 	if (tsp == NULL && ticksp == NULL)
358 		return;
359 
360 	ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
361 	if (tsp != NULL)
362 		*tsp = ncp_ts->nc_time;
363 	if (ticksp != NULL)
364 		*ticksp = ncp_ts->nc_ticks;
365 }
366 
367 #ifdef DEBUG_CACHE
368 static int __read_mostly	doingcache = 1;	/* 1 => enable the cache */
369 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
370     "VFS namecache enabled");
371 #endif
372 
373 /* Export size information to userland */
374 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
375     sizeof(struct namecache), "sizeof(struct namecache)");
376 
377 /*
378  * The new name cache statistics
379  */
380 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
381     "Name cache statistics");
382 #define STATNODE_ULONG(name, descr)					\
383 	SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, descr);
384 #define STATNODE_COUNTER(name, descr)					\
385 	static COUNTER_U64_DEFINE_EARLY(name);				\
386 	SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, \
387 	    descr);
388 STATNODE_ULONG(numneg, "Number of negative cache entries");
389 STATNODE_ULONG(numcache, "Number of cache entries");
390 STATNODE_COUNTER(numcachehv, "Number of namecache entries with vnodes held");
391 STATNODE_COUNTER(numcalls, "Number of cache lookups");
392 STATNODE_COUNTER(dothits, "Number of '.' hits");
393 STATNODE_COUNTER(dotdothits, "Number of '..' hits");
394 STATNODE_COUNTER(numchecks, "Number of checks in lookup");
395 STATNODE_COUNTER(nummiss, "Number of cache misses");
396 STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache");
397 STATNODE_COUNTER(numposzaps,
398     "Number of cache hits (positive) we do not want to cache");
399 STATNODE_COUNTER(numposhits, "Number of cache hits (positive)");
400 STATNODE_COUNTER(numnegzaps,
401     "Number of cache hits (negative) we do not want to cache");
402 STATNODE_COUNTER(numneghits, "Number of cache hits (negative)");
403 /* These count for vn_getcwd(), too. */
404 STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls");
405 STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
406 STATNODE_COUNTER(numfullpathfail2,
407     "Number of fullpath search errors (VOP_VPTOCNP failures)");
408 STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
409 STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls");
410 STATNODE_COUNTER(zap_and_exit_bucket_relock_success,
411     "Number of successful removals after relocking");
412 static long zap_and_exit_bucket_fail; STATNODE_ULONG(zap_and_exit_bucket_fail,
413     "Number of times zap_and_exit failed to lock");
414 static long zap_and_exit_bucket_fail2; STATNODE_ULONG(zap_and_exit_bucket_fail2,
415     "Number of times zap_and_exit failed to lock");
416 static long cache_lock_vnodes_cel_3_failures;
417 STATNODE_ULONG(cache_lock_vnodes_cel_3_failures,
418     "Number of times 3-way vnode locking failed");
419 STATNODE_ULONG(numhotneg, "Number of hot negative entries");
420 STATNODE_COUNTER(numneg_evicted,
421     "Number of negative entries evicted when adding a new entry");
422 STATNODE_COUNTER(shrinking_skipped,
423     "Number of times shrinking was already in progress");
424 
425 static void cache_zap_locked(struct namecache *ncp, bool neg_locked);
426 static int vn_fullpath_hardlink(struct thread *td, struct nameidata *ndp, char **retbuf,
427     char **freebuf, size_t *buflen);
428 static int vn_fullpath_any(struct thread *td, struct vnode *vp, struct vnode *rdir,
429     char *buf, char **retbuf, size_t *buflen);
430 static int vn_fullpath_dir(struct thread *td, struct vnode *vp, struct vnode *rdir,
431     char *buf, char **retbuf, size_t *len, bool slash_prefixed, size_t addend);
432 
433 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
434 
435 static int cache_yield;
436 SYSCTL_INT(_vfs_cache, OID_AUTO, yield, CTLFLAG_RD, &cache_yield, 0,
437     "Number of times cache called yield");
438 
439 static void __noinline
440 cache_maybe_yield(void)
441 {
442 
443 	if (should_yield()) {
444 		cache_yield++;
445 		kern_yield(PRI_USER);
446 	}
447 }
448 
449 static inline void
450 cache_assert_vlp_locked(struct mtx *vlp)
451 {
452 
453 	if (vlp != NULL)
454 		mtx_assert(vlp, MA_OWNED);
455 }
456 
457 static inline void
458 cache_assert_vnode_locked(struct vnode *vp)
459 {
460 	struct mtx *vlp;
461 
462 	vlp = VP2VNODELOCK(vp);
463 	cache_assert_vlp_locked(vlp);
464 }
465 
466 static uint32_t
467 cache_get_hash(char *name, u_char len, struct vnode *dvp)
468 {
469 	uint32_t hash;
470 
471 	hash = fnv_32_buf(name, len, FNV1_32_INIT);
472 	hash = fnv_32_buf(&dvp, sizeof(dvp), hash);
473 	return (hash);
474 }
475 
476 static inline struct rwlock *
477 NCP2BUCKETLOCK(struct namecache *ncp)
478 {
479 	uint32_t hash;
480 
481 	hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
482 	return (HASH2BUCKETLOCK(hash));
483 }
484 
485 #ifdef INVARIANTS
486 static void
487 cache_assert_bucket_locked(struct namecache *ncp, int mode)
488 {
489 	struct rwlock *blp;
490 
491 	blp = NCP2BUCKETLOCK(ncp);
492 	rw_assert(blp, mode);
493 }
494 #else
495 #define cache_assert_bucket_locked(x, y) do { } while (0)
496 #endif
497 
498 #define cache_sort_vnodes(x, y)	_cache_sort_vnodes((void **)(x), (void **)(y))
499 static void
500 _cache_sort_vnodes(void **p1, void **p2)
501 {
502 	void *tmp;
503 
504 	MPASS(*p1 != NULL || *p2 != NULL);
505 
506 	if (*p1 > *p2) {
507 		tmp = *p2;
508 		*p2 = *p1;
509 		*p1 = tmp;
510 	}
511 }
512 
513 static void
514 cache_lock_all_buckets(void)
515 {
516 	u_int i;
517 
518 	for (i = 0; i < numbucketlocks; i++)
519 		rw_wlock(&bucketlocks[i]);
520 }
521 
522 static void
523 cache_unlock_all_buckets(void)
524 {
525 	u_int i;
526 
527 	for (i = 0; i < numbucketlocks; i++)
528 		rw_wunlock(&bucketlocks[i]);
529 }
530 
531 static void
532 cache_lock_all_vnodes(void)
533 {
534 	u_int i;
535 
536 	for (i = 0; i < numvnodelocks; i++)
537 		mtx_lock(&vnodelocks[i]);
538 }
539 
540 static void
541 cache_unlock_all_vnodes(void)
542 {
543 	u_int i;
544 
545 	for (i = 0; i < numvnodelocks; i++)
546 		mtx_unlock(&vnodelocks[i]);
547 }
548 
549 static int
550 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
551 {
552 
553 	cache_sort_vnodes(&vlp1, &vlp2);
554 
555 	if (vlp1 != NULL) {
556 		if (!mtx_trylock(vlp1))
557 			return (EAGAIN);
558 	}
559 	if (!mtx_trylock(vlp2)) {
560 		if (vlp1 != NULL)
561 			mtx_unlock(vlp1);
562 		return (EAGAIN);
563 	}
564 
565 	return (0);
566 }
567 
568 static void
569 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
570 {
571 
572 	MPASS(vlp1 != NULL || vlp2 != NULL);
573 	MPASS(vlp1 <= vlp2);
574 
575 	if (vlp1 != NULL)
576 		mtx_lock(vlp1);
577 	if (vlp2 != NULL)
578 		mtx_lock(vlp2);
579 }
580 
581 static void
582 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
583 {
584 
585 	MPASS(vlp1 != NULL || vlp2 != NULL);
586 
587 	if (vlp1 != NULL)
588 		mtx_unlock(vlp1);
589 	if (vlp2 != NULL)
590 		mtx_unlock(vlp2);
591 }
592 
593 static int
594 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
595 {
596 	struct nchstats snap;
597 
598 	if (req->oldptr == NULL)
599 		return (SYSCTL_OUT(req, 0, sizeof(snap)));
600 
601 	snap = nchstats;
602 	snap.ncs_goodhits = counter_u64_fetch(numposhits);
603 	snap.ncs_neghits = counter_u64_fetch(numneghits);
604 	snap.ncs_badhits = counter_u64_fetch(numposzaps) +
605 	    counter_u64_fetch(numnegzaps);
606 	snap.ncs_miss = counter_u64_fetch(nummisszap) +
607 	    counter_u64_fetch(nummiss);
608 
609 	return (SYSCTL_OUT(req, &snap, sizeof(snap)));
610 }
611 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
612     CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
613     "VFS cache effectiveness statistics");
614 
615 #ifdef DIAGNOSTIC
616 /*
617  * Grab an atomic snapshot of the name cache hash chain lengths
618  */
619 static SYSCTL_NODE(_debug, OID_AUTO, hashstat,
620     CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
621     "hash table stats");
622 
623 static int
624 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
625 {
626 	struct nchashhead *ncpp;
627 	struct namecache *ncp;
628 	int i, error, n_nchash, *cntbuf;
629 
630 retry:
631 	n_nchash = nchash + 1;	/* nchash is max index, not count */
632 	if (req->oldptr == NULL)
633 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
634 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
635 	cache_lock_all_buckets();
636 	if (n_nchash != nchash + 1) {
637 		cache_unlock_all_buckets();
638 		free(cntbuf, M_TEMP);
639 		goto retry;
640 	}
641 	/* Scan hash tables counting entries */
642 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
643 		CK_LIST_FOREACH(ncp, ncpp, nc_hash)
644 			cntbuf[i]++;
645 	cache_unlock_all_buckets();
646 	for (error = 0, i = 0; i < n_nchash; i++)
647 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
648 			break;
649 	free(cntbuf, M_TEMP);
650 	return (error);
651 }
652 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
653     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
654     "nchash chain lengths");
655 
656 static int
657 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
658 {
659 	int error;
660 	struct nchashhead *ncpp;
661 	struct namecache *ncp;
662 	int n_nchash;
663 	int count, maxlength, used, pct;
664 
665 	if (!req->oldptr)
666 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
667 
668 	cache_lock_all_buckets();
669 	n_nchash = nchash + 1;	/* nchash is max index, not count */
670 	used = 0;
671 	maxlength = 0;
672 
673 	/* Scan hash tables for applicable entries */
674 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
675 		count = 0;
676 		CK_LIST_FOREACH(ncp, ncpp, nc_hash) {
677 			count++;
678 		}
679 		if (count)
680 			used++;
681 		if (maxlength < count)
682 			maxlength = count;
683 	}
684 	n_nchash = nchash + 1;
685 	cache_unlock_all_buckets();
686 	pct = (used * 100) / (n_nchash / 100);
687 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
688 	if (error)
689 		return (error);
690 	error = SYSCTL_OUT(req, &used, sizeof(used));
691 	if (error)
692 		return (error);
693 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
694 	if (error)
695 		return (error);
696 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
697 	if (error)
698 		return (error);
699 	return (0);
700 }
701 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
702     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
703     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
704 #endif
705 
706 /*
707  * Negative entries management
708  *
709  * A variation of LRU scheme is used. New entries are hashed into one of
710  * numneglists cold lists. Entries get promoted to the hot list on first hit.
711  *
712  * The shrinker will demote hot list head and evict from the cold list in a
713  * round-robin manner.
714  */
715 static void
716 cache_negative_hit(struct namecache *ncp)
717 {
718 	struct neglist *neglist;
719 
720 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
721 	if (ncp->nc_flag & NCF_HOTNEGATIVE)
722 		return;
723 	neglist = NCP2NEGLIST(ncp);
724 	mtx_lock(&ncneg_hot.nl_lock);
725 	mtx_lock(&neglist->nl_lock);
726 	if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
727 		numhotneg++;
728 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
729 		TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
730 		ncp->nc_flag |= NCF_HOTNEGATIVE;
731 	}
732 	mtx_unlock(&neglist->nl_lock);
733 	mtx_unlock(&ncneg_hot.nl_lock);
734 }
735 
736 static void
737 cache_negative_insert(struct namecache *ncp, bool neg_locked)
738 {
739 	struct neglist *neglist;
740 
741 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
742 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
743 	neglist = NCP2NEGLIST(ncp);
744 	if (!neg_locked) {
745 		mtx_lock(&neglist->nl_lock);
746 	} else {
747 		mtx_assert(&neglist->nl_lock, MA_OWNED);
748 	}
749 	TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
750 	if (!neg_locked)
751 		mtx_unlock(&neglist->nl_lock);
752 	atomic_add_rel_long(&numneg, 1);
753 }
754 
755 static void
756 cache_negative_remove(struct namecache *ncp, bool neg_locked)
757 {
758 	struct neglist *neglist;
759 	bool hot_locked = false;
760 	bool list_locked = false;
761 
762 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
763 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
764 	neglist = NCP2NEGLIST(ncp);
765 	if (!neg_locked) {
766 		if (ncp->nc_flag & NCF_HOTNEGATIVE) {
767 			hot_locked = true;
768 			mtx_lock(&ncneg_hot.nl_lock);
769 			if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
770 				list_locked = true;
771 				mtx_lock(&neglist->nl_lock);
772 			}
773 		} else {
774 			list_locked = true;
775 			mtx_lock(&neglist->nl_lock);
776 		}
777 	}
778 	if (ncp->nc_flag & NCF_HOTNEGATIVE) {
779 		mtx_assert(&ncneg_hot.nl_lock, MA_OWNED);
780 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
781 		numhotneg--;
782 	} else {
783 		mtx_assert(&neglist->nl_lock, MA_OWNED);
784 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
785 	}
786 	if (list_locked)
787 		mtx_unlock(&neglist->nl_lock);
788 	if (hot_locked)
789 		mtx_unlock(&ncneg_hot.nl_lock);
790 	atomic_subtract_rel_long(&numneg, 1);
791 }
792 
793 static void
794 cache_negative_shrink_select(int start, struct namecache **ncpp,
795     struct neglist **neglistpp)
796 {
797 	struct neglist *neglist;
798 	struct namecache *ncp;
799 	int i;
800 
801 	*ncpp = ncp = NULL;
802 	neglist = NULL;
803 
804 	for (i = start; i < numneglists; i++) {
805 		neglist = &neglists[i];
806 		if (TAILQ_FIRST(&neglist->nl_list) == NULL)
807 			continue;
808 		mtx_lock(&neglist->nl_lock);
809 		ncp = TAILQ_FIRST(&neglist->nl_list);
810 		if (ncp != NULL)
811 			break;
812 		mtx_unlock(&neglist->nl_lock);
813 	}
814 
815 	*neglistpp = neglist;
816 	*ncpp = ncp;
817 }
818 
819 static void
820 cache_negative_zap_one(void)
821 {
822 	struct namecache *ncp, *ncp2;
823 	struct neglist *neglist;
824 	struct mtx *dvlp;
825 	struct rwlock *blp;
826 
827 	if (mtx_owner(&ncneg_shrink_lock) != NULL ||
828 	    !mtx_trylock(&ncneg_shrink_lock)) {
829 		counter_u64_add(shrinking_skipped, 1);
830 		return;
831 	}
832 
833 	mtx_lock(&ncneg_hot.nl_lock);
834 	ncp = TAILQ_FIRST(&ncneg_hot.nl_list);
835 	if (ncp != NULL) {
836 		neglist = NCP2NEGLIST(ncp);
837 		mtx_lock(&neglist->nl_lock);
838 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
839 		TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
840 		ncp->nc_flag &= ~NCF_HOTNEGATIVE;
841 		numhotneg--;
842 		mtx_unlock(&neglist->nl_lock);
843 	}
844 	mtx_unlock(&ncneg_hot.nl_lock);
845 
846 	cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
847 	shrink_list_turn++;
848 	if (shrink_list_turn == numneglists)
849 		shrink_list_turn = 0;
850 	if (ncp == NULL && shrink_list_turn == 0)
851 		cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
852 	mtx_unlock(&ncneg_shrink_lock);
853 	if (ncp == NULL)
854 		return;
855 
856 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
857 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
858 	blp = NCP2BUCKETLOCK(ncp);
859 	mtx_unlock(&neglist->nl_lock);
860 	mtx_lock(dvlp);
861 	rw_wlock(blp);
862 	mtx_lock(&neglist->nl_lock);
863 	ncp2 = TAILQ_FIRST(&neglist->nl_list);
864 	if (ncp != ncp2 || dvlp != VP2VNODELOCK(ncp2->nc_dvp) ||
865 	    blp != NCP2BUCKETLOCK(ncp2) || !(ncp2->nc_flag & NCF_NEGATIVE)) {
866 		ncp = NULL;
867 	} else {
868 		SDT_PROBE2(vfs, namecache, shrink_negative, done, ncp->nc_dvp,
869 		    ncp->nc_name);
870 
871 		cache_zap_locked(ncp, true);
872 		counter_u64_add(numneg_evicted, 1);
873 	}
874 	mtx_unlock(&neglist->nl_lock);
875 	rw_wunlock(blp);
876 	mtx_unlock(dvlp);
877 	cache_free(ncp);
878 }
879 
880 /*
881  * cache_zap_locked():
882  *
883  *   Removes a namecache entry from cache, whether it contains an actual
884  *   pointer to a vnode or if it is just a negative cache entry.
885  */
886 static void
887 cache_zap_locked(struct namecache *ncp, bool neg_locked)
888 {
889 
890 	if (!(ncp->nc_flag & NCF_NEGATIVE))
891 		cache_assert_vnode_locked(ncp->nc_vp);
892 	cache_assert_vnode_locked(ncp->nc_dvp);
893 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
894 
895 	CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp,
896 	    (ncp->nc_flag & NCF_NEGATIVE) ? NULL : ncp->nc_vp);
897 
898 	cache_ncp_invalidate(ncp);
899 
900 	CK_LIST_REMOVE(ncp, nc_hash);
901 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
902 		SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
903 		    ncp->nc_name, ncp->nc_vp);
904 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
905 		if (ncp == ncp->nc_vp->v_cache_dd)
906 			ncp->nc_vp->v_cache_dd = NULL;
907 	} else {
908 		SDT_PROBE2(vfs, namecache, zap_negative, done, ncp->nc_dvp,
909 		    ncp->nc_name);
910 		cache_negative_remove(ncp, neg_locked);
911 	}
912 	if (ncp->nc_flag & NCF_ISDOTDOT) {
913 		if (ncp == ncp->nc_dvp->v_cache_dd)
914 			ncp->nc_dvp->v_cache_dd = NULL;
915 	} else {
916 		LIST_REMOVE(ncp, nc_src);
917 		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
918 			ncp->nc_flag |= NCF_DVDROP;
919 			counter_u64_add(numcachehv, -1);
920 		}
921 	}
922 	atomic_subtract_rel_long(&numcache, 1);
923 }
924 
925 static void
926 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
927 {
928 	struct rwlock *blp;
929 
930 	MPASS(ncp->nc_dvp == vp);
931 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
932 	cache_assert_vnode_locked(vp);
933 
934 	blp = NCP2BUCKETLOCK(ncp);
935 	rw_wlock(blp);
936 	cache_zap_locked(ncp, false);
937 	rw_wunlock(blp);
938 }
939 
940 static bool
941 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
942     struct mtx **vlpp)
943 {
944 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
945 	struct rwlock *blp;
946 
947 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
948 	cache_assert_vnode_locked(vp);
949 
950 	if (ncp->nc_flag & NCF_NEGATIVE) {
951 		if (*vlpp != NULL) {
952 			mtx_unlock(*vlpp);
953 			*vlpp = NULL;
954 		}
955 		cache_zap_negative_locked_vnode_kl(ncp, vp);
956 		return (true);
957 	}
958 
959 	pvlp = VP2VNODELOCK(vp);
960 	blp = NCP2BUCKETLOCK(ncp);
961 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
962 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
963 
964 	if (*vlpp == vlp1 || *vlpp == vlp2) {
965 		to_unlock = *vlpp;
966 		*vlpp = NULL;
967 	} else {
968 		if (*vlpp != NULL) {
969 			mtx_unlock(*vlpp);
970 			*vlpp = NULL;
971 		}
972 		cache_sort_vnodes(&vlp1, &vlp2);
973 		if (vlp1 == pvlp) {
974 			mtx_lock(vlp2);
975 			to_unlock = vlp2;
976 		} else {
977 			if (!mtx_trylock(vlp1))
978 				goto out_relock;
979 			to_unlock = vlp1;
980 		}
981 	}
982 	rw_wlock(blp);
983 	cache_zap_locked(ncp, false);
984 	rw_wunlock(blp);
985 	if (to_unlock != NULL)
986 		mtx_unlock(to_unlock);
987 	return (true);
988 
989 out_relock:
990 	mtx_unlock(vlp2);
991 	mtx_lock(vlp1);
992 	mtx_lock(vlp2);
993 	MPASS(*vlpp == NULL);
994 	*vlpp = vlp1;
995 	return (false);
996 }
997 
998 static int __noinline
999 cache_zap_locked_vnode(struct namecache *ncp, struct vnode *vp)
1000 {
1001 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
1002 	struct rwlock *blp;
1003 	int error = 0;
1004 
1005 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
1006 	cache_assert_vnode_locked(vp);
1007 
1008 	pvlp = VP2VNODELOCK(vp);
1009 	if (ncp->nc_flag & NCF_NEGATIVE) {
1010 		cache_zap_negative_locked_vnode_kl(ncp, vp);
1011 		goto out;
1012 	}
1013 
1014 	blp = NCP2BUCKETLOCK(ncp);
1015 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
1016 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
1017 	cache_sort_vnodes(&vlp1, &vlp2);
1018 	if (vlp1 == pvlp) {
1019 		mtx_lock(vlp2);
1020 		to_unlock = vlp2;
1021 	} else {
1022 		if (!mtx_trylock(vlp1)) {
1023 			error = EAGAIN;
1024 			goto out;
1025 		}
1026 		to_unlock = vlp1;
1027 	}
1028 	rw_wlock(blp);
1029 	cache_zap_locked(ncp, false);
1030 	rw_wunlock(blp);
1031 	mtx_unlock(to_unlock);
1032 out:
1033 	mtx_unlock(pvlp);
1034 	return (error);
1035 }
1036 
1037 /*
1038  * If trylocking failed we can get here. We know enough to take all needed locks
1039  * in the right order and re-lookup the entry.
1040  */
1041 static int
1042 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1043     struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1044     struct rwlock *blp)
1045 {
1046 	struct namecache *rncp;
1047 
1048 	cache_assert_bucket_locked(ncp, RA_UNLOCKED);
1049 
1050 	cache_sort_vnodes(&dvlp, &vlp);
1051 	cache_lock_vnodes(dvlp, vlp);
1052 	rw_wlock(blp);
1053 	CK_LIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1054 		if (rncp == ncp && rncp->nc_dvp == dvp &&
1055 		    rncp->nc_nlen == cnp->cn_namelen &&
1056 		    !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen))
1057 			break;
1058 	}
1059 	if (rncp != NULL) {
1060 		cache_zap_locked(rncp, false);
1061 		rw_wunlock(blp);
1062 		cache_unlock_vnodes(dvlp, vlp);
1063 		counter_u64_add(zap_and_exit_bucket_relock_success, 1);
1064 		return (0);
1065 	}
1066 
1067 	rw_wunlock(blp);
1068 	cache_unlock_vnodes(dvlp, vlp);
1069 	return (EAGAIN);
1070 }
1071 
1072 static int __noinline
1073 cache_zap_wlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1074     uint32_t hash, struct rwlock *blp)
1075 {
1076 	struct mtx *dvlp, *vlp;
1077 	struct vnode *dvp;
1078 
1079 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1080 
1081 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1082 	vlp = NULL;
1083 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1084 		vlp = VP2VNODELOCK(ncp->nc_vp);
1085 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1086 		cache_zap_locked(ncp, false);
1087 		rw_wunlock(blp);
1088 		cache_unlock_vnodes(dvlp, vlp);
1089 		return (0);
1090 	}
1091 
1092 	dvp = ncp->nc_dvp;
1093 	rw_wunlock(blp);
1094 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1095 }
1096 
1097 static int __noinline
1098 cache_zap_rlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1099     uint32_t hash, struct rwlock *blp)
1100 {
1101 	struct mtx *dvlp, *vlp;
1102 	struct vnode *dvp;
1103 
1104 	cache_assert_bucket_locked(ncp, RA_RLOCKED);
1105 
1106 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1107 	vlp = NULL;
1108 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1109 		vlp = VP2VNODELOCK(ncp->nc_vp);
1110 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1111 		rw_runlock(blp);
1112 		rw_wlock(blp);
1113 		cache_zap_locked(ncp, false);
1114 		rw_wunlock(blp);
1115 		cache_unlock_vnodes(dvlp, vlp);
1116 		return (0);
1117 	}
1118 
1119 	dvp = ncp->nc_dvp;
1120 	rw_runlock(blp);
1121 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1122 }
1123 
1124 static int
1125 cache_zap_wlocked_bucket_kl(struct namecache *ncp, struct rwlock *blp,
1126     struct mtx **vlpp1, struct mtx **vlpp2)
1127 {
1128 	struct mtx *dvlp, *vlp;
1129 
1130 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1131 
1132 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1133 	vlp = NULL;
1134 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1135 		vlp = VP2VNODELOCK(ncp->nc_vp);
1136 	cache_sort_vnodes(&dvlp, &vlp);
1137 
1138 	if (*vlpp1 == dvlp && *vlpp2 == vlp) {
1139 		cache_zap_locked(ncp, false);
1140 		cache_unlock_vnodes(dvlp, vlp);
1141 		*vlpp1 = NULL;
1142 		*vlpp2 = NULL;
1143 		return (0);
1144 	}
1145 
1146 	if (*vlpp1 != NULL)
1147 		mtx_unlock(*vlpp1);
1148 	if (*vlpp2 != NULL)
1149 		mtx_unlock(*vlpp2);
1150 	*vlpp1 = NULL;
1151 	*vlpp2 = NULL;
1152 
1153 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1154 		cache_zap_locked(ncp, false);
1155 		cache_unlock_vnodes(dvlp, vlp);
1156 		return (0);
1157 	}
1158 
1159 	rw_wunlock(blp);
1160 	*vlpp1 = dvlp;
1161 	*vlpp2 = vlp;
1162 	if (*vlpp1 != NULL)
1163 		mtx_lock(*vlpp1);
1164 	mtx_lock(*vlpp2);
1165 	rw_wlock(blp);
1166 	return (EAGAIN);
1167 }
1168 
1169 static void
1170 cache_lookup_unlock(struct rwlock *blp, struct mtx *vlp)
1171 {
1172 
1173 	if (blp != NULL) {
1174 		rw_runlock(blp);
1175 	} else {
1176 		mtx_unlock(vlp);
1177 	}
1178 }
1179 
1180 static int __noinline
1181 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1182     struct timespec *tsp, int *ticksp)
1183 {
1184 	int ltype;
1185 
1186 	*vpp = dvp;
1187 	CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .",
1188 			dvp, cnp->cn_nameptr);
1189 	counter_u64_add(dothits, 1);
1190 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1191 	if (tsp != NULL)
1192 		timespecclear(tsp);
1193 	if (ticksp != NULL)
1194 		*ticksp = ticks;
1195 	vrefact(*vpp);
1196 	/*
1197 	 * When we lookup "." we still can be asked to lock it
1198 	 * differently...
1199 	 */
1200 	ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1201 	if (ltype != VOP_ISLOCKED(*vpp)) {
1202 		if (ltype == LK_EXCLUSIVE) {
1203 			vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1204 			if (VN_IS_DOOMED((*vpp))) {
1205 				/* forced unmount */
1206 				vrele(*vpp);
1207 				*vpp = NULL;
1208 				return (ENOENT);
1209 			}
1210 		} else
1211 			vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1212 	}
1213 	return (-1);
1214 }
1215 
1216 static __noinline int
1217 cache_lookup_nomakeentry(struct vnode *dvp, struct vnode **vpp,
1218     struct componentname *cnp, struct timespec *tsp, int *ticksp)
1219 {
1220 	struct namecache *ncp;
1221 	struct rwlock *blp;
1222 	struct mtx *dvlp, *dvlp2;
1223 	uint32_t hash;
1224 	int error;
1225 
1226 	if (cnp->cn_namelen == 2 &&
1227 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1228 		counter_u64_add(dotdothits, 1);
1229 		dvlp = VP2VNODELOCK(dvp);
1230 		dvlp2 = NULL;
1231 		mtx_lock(dvlp);
1232 retry_dotdot:
1233 		ncp = dvp->v_cache_dd;
1234 		if (ncp == NULL) {
1235 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1236 			    "..", NULL);
1237 			mtx_unlock(dvlp);
1238 			if (dvlp2 != NULL)
1239 				mtx_unlock(dvlp2);
1240 			return (0);
1241 		}
1242 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1243 			if (ncp->nc_dvp != dvp)
1244 				panic("dvp %p v_cache_dd %p\n", dvp, ncp);
1245 			if (!cache_zap_locked_vnode_kl2(ncp,
1246 			    dvp, &dvlp2))
1247 				goto retry_dotdot;
1248 			MPASS(dvp->v_cache_dd == NULL);
1249 			mtx_unlock(dvlp);
1250 			if (dvlp2 != NULL)
1251 				mtx_unlock(dvlp2);
1252 			cache_free(ncp);
1253 		} else {
1254 			dvp->v_cache_dd = NULL;
1255 			mtx_unlock(dvlp);
1256 			if (dvlp2 != NULL)
1257 				mtx_unlock(dvlp2);
1258 		}
1259 		return (0);
1260 	}
1261 
1262 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1263 	blp = HASH2BUCKETLOCK(hash);
1264 retry:
1265 	if (CK_LIST_EMPTY(NCHHASH(hash)))
1266 		goto out_no_entry;
1267 
1268 	rw_wlock(blp);
1269 
1270 	CK_LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1271 		counter_u64_add(numchecks, 1);
1272 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1273 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1274 			break;
1275 	}
1276 
1277 	/* We failed to find an entry */
1278 	if (ncp == NULL) {
1279 		rw_wunlock(blp);
1280 		goto out_no_entry;
1281 	}
1282 
1283 	error = cache_zap_wlocked_bucket(ncp, cnp, hash, blp);
1284 	if (__predict_false(error != 0)) {
1285 		zap_and_exit_bucket_fail++;
1286 		cache_maybe_yield();
1287 		goto retry;
1288 	}
1289 	counter_u64_add(numposzaps, 1);
1290 	cache_free(ncp);
1291 	return (0);
1292 out_no_entry:
1293 	SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr, NULL);
1294 	counter_u64_add(nummisszap, 1);
1295 	return (0);
1296 }
1297 
1298 /**
1299  * Lookup a name in the name cache
1300  *
1301  * # Arguments
1302  *
1303  * - dvp:	Parent directory in which to search.
1304  * - vpp:	Return argument.  Will contain desired vnode on cache hit.
1305  * - cnp:	Parameters of the name search.  The most interesting bits of
1306  *   		the cn_flags field have the following meanings:
1307  *   	- MAKEENTRY:	If clear, free an entry from the cache rather than look
1308  *   			it up.
1309  *   	- ISDOTDOT:	Must be set if and only if cn_nameptr == ".."
1310  * - tsp:	Return storage for cache timestamp.  On a successful (positive
1311  *   		or negative) lookup, tsp will be filled with any timespec that
1312  *   		was stored when this cache entry was created.  However, it will
1313  *   		be clear for "." entries.
1314  * - ticks:	Return storage for alternate cache timestamp.  On a successful
1315  *   		(positive or negative) lookup, it will contain the ticks value
1316  *   		that was current when the cache entry was created, unless cnp
1317  *   		was ".".
1318  *
1319  * # Returns
1320  *
1321  * - -1:	A positive cache hit.  vpp will contain the desired vnode.
1322  * - ENOENT:	A negative cache hit, or dvp was recycled out from under us due
1323  *		to a forced unmount.  vpp will not be modified.  If the entry
1324  *		is a whiteout, then the ISWHITEOUT flag will be set in
1325  *		cnp->cn_flags.
1326  * - 0:		A cache miss.  vpp will not be modified.
1327  *
1328  * # Locking
1329  *
1330  * On a cache hit, vpp will be returned locked and ref'd.  If we're looking up
1331  * .., dvp is unlocked.  If we're looking up . an extra ref is taken, but the
1332  * lock is not recursively acquired.
1333  */
1334 int
1335 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1336     struct timespec *tsp, int *ticksp)
1337 {
1338 	struct namecache_ts *ncp_ts;
1339 	struct namecache *ncp;
1340 	struct rwlock *blp;
1341 	struct mtx *dvlp;
1342 	uint32_t hash;
1343 	enum vgetstate vs;
1344 	int error, ltype;
1345 	bool try_smr, doing_smr, whiteout;
1346 
1347 #ifdef DEBUG_CACHE
1348 	if (__predict_false(!doingcache)) {
1349 		cnp->cn_flags &= ~MAKEENTRY;
1350 		return (0);
1351 	}
1352 #endif
1353 
1354 	counter_u64_add(numcalls, 1);
1355 
1356 	if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.'))
1357 		return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1358 
1359 	if ((cnp->cn_flags & MAKEENTRY) == 0)
1360 		return (cache_lookup_nomakeentry(dvp, vpp, cnp, tsp, ticksp));
1361 
1362 	try_smr = true;
1363 	if (cnp->cn_nameiop == CREATE)
1364 		try_smr = false;
1365 retry:
1366 	doing_smr = false;
1367 	blp = NULL;
1368 	dvlp = NULL;
1369 	error = 0;
1370 	if (cnp->cn_namelen == 2 &&
1371 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1372 		counter_u64_add(dotdothits, 1);
1373 		dvlp = VP2VNODELOCK(dvp);
1374 		mtx_lock(dvlp);
1375 		ncp = dvp->v_cache_dd;
1376 		if (ncp == NULL) {
1377 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1378 			    "..", NULL);
1379 			mtx_unlock(dvlp);
1380 			return (0);
1381 		}
1382 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1383 			if (ncp->nc_flag & NCF_NEGATIVE)
1384 				*vpp = NULL;
1385 			else
1386 				*vpp = ncp->nc_vp;
1387 		} else
1388 			*vpp = ncp->nc_dvp;
1389 		/* Return failure if negative entry was found. */
1390 		if (*vpp == NULL)
1391 			goto negative_success;
1392 		CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
1393 		    dvp, cnp->cn_nameptr, *vpp);
1394 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..",
1395 		    *vpp);
1396 		cache_out_ts(ncp, tsp, ticksp);
1397 		if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1398 		    NCF_DTS && tsp != NULL) {
1399 			ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1400 			*tsp = ncp_ts->nc_dotdottime;
1401 		}
1402 		goto success;
1403 	}
1404 
1405 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1406 retry_hashed:
1407 	if (try_smr) {
1408 		vfs_smr_enter();
1409 		doing_smr = true;
1410 		try_smr = false;
1411 	} else {
1412 		blp = HASH2BUCKETLOCK(hash);
1413 		rw_rlock(blp);
1414 	}
1415 
1416 	CK_LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1417 		counter_u64_add(numchecks, 1);
1418 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1419 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1420 			break;
1421 	}
1422 
1423 	/* We failed to find an entry */
1424 	if (__predict_false(ncp == NULL)) {
1425 		if (doing_smr)
1426 			vfs_smr_exit();
1427 		else
1428 			rw_runlock(blp);
1429 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1430 		    NULL);
1431 		counter_u64_add(nummiss, 1);
1432 		return (0);
1433 	}
1434 
1435 	if (ncp->nc_flag & NCF_NEGATIVE)
1436 		goto negative_success;
1437 
1438 	/* We found a "positive" match, return the vnode */
1439 	counter_u64_add(numposhits, 1);
1440 	*vpp = ncp->nc_vp;
1441 	CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p",
1442 	    dvp, cnp->cn_nameptr, *vpp, ncp);
1443 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name,
1444 	    *vpp);
1445 	cache_out_ts(ncp, tsp, ticksp);
1446 success:
1447 	/*
1448 	 * On success we return a locked and ref'd vnode as per the lookup
1449 	 * protocol.
1450 	 */
1451 	MPASS(dvp != *vpp);
1452 	ltype = 0;	/* silence gcc warning */
1453 	if (cnp->cn_flags & ISDOTDOT) {
1454 		ltype = VOP_ISLOCKED(dvp);
1455 		VOP_UNLOCK(dvp);
1456 	}
1457 	if (doing_smr) {
1458 		if (cache_ncp_invalid(ncp)) {
1459 			vfs_smr_exit();
1460 			*vpp = NULL;
1461 			goto retry;
1462 		}
1463 		vs = vget_prep_smr(*vpp);
1464 		vfs_smr_exit();
1465 		if (vs == VGET_NONE) {
1466 			*vpp = NULL;
1467 			goto retry;
1468 		}
1469 	} else {
1470 		vs = vget_prep(*vpp);
1471 		cache_lookup_unlock(blp, dvlp);
1472 	}
1473 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1474 	if (cnp->cn_flags & ISDOTDOT) {
1475 		vn_lock(dvp, ltype | LK_RETRY);
1476 		if (VN_IS_DOOMED(dvp)) {
1477 			if (error == 0)
1478 				vput(*vpp);
1479 			*vpp = NULL;
1480 			return (ENOENT);
1481 		}
1482 	}
1483 	if (error) {
1484 		*vpp = NULL;
1485 		goto retry;
1486 	}
1487 	if ((cnp->cn_flags & ISLASTCN) &&
1488 	    (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) {
1489 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
1490 	}
1491 	return (-1);
1492 
1493 negative_success:
1494 	/* We found a negative match, and want to create it, so purge */
1495 	if (cnp->cn_nameiop == CREATE) {
1496 		MPASS(!doing_smr);
1497 		counter_u64_add(numnegzaps, 1);
1498 		goto zap_and_exit;
1499 	}
1500 
1501 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp, ncp->nc_name);
1502 	cache_out_ts(ncp, tsp, ticksp);
1503 	counter_u64_add(numneghits, 1);
1504 	whiteout = (ncp->nc_flag & NCF_WHITE);
1505 
1506 	if (doing_smr) {
1507 		/*
1508 		 * We need to take locks to promote an entry.
1509 		 */
1510 		if ((ncp->nc_flag & NCF_HOTNEGATIVE) == 0 ||
1511 		    cache_ncp_invalid(ncp)) {
1512 			vfs_smr_exit();
1513 			doing_smr = false;
1514 			goto retry_hashed;
1515 		}
1516 		vfs_smr_exit();
1517 	} else {
1518 		cache_negative_hit(ncp);
1519 		cache_lookup_unlock(blp, dvlp);
1520 	}
1521 	if (whiteout)
1522 		cnp->cn_flags |= ISWHITEOUT;
1523 	return (ENOENT);
1524 
1525 zap_and_exit:
1526 	MPASS(!doing_smr);
1527 	if (blp != NULL)
1528 		error = cache_zap_rlocked_bucket(ncp, cnp, hash, blp);
1529 	else
1530 		error = cache_zap_locked_vnode(ncp, dvp);
1531 	if (__predict_false(error != 0)) {
1532 		zap_and_exit_bucket_fail2++;
1533 		cache_maybe_yield();
1534 		goto retry;
1535 	}
1536 	cache_free(ncp);
1537 	return (0);
1538 }
1539 
1540 struct celockstate {
1541 	struct mtx *vlp[3];
1542 	struct rwlock *blp[2];
1543 };
1544 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1545 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1546 
1547 static inline void
1548 cache_celockstate_init(struct celockstate *cel)
1549 {
1550 
1551 	bzero(cel, sizeof(*cel));
1552 }
1553 
1554 static void
1555 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1556     struct vnode *dvp)
1557 {
1558 	struct mtx *vlp1, *vlp2;
1559 
1560 	MPASS(cel->vlp[0] == NULL);
1561 	MPASS(cel->vlp[1] == NULL);
1562 	MPASS(cel->vlp[2] == NULL);
1563 
1564 	MPASS(vp != NULL || dvp != NULL);
1565 
1566 	vlp1 = VP2VNODELOCK(vp);
1567 	vlp2 = VP2VNODELOCK(dvp);
1568 	cache_sort_vnodes(&vlp1, &vlp2);
1569 
1570 	if (vlp1 != NULL) {
1571 		mtx_lock(vlp1);
1572 		cel->vlp[0] = vlp1;
1573 	}
1574 	mtx_lock(vlp2);
1575 	cel->vlp[1] = vlp2;
1576 }
1577 
1578 static void
1579 cache_unlock_vnodes_cel(struct celockstate *cel)
1580 {
1581 
1582 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
1583 
1584 	if (cel->vlp[0] != NULL)
1585 		mtx_unlock(cel->vlp[0]);
1586 	if (cel->vlp[1] != NULL)
1587 		mtx_unlock(cel->vlp[1]);
1588 	if (cel->vlp[2] != NULL)
1589 		mtx_unlock(cel->vlp[2]);
1590 }
1591 
1592 static bool
1593 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
1594 {
1595 	struct mtx *vlp;
1596 	bool ret;
1597 
1598 	cache_assert_vlp_locked(cel->vlp[0]);
1599 	cache_assert_vlp_locked(cel->vlp[1]);
1600 	MPASS(cel->vlp[2] == NULL);
1601 
1602 	MPASS(vp != NULL);
1603 	vlp = VP2VNODELOCK(vp);
1604 
1605 	ret = true;
1606 	if (vlp >= cel->vlp[1]) {
1607 		mtx_lock(vlp);
1608 	} else {
1609 		if (mtx_trylock(vlp))
1610 			goto out;
1611 		cache_lock_vnodes_cel_3_failures++;
1612 		cache_unlock_vnodes_cel(cel);
1613 		if (vlp < cel->vlp[0]) {
1614 			mtx_lock(vlp);
1615 			mtx_lock(cel->vlp[0]);
1616 			mtx_lock(cel->vlp[1]);
1617 		} else {
1618 			if (cel->vlp[0] != NULL)
1619 				mtx_lock(cel->vlp[0]);
1620 			mtx_lock(vlp);
1621 			mtx_lock(cel->vlp[1]);
1622 		}
1623 		ret = false;
1624 	}
1625 out:
1626 	cel->vlp[2] = vlp;
1627 	return (ret);
1628 }
1629 
1630 static void
1631 cache_lock_buckets_cel(struct celockstate *cel, struct rwlock *blp1,
1632     struct rwlock *blp2)
1633 {
1634 
1635 	MPASS(cel->blp[0] == NULL);
1636 	MPASS(cel->blp[1] == NULL);
1637 
1638 	cache_sort_vnodes(&blp1, &blp2);
1639 
1640 	if (blp1 != NULL) {
1641 		rw_wlock(blp1);
1642 		cel->blp[0] = blp1;
1643 	}
1644 	rw_wlock(blp2);
1645 	cel->blp[1] = blp2;
1646 }
1647 
1648 static void
1649 cache_unlock_buckets_cel(struct celockstate *cel)
1650 {
1651 
1652 	if (cel->blp[0] != NULL)
1653 		rw_wunlock(cel->blp[0]);
1654 	rw_wunlock(cel->blp[1]);
1655 }
1656 
1657 /*
1658  * Lock part of the cache affected by the insertion.
1659  *
1660  * This means vnodelocks for dvp, vp and the relevant bucketlock.
1661  * However, insertion can result in removal of an old entry. In this
1662  * case we have an additional vnode and bucketlock pair to lock. If the
1663  * entry is negative, ncelock is locked instead of the vnode.
1664  *
1665  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
1666  * preserving the locking order (smaller address first).
1667  */
1668 static void
1669 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1670     uint32_t hash)
1671 {
1672 	struct namecache *ncp;
1673 	struct rwlock *blps[2];
1674 
1675 	blps[0] = HASH2BUCKETLOCK(hash);
1676 	for (;;) {
1677 		blps[1] = NULL;
1678 		cache_lock_vnodes_cel(cel, dvp, vp);
1679 		if (vp == NULL || vp->v_type != VDIR)
1680 			break;
1681 		ncp = vp->v_cache_dd;
1682 		if (ncp == NULL)
1683 			break;
1684 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1685 			break;
1686 		MPASS(ncp->nc_dvp == vp);
1687 		blps[1] = NCP2BUCKETLOCK(ncp);
1688 		if (ncp->nc_flag & NCF_NEGATIVE)
1689 			break;
1690 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1691 			break;
1692 		/*
1693 		 * All vnodes got re-locked. Re-validate the state and if
1694 		 * nothing changed we are done. Otherwise restart.
1695 		 */
1696 		if (ncp == vp->v_cache_dd &&
1697 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1698 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1699 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1700 			break;
1701 		cache_unlock_vnodes_cel(cel);
1702 		cel->vlp[0] = NULL;
1703 		cel->vlp[1] = NULL;
1704 		cel->vlp[2] = NULL;
1705 	}
1706 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1707 }
1708 
1709 static void
1710 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1711     uint32_t hash)
1712 {
1713 	struct namecache *ncp;
1714 	struct rwlock *blps[2];
1715 
1716 	blps[0] = HASH2BUCKETLOCK(hash);
1717 	for (;;) {
1718 		blps[1] = NULL;
1719 		cache_lock_vnodes_cel(cel, dvp, vp);
1720 		ncp = dvp->v_cache_dd;
1721 		if (ncp == NULL)
1722 			break;
1723 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1724 			break;
1725 		MPASS(ncp->nc_dvp == dvp);
1726 		blps[1] = NCP2BUCKETLOCK(ncp);
1727 		if (ncp->nc_flag & NCF_NEGATIVE)
1728 			break;
1729 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1730 			break;
1731 		if (ncp == dvp->v_cache_dd &&
1732 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1733 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1734 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1735 			break;
1736 		cache_unlock_vnodes_cel(cel);
1737 		cel->vlp[0] = NULL;
1738 		cel->vlp[1] = NULL;
1739 		cel->vlp[2] = NULL;
1740 	}
1741 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1742 }
1743 
1744 static void
1745 cache_enter_unlock(struct celockstate *cel)
1746 {
1747 
1748 	cache_unlock_buckets_cel(cel);
1749 	cache_unlock_vnodes_cel(cel);
1750 }
1751 
1752 static void __noinline
1753 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp,
1754     struct componentname *cnp)
1755 {
1756 	struct celockstate cel;
1757 	struct namecache *ncp;
1758 	uint32_t hash;
1759 	int len;
1760 
1761 	if (dvp->v_cache_dd == NULL)
1762 		return;
1763 	len = cnp->cn_namelen;
1764 	cache_celockstate_init(&cel);
1765 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1766 	cache_enter_lock_dd(&cel, dvp, vp, hash);
1767 	ncp = dvp->v_cache_dd;
1768 	if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) {
1769 		KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent"));
1770 		cache_zap_locked(ncp, false);
1771 	} else {
1772 		ncp = NULL;
1773 	}
1774 	dvp->v_cache_dd = NULL;
1775 	cache_enter_unlock(&cel);
1776 	cache_free(ncp);
1777 }
1778 
1779 /*
1780  * Add an entry to the cache.
1781  */
1782 void
1783 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
1784     struct timespec *tsp, struct timespec *dtsp)
1785 {
1786 	struct celockstate cel;
1787 	struct namecache *ncp, *n2, *ndd;
1788 	struct namecache_ts *ncp_ts, *n2_ts;
1789 	struct nchashhead *ncpp;
1790 	uint32_t hash;
1791 	int flag;
1792 	int len;
1793 	u_long lnumcache;
1794 
1795 	CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr);
1796 	VNASSERT(vp == NULL || !VN_IS_DOOMED(vp), vp,
1797 	    ("cache_enter: Adding a doomed vnode"));
1798 	VNASSERT(dvp == NULL || !VN_IS_DOOMED(dvp), dvp,
1799 	    ("cache_enter: Doomed vnode used as src"));
1800 
1801 #ifdef DEBUG_CACHE
1802 	if (__predict_false(!doingcache))
1803 		return;
1804 #endif
1805 
1806 	flag = 0;
1807 	if (__predict_false(cnp->cn_nameptr[0] == '.')) {
1808 		if (cnp->cn_namelen == 1)
1809 			return;
1810 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1811 			cache_enter_dotdot_prep(dvp, vp, cnp);
1812 			flag = NCF_ISDOTDOT;
1813 		}
1814 	}
1815 
1816 	/*
1817 	 * Avoid blowout in namecache entries.
1818 	 */
1819 	lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
1820 	if (__predict_false(lnumcache >= ncsize)) {
1821 		atomic_add_long(&numcache, -1);
1822 		return;
1823 	}
1824 
1825 	cache_celockstate_init(&cel);
1826 	ndd = NULL;
1827 	ncp_ts = NULL;
1828 
1829 	/*
1830 	 * Calculate the hash key and setup as much of the new
1831 	 * namecache entry as possible before acquiring the lock.
1832 	 */
1833 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
1834 	ncp->nc_flag = flag;
1835 	ncp->nc_vp = vp;
1836 	if (vp == NULL)
1837 		ncp->nc_flag |= NCF_NEGATIVE;
1838 	ncp->nc_dvp = dvp;
1839 	if (tsp != NULL) {
1840 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1841 		ncp_ts->nc_time = *tsp;
1842 		ncp_ts->nc_ticks = ticks;
1843 		ncp_ts->nc_nc.nc_flag |= NCF_TS;
1844 		if (dtsp != NULL) {
1845 			ncp_ts->nc_dotdottime = *dtsp;
1846 			ncp_ts->nc_nc.nc_flag |= NCF_DTS;
1847 		}
1848 	}
1849 	len = ncp->nc_nlen = cnp->cn_namelen;
1850 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1851 	strlcpy(ncp->nc_name, cnp->cn_nameptr, len + 1);
1852 	cache_enter_lock(&cel, dvp, vp, hash);
1853 
1854 	/*
1855 	 * See if this vnode or negative entry is already in the cache
1856 	 * with this name.  This can happen with concurrent lookups of
1857 	 * the same path name.
1858 	 */
1859 	ncpp = NCHHASH(hash);
1860 	CK_LIST_FOREACH(n2, ncpp, nc_hash) {
1861 		if (n2->nc_dvp == dvp &&
1862 		    n2->nc_nlen == cnp->cn_namelen &&
1863 		    !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
1864 			if (tsp != NULL) {
1865 				KASSERT((n2->nc_flag & NCF_TS) != 0,
1866 				    ("no NCF_TS"));
1867 				n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
1868 				n2_ts->nc_time = ncp_ts->nc_time;
1869 				n2_ts->nc_ticks = ncp_ts->nc_ticks;
1870 				if (dtsp != NULL) {
1871 					n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
1872 					if (ncp->nc_flag & NCF_NEGATIVE)
1873 						mtx_lock(&ncneg_hot.nl_lock);
1874 					n2_ts->nc_nc.nc_flag |= NCF_DTS;
1875 					if (ncp->nc_flag & NCF_NEGATIVE)
1876 						mtx_unlock(&ncneg_hot.nl_lock);
1877 				}
1878 			}
1879 			goto out_unlock_free;
1880 		}
1881 	}
1882 
1883 	if (flag == NCF_ISDOTDOT) {
1884 		/*
1885 		 * See if we are trying to add .. entry, but some other lookup
1886 		 * has populated v_cache_dd pointer already.
1887 		 */
1888 		if (dvp->v_cache_dd != NULL)
1889 			goto out_unlock_free;
1890 		KASSERT(vp == NULL || vp->v_type == VDIR,
1891 		    ("wrong vnode type %p", vp));
1892 		dvp->v_cache_dd = ncp;
1893 	}
1894 
1895 	if (vp != NULL) {
1896 		if (vp->v_type == VDIR) {
1897 			if (flag != NCF_ISDOTDOT) {
1898 				/*
1899 				 * For this case, the cache entry maps both the
1900 				 * directory name in it and the name ".." for the
1901 				 * directory's parent.
1902 				 */
1903 				if ((ndd = vp->v_cache_dd) != NULL) {
1904 					if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
1905 						cache_zap_locked(ndd, false);
1906 					else
1907 						ndd = NULL;
1908 				}
1909 				vp->v_cache_dd = ncp;
1910 			}
1911 		} else {
1912 			vp->v_cache_dd = NULL;
1913 		}
1914 	}
1915 
1916 	if (flag != NCF_ISDOTDOT) {
1917 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1918 			vhold(dvp);
1919 			counter_u64_add(numcachehv, 1);
1920 		}
1921 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
1922 	}
1923 
1924 	/*
1925 	 * If the entry is "negative", we place it into the
1926 	 * "negative" cache queue, otherwise, we place it into the
1927 	 * destination vnode's cache entries queue.
1928 	 */
1929 	if (vp != NULL) {
1930 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
1931 		SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
1932 		    vp);
1933 	} else {
1934 		if (cnp->cn_flags & ISWHITEOUT)
1935 			ncp->nc_flag |= NCF_WHITE;
1936 		cache_negative_insert(ncp, false);
1937 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
1938 		    ncp->nc_name);
1939 	}
1940 
1941 	atomic_thread_fence_rel();
1942 	/*
1943 	 * Insert the new namecache entry into the appropriate chain
1944 	 * within the cache entries table.
1945 	 */
1946 	CK_LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
1947 
1948 	cache_enter_unlock(&cel);
1949 	if (numneg * ncnegfactor > lnumcache)
1950 		cache_negative_zap_one();
1951 	cache_free(ndd);
1952 	return;
1953 out_unlock_free:
1954 	cache_enter_unlock(&cel);
1955 	cache_free(ncp);
1956 	return;
1957 }
1958 
1959 static u_int
1960 cache_roundup_2(u_int val)
1961 {
1962 	u_int res;
1963 
1964 	for (res = 1; res <= val; res <<= 1)
1965 		continue;
1966 
1967 	return (res);
1968 }
1969 
1970 /*
1971  * Name cache initialization, from vfs_init() when we are booting
1972  */
1973 static void
1974 nchinit(void *dummy __unused)
1975 {
1976 	u_int i;
1977 
1978 	cache_zone_small = uma_zcreate("S VFS Cache",
1979 	    sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1,
1980 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1981 	    UMA_ZONE_ZINIT);
1982 	cache_zone_small_ts = uma_zcreate("STS VFS Cache",
1983 	    sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1,
1984 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1985 	    UMA_ZONE_ZINIT);
1986 	cache_zone_large = uma_zcreate("L VFS Cache",
1987 	    sizeof(struct namecache) + NAME_MAX + 1,
1988 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1989 	    UMA_ZONE_ZINIT);
1990 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache",
1991 	    sizeof(struct namecache_ts) + NAME_MAX + 1,
1992 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1993 	    UMA_ZONE_ZINIT);
1994 
1995 	VFS_SMR_ZONE_SET(cache_zone_small);
1996 	VFS_SMR_ZONE_SET(cache_zone_small_ts);
1997 	VFS_SMR_ZONE_SET(cache_zone_large);
1998 	VFS_SMR_ZONE_SET(cache_zone_large_ts);
1999 
2000 	ncsize = desiredvnodes * ncsizefactor;
2001 	nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash);
2002 	ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
2003 	if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
2004 		ncbuckethash = 7;
2005 	if (ncbuckethash > nchash)
2006 		ncbuckethash = nchash;
2007 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
2008 	    M_WAITOK | M_ZERO);
2009 	for (i = 0; i < numbucketlocks; i++)
2010 		rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK | RW_RECURSE);
2011 	ncvnodehash = ncbuckethash;
2012 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
2013 	    M_WAITOK | M_ZERO);
2014 	for (i = 0; i < numvnodelocks; i++)
2015 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
2016 	ncpurgeminvnodes = numbucketlocks * 2;
2017 
2018 	ncneghash = 3;
2019 	neglists = malloc(sizeof(*neglists) * numneglists, M_VFSCACHE,
2020 	    M_WAITOK | M_ZERO);
2021 	for (i = 0; i < numneglists; i++) {
2022 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
2023 		TAILQ_INIT(&neglists[i].nl_list);
2024 	}
2025 	mtx_init(&ncneg_hot.nl_lock, "ncneglh", NULL, MTX_DEF);
2026 	TAILQ_INIT(&ncneg_hot.nl_list);
2027 
2028 	mtx_init(&ncneg_shrink_lock, "ncnegs", NULL, MTX_DEF);
2029 }
2030 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
2031 
2032 void
2033 cache_changesize(u_long newmaxvnodes)
2034 {
2035 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
2036 	u_long new_nchash, old_nchash;
2037 	struct namecache *ncp;
2038 	uint32_t hash;
2039 	u_long newncsize;
2040 	int i;
2041 
2042 	newncsize = newmaxvnodes * ncsizefactor;
2043 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
2044 	if (newmaxvnodes < numbucketlocks)
2045 		newmaxvnodes = numbucketlocks;
2046 
2047 	new_nchashtbl = hashinit(newmaxvnodes, M_VFSCACHE, &new_nchash);
2048 	/* If same hash table size, nothing to do */
2049 	if (nchash == new_nchash) {
2050 		free(new_nchashtbl, M_VFSCACHE);
2051 		return;
2052 	}
2053 	/*
2054 	 * Move everything from the old hash table to the new table.
2055 	 * None of the namecache entries in the table can be removed
2056 	 * because to do so, they have to be removed from the hash table.
2057 	 */
2058 	cache_lock_all_vnodes();
2059 	cache_lock_all_buckets();
2060 	old_nchashtbl = nchashtbl;
2061 	old_nchash = nchash;
2062 	nchashtbl = new_nchashtbl;
2063 	nchash = new_nchash;
2064 	for (i = 0; i <= old_nchash; i++) {
2065 		while ((ncp = CK_LIST_FIRST(&old_nchashtbl[i])) != NULL) {
2066 			hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
2067 			    ncp->nc_dvp);
2068 			CK_LIST_REMOVE(ncp, nc_hash);
2069 			CK_LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
2070 		}
2071 	}
2072 	ncsize = newncsize;
2073 	cache_unlock_all_buckets();
2074 	cache_unlock_all_vnodes();
2075 	free(old_nchashtbl, M_VFSCACHE);
2076 }
2077 
2078 /*
2079  * Invalidate all entries from and to a particular vnode.
2080  */
2081 void
2082 cache_purge(struct vnode *vp)
2083 {
2084 	TAILQ_HEAD(, namecache) ncps;
2085 	struct namecache *ncp, *nnp;
2086 	struct mtx *vlp, *vlp2;
2087 
2088 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
2089 	SDT_PROBE1(vfs, namecache, purge, done, vp);
2090 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2091 	    vp->v_cache_dd == NULL)
2092 		return;
2093 	TAILQ_INIT(&ncps);
2094 	vlp = VP2VNODELOCK(vp);
2095 	vlp2 = NULL;
2096 	mtx_lock(vlp);
2097 retry:
2098 	while (!LIST_EMPTY(&vp->v_cache_src)) {
2099 		ncp = LIST_FIRST(&vp->v_cache_src);
2100 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2101 			goto retry;
2102 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2103 	}
2104 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2105 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
2106 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2107 			goto retry;
2108 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2109 	}
2110 	ncp = vp->v_cache_dd;
2111 	if (ncp != NULL) {
2112 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2113 		   ("lost dotdot link"));
2114 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2115 			goto retry;
2116 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2117 	}
2118 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2119 	mtx_unlock(vlp);
2120 	if (vlp2 != NULL)
2121 		mtx_unlock(vlp2);
2122 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2123 		cache_free(ncp);
2124 	}
2125 }
2126 
2127 /*
2128  * Invalidate all negative entries for a particular directory vnode.
2129  */
2130 void
2131 cache_purge_negative(struct vnode *vp)
2132 {
2133 	TAILQ_HEAD(, namecache) ncps;
2134 	struct namecache *ncp, *nnp;
2135 	struct mtx *vlp;
2136 
2137 	CTR1(KTR_VFS, "cache_purge_negative(%p)", vp);
2138 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2139 	if (LIST_EMPTY(&vp->v_cache_src))
2140 		return;
2141 	TAILQ_INIT(&ncps);
2142 	vlp = VP2VNODELOCK(vp);
2143 	mtx_lock(vlp);
2144 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2145 		if (!(ncp->nc_flag & NCF_NEGATIVE))
2146 			continue;
2147 		cache_zap_negative_locked_vnode_kl(ncp, vp);
2148 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2149 	}
2150 	mtx_unlock(vlp);
2151 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2152 		cache_free(ncp);
2153 	}
2154 }
2155 
2156 /*
2157  * Flush all entries referencing a particular filesystem.
2158  */
2159 void
2160 cache_purgevfs(struct mount *mp, bool force)
2161 {
2162 	TAILQ_HEAD(, namecache) ncps;
2163 	struct mtx *vlp1, *vlp2;
2164 	struct rwlock *blp;
2165 	struct nchashhead *bucket;
2166 	struct namecache *ncp, *nnp;
2167 	u_long i, j, n_nchash;
2168 	int error;
2169 
2170 	/* Scan hash tables for applicable entries */
2171 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2172 	if (!force && mp->mnt_nvnodelistsize <= ncpurgeminvnodes)
2173 		return;
2174 	TAILQ_INIT(&ncps);
2175 	n_nchash = nchash + 1;
2176 	vlp1 = vlp2 = NULL;
2177 	for (i = 0; i < numbucketlocks; i++) {
2178 		blp = (struct rwlock *)&bucketlocks[i];
2179 		rw_wlock(blp);
2180 		for (j = i; j < n_nchash; j += numbucketlocks) {
2181 retry:
2182 			bucket = &nchashtbl[j];
2183 			CK_LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) {
2184 				cache_assert_bucket_locked(ncp, RA_WLOCKED);
2185 				if (ncp->nc_dvp->v_mount != mp)
2186 					continue;
2187 				error = cache_zap_wlocked_bucket_kl(ncp, blp,
2188 				    &vlp1, &vlp2);
2189 				if (error != 0)
2190 					goto retry;
2191 				TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst);
2192 			}
2193 		}
2194 		rw_wunlock(blp);
2195 		if (vlp1 == NULL && vlp2 == NULL)
2196 			cache_maybe_yield();
2197 	}
2198 	if (vlp1 != NULL)
2199 		mtx_unlock(vlp1);
2200 	if (vlp2 != NULL)
2201 		mtx_unlock(vlp2);
2202 
2203 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2204 		cache_free(ncp);
2205 	}
2206 }
2207 
2208 /*
2209  * Perform canonical checks and cache lookup and pass on to filesystem
2210  * through the vop_cachedlookup only if needed.
2211  */
2212 
2213 int
2214 vfs_cache_lookup(struct vop_lookup_args *ap)
2215 {
2216 	struct vnode *dvp;
2217 	int error;
2218 	struct vnode **vpp = ap->a_vpp;
2219 	struct componentname *cnp = ap->a_cnp;
2220 	int flags = cnp->cn_flags;
2221 
2222 	*vpp = NULL;
2223 	dvp = ap->a_dvp;
2224 
2225 	if (dvp->v_type != VDIR)
2226 		return (ENOTDIR);
2227 
2228 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2229 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2230 		return (EROFS);
2231 
2232 	error = vn_dir_check_exec(dvp, cnp);
2233 	if (error != 0)
2234 		return (error);
2235 
2236 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2237 	if (error == 0)
2238 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2239 	if (error == -1)
2240 		return (0);
2241 	return (error);
2242 }
2243 
2244 /* Implementation of the getcwd syscall. */
2245 int
2246 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2247 {
2248 	char *buf, *retbuf;
2249 	size_t buflen;
2250 	int error;
2251 
2252 	buflen = uap->buflen;
2253 	if (__predict_false(buflen < 2))
2254 		return (EINVAL);
2255 	if (buflen > MAXPATHLEN)
2256 		buflen = MAXPATHLEN;
2257 
2258 	buf = malloc(buflen, M_TEMP, M_WAITOK);
2259 	error = vn_getcwd(td, buf, &retbuf, &buflen);
2260 	if (error == 0)
2261 		error = copyout(retbuf, uap->buf, buflen);
2262 	free(buf, M_TEMP);
2263 	return (error);
2264 }
2265 
2266 int
2267 vn_getcwd(struct thread *td, char *buf, char **retbuf, size_t *buflen)
2268 {
2269 	struct pwd *pwd;
2270 	int error;
2271 
2272 	pwd = pwd_hold(td);
2273 	error = vn_fullpath_any(td, pwd->pwd_cdir, pwd->pwd_rdir, buf, retbuf, buflen);
2274 	pwd_drop(pwd);
2275 
2276 #ifdef KTRACE
2277 	if (KTRPOINT(curthread, KTR_NAMEI) && error == 0)
2278 		ktrnamei(*retbuf);
2279 #endif
2280 	return (error);
2281 }
2282 
2283 static int
2284 kern___realpathat(struct thread *td, int fd, const char *path, char *buf,
2285     size_t size, int flags, enum uio_seg pathseg)
2286 {
2287 	struct nameidata nd;
2288 	char *retbuf, *freebuf;
2289 	int error;
2290 
2291 	if (flags != 0)
2292 		return (EINVAL);
2293 	NDINIT_ATRIGHTS(&nd, LOOKUP, FOLLOW | SAVENAME | WANTPARENT | AUDITVNODE1,
2294 	    pathseg, path, fd, &cap_fstat_rights, td);
2295 	if ((error = namei(&nd)) != 0)
2296 		return (error);
2297 	error = vn_fullpath_hardlink(td, &nd, &retbuf, &freebuf, &size);
2298 	if (error == 0) {
2299 		error = copyout(retbuf, buf, size);
2300 		free(freebuf, M_TEMP);
2301 	}
2302 	NDFREE(&nd, 0);
2303 	return (error);
2304 }
2305 
2306 int
2307 sys___realpathat(struct thread *td, struct __realpathat_args *uap)
2308 {
2309 
2310 	return (kern___realpathat(td, uap->fd, uap->path, uap->buf, uap->size,
2311 	    uap->flags, UIO_USERSPACE));
2312 }
2313 
2314 /*
2315  * Retrieve the full filesystem path that correspond to a vnode from the name
2316  * cache (if available)
2317  */
2318 int
2319 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf)
2320 {
2321 	struct pwd *pwd;
2322 	char *buf;
2323 	size_t buflen;
2324 	int error;
2325 
2326 	if (__predict_false(vn == NULL))
2327 		return (EINVAL);
2328 
2329 	buflen = MAXPATHLEN;
2330 	buf = malloc(buflen, M_TEMP, M_WAITOK);
2331 	pwd = pwd_hold(td);
2332 	error = vn_fullpath_any(td, vn, pwd->pwd_rdir, buf, retbuf, &buflen);
2333 	pwd_drop(pwd);
2334 
2335 	if (!error)
2336 		*freebuf = buf;
2337 	else
2338 		free(buf, M_TEMP);
2339 	return (error);
2340 }
2341 
2342 /*
2343  * This function is similar to vn_fullpath, but it attempts to lookup the
2344  * pathname relative to the global root mount point.  This is required for the
2345  * auditing sub-system, as audited pathnames must be absolute, relative to the
2346  * global root mount point.
2347  */
2348 int
2349 vn_fullpath_global(struct thread *td, struct vnode *vn,
2350     char **retbuf, char **freebuf)
2351 {
2352 	char *buf;
2353 	size_t buflen;
2354 	int error;
2355 
2356 	if (__predict_false(vn == NULL))
2357 		return (EINVAL);
2358 	buflen = MAXPATHLEN;
2359 	buf = malloc(buflen, M_TEMP, M_WAITOK);
2360 	error = vn_fullpath_any(td, vn, rootvnode, buf, retbuf, &buflen);
2361 	if (!error)
2362 		*freebuf = buf;
2363 	else
2364 		free(buf, M_TEMP);
2365 	return (error);
2366 }
2367 
2368 int
2369 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, size_t *buflen)
2370 {
2371 	struct vnode *dvp;
2372 	struct namecache *ncp;
2373 	struct mtx *vlp;
2374 	int error;
2375 
2376 	vlp = VP2VNODELOCK(*vp);
2377 	mtx_lock(vlp);
2378 	TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) {
2379 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2380 			break;
2381 	}
2382 	if (ncp != NULL) {
2383 		if (*buflen < ncp->nc_nlen) {
2384 			mtx_unlock(vlp);
2385 			vrele(*vp);
2386 			counter_u64_add(numfullpathfail4, 1);
2387 			error = ENOMEM;
2388 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2389 			    vp, NULL);
2390 			return (error);
2391 		}
2392 		*buflen -= ncp->nc_nlen;
2393 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2394 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2395 		    ncp->nc_name, vp);
2396 		dvp = *vp;
2397 		*vp = ncp->nc_dvp;
2398 		vref(*vp);
2399 		mtx_unlock(vlp);
2400 		vrele(dvp);
2401 		return (0);
2402 	}
2403 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2404 
2405 	mtx_unlock(vlp);
2406 	vn_lock(*vp, LK_SHARED | LK_RETRY);
2407 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
2408 	vput(*vp);
2409 	if (error) {
2410 		counter_u64_add(numfullpathfail2, 1);
2411 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
2412 		return (error);
2413 	}
2414 
2415 	*vp = dvp;
2416 	if (VN_IS_DOOMED(dvp)) {
2417 		/* forced unmount */
2418 		vrele(dvp);
2419 		error = ENOENT;
2420 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2421 		return (error);
2422 	}
2423 	/*
2424 	 * *vp has its use count incremented still.
2425 	 */
2426 
2427 	return (0);
2428 }
2429 
2430 /*
2431  * Resolve a directory to a pathname.
2432  *
2433  * The name of the directory can always be found in the namecache or fetched
2434  * from the filesystem. There is also guaranteed to be only one parent, meaning
2435  * we can just follow vnodes up until we find the root.
2436  *
2437  * The vnode must be referenced.
2438  */
2439 static int
2440 vn_fullpath_dir(struct thread *td, struct vnode *vp, struct vnode *rdir,
2441     char *buf, char **retbuf, size_t *len, bool slash_prefixed, size_t addend)
2442 {
2443 #ifdef KDTRACE_HOOKS
2444 	struct vnode *startvp = vp;
2445 #endif
2446 	struct vnode *vp1;
2447 	size_t buflen;
2448 	int error;
2449 
2450 	VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp);
2451 	VNPASS(vp->v_usecount > 0, vp);
2452 
2453 	buflen = *len;
2454 
2455 	if (!slash_prefixed) {
2456 		MPASS(*len >= 2);
2457 		buflen--;
2458 		buf[buflen] = '\0';
2459 	}
2460 
2461 	error = 0;
2462 
2463 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
2464 	counter_u64_add(numfullpathcalls, 1);
2465 	while (vp != rdir && vp != rootvnode) {
2466 		/*
2467 		 * The vp vnode must be already fully constructed,
2468 		 * since it is either found in namecache or obtained
2469 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
2470 		 * without obtaining the vnode lock.
2471 		 */
2472 		if ((vp->v_vflag & VV_ROOT) != 0) {
2473 			vn_lock(vp, LK_RETRY | LK_SHARED);
2474 
2475 			/*
2476 			 * With the vnode locked, check for races with
2477 			 * unmount, forced or not.  Note that we
2478 			 * already verified that vp is not equal to
2479 			 * the root vnode, which means that
2480 			 * mnt_vnodecovered can be NULL only for the
2481 			 * case of unmount.
2482 			 */
2483 			if (VN_IS_DOOMED(vp) ||
2484 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
2485 			    vp1->v_mountedhere != vp->v_mount) {
2486 				vput(vp);
2487 				error = ENOENT;
2488 				SDT_PROBE3(vfs, namecache, fullpath, return,
2489 				    error, vp, NULL);
2490 				break;
2491 			}
2492 
2493 			vref(vp1);
2494 			vput(vp);
2495 			vp = vp1;
2496 			continue;
2497 		}
2498 		if (vp->v_type != VDIR) {
2499 			vrele(vp);
2500 			counter_u64_add(numfullpathfail1, 1);
2501 			error = ENOTDIR;
2502 			SDT_PROBE3(vfs, namecache, fullpath, return,
2503 			    error, vp, NULL);
2504 			break;
2505 		}
2506 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2507 		if (error)
2508 			break;
2509 		if (buflen == 0) {
2510 			vrele(vp);
2511 			error = ENOMEM;
2512 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2513 			    startvp, NULL);
2514 			break;
2515 		}
2516 		buf[--buflen] = '/';
2517 		slash_prefixed = true;
2518 	}
2519 	if (error)
2520 		return (error);
2521 	if (!slash_prefixed) {
2522 		if (buflen == 0) {
2523 			vrele(vp);
2524 			counter_u64_add(numfullpathfail4, 1);
2525 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
2526 			    startvp, NULL);
2527 			return (ENOMEM);
2528 		}
2529 		buf[--buflen] = '/';
2530 	}
2531 	counter_u64_add(numfullpathfound, 1);
2532 	vrele(vp);
2533 
2534 	*retbuf = buf + buflen;
2535 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, *retbuf);
2536 	*len -= buflen;
2537 	*len += addend;
2538 	return (0);
2539 }
2540 
2541 /*
2542  * Resolve an arbitrary vnode to a pathname.
2543  *
2544  * Note 2 caveats:
2545  * - hardlinks are not tracked, thus if the vnode is not a directory this can
2546  *   resolve to a different path than the one used to find it
2547  * - namecache is not mandatory, meaning names are not guaranteed to be added
2548  *   (in which case resolving fails)
2549  */
2550 static int
2551 vn_fullpath_any(struct thread *td, struct vnode *vp, struct vnode *rdir,
2552     char *buf, char **retbuf, size_t *buflen)
2553 {
2554 	size_t orig_buflen;
2555 	bool slash_prefixed;
2556 	int error;
2557 
2558 	if (*buflen < 2)
2559 		return (EINVAL);
2560 
2561 	orig_buflen = *buflen;
2562 
2563 	vref(vp);
2564 	slash_prefixed = false;
2565 	if (vp->v_type != VDIR) {
2566 		*buflen -= 1;
2567 		buf[*buflen] = '\0';
2568 		error = vn_vptocnp(&vp, td->td_ucred, buf, buflen);
2569 		if (error)
2570 			return (error);
2571 		if (*buflen == 0) {
2572 			vrele(vp);
2573 			return (ENOMEM);
2574 		}
2575 		*buflen -= 1;
2576 		buf[*buflen] = '/';
2577 		slash_prefixed = true;
2578 	}
2579 
2580 	return (vn_fullpath_dir(td, vp, rdir, buf, retbuf, buflen, slash_prefixed,
2581 	    orig_buflen - *buflen));
2582 }
2583 
2584 /*
2585  * Resolve an arbitrary vnode to a pathname (taking care of hardlinks).
2586  *
2587  * Since the namecache does not track handlings, the caller is expected to first
2588  * look up the target vnode with SAVENAME | WANTPARENT flags passed to namei.
2589  *
2590  * Then we have 2 cases:
2591  * - if the found vnode is a directory, the path can be constructed just by
2592  *   fullowing names up the chain
2593  * - otherwise we populate the buffer with the saved name and start resolving
2594  *   from the parent
2595  */
2596 static int
2597 vn_fullpath_hardlink(struct thread *td, struct nameidata *ndp, char **retbuf,
2598     char **freebuf, size_t *buflen)
2599 {
2600 	char *buf, *tmpbuf;
2601 	struct pwd *pwd;
2602 	struct componentname *cnp;
2603 	struct vnode *vp;
2604 	size_t addend;
2605 	int error;
2606 	bool slash_prefixed;
2607 
2608 	if (*buflen < 2)
2609 		return (EINVAL);
2610 	if (*buflen > MAXPATHLEN)
2611 		*buflen = MAXPATHLEN;
2612 
2613 	slash_prefixed = false;
2614 
2615 	buf = malloc(*buflen, M_TEMP, M_WAITOK);
2616 	pwd = pwd_hold(td);
2617 
2618 	addend = 0;
2619 	vp = ndp->ni_vp;
2620 	if (vp->v_type != VDIR) {
2621 		cnp = &ndp->ni_cnd;
2622 		addend = cnp->cn_namelen + 2;
2623 		if (*buflen < addend) {
2624 			error = ENOMEM;
2625 			goto out_bad;
2626 		}
2627 		*buflen -= addend;
2628 		tmpbuf = buf + *buflen;
2629 		tmpbuf[0] = '/';
2630 		memcpy(&tmpbuf[1], cnp->cn_nameptr, cnp->cn_namelen);
2631 		tmpbuf[addend - 1] = '\0';
2632 		slash_prefixed = true;
2633 		vp = ndp->ni_dvp;
2634 	}
2635 
2636 	vref(vp);
2637 	error = vn_fullpath_dir(td, vp, pwd->pwd_rdir, buf, retbuf, buflen,
2638 	    slash_prefixed, addend);
2639 	if (error != 0)
2640 		goto out_bad;
2641 
2642 	pwd_drop(pwd);
2643 	*freebuf = buf;
2644 
2645 	return (0);
2646 out_bad:
2647 	pwd_drop(pwd);
2648 	free(buf, M_TEMP);
2649 	return (error);
2650 }
2651 
2652 struct vnode *
2653 vn_dir_dd_ino(struct vnode *vp)
2654 {
2655 	struct namecache *ncp;
2656 	struct vnode *ddvp;
2657 	struct mtx *vlp;
2658 	enum vgetstate vs;
2659 
2660 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
2661 	vlp = VP2VNODELOCK(vp);
2662 	mtx_lock(vlp);
2663 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
2664 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
2665 			continue;
2666 		ddvp = ncp->nc_dvp;
2667 		vs = vget_prep(ddvp);
2668 		mtx_unlock(vlp);
2669 		if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
2670 			return (NULL);
2671 		return (ddvp);
2672 	}
2673 	mtx_unlock(vlp);
2674 	return (NULL);
2675 }
2676 
2677 int
2678 vn_commname(struct vnode *vp, char *buf, u_int buflen)
2679 {
2680 	struct namecache *ncp;
2681 	struct mtx *vlp;
2682 	int l;
2683 
2684 	vlp = VP2VNODELOCK(vp);
2685 	mtx_lock(vlp);
2686 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
2687 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2688 			break;
2689 	if (ncp == NULL) {
2690 		mtx_unlock(vlp);
2691 		return (ENOENT);
2692 	}
2693 	l = min(ncp->nc_nlen, buflen - 1);
2694 	memcpy(buf, ncp->nc_name, l);
2695 	mtx_unlock(vlp);
2696 	buf[l] = '\0';
2697 	return (0);
2698 }
2699 
2700 /*
2701  * This function updates path string to vnode's full global path
2702  * and checks the size of the new path string against the pathlen argument.
2703  *
2704  * Requires a locked, referenced vnode.
2705  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
2706  *
2707  * If vp is a directory, the call to vn_fullpath_global() always succeeds
2708  * because it falls back to the ".." lookup if the namecache lookup fails.
2709  */
2710 int
2711 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
2712     u_int pathlen)
2713 {
2714 	struct nameidata nd;
2715 	struct vnode *vp1;
2716 	char *rpath, *fbuf;
2717 	int error;
2718 
2719 	ASSERT_VOP_ELOCKED(vp, __func__);
2720 
2721 	/* Construct global filesystem path from vp. */
2722 	VOP_UNLOCK(vp);
2723 	error = vn_fullpath_global(td, vp, &rpath, &fbuf);
2724 
2725 	if (error != 0) {
2726 		vrele(vp);
2727 		return (error);
2728 	}
2729 
2730 	if (strlen(rpath) >= pathlen) {
2731 		vrele(vp);
2732 		error = ENAMETOOLONG;
2733 		goto out;
2734 	}
2735 
2736 	/*
2737 	 * Re-lookup the vnode by path to detect a possible rename.
2738 	 * As a side effect, the vnode is relocked.
2739 	 * If vnode was renamed, return ENOENT.
2740 	 */
2741 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
2742 	    UIO_SYSSPACE, path, td);
2743 	error = namei(&nd);
2744 	if (error != 0) {
2745 		vrele(vp);
2746 		goto out;
2747 	}
2748 	NDFREE(&nd, NDF_ONLY_PNBUF);
2749 	vp1 = nd.ni_vp;
2750 	vrele(vp);
2751 	if (vp1 == vp)
2752 		strcpy(path, rpath);
2753 	else {
2754 		vput(vp1);
2755 		error = ENOENT;
2756 	}
2757 
2758 out:
2759 	free(fbuf, M_TEMP);
2760 	return (error);
2761 }
2762 
2763 #ifdef DDB
2764 static void
2765 db_print_vpath(struct vnode *vp)
2766 {
2767 
2768 	while (vp != NULL) {
2769 		db_printf("%p: ", vp);
2770 		if (vp == rootvnode) {
2771 			db_printf("/");
2772 			vp = NULL;
2773 		} else {
2774 			if (vp->v_vflag & VV_ROOT) {
2775 				db_printf("<mount point>");
2776 				vp = vp->v_mount->mnt_vnodecovered;
2777 			} else {
2778 				struct namecache *ncp;
2779 				char *ncn;
2780 				int i;
2781 
2782 				ncp = TAILQ_FIRST(&vp->v_cache_dst);
2783 				if (ncp != NULL) {
2784 					ncn = ncp->nc_name;
2785 					for (i = 0; i < ncp->nc_nlen; i++)
2786 						db_printf("%c", *ncn++);
2787 					vp = ncp->nc_dvp;
2788 				} else {
2789 					vp = NULL;
2790 				}
2791 			}
2792 		}
2793 		db_printf("\n");
2794 	}
2795 
2796 	return;
2797 }
2798 
2799 DB_SHOW_COMMAND(vpath, db_show_vpath)
2800 {
2801 	struct vnode *vp;
2802 
2803 	if (!have_addr) {
2804 		db_printf("usage: show vpath <struct vnode *>\n");
2805 		return;
2806 	}
2807 
2808 	vp = (struct vnode *)addr;
2809 	db_print_vpath(vp);
2810 }
2811 
2812 #endif
2813