xref: /freebsd/sys/kern/vfs_cache.c (revision 7be8de4271d5cb5d441e2757912c1824f6c3dc3b)
1 /*-
2  * Copyright (c) 1989, 1993, 1995
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * Poul-Henning Kamp of the FreeBSD Project.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 4. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	@(#)vfs_cache.c	8.5 (Berkeley) 3/22/95
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include "opt_ktrace.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/counter.h>
43 #include <sys/filedesc.h>
44 #include <sys/fnv_hash.h>
45 #include <sys/kernel.h>
46 #include <sys/lock.h>
47 #include <sys/malloc.h>
48 #include <sys/fcntl.h>
49 #include <sys/mount.h>
50 #include <sys/namei.h>
51 #include <sys/proc.h>
52 #include <sys/rwlock.h>
53 #include <sys/sdt.h>
54 #include <sys/syscallsubr.h>
55 #include <sys/sysctl.h>
56 #include <sys/sysproto.h>
57 #include <sys/vnode.h>
58 #ifdef KTRACE
59 #include <sys/ktrace.h>
60 #endif
61 
62 #include <vm/uma.h>
63 
64 SDT_PROVIDER_DECLARE(vfs);
65 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
66     "struct vnode *");
67 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
68     "char *");
69 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
70 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
71     "char *", "struct vnode *");
72 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
73 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
74     "struct vnode *", "char *");
75 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
76     "struct vnode *");
77 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
78     "struct vnode *", "char *");
79 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
80     "char *");
81 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *");
82 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
83 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
84 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
85     "struct vnode *");
86 SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *",
87     "char *");
88 
89 /*
90  * This structure describes the elements in the cache of recent
91  * names looked up by namei.
92  */
93 
94 struct	namecache {
95 	LIST_ENTRY(namecache) nc_hash;	/* hash chain */
96 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
97 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
98 	struct	vnode *nc_dvp;		/* vnode of parent of name */
99 	struct	vnode *nc_vp;		/* vnode the name refers to */
100 	u_char	nc_flag;		/* flag bits */
101 	u_char	nc_nlen;		/* length of name */
102 	char	nc_name[0];		/* segment name + nul */
103 };
104 
105 /*
106  * struct namecache_ts repeats struct namecache layout up to the
107  * nc_nlen member.
108  * struct namecache_ts is used in place of struct namecache when time(s) need
109  * to be stored.  The nc_dotdottime field is used when a cache entry is mapping
110  * both a non-dotdot directory name plus dotdot for the directory's
111  * parent.
112  */
113 struct	namecache_ts {
114 	LIST_ENTRY(namecache) nc_hash;	/* hash chain */
115 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
116 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
117 	struct	vnode *nc_dvp;		/* vnode of parent of name */
118 	struct	vnode *nc_vp;		/* vnode the name refers to */
119 	u_char	nc_flag;		/* flag bits */
120 	u_char	nc_nlen;		/* length of name */
121 	struct	timespec nc_time;	/* timespec provided by fs */
122 	struct	timespec nc_dotdottime;	/* dotdot timespec provided by fs */
123 	int	nc_ticks;		/* ticks value when entry was added */
124 	char	nc_name[0];		/* segment name + nul */
125 };
126 
127 /*
128  * Flags in namecache.nc_flag
129  */
130 #define NCF_WHITE	0x01
131 #define NCF_ISDOTDOT	0x02
132 #define	NCF_TS		0x04
133 #define	NCF_DTS		0x08
134 #define	NCF_DVDROP	0x10
135 
136 /*
137  * Name caching works as follows:
138  *
139  * Names found by directory scans are retained in a cache
140  * for future reference.  It is managed LRU, so frequently
141  * used names will hang around.  Cache is indexed by hash value
142  * obtained from (vp, name) where vp refers to the directory
143  * containing name.
144  *
145  * If it is a "negative" entry, (i.e. for a name that is known NOT to
146  * exist) the vnode pointer will be NULL.
147  *
148  * Upon reaching the last segment of a path, if the reference
149  * is for DELETE, or NOCACHE is set (rewrite), and the
150  * name is located in the cache, it will be dropped.
151  */
152 
153 /*
154  * Structures associated with name caching.
155  */
156 #define NCHHASH(hash) \
157 	(&nchashtbl[(hash) & nchash])
158 static LIST_HEAD(nchashhead, namecache) *nchashtbl;	/* Hash Table */
159 static TAILQ_HEAD(, namecache) ncneg;	/* Hash Table */
160 static u_long	nchash;			/* size of hash table */
161 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
162     "Size of namecache hash table");
163 static u_long	ncnegfactor = 16;	/* ratio of negative entries */
164 SYSCTL_ULONG(_vfs, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
165     "Ratio of negative namecache entries");
166 static u_long	numneg;			/* number of negative entries allocated */
167 SYSCTL_ULONG(_debug, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0,
168     "Number of negative entries in namecache");
169 static u_long	numcache;		/* number of cache entries allocated */
170 SYSCTL_ULONG(_debug, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0,
171     "Number of namecache entries");
172 static u_long	numcachehv;		/* number of cache entries with vnodes held */
173 SYSCTL_ULONG(_debug, OID_AUTO, numcachehv, CTLFLAG_RD, &numcachehv, 0,
174     "Number of namecache entries with vnodes held");
175 u_int	ncsizefactor = 2;
176 SYSCTL_UINT(_vfs, OID_AUTO, ncsizefactor, CTLFLAG_RW, &ncsizefactor, 0,
177     "Size factor for namecache");
178 
179 struct nchstats	nchstats;		/* cache effectiveness statistics */
180 
181 static struct rwlock cache_lock;
182 RW_SYSINIT(vfscache, &cache_lock, "Name Cache");
183 
184 #define	CACHE_UPGRADE_LOCK()	rw_try_upgrade(&cache_lock)
185 #define	CACHE_RLOCK()		rw_rlock(&cache_lock)
186 #define	CACHE_RUNLOCK()		rw_runlock(&cache_lock)
187 #define	CACHE_WLOCK()		rw_wlock(&cache_lock)
188 #define	CACHE_WUNLOCK()		rw_wunlock(&cache_lock)
189 
190 static struct mtx_padalign ncneg_mtx;
191 MTX_SYSINIT(vfscache_neg, &ncneg_mtx, "Name Cache neg", MTX_DEF);
192 
193 /*
194  * UMA zones for the VFS cache.
195  *
196  * The small cache is used for entries with short names, which are the
197  * most common.  The large cache is used for entries which are too big to
198  * fit in the small cache.
199  */
200 static uma_zone_t cache_zone_small;
201 static uma_zone_t cache_zone_small_ts;
202 static uma_zone_t cache_zone_large;
203 static uma_zone_t cache_zone_large_ts;
204 
205 #define	CACHE_PATH_CUTOFF	35
206 
207 static struct namecache *
208 cache_alloc(int len, int ts)
209 {
210 
211 	if (len > CACHE_PATH_CUTOFF) {
212 		if (ts)
213 			return (uma_zalloc(cache_zone_large_ts, M_WAITOK));
214 		else
215 			return (uma_zalloc(cache_zone_large, M_WAITOK));
216 	}
217 	if (ts)
218 		return (uma_zalloc(cache_zone_small_ts, M_WAITOK));
219 	else
220 		return (uma_zalloc(cache_zone_small, M_WAITOK));
221 }
222 
223 static void
224 cache_free(struct namecache *ncp)
225 {
226 	int ts;
227 
228 	if (ncp == NULL)
229 		return;
230 	ts = ncp->nc_flag & NCF_TS;
231 	if ((ncp->nc_flag & NCF_DVDROP) != 0)
232 		vdrop(ncp->nc_dvp);
233 	if (ncp->nc_nlen <= CACHE_PATH_CUTOFF) {
234 		if (ts)
235 			uma_zfree(cache_zone_small_ts, ncp);
236 		else
237 			uma_zfree(cache_zone_small, ncp);
238 	} else if (ts)
239 		uma_zfree(cache_zone_large_ts, ncp);
240 	else
241 		uma_zfree(cache_zone_large, ncp);
242 }
243 
244 static char *
245 nc_get_name(struct namecache *ncp)
246 {
247 	struct namecache_ts *ncp_ts;
248 
249 	if ((ncp->nc_flag & NCF_TS) == 0)
250 		return (ncp->nc_name);
251 	ncp_ts = (struct namecache_ts *)ncp;
252 	return (ncp_ts->nc_name);
253 }
254 
255 static void
256 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
257 {
258 
259 	KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
260 	    (tsp == NULL && ticksp == NULL),
261 	    ("No NCF_TS"));
262 
263 	if (tsp != NULL)
264 		*tsp = ((struct namecache_ts *)ncp)->nc_time;
265 	if (ticksp != NULL)
266 		*ticksp = ((struct namecache_ts *)ncp)->nc_ticks;
267 }
268 
269 static int	doingcache = 1;		/* 1 => enable the cache */
270 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
271     "VFS namecache enabled");
272 
273 /* Export size information to userland */
274 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
275     sizeof(struct namecache), "sizeof(struct namecache)");
276 
277 /*
278  * The new name cache statistics
279  */
280 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0,
281     "Name cache statistics");
282 #define STATNODE_ULONG(name, descr)	\
283 	SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, descr);
284 #define STATNODE_COUNTER(name, descr)	\
285 	static counter_u64_t name;	\
286 	SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, descr);
287 STATNODE_ULONG(numneg, "Number of negative cache entries");
288 STATNODE_ULONG(numcache, "Number of cache entries");
289 STATNODE_COUNTER(numcalls, "Number of cache lookups");
290 STATNODE_COUNTER(dothits, "Number of '.' hits");
291 STATNODE_COUNTER(dotdothits, "Number of '..' hits");
292 STATNODE_COUNTER(numchecks, "Number of checks in lookup");
293 STATNODE_COUNTER(nummiss, "Number of cache misses");
294 STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache");
295 STATNODE_COUNTER(numposzaps,
296     "Number of cache hits (positive) we do not want to cache");
297 STATNODE_COUNTER(numposhits, "Number of cache hits (positive)");
298 STATNODE_COUNTER(numnegzaps,
299     "Number of cache hits (negative) we do not want to cache");
300 STATNODE_COUNTER(numneghits, "Number of cache hits (negative)");
301 /* These count for kern___getcwd(), too. */
302 STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls");
303 STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
304 STATNODE_COUNTER(numfullpathfail2,
305     "Number of fullpath search errors (VOP_VPTOCNP failures)");
306 STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
307 STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls");
308 static long numupgrades; STATNODE_ULONG(numupgrades,
309     "Number of updates of the cache after lookup (write lock + retry)");
310 
311 static void cache_zap(struct namecache *ncp);
312 static int vn_vptocnp_locked(struct vnode **vp, struct ucred *cred, char *buf,
313     u_int *buflen);
314 static int vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
315     char *buf, char **retbuf, u_int buflen);
316 
317 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
318 
319 static uint32_t
320 cache_get_hash(char *name, u_char len, struct vnode *dvp)
321 {
322 	uint32_t hash;
323 
324 	hash = fnv_32_buf(name, len, FNV1_32_INIT);
325 	hash = fnv_32_buf(&dvp, sizeof(dvp), hash);
326 	return (hash);
327 }
328 
329 static int
330 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
331 {
332 	struct nchstats snap;
333 
334 	if (req->oldptr == NULL)
335 		return (SYSCTL_OUT(req, 0, sizeof(snap)));
336 
337 	snap = nchstats;
338 	snap.ncs_goodhits = counter_u64_fetch(numposhits);
339 	snap.ncs_neghits = counter_u64_fetch(numneghits);
340 	snap.ncs_badhits = counter_u64_fetch(numposzaps) +
341 	    counter_u64_fetch(numnegzaps);
342 	snap.ncs_miss = counter_u64_fetch(nummisszap) +
343 	    counter_u64_fetch(nummiss);
344 
345 	return (SYSCTL_OUT(req, &snap, sizeof(snap)));
346 }
347 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
348     CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
349     "VFS cache effectiveness statistics");
350 
351 #ifdef DIAGNOSTIC
352 /*
353  * Grab an atomic snapshot of the name cache hash chain lengths
354  */
355 static SYSCTL_NODE(_debug, OID_AUTO, hashstat, CTLFLAG_RW, NULL,
356     "hash table stats");
357 
358 static int
359 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
360 {
361 	struct nchashhead *ncpp;
362 	struct namecache *ncp;
363 	int i, error, n_nchash, *cntbuf;
364 
365 retry:
366 	n_nchash = nchash + 1;	/* nchash is max index, not count */
367 	if (req->oldptr == NULL)
368 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
369 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
370 	CACHE_RLOCK();
371 	if (n_nchash != nchash + 1) {
372 		CACHE_RUNLOCK();
373 		free(cntbuf, M_TEMP);
374 		goto retry;
375 	}
376 	/* Scan hash tables counting entries */
377 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
378 		LIST_FOREACH(ncp, ncpp, nc_hash)
379 			cntbuf[i]++;
380 	CACHE_RUNLOCK();
381 	for (error = 0, i = 0; i < n_nchash; i++)
382 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
383 			break;
384 	free(cntbuf, M_TEMP);
385 	return (error);
386 }
387 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
388     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
389     "nchash chain lengths");
390 
391 static int
392 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
393 {
394 	int error;
395 	struct nchashhead *ncpp;
396 	struct namecache *ncp;
397 	int n_nchash;
398 	int count, maxlength, used, pct;
399 
400 	if (!req->oldptr)
401 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
402 
403 	CACHE_RLOCK();
404 	n_nchash = nchash + 1;	/* nchash is max index, not count */
405 	used = 0;
406 	maxlength = 0;
407 
408 	/* Scan hash tables for applicable entries */
409 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
410 		count = 0;
411 		LIST_FOREACH(ncp, ncpp, nc_hash) {
412 			count++;
413 		}
414 		if (count)
415 			used++;
416 		if (maxlength < count)
417 			maxlength = count;
418 	}
419 	n_nchash = nchash + 1;
420 	CACHE_RUNLOCK();
421 	pct = (used * 100) / (n_nchash / 100);
422 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
423 	if (error)
424 		return (error);
425 	error = SYSCTL_OUT(req, &used, sizeof(used));
426 	if (error)
427 		return (error);
428 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
429 	if (error)
430 		return (error);
431 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
432 	if (error)
433 		return (error);
434 	return (0);
435 }
436 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
437     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
438     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
439 #endif
440 
441 /*
442  * Negative entries management
443  */
444 static void
445 cache_negative_hit(struct namecache *ncp, int wlocked)
446 {
447 
448 	if (!wlocked) {
449 		rw_assert(&cache_lock, RA_RLOCKED);
450 		mtx_lock(&ncneg_mtx);
451 	} else {
452 		rw_assert(&cache_lock, RA_WLOCKED);
453 	}
454 
455 	TAILQ_REMOVE(&ncneg, ncp, nc_dst);
456 	TAILQ_INSERT_TAIL(&ncneg, ncp, nc_dst);
457 
458 	if (!wlocked)
459 		mtx_unlock(&ncneg_mtx);
460 }
461 
462 static void
463 cache_negative_insert(struct namecache *ncp)
464 {
465 
466 	rw_assert(&cache_lock, RA_WLOCKED);
467 	MPASS(ncp->nc_vp == NULL);
468 	TAILQ_INSERT_TAIL(&ncneg, ncp, nc_dst);
469 	numneg++;
470 }
471 
472 static void
473 cache_negative_remove(struct namecache *ncp)
474 {
475 
476 	rw_assert(&cache_lock, RA_WLOCKED);
477 	MPASS(ncp->nc_vp == NULL);
478 	TAILQ_REMOVE(&ncneg, ncp, nc_dst);
479 	numneg--;
480 }
481 
482 static struct namecache *
483 cache_negative_zap_one(void)
484 {
485 	struct namecache *ncp;
486 
487 	rw_assert(&cache_lock, RA_WLOCKED);
488 	ncp = TAILQ_FIRST(&ncneg);
489 	KASSERT(ncp->nc_vp == NULL, ("ncp %p vp %p on ncneg",
490 	    ncp, ncp->nc_vp));
491 	cache_zap(ncp);
492 	return (ncp);
493 }
494 
495 /*
496  * cache_zap():
497  *
498  *   Removes a namecache entry from cache, whether it contains an actual
499  *   pointer to a vnode or if it is just a negative cache entry.
500  */
501 static void
502 cache_zap(struct namecache *ncp)
503 {
504 
505 	rw_assert(&cache_lock, RA_WLOCKED);
506 	CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp, ncp->nc_vp);
507 	if (ncp->nc_vp != NULL) {
508 		SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
509 		    nc_get_name(ncp), ncp->nc_vp);
510 	} else {
511 		SDT_PROBE2(vfs, namecache, zap_negative, done, ncp->nc_dvp,
512 		    nc_get_name(ncp));
513 	}
514 	LIST_REMOVE(ncp, nc_hash);
515 	if (ncp->nc_flag & NCF_ISDOTDOT) {
516 		if (ncp == ncp->nc_dvp->v_cache_dd)
517 			ncp->nc_dvp->v_cache_dd = NULL;
518 	} else {
519 		LIST_REMOVE(ncp, nc_src);
520 		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
521 			ncp->nc_flag |= NCF_DVDROP;
522 			numcachehv--;
523 		}
524 	}
525 	if (ncp->nc_vp) {
526 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
527 		if (ncp == ncp->nc_vp->v_cache_dd)
528 			ncp->nc_vp->v_cache_dd = NULL;
529 	} else {
530 		cache_negative_remove(ncp);
531 	}
532 	numcache--;
533 }
534 
535 /*
536  * Lookup an entry in the cache
537  *
538  * Lookup is called with dvp pointing to the directory to search,
539  * cnp pointing to the name of the entry being sought. If the lookup
540  * succeeds, the vnode is returned in *vpp, and a status of -1 is
541  * returned. If the lookup determines that the name does not exist
542  * (negative caching), a status of ENOENT is returned. If the lookup
543  * fails, a status of zero is returned.  If the directory vnode is
544  * recycled out from under us due to a forced unmount, a status of
545  * ENOENT is returned.
546  *
547  * vpp is locked and ref'd on return.  If we're looking up DOTDOT, dvp is
548  * unlocked.  If we're looking up . an extra ref is taken, but the lock is
549  * not recursively acquired.
550  */
551 
552 int
553 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
554     struct timespec *tsp, int *ticksp)
555 {
556 	struct namecache *ncp;
557 	uint32_t hash;
558 	int error, ltype, wlocked;
559 
560 	if (!doingcache) {
561 		cnp->cn_flags &= ~MAKEENTRY;
562 		return (0);
563 	}
564 retry:
565 	wlocked = 0;
566 	counter_u64_add(numcalls, 1);
567 	error = 0;
568 
569 retry_wlocked:
570 	if (cnp->cn_nameptr[0] == '.') {
571 		if (cnp->cn_namelen == 1) {
572 			*vpp = dvp;
573 			CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .",
574 			    dvp, cnp->cn_nameptr);
575 			counter_u64_add(dothits, 1);
576 			SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
577 			if (tsp != NULL)
578 				timespecclear(tsp);
579 			if (ticksp != NULL)
580 				*ticksp = ticks;
581 			VREF(*vpp);
582 			/*
583 			 * When we lookup "." we still can be asked to lock it
584 			 * differently...
585 			 */
586 			ltype = cnp->cn_lkflags & LK_TYPE_MASK;
587 			if (ltype != VOP_ISLOCKED(*vpp)) {
588 				if (ltype == LK_EXCLUSIVE) {
589 					vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
590 					if ((*vpp)->v_iflag & VI_DOOMED) {
591 						/* forced unmount */
592 						vrele(*vpp);
593 						*vpp = NULL;
594 						return (ENOENT);
595 					}
596 				} else
597 					vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
598 			}
599 			return (-1);
600 		}
601 		if (!wlocked)
602 			CACHE_RLOCK();
603 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
604 			counter_u64_add(dotdothits, 1);
605 			if (dvp->v_cache_dd == NULL) {
606 				SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
607 				    "..", NULL);
608 				goto unlock;
609 			}
610 			if ((cnp->cn_flags & MAKEENTRY) == 0) {
611 				if (!wlocked && !CACHE_UPGRADE_LOCK())
612 					goto wlock;
613 				ncp = NULL;
614 				if (dvp->v_cache_dd->nc_flag & NCF_ISDOTDOT) {
615 					ncp = dvp->v_cache_dd;
616 					cache_zap(ncp);
617 				}
618 				dvp->v_cache_dd = NULL;
619 				CACHE_WUNLOCK();
620 				cache_free(ncp);
621 				return (0);
622 			}
623 			ncp = dvp->v_cache_dd;
624 			if (ncp->nc_flag & NCF_ISDOTDOT)
625 				*vpp = ncp->nc_vp;
626 			else
627 				*vpp = ncp->nc_dvp;
628 			/* Return failure if negative entry was found. */
629 			if (*vpp == NULL)
630 				goto negative_success;
631 			CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
632 			    dvp, cnp->cn_nameptr, *vpp);
633 			SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..",
634 			    *vpp);
635 			cache_out_ts(ncp, tsp, ticksp);
636 			if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
637 			    NCF_DTS && tsp != NULL)
638 				*tsp = ((struct namecache_ts *)ncp)->
639 				    nc_dotdottime;
640 			goto success;
641 		}
642 	} else if (!wlocked)
643 		CACHE_RLOCK();
644 
645 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
646 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
647 		counter_u64_add(numchecks, 1);
648 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
649 		    !bcmp(nc_get_name(ncp), cnp->cn_nameptr, ncp->nc_nlen))
650 			break;
651 	}
652 
653 	/* We failed to find an entry */
654 	if (ncp == NULL) {
655 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
656 		    NULL);
657 		if ((cnp->cn_flags & MAKEENTRY) == 0) {
658 			counter_u64_add(nummisszap, 1);
659 		} else {
660 			counter_u64_add(nummiss, 1);
661 		}
662 		goto unlock;
663 	}
664 
665 	/* We don't want to have an entry, so dump it */
666 	if ((cnp->cn_flags & MAKEENTRY) == 0) {
667 		counter_u64_add(numposzaps, 1);
668 		if (!wlocked && !CACHE_UPGRADE_LOCK())
669 			goto wlock;
670 		cache_zap(ncp);
671 		CACHE_WUNLOCK();
672 		cache_free(ncp);
673 		return (0);
674 	}
675 
676 	/* We found a "positive" match, return the vnode */
677 	if (ncp->nc_vp) {
678 		counter_u64_add(numposhits, 1);
679 		*vpp = ncp->nc_vp;
680 		CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p",
681 		    dvp, cnp->cn_nameptr, *vpp, ncp);
682 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, nc_get_name(ncp),
683 		    *vpp);
684 		cache_out_ts(ncp, tsp, ticksp);
685 		goto success;
686 	}
687 
688 negative_success:
689 	/* We found a negative match, and want to create it, so purge */
690 	if (cnp->cn_nameiop == CREATE) {
691 		counter_u64_add(numnegzaps, 1);
692 		if (!wlocked && !CACHE_UPGRADE_LOCK())
693 			goto wlock;
694 		cache_zap(ncp);
695 		CACHE_WUNLOCK();
696 		cache_free(ncp);
697 		return (0);
698 	}
699 
700 	counter_u64_add(numneghits, 1);
701 	cache_negative_hit(ncp, wlocked);
702 	if (ncp->nc_flag & NCF_WHITE)
703 		cnp->cn_flags |= ISWHITEOUT;
704 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp,
705 	    nc_get_name(ncp));
706 	cache_out_ts(ncp, tsp, ticksp);
707 	if (wlocked)
708 		CACHE_WUNLOCK();
709 	else
710 		CACHE_RUNLOCK();
711 	return (ENOENT);
712 
713 wlock:
714 	/*
715 	 * We need to update the cache after our lookup, so upgrade to
716 	 * a write lock and retry the operation.
717 	 */
718 	CACHE_RUNLOCK();
719 	CACHE_WLOCK();
720 	numupgrades++;
721 	wlocked = 1;
722 	goto retry_wlocked;
723 
724 success:
725 	/*
726 	 * On success we return a locked and ref'd vnode as per the lookup
727 	 * protocol.
728 	 */
729 	MPASS(dvp != *vpp);
730 	ltype = 0;	/* silence gcc warning */
731 	if (cnp->cn_flags & ISDOTDOT) {
732 		ltype = VOP_ISLOCKED(dvp);
733 		VOP_UNLOCK(dvp, 0);
734 	}
735 	vhold(*vpp);
736 	if (wlocked)
737 		CACHE_WUNLOCK();
738 	else
739 		CACHE_RUNLOCK();
740 	error = vget(*vpp, cnp->cn_lkflags | LK_VNHELD, cnp->cn_thread);
741 	if (cnp->cn_flags & ISDOTDOT) {
742 		vn_lock(dvp, ltype | LK_RETRY);
743 		if (dvp->v_iflag & VI_DOOMED) {
744 			if (error == 0)
745 				vput(*vpp);
746 			*vpp = NULL;
747 			return (ENOENT);
748 		}
749 	}
750 	if (error) {
751 		*vpp = NULL;
752 		goto retry;
753 	}
754 	if ((cnp->cn_flags & ISLASTCN) &&
755 	    (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) {
756 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
757 	}
758 	return (-1);
759 
760 unlock:
761 	if (wlocked)
762 		CACHE_WUNLOCK();
763 	else
764 		CACHE_RUNLOCK();
765 	return (0);
766 }
767 
768 /*
769  * Add an entry to the cache.
770  */
771 void
772 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
773     struct timespec *tsp, struct timespec *dtsp)
774 {
775 	struct namecache *ncp, *n2, *ndd, *nneg;
776 	struct namecache_ts *n3;
777 	struct nchashhead *ncpp;
778 	uint32_t hash;
779 	int flag;
780 	int len;
781 
782 	CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr);
783 	VNASSERT(vp == NULL || (vp->v_iflag & VI_DOOMED) == 0, vp,
784 	    ("cache_enter: Adding a doomed vnode"));
785 	VNASSERT(dvp == NULL || (dvp->v_iflag & VI_DOOMED) == 0, dvp,
786 	    ("cache_enter: Doomed vnode used as src"));
787 
788 	if (!doingcache)
789 		return;
790 
791 	/*
792 	 * Avoid blowout in namecache entries.
793 	 */
794 	if (numcache >= desiredvnodes * ncsizefactor)
795 		return;
796 
797 	ndd = nneg = NULL;
798 	flag = 0;
799 	if (cnp->cn_nameptr[0] == '.') {
800 		if (cnp->cn_namelen == 1)
801 			return;
802 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
803 			CACHE_WLOCK();
804 			/*
805 			 * If dotdot entry already exists, just retarget it
806 			 * to new parent vnode, otherwise continue with new
807 			 * namecache entry allocation.
808 			 */
809 			if ((ncp = dvp->v_cache_dd) != NULL &&
810 			    ncp->nc_flag & NCF_ISDOTDOT) {
811 				KASSERT(ncp->nc_dvp == dvp,
812 				    ("wrong isdotdot parent"));
813 				if (ncp->nc_vp != NULL) {
814 					TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst,
815 					    ncp, nc_dst);
816 				} else {
817 					cache_negative_remove(ncp);
818 				}
819 				if (vp != NULL) {
820 					TAILQ_INSERT_HEAD(&vp->v_cache_dst,
821 					    ncp, nc_dst);
822 				} else {
823 					cache_negative_insert(ncp);
824 				}
825 				ncp->nc_vp = vp;
826 				CACHE_WUNLOCK();
827 				return;
828 			}
829 			dvp->v_cache_dd = NULL;
830 			SDT_PROBE3(vfs, namecache, enter, done, dvp, "..", vp);
831 			CACHE_WUNLOCK();
832 			flag = NCF_ISDOTDOT;
833 		}
834 	}
835 
836 	/*
837 	 * Calculate the hash key and setup as much of the new
838 	 * namecache entry as possible before acquiring the lock.
839 	 */
840 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
841 	ncp->nc_vp = vp;
842 	ncp->nc_dvp = dvp;
843 	ncp->nc_flag = flag;
844 	if (tsp != NULL) {
845 		n3 = (struct namecache_ts *)ncp;
846 		n3->nc_time = *tsp;
847 		n3->nc_ticks = ticks;
848 		n3->nc_flag |= NCF_TS;
849 		if (dtsp != NULL) {
850 			n3->nc_dotdottime = *dtsp;
851 			n3->nc_flag |= NCF_DTS;
852 		}
853 	}
854 	len = ncp->nc_nlen = cnp->cn_namelen;
855 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
856 	strlcpy(nc_get_name(ncp), cnp->cn_nameptr, len + 1);
857 	CACHE_WLOCK();
858 
859 	/*
860 	 * See if this vnode or negative entry is already in the cache
861 	 * with this name.  This can happen with concurrent lookups of
862 	 * the same path name.
863 	 */
864 	ncpp = NCHHASH(hash);
865 	LIST_FOREACH(n2, ncpp, nc_hash) {
866 		if (n2->nc_dvp == dvp &&
867 		    n2->nc_nlen == cnp->cn_namelen &&
868 		    !bcmp(nc_get_name(n2), cnp->cn_nameptr, n2->nc_nlen)) {
869 			if (tsp != NULL) {
870 				KASSERT((n2->nc_flag & NCF_TS) != 0,
871 				    ("no NCF_TS"));
872 				n3 = (struct namecache_ts *)n2;
873 				n3->nc_time =
874 				    ((struct namecache_ts *)ncp)->nc_time;
875 				n3->nc_ticks =
876 				    ((struct namecache_ts *)ncp)->nc_ticks;
877 				if (dtsp != NULL) {
878 					n3->nc_dotdottime =
879 					    ((struct namecache_ts *)ncp)->
880 					    nc_dotdottime;
881 					n3->nc_flag |= NCF_DTS;
882 				}
883 			}
884 			CACHE_WUNLOCK();
885 			cache_free(ncp);
886 			return;
887 		}
888 	}
889 
890 	if (flag == NCF_ISDOTDOT) {
891 		/*
892 		 * See if we are trying to add .. entry, but some other lookup
893 		 * has populated v_cache_dd pointer already.
894 		 */
895 		if (dvp->v_cache_dd != NULL) {
896 			CACHE_WUNLOCK();
897 			cache_free(ncp);
898 			return;
899 		}
900 		KASSERT(vp == NULL || vp->v_type == VDIR,
901 		    ("wrong vnode type %p", vp));
902 		dvp->v_cache_dd = ncp;
903 	}
904 
905 	numcache++;
906 	if (vp != NULL) {
907 		if (vp->v_type == VDIR) {
908 			if (flag != NCF_ISDOTDOT) {
909 				/*
910 				 * For this case, the cache entry maps both the
911 				 * directory name in it and the name ".." for the
912 				 * directory's parent.
913 				 */
914 				if ((ndd = vp->v_cache_dd) != NULL) {
915 					if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
916 						cache_zap(ndd);
917 					else
918 						ndd = NULL;
919 				}
920 				vp->v_cache_dd = ncp;
921 			}
922 		} else {
923 			vp->v_cache_dd = NULL;
924 		}
925 	}
926 
927 	/*
928 	 * Insert the new namecache entry into the appropriate chain
929 	 * within the cache entries table.
930 	 */
931 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
932 	if (flag != NCF_ISDOTDOT) {
933 		if (LIST_EMPTY(&dvp->v_cache_src)) {
934 			vhold(dvp);
935 			numcachehv++;
936 		}
937 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
938 	}
939 
940 	/*
941 	 * If the entry is "negative", we place it into the
942 	 * "negative" cache queue, otherwise, we place it into the
943 	 * destination vnode's cache entries queue.
944 	 */
945 	if (vp != NULL) {
946 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
947 		SDT_PROBE3(vfs, namecache, enter, done, dvp, nc_get_name(ncp),
948 		    vp);
949 	} else {
950 		if (cnp->cn_flags & ISWHITEOUT)
951 			ncp->nc_flag |= NCF_WHITE;
952 		cache_negative_insert(ncp);
953 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
954 		    nc_get_name(ncp));
955 	}
956 	if (numneg * ncnegfactor > numcache)
957 		nneg = cache_negative_zap_one();
958 	CACHE_WUNLOCK();
959 	cache_free(ndd);
960 	cache_free(nneg);
961 }
962 
963 /*
964  * Name cache initialization, from vfs_init() when we are booting
965  */
966 static void
967 nchinit(void *dummy __unused)
968 {
969 
970 	TAILQ_INIT(&ncneg);
971 
972 	cache_zone_small = uma_zcreate("S VFS Cache",
973 	    sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1,
974 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
975 	cache_zone_small_ts = uma_zcreate("STS VFS Cache",
976 	    sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1,
977 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
978 	cache_zone_large = uma_zcreate("L VFS Cache",
979 	    sizeof(struct namecache) + NAME_MAX + 1,
980 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
981 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache",
982 	    sizeof(struct namecache_ts) + NAME_MAX + 1,
983 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
984 
985 	nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash);
986 
987 	numcalls = counter_u64_alloc(M_WAITOK);
988 	dothits = counter_u64_alloc(M_WAITOK);
989 	dotdothits = counter_u64_alloc(M_WAITOK);
990 	numchecks = counter_u64_alloc(M_WAITOK);
991 	nummiss = counter_u64_alloc(M_WAITOK);
992 	nummisszap = counter_u64_alloc(M_WAITOK);
993 	numposzaps = counter_u64_alloc(M_WAITOK);
994 	numposhits = counter_u64_alloc(M_WAITOK);
995 	numnegzaps = counter_u64_alloc(M_WAITOK);
996 	numneghits = counter_u64_alloc(M_WAITOK);
997 	numfullpathcalls = counter_u64_alloc(M_WAITOK);
998 	numfullpathfail1 = counter_u64_alloc(M_WAITOK);
999 	numfullpathfail2 = counter_u64_alloc(M_WAITOK);
1000 	numfullpathfail4 = counter_u64_alloc(M_WAITOK);
1001 	numfullpathfound = counter_u64_alloc(M_WAITOK);
1002 }
1003 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
1004 
1005 void
1006 cache_changesize(int newmaxvnodes)
1007 {
1008 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
1009 	u_long new_nchash, old_nchash;
1010 	struct namecache *ncp;
1011 	uint32_t hash;
1012 	int i;
1013 
1014 	new_nchashtbl = hashinit(newmaxvnodes * 2, M_VFSCACHE, &new_nchash);
1015 	/* If same hash table size, nothing to do */
1016 	if (nchash == new_nchash) {
1017 		free(new_nchashtbl, M_VFSCACHE);
1018 		return;
1019 	}
1020 	/*
1021 	 * Move everything from the old hash table to the new table.
1022 	 * None of the namecache entries in the table can be removed
1023 	 * because to do so, they have to be removed from the hash table.
1024 	 */
1025 	CACHE_WLOCK();
1026 	old_nchashtbl = nchashtbl;
1027 	old_nchash = nchash;
1028 	nchashtbl = new_nchashtbl;
1029 	nchash = new_nchash;
1030 	for (i = 0; i <= old_nchash; i++) {
1031 		while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) {
1032 			hash = cache_get_hash(nc_get_name(ncp), ncp->nc_nlen,
1033 			    ncp->nc_dvp);
1034 			LIST_REMOVE(ncp, nc_hash);
1035 			LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
1036 		}
1037 	}
1038 	CACHE_WUNLOCK();
1039 	free(old_nchashtbl, M_VFSCACHE);
1040 }
1041 
1042 /*
1043  * Invalidate all entries to a particular vnode.
1044  */
1045 void
1046 cache_purge(struct vnode *vp)
1047 {
1048 	TAILQ_HEAD(, namecache) ncps;
1049 	struct namecache *ncp, *nnp;
1050 
1051 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
1052 	SDT_PROBE1(vfs, namecache, purge, done, vp);
1053 	TAILQ_INIT(&ncps);
1054 	CACHE_WLOCK();
1055 	while (!LIST_EMPTY(&vp->v_cache_src)) {
1056 		ncp = LIST_FIRST(&vp->v_cache_src);
1057 		cache_zap(ncp);
1058 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1059 	}
1060 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
1061 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
1062 		cache_zap(ncp);
1063 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1064 	}
1065 	if (vp->v_cache_dd != NULL) {
1066 		ncp = vp->v_cache_dd;
1067 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
1068 		   ("lost dotdot link"));
1069 		cache_zap(ncp);
1070 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1071 	}
1072 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
1073 	CACHE_WUNLOCK();
1074 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
1075 		cache_free(ncp);
1076 	}
1077 }
1078 
1079 /*
1080  * Invalidate all negative entries for a particular directory vnode.
1081  */
1082 void
1083 cache_purge_negative(struct vnode *vp)
1084 {
1085 	TAILQ_HEAD(, namecache) ncps;
1086 	struct namecache *ncp, *nnp;
1087 
1088 	CTR1(KTR_VFS, "cache_purge_negative(%p)", vp);
1089 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
1090 	TAILQ_INIT(&ncps);
1091 	CACHE_WLOCK();
1092 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
1093 		if (ncp->nc_vp != NULL)
1094 			continue;
1095 		cache_zap(ncp);
1096 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1097 	}
1098 	CACHE_WUNLOCK();
1099 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
1100 		cache_free(ncp);
1101 	}
1102 }
1103 
1104 /*
1105  * Flush all entries referencing a particular filesystem.
1106  */
1107 void
1108 cache_purgevfs(struct mount *mp)
1109 {
1110 	TAILQ_HEAD(, namecache) ncps;
1111 	struct nchashhead *ncpp;
1112 	struct namecache *ncp, *nnp;
1113 
1114 	/* Scan hash tables for applicable entries */
1115 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
1116 	TAILQ_INIT(&ncps);
1117 	CACHE_WLOCK();
1118 	for (ncpp = &nchashtbl[nchash]; ncpp >= nchashtbl; ncpp--) {
1119 		LIST_FOREACH_SAFE(ncp, ncpp, nc_hash, nnp) {
1120 			if (ncp->nc_dvp->v_mount != mp)
1121 				continue;
1122 			cache_zap(ncp);
1123 			TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1124 		}
1125 	}
1126 	CACHE_WUNLOCK();
1127 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
1128 		cache_free(ncp);
1129 	}
1130 }
1131 
1132 /*
1133  * Perform canonical checks and cache lookup and pass on to filesystem
1134  * through the vop_cachedlookup only if needed.
1135  */
1136 
1137 int
1138 vfs_cache_lookup(struct vop_lookup_args *ap)
1139 {
1140 	struct vnode *dvp;
1141 	int error;
1142 	struct vnode **vpp = ap->a_vpp;
1143 	struct componentname *cnp = ap->a_cnp;
1144 	struct ucred *cred = cnp->cn_cred;
1145 	int flags = cnp->cn_flags;
1146 	struct thread *td = cnp->cn_thread;
1147 
1148 	*vpp = NULL;
1149 	dvp = ap->a_dvp;
1150 
1151 	if (dvp->v_type != VDIR)
1152 		return (ENOTDIR);
1153 
1154 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
1155 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
1156 		return (EROFS);
1157 
1158 	error = VOP_ACCESS(dvp, VEXEC, cred, td);
1159 	if (error)
1160 		return (error);
1161 
1162 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
1163 	if (error == 0)
1164 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
1165 	if (error == -1)
1166 		return (0);
1167 	return (error);
1168 }
1169 
1170 /*
1171  * XXX All of these sysctls would probably be more productive dead.
1172  */
1173 static int disablecwd;
1174 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
1175    "Disable the getcwd syscall");
1176 
1177 /* Implementation of the getcwd syscall. */
1178 int
1179 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
1180 {
1181 
1182 	return (kern___getcwd(td, uap->buf, UIO_USERSPACE, uap->buflen,
1183 	    MAXPATHLEN));
1184 }
1185 
1186 int
1187 kern___getcwd(struct thread *td, char *buf, enum uio_seg bufseg, u_int buflen,
1188     u_int path_max)
1189 {
1190 	char *bp, *tmpbuf;
1191 	struct filedesc *fdp;
1192 	struct vnode *cdir, *rdir;
1193 	int error;
1194 
1195 	if (disablecwd)
1196 		return (ENODEV);
1197 	if (buflen < 2)
1198 		return (EINVAL);
1199 	if (buflen > path_max)
1200 		buflen = path_max;
1201 
1202 	tmpbuf = malloc(buflen, M_TEMP, M_WAITOK);
1203 	fdp = td->td_proc->p_fd;
1204 	FILEDESC_SLOCK(fdp);
1205 	cdir = fdp->fd_cdir;
1206 	VREF(cdir);
1207 	rdir = fdp->fd_rdir;
1208 	VREF(rdir);
1209 	FILEDESC_SUNLOCK(fdp);
1210 	error = vn_fullpath1(td, cdir, rdir, tmpbuf, &bp, buflen);
1211 	vrele(rdir);
1212 	vrele(cdir);
1213 
1214 	if (!error) {
1215 		if (bufseg == UIO_SYSSPACE)
1216 			bcopy(bp, buf, strlen(bp) + 1);
1217 		else
1218 			error = copyout(bp, buf, strlen(bp) + 1);
1219 #ifdef KTRACE
1220 	if (KTRPOINT(curthread, KTR_NAMEI))
1221 		ktrnamei(bp);
1222 #endif
1223 	}
1224 	free(tmpbuf, M_TEMP);
1225 	return (error);
1226 }
1227 
1228 /*
1229  * Thus begins the fullpath magic.
1230  */
1231 
1232 static int disablefullpath;
1233 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, &disablefullpath, 0,
1234     "Disable the vn_fullpath function");
1235 
1236 /*
1237  * Retrieve the full filesystem path that correspond to a vnode from the name
1238  * cache (if available)
1239  */
1240 int
1241 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf)
1242 {
1243 	char *buf;
1244 	struct filedesc *fdp;
1245 	struct vnode *rdir;
1246 	int error;
1247 
1248 	if (disablefullpath)
1249 		return (ENODEV);
1250 	if (vn == NULL)
1251 		return (EINVAL);
1252 
1253 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1254 	fdp = td->td_proc->p_fd;
1255 	FILEDESC_SLOCK(fdp);
1256 	rdir = fdp->fd_rdir;
1257 	VREF(rdir);
1258 	FILEDESC_SUNLOCK(fdp);
1259 	error = vn_fullpath1(td, vn, rdir, buf, retbuf, MAXPATHLEN);
1260 	vrele(rdir);
1261 
1262 	if (!error)
1263 		*freebuf = buf;
1264 	else
1265 		free(buf, M_TEMP);
1266 	return (error);
1267 }
1268 
1269 /*
1270  * This function is similar to vn_fullpath, but it attempts to lookup the
1271  * pathname relative to the global root mount point.  This is required for the
1272  * auditing sub-system, as audited pathnames must be absolute, relative to the
1273  * global root mount point.
1274  */
1275 int
1276 vn_fullpath_global(struct thread *td, struct vnode *vn,
1277     char **retbuf, char **freebuf)
1278 {
1279 	char *buf;
1280 	int error;
1281 
1282 	if (disablefullpath)
1283 		return (ENODEV);
1284 	if (vn == NULL)
1285 		return (EINVAL);
1286 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1287 	error = vn_fullpath1(td, vn, rootvnode, buf, retbuf, MAXPATHLEN);
1288 	if (!error)
1289 		*freebuf = buf;
1290 	else
1291 		free(buf, M_TEMP);
1292 	return (error);
1293 }
1294 
1295 int
1296 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen)
1297 {
1298 	int error;
1299 
1300 	CACHE_RLOCK();
1301 	error = vn_vptocnp_locked(vp, cred, buf, buflen);
1302 	if (error == 0)
1303 		CACHE_RUNLOCK();
1304 	return (error);
1305 }
1306 
1307 static int
1308 vn_vptocnp_locked(struct vnode **vp, struct ucred *cred, char *buf,
1309     u_int *buflen)
1310 {
1311 	struct vnode *dvp;
1312 	struct namecache *ncp;
1313 	int error;
1314 
1315 	TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) {
1316 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1317 			break;
1318 	}
1319 	if (ncp != NULL) {
1320 		if (*buflen < ncp->nc_nlen) {
1321 			CACHE_RUNLOCK();
1322 			vrele(*vp);
1323 			counter_u64_add(numfullpathfail4, 1);
1324 			error = ENOMEM;
1325 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
1326 			    vp, NULL);
1327 			return (error);
1328 		}
1329 		*buflen -= ncp->nc_nlen;
1330 		memcpy(buf + *buflen, nc_get_name(ncp), ncp->nc_nlen);
1331 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
1332 		    nc_get_name(ncp), vp);
1333 		dvp = *vp;
1334 		*vp = ncp->nc_dvp;
1335 		vref(*vp);
1336 		CACHE_RUNLOCK();
1337 		vrele(dvp);
1338 		CACHE_RLOCK();
1339 		return (0);
1340 	}
1341 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
1342 
1343 	CACHE_RUNLOCK();
1344 	vn_lock(*vp, LK_SHARED | LK_RETRY);
1345 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
1346 	vput(*vp);
1347 	if (error) {
1348 		counter_u64_add(numfullpathfail2, 1);
1349 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
1350 		return (error);
1351 	}
1352 
1353 	*vp = dvp;
1354 	CACHE_RLOCK();
1355 	if (dvp->v_iflag & VI_DOOMED) {
1356 		/* forced unmount */
1357 		CACHE_RUNLOCK();
1358 		vrele(dvp);
1359 		error = ENOENT;
1360 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
1361 		return (error);
1362 	}
1363 	/*
1364 	 * *vp has its use count incremented still.
1365 	 */
1366 
1367 	return (0);
1368 }
1369 
1370 /*
1371  * The magic behind kern___getcwd() and vn_fullpath().
1372  */
1373 static int
1374 vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
1375     char *buf, char **retbuf, u_int buflen)
1376 {
1377 	int error, slash_prefixed;
1378 #ifdef KDTRACE_HOOKS
1379 	struct vnode *startvp = vp;
1380 #endif
1381 	struct vnode *vp1;
1382 
1383 	buflen--;
1384 	buf[buflen] = '\0';
1385 	error = 0;
1386 	slash_prefixed = 0;
1387 
1388 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
1389 	counter_u64_add(numfullpathcalls, 1);
1390 	vref(vp);
1391 	CACHE_RLOCK();
1392 	if (vp->v_type != VDIR) {
1393 		error = vn_vptocnp_locked(&vp, td->td_ucred, buf, &buflen);
1394 		if (error)
1395 			return (error);
1396 		if (buflen == 0) {
1397 			CACHE_RUNLOCK();
1398 			vrele(vp);
1399 			return (ENOMEM);
1400 		}
1401 		buf[--buflen] = '/';
1402 		slash_prefixed = 1;
1403 	}
1404 	while (vp != rdir && vp != rootvnode) {
1405 		if (vp->v_vflag & VV_ROOT) {
1406 			if (vp->v_iflag & VI_DOOMED) {	/* forced unmount */
1407 				CACHE_RUNLOCK();
1408 				vrele(vp);
1409 				error = ENOENT;
1410 				SDT_PROBE3(vfs, namecache, fullpath, return,
1411 				    error, vp, NULL);
1412 				break;
1413 			}
1414 			vp1 = vp->v_mount->mnt_vnodecovered;
1415 			vref(vp1);
1416 			CACHE_RUNLOCK();
1417 			vrele(vp);
1418 			vp = vp1;
1419 			CACHE_RLOCK();
1420 			continue;
1421 		}
1422 		if (vp->v_type != VDIR) {
1423 			CACHE_RUNLOCK();
1424 			vrele(vp);
1425 			counter_u64_add(numfullpathfail1, 1);
1426 			error = ENOTDIR;
1427 			SDT_PROBE3(vfs, namecache, fullpath, return,
1428 			    error, vp, NULL);
1429 			break;
1430 		}
1431 		error = vn_vptocnp_locked(&vp, td->td_ucred, buf, &buflen);
1432 		if (error)
1433 			break;
1434 		if (buflen == 0) {
1435 			CACHE_RUNLOCK();
1436 			vrele(vp);
1437 			error = ENOMEM;
1438 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
1439 			    startvp, NULL);
1440 			break;
1441 		}
1442 		buf[--buflen] = '/';
1443 		slash_prefixed = 1;
1444 	}
1445 	if (error)
1446 		return (error);
1447 	if (!slash_prefixed) {
1448 		if (buflen == 0) {
1449 			CACHE_RUNLOCK();
1450 			vrele(vp);
1451 			counter_u64_add(numfullpathfail4, 1);
1452 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
1453 			    startvp, NULL);
1454 			return (ENOMEM);
1455 		}
1456 		buf[--buflen] = '/';
1457 	}
1458 	counter_u64_add(numfullpathfound, 1);
1459 	CACHE_RUNLOCK();
1460 	vrele(vp);
1461 
1462 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, buf + buflen);
1463 	*retbuf = buf + buflen;
1464 	return (0);
1465 }
1466 
1467 struct vnode *
1468 vn_dir_dd_ino(struct vnode *vp)
1469 {
1470 	struct namecache *ncp;
1471 	struct vnode *ddvp;
1472 
1473 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
1474 	CACHE_RLOCK();
1475 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
1476 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
1477 			continue;
1478 		ddvp = ncp->nc_dvp;
1479 		vhold(ddvp);
1480 		CACHE_RUNLOCK();
1481 		if (vget(ddvp, LK_SHARED | LK_NOWAIT | LK_VNHELD, curthread))
1482 			return (NULL);
1483 		return (ddvp);
1484 	}
1485 	CACHE_RUNLOCK();
1486 	return (NULL);
1487 }
1488 
1489 int
1490 vn_commname(struct vnode *vp, char *buf, u_int buflen)
1491 {
1492 	struct namecache *ncp;
1493 	int l;
1494 
1495 	CACHE_RLOCK();
1496 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
1497 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1498 			break;
1499 	if (ncp == NULL) {
1500 		CACHE_RUNLOCK();
1501 		return (ENOENT);
1502 	}
1503 	l = min(ncp->nc_nlen, buflen - 1);
1504 	memcpy(buf, nc_get_name(ncp), l);
1505 	CACHE_RUNLOCK();
1506 	buf[l] = '\0';
1507 	return (0);
1508 }
1509 
1510 /* ABI compat shims for old kernel modules. */
1511 #undef cache_enter
1512 
1513 void	cache_enter(struct vnode *dvp, struct vnode *vp,
1514 	    struct componentname *cnp);
1515 
1516 void
1517 cache_enter(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
1518 {
1519 
1520 	cache_enter_time(dvp, vp, cnp, NULL, NULL);
1521 }
1522 
1523 /*
1524  * This function updates path string to vnode's full global path
1525  * and checks the size of the new path string against the pathlen argument.
1526  *
1527  * Requires a locked, referenced vnode.
1528  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
1529  *
1530  * If sysctl debug.disablefullpath is set, ENODEV is returned,
1531  * vnode is left locked and path remain untouched.
1532  *
1533  * If vp is a directory, the call to vn_fullpath_global() always succeeds
1534  * because it falls back to the ".." lookup if the namecache lookup fails.
1535  */
1536 int
1537 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
1538     u_int pathlen)
1539 {
1540 	struct nameidata nd;
1541 	struct vnode *vp1;
1542 	char *rpath, *fbuf;
1543 	int error;
1544 
1545 	ASSERT_VOP_ELOCKED(vp, __func__);
1546 
1547 	/* Return ENODEV if sysctl debug.disablefullpath==1 */
1548 	if (disablefullpath)
1549 		return (ENODEV);
1550 
1551 	/* Construct global filesystem path from vp. */
1552 	VOP_UNLOCK(vp, 0);
1553 	error = vn_fullpath_global(td, vp, &rpath, &fbuf);
1554 
1555 	if (error != 0) {
1556 		vrele(vp);
1557 		return (error);
1558 	}
1559 
1560 	if (strlen(rpath) >= pathlen) {
1561 		vrele(vp);
1562 		error = ENAMETOOLONG;
1563 		goto out;
1564 	}
1565 
1566 	/*
1567 	 * Re-lookup the vnode by path to detect a possible rename.
1568 	 * As a side effect, the vnode is relocked.
1569 	 * If vnode was renamed, return ENOENT.
1570 	 */
1571 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
1572 	    UIO_SYSSPACE, path, td);
1573 	error = namei(&nd);
1574 	if (error != 0) {
1575 		vrele(vp);
1576 		goto out;
1577 	}
1578 	NDFREE(&nd, NDF_ONLY_PNBUF);
1579 	vp1 = nd.ni_vp;
1580 	vrele(vp);
1581 	if (vp1 == vp)
1582 		strcpy(path, rpath);
1583 	else {
1584 		vput(vp1);
1585 		error = ENOENT;
1586 	}
1587 
1588 out:
1589 	free(fbuf, M_TEMP);
1590 	return (error);
1591 }
1592