xref: /freebsd/sys/kern/vfs_cache.c (revision 036d2e814bf0f5d88ffb4b24c159320894541757)
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 SYSCTL_ULONG(_debug, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0,
208     "Number of negative entries in namecache");
209 static u_long __exclusive_cache_line	numcache;/* number of cache entries allocated */
210 SYSCTL_ULONG(_debug, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0,
211     "Number of namecache entries");
212 static u_long __exclusive_cache_line	numcachehv;/* number of cache entries with vnodes held */
213 SYSCTL_ULONG(_debug, OID_AUTO, numcachehv, CTLFLAG_RD, &numcachehv, 0,
214     "Number of namecache entries with vnodes held");
215 u_int ncsizefactor = 2;
216 SYSCTL_UINT(_vfs, OID_AUTO, ncsizefactor, CTLFLAG_RW, &ncsizefactor, 0,
217     "Size factor for namecache");
218 static u_int __read_mostly	ncpurgeminvnodes;
219 SYSCTL_UINT(_vfs, OID_AUTO, ncpurgeminvnodes, CTLFLAG_RW, &ncpurgeminvnodes, 0,
220     "Number of vnodes below which purgevfs ignores the request");
221 static u_int __read_mostly	ncsize; /* the size as computed on creation or resizing */
222 
223 struct nchstats	nchstats;		/* cache effectiveness statistics */
224 
225 static struct mtx __exclusive_cache_line	ncneg_shrink_lock;
226 static int	shrink_list_turn;
227 
228 struct neglist {
229 	struct mtx		nl_lock;
230 	TAILQ_HEAD(, namecache) nl_list;
231 } __aligned(CACHE_LINE_SIZE);
232 
233 static struct neglist __read_mostly	*neglists;
234 static struct neglist ncneg_hot;
235 static u_long numhotneg;
236 
237 #define	numneglists (ncneghash + 1)
238 static u_int __read_mostly	ncneghash;
239 static inline struct neglist *
240 NCP2NEGLIST(struct namecache *ncp)
241 {
242 
243 	return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
244 }
245 
246 #define	numbucketlocks (ncbuckethash + 1)
247 static u_int __read_mostly  ncbuckethash;
248 static struct rwlock_padalign __read_mostly  *bucketlocks;
249 #define	HASH2BUCKETLOCK(hash) \
250 	((struct rwlock *)(&bucketlocks[((hash) & ncbuckethash)]))
251 
252 #define	numvnodelocks (ncvnodehash + 1)
253 static u_int __read_mostly  ncvnodehash;
254 static struct mtx __read_mostly *vnodelocks;
255 static inline struct mtx *
256 VP2VNODELOCK(struct vnode *vp)
257 {
258 
259 	return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
260 }
261 
262 /*
263  * UMA zones for the VFS cache.
264  *
265  * The small cache is used for entries with short names, which are the
266  * most common.  The large cache is used for entries which are too big to
267  * fit in the small cache.
268  */
269 static uma_zone_t __read_mostly cache_zone_small;
270 static uma_zone_t __read_mostly cache_zone_small_ts;
271 static uma_zone_t __read_mostly cache_zone_large;
272 static uma_zone_t __read_mostly cache_zone_large_ts;
273 
274 #define	CACHE_PATH_CUTOFF	35
275 
276 static struct namecache *
277 cache_alloc(int len, int ts)
278 {
279 	struct namecache_ts *ncp_ts;
280 	struct namecache *ncp;
281 
282 	if (__predict_false(ts)) {
283 		if (len <= CACHE_PATH_CUTOFF)
284 			ncp_ts = uma_zalloc(cache_zone_small_ts, M_WAITOK);
285 		else
286 			ncp_ts = uma_zalloc(cache_zone_large_ts, M_WAITOK);
287 		ncp = &ncp_ts->nc_nc;
288 	} else {
289 		if (len <= CACHE_PATH_CUTOFF)
290 			ncp = uma_zalloc(cache_zone_small, M_WAITOK);
291 		else
292 			ncp = uma_zalloc(cache_zone_large, M_WAITOK);
293 	}
294 	return (ncp);
295 }
296 
297 static void
298 cache_free(struct namecache *ncp)
299 {
300 	struct namecache_ts *ncp_ts;
301 
302 	if (ncp == NULL)
303 		return;
304 	if ((ncp->nc_flag & NCF_DVDROP) != 0)
305 		vdrop(ncp->nc_dvp);
306 	if (__predict_false(ncp->nc_flag & NCF_TS)) {
307 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
308 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
309 			uma_zfree(cache_zone_small_ts, ncp_ts);
310 		else
311 			uma_zfree(cache_zone_large_ts, ncp_ts);
312 	} else {
313 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
314 			uma_zfree(cache_zone_small, ncp);
315 		else
316 			uma_zfree(cache_zone_large, ncp);
317 	}
318 }
319 
320 static void
321 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
322 {
323 	struct namecache_ts *ncp_ts;
324 
325 	KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
326 	    (tsp == NULL && ticksp == NULL),
327 	    ("No NCF_TS"));
328 
329 	if (tsp == NULL && ticksp == NULL)
330 		return;
331 
332 	ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
333 	if (tsp != NULL)
334 		*tsp = ncp_ts->nc_time;
335 	if (ticksp != NULL)
336 		*ticksp = ncp_ts->nc_ticks;
337 }
338 
339 static int __read_mostly	doingcache = 1;	/* 1 => enable the cache */
340 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
341     "VFS namecache enabled");
342 
343 /* Export size information to userland */
344 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
345     sizeof(struct namecache), "sizeof(struct namecache)");
346 
347 /*
348  * The new name cache statistics
349  */
350 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0,
351     "Name cache statistics");
352 #define STATNODE_ULONG(name, descr)	\
353 	SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, descr);
354 #define STATNODE_COUNTER(name, descr)	\
355 	static counter_u64_t __read_mostly name; \
356 	SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, descr);
357 STATNODE_ULONG(numneg, "Number of negative cache entries");
358 STATNODE_ULONG(numcache, "Number of cache entries");
359 STATNODE_COUNTER(numcalls, "Number of cache lookups");
360 STATNODE_COUNTER(dothits, "Number of '.' hits");
361 STATNODE_COUNTER(dotdothits, "Number of '..' hits");
362 STATNODE_COUNTER(numchecks, "Number of checks in lookup");
363 STATNODE_COUNTER(nummiss, "Number of cache misses");
364 STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache");
365 STATNODE_COUNTER(numposzaps,
366     "Number of cache hits (positive) we do not want to cache");
367 STATNODE_COUNTER(numposhits, "Number of cache hits (positive)");
368 STATNODE_COUNTER(numnegzaps,
369     "Number of cache hits (negative) we do not want to cache");
370 STATNODE_COUNTER(numneghits, "Number of cache hits (negative)");
371 /* These count for kern___getcwd(), too. */
372 STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls");
373 STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
374 STATNODE_COUNTER(numfullpathfail2,
375     "Number of fullpath search errors (VOP_VPTOCNP failures)");
376 STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
377 STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls");
378 STATNODE_COUNTER(zap_and_exit_bucket_relock_success,
379     "Number of successful removals after relocking");
380 static long zap_and_exit_bucket_fail; STATNODE_ULONG(zap_and_exit_bucket_fail,
381     "Number of times zap_and_exit failed to lock");
382 static long zap_and_exit_bucket_fail2; STATNODE_ULONG(zap_and_exit_bucket_fail2,
383     "Number of times zap_and_exit failed to lock");
384 static long cache_lock_vnodes_cel_3_failures;
385 STATNODE_ULONG(cache_lock_vnodes_cel_3_failures,
386     "Number of times 3-way vnode locking failed");
387 STATNODE_ULONG(numhotneg, "Number of hot negative entries");
388 STATNODE_COUNTER(numneg_evicted,
389     "Number of negative entries evicted when adding a new entry");
390 STATNODE_COUNTER(shrinking_skipped,
391     "Number of times shrinking was already in progress");
392 
393 static void cache_zap_locked(struct namecache *ncp, bool neg_locked);
394 static int vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
395     char *buf, char **retbuf, u_int buflen);
396 
397 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
398 
399 static int cache_yield;
400 SYSCTL_INT(_vfs_cache, OID_AUTO, yield, CTLFLAG_RD, &cache_yield, 0,
401     "Number of times cache called yield");
402 
403 static void __noinline
404 cache_maybe_yield(void)
405 {
406 
407 	if (should_yield()) {
408 		cache_yield++;
409 		kern_yield(PRI_USER);
410 	}
411 }
412 
413 static inline void
414 cache_assert_vlp_locked(struct mtx *vlp)
415 {
416 
417 	if (vlp != NULL)
418 		mtx_assert(vlp, MA_OWNED);
419 }
420 
421 static inline void
422 cache_assert_vnode_locked(struct vnode *vp)
423 {
424 	struct mtx *vlp;
425 
426 	vlp = VP2VNODELOCK(vp);
427 	cache_assert_vlp_locked(vlp);
428 }
429 
430 static uint32_t
431 cache_get_hash(char *name, u_char len, struct vnode *dvp)
432 {
433 	uint32_t hash;
434 
435 	hash = fnv_32_buf(name, len, FNV1_32_INIT);
436 	hash = fnv_32_buf(&dvp, sizeof(dvp), hash);
437 	return (hash);
438 }
439 
440 static inline struct rwlock *
441 NCP2BUCKETLOCK(struct namecache *ncp)
442 {
443 	uint32_t hash;
444 
445 	hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
446 	return (HASH2BUCKETLOCK(hash));
447 }
448 
449 #ifdef INVARIANTS
450 static void
451 cache_assert_bucket_locked(struct namecache *ncp, int mode)
452 {
453 	struct rwlock *blp;
454 
455 	blp = NCP2BUCKETLOCK(ncp);
456 	rw_assert(blp, mode);
457 }
458 #else
459 #define cache_assert_bucket_locked(x, y) do { } while (0)
460 #endif
461 
462 #define cache_sort_vnodes(x, y)	_cache_sort_vnodes((void **)(x), (void **)(y))
463 static void
464 _cache_sort_vnodes(void **p1, void **p2)
465 {
466 	void *tmp;
467 
468 	MPASS(*p1 != NULL || *p2 != NULL);
469 
470 	if (*p1 > *p2) {
471 		tmp = *p2;
472 		*p2 = *p1;
473 		*p1 = tmp;
474 	}
475 }
476 
477 static void
478 cache_lock_all_buckets(void)
479 {
480 	u_int i;
481 
482 	for (i = 0; i < numbucketlocks; i++)
483 		rw_wlock(&bucketlocks[i]);
484 }
485 
486 static void
487 cache_unlock_all_buckets(void)
488 {
489 	u_int i;
490 
491 	for (i = 0; i < numbucketlocks; i++)
492 		rw_wunlock(&bucketlocks[i]);
493 }
494 
495 static void
496 cache_lock_all_vnodes(void)
497 {
498 	u_int i;
499 
500 	for (i = 0; i < numvnodelocks; i++)
501 		mtx_lock(&vnodelocks[i]);
502 }
503 
504 static void
505 cache_unlock_all_vnodes(void)
506 {
507 	u_int i;
508 
509 	for (i = 0; i < numvnodelocks; i++)
510 		mtx_unlock(&vnodelocks[i]);
511 }
512 
513 static int
514 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
515 {
516 
517 	cache_sort_vnodes(&vlp1, &vlp2);
518 
519 	if (vlp1 != NULL) {
520 		if (!mtx_trylock(vlp1))
521 			return (EAGAIN);
522 	}
523 	if (!mtx_trylock(vlp2)) {
524 		if (vlp1 != NULL)
525 			mtx_unlock(vlp1);
526 		return (EAGAIN);
527 	}
528 
529 	return (0);
530 }
531 
532 static void
533 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
534 {
535 
536 	MPASS(vlp1 != NULL || vlp2 != NULL);
537 	MPASS(vlp1 <= vlp2);
538 
539 	if (vlp1 != NULL)
540 		mtx_lock(vlp1);
541 	if (vlp2 != NULL)
542 		mtx_lock(vlp2);
543 }
544 
545 static void
546 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
547 {
548 
549 	MPASS(vlp1 != NULL || vlp2 != NULL);
550 
551 	if (vlp1 != NULL)
552 		mtx_unlock(vlp1);
553 	if (vlp2 != NULL)
554 		mtx_unlock(vlp2);
555 }
556 
557 static int
558 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
559 {
560 	struct nchstats snap;
561 
562 	if (req->oldptr == NULL)
563 		return (SYSCTL_OUT(req, 0, sizeof(snap)));
564 
565 	snap = nchstats;
566 	snap.ncs_goodhits = counter_u64_fetch(numposhits);
567 	snap.ncs_neghits = counter_u64_fetch(numneghits);
568 	snap.ncs_badhits = counter_u64_fetch(numposzaps) +
569 	    counter_u64_fetch(numnegzaps);
570 	snap.ncs_miss = counter_u64_fetch(nummisszap) +
571 	    counter_u64_fetch(nummiss);
572 
573 	return (SYSCTL_OUT(req, &snap, sizeof(snap)));
574 }
575 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
576     CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
577     "VFS cache effectiveness statistics");
578 
579 #ifdef DIAGNOSTIC
580 /*
581  * Grab an atomic snapshot of the name cache hash chain lengths
582  */
583 static SYSCTL_NODE(_debug, OID_AUTO, hashstat, CTLFLAG_RW, NULL,
584     "hash table stats");
585 
586 static int
587 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
588 {
589 	struct nchashhead *ncpp;
590 	struct namecache *ncp;
591 	int i, error, n_nchash, *cntbuf;
592 
593 retry:
594 	n_nchash = nchash + 1;	/* nchash is max index, not count */
595 	if (req->oldptr == NULL)
596 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
597 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
598 	cache_lock_all_buckets();
599 	if (n_nchash != nchash + 1) {
600 		cache_unlock_all_buckets();
601 		free(cntbuf, M_TEMP);
602 		goto retry;
603 	}
604 	/* Scan hash tables counting entries */
605 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
606 		LIST_FOREACH(ncp, ncpp, nc_hash)
607 			cntbuf[i]++;
608 	cache_unlock_all_buckets();
609 	for (error = 0, i = 0; i < n_nchash; i++)
610 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
611 			break;
612 	free(cntbuf, M_TEMP);
613 	return (error);
614 }
615 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
616     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
617     "nchash chain lengths");
618 
619 static int
620 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
621 {
622 	int error;
623 	struct nchashhead *ncpp;
624 	struct namecache *ncp;
625 	int n_nchash;
626 	int count, maxlength, used, pct;
627 
628 	if (!req->oldptr)
629 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
630 
631 	cache_lock_all_buckets();
632 	n_nchash = nchash + 1;	/* nchash is max index, not count */
633 	used = 0;
634 	maxlength = 0;
635 
636 	/* Scan hash tables for applicable entries */
637 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
638 		count = 0;
639 		LIST_FOREACH(ncp, ncpp, nc_hash) {
640 			count++;
641 		}
642 		if (count)
643 			used++;
644 		if (maxlength < count)
645 			maxlength = count;
646 	}
647 	n_nchash = nchash + 1;
648 	cache_unlock_all_buckets();
649 	pct = (used * 100) / (n_nchash / 100);
650 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
651 	if (error)
652 		return (error);
653 	error = SYSCTL_OUT(req, &used, sizeof(used));
654 	if (error)
655 		return (error);
656 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
657 	if (error)
658 		return (error);
659 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
660 	if (error)
661 		return (error);
662 	return (0);
663 }
664 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
665     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
666     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
667 #endif
668 
669 /*
670  * Negative entries management
671  *
672  * A variation of LRU scheme is used. New entries are hashed into one of
673  * numneglists cold lists. Entries get promoted to the hot list on first hit.
674  *
675  * The shrinker will demote hot list head and evict from the cold list in a
676  * round-robin manner.
677  */
678 static void
679 cache_negative_hit(struct namecache *ncp)
680 {
681 	struct neglist *neglist;
682 
683 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
684 	if (ncp->nc_flag & NCF_HOTNEGATIVE)
685 		return;
686 	neglist = NCP2NEGLIST(ncp);
687 	mtx_lock(&ncneg_hot.nl_lock);
688 	mtx_lock(&neglist->nl_lock);
689 	if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
690 		numhotneg++;
691 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
692 		TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
693 		ncp->nc_flag |= NCF_HOTNEGATIVE;
694 	}
695 	mtx_unlock(&neglist->nl_lock);
696 	mtx_unlock(&ncneg_hot.nl_lock);
697 }
698 
699 static void
700 cache_negative_insert(struct namecache *ncp, bool neg_locked)
701 {
702 	struct neglist *neglist;
703 
704 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
705 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
706 	neglist = NCP2NEGLIST(ncp);
707 	if (!neg_locked) {
708 		mtx_lock(&neglist->nl_lock);
709 	} else {
710 		mtx_assert(&neglist->nl_lock, MA_OWNED);
711 	}
712 	TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
713 	if (!neg_locked)
714 		mtx_unlock(&neglist->nl_lock);
715 	atomic_add_rel_long(&numneg, 1);
716 }
717 
718 static void
719 cache_negative_remove(struct namecache *ncp, bool neg_locked)
720 {
721 	struct neglist *neglist;
722 	bool hot_locked = false;
723 	bool list_locked = false;
724 
725 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
726 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
727 	neglist = NCP2NEGLIST(ncp);
728 	if (!neg_locked) {
729 		if (ncp->nc_flag & NCF_HOTNEGATIVE) {
730 			hot_locked = true;
731 			mtx_lock(&ncneg_hot.nl_lock);
732 			if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
733 				list_locked = true;
734 				mtx_lock(&neglist->nl_lock);
735 			}
736 		} else {
737 			list_locked = true;
738 			mtx_lock(&neglist->nl_lock);
739 		}
740 	}
741 	if (ncp->nc_flag & NCF_HOTNEGATIVE) {
742 		mtx_assert(&ncneg_hot.nl_lock, MA_OWNED);
743 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
744 		numhotneg--;
745 	} else {
746 		mtx_assert(&neglist->nl_lock, MA_OWNED);
747 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
748 	}
749 	if (list_locked)
750 		mtx_unlock(&neglist->nl_lock);
751 	if (hot_locked)
752 		mtx_unlock(&ncneg_hot.nl_lock);
753 	atomic_subtract_rel_long(&numneg, 1);
754 }
755 
756 static void
757 cache_negative_shrink_select(int start, struct namecache **ncpp,
758     struct neglist **neglistpp)
759 {
760 	struct neglist *neglist;
761 	struct namecache *ncp;
762 	int i;
763 
764 	*ncpp = ncp = NULL;
765 	neglist = NULL;
766 
767 	for (i = start; i < numneglists; i++) {
768 		neglist = &neglists[i];
769 		if (TAILQ_FIRST(&neglist->nl_list) == NULL)
770 			continue;
771 		mtx_lock(&neglist->nl_lock);
772 		ncp = TAILQ_FIRST(&neglist->nl_list);
773 		if (ncp != NULL)
774 			break;
775 		mtx_unlock(&neglist->nl_lock);
776 	}
777 
778 	*neglistpp = neglist;
779 	*ncpp = ncp;
780 }
781 
782 static void
783 cache_negative_zap_one(void)
784 {
785 	struct namecache *ncp, *ncp2;
786 	struct neglist *neglist;
787 	struct mtx *dvlp;
788 	struct rwlock *blp;
789 
790 	if (mtx_owner(&ncneg_shrink_lock) != NULL ||
791 	    !mtx_trylock(&ncneg_shrink_lock)) {
792 		counter_u64_add(shrinking_skipped, 1);
793 		return;
794 	}
795 
796 	mtx_lock(&ncneg_hot.nl_lock);
797 	ncp = TAILQ_FIRST(&ncneg_hot.nl_list);
798 	if (ncp != NULL) {
799 		neglist = NCP2NEGLIST(ncp);
800 		mtx_lock(&neglist->nl_lock);
801 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
802 		TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
803 		ncp->nc_flag &= ~NCF_HOTNEGATIVE;
804 		numhotneg--;
805 		mtx_unlock(&neglist->nl_lock);
806 	}
807 	mtx_unlock(&ncneg_hot.nl_lock);
808 
809 	cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
810 	shrink_list_turn++;
811 	if (shrink_list_turn == numneglists)
812 		shrink_list_turn = 0;
813 	if (ncp == NULL && shrink_list_turn == 0)
814 		cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
815 	mtx_unlock(&ncneg_shrink_lock);
816 	if (ncp == NULL)
817 		return;
818 
819 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
820 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
821 	blp = NCP2BUCKETLOCK(ncp);
822 	mtx_unlock(&neglist->nl_lock);
823 	mtx_lock(dvlp);
824 	rw_wlock(blp);
825 	mtx_lock(&neglist->nl_lock);
826 	ncp2 = TAILQ_FIRST(&neglist->nl_list);
827 	if (ncp != ncp2 || dvlp != VP2VNODELOCK(ncp2->nc_dvp) ||
828 	    blp != NCP2BUCKETLOCK(ncp2) || !(ncp2->nc_flag & NCF_NEGATIVE)) {
829 		ncp = NULL;
830 	} else {
831 		SDT_PROBE2(vfs, namecache, shrink_negative, done, ncp->nc_dvp,
832 		    ncp->nc_name);
833 
834 		cache_zap_locked(ncp, true);
835 		counter_u64_add(numneg_evicted, 1);
836 	}
837 	mtx_unlock(&neglist->nl_lock);
838 	rw_wunlock(blp);
839 	mtx_unlock(dvlp);
840 	cache_free(ncp);
841 }
842 
843 /*
844  * cache_zap_locked():
845  *
846  *   Removes a namecache entry from cache, whether it contains an actual
847  *   pointer to a vnode or if it is just a negative cache entry.
848  */
849 static void
850 cache_zap_locked(struct namecache *ncp, bool neg_locked)
851 {
852 
853 	if (!(ncp->nc_flag & NCF_NEGATIVE))
854 		cache_assert_vnode_locked(ncp->nc_vp);
855 	cache_assert_vnode_locked(ncp->nc_dvp);
856 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
857 
858 	CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp,
859 	    (ncp->nc_flag & NCF_NEGATIVE) ? NULL : ncp->nc_vp);
860 	LIST_REMOVE(ncp, nc_hash);
861 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
862 		SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
863 		    ncp->nc_name, ncp->nc_vp);
864 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
865 		if (ncp == ncp->nc_vp->v_cache_dd)
866 			ncp->nc_vp->v_cache_dd = NULL;
867 	} else {
868 		SDT_PROBE2(vfs, namecache, zap_negative, done, ncp->nc_dvp,
869 		    ncp->nc_name);
870 		cache_negative_remove(ncp, neg_locked);
871 	}
872 	if (ncp->nc_flag & NCF_ISDOTDOT) {
873 		if (ncp == ncp->nc_dvp->v_cache_dd)
874 			ncp->nc_dvp->v_cache_dd = NULL;
875 	} else {
876 		LIST_REMOVE(ncp, nc_src);
877 		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
878 			ncp->nc_flag |= NCF_DVDROP;
879 			atomic_subtract_rel_long(&numcachehv, 1);
880 		}
881 	}
882 	atomic_subtract_rel_long(&numcache, 1);
883 }
884 
885 static void
886 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
887 {
888 	struct rwlock *blp;
889 
890 	MPASS(ncp->nc_dvp == vp);
891 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
892 	cache_assert_vnode_locked(vp);
893 
894 	blp = NCP2BUCKETLOCK(ncp);
895 	rw_wlock(blp);
896 	cache_zap_locked(ncp, false);
897 	rw_wunlock(blp);
898 }
899 
900 static bool
901 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
902     struct mtx **vlpp)
903 {
904 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
905 	struct rwlock *blp;
906 
907 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
908 	cache_assert_vnode_locked(vp);
909 
910 	if (ncp->nc_flag & NCF_NEGATIVE) {
911 		if (*vlpp != NULL) {
912 			mtx_unlock(*vlpp);
913 			*vlpp = NULL;
914 		}
915 		cache_zap_negative_locked_vnode_kl(ncp, vp);
916 		return (true);
917 	}
918 
919 	pvlp = VP2VNODELOCK(vp);
920 	blp = NCP2BUCKETLOCK(ncp);
921 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
922 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
923 
924 	if (*vlpp == vlp1 || *vlpp == vlp2) {
925 		to_unlock = *vlpp;
926 		*vlpp = NULL;
927 	} else {
928 		if (*vlpp != NULL) {
929 			mtx_unlock(*vlpp);
930 			*vlpp = NULL;
931 		}
932 		cache_sort_vnodes(&vlp1, &vlp2);
933 		if (vlp1 == pvlp) {
934 			mtx_lock(vlp2);
935 			to_unlock = vlp2;
936 		} else {
937 			if (!mtx_trylock(vlp1))
938 				goto out_relock;
939 			to_unlock = vlp1;
940 		}
941 	}
942 	rw_wlock(blp);
943 	cache_zap_locked(ncp, false);
944 	rw_wunlock(blp);
945 	if (to_unlock != NULL)
946 		mtx_unlock(to_unlock);
947 	return (true);
948 
949 out_relock:
950 	mtx_unlock(vlp2);
951 	mtx_lock(vlp1);
952 	mtx_lock(vlp2);
953 	MPASS(*vlpp == NULL);
954 	*vlpp = vlp1;
955 	return (false);
956 }
957 
958 static int __noinline
959 cache_zap_locked_vnode(struct namecache *ncp, struct vnode *vp)
960 {
961 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
962 	struct rwlock *blp;
963 	int error = 0;
964 
965 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
966 	cache_assert_vnode_locked(vp);
967 
968 	pvlp = VP2VNODELOCK(vp);
969 	if (ncp->nc_flag & NCF_NEGATIVE) {
970 		cache_zap_negative_locked_vnode_kl(ncp, vp);
971 		goto out;
972 	}
973 
974 	blp = NCP2BUCKETLOCK(ncp);
975 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
976 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
977 	cache_sort_vnodes(&vlp1, &vlp2);
978 	if (vlp1 == pvlp) {
979 		mtx_lock(vlp2);
980 		to_unlock = vlp2;
981 	} else {
982 		if (!mtx_trylock(vlp1)) {
983 			error = EAGAIN;
984 			goto out;
985 		}
986 		to_unlock = vlp1;
987 	}
988 	rw_wlock(blp);
989 	cache_zap_locked(ncp, false);
990 	rw_wunlock(blp);
991 	mtx_unlock(to_unlock);
992 out:
993 	mtx_unlock(pvlp);
994 	return (error);
995 }
996 
997 /*
998  * If trylocking failed we can get here. We know enough to take all needed locks
999  * in the right order and re-lookup the entry.
1000  */
1001 static int
1002 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1003     struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1004     struct rwlock *blp)
1005 {
1006 	struct namecache *rncp;
1007 
1008 	cache_assert_bucket_locked(ncp, RA_UNLOCKED);
1009 
1010 	cache_sort_vnodes(&dvlp, &vlp);
1011 	cache_lock_vnodes(dvlp, vlp);
1012 	rw_wlock(blp);
1013 	LIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1014 		if (rncp == ncp && rncp->nc_dvp == dvp &&
1015 		    rncp->nc_nlen == cnp->cn_namelen &&
1016 		    !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen))
1017 			break;
1018 	}
1019 	if (rncp != NULL) {
1020 		cache_zap_locked(rncp, false);
1021 		rw_wunlock(blp);
1022 		cache_unlock_vnodes(dvlp, vlp);
1023 		counter_u64_add(zap_and_exit_bucket_relock_success, 1);
1024 		return (0);
1025 	}
1026 
1027 	rw_wunlock(blp);
1028 	cache_unlock_vnodes(dvlp, vlp);
1029 	return (EAGAIN);
1030 }
1031 
1032 static int __noinline
1033 cache_zap_wlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1034     uint32_t hash, struct rwlock *blp)
1035 {
1036 	struct mtx *dvlp, *vlp;
1037 	struct vnode *dvp;
1038 
1039 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1040 
1041 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1042 	vlp = NULL;
1043 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1044 		vlp = VP2VNODELOCK(ncp->nc_vp);
1045 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1046 		cache_zap_locked(ncp, false);
1047 		rw_wunlock(blp);
1048 		cache_unlock_vnodes(dvlp, vlp);
1049 		return (0);
1050 	}
1051 
1052 	dvp = ncp->nc_dvp;
1053 	rw_wunlock(blp);
1054 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1055 }
1056 
1057 static int __noinline
1058 cache_zap_rlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1059     uint32_t hash, struct rwlock *blp)
1060 {
1061 	struct mtx *dvlp, *vlp;
1062 	struct vnode *dvp;
1063 
1064 	cache_assert_bucket_locked(ncp, RA_RLOCKED);
1065 
1066 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1067 	vlp = NULL;
1068 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1069 		vlp = VP2VNODELOCK(ncp->nc_vp);
1070 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1071 		rw_runlock(blp);
1072 		rw_wlock(blp);
1073 		cache_zap_locked(ncp, false);
1074 		rw_wunlock(blp);
1075 		cache_unlock_vnodes(dvlp, vlp);
1076 		return (0);
1077 	}
1078 
1079 	dvp = ncp->nc_dvp;
1080 	rw_runlock(blp);
1081 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1082 }
1083 
1084 static int
1085 cache_zap_wlocked_bucket_kl(struct namecache *ncp, struct rwlock *blp,
1086     struct mtx **vlpp1, struct mtx **vlpp2)
1087 {
1088 	struct mtx *dvlp, *vlp;
1089 
1090 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1091 
1092 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1093 	vlp = NULL;
1094 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1095 		vlp = VP2VNODELOCK(ncp->nc_vp);
1096 	cache_sort_vnodes(&dvlp, &vlp);
1097 
1098 	if (*vlpp1 == dvlp && *vlpp2 == vlp) {
1099 		cache_zap_locked(ncp, false);
1100 		cache_unlock_vnodes(dvlp, vlp);
1101 		*vlpp1 = NULL;
1102 		*vlpp2 = NULL;
1103 		return (0);
1104 	}
1105 
1106 	if (*vlpp1 != NULL)
1107 		mtx_unlock(*vlpp1);
1108 	if (*vlpp2 != NULL)
1109 		mtx_unlock(*vlpp2);
1110 	*vlpp1 = NULL;
1111 	*vlpp2 = NULL;
1112 
1113 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1114 		cache_zap_locked(ncp, false);
1115 		cache_unlock_vnodes(dvlp, vlp);
1116 		return (0);
1117 	}
1118 
1119 	rw_wunlock(blp);
1120 	*vlpp1 = dvlp;
1121 	*vlpp2 = vlp;
1122 	if (*vlpp1 != NULL)
1123 		mtx_lock(*vlpp1);
1124 	mtx_lock(*vlpp2);
1125 	rw_wlock(blp);
1126 	return (EAGAIN);
1127 }
1128 
1129 static void
1130 cache_lookup_unlock(struct rwlock *blp, struct mtx *vlp)
1131 {
1132 
1133 	if (blp != NULL) {
1134 		rw_runlock(blp);
1135 	} else {
1136 		mtx_unlock(vlp);
1137 	}
1138 }
1139 
1140 static int __noinline
1141 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1142     struct timespec *tsp, int *ticksp)
1143 {
1144 	int ltype;
1145 
1146 	*vpp = dvp;
1147 	CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .",
1148 			dvp, cnp->cn_nameptr);
1149 	counter_u64_add(dothits, 1);
1150 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1151 	if (tsp != NULL)
1152 		timespecclear(tsp);
1153 	if (ticksp != NULL)
1154 		*ticksp = ticks;
1155 	vrefact(*vpp);
1156 	/*
1157 	 * When we lookup "." we still can be asked to lock it
1158 	 * differently...
1159 	 */
1160 	ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1161 	if (ltype != VOP_ISLOCKED(*vpp)) {
1162 		if (ltype == LK_EXCLUSIVE) {
1163 			vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1164 			if ((*vpp)->v_iflag & VI_DOOMED) {
1165 				/* forced unmount */
1166 				vrele(*vpp);
1167 				*vpp = NULL;
1168 				return (ENOENT);
1169 			}
1170 		} else
1171 			vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1172 	}
1173 	return (-1);
1174 }
1175 
1176 static __noinline int
1177 cache_lookup_nomakeentry(struct vnode *dvp, struct vnode **vpp,
1178     struct componentname *cnp, struct timespec *tsp, int *ticksp)
1179 {
1180 	struct namecache *ncp;
1181 	struct rwlock *blp;
1182 	struct mtx *dvlp, *dvlp2;
1183 	uint32_t hash;
1184 	int error;
1185 
1186 	if (cnp->cn_namelen == 2 &&
1187 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1188 		counter_u64_add(dotdothits, 1);
1189 		dvlp = VP2VNODELOCK(dvp);
1190 		dvlp2 = NULL;
1191 		mtx_lock(dvlp);
1192 retry_dotdot:
1193 		ncp = dvp->v_cache_dd;
1194 		if (ncp == NULL) {
1195 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1196 			    "..", NULL);
1197 			mtx_unlock(dvlp);
1198 			if (dvlp2 != NULL)
1199 				mtx_unlock(dvlp2);
1200 			return (0);
1201 		}
1202 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1203 			if (ncp->nc_dvp != dvp)
1204 				panic("dvp %p v_cache_dd %p\n", dvp, ncp);
1205 			if (!cache_zap_locked_vnode_kl2(ncp,
1206 			    dvp, &dvlp2))
1207 				goto retry_dotdot;
1208 			MPASS(dvp->v_cache_dd == NULL);
1209 			mtx_unlock(dvlp);
1210 			if (dvlp2 != NULL)
1211 				mtx_unlock(dvlp2);
1212 			cache_free(ncp);
1213 		} else {
1214 			dvp->v_cache_dd = NULL;
1215 			mtx_unlock(dvlp);
1216 			if (dvlp2 != NULL)
1217 				mtx_unlock(dvlp2);
1218 		}
1219 		return (0);
1220 	}
1221 
1222 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1223 	blp = HASH2BUCKETLOCK(hash);
1224 retry:
1225 	if (LIST_EMPTY(NCHHASH(hash)))
1226 		goto out_no_entry;
1227 
1228 	rw_wlock(blp);
1229 
1230 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1231 		counter_u64_add(numchecks, 1);
1232 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1233 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1234 			break;
1235 	}
1236 
1237 	/* We failed to find an entry */
1238 	if (ncp == NULL) {
1239 		rw_wunlock(blp);
1240 		goto out_no_entry;
1241 	}
1242 
1243 	error = cache_zap_wlocked_bucket(ncp, cnp, hash, blp);
1244 	if (__predict_false(error != 0)) {
1245 		zap_and_exit_bucket_fail++;
1246 		cache_maybe_yield();
1247 		goto retry;
1248 	}
1249 	counter_u64_add(numposzaps, 1);
1250 	cache_free(ncp);
1251 	return (0);
1252 out_no_entry:
1253 	SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr, NULL);
1254 	counter_u64_add(nummisszap, 1);
1255 	return (0);
1256 }
1257 
1258 /**
1259  * Lookup a name in the name cache
1260  *
1261  * # Arguments
1262  *
1263  * - dvp:	Parent directory in which to search.
1264  * - vpp:	Return argument.  Will contain desired vnode on cache hit.
1265  * - cnp:	Parameters of the name search.  The most interesting bits of
1266  *   		the cn_flags field have the following meanings:
1267  *   	- MAKEENTRY:	If clear, free an entry from the cache rather than look
1268  *   			it up.
1269  *   	- ISDOTDOT:	Must be set if and only if cn_nameptr == ".."
1270  * - tsp:	Return storage for cache timestamp.  On a successful (positive
1271  *   		or negative) lookup, tsp will be filled with any timespec that
1272  *   		was stored when this cache entry was created.  However, it will
1273  *   		be clear for "." entries.
1274  * - ticks:	Return storage for alternate cache timestamp.  On a successful
1275  *   		(positive or negative) lookup, it will contain the ticks value
1276  *   		that was current when the cache entry was created, unless cnp
1277  *   		was ".".
1278  *
1279  * # Returns
1280  *
1281  * - -1:	A positive cache hit.  vpp will contain the desired vnode.
1282  * - ENOENT:	A negative cache hit, or dvp was recycled out from under us due
1283  *		to a forced unmount.  vpp will not be modified.  If the entry
1284  *		is a whiteout, then the ISWHITEOUT flag will be set in
1285  *		cnp->cn_flags.
1286  * - 0:		A cache miss.  vpp will not be modified.
1287  *
1288  * # Locking
1289  *
1290  * On a cache hit, vpp will be returned locked and ref'd.  If we're looking up
1291  * .., dvp is unlocked.  If we're looking up . an extra ref is taken, but the
1292  * lock is not recursively acquired.
1293  */
1294 int
1295 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1296     struct timespec *tsp, int *ticksp)
1297 {
1298 	struct namecache_ts *ncp_ts;
1299 	struct namecache *ncp;
1300 	struct rwlock *blp;
1301 	struct mtx *dvlp;
1302 	uint32_t hash;
1303 	enum vgetstate vs;
1304 	int error, ltype;
1305 
1306 	if (__predict_false(!doingcache)) {
1307 		cnp->cn_flags &= ~MAKEENTRY;
1308 		return (0);
1309 	}
1310 
1311 	counter_u64_add(numcalls, 1);
1312 
1313 	if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.'))
1314 		return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1315 
1316 	if ((cnp->cn_flags & MAKEENTRY) == 0)
1317 		return (cache_lookup_nomakeentry(dvp, vpp, cnp, tsp, ticksp));
1318 
1319 retry:
1320 	blp = NULL;
1321 	dvlp = NULL;
1322 	error = 0;
1323 	if (cnp->cn_namelen == 2 &&
1324 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1325 		counter_u64_add(dotdothits, 1);
1326 		dvlp = VP2VNODELOCK(dvp);
1327 		mtx_lock(dvlp);
1328 		ncp = dvp->v_cache_dd;
1329 		if (ncp == NULL) {
1330 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1331 			    "..", NULL);
1332 			mtx_unlock(dvlp);
1333 			return (0);
1334 		}
1335 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1336 			if (ncp->nc_flag & NCF_NEGATIVE)
1337 				*vpp = NULL;
1338 			else
1339 				*vpp = ncp->nc_vp;
1340 		} else
1341 			*vpp = ncp->nc_dvp;
1342 		/* Return failure if negative entry was found. */
1343 		if (*vpp == NULL)
1344 			goto negative_success;
1345 		CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
1346 		    dvp, cnp->cn_nameptr, *vpp);
1347 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..",
1348 		    *vpp);
1349 		cache_out_ts(ncp, tsp, ticksp);
1350 		if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1351 		    NCF_DTS && tsp != NULL) {
1352 			ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1353 			*tsp = ncp_ts->nc_dotdottime;
1354 		}
1355 		goto success;
1356 	}
1357 
1358 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1359 	blp = HASH2BUCKETLOCK(hash);
1360 	rw_rlock(blp);
1361 
1362 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1363 		counter_u64_add(numchecks, 1);
1364 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1365 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1366 			break;
1367 	}
1368 
1369 	/* We failed to find an entry */
1370 	if (__predict_false(ncp == NULL)) {
1371 		rw_runlock(blp);
1372 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1373 		    NULL);
1374 		counter_u64_add(nummiss, 1);
1375 		return (0);
1376 	}
1377 
1378 	if (ncp->nc_flag & NCF_NEGATIVE)
1379 		goto negative_success;
1380 
1381 	/* We found a "positive" match, return the vnode */
1382 	counter_u64_add(numposhits, 1);
1383 	*vpp = ncp->nc_vp;
1384 	CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p",
1385 	    dvp, cnp->cn_nameptr, *vpp, ncp);
1386 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name,
1387 	    *vpp);
1388 	cache_out_ts(ncp, tsp, ticksp);
1389 success:
1390 	/*
1391 	 * On success we return a locked and ref'd vnode as per the lookup
1392 	 * protocol.
1393 	 */
1394 	MPASS(dvp != *vpp);
1395 	ltype = 0;	/* silence gcc warning */
1396 	if (cnp->cn_flags & ISDOTDOT) {
1397 		ltype = VOP_ISLOCKED(dvp);
1398 		VOP_UNLOCK(dvp, 0);
1399 	}
1400 	vs = vget_prep(*vpp);
1401 	cache_lookup_unlock(blp, dvlp);
1402 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1403 	if (cnp->cn_flags & ISDOTDOT) {
1404 		vn_lock(dvp, ltype | LK_RETRY);
1405 		if (dvp->v_iflag & VI_DOOMED) {
1406 			if (error == 0)
1407 				vput(*vpp);
1408 			*vpp = NULL;
1409 			return (ENOENT);
1410 		}
1411 	}
1412 	if (error) {
1413 		*vpp = NULL;
1414 		goto retry;
1415 	}
1416 	if ((cnp->cn_flags & ISLASTCN) &&
1417 	    (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) {
1418 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
1419 	}
1420 	return (-1);
1421 
1422 negative_success:
1423 	/* We found a negative match, and want to create it, so purge */
1424 	if (cnp->cn_nameiop == CREATE) {
1425 		counter_u64_add(numnegzaps, 1);
1426 		goto zap_and_exit;
1427 	}
1428 
1429 	counter_u64_add(numneghits, 1);
1430 	cache_negative_hit(ncp);
1431 	if (ncp->nc_flag & NCF_WHITE)
1432 		cnp->cn_flags |= ISWHITEOUT;
1433 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp,
1434 	    ncp->nc_name);
1435 	cache_out_ts(ncp, tsp, ticksp);
1436 	cache_lookup_unlock(blp, dvlp);
1437 	return (ENOENT);
1438 
1439 zap_and_exit:
1440 	if (blp != NULL)
1441 		error = cache_zap_rlocked_bucket(ncp, cnp, hash, blp);
1442 	else
1443 		error = cache_zap_locked_vnode(ncp, dvp);
1444 	if (__predict_false(error != 0)) {
1445 		zap_and_exit_bucket_fail2++;
1446 		cache_maybe_yield();
1447 		goto retry;
1448 	}
1449 	cache_free(ncp);
1450 	return (0);
1451 }
1452 
1453 struct celockstate {
1454 	struct mtx *vlp[3];
1455 	struct rwlock *blp[2];
1456 };
1457 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1458 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1459 
1460 static inline void
1461 cache_celockstate_init(struct celockstate *cel)
1462 {
1463 
1464 	bzero(cel, sizeof(*cel));
1465 }
1466 
1467 static void
1468 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1469     struct vnode *dvp)
1470 {
1471 	struct mtx *vlp1, *vlp2;
1472 
1473 	MPASS(cel->vlp[0] == NULL);
1474 	MPASS(cel->vlp[1] == NULL);
1475 	MPASS(cel->vlp[2] == NULL);
1476 
1477 	MPASS(vp != NULL || dvp != NULL);
1478 
1479 	vlp1 = VP2VNODELOCK(vp);
1480 	vlp2 = VP2VNODELOCK(dvp);
1481 	cache_sort_vnodes(&vlp1, &vlp2);
1482 
1483 	if (vlp1 != NULL) {
1484 		mtx_lock(vlp1);
1485 		cel->vlp[0] = vlp1;
1486 	}
1487 	mtx_lock(vlp2);
1488 	cel->vlp[1] = vlp2;
1489 }
1490 
1491 static void
1492 cache_unlock_vnodes_cel(struct celockstate *cel)
1493 {
1494 
1495 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
1496 
1497 	if (cel->vlp[0] != NULL)
1498 		mtx_unlock(cel->vlp[0]);
1499 	if (cel->vlp[1] != NULL)
1500 		mtx_unlock(cel->vlp[1]);
1501 	if (cel->vlp[2] != NULL)
1502 		mtx_unlock(cel->vlp[2]);
1503 }
1504 
1505 static bool
1506 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
1507 {
1508 	struct mtx *vlp;
1509 	bool ret;
1510 
1511 	cache_assert_vlp_locked(cel->vlp[0]);
1512 	cache_assert_vlp_locked(cel->vlp[1]);
1513 	MPASS(cel->vlp[2] == NULL);
1514 
1515 	MPASS(vp != NULL);
1516 	vlp = VP2VNODELOCK(vp);
1517 
1518 	ret = true;
1519 	if (vlp >= cel->vlp[1]) {
1520 		mtx_lock(vlp);
1521 	} else {
1522 		if (mtx_trylock(vlp))
1523 			goto out;
1524 		cache_lock_vnodes_cel_3_failures++;
1525 		cache_unlock_vnodes_cel(cel);
1526 		if (vlp < cel->vlp[0]) {
1527 			mtx_lock(vlp);
1528 			mtx_lock(cel->vlp[0]);
1529 			mtx_lock(cel->vlp[1]);
1530 		} else {
1531 			if (cel->vlp[0] != NULL)
1532 				mtx_lock(cel->vlp[0]);
1533 			mtx_lock(vlp);
1534 			mtx_lock(cel->vlp[1]);
1535 		}
1536 		ret = false;
1537 	}
1538 out:
1539 	cel->vlp[2] = vlp;
1540 	return (ret);
1541 }
1542 
1543 static void
1544 cache_lock_buckets_cel(struct celockstate *cel, struct rwlock *blp1,
1545     struct rwlock *blp2)
1546 {
1547 
1548 	MPASS(cel->blp[0] == NULL);
1549 	MPASS(cel->blp[1] == NULL);
1550 
1551 	cache_sort_vnodes(&blp1, &blp2);
1552 
1553 	if (blp1 != NULL) {
1554 		rw_wlock(blp1);
1555 		cel->blp[0] = blp1;
1556 	}
1557 	rw_wlock(blp2);
1558 	cel->blp[1] = blp2;
1559 }
1560 
1561 static void
1562 cache_unlock_buckets_cel(struct celockstate *cel)
1563 {
1564 
1565 	if (cel->blp[0] != NULL)
1566 		rw_wunlock(cel->blp[0]);
1567 	rw_wunlock(cel->blp[1]);
1568 }
1569 
1570 /*
1571  * Lock part of the cache affected by the insertion.
1572  *
1573  * This means vnodelocks for dvp, vp and the relevant bucketlock.
1574  * However, insertion can result in removal of an old entry. In this
1575  * case we have an additional vnode and bucketlock pair to lock. If the
1576  * entry is negative, ncelock is locked instead of the vnode.
1577  *
1578  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
1579  * preserving the locking order (smaller address first).
1580  */
1581 static void
1582 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1583     uint32_t hash)
1584 {
1585 	struct namecache *ncp;
1586 	struct rwlock *blps[2];
1587 
1588 	blps[0] = HASH2BUCKETLOCK(hash);
1589 	for (;;) {
1590 		blps[1] = NULL;
1591 		cache_lock_vnodes_cel(cel, dvp, vp);
1592 		if (vp == NULL || vp->v_type != VDIR)
1593 			break;
1594 		ncp = vp->v_cache_dd;
1595 		if (ncp == NULL)
1596 			break;
1597 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1598 			break;
1599 		MPASS(ncp->nc_dvp == vp);
1600 		blps[1] = NCP2BUCKETLOCK(ncp);
1601 		if (ncp->nc_flag & NCF_NEGATIVE)
1602 			break;
1603 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1604 			break;
1605 		/*
1606 		 * All vnodes got re-locked. Re-validate the state and if
1607 		 * nothing changed we are done. Otherwise restart.
1608 		 */
1609 		if (ncp == vp->v_cache_dd &&
1610 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1611 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1612 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1613 			break;
1614 		cache_unlock_vnodes_cel(cel);
1615 		cel->vlp[0] = NULL;
1616 		cel->vlp[1] = NULL;
1617 		cel->vlp[2] = NULL;
1618 	}
1619 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1620 }
1621 
1622 static void
1623 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1624     uint32_t hash)
1625 {
1626 	struct namecache *ncp;
1627 	struct rwlock *blps[2];
1628 
1629 	blps[0] = HASH2BUCKETLOCK(hash);
1630 	for (;;) {
1631 		blps[1] = NULL;
1632 		cache_lock_vnodes_cel(cel, dvp, vp);
1633 		ncp = dvp->v_cache_dd;
1634 		if (ncp == NULL)
1635 			break;
1636 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1637 			break;
1638 		MPASS(ncp->nc_dvp == dvp);
1639 		blps[1] = NCP2BUCKETLOCK(ncp);
1640 		if (ncp->nc_flag & NCF_NEGATIVE)
1641 			break;
1642 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1643 			break;
1644 		if (ncp == dvp->v_cache_dd &&
1645 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1646 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1647 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1648 			break;
1649 		cache_unlock_vnodes_cel(cel);
1650 		cel->vlp[0] = NULL;
1651 		cel->vlp[1] = NULL;
1652 		cel->vlp[2] = NULL;
1653 	}
1654 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1655 }
1656 
1657 static void
1658 cache_enter_unlock(struct celockstate *cel)
1659 {
1660 
1661 	cache_unlock_buckets_cel(cel);
1662 	cache_unlock_vnodes_cel(cel);
1663 }
1664 
1665 /*
1666  * Add an entry to the cache.
1667  */
1668 void
1669 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
1670     struct timespec *tsp, struct timespec *dtsp)
1671 {
1672 	struct celockstate cel;
1673 	struct namecache *ncp, *n2, *ndd;
1674 	struct namecache_ts *ncp_ts, *n2_ts;
1675 	struct nchashhead *ncpp;
1676 	struct neglist *neglist;
1677 	uint32_t hash;
1678 	int flag;
1679 	int len;
1680 	bool neg_locked, held_dvp;
1681 	u_long lnumcache;
1682 
1683 	CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr);
1684 	VNASSERT(vp == NULL || (vp->v_iflag & VI_DOOMED) == 0, vp,
1685 	    ("cache_enter: Adding a doomed vnode"));
1686 	VNASSERT(dvp == NULL || (dvp->v_iflag & VI_DOOMED) == 0, dvp,
1687 	    ("cache_enter: Doomed vnode used as src"));
1688 
1689 	if (__predict_false(!doingcache))
1690 		return;
1691 
1692 	/*
1693 	 * Avoid blowout in namecache entries.
1694 	 */
1695 	lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
1696 	if (__predict_false(lnumcache >= ncsize)) {
1697 		atomic_add_long(&numcache, -1);
1698 		return;
1699 	}
1700 
1701 	cache_celockstate_init(&cel);
1702 	ndd = NULL;
1703 	ncp_ts = NULL;
1704 	flag = 0;
1705 	if (cnp->cn_nameptr[0] == '.') {
1706 		if (cnp->cn_namelen == 1)
1707 			return;
1708 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1709 			len = cnp->cn_namelen;
1710 			hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1711 			cache_enter_lock_dd(&cel, dvp, vp, hash);
1712 			/*
1713 			 * If dotdot entry already exists, just retarget it
1714 			 * to new parent vnode, otherwise continue with new
1715 			 * namecache entry allocation.
1716 			 */
1717 			if ((ncp = dvp->v_cache_dd) != NULL &&
1718 			    ncp->nc_flag & NCF_ISDOTDOT) {
1719 				KASSERT(ncp->nc_dvp == dvp,
1720 				    ("wrong isdotdot parent"));
1721 				neg_locked = false;
1722 				if (ncp->nc_flag & NCF_NEGATIVE || vp == NULL) {
1723 					neglist = NCP2NEGLIST(ncp);
1724 					mtx_lock(&ncneg_hot.nl_lock);
1725 					mtx_lock(&neglist->nl_lock);
1726 					neg_locked = true;
1727 				}
1728 				if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1729 					TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst,
1730 					    ncp, nc_dst);
1731 				} else {
1732 					cache_negative_remove(ncp, true);
1733 				}
1734 				if (vp != NULL) {
1735 					TAILQ_INSERT_HEAD(&vp->v_cache_dst,
1736 					    ncp, nc_dst);
1737 					if (ncp->nc_flag & NCF_HOTNEGATIVE)
1738 						numhotneg--;
1739 					ncp->nc_flag &= ~(NCF_NEGATIVE|NCF_HOTNEGATIVE);
1740 				} else {
1741 					if (ncp->nc_flag & NCF_HOTNEGATIVE) {
1742 						numhotneg--;
1743 						ncp->nc_flag &= ~(NCF_HOTNEGATIVE);
1744 					}
1745 					ncp->nc_flag |= NCF_NEGATIVE;
1746 					cache_negative_insert(ncp, true);
1747 				}
1748 				if (neg_locked) {
1749 					mtx_unlock(&neglist->nl_lock);
1750 					mtx_unlock(&ncneg_hot.nl_lock);
1751 				}
1752 				ncp->nc_vp = vp;
1753 				cache_enter_unlock(&cel);
1754 				return;
1755 			}
1756 			dvp->v_cache_dd = NULL;
1757 			cache_enter_unlock(&cel);
1758 			cache_celockstate_init(&cel);
1759 			SDT_PROBE3(vfs, namecache, enter, done, dvp, "..", vp);
1760 			flag = NCF_ISDOTDOT;
1761 		}
1762 	}
1763 
1764 	held_dvp = false;
1765 	if (LIST_EMPTY(&dvp->v_cache_src) && flag != NCF_ISDOTDOT) {
1766 		vhold(dvp);
1767 		atomic_add_long(&numcachehv, 1);
1768 		held_dvp = true;
1769 	}
1770 
1771 	/*
1772 	 * Calculate the hash key and setup as much of the new
1773 	 * namecache entry as possible before acquiring the lock.
1774 	 */
1775 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
1776 	ncp->nc_flag = flag;
1777 	ncp->nc_vp = vp;
1778 	if (vp == NULL)
1779 		ncp->nc_flag |= NCF_NEGATIVE;
1780 	ncp->nc_dvp = dvp;
1781 	if (tsp != NULL) {
1782 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1783 		ncp_ts->nc_time = *tsp;
1784 		ncp_ts->nc_ticks = ticks;
1785 		ncp_ts->nc_nc.nc_flag |= NCF_TS;
1786 		if (dtsp != NULL) {
1787 			ncp_ts->nc_dotdottime = *dtsp;
1788 			ncp_ts->nc_nc.nc_flag |= NCF_DTS;
1789 		}
1790 	}
1791 	len = ncp->nc_nlen = cnp->cn_namelen;
1792 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1793 	strlcpy(ncp->nc_name, cnp->cn_nameptr, len + 1);
1794 	cache_enter_lock(&cel, dvp, vp, hash);
1795 
1796 	/*
1797 	 * See if this vnode or negative entry is already in the cache
1798 	 * with this name.  This can happen with concurrent lookups of
1799 	 * the same path name.
1800 	 */
1801 	ncpp = NCHHASH(hash);
1802 	LIST_FOREACH(n2, ncpp, nc_hash) {
1803 		if (n2->nc_dvp == dvp &&
1804 		    n2->nc_nlen == cnp->cn_namelen &&
1805 		    !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
1806 			if (tsp != NULL) {
1807 				KASSERT((n2->nc_flag & NCF_TS) != 0,
1808 				    ("no NCF_TS"));
1809 				n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
1810 				n2_ts->nc_time = ncp_ts->nc_time;
1811 				n2_ts->nc_ticks = ncp_ts->nc_ticks;
1812 				if (dtsp != NULL) {
1813 					n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
1814 					if (ncp->nc_flag & NCF_NEGATIVE)
1815 						mtx_lock(&ncneg_hot.nl_lock);
1816 					n2_ts->nc_nc.nc_flag |= NCF_DTS;
1817 					if (ncp->nc_flag & NCF_NEGATIVE)
1818 						mtx_unlock(&ncneg_hot.nl_lock);
1819 				}
1820 			}
1821 			goto out_unlock_free;
1822 		}
1823 	}
1824 
1825 	if (flag == NCF_ISDOTDOT) {
1826 		/*
1827 		 * See if we are trying to add .. entry, but some other lookup
1828 		 * has populated v_cache_dd pointer already.
1829 		 */
1830 		if (dvp->v_cache_dd != NULL)
1831 			goto out_unlock_free;
1832 		KASSERT(vp == NULL || vp->v_type == VDIR,
1833 		    ("wrong vnode type %p", vp));
1834 		dvp->v_cache_dd = ncp;
1835 	}
1836 
1837 	if (vp != NULL) {
1838 		if (vp->v_type == VDIR) {
1839 			if (flag != NCF_ISDOTDOT) {
1840 				/*
1841 				 * For this case, the cache entry maps both the
1842 				 * directory name in it and the name ".." for the
1843 				 * directory's parent.
1844 				 */
1845 				if ((ndd = vp->v_cache_dd) != NULL) {
1846 					if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
1847 						cache_zap_locked(ndd, false);
1848 					else
1849 						ndd = NULL;
1850 				}
1851 				vp->v_cache_dd = ncp;
1852 			}
1853 		} else {
1854 			vp->v_cache_dd = NULL;
1855 		}
1856 	}
1857 
1858 	if (flag != NCF_ISDOTDOT) {
1859 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1860 			if (!held_dvp) {
1861 				vhold(dvp);
1862 				atomic_add_long(&numcachehv, 1);
1863 			}
1864 		} else {
1865 			if (held_dvp) {
1866 				/*
1867 				 * This will not take the interlock as someone
1868 				 * else already holds the vnode on account of
1869 				 * the namecache and we hold locks preventing
1870 				 * this from changing.
1871 				 */
1872 				vdrop(dvp);
1873 				atomic_subtract_long(&numcachehv, 1);
1874 			}
1875 		}
1876 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
1877 	}
1878 
1879 	/*
1880 	 * Insert the new namecache entry into the appropriate chain
1881 	 * within the cache entries table.
1882 	 */
1883 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
1884 
1885 	/*
1886 	 * If the entry is "negative", we place it into the
1887 	 * "negative" cache queue, otherwise, we place it into the
1888 	 * destination vnode's cache entries queue.
1889 	 */
1890 	if (vp != NULL) {
1891 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
1892 		SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
1893 		    vp);
1894 	} else {
1895 		if (cnp->cn_flags & ISWHITEOUT)
1896 			ncp->nc_flag |= NCF_WHITE;
1897 		cache_negative_insert(ncp, false);
1898 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
1899 		    ncp->nc_name);
1900 	}
1901 	cache_enter_unlock(&cel);
1902 	if (numneg * ncnegfactor > lnumcache)
1903 		cache_negative_zap_one();
1904 	cache_free(ndd);
1905 	return;
1906 out_unlock_free:
1907 	cache_enter_unlock(&cel);
1908 	cache_free(ncp);
1909 	if (held_dvp) {
1910 		vdrop(dvp);
1911 		atomic_subtract_long(&numcachehv, 1);
1912 	}
1913 	return;
1914 }
1915 
1916 static u_int
1917 cache_roundup_2(u_int val)
1918 {
1919 	u_int res;
1920 
1921 	for (res = 1; res <= val; res <<= 1)
1922 		continue;
1923 
1924 	return (res);
1925 }
1926 
1927 /*
1928  * Name cache initialization, from vfs_init() when we are booting
1929  */
1930 static void
1931 nchinit(void *dummy __unused)
1932 {
1933 	u_int i;
1934 
1935 	cache_zone_small = uma_zcreate("S VFS Cache",
1936 	    sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1,
1937 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1938 	    UMA_ZONE_ZINIT);
1939 	cache_zone_small_ts = uma_zcreate("STS VFS Cache",
1940 	    sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1,
1941 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1942 	    UMA_ZONE_ZINIT);
1943 	cache_zone_large = uma_zcreate("L VFS Cache",
1944 	    sizeof(struct namecache) + NAME_MAX + 1,
1945 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1946 	    UMA_ZONE_ZINIT);
1947 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache",
1948 	    sizeof(struct namecache_ts) + NAME_MAX + 1,
1949 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1950 	    UMA_ZONE_ZINIT);
1951 
1952 	ncsize = desiredvnodes * ncsizefactor;
1953 	nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash);
1954 	ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
1955 	if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
1956 		ncbuckethash = 7;
1957 	if (ncbuckethash > nchash)
1958 		ncbuckethash = nchash;
1959 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
1960 	    M_WAITOK | M_ZERO);
1961 	for (i = 0; i < numbucketlocks; i++)
1962 		rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK | RW_RECURSE);
1963 	ncvnodehash = ncbuckethash;
1964 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
1965 	    M_WAITOK | M_ZERO);
1966 	for (i = 0; i < numvnodelocks; i++)
1967 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
1968 	ncpurgeminvnodes = numbucketlocks * 2;
1969 
1970 	ncneghash = 3;
1971 	neglists = malloc(sizeof(*neglists) * numneglists, M_VFSCACHE,
1972 	    M_WAITOK | M_ZERO);
1973 	for (i = 0; i < numneglists; i++) {
1974 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
1975 		TAILQ_INIT(&neglists[i].nl_list);
1976 	}
1977 	mtx_init(&ncneg_hot.nl_lock, "ncneglh", NULL, MTX_DEF);
1978 	TAILQ_INIT(&ncneg_hot.nl_list);
1979 
1980 	mtx_init(&ncneg_shrink_lock, "ncnegs", NULL, MTX_DEF);
1981 
1982 	numcalls = counter_u64_alloc(M_WAITOK);
1983 	dothits = counter_u64_alloc(M_WAITOK);
1984 	dotdothits = counter_u64_alloc(M_WAITOK);
1985 	numchecks = counter_u64_alloc(M_WAITOK);
1986 	nummiss = counter_u64_alloc(M_WAITOK);
1987 	nummisszap = counter_u64_alloc(M_WAITOK);
1988 	numposzaps = counter_u64_alloc(M_WAITOK);
1989 	numposhits = counter_u64_alloc(M_WAITOK);
1990 	numnegzaps = counter_u64_alloc(M_WAITOK);
1991 	numneghits = counter_u64_alloc(M_WAITOK);
1992 	numfullpathcalls = counter_u64_alloc(M_WAITOK);
1993 	numfullpathfail1 = counter_u64_alloc(M_WAITOK);
1994 	numfullpathfail2 = counter_u64_alloc(M_WAITOK);
1995 	numfullpathfail4 = counter_u64_alloc(M_WAITOK);
1996 	numfullpathfound = counter_u64_alloc(M_WAITOK);
1997 	zap_and_exit_bucket_relock_success = counter_u64_alloc(M_WAITOK);
1998 	numneg_evicted = counter_u64_alloc(M_WAITOK);
1999 	shrinking_skipped = counter_u64_alloc(M_WAITOK);
2000 }
2001 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
2002 
2003 void
2004 cache_changesize(int newmaxvnodes)
2005 {
2006 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
2007 	u_long new_nchash, old_nchash;
2008 	struct namecache *ncp;
2009 	uint32_t hash;
2010 	int newncsize;
2011 	int i;
2012 
2013 	newncsize = newmaxvnodes * ncsizefactor;
2014 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
2015 	if (newmaxvnodes < numbucketlocks)
2016 		newmaxvnodes = numbucketlocks;
2017 
2018 	new_nchashtbl = hashinit(newmaxvnodes, M_VFSCACHE, &new_nchash);
2019 	/* If same hash table size, nothing to do */
2020 	if (nchash == new_nchash) {
2021 		free(new_nchashtbl, M_VFSCACHE);
2022 		return;
2023 	}
2024 	/*
2025 	 * Move everything from the old hash table to the new table.
2026 	 * None of the namecache entries in the table can be removed
2027 	 * because to do so, they have to be removed from the hash table.
2028 	 */
2029 	cache_lock_all_vnodes();
2030 	cache_lock_all_buckets();
2031 	old_nchashtbl = nchashtbl;
2032 	old_nchash = nchash;
2033 	nchashtbl = new_nchashtbl;
2034 	nchash = new_nchash;
2035 	for (i = 0; i <= old_nchash; i++) {
2036 		while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) {
2037 			hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
2038 			    ncp->nc_dvp);
2039 			LIST_REMOVE(ncp, nc_hash);
2040 			LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
2041 		}
2042 	}
2043 	ncsize = newncsize;
2044 	cache_unlock_all_buckets();
2045 	cache_unlock_all_vnodes();
2046 	free(old_nchashtbl, M_VFSCACHE);
2047 }
2048 
2049 /*
2050  * Invalidate all entries from and to a particular vnode.
2051  */
2052 void
2053 cache_purge(struct vnode *vp)
2054 {
2055 	TAILQ_HEAD(, namecache) ncps;
2056 	struct namecache *ncp, *nnp;
2057 	struct mtx *vlp, *vlp2;
2058 
2059 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
2060 	SDT_PROBE1(vfs, namecache, purge, done, vp);
2061 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2062 	    vp->v_cache_dd == NULL)
2063 		return;
2064 	TAILQ_INIT(&ncps);
2065 	vlp = VP2VNODELOCK(vp);
2066 	vlp2 = NULL;
2067 	mtx_lock(vlp);
2068 retry:
2069 	while (!LIST_EMPTY(&vp->v_cache_src)) {
2070 		ncp = LIST_FIRST(&vp->v_cache_src);
2071 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2072 			goto retry;
2073 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2074 	}
2075 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2076 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
2077 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2078 			goto retry;
2079 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2080 	}
2081 	ncp = vp->v_cache_dd;
2082 	if (ncp != NULL) {
2083 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2084 		   ("lost dotdot link"));
2085 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2086 			goto retry;
2087 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2088 	}
2089 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2090 	mtx_unlock(vlp);
2091 	if (vlp2 != NULL)
2092 		mtx_unlock(vlp2);
2093 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2094 		cache_free(ncp);
2095 	}
2096 }
2097 
2098 /*
2099  * Invalidate all negative entries for a particular directory vnode.
2100  */
2101 void
2102 cache_purge_negative(struct vnode *vp)
2103 {
2104 	TAILQ_HEAD(, namecache) ncps;
2105 	struct namecache *ncp, *nnp;
2106 	struct mtx *vlp;
2107 
2108 	CTR1(KTR_VFS, "cache_purge_negative(%p)", vp);
2109 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2110 	if (LIST_EMPTY(&vp->v_cache_src))
2111 		return;
2112 	TAILQ_INIT(&ncps);
2113 	vlp = VP2VNODELOCK(vp);
2114 	mtx_lock(vlp);
2115 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2116 		if (!(ncp->nc_flag & NCF_NEGATIVE))
2117 			continue;
2118 		cache_zap_negative_locked_vnode_kl(ncp, vp);
2119 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2120 	}
2121 	mtx_unlock(vlp);
2122 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2123 		cache_free(ncp);
2124 	}
2125 }
2126 
2127 /*
2128  * Flush all entries referencing a particular filesystem.
2129  */
2130 void
2131 cache_purgevfs(struct mount *mp, bool force)
2132 {
2133 	TAILQ_HEAD(, namecache) ncps;
2134 	struct mtx *vlp1, *vlp2;
2135 	struct rwlock *blp;
2136 	struct nchashhead *bucket;
2137 	struct namecache *ncp, *nnp;
2138 	u_long i, j, n_nchash;
2139 	int error;
2140 
2141 	/* Scan hash tables for applicable entries */
2142 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2143 	if (!force && mp->mnt_nvnodelistsize <= ncpurgeminvnodes)
2144 		return;
2145 	TAILQ_INIT(&ncps);
2146 	n_nchash = nchash + 1;
2147 	vlp1 = vlp2 = NULL;
2148 	for (i = 0; i < numbucketlocks; i++) {
2149 		blp = (struct rwlock *)&bucketlocks[i];
2150 		rw_wlock(blp);
2151 		for (j = i; j < n_nchash; j += numbucketlocks) {
2152 retry:
2153 			bucket = &nchashtbl[j];
2154 			LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) {
2155 				cache_assert_bucket_locked(ncp, RA_WLOCKED);
2156 				if (ncp->nc_dvp->v_mount != mp)
2157 					continue;
2158 				error = cache_zap_wlocked_bucket_kl(ncp, blp,
2159 				    &vlp1, &vlp2);
2160 				if (error != 0)
2161 					goto retry;
2162 				TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst);
2163 			}
2164 		}
2165 		rw_wunlock(blp);
2166 		if (vlp1 == NULL && vlp2 == NULL)
2167 			cache_maybe_yield();
2168 	}
2169 	if (vlp1 != NULL)
2170 		mtx_unlock(vlp1);
2171 	if (vlp2 != NULL)
2172 		mtx_unlock(vlp2);
2173 
2174 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2175 		cache_free(ncp);
2176 	}
2177 }
2178 
2179 /*
2180  * Perform canonical checks and cache lookup and pass on to filesystem
2181  * through the vop_cachedlookup only if needed.
2182  */
2183 
2184 int
2185 vfs_cache_lookup(struct vop_lookup_args *ap)
2186 {
2187 	struct vnode *dvp;
2188 	int error;
2189 	struct vnode **vpp = ap->a_vpp;
2190 	struct componentname *cnp = ap->a_cnp;
2191 	struct ucred *cred = cnp->cn_cred;
2192 	int flags = cnp->cn_flags;
2193 	struct thread *td = cnp->cn_thread;
2194 
2195 	*vpp = NULL;
2196 	dvp = ap->a_dvp;
2197 
2198 	if (dvp->v_type != VDIR)
2199 		return (ENOTDIR);
2200 
2201 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2202 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2203 		return (EROFS);
2204 
2205 	error = VOP_ACCESS(dvp, VEXEC, cred, td);
2206 	if (error)
2207 		return (error);
2208 
2209 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2210 	if (error == 0)
2211 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2212 	if (error == -1)
2213 		return (0);
2214 	return (error);
2215 }
2216 
2217 /*
2218  * XXX All of these sysctls would probably be more productive dead.
2219  */
2220 static int __read_mostly disablecwd;
2221 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
2222    "Disable the getcwd syscall");
2223 
2224 /* Implementation of the getcwd syscall. */
2225 int
2226 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2227 {
2228 
2229 	return (kern___getcwd(td, uap->buf, UIO_USERSPACE, uap->buflen,
2230 	    MAXPATHLEN));
2231 }
2232 
2233 int
2234 kern___getcwd(struct thread *td, char *buf, enum uio_seg bufseg, size_t buflen,
2235     size_t path_max)
2236 {
2237 	char *bp, *tmpbuf;
2238 	struct filedesc *fdp;
2239 	struct vnode *cdir, *rdir;
2240 	int error;
2241 
2242 	if (__predict_false(disablecwd))
2243 		return (ENODEV);
2244 	if (__predict_false(buflen < 2))
2245 		return (EINVAL);
2246 	if (buflen > path_max)
2247 		buflen = path_max;
2248 
2249 	tmpbuf = malloc(buflen, M_TEMP, M_WAITOK);
2250 	fdp = td->td_proc->p_fd;
2251 	FILEDESC_SLOCK(fdp);
2252 	cdir = fdp->fd_cdir;
2253 	vrefact(cdir);
2254 	rdir = fdp->fd_rdir;
2255 	vrefact(rdir);
2256 	FILEDESC_SUNLOCK(fdp);
2257 	error = vn_fullpath1(td, cdir, rdir, tmpbuf, &bp, buflen);
2258 	vrele(rdir);
2259 	vrele(cdir);
2260 
2261 	if (!error) {
2262 		if (bufseg == UIO_SYSSPACE)
2263 			bcopy(bp, buf, strlen(bp) + 1);
2264 		else
2265 			error = copyout(bp, buf, strlen(bp) + 1);
2266 #ifdef KTRACE
2267 	if (KTRPOINT(curthread, KTR_NAMEI))
2268 		ktrnamei(bp);
2269 #endif
2270 	}
2271 	free(tmpbuf, M_TEMP);
2272 	return (error);
2273 }
2274 
2275 /*
2276  * Thus begins the fullpath magic.
2277  */
2278 
2279 static int __read_mostly disablefullpath;
2280 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, &disablefullpath, 0,
2281     "Disable the vn_fullpath function");
2282 
2283 /*
2284  * Retrieve the full filesystem path that correspond to a vnode from the name
2285  * cache (if available)
2286  */
2287 int
2288 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf)
2289 {
2290 	char *buf;
2291 	struct filedesc *fdp;
2292 	struct vnode *rdir;
2293 	int error;
2294 
2295 	if (__predict_false(disablefullpath))
2296 		return (ENODEV);
2297 	if (__predict_false(vn == NULL))
2298 		return (EINVAL);
2299 
2300 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2301 	fdp = td->td_proc->p_fd;
2302 	FILEDESC_SLOCK(fdp);
2303 	rdir = fdp->fd_rdir;
2304 	vrefact(rdir);
2305 	FILEDESC_SUNLOCK(fdp);
2306 	error = vn_fullpath1(td, vn, rdir, buf, retbuf, MAXPATHLEN);
2307 	vrele(rdir);
2308 
2309 	if (!error)
2310 		*freebuf = buf;
2311 	else
2312 		free(buf, M_TEMP);
2313 	return (error);
2314 }
2315 
2316 /*
2317  * This function is similar to vn_fullpath, but it attempts to lookup the
2318  * pathname relative to the global root mount point.  This is required for the
2319  * auditing sub-system, as audited pathnames must be absolute, relative to the
2320  * global root mount point.
2321  */
2322 int
2323 vn_fullpath_global(struct thread *td, struct vnode *vn,
2324     char **retbuf, char **freebuf)
2325 {
2326 	char *buf;
2327 	int error;
2328 
2329 	if (__predict_false(disablefullpath))
2330 		return (ENODEV);
2331 	if (__predict_false(vn == NULL))
2332 		return (EINVAL);
2333 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2334 	error = vn_fullpath1(td, vn, rootvnode, buf, retbuf, MAXPATHLEN);
2335 	if (!error)
2336 		*freebuf = buf;
2337 	else
2338 		free(buf, M_TEMP);
2339 	return (error);
2340 }
2341 
2342 int
2343 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen)
2344 {
2345 	struct vnode *dvp;
2346 	struct namecache *ncp;
2347 	struct mtx *vlp;
2348 	int error;
2349 
2350 	vlp = VP2VNODELOCK(*vp);
2351 	mtx_lock(vlp);
2352 	TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) {
2353 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2354 			break;
2355 	}
2356 	if (ncp != NULL) {
2357 		if (*buflen < ncp->nc_nlen) {
2358 			mtx_unlock(vlp);
2359 			vrele(*vp);
2360 			counter_u64_add(numfullpathfail4, 1);
2361 			error = ENOMEM;
2362 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2363 			    vp, NULL);
2364 			return (error);
2365 		}
2366 		*buflen -= ncp->nc_nlen;
2367 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2368 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2369 		    ncp->nc_name, vp);
2370 		dvp = *vp;
2371 		*vp = ncp->nc_dvp;
2372 		vref(*vp);
2373 		mtx_unlock(vlp);
2374 		vrele(dvp);
2375 		return (0);
2376 	}
2377 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2378 
2379 	mtx_unlock(vlp);
2380 	vn_lock(*vp, LK_SHARED | LK_RETRY);
2381 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
2382 	vput(*vp);
2383 	if (error) {
2384 		counter_u64_add(numfullpathfail2, 1);
2385 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
2386 		return (error);
2387 	}
2388 
2389 	*vp = dvp;
2390 	if (dvp->v_iflag & VI_DOOMED) {
2391 		/* forced unmount */
2392 		vrele(dvp);
2393 		error = ENOENT;
2394 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2395 		return (error);
2396 	}
2397 	/*
2398 	 * *vp has its use count incremented still.
2399 	 */
2400 
2401 	return (0);
2402 }
2403 
2404 /*
2405  * The magic behind kern___getcwd() and vn_fullpath().
2406  */
2407 static int
2408 vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
2409     char *buf, char **retbuf, u_int buflen)
2410 {
2411 	int error, slash_prefixed;
2412 #ifdef KDTRACE_HOOKS
2413 	struct vnode *startvp = vp;
2414 #endif
2415 	struct vnode *vp1;
2416 
2417 	buflen--;
2418 	buf[buflen] = '\0';
2419 	error = 0;
2420 	slash_prefixed = 0;
2421 
2422 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
2423 	counter_u64_add(numfullpathcalls, 1);
2424 	vref(vp);
2425 	if (vp->v_type != VDIR) {
2426 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2427 		if (error)
2428 			return (error);
2429 		if (buflen == 0) {
2430 			vrele(vp);
2431 			return (ENOMEM);
2432 		}
2433 		buf[--buflen] = '/';
2434 		slash_prefixed = 1;
2435 	}
2436 	while (vp != rdir && vp != rootvnode) {
2437 		/*
2438 		 * The vp vnode must be already fully constructed,
2439 		 * since it is either found in namecache or obtained
2440 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
2441 		 * without obtaining the vnode lock.
2442 		 */
2443 		if ((vp->v_vflag & VV_ROOT) != 0) {
2444 			vn_lock(vp, LK_RETRY | LK_SHARED);
2445 
2446 			/*
2447 			 * With the vnode locked, check for races with
2448 			 * unmount, forced or not.  Note that we
2449 			 * already verified that vp is not equal to
2450 			 * the root vnode, which means that
2451 			 * mnt_vnodecovered can be NULL only for the
2452 			 * case of unmount.
2453 			 */
2454 			if ((vp->v_iflag & VI_DOOMED) != 0 ||
2455 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
2456 			    vp1->v_mountedhere != vp->v_mount) {
2457 				vput(vp);
2458 				error = ENOENT;
2459 				SDT_PROBE3(vfs, namecache, fullpath, return,
2460 				    error, vp, NULL);
2461 				break;
2462 			}
2463 
2464 			vref(vp1);
2465 			vput(vp);
2466 			vp = vp1;
2467 			continue;
2468 		}
2469 		if (vp->v_type != VDIR) {
2470 			vrele(vp);
2471 			counter_u64_add(numfullpathfail1, 1);
2472 			error = ENOTDIR;
2473 			SDT_PROBE3(vfs, namecache, fullpath, return,
2474 			    error, vp, NULL);
2475 			break;
2476 		}
2477 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2478 		if (error)
2479 			break;
2480 		if (buflen == 0) {
2481 			vrele(vp);
2482 			error = ENOMEM;
2483 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2484 			    startvp, NULL);
2485 			break;
2486 		}
2487 		buf[--buflen] = '/';
2488 		slash_prefixed = 1;
2489 	}
2490 	if (error)
2491 		return (error);
2492 	if (!slash_prefixed) {
2493 		if (buflen == 0) {
2494 			vrele(vp);
2495 			counter_u64_add(numfullpathfail4, 1);
2496 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
2497 			    startvp, NULL);
2498 			return (ENOMEM);
2499 		}
2500 		buf[--buflen] = '/';
2501 	}
2502 	counter_u64_add(numfullpathfound, 1);
2503 	vrele(vp);
2504 
2505 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, buf + buflen);
2506 	*retbuf = buf + buflen;
2507 	return (0);
2508 }
2509 
2510 struct vnode *
2511 vn_dir_dd_ino(struct vnode *vp)
2512 {
2513 	struct namecache *ncp;
2514 	struct vnode *ddvp;
2515 	struct mtx *vlp;
2516 	enum vgetstate vs;
2517 
2518 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
2519 	vlp = VP2VNODELOCK(vp);
2520 	mtx_lock(vlp);
2521 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
2522 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
2523 			continue;
2524 		ddvp = ncp->nc_dvp;
2525 		vs = vget_prep(ddvp);
2526 		mtx_unlock(vlp);
2527 		if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
2528 			return (NULL);
2529 		return (ddvp);
2530 	}
2531 	mtx_unlock(vlp);
2532 	return (NULL);
2533 }
2534 
2535 int
2536 vn_commname(struct vnode *vp, char *buf, u_int buflen)
2537 {
2538 	struct namecache *ncp;
2539 	struct mtx *vlp;
2540 	int l;
2541 
2542 	vlp = VP2VNODELOCK(vp);
2543 	mtx_lock(vlp);
2544 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
2545 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2546 			break;
2547 	if (ncp == NULL) {
2548 		mtx_unlock(vlp);
2549 		return (ENOENT);
2550 	}
2551 	l = min(ncp->nc_nlen, buflen - 1);
2552 	memcpy(buf, ncp->nc_name, l);
2553 	mtx_unlock(vlp);
2554 	buf[l] = '\0';
2555 	return (0);
2556 }
2557 
2558 /*
2559  * This function updates path string to vnode's full global path
2560  * and checks the size of the new path string against the pathlen argument.
2561  *
2562  * Requires a locked, referenced vnode.
2563  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
2564  *
2565  * If sysctl debug.disablefullpath is set, ENODEV is returned,
2566  * vnode is left locked and path remain untouched.
2567  *
2568  * If vp is a directory, the call to vn_fullpath_global() always succeeds
2569  * because it falls back to the ".." lookup if the namecache lookup fails.
2570  */
2571 int
2572 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
2573     u_int pathlen)
2574 {
2575 	struct nameidata nd;
2576 	struct vnode *vp1;
2577 	char *rpath, *fbuf;
2578 	int error;
2579 
2580 	ASSERT_VOP_ELOCKED(vp, __func__);
2581 
2582 	/* Return ENODEV if sysctl debug.disablefullpath==1 */
2583 	if (__predict_false(disablefullpath))
2584 		return (ENODEV);
2585 
2586 	/* Construct global filesystem path from vp. */
2587 	VOP_UNLOCK(vp, 0);
2588 	error = vn_fullpath_global(td, vp, &rpath, &fbuf);
2589 
2590 	if (error != 0) {
2591 		vrele(vp);
2592 		return (error);
2593 	}
2594 
2595 	if (strlen(rpath) >= pathlen) {
2596 		vrele(vp);
2597 		error = ENAMETOOLONG;
2598 		goto out;
2599 	}
2600 
2601 	/*
2602 	 * Re-lookup the vnode by path to detect a possible rename.
2603 	 * As a side effect, the vnode is relocked.
2604 	 * If vnode was renamed, return ENOENT.
2605 	 */
2606 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
2607 	    UIO_SYSSPACE, path, td);
2608 	error = namei(&nd);
2609 	if (error != 0) {
2610 		vrele(vp);
2611 		goto out;
2612 	}
2613 	NDFREE(&nd, NDF_ONLY_PNBUF);
2614 	vp1 = nd.ni_vp;
2615 	vrele(vp);
2616 	if (vp1 == vp)
2617 		strcpy(path, rpath);
2618 	else {
2619 		vput(vp1);
2620 		error = ENOENT;
2621 	}
2622 
2623 out:
2624 	free(fbuf, M_TEMP);
2625 	return (error);
2626 }
2627 
2628 #ifdef DDB
2629 static void
2630 db_print_vpath(struct vnode *vp)
2631 {
2632 
2633 	while (vp != NULL) {
2634 		db_printf("%p: ", vp);
2635 		if (vp == rootvnode) {
2636 			db_printf("/");
2637 			vp = NULL;
2638 		} else {
2639 			if (vp->v_vflag & VV_ROOT) {
2640 				db_printf("<mount point>");
2641 				vp = vp->v_mount->mnt_vnodecovered;
2642 			} else {
2643 				struct namecache *ncp;
2644 				char *ncn;
2645 				int i;
2646 
2647 				ncp = TAILQ_FIRST(&vp->v_cache_dst);
2648 				if (ncp != NULL) {
2649 					ncn = ncp->nc_name;
2650 					for (i = 0; i < ncp->nc_nlen; i++)
2651 						db_printf("%c", *ncn++);
2652 					vp = ncp->nc_dvp;
2653 				} else {
2654 					vp = NULL;
2655 				}
2656 			}
2657 		}
2658 		db_printf("\n");
2659 	}
2660 
2661 	return;
2662 }
2663 
2664 DB_SHOW_COMMAND(vpath, db_show_vpath)
2665 {
2666 	struct vnode *vp;
2667 
2668 	if (!have_addr) {
2669 		db_printf("usage: show vpath <struct vnode *>\n");
2670 		return;
2671 	}
2672 
2673 	vp = (struct vnode *)addr;
2674 	db_print_vpath(vp);
2675 }
2676 
2677 #endif
2678