xref: /freebsd/sys/kern/vfs_cache.c (revision 97bb4528d985130fe091155c4e485ff12322b197)
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  * 3. 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/smp.h>
55 #include <sys/syscallsubr.h>
56 #include <sys/sysctl.h>
57 #include <sys/sysproto.h>
58 #include <sys/vnode.h>
59 #ifdef KTRACE
60 #include <sys/ktrace.h>
61 #endif
62 
63 #include <vm/uma.h>
64 
65 SDT_PROVIDER_DECLARE(vfs);
66 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
67     "struct vnode *");
68 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
69     "char *");
70 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
71 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
72     "char *", "struct vnode *");
73 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
74 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
75     "struct vnode *", "char *");
76 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
77     "struct vnode *");
78 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
79     "struct vnode *", "char *");
80 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
81     "char *");
82 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *");
83 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
84 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
85 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
86     "struct vnode *");
87 SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *",
88     "char *");
89 
90 /*
91  * This structure describes the elements in the cache of recent
92  * names looked up by namei.
93  */
94 
95 struct	namecache {
96 	LIST_ENTRY(namecache) nc_hash;	/* hash chain */
97 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
98 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
99 	struct	vnode *nc_dvp;		/* vnode of parent of name */
100 	struct	vnode *nc_vp;		/* vnode the name refers to */
101 	u_char	nc_flag;		/* flag bits */
102 	u_char	nc_nlen;		/* length of name */
103 	char	nc_name[0];		/* segment name + nul */
104 };
105 
106 /*
107  * struct namecache_ts repeats struct namecache layout up to the
108  * nc_nlen member.
109  * struct namecache_ts is used in place of struct namecache when time(s) need
110  * to be stored.  The nc_dotdottime field is used when a cache entry is mapping
111  * both a non-dotdot directory name plus dotdot for the directory's
112  * parent.
113  */
114 struct	namecache_ts {
115 	LIST_ENTRY(namecache) nc_hash;	/* hash chain */
116 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
117 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
118 	struct	vnode *nc_dvp;		/* vnode of parent of name */
119 	struct	vnode *nc_vp;		/* vnode the name refers to */
120 	u_char	nc_flag;		/* flag bits */
121 	u_char	nc_nlen;		/* length of name */
122 	struct	timespec nc_time;	/* timespec provided by fs */
123 	struct	timespec nc_dotdottime;	/* dotdot timespec provided by fs */
124 	int	nc_ticks;		/* ticks value when entry was added */
125 	char	nc_name[0];		/* segment name + nul */
126 };
127 
128 /*
129  * Flags in namecache.nc_flag
130  */
131 #define NCF_WHITE	0x01
132 #define NCF_ISDOTDOT	0x02
133 #define	NCF_TS		0x04
134 #define	NCF_DTS		0x08
135 #define	NCF_DVDROP	0x10
136 
137 /*
138  * Name caching works as follows:
139  *
140  * Names found by directory scans are retained in a cache
141  * for future reference.  It is managed LRU, so frequently
142  * used names will hang around.  Cache is indexed by hash value
143  * obtained from (vp, name) where vp refers to the directory
144  * containing name.
145  *
146  * If it is a "negative" entry, (i.e. for a name that is known NOT to
147  * exist) the vnode pointer will be NULL.
148  *
149  * Upon reaching the last segment of a path, if the reference
150  * is for DELETE, or NOCACHE is set (rewrite), and the
151  * name is located in the cache, it will be dropped.
152  *
153  * These locks are used (in the order in which they can be taken):
154  * NAME         TYPE    ROLE
155  * cache_lock   rwlock  global, needed for all modifications
156  * bucketlock   rwlock  for access to given hash bucket
157  * ncneg_mtx    mtx     negative entry LRU management
158  *
159  * A name -> vnode lookup can be safely performed by either locking cache_lock
160  * or the relevant hash bucket.
161  *
162  * ".." and vnode -> name lookups require cache_lock.
163  *
164  * Modifications require both cache_lock and relevant bucketlock taken for
165  * writing.
166  *
167  * Negative entry LRU management requires ncneg_mtx taken on top of either
168  * cache_lock or bucketlock.
169  */
170 
171 /*
172  * Structures associated with name caching.
173  */
174 #define NCHHASH(hash) \
175 	(&nchashtbl[(hash) & nchash])
176 static LIST_HEAD(nchashhead, namecache) *nchashtbl;	/* Hash Table */
177 static TAILQ_HEAD(, namecache) ncneg;	/* Hash Table */
178 static u_long	nchash;			/* size of hash table */
179 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
180     "Size of namecache hash table");
181 static u_long	ncnegfactor = 16;	/* ratio of negative entries */
182 SYSCTL_ULONG(_vfs, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
183     "Ratio of negative namecache entries");
184 static u_long	numneg;			/* number of negative entries allocated */
185 SYSCTL_ULONG(_debug, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0,
186     "Number of negative entries in namecache");
187 static u_long	numcache;		/* number of cache entries allocated */
188 SYSCTL_ULONG(_debug, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0,
189     "Number of namecache entries");
190 static u_long	numcachehv;		/* number of cache entries with vnodes held */
191 SYSCTL_ULONG(_debug, OID_AUTO, numcachehv, CTLFLAG_RD, &numcachehv, 0,
192     "Number of namecache entries with vnodes held");
193 u_int	ncsizefactor = 2;
194 SYSCTL_UINT(_vfs, OID_AUTO, ncsizefactor, CTLFLAG_RW, &ncsizefactor, 0,
195     "Size factor for namecache");
196 
197 struct nchstats	nchstats;		/* cache effectiveness statistics */
198 
199 static struct rwlock cache_lock;
200 RW_SYSINIT(vfscache, &cache_lock, "ncglobal");
201 
202 #define	CACHE_TRY_WLOCK()	rw_try_wlock(&cache_lock)
203 #define	CACHE_UPGRADE_LOCK()	rw_try_upgrade(&cache_lock)
204 #define	CACHE_RLOCK()		rw_rlock(&cache_lock)
205 #define	CACHE_RUNLOCK()		rw_runlock(&cache_lock)
206 #define	CACHE_WLOCK()		rw_wlock(&cache_lock)
207 #define	CACHE_WUNLOCK()		rw_wunlock(&cache_lock)
208 
209 static struct mtx_padalign ncneg_mtx;
210 MTX_SYSINIT(vfscache_neg, &ncneg_mtx, "ncneg", MTX_DEF);
211 
212 static u_int   numbucketlocks;
213 static struct rwlock_padalign  *bucketlocks;
214 #define	HASH2BUCKETLOCK(hash) \
215 	((struct rwlock *)(&bucketlocks[((hash) % numbucketlocks)]))
216 
217 /*
218  * UMA zones for the VFS cache.
219  *
220  * The small cache is used for entries with short names, which are the
221  * most common.  The large cache is used for entries which are too big to
222  * fit in the small cache.
223  */
224 static uma_zone_t cache_zone_small;
225 static uma_zone_t cache_zone_small_ts;
226 static uma_zone_t cache_zone_large;
227 static uma_zone_t cache_zone_large_ts;
228 
229 #define	CACHE_PATH_CUTOFF	35
230 
231 static struct namecache *
232 cache_alloc(int len, int ts)
233 {
234 
235 	if (len > CACHE_PATH_CUTOFF) {
236 		if (ts)
237 			return (uma_zalloc(cache_zone_large_ts, M_WAITOK));
238 		else
239 			return (uma_zalloc(cache_zone_large, M_WAITOK));
240 	}
241 	if (ts)
242 		return (uma_zalloc(cache_zone_small_ts, M_WAITOK));
243 	else
244 		return (uma_zalloc(cache_zone_small, M_WAITOK));
245 }
246 
247 static void
248 cache_free(struct namecache *ncp)
249 {
250 	int ts;
251 
252 	if (ncp == NULL)
253 		return;
254 	ts = ncp->nc_flag & NCF_TS;
255 	if ((ncp->nc_flag & NCF_DVDROP) != 0)
256 		vdrop(ncp->nc_dvp);
257 	if (ncp->nc_nlen <= CACHE_PATH_CUTOFF) {
258 		if (ts)
259 			uma_zfree(cache_zone_small_ts, ncp);
260 		else
261 			uma_zfree(cache_zone_small, ncp);
262 	} else if (ts)
263 		uma_zfree(cache_zone_large_ts, ncp);
264 	else
265 		uma_zfree(cache_zone_large, ncp);
266 }
267 
268 static char *
269 nc_get_name(struct namecache *ncp)
270 {
271 	struct namecache_ts *ncp_ts;
272 
273 	if ((ncp->nc_flag & NCF_TS) == 0)
274 		return (ncp->nc_name);
275 	ncp_ts = (struct namecache_ts *)ncp;
276 	return (ncp_ts->nc_name);
277 }
278 
279 static void
280 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
281 {
282 
283 	KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
284 	    (tsp == NULL && ticksp == NULL),
285 	    ("No NCF_TS"));
286 
287 	if (tsp != NULL)
288 		*tsp = ((struct namecache_ts *)ncp)->nc_time;
289 	if (ticksp != NULL)
290 		*ticksp = ((struct namecache_ts *)ncp)->nc_ticks;
291 }
292 
293 static int	doingcache = 1;		/* 1 => enable the cache */
294 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
295     "VFS namecache enabled");
296 
297 /* Export size information to userland */
298 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
299     sizeof(struct namecache), "sizeof(struct namecache)");
300 
301 /*
302  * The new name cache statistics
303  */
304 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0,
305     "Name cache statistics");
306 #define STATNODE_ULONG(name, descr)	\
307 	SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, descr);
308 #define STATNODE_COUNTER(name, descr)	\
309 	static counter_u64_t name;	\
310 	SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, descr);
311 STATNODE_ULONG(numneg, "Number of negative cache entries");
312 STATNODE_ULONG(numcache, "Number of cache entries");
313 STATNODE_COUNTER(numcalls, "Number of cache lookups");
314 STATNODE_COUNTER(dothits, "Number of '.' hits");
315 STATNODE_COUNTER(dotdothits, "Number of '..' hits");
316 STATNODE_COUNTER(numchecks, "Number of checks in lookup");
317 STATNODE_COUNTER(nummiss, "Number of cache misses");
318 STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache");
319 STATNODE_COUNTER(numposzaps,
320     "Number of cache hits (positive) we do not want to cache");
321 STATNODE_COUNTER(numposhits, "Number of cache hits (positive)");
322 STATNODE_COUNTER(numnegzaps,
323     "Number of cache hits (negative) we do not want to cache");
324 STATNODE_COUNTER(numneghits, "Number of cache hits (negative)");
325 /* These count for kern___getcwd(), too. */
326 STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls");
327 STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
328 STATNODE_COUNTER(numfullpathfail2,
329     "Number of fullpath search errors (VOP_VPTOCNP failures)");
330 STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
331 STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls");
332 static long numupgrades; STATNODE_ULONG(numupgrades,
333     "Number of updates of the cache after lookup (write lock + retry)");
334 static long zap_and_exit_bucket_fail; STATNODE_ULONG(zap_and_exit_bucket_fail,
335     "Number of times bucketlocked zap_and_exit case failed to writelock");
336 
337 static void cache_zap(struct namecache *ncp);
338 static int vn_vptocnp_locked(struct vnode **vp, struct ucred *cred, char *buf,
339     u_int *buflen);
340 static int vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
341     char *buf, char **retbuf, u_int buflen);
342 
343 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
344 
345 static uint32_t
346 cache_get_hash(char *name, u_char len, struct vnode *dvp)
347 {
348 	uint32_t hash;
349 
350 	hash = fnv_32_buf(name, len, FNV1_32_INIT);
351 	hash = fnv_32_buf(&dvp, sizeof(dvp), hash);
352 	return (hash);
353 }
354 
355 #ifdef INVARIANTS
356 static void
357 cache_assert_bucket_locked(struct namecache *ncp, int mode)
358 {
359 	struct rwlock *bucketlock;
360 	uint32_t hash;
361 
362 	hash = cache_get_hash(nc_get_name(ncp), ncp->nc_nlen, ncp->nc_dvp);
363 	bucketlock = HASH2BUCKETLOCK(hash);
364 	rw_assert(bucketlock, mode);
365 }
366 #else
367 #define cache_assert_bucket_locked(x, y) do { } while (0)
368 #endif
369 
370 static void
371 cache_lock_all_buckets(void)
372 {
373 	u_int i;
374 
375 	for (i = 0; i < numbucketlocks; i++)
376 		rw_wlock(&bucketlocks[i]);
377 }
378 
379 static void
380 cache_unlock_all_buckets(void)
381 {
382 	u_int i;
383 
384 	for (i = 0; i < numbucketlocks; i++)
385 		rw_wunlock(&bucketlocks[i]);
386 }
387 
388 static int
389 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
390 {
391 	struct nchstats snap;
392 
393 	if (req->oldptr == NULL)
394 		return (SYSCTL_OUT(req, 0, sizeof(snap)));
395 
396 	snap = nchstats;
397 	snap.ncs_goodhits = counter_u64_fetch(numposhits);
398 	snap.ncs_neghits = counter_u64_fetch(numneghits);
399 	snap.ncs_badhits = counter_u64_fetch(numposzaps) +
400 	    counter_u64_fetch(numnegzaps);
401 	snap.ncs_miss = counter_u64_fetch(nummisszap) +
402 	    counter_u64_fetch(nummiss);
403 
404 	return (SYSCTL_OUT(req, &snap, sizeof(snap)));
405 }
406 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
407     CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
408     "VFS cache effectiveness statistics");
409 
410 #ifdef DIAGNOSTIC
411 /*
412  * Grab an atomic snapshot of the name cache hash chain lengths
413  */
414 static SYSCTL_NODE(_debug, OID_AUTO, hashstat, CTLFLAG_RW, NULL,
415     "hash table stats");
416 
417 static int
418 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
419 {
420 	struct nchashhead *ncpp;
421 	struct namecache *ncp;
422 	int i, error, n_nchash, *cntbuf;
423 
424 retry:
425 	n_nchash = nchash + 1;	/* nchash is max index, not count */
426 	if (req->oldptr == NULL)
427 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
428 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
429 	CACHE_RLOCK();
430 	if (n_nchash != nchash + 1) {
431 		CACHE_RUNLOCK();
432 		free(cntbuf, M_TEMP);
433 		goto retry;
434 	}
435 	/* Scan hash tables counting entries */
436 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
437 		LIST_FOREACH(ncp, ncpp, nc_hash)
438 			cntbuf[i]++;
439 	CACHE_RUNLOCK();
440 	for (error = 0, i = 0; i < n_nchash; i++)
441 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
442 			break;
443 	free(cntbuf, M_TEMP);
444 	return (error);
445 }
446 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
447     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
448     "nchash chain lengths");
449 
450 static int
451 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
452 {
453 	int error;
454 	struct nchashhead *ncpp;
455 	struct namecache *ncp;
456 	int n_nchash;
457 	int count, maxlength, used, pct;
458 
459 	if (!req->oldptr)
460 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
461 
462 	CACHE_RLOCK();
463 	n_nchash = nchash + 1;	/* nchash is max index, not count */
464 	used = 0;
465 	maxlength = 0;
466 
467 	/* Scan hash tables for applicable entries */
468 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
469 		count = 0;
470 		LIST_FOREACH(ncp, ncpp, nc_hash) {
471 			count++;
472 		}
473 		if (count)
474 			used++;
475 		if (maxlength < count)
476 			maxlength = count;
477 	}
478 	n_nchash = nchash + 1;
479 	CACHE_RUNLOCK();
480 	pct = (used * 100) / (n_nchash / 100);
481 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
482 	if (error)
483 		return (error);
484 	error = SYSCTL_OUT(req, &used, sizeof(used));
485 	if (error)
486 		return (error);
487 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
488 	if (error)
489 		return (error);
490 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
491 	if (error)
492 		return (error);
493 	return (0);
494 }
495 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
496     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
497     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
498 #endif
499 
500 /*
501  * Negative entries management
502  */
503 static void
504 cache_negative_hit(struct namecache *ncp)
505 {
506 
507 	mtx_lock(&ncneg_mtx);
508 	TAILQ_REMOVE(&ncneg, ncp, nc_dst);
509 	TAILQ_INSERT_TAIL(&ncneg, ncp, nc_dst);
510 	mtx_unlock(&ncneg_mtx);
511 }
512 
513 static void
514 cache_negative_insert(struct namecache *ncp)
515 {
516 
517 	rw_assert(&cache_lock, RA_WLOCKED);
518 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
519 	MPASS(ncp->nc_vp == NULL);
520 	mtx_lock(&ncneg_mtx);
521 	TAILQ_INSERT_TAIL(&ncneg, ncp, nc_dst);
522 	numneg++;
523 	mtx_unlock(&ncneg_mtx);
524 }
525 
526 static void
527 cache_negative_remove(struct namecache *ncp)
528 {
529 
530 	rw_assert(&cache_lock, RA_WLOCKED);
531 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
532 	MPASS(ncp->nc_vp == NULL);
533 	mtx_lock(&ncneg_mtx);
534 	TAILQ_REMOVE(&ncneg, ncp, nc_dst);
535 	numneg--;
536 	mtx_unlock(&ncneg_mtx);
537 }
538 
539 static struct namecache *
540 cache_negative_zap_one(void)
541 {
542 	struct namecache *ncp;
543 
544 	rw_assert(&cache_lock, RA_WLOCKED);
545 	ncp = TAILQ_FIRST(&ncneg);
546 	KASSERT(ncp->nc_vp == NULL, ("ncp %p vp %p on ncneg",
547 	    ncp, ncp->nc_vp));
548 	cache_zap(ncp);
549 	return (ncp);
550 }
551 
552 /*
553  * cache_zap():
554  *
555  *   Removes a namecache entry from cache, whether it contains an actual
556  *   pointer to a vnode or if it is just a negative cache entry.
557  */
558 static void
559 cache_zap_locked(struct namecache *ncp)
560 {
561 
562 	rw_assert(&cache_lock, RA_WLOCKED);
563 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
564 	CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp, ncp->nc_vp);
565 	if (ncp->nc_vp != NULL) {
566 		SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
567 		    nc_get_name(ncp), ncp->nc_vp);
568 	} else {
569 		SDT_PROBE2(vfs, namecache, zap_negative, done, ncp->nc_dvp,
570 		    nc_get_name(ncp));
571 	}
572 	LIST_REMOVE(ncp, nc_hash);
573 	if (ncp->nc_flag & NCF_ISDOTDOT) {
574 		if (ncp == ncp->nc_dvp->v_cache_dd)
575 			ncp->nc_dvp->v_cache_dd = NULL;
576 	} else {
577 		LIST_REMOVE(ncp, nc_src);
578 		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
579 			ncp->nc_flag |= NCF_DVDROP;
580 			numcachehv--;
581 		}
582 	}
583 	if (ncp->nc_vp) {
584 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
585 		if (ncp == ncp->nc_vp->v_cache_dd)
586 			ncp->nc_vp->v_cache_dd = NULL;
587 	} else {
588 		cache_negative_remove(ncp);
589 	}
590 	numcache--;
591 }
592 
593 static void
594 cache_zap(struct namecache *ncp)
595 {
596 	struct rwlock *bucketlock;
597 	uint32_t hash;
598 
599 	rw_assert(&cache_lock, RA_WLOCKED);
600 
601 	hash = cache_get_hash(nc_get_name(ncp), ncp->nc_nlen, ncp->nc_dvp);
602 	bucketlock = HASH2BUCKETLOCK(hash);
603 	rw_wlock(bucketlock);
604 	cache_zap_locked(ncp);
605 	rw_wunlock(bucketlock);
606 }
607 
608 /*
609  * Lookup an entry in the cache
610  *
611  * Lookup is called with dvp pointing to the directory to search,
612  * cnp pointing to the name of the entry being sought. If the lookup
613  * succeeds, the vnode is returned in *vpp, and a status of -1 is
614  * returned. If the lookup determines that the name does not exist
615  * (negative caching), a status of ENOENT is returned. If the lookup
616  * fails, a status of zero is returned.  If the directory vnode is
617  * recycled out from under us due to a forced unmount, a status of
618  * ENOENT is returned.
619  *
620  * vpp is locked and ref'd on return.  If we're looking up DOTDOT, dvp is
621  * unlocked.  If we're looking up . an extra ref is taken, but the lock is
622  * not recursively acquired.
623  */
624 
625 enum { UNLOCKED, WLOCKED, RLOCKED };
626 
627 static void
628 cache_unlock(int cache_locked)
629 {
630 
631 	switch (cache_locked) {
632 	case UNLOCKED:
633 		break;
634 	case WLOCKED:
635 		CACHE_WUNLOCK();
636 		break;
637 	case RLOCKED:
638 		CACHE_RUNLOCK();
639 		break;
640 	}
641 }
642 
643 int
644 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
645     struct timespec *tsp, int *ticksp)
646 {
647 	struct rwlock *bucketlock;
648 	struct namecache *ncp;
649 	uint32_t hash;
650 	int error, ltype, cache_locked;
651 
652 	if (!doingcache) {
653 		cnp->cn_flags &= ~MAKEENTRY;
654 		return (0);
655 	}
656 retry:
657 	bucketlock = NULL;
658 	cache_locked = UNLOCKED;
659 	error = 0;
660 	counter_u64_add(numcalls, 1);
661 
662 retry_wlocked:
663 	if (cnp->cn_nameptr[0] == '.') {
664 		if (cnp->cn_namelen == 1) {
665 			*vpp = dvp;
666 			CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .",
667 			    dvp, cnp->cn_nameptr);
668 			counter_u64_add(dothits, 1);
669 			SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
670 			if (tsp != NULL)
671 				timespecclear(tsp);
672 			if (ticksp != NULL)
673 				*ticksp = ticks;
674 			VREF(*vpp);
675 			/*
676 			 * When we lookup "." we still can be asked to lock it
677 			 * differently...
678 			 */
679 			ltype = cnp->cn_lkflags & LK_TYPE_MASK;
680 			if (ltype != VOP_ISLOCKED(*vpp)) {
681 				if (ltype == LK_EXCLUSIVE) {
682 					vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
683 					if ((*vpp)->v_iflag & VI_DOOMED) {
684 						/* forced unmount */
685 						vrele(*vpp);
686 						*vpp = NULL;
687 						return (ENOENT);
688 					}
689 				} else
690 					vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
691 			}
692 			return (-1);
693 		}
694 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
695 			counter_u64_add(dotdothits, 1);
696 			if (cache_locked == UNLOCKED) {
697 				CACHE_RLOCK();
698 				cache_locked = RLOCKED;
699 			}
700 
701 			if (dvp->v_cache_dd == NULL) {
702 				SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
703 				    "..", NULL);
704 				goto unlock;
705 			}
706 			if ((cnp->cn_flags & MAKEENTRY) == 0) {
707 				if (cache_locked != WLOCKED &&
708 				    !CACHE_UPGRADE_LOCK())
709 					goto wlock;
710 				ncp = NULL;
711 				if (dvp->v_cache_dd->nc_flag & NCF_ISDOTDOT) {
712 					ncp = dvp->v_cache_dd;
713 					cache_zap(ncp);
714 				}
715 				dvp->v_cache_dd = NULL;
716 				CACHE_WUNLOCK();
717 				cache_free(ncp);
718 				return (0);
719 			}
720 			ncp = dvp->v_cache_dd;
721 			if (ncp->nc_flag & NCF_ISDOTDOT)
722 				*vpp = ncp->nc_vp;
723 			else
724 				*vpp = ncp->nc_dvp;
725 			/* Return failure if negative entry was found. */
726 			if (*vpp == NULL)
727 				goto negative_success;
728 			CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
729 			    dvp, cnp->cn_nameptr, *vpp);
730 			SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..",
731 			    *vpp);
732 			cache_out_ts(ncp, tsp, ticksp);
733 			if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
734 			    NCF_DTS && tsp != NULL)
735 				*tsp = ((struct namecache_ts *)ncp)->
736 				    nc_dotdottime;
737 			goto success;
738 		}
739 	}
740 
741 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
742 	if (cache_locked == UNLOCKED) {
743 		bucketlock = HASH2BUCKETLOCK(hash);
744 		rw_rlock(bucketlock);
745 	}
746 
747 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
748 		counter_u64_add(numchecks, 1);
749 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
750 		    !bcmp(nc_get_name(ncp), cnp->cn_nameptr, ncp->nc_nlen))
751 			break;
752 	}
753 
754 	/* We failed to find an entry */
755 	if (ncp == NULL) {
756 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
757 		    NULL);
758 		if ((cnp->cn_flags & MAKEENTRY) == 0) {
759 			counter_u64_add(nummisszap, 1);
760 		} else {
761 			counter_u64_add(nummiss, 1);
762 		}
763 		goto unlock;
764 	}
765 
766 	/* We don't want to have an entry, so dump it */
767 	if ((cnp->cn_flags & MAKEENTRY) == 0) {
768 		counter_u64_add(numposzaps, 1);
769 		goto zap_and_exit;
770 	}
771 
772 	/* We found a "positive" match, return the vnode */
773 	if (ncp->nc_vp) {
774 		counter_u64_add(numposhits, 1);
775 		*vpp = ncp->nc_vp;
776 		CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p",
777 		    dvp, cnp->cn_nameptr, *vpp, ncp);
778 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, nc_get_name(ncp),
779 		    *vpp);
780 		cache_out_ts(ncp, tsp, ticksp);
781 		goto success;
782 	}
783 
784 negative_success:
785 	/* We found a negative match, and want to create it, so purge */
786 	if (cnp->cn_nameiop == CREATE) {
787 		counter_u64_add(numnegzaps, 1);
788 		goto zap_and_exit;
789 	}
790 
791 	counter_u64_add(numneghits, 1);
792 	cache_negative_hit(ncp);
793 	if (ncp->nc_flag & NCF_WHITE)
794 		cnp->cn_flags |= ISWHITEOUT;
795 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp,
796 	    nc_get_name(ncp));
797 	cache_out_ts(ncp, tsp, ticksp);
798 	MPASS(bucketlock != NULL || cache_locked != UNLOCKED);
799 	if (bucketlock != NULL)
800 		rw_runlock(bucketlock);
801 	cache_unlock(cache_locked);
802 	return (ENOENT);
803 
804 wlock:
805 	/*
806 	 * We need to update the cache after our lookup, so upgrade to
807 	 * a write lock and retry the operation.
808 	 */
809 	CACHE_RUNLOCK();
810 wlock_unlocked:
811 	CACHE_WLOCK();
812 	numupgrades++;
813 	cache_locked = WLOCKED;
814 	goto retry_wlocked;
815 
816 success:
817 	/*
818 	 * On success we return a locked and ref'd vnode as per the lookup
819 	 * protocol.
820 	 */
821 	MPASS(dvp != *vpp);
822 	ltype = 0;	/* silence gcc warning */
823 	if (cnp->cn_flags & ISDOTDOT) {
824 		ltype = VOP_ISLOCKED(dvp);
825 		VOP_UNLOCK(dvp, 0);
826 	}
827 	vhold(*vpp);
828 	MPASS(bucketlock != NULL || cache_locked != UNLOCKED);
829 	if (bucketlock != NULL)
830 		rw_runlock(bucketlock);
831 	cache_unlock(cache_locked);
832 	error = vget(*vpp, cnp->cn_lkflags | LK_VNHELD, cnp->cn_thread);
833 	if (cnp->cn_flags & ISDOTDOT) {
834 		vn_lock(dvp, ltype | LK_RETRY);
835 		if (dvp->v_iflag & VI_DOOMED) {
836 			if (error == 0)
837 				vput(*vpp);
838 			*vpp = NULL;
839 			return (ENOENT);
840 		}
841 	}
842 	if (error) {
843 		*vpp = NULL;
844 		goto retry;
845 	}
846 	if ((cnp->cn_flags & ISLASTCN) &&
847 	    (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) {
848 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
849 	}
850 	return (-1);
851 
852 unlock:
853 	MPASS(bucketlock != NULL || cache_locked != UNLOCKED);
854 	if (bucketlock != NULL)
855 		rw_runlock(bucketlock);
856 	cache_unlock(cache_locked);
857 	return (0);
858 
859 zap_and_exit:
860 	if (bucketlock != NULL) {
861 		rw_assert(&cache_lock, RA_UNLOCKED);
862 		if (!CACHE_TRY_WLOCK()) {
863 			rw_runlock(bucketlock);
864 			bucketlock = NULL;
865 			zap_and_exit_bucket_fail++;
866 			goto wlock_unlocked;
867 		}
868 		cache_locked = WLOCKED;
869 		rw_runlock(bucketlock);
870 		bucketlock = NULL;
871 	} else if (cache_locked != WLOCKED && !CACHE_UPGRADE_LOCK())
872 		goto wlock;
873 	cache_zap(ncp);
874 	CACHE_WUNLOCK();
875 	cache_free(ncp);
876 	return (0);
877 }
878 
879 /*
880  * Add an entry to the cache.
881  */
882 void
883 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
884     struct timespec *tsp, struct timespec *dtsp)
885 {
886 	struct rwlock *bucketlock;
887 	struct namecache *ncp, *n2, *ndd, *nneg;
888 	struct namecache_ts *n3;
889 	struct nchashhead *ncpp;
890 	uint32_t hash;
891 	int flag;
892 	int len;
893 
894 	CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr);
895 	VNASSERT(vp == NULL || (vp->v_iflag & VI_DOOMED) == 0, vp,
896 	    ("cache_enter: Adding a doomed vnode"));
897 	VNASSERT(dvp == NULL || (dvp->v_iflag & VI_DOOMED) == 0, dvp,
898 	    ("cache_enter: Doomed vnode used as src"));
899 
900 	if (!doingcache)
901 		return;
902 
903 	/*
904 	 * Avoid blowout in namecache entries.
905 	 */
906 	if (numcache >= desiredvnodes * ncsizefactor)
907 		return;
908 
909 	ndd = nneg = NULL;
910 	flag = 0;
911 	if (cnp->cn_nameptr[0] == '.') {
912 		if (cnp->cn_namelen == 1)
913 			return;
914 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
915 			CACHE_WLOCK();
916 			/*
917 			 * If dotdot entry already exists, just retarget it
918 			 * to new parent vnode, otherwise continue with new
919 			 * namecache entry allocation.
920 			 */
921 			if ((ncp = dvp->v_cache_dd) != NULL &&
922 			    ncp->nc_flag & NCF_ISDOTDOT) {
923 				KASSERT(ncp->nc_dvp == dvp,
924 				    ("wrong isdotdot parent"));
925 				if (ncp->nc_vp != NULL) {
926 					TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst,
927 					    ncp, nc_dst);
928 				} else {
929 					cache_negative_remove(ncp);
930 				}
931 				if (vp != NULL) {
932 					TAILQ_INSERT_HEAD(&vp->v_cache_dst,
933 					    ncp, nc_dst);
934 				} else {
935 					cache_negative_insert(ncp);
936 				}
937 				ncp->nc_vp = vp;
938 				CACHE_WUNLOCK();
939 				return;
940 			}
941 			dvp->v_cache_dd = NULL;
942 			SDT_PROBE3(vfs, namecache, enter, done, dvp, "..", vp);
943 			CACHE_WUNLOCK();
944 			flag = NCF_ISDOTDOT;
945 		}
946 	}
947 
948 	/*
949 	 * Calculate the hash key and setup as much of the new
950 	 * namecache entry as possible before acquiring the lock.
951 	 */
952 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
953 	ncp->nc_vp = vp;
954 	ncp->nc_dvp = dvp;
955 	ncp->nc_flag = flag;
956 	if (tsp != NULL) {
957 		n3 = (struct namecache_ts *)ncp;
958 		n3->nc_time = *tsp;
959 		n3->nc_ticks = ticks;
960 		n3->nc_flag |= NCF_TS;
961 		if (dtsp != NULL) {
962 			n3->nc_dotdottime = *dtsp;
963 			n3->nc_flag |= NCF_DTS;
964 		}
965 	}
966 	len = ncp->nc_nlen = cnp->cn_namelen;
967 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
968 	strlcpy(nc_get_name(ncp), cnp->cn_nameptr, len + 1);
969 	CACHE_WLOCK();
970 
971 	/*
972 	 * See if this vnode or negative entry is already in the cache
973 	 * with this name.  This can happen with concurrent lookups of
974 	 * the same path name.
975 	 */
976 	ncpp = NCHHASH(hash);
977 	LIST_FOREACH(n2, ncpp, nc_hash) {
978 		if (n2->nc_dvp == dvp &&
979 		    n2->nc_nlen == cnp->cn_namelen &&
980 		    !bcmp(nc_get_name(n2), cnp->cn_nameptr, n2->nc_nlen)) {
981 			if (tsp != NULL) {
982 				KASSERT((n2->nc_flag & NCF_TS) != 0,
983 				    ("no NCF_TS"));
984 				n3 = (struct namecache_ts *)n2;
985 				n3->nc_time =
986 				    ((struct namecache_ts *)ncp)->nc_time;
987 				n3->nc_ticks =
988 				    ((struct namecache_ts *)ncp)->nc_ticks;
989 				if (dtsp != NULL) {
990 					n3->nc_dotdottime =
991 					    ((struct namecache_ts *)ncp)->
992 					    nc_dotdottime;
993 					n3->nc_flag |= NCF_DTS;
994 				}
995 			}
996 			CACHE_WUNLOCK();
997 			cache_free(ncp);
998 			return;
999 		}
1000 	}
1001 
1002 	if (flag == NCF_ISDOTDOT) {
1003 		/*
1004 		 * See if we are trying to add .. entry, but some other lookup
1005 		 * has populated v_cache_dd pointer already.
1006 		 */
1007 		if (dvp->v_cache_dd != NULL) {
1008 			CACHE_WUNLOCK();
1009 			cache_free(ncp);
1010 			return;
1011 		}
1012 		KASSERT(vp == NULL || vp->v_type == VDIR,
1013 		    ("wrong vnode type %p", vp));
1014 		dvp->v_cache_dd = ncp;
1015 	}
1016 
1017 	numcache++;
1018 	if (vp != NULL) {
1019 		if (vp->v_type == VDIR) {
1020 			if (flag != NCF_ISDOTDOT) {
1021 				/*
1022 				 * For this case, the cache entry maps both the
1023 				 * directory name in it and the name ".." for the
1024 				 * directory's parent.
1025 				 */
1026 				if ((ndd = vp->v_cache_dd) != NULL) {
1027 					if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
1028 						cache_zap(ndd);
1029 					else
1030 						ndd = NULL;
1031 				}
1032 				vp->v_cache_dd = ncp;
1033 			}
1034 		} else {
1035 			vp->v_cache_dd = NULL;
1036 		}
1037 	}
1038 
1039 	if (flag != NCF_ISDOTDOT) {
1040 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1041 			vhold(dvp);
1042 			numcachehv++;
1043 		}
1044 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
1045 	}
1046 
1047 	bucketlock = HASH2BUCKETLOCK(hash);
1048 	rw_wlock(bucketlock);
1049 
1050 	/*
1051 	 * Insert the new namecache entry into the appropriate chain
1052 	 * within the cache entries table.
1053 	 */
1054 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
1055 
1056 	/*
1057 	 * If the entry is "negative", we place it into the
1058 	 * "negative" cache queue, otherwise, we place it into the
1059 	 * destination vnode's cache entries queue.
1060 	 */
1061 	if (vp != NULL) {
1062 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
1063 		SDT_PROBE3(vfs, namecache, enter, done, dvp, nc_get_name(ncp),
1064 		    vp);
1065 	} else {
1066 		if (cnp->cn_flags & ISWHITEOUT)
1067 			ncp->nc_flag |= NCF_WHITE;
1068 		cache_negative_insert(ncp);
1069 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
1070 		    nc_get_name(ncp));
1071 	}
1072 	rw_wunlock(bucketlock);
1073 	if (numneg * ncnegfactor > numcache)
1074 		nneg = cache_negative_zap_one();
1075 	CACHE_WUNLOCK();
1076 	cache_free(ndd);
1077 	cache_free(nneg);
1078 }
1079 
1080 static u_int
1081 cache_roundup_2(u_int val)
1082 {
1083 	u_int res;
1084 
1085 	for (res = 1; res <= val; res <<= 1)
1086 		continue;
1087 
1088 	return (res);
1089 }
1090 
1091 /*
1092  * Name cache initialization, from vfs_init() when we are booting
1093  */
1094 static void
1095 nchinit(void *dummy __unused)
1096 {
1097 	u_int i;
1098 
1099 	TAILQ_INIT(&ncneg);
1100 
1101 	cache_zone_small = uma_zcreate("S VFS Cache",
1102 	    sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1,
1103 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
1104 	cache_zone_small_ts = uma_zcreate("STS VFS Cache",
1105 	    sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1,
1106 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
1107 	cache_zone_large = uma_zcreate("L VFS Cache",
1108 	    sizeof(struct namecache) + NAME_MAX + 1,
1109 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
1110 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache",
1111 	    sizeof(struct namecache_ts) + NAME_MAX + 1,
1112 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
1113 
1114 	nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash);
1115 	numbucketlocks = cache_roundup_2(mp_ncpus * 16);
1116 	if (numbucketlocks > nchash)
1117 		numbucketlocks = nchash;
1118 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
1119 	    M_WAITOK | M_ZERO);
1120 	for (i = 0; i < numbucketlocks; i++)
1121 		rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK);
1122 
1123 	numcalls = counter_u64_alloc(M_WAITOK);
1124 	dothits = counter_u64_alloc(M_WAITOK);
1125 	dotdothits = counter_u64_alloc(M_WAITOK);
1126 	numchecks = counter_u64_alloc(M_WAITOK);
1127 	nummiss = counter_u64_alloc(M_WAITOK);
1128 	nummisszap = counter_u64_alloc(M_WAITOK);
1129 	numposzaps = counter_u64_alloc(M_WAITOK);
1130 	numposhits = counter_u64_alloc(M_WAITOK);
1131 	numnegzaps = counter_u64_alloc(M_WAITOK);
1132 	numneghits = counter_u64_alloc(M_WAITOK);
1133 	numfullpathcalls = counter_u64_alloc(M_WAITOK);
1134 	numfullpathfail1 = counter_u64_alloc(M_WAITOK);
1135 	numfullpathfail2 = counter_u64_alloc(M_WAITOK);
1136 	numfullpathfail4 = counter_u64_alloc(M_WAITOK);
1137 	numfullpathfound = counter_u64_alloc(M_WAITOK);
1138 }
1139 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
1140 
1141 void
1142 cache_changesize(int newmaxvnodes)
1143 {
1144 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
1145 	u_long new_nchash, old_nchash;
1146 	struct namecache *ncp;
1147 	uint32_t hash;
1148 	int i;
1149 
1150 	new_nchashtbl = hashinit(newmaxvnodes * 2, M_VFSCACHE, &new_nchash);
1151 	/* If same hash table size, nothing to do */
1152 	if (nchash == new_nchash) {
1153 		free(new_nchashtbl, M_VFSCACHE);
1154 		return;
1155 	}
1156 	/*
1157 	 * Move everything from the old hash table to the new table.
1158 	 * None of the namecache entries in the table can be removed
1159 	 * because to do so, they have to be removed from the hash table.
1160 	 */
1161 	CACHE_WLOCK();
1162 	cache_lock_all_buckets();
1163 	old_nchashtbl = nchashtbl;
1164 	old_nchash = nchash;
1165 	nchashtbl = new_nchashtbl;
1166 	nchash = new_nchash;
1167 	for (i = 0; i <= old_nchash; i++) {
1168 		while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) {
1169 			hash = cache_get_hash(nc_get_name(ncp), ncp->nc_nlen,
1170 			    ncp->nc_dvp);
1171 			LIST_REMOVE(ncp, nc_hash);
1172 			LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
1173 		}
1174 	}
1175 	cache_unlock_all_buckets();
1176 	CACHE_WUNLOCK();
1177 	free(old_nchashtbl, M_VFSCACHE);
1178 }
1179 
1180 /*
1181  * Invalidate all entries to a particular vnode.
1182  */
1183 void
1184 cache_purge(struct vnode *vp)
1185 {
1186 	TAILQ_HEAD(, namecache) ncps;
1187 	struct namecache *ncp, *nnp;
1188 
1189 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
1190 	SDT_PROBE1(vfs, namecache, purge, done, vp);
1191 	TAILQ_INIT(&ncps);
1192 	CACHE_WLOCK();
1193 	while (!LIST_EMPTY(&vp->v_cache_src)) {
1194 		ncp = LIST_FIRST(&vp->v_cache_src);
1195 		cache_zap(ncp);
1196 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1197 	}
1198 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
1199 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
1200 		cache_zap(ncp);
1201 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1202 	}
1203 	if (vp->v_cache_dd != NULL) {
1204 		ncp = vp->v_cache_dd;
1205 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
1206 		   ("lost dotdot link"));
1207 		cache_zap(ncp);
1208 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1209 	}
1210 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
1211 	CACHE_WUNLOCK();
1212 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
1213 		cache_free(ncp);
1214 	}
1215 }
1216 
1217 /*
1218  * Invalidate all negative entries for a particular directory vnode.
1219  */
1220 void
1221 cache_purge_negative(struct vnode *vp)
1222 {
1223 	TAILQ_HEAD(, namecache) ncps;
1224 	struct namecache *ncp, *nnp;
1225 
1226 	CTR1(KTR_VFS, "cache_purge_negative(%p)", vp);
1227 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
1228 	TAILQ_INIT(&ncps);
1229 	CACHE_WLOCK();
1230 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
1231 		if (ncp->nc_vp != NULL)
1232 			continue;
1233 		cache_zap(ncp);
1234 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
1235 	}
1236 	CACHE_WUNLOCK();
1237 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
1238 		cache_free(ncp);
1239 	}
1240 }
1241 
1242 /*
1243  * Flush all entries referencing a particular filesystem.
1244  */
1245 void
1246 cache_purgevfs(struct mount *mp)
1247 {
1248 	TAILQ_HEAD(, namecache) ncps;
1249 	struct rwlock *bucketlock;
1250 	struct nchashhead *bucket;
1251 	struct namecache *ncp, *nnp;
1252 	u_long i, j, n_nchash;
1253 
1254 	/* Scan hash tables for applicable entries */
1255 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
1256 	TAILQ_INIT(&ncps);
1257 	CACHE_WLOCK();
1258 	n_nchash = nchash + 1;
1259 	for (i = 0; i < numbucketlocks; i++) {
1260 		bucketlock = (struct rwlock *)&bucketlocks[i];
1261 		rw_wlock(bucketlock);
1262 		for (j = i; j < n_nchash; j += numbucketlocks) {
1263 			bucket = &nchashtbl[j];
1264 			LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) {
1265 				cache_assert_bucket_locked(ncp, RA_WLOCKED);
1266 				if (ncp->nc_dvp->v_mount != mp)
1267 					continue;
1268 				cache_zap_locked(ncp);
1269 				TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst);
1270 			}
1271 		}
1272 		rw_wunlock(bucketlock);
1273 	}
1274 	CACHE_WUNLOCK();
1275 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
1276 		cache_free(ncp);
1277 	}
1278 }
1279 
1280 /*
1281  * Perform canonical checks and cache lookup and pass on to filesystem
1282  * through the vop_cachedlookup only if needed.
1283  */
1284 
1285 int
1286 vfs_cache_lookup(struct vop_lookup_args *ap)
1287 {
1288 	struct vnode *dvp;
1289 	int error;
1290 	struct vnode **vpp = ap->a_vpp;
1291 	struct componentname *cnp = ap->a_cnp;
1292 	struct ucred *cred = cnp->cn_cred;
1293 	int flags = cnp->cn_flags;
1294 	struct thread *td = cnp->cn_thread;
1295 
1296 	*vpp = NULL;
1297 	dvp = ap->a_dvp;
1298 
1299 	if (dvp->v_type != VDIR)
1300 		return (ENOTDIR);
1301 
1302 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
1303 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
1304 		return (EROFS);
1305 
1306 	error = VOP_ACCESS(dvp, VEXEC, cred, td);
1307 	if (error)
1308 		return (error);
1309 
1310 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
1311 	if (error == 0)
1312 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
1313 	if (error == -1)
1314 		return (0);
1315 	return (error);
1316 }
1317 
1318 /*
1319  * XXX All of these sysctls would probably be more productive dead.
1320  */
1321 static int disablecwd;
1322 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
1323    "Disable the getcwd syscall");
1324 
1325 /* Implementation of the getcwd syscall. */
1326 int
1327 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
1328 {
1329 
1330 	return (kern___getcwd(td, uap->buf, UIO_USERSPACE, uap->buflen,
1331 	    MAXPATHLEN));
1332 }
1333 
1334 int
1335 kern___getcwd(struct thread *td, char *buf, enum uio_seg bufseg, u_int buflen,
1336     u_int path_max)
1337 {
1338 	char *bp, *tmpbuf;
1339 	struct filedesc *fdp;
1340 	struct vnode *cdir, *rdir;
1341 	int error;
1342 
1343 	if (disablecwd)
1344 		return (ENODEV);
1345 	if (buflen < 2)
1346 		return (EINVAL);
1347 	if (buflen > path_max)
1348 		buflen = path_max;
1349 
1350 	tmpbuf = malloc(buflen, M_TEMP, M_WAITOK);
1351 	fdp = td->td_proc->p_fd;
1352 	FILEDESC_SLOCK(fdp);
1353 	cdir = fdp->fd_cdir;
1354 	VREF(cdir);
1355 	rdir = fdp->fd_rdir;
1356 	VREF(rdir);
1357 	FILEDESC_SUNLOCK(fdp);
1358 	error = vn_fullpath1(td, cdir, rdir, tmpbuf, &bp, buflen);
1359 	vrele(rdir);
1360 	vrele(cdir);
1361 
1362 	if (!error) {
1363 		if (bufseg == UIO_SYSSPACE)
1364 			bcopy(bp, buf, strlen(bp) + 1);
1365 		else
1366 			error = copyout(bp, buf, strlen(bp) + 1);
1367 #ifdef KTRACE
1368 	if (KTRPOINT(curthread, KTR_NAMEI))
1369 		ktrnamei(bp);
1370 #endif
1371 	}
1372 	free(tmpbuf, M_TEMP);
1373 	return (error);
1374 }
1375 
1376 /*
1377  * Thus begins the fullpath magic.
1378  */
1379 
1380 static int disablefullpath;
1381 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, &disablefullpath, 0,
1382     "Disable the vn_fullpath function");
1383 
1384 /*
1385  * Retrieve the full filesystem path that correspond to a vnode from the name
1386  * cache (if available)
1387  */
1388 int
1389 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf)
1390 {
1391 	char *buf;
1392 	struct filedesc *fdp;
1393 	struct vnode *rdir;
1394 	int error;
1395 
1396 	if (disablefullpath)
1397 		return (ENODEV);
1398 	if (vn == NULL)
1399 		return (EINVAL);
1400 
1401 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1402 	fdp = td->td_proc->p_fd;
1403 	FILEDESC_SLOCK(fdp);
1404 	rdir = fdp->fd_rdir;
1405 	VREF(rdir);
1406 	FILEDESC_SUNLOCK(fdp);
1407 	error = vn_fullpath1(td, vn, rdir, buf, retbuf, MAXPATHLEN);
1408 	vrele(rdir);
1409 
1410 	if (!error)
1411 		*freebuf = buf;
1412 	else
1413 		free(buf, M_TEMP);
1414 	return (error);
1415 }
1416 
1417 /*
1418  * This function is similar to vn_fullpath, but it attempts to lookup the
1419  * pathname relative to the global root mount point.  This is required for the
1420  * auditing sub-system, as audited pathnames must be absolute, relative to the
1421  * global root mount point.
1422  */
1423 int
1424 vn_fullpath_global(struct thread *td, struct vnode *vn,
1425     char **retbuf, char **freebuf)
1426 {
1427 	char *buf;
1428 	int error;
1429 
1430 	if (disablefullpath)
1431 		return (ENODEV);
1432 	if (vn == NULL)
1433 		return (EINVAL);
1434 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
1435 	error = vn_fullpath1(td, vn, rootvnode, buf, retbuf, MAXPATHLEN);
1436 	if (!error)
1437 		*freebuf = buf;
1438 	else
1439 		free(buf, M_TEMP);
1440 	return (error);
1441 }
1442 
1443 int
1444 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen)
1445 {
1446 	int error;
1447 
1448 	CACHE_RLOCK();
1449 	error = vn_vptocnp_locked(vp, cred, buf, buflen);
1450 	if (error == 0)
1451 		CACHE_RUNLOCK();
1452 	return (error);
1453 }
1454 
1455 static int
1456 vn_vptocnp_locked(struct vnode **vp, struct ucred *cred, char *buf,
1457     u_int *buflen)
1458 {
1459 	struct vnode *dvp;
1460 	struct namecache *ncp;
1461 	int error;
1462 
1463 	TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) {
1464 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1465 			break;
1466 	}
1467 	if (ncp != NULL) {
1468 		if (*buflen < ncp->nc_nlen) {
1469 			CACHE_RUNLOCK();
1470 			vrele(*vp);
1471 			counter_u64_add(numfullpathfail4, 1);
1472 			error = ENOMEM;
1473 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
1474 			    vp, NULL);
1475 			return (error);
1476 		}
1477 		*buflen -= ncp->nc_nlen;
1478 		memcpy(buf + *buflen, nc_get_name(ncp), ncp->nc_nlen);
1479 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
1480 		    nc_get_name(ncp), vp);
1481 		dvp = *vp;
1482 		*vp = ncp->nc_dvp;
1483 		vref(*vp);
1484 		CACHE_RUNLOCK();
1485 		vrele(dvp);
1486 		CACHE_RLOCK();
1487 		return (0);
1488 	}
1489 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
1490 
1491 	CACHE_RUNLOCK();
1492 	vn_lock(*vp, LK_SHARED | LK_RETRY);
1493 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
1494 	vput(*vp);
1495 	if (error) {
1496 		counter_u64_add(numfullpathfail2, 1);
1497 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
1498 		return (error);
1499 	}
1500 
1501 	*vp = dvp;
1502 	CACHE_RLOCK();
1503 	if (dvp->v_iflag & VI_DOOMED) {
1504 		/* forced unmount */
1505 		CACHE_RUNLOCK();
1506 		vrele(dvp);
1507 		error = ENOENT;
1508 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
1509 		return (error);
1510 	}
1511 	/*
1512 	 * *vp has its use count incremented still.
1513 	 */
1514 
1515 	return (0);
1516 }
1517 
1518 /*
1519  * The magic behind kern___getcwd() and vn_fullpath().
1520  */
1521 static int
1522 vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
1523     char *buf, char **retbuf, u_int buflen)
1524 {
1525 	int error, slash_prefixed;
1526 #ifdef KDTRACE_HOOKS
1527 	struct vnode *startvp = vp;
1528 #endif
1529 	struct vnode *vp1;
1530 
1531 	buflen--;
1532 	buf[buflen] = '\0';
1533 	error = 0;
1534 	slash_prefixed = 0;
1535 
1536 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
1537 	counter_u64_add(numfullpathcalls, 1);
1538 	vref(vp);
1539 	CACHE_RLOCK();
1540 	if (vp->v_type != VDIR) {
1541 		error = vn_vptocnp_locked(&vp, td->td_ucred, buf, &buflen);
1542 		if (error)
1543 			return (error);
1544 		if (buflen == 0) {
1545 			CACHE_RUNLOCK();
1546 			vrele(vp);
1547 			return (ENOMEM);
1548 		}
1549 		buf[--buflen] = '/';
1550 		slash_prefixed = 1;
1551 	}
1552 	while (vp != rdir && vp != rootvnode) {
1553 		if (vp->v_vflag & VV_ROOT) {
1554 			if (vp->v_iflag & VI_DOOMED) {	/* forced unmount */
1555 				CACHE_RUNLOCK();
1556 				vrele(vp);
1557 				error = ENOENT;
1558 				SDT_PROBE3(vfs, namecache, fullpath, return,
1559 				    error, vp, NULL);
1560 				break;
1561 			}
1562 			vp1 = vp->v_mount->mnt_vnodecovered;
1563 			vref(vp1);
1564 			CACHE_RUNLOCK();
1565 			vrele(vp);
1566 			vp = vp1;
1567 			CACHE_RLOCK();
1568 			continue;
1569 		}
1570 		if (vp->v_type != VDIR) {
1571 			CACHE_RUNLOCK();
1572 			vrele(vp);
1573 			counter_u64_add(numfullpathfail1, 1);
1574 			error = ENOTDIR;
1575 			SDT_PROBE3(vfs, namecache, fullpath, return,
1576 			    error, vp, NULL);
1577 			break;
1578 		}
1579 		error = vn_vptocnp_locked(&vp, td->td_ucred, buf, &buflen);
1580 		if (error)
1581 			break;
1582 		if (buflen == 0) {
1583 			CACHE_RUNLOCK();
1584 			vrele(vp);
1585 			error = ENOMEM;
1586 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
1587 			    startvp, NULL);
1588 			break;
1589 		}
1590 		buf[--buflen] = '/';
1591 		slash_prefixed = 1;
1592 	}
1593 	if (error)
1594 		return (error);
1595 	if (!slash_prefixed) {
1596 		if (buflen == 0) {
1597 			CACHE_RUNLOCK();
1598 			vrele(vp);
1599 			counter_u64_add(numfullpathfail4, 1);
1600 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
1601 			    startvp, NULL);
1602 			return (ENOMEM);
1603 		}
1604 		buf[--buflen] = '/';
1605 	}
1606 	counter_u64_add(numfullpathfound, 1);
1607 	CACHE_RUNLOCK();
1608 	vrele(vp);
1609 
1610 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, buf + buflen);
1611 	*retbuf = buf + buflen;
1612 	return (0);
1613 }
1614 
1615 struct vnode *
1616 vn_dir_dd_ino(struct vnode *vp)
1617 {
1618 	struct namecache *ncp;
1619 	struct vnode *ddvp;
1620 
1621 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
1622 	CACHE_RLOCK();
1623 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
1624 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
1625 			continue;
1626 		ddvp = ncp->nc_dvp;
1627 		vhold(ddvp);
1628 		CACHE_RUNLOCK();
1629 		if (vget(ddvp, LK_SHARED | LK_NOWAIT | LK_VNHELD, curthread))
1630 			return (NULL);
1631 		return (ddvp);
1632 	}
1633 	CACHE_RUNLOCK();
1634 	return (NULL);
1635 }
1636 
1637 int
1638 vn_commname(struct vnode *vp, char *buf, u_int buflen)
1639 {
1640 	struct namecache *ncp;
1641 	int l;
1642 
1643 	CACHE_RLOCK();
1644 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
1645 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1646 			break;
1647 	if (ncp == NULL) {
1648 		CACHE_RUNLOCK();
1649 		return (ENOENT);
1650 	}
1651 	l = min(ncp->nc_nlen, buflen - 1);
1652 	memcpy(buf, nc_get_name(ncp), l);
1653 	CACHE_RUNLOCK();
1654 	buf[l] = '\0';
1655 	return (0);
1656 }
1657 
1658 /* ABI compat shims for old kernel modules. */
1659 #undef cache_enter
1660 
1661 void	cache_enter(struct vnode *dvp, struct vnode *vp,
1662 	    struct componentname *cnp);
1663 
1664 void
1665 cache_enter(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
1666 {
1667 
1668 	cache_enter_time(dvp, vp, cnp, NULL, NULL);
1669 }
1670 
1671 /*
1672  * This function updates path string to vnode's full global path
1673  * and checks the size of the new path string against the pathlen argument.
1674  *
1675  * Requires a locked, referenced vnode.
1676  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
1677  *
1678  * If sysctl debug.disablefullpath is set, ENODEV is returned,
1679  * vnode is left locked and path remain untouched.
1680  *
1681  * If vp is a directory, the call to vn_fullpath_global() always succeeds
1682  * because it falls back to the ".." lookup if the namecache lookup fails.
1683  */
1684 int
1685 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
1686     u_int pathlen)
1687 {
1688 	struct nameidata nd;
1689 	struct vnode *vp1;
1690 	char *rpath, *fbuf;
1691 	int error;
1692 
1693 	ASSERT_VOP_ELOCKED(vp, __func__);
1694 
1695 	/* Return ENODEV if sysctl debug.disablefullpath==1 */
1696 	if (disablefullpath)
1697 		return (ENODEV);
1698 
1699 	/* Construct global filesystem path from vp. */
1700 	VOP_UNLOCK(vp, 0);
1701 	error = vn_fullpath_global(td, vp, &rpath, &fbuf);
1702 
1703 	if (error != 0) {
1704 		vrele(vp);
1705 		return (error);
1706 	}
1707 
1708 	if (strlen(rpath) >= pathlen) {
1709 		vrele(vp);
1710 		error = ENAMETOOLONG;
1711 		goto out;
1712 	}
1713 
1714 	/*
1715 	 * Re-lookup the vnode by path to detect a possible rename.
1716 	 * As a side effect, the vnode is relocked.
1717 	 * If vnode was renamed, return ENOENT.
1718 	 */
1719 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
1720 	    UIO_SYSSPACE, path, td);
1721 	error = namei(&nd);
1722 	if (error != 0) {
1723 		vrele(vp);
1724 		goto out;
1725 	}
1726 	NDFREE(&nd, NDF_ONLY_PNBUF);
1727 	vp1 = nd.ni_vp;
1728 	vrele(vp);
1729 	if (vp1 == vp)
1730 		strcpy(path, rpath);
1731 	else {
1732 		vput(vp1);
1733 		error = ENOENT;
1734 	}
1735 
1736 out:
1737 	free(fbuf, M_TEMP);
1738 	return (error);
1739 }
1740