xref: /freebsd/sys/kern/vfs_cache.c (revision 4b50c451720d8b427757a6da1dd2bb4c52cd9e35)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993, 1995
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * Poul-Henning Kamp of the FreeBSD Project.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)vfs_cache.c	8.5 (Berkeley) 3/22/95
35  */
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include "opt_ddb.h"
41 #include "opt_ktrace.h"
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/counter.h>
46 #include <sys/filedesc.h>
47 #include <sys/fnv_hash.h>
48 #include <sys/kernel.h>
49 #include <sys/ktr.h>
50 #include <sys/lock.h>
51 #include <sys/malloc.h>
52 #include <sys/fcntl.h>
53 #include <sys/mount.h>
54 #include <sys/namei.h>
55 #include <sys/proc.h>
56 #include <sys/rwlock.h>
57 #include <sys/sdt.h>
58 #include <sys/smp.h>
59 #include <sys/syscallsubr.h>
60 #include <sys/sysctl.h>
61 #include <sys/sysproto.h>
62 #include <sys/vnode.h>
63 #ifdef KTRACE
64 #include <sys/ktrace.h>
65 #endif
66 
67 #ifdef DDB
68 #include <ddb/ddb.h>
69 #endif
70 
71 #include <vm/uma.h>
72 
73 SDT_PROVIDER_DECLARE(vfs);
74 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
75     "struct vnode *");
76 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
77     "char *");
78 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
79 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
80     "char *", "struct vnode *");
81 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
82 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
83     "struct vnode *", "char *");
84 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
85     "struct vnode *");
86 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
87     "struct vnode *", "char *");
88 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
89     "char *");
90 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *");
91 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
92 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
93 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
94     "struct vnode *");
95 SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *",
96     "char *");
97 SDT_PROBE_DEFINE2(vfs, namecache, shrink_negative, done, "struct vnode *",
98     "char *");
99 
100 /*
101  * This structure describes the elements in the cache of recent
102  * names looked up by namei.
103  */
104 
105 struct	namecache {
106 	LIST_ENTRY(namecache) nc_hash;	/* hash chain */
107 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
108 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
109 	struct	vnode *nc_dvp;		/* vnode of parent of name */
110 	union {
111 		struct	vnode *nu_vp;	/* vnode the name refers to */
112 	} n_un;
113 	u_char	nc_flag;		/* flag bits */
114 	u_char	nc_nlen;		/* length of name */
115 	char	nc_name[0];		/* segment name + nul */
116 };
117 
118 /*
119  * struct namecache_ts repeats struct namecache layout up to the
120  * nc_nlen member.
121  * struct namecache_ts is used in place of struct namecache when time(s) need
122  * to be stored.  The nc_dotdottime field is used when a cache entry is mapping
123  * both a non-dotdot directory name plus dotdot for the directory's
124  * parent.
125  */
126 struct	namecache_ts {
127 	struct	timespec nc_time;	/* timespec provided by fs */
128 	struct	timespec nc_dotdottime;	/* dotdot timespec provided by fs */
129 	int	nc_ticks;		/* ticks value when entry was added */
130 	struct namecache nc_nc;
131 };
132 
133 #define	nc_vp		n_un.nu_vp
134 
135 /*
136  * Flags in namecache.nc_flag
137  */
138 #define NCF_WHITE	0x01
139 #define NCF_ISDOTDOT	0x02
140 #define	NCF_TS		0x04
141 #define	NCF_DTS		0x08
142 #define	NCF_DVDROP	0x10
143 #define	NCF_NEGATIVE	0x20
144 #define	NCF_HOTNEGATIVE	0x40
145 
146 /*
147  * Name caching works as follows:
148  *
149  * Names found by directory scans are retained in a cache
150  * for future reference.  It is managed LRU, so frequently
151  * used names will hang around.  Cache is indexed by hash value
152  * obtained from (dvp, name) where dvp refers to the directory
153  * containing name.
154  *
155  * If it is a "negative" entry, (i.e. for a name that is known NOT to
156  * exist) the vnode pointer will be NULL.
157  *
158  * Upon reaching the last segment of a path, if the reference
159  * is for DELETE, or NOCACHE is set (rewrite), and the
160  * name is located in the cache, it will be dropped.
161  *
162  * These locks are used (in the order in which they can be taken):
163  * NAME		TYPE	ROLE
164  * vnodelock	mtx	vnode lists and v_cache_dd field protection
165  * bucketlock	rwlock	for access to given set of hash buckets
166  * neglist	mtx	negative entry LRU management
167  *
168  * Additionally, ncneg_shrink_lock mtx is used to have at most one thread
169  * shrinking the LRU list.
170  *
171  * It is legal to take multiple vnodelock and bucketlock locks. The locking
172  * order is lower address first. Both are recursive.
173  *
174  * "." lookups are lockless.
175  *
176  * ".." and vnode -> name lookups require vnodelock.
177  *
178  * name -> vnode lookup requires the relevant bucketlock to be held for reading.
179  *
180  * Insertions and removals of entries require involved vnodes and bucketlocks
181  * to be write-locked to prevent other threads from seeing the entry.
182  *
183  * Some lookups result in removal of the found entry (e.g. getting rid of a
184  * negative entry with the intent to create a positive one), which poses a
185  * problem when multiple threads reach the state. Similarly, two different
186  * threads can purge two different vnodes and try to remove the same name.
187  *
188  * If the already held vnode lock is lower than the second required lock, we
189  * can just take the other lock. However, in the opposite case, this could
190  * deadlock. As such, this is resolved by trylocking and if that fails unlocking
191  * the first node, locking everything in order and revalidating the state.
192  */
193 
194 /*
195  * Structures associated with name caching.
196  */
197 #define NCHHASH(hash) \
198 	(&nchashtbl[(hash) & nchash])
199 static __read_mostly LIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */
200 static u_long __read_mostly	nchash;			/* size of hash table */
201 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
202     "Size of namecache hash table");
203 static u_long __read_mostly	ncnegfactor = 5; /* ratio of negative entries */
204 SYSCTL_ULONG(_vfs, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
205     "Ratio of negative namecache entries");
206 static u_long __exclusive_cache_line	numneg;	/* number of negative entries allocated */
207 static u_long __exclusive_cache_line	numcache;/* number of cache entries allocated */
208 static u_long __exclusive_cache_line	numcachehv;/* number of cache entries with vnodes held */
209 u_int ncsizefactor = 2;
210 SYSCTL_UINT(_vfs, OID_AUTO, ncsizefactor, CTLFLAG_RW, &ncsizefactor, 0,
211     "Size factor for namecache");
212 static u_int __read_mostly	ncpurgeminvnodes;
213 SYSCTL_UINT(_vfs, OID_AUTO, ncpurgeminvnodes, CTLFLAG_RW, &ncpurgeminvnodes, 0,
214     "Number of vnodes below which purgevfs ignores the request");
215 static u_int __read_mostly	ncsize; /* the size as computed on creation or resizing */
216 
217 struct nchstats	nchstats;		/* cache effectiveness statistics */
218 
219 static struct mtx __exclusive_cache_line	ncneg_shrink_lock;
220 static int	shrink_list_turn;
221 
222 struct neglist {
223 	struct mtx		nl_lock;
224 	TAILQ_HEAD(, namecache) nl_list;
225 } __aligned(CACHE_LINE_SIZE);
226 
227 static struct neglist __read_mostly	*neglists;
228 static struct neglist ncneg_hot;
229 static u_long numhotneg;
230 
231 #define	numneglists (ncneghash + 1)
232 static u_int __read_mostly	ncneghash;
233 static inline struct neglist *
234 NCP2NEGLIST(struct namecache *ncp)
235 {
236 
237 	return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
238 }
239 
240 #define	numbucketlocks (ncbuckethash + 1)
241 static u_int __read_mostly  ncbuckethash;
242 static struct rwlock_padalign __read_mostly  *bucketlocks;
243 #define	HASH2BUCKETLOCK(hash) \
244 	((struct rwlock *)(&bucketlocks[((hash) & ncbuckethash)]))
245 
246 #define	numvnodelocks (ncvnodehash + 1)
247 static u_int __read_mostly  ncvnodehash;
248 static struct mtx __read_mostly *vnodelocks;
249 static inline struct mtx *
250 VP2VNODELOCK(struct vnode *vp)
251 {
252 
253 	return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
254 }
255 
256 /*
257  * UMA zones for the VFS cache.
258  *
259  * The small cache is used for entries with short names, which are the
260  * most common.  The large cache is used for entries which are too big to
261  * fit in the small cache.
262  */
263 static uma_zone_t __read_mostly cache_zone_small;
264 static uma_zone_t __read_mostly cache_zone_small_ts;
265 static uma_zone_t __read_mostly cache_zone_large;
266 static uma_zone_t __read_mostly cache_zone_large_ts;
267 
268 #define	CACHE_PATH_CUTOFF	35
269 
270 static struct namecache *
271 cache_alloc(int len, int ts)
272 {
273 	struct namecache_ts *ncp_ts;
274 	struct namecache *ncp;
275 
276 	if (__predict_false(ts)) {
277 		if (len <= CACHE_PATH_CUTOFF)
278 			ncp_ts = uma_zalloc(cache_zone_small_ts, M_WAITOK);
279 		else
280 			ncp_ts = uma_zalloc(cache_zone_large_ts, M_WAITOK);
281 		ncp = &ncp_ts->nc_nc;
282 	} else {
283 		if (len <= CACHE_PATH_CUTOFF)
284 			ncp = uma_zalloc(cache_zone_small, M_WAITOK);
285 		else
286 			ncp = uma_zalloc(cache_zone_large, M_WAITOK);
287 	}
288 	return (ncp);
289 }
290 
291 static void
292 cache_free(struct namecache *ncp)
293 {
294 	struct namecache_ts *ncp_ts;
295 
296 	if (ncp == NULL)
297 		return;
298 	if ((ncp->nc_flag & NCF_DVDROP) != 0)
299 		vdrop(ncp->nc_dvp);
300 	if (__predict_false(ncp->nc_flag & NCF_TS)) {
301 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
302 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
303 			uma_zfree(cache_zone_small_ts, ncp_ts);
304 		else
305 			uma_zfree(cache_zone_large_ts, ncp_ts);
306 	} else {
307 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
308 			uma_zfree(cache_zone_small, ncp);
309 		else
310 			uma_zfree(cache_zone_large, ncp);
311 	}
312 }
313 
314 static void
315 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
316 {
317 	struct namecache_ts *ncp_ts;
318 
319 	KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
320 	    (tsp == NULL && ticksp == NULL),
321 	    ("No NCF_TS"));
322 
323 	if (tsp == NULL && ticksp == NULL)
324 		return;
325 
326 	ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
327 	if (tsp != NULL)
328 		*tsp = ncp_ts->nc_time;
329 	if (ticksp != NULL)
330 		*ticksp = ncp_ts->nc_ticks;
331 }
332 
333 #ifdef DEBUG_CACHE
334 static int __read_mostly	doingcache = 1;	/* 1 => enable the cache */
335 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
336     "VFS namecache enabled");
337 #endif
338 
339 /* Export size information to userland */
340 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
341     sizeof(struct namecache), "sizeof(struct namecache)");
342 
343 /*
344  * The new name cache statistics
345  */
346 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0,
347     "Name cache statistics");
348 #define STATNODE_ULONG(name, descr)	\
349 	SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, descr);
350 #define STATNODE_COUNTER(name, descr)	\
351 	static counter_u64_t __read_mostly name; \
352 	SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, descr);
353 STATNODE_ULONG(numneg, "Number of negative cache entries");
354 STATNODE_ULONG(numcache, "Number of cache entries");
355 STATNODE_ULONG(numcachehv, "Number of namecache entries with vnodes held");
356 STATNODE_COUNTER(numcalls, "Number of cache lookups");
357 STATNODE_COUNTER(dothits, "Number of '.' hits");
358 STATNODE_COUNTER(dotdothits, "Number of '..' hits");
359 STATNODE_COUNTER(numchecks, "Number of checks in lookup");
360 STATNODE_COUNTER(nummiss, "Number of cache misses");
361 STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache");
362 STATNODE_COUNTER(numposzaps,
363     "Number of cache hits (positive) we do not want to cache");
364 STATNODE_COUNTER(numposhits, "Number of cache hits (positive)");
365 STATNODE_COUNTER(numnegzaps,
366     "Number of cache hits (negative) we do not want to cache");
367 STATNODE_COUNTER(numneghits, "Number of cache hits (negative)");
368 /* These count for kern___getcwd(), too. */
369 STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls");
370 STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
371 STATNODE_COUNTER(numfullpathfail2,
372     "Number of fullpath search errors (VOP_VPTOCNP failures)");
373 STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
374 STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls");
375 STATNODE_COUNTER(zap_and_exit_bucket_relock_success,
376     "Number of successful removals after relocking");
377 static long zap_and_exit_bucket_fail; STATNODE_ULONG(zap_and_exit_bucket_fail,
378     "Number of times zap_and_exit failed to lock");
379 static long zap_and_exit_bucket_fail2; STATNODE_ULONG(zap_and_exit_bucket_fail2,
380     "Number of times zap_and_exit failed to lock");
381 static long cache_lock_vnodes_cel_3_failures;
382 STATNODE_ULONG(cache_lock_vnodes_cel_3_failures,
383     "Number of times 3-way vnode locking failed");
384 STATNODE_ULONG(numhotneg, "Number of hot negative entries");
385 STATNODE_COUNTER(numneg_evicted,
386     "Number of negative entries evicted when adding a new entry");
387 STATNODE_COUNTER(shrinking_skipped,
388     "Number of times shrinking was already in progress");
389 
390 static void cache_zap_locked(struct namecache *ncp, bool neg_locked);
391 static int vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
392     char *buf, char **retbuf, u_int buflen);
393 
394 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
395 
396 static int cache_yield;
397 SYSCTL_INT(_vfs_cache, OID_AUTO, yield, CTLFLAG_RD, &cache_yield, 0,
398     "Number of times cache called yield");
399 
400 static void __noinline
401 cache_maybe_yield(void)
402 {
403 
404 	if (should_yield()) {
405 		cache_yield++;
406 		kern_yield(PRI_USER);
407 	}
408 }
409 
410 static inline void
411 cache_assert_vlp_locked(struct mtx *vlp)
412 {
413 
414 	if (vlp != NULL)
415 		mtx_assert(vlp, MA_OWNED);
416 }
417 
418 static inline void
419 cache_assert_vnode_locked(struct vnode *vp)
420 {
421 	struct mtx *vlp;
422 
423 	vlp = VP2VNODELOCK(vp);
424 	cache_assert_vlp_locked(vlp);
425 }
426 
427 static uint32_t
428 cache_get_hash(char *name, u_char len, struct vnode *dvp)
429 {
430 	uint32_t hash;
431 
432 	hash = fnv_32_buf(name, len, FNV1_32_INIT);
433 	hash = fnv_32_buf(&dvp, sizeof(dvp), hash);
434 	return (hash);
435 }
436 
437 static inline struct rwlock *
438 NCP2BUCKETLOCK(struct namecache *ncp)
439 {
440 	uint32_t hash;
441 
442 	hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
443 	return (HASH2BUCKETLOCK(hash));
444 }
445 
446 #ifdef INVARIANTS
447 static void
448 cache_assert_bucket_locked(struct namecache *ncp, int mode)
449 {
450 	struct rwlock *blp;
451 
452 	blp = NCP2BUCKETLOCK(ncp);
453 	rw_assert(blp, mode);
454 }
455 #else
456 #define cache_assert_bucket_locked(x, y) do { } while (0)
457 #endif
458 
459 #define cache_sort_vnodes(x, y)	_cache_sort_vnodes((void **)(x), (void **)(y))
460 static void
461 _cache_sort_vnodes(void **p1, void **p2)
462 {
463 	void *tmp;
464 
465 	MPASS(*p1 != NULL || *p2 != NULL);
466 
467 	if (*p1 > *p2) {
468 		tmp = *p2;
469 		*p2 = *p1;
470 		*p1 = tmp;
471 	}
472 }
473 
474 static void
475 cache_lock_all_buckets(void)
476 {
477 	u_int i;
478 
479 	for (i = 0; i < numbucketlocks; i++)
480 		rw_wlock(&bucketlocks[i]);
481 }
482 
483 static void
484 cache_unlock_all_buckets(void)
485 {
486 	u_int i;
487 
488 	for (i = 0; i < numbucketlocks; i++)
489 		rw_wunlock(&bucketlocks[i]);
490 }
491 
492 static void
493 cache_lock_all_vnodes(void)
494 {
495 	u_int i;
496 
497 	for (i = 0; i < numvnodelocks; i++)
498 		mtx_lock(&vnodelocks[i]);
499 }
500 
501 static void
502 cache_unlock_all_vnodes(void)
503 {
504 	u_int i;
505 
506 	for (i = 0; i < numvnodelocks; i++)
507 		mtx_unlock(&vnodelocks[i]);
508 }
509 
510 static int
511 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
512 {
513 
514 	cache_sort_vnodes(&vlp1, &vlp2);
515 
516 	if (vlp1 != NULL) {
517 		if (!mtx_trylock(vlp1))
518 			return (EAGAIN);
519 	}
520 	if (!mtx_trylock(vlp2)) {
521 		if (vlp1 != NULL)
522 			mtx_unlock(vlp1);
523 		return (EAGAIN);
524 	}
525 
526 	return (0);
527 }
528 
529 static void
530 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
531 {
532 
533 	MPASS(vlp1 != NULL || vlp2 != NULL);
534 	MPASS(vlp1 <= vlp2);
535 
536 	if (vlp1 != NULL)
537 		mtx_lock(vlp1);
538 	if (vlp2 != NULL)
539 		mtx_lock(vlp2);
540 }
541 
542 static void
543 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
544 {
545 
546 	MPASS(vlp1 != NULL || vlp2 != NULL);
547 
548 	if (vlp1 != NULL)
549 		mtx_unlock(vlp1);
550 	if (vlp2 != NULL)
551 		mtx_unlock(vlp2);
552 }
553 
554 static int
555 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
556 {
557 	struct nchstats snap;
558 
559 	if (req->oldptr == NULL)
560 		return (SYSCTL_OUT(req, 0, sizeof(snap)));
561 
562 	snap = nchstats;
563 	snap.ncs_goodhits = counter_u64_fetch(numposhits);
564 	snap.ncs_neghits = counter_u64_fetch(numneghits);
565 	snap.ncs_badhits = counter_u64_fetch(numposzaps) +
566 	    counter_u64_fetch(numnegzaps);
567 	snap.ncs_miss = counter_u64_fetch(nummisszap) +
568 	    counter_u64_fetch(nummiss);
569 
570 	return (SYSCTL_OUT(req, &snap, sizeof(snap)));
571 }
572 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
573     CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
574     "VFS cache effectiveness statistics");
575 
576 #ifdef DIAGNOSTIC
577 /*
578  * Grab an atomic snapshot of the name cache hash chain lengths
579  */
580 static SYSCTL_NODE(_debug, OID_AUTO, hashstat, CTLFLAG_RW, NULL,
581     "hash table stats");
582 
583 static int
584 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
585 {
586 	struct nchashhead *ncpp;
587 	struct namecache *ncp;
588 	int i, error, n_nchash, *cntbuf;
589 
590 retry:
591 	n_nchash = nchash + 1;	/* nchash is max index, not count */
592 	if (req->oldptr == NULL)
593 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
594 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
595 	cache_lock_all_buckets();
596 	if (n_nchash != nchash + 1) {
597 		cache_unlock_all_buckets();
598 		free(cntbuf, M_TEMP);
599 		goto retry;
600 	}
601 	/* Scan hash tables counting entries */
602 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
603 		LIST_FOREACH(ncp, ncpp, nc_hash)
604 			cntbuf[i]++;
605 	cache_unlock_all_buckets();
606 	for (error = 0, i = 0; i < n_nchash; i++)
607 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
608 			break;
609 	free(cntbuf, M_TEMP);
610 	return (error);
611 }
612 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
613     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
614     "nchash chain lengths");
615 
616 static int
617 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
618 {
619 	int error;
620 	struct nchashhead *ncpp;
621 	struct namecache *ncp;
622 	int n_nchash;
623 	int count, maxlength, used, pct;
624 
625 	if (!req->oldptr)
626 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
627 
628 	cache_lock_all_buckets();
629 	n_nchash = nchash + 1;	/* nchash is max index, not count */
630 	used = 0;
631 	maxlength = 0;
632 
633 	/* Scan hash tables for applicable entries */
634 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
635 		count = 0;
636 		LIST_FOREACH(ncp, ncpp, nc_hash) {
637 			count++;
638 		}
639 		if (count)
640 			used++;
641 		if (maxlength < count)
642 			maxlength = count;
643 	}
644 	n_nchash = nchash + 1;
645 	cache_unlock_all_buckets();
646 	pct = (used * 100) / (n_nchash / 100);
647 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
648 	if (error)
649 		return (error);
650 	error = SYSCTL_OUT(req, &used, sizeof(used));
651 	if (error)
652 		return (error);
653 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
654 	if (error)
655 		return (error);
656 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
657 	if (error)
658 		return (error);
659 	return (0);
660 }
661 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
662     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
663     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
664 #endif
665 
666 /*
667  * Negative entries management
668  *
669  * A variation of LRU scheme is used. New entries are hashed into one of
670  * numneglists cold lists. Entries get promoted to the hot list on first hit.
671  *
672  * The shrinker will demote hot list head and evict from the cold list in a
673  * round-robin manner.
674  */
675 static void
676 cache_negative_hit(struct namecache *ncp)
677 {
678 	struct neglist *neglist;
679 
680 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
681 	if (ncp->nc_flag & NCF_HOTNEGATIVE)
682 		return;
683 	neglist = NCP2NEGLIST(ncp);
684 	mtx_lock(&ncneg_hot.nl_lock);
685 	mtx_lock(&neglist->nl_lock);
686 	if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
687 		numhotneg++;
688 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
689 		TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst);
690 		ncp->nc_flag |= NCF_HOTNEGATIVE;
691 	}
692 	mtx_unlock(&neglist->nl_lock);
693 	mtx_unlock(&ncneg_hot.nl_lock);
694 }
695 
696 static void
697 cache_negative_insert(struct namecache *ncp, bool neg_locked)
698 {
699 	struct neglist *neglist;
700 
701 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
702 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
703 	neglist = NCP2NEGLIST(ncp);
704 	if (!neg_locked) {
705 		mtx_lock(&neglist->nl_lock);
706 	} else {
707 		mtx_assert(&neglist->nl_lock, MA_OWNED);
708 	}
709 	TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
710 	if (!neg_locked)
711 		mtx_unlock(&neglist->nl_lock);
712 	atomic_add_rel_long(&numneg, 1);
713 }
714 
715 static void
716 cache_negative_remove(struct namecache *ncp, bool neg_locked)
717 {
718 	struct neglist *neglist;
719 	bool hot_locked = false;
720 	bool list_locked = false;
721 
722 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
723 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
724 	neglist = NCP2NEGLIST(ncp);
725 	if (!neg_locked) {
726 		if (ncp->nc_flag & NCF_HOTNEGATIVE) {
727 			hot_locked = true;
728 			mtx_lock(&ncneg_hot.nl_lock);
729 			if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) {
730 				list_locked = true;
731 				mtx_lock(&neglist->nl_lock);
732 			}
733 		} else {
734 			list_locked = true;
735 			mtx_lock(&neglist->nl_lock);
736 		}
737 	}
738 	if (ncp->nc_flag & NCF_HOTNEGATIVE) {
739 		mtx_assert(&ncneg_hot.nl_lock, MA_OWNED);
740 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
741 		numhotneg--;
742 	} else {
743 		mtx_assert(&neglist->nl_lock, MA_OWNED);
744 		TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst);
745 	}
746 	if (list_locked)
747 		mtx_unlock(&neglist->nl_lock);
748 	if (hot_locked)
749 		mtx_unlock(&ncneg_hot.nl_lock);
750 	atomic_subtract_rel_long(&numneg, 1);
751 }
752 
753 static void
754 cache_negative_shrink_select(int start, struct namecache **ncpp,
755     struct neglist **neglistpp)
756 {
757 	struct neglist *neglist;
758 	struct namecache *ncp;
759 	int i;
760 
761 	*ncpp = ncp = NULL;
762 	neglist = NULL;
763 
764 	for (i = start; i < numneglists; i++) {
765 		neglist = &neglists[i];
766 		if (TAILQ_FIRST(&neglist->nl_list) == NULL)
767 			continue;
768 		mtx_lock(&neglist->nl_lock);
769 		ncp = TAILQ_FIRST(&neglist->nl_list);
770 		if (ncp != NULL)
771 			break;
772 		mtx_unlock(&neglist->nl_lock);
773 	}
774 
775 	*neglistpp = neglist;
776 	*ncpp = ncp;
777 }
778 
779 static void
780 cache_negative_zap_one(void)
781 {
782 	struct namecache *ncp, *ncp2;
783 	struct neglist *neglist;
784 	struct mtx *dvlp;
785 	struct rwlock *blp;
786 
787 	if (mtx_owner(&ncneg_shrink_lock) != NULL ||
788 	    !mtx_trylock(&ncneg_shrink_lock)) {
789 		counter_u64_add(shrinking_skipped, 1);
790 		return;
791 	}
792 
793 	mtx_lock(&ncneg_hot.nl_lock);
794 	ncp = TAILQ_FIRST(&ncneg_hot.nl_list);
795 	if (ncp != NULL) {
796 		neglist = NCP2NEGLIST(ncp);
797 		mtx_lock(&neglist->nl_lock);
798 		TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst);
799 		TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst);
800 		ncp->nc_flag &= ~NCF_HOTNEGATIVE;
801 		numhotneg--;
802 		mtx_unlock(&neglist->nl_lock);
803 	}
804 	mtx_unlock(&ncneg_hot.nl_lock);
805 
806 	cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
807 	shrink_list_turn++;
808 	if (shrink_list_turn == numneglists)
809 		shrink_list_turn = 0;
810 	if (ncp == NULL && shrink_list_turn == 0)
811 		cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist);
812 	mtx_unlock(&ncneg_shrink_lock);
813 	if (ncp == NULL)
814 		return;
815 
816 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
817 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
818 	blp = NCP2BUCKETLOCK(ncp);
819 	mtx_unlock(&neglist->nl_lock);
820 	mtx_lock(dvlp);
821 	rw_wlock(blp);
822 	mtx_lock(&neglist->nl_lock);
823 	ncp2 = TAILQ_FIRST(&neglist->nl_list);
824 	if (ncp != ncp2 || dvlp != VP2VNODELOCK(ncp2->nc_dvp) ||
825 	    blp != NCP2BUCKETLOCK(ncp2) || !(ncp2->nc_flag & NCF_NEGATIVE)) {
826 		ncp = NULL;
827 	} else {
828 		SDT_PROBE2(vfs, namecache, shrink_negative, done, ncp->nc_dvp,
829 		    ncp->nc_name);
830 
831 		cache_zap_locked(ncp, true);
832 		counter_u64_add(numneg_evicted, 1);
833 	}
834 	mtx_unlock(&neglist->nl_lock);
835 	rw_wunlock(blp);
836 	mtx_unlock(dvlp);
837 	cache_free(ncp);
838 }
839 
840 /*
841  * cache_zap_locked():
842  *
843  *   Removes a namecache entry from cache, whether it contains an actual
844  *   pointer to a vnode or if it is just a negative cache entry.
845  */
846 static void
847 cache_zap_locked(struct namecache *ncp, bool neg_locked)
848 {
849 
850 	if (!(ncp->nc_flag & NCF_NEGATIVE))
851 		cache_assert_vnode_locked(ncp->nc_vp);
852 	cache_assert_vnode_locked(ncp->nc_dvp);
853 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
854 
855 	CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp,
856 	    (ncp->nc_flag & NCF_NEGATIVE) ? NULL : ncp->nc_vp);
857 	LIST_REMOVE(ncp, nc_hash);
858 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
859 		SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
860 		    ncp->nc_name, ncp->nc_vp);
861 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
862 		if (ncp == ncp->nc_vp->v_cache_dd)
863 			ncp->nc_vp->v_cache_dd = NULL;
864 	} else {
865 		SDT_PROBE2(vfs, namecache, zap_negative, done, ncp->nc_dvp,
866 		    ncp->nc_name);
867 		cache_negative_remove(ncp, neg_locked);
868 	}
869 	if (ncp->nc_flag & NCF_ISDOTDOT) {
870 		if (ncp == ncp->nc_dvp->v_cache_dd)
871 			ncp->nc_dvp->v_cache_dd = NULL;
872 	} else {
873 		LIST_REMOVE(ncp, nc_src);
874 		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
875 			ncp->nc_flag |= NCF_DVDROP;
876 			atomic_subtract_rel_long(&numcachehv, 1);
877 		}
878 	}
879 	atomic_subtract_rel_long(&numcache, 1);
880 }
881 
882 static void
883 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
884 {
885 	struct rwlock *blp;
886 
887 	MPASS(ncp->nc_dvp == vp);
888 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
889 	cache_assert_vnode_locked(vp);
890 
891 	blp = NCP2BUCKETLOCK(ncp);
892 	rw_wlock(blp);
893 	cache_zap_locked(ncp, false);
894 	rw_wunlock(blp);
895 }
896 
897 static bool
898 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
899     struct mtx **vlpp)
900 {
901 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
902 	struct rwlock *blp;
903 
904 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
905 	cache_assert_vnode_locked(vp);
906 
907 	if (ncp->nc_flag & NCF_NEGATIVE) {
908 		if (*vlpp != NULL) {
909 			mtx_unlock(*vlpp);
910 			*vlpp = NULL;
911 		}
912 		cache_zap_negative_locked_vnode_kl(ncp, vp);
913 		return (true);
914 	}
915 
916 	pvlp = VP2VNODELOCK(vp);
917 	blp = NCP2BUCKETLOCK(ncp);
918 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
919 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
920 
921 	if (*vlpp == vlp1 || *vlpp == vlp2) {
922 		to_unlock = *vlpp;
923 		*vlpp = NULL;
924 	} else {
925 		if (*vlpp != NULL) {
926 			mtx_unlock(*vlpp);
927 			*vlpp = NULL;
928 		}
929 		cache_sort_vnodes(&vlp1, &vlp2);
930 		if (vlp1 == pvlp) {
931 			mtx_lock(vlp2);
932 			to_unlock = vlp2;
933 		} else {
934 			if (!mtx_trylock(vlp1))
935 				goto out_relock;
936 			to_unlock = vlp1;
937 		}
938 	}
939 	rw_wlock(blp);
940 	cache_zap_locked(ncp, false);
941 	rw_wunlock(blp);
942 	if (to_unlock != NULL)
943 		mtx_unlock(to_unlock);
944 	return (true);
945 
946 out_relock:
947 	mtx_unlock(vlp2);
948 	mtx_lock(vlp1);
949 	mtx_lock(vlp2);
950 	MPASS(*vlpp == NULL);
951 	*vlpp = vlp1;
952 	return (false);
953 }
954 
955 static int __noinline
956 cache_zap_locked_vnode(struct namecache *ncp, struct vnode *vp)
957 {
958 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
959 	struct rwlock *blp;
960 	int error = 0;
961 
962 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
963 	cache_assert_vnode_locked(vp);
964 
965 	pvlp = VP2VNODELOCK(vp);
966 	if (ncp->nc_flag & NCF_NEGATIVE) {
967 		cache_zap_negative_locked_vnode_kl(ncp, vp);
968 		goto out;
969 	}
970 
971 	blp = NCP2BUCKETLOCK(ncp);
972 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
973 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
974 	cache_sort_vnodes(&vlp1, &vlp2);
975 	if (vlp1 == pvlp) {
976 		mtx_lock(vlp2);
977 		to_unlock = vlp2;
978 	} else {
979 		if (!mtx_trylock(vlp1)) {
980 			error = EAGAIN;
981 			goto out;
982 		}
983 		to_unlock = vlp1;
984 	}
985 	rw_wlock(blp);
986 	cache_zap_locked(ncp, false);
987 	rw_wunlock(blp);
988 	mtx_unlock(to_unlock);
989 out:
990 	mtx_unlock(pvlp);
991 	return (error);
992 }
993 
994 /*
995  * If trylocking failed we can get here. We know enough to take all needed locks
996  * in the right order and re-lookup the entry.
997  */
998 static int
999 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1000     struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1001     struct rwlock *blp)
1002 {
1003 	struct namecache *rncp;
1004 
1005 	cache_assert_bucket_locked(ncp, RA_UNLOCKED);
1006 
1007 	cache_sort_vnodes(&dvlp, &vlp);
1008 	cache_lock_vnodes(dvlp, vlp);
1009 	rw_wlock(blp);
1010 	LIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1011 		if (rncp == ncp && rncp->nc_dvp == dvp &&
1012 		    rncp->nc_nlen == cnp->cn_namelen &&
1013 		    !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen))
1014 			break;
1015 	}
1016 	if (rncp != NULL) {
1017 		cache_zap_locked(rncp, false);
1018 		rw_wunlock(blp);
1019 		cache_unlock_vnodes(dvlp, vlp);
1020 		counter_u64_add(zap_and_exit_bucket_relock_success, 1);
1021 		return (0);
1022 	}
1023 
1024 	rw_wunlock(blp);
1025 	cache_unlock_vnodes(dvlp, vlp);
1026 	return (EAGAIN);
1027 }
1028 
1029 static int __noinline
1030 cache_zap_wlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1031     uint32_t hash, struct rwlock *blp)
1032 {
1033 	struct mtx *dvlp, *vlp;
1034 	struct vnode *dvp;
1035 
1036 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1037 
1038 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1039 	vlp = NULL;
1040 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1041 		vlp = VP2VNODELOCK(ncp->nc_vp);
1042 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1043 		cache_zap_locked(ncp, false);
1044 		rw_wunlock(blp);
1045 		cache_unlock_vnodes(dvlp, vlp);
1046 		return (0);
1047 	}
1048 
1049 	dvp = ncp->nc_dvp;
1050 	rw_wunlock(blp);
1051 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1052 }
1053 
1054 static int __noinline
1055 cache_zap_rlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1056     uint32_t hash, struct rwlock *blp)
1057 {
1058 	struct mtx *dvlp, *vlp;
1059 	struct vnode *dvp;
1060 
1061 	cache_assert_bucket_locked(ncp, RA_RLOCKED);
1062 
1063 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1064 	vlp = NULL;
1065 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1066 		vlp = VP2VNODELOCK(ncp->nc_vp);
1067 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1068 		rw_runlock(blp);
1069 		rw_wlock(blp);
1070 		cache_zap_locked(ncp, false);
1071 		rw_wunlock(blp);
1072 		cache_unlock_vnodes(dvlp, vlp);
1073 		return (0);
1074 	}
1075 
1076 	dvp = ncp->nc_dvp;
1077 	rw_runlock(blp);
1078 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1079 }
1080 
1081 static int
1082 cache_zap_wlocked_bucket_kl(struct namecache *ncp, struct rwlock *blp,
1083     struct mtx **vlpp1, struct mtx **vlpp2)
1084 {
1085 	struct mtx *dvlp, *vlp;
1086 
1087 	cache_assert_bucket_locked(ncp, RA_WLOCKED);
1088 
1089 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1090 	vlp = NULL;
1091 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1092 		vlp = VP2VNODELOCK(ncp->nc_vp);
1093 	cache_sort_vnodes(&dvlp, &vlp);
1094 
1095 	if (*vlpp1 == dvlp && *vlpp2 == vlp) {
1096 		cache_zap_locked(ncp, false);
1097 		cache_unlock_vnodes(dvlp, vlp);
1098 		*vlpp1 = NULL;
1099 		*vlpp2 = NULL;
1100 		return (0);
1101 	}
1102 
1103 	if (*vlpp1 != NULL)
1104 		mtx_unlock(*vlpp1);
1105 	if (*vlpp2 != NULL)
1106 		mtx_unlock(*vlpp2);
1107 	*vlpp1 = NULL;
1108 	*vlpp2 = NULL;
1109 
1110 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1111 		cache_zap_locked(ncp, false);
1112 		cache_unlock_vnodes(dvlp, vlp);
1113 		return (0);
1114 	}
1115 
1116 	rw_wunlock(blp);
1117 	*vlpp1 = dvlp;
1118 	*vlpp2 = vlp;
1119 	if (*vlpp1 != NULL)
1120 		mtx_lock(*vlpp1);
1121 	mtx_lock(*vlpp2);
1122 	rw_wlock(blp);
1123 	return (EAGAIN);
1124 }
1125 
1126 static void
1127 cache_lookup_unlock(struct rwlock *blp, struct mtx *vlp)
1128 {
1129 
1130 	if (blp != NULL) {
1131 		rw_runlock(blp);
1132 	} else {
1133 		mtx_unlock(vlp);
1134 	}
1135 }
1136 
1137 static int __noinline
1138 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1139     struct timespec *tsp, int *ticksp)
1140 {
1141 	int ltype;
1142 
1143 	*vpp = dvp;
1144 	CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .",
1145 			dvp, cnp->cn_nameptr);
1146 	counter_u64_add(dothits, 1);
1147 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1148 	if (tsp != NULL)
1149 		timespecclear(tsp);
1150 	if (ticksp != NULL)
1151 		*ticksp = ticks;
1152 	vrefact(*vpp);
1153 	/*
1154 	 * When we lookup "." we still can be asked to lock it
1155 	 * differently...
1156 	 */
1157 	ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1158 	if (ltype != VOP_ISLOCKED(*vpp)) {
1159 		if (ltype == LK_EXCLUSIVE) {
1160 			vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1161 			if (VN_IS_DOOMED((*vpp))) {
1162 				/* forced unmount */
1163 				vrele(*vpp);
1164 				*vpp = NULL;
1165 				return (ENOENT);
1166 			}
1167 		} else
1168 			vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1169 	}
1170 	return (-1);
1171 }
1172 
1173 static __noinline int
1174 cache_lookup_nomakeentry(struct vnode *dvp, struct vnode **vpp,
1175     struct componentname *cnp, struct timespec *tsp, int *ticksp)
1176 {
1177 	struct namecache *ncp;
1178 	struct rwlock *blp;
1179 	struct mtx *dvlp, *dvlp2;
1180 	uint32_t hash;
1181 	int error;
1182 
1183 	if (cnp->cn_namelen == 2 &&
1184 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1185 		counter_u64_add(dotdothits, 1);
1186 		dvlp = VP2VNODELOCK(dvp);
1187 		dvlp2 = NULL;
1188 		mtx_lock(dvlp);
1189 retry_dotdot:
1190 		ncp = dvp->v_cache_dd;
1191 		if (ncp == NULL) {
1192 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1193 			    "..", NULL);
1194 			mtx_unlock(dvlp);
1195 			if (dvlp2 != NULL)
1196 				mtx_unlock(dvlp2);
1197 			return (0);
1198 		}
1199 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1200 			if (ncp->nc_dvp != dvp)
1201 				panic("dvp %p v_cache_dd %p\n", dvp, ncp);
1202 			if (!cache_zap_locked_vnode_kl2(ncp,
1203 			    dvp, &dvlp2))
1204 				goto retry_dotdot;
1205 			MPASS(dvp->v_cache_dd == NULL);
1206 			mtx_unlock(dvlp);
1207 			if (dvlp2 != NULL)
1208 				mtx_unlock(dvlp2);
1209 			cache_free(ncp);
1210 		} else {
1211 			dvp->v_cache_dd = NULL;
1212 			mtx_unlock(dvlp);
1213 			if (dvlp2 != NULL)
1214 				mtx_unlock(dvlp2);
1215 		}
1216 		return (0);
1217 	}
1218 
1219 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1220 	blp = HASH2BUCKETLOCK(hash);
1221 retry:
1222 	if (LIST_EMPTY(NCHHASH(hash)))
1223 		goto out_no_entry;
1224 
1225 	rw_wlock(blp);
1226 
1227 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1228 		counter_u64_add(numchecks, 1);
1229 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1230 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1231 			break;
1232 	}
1233 
1234 	/* We failed to find an entry */
1235 	if (ncp == NULL) {
1236 		rw_wunlock(blp);
1237 		goto out_no_entry;
1238 	}
1239 
1240 	error = cache_zap_wlocked_bucket(ncp, cnp, hash, blp);
1241 	if (__predict_false(error != 0)) {
1242 		zap_and_exit_bucket_fail++;
1243 		cache_maybe_yield();
1244 		goto retry;
1245 	}
1246 	counter_u64_add(numposzaps, 1);
1247 	cache_free(ncp);
1248 	return (0);
1249 out_no_entry:
1250 	SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr, NULL);
1251 	counter_u64_add(nummisszap, 1);
1252 	return (0);
1253 }
1254 
1255 /**
1256  * Lookup a name in the name cache
1257  *
1258  * # Arguments
1259  *
1260  * - dvp:	Parent directory in which to search.
1261  * - vpp:	Return argument.  Will contain desired vnode on cache hit.
1262  * - cnp:	Parameters of the name search.  The most interesting bits of
1263  *   		the cn_flags field have the following meanings:
1264  *   	- MAKEENTRY:	If clear, free an entry from the cache rather than look
1265  *   			it up.
1266  *   	- ISDOTDOT:	Must be set if and only if cn_nameptr == ".."
1267  * - tsp:	Return storage for cache timestamp.  On a successful (positive
1268  *   		or negative) lookup, tsp will be filled with any timespec that
1269  *   		was stored when this cache entry was created.  However, it will
1270  *   		be clear for "." entries.
1271  * - ticks:	Return storage for alternate cache timestamp.  On a successful
1272  *   		(positive or negative) lookup, it will contain the ticks value
1273  *   		that was current when the cache entry was created, unless cnp
1274  *   		was ".".
1275  *
1276  * # Returns
1277  *
1278  * - -1:	A positive cache hit.  vpp will contain the desired vnode.
1279  * - ENOENT:	A negative cache hit, or dvp was recycled out from under us due
1280  *		to a forced unmount.  vpp will not be modified.  If the entry
1281  *		is a whiteout, then the ISWHITEOUT flag will be set in
1282  *		cnp->cn_flags.
1283  * - 0:		A cache miss.  vpp will not be modified.
1284  *
1285  * # Locking
1286  *
1287  * On a cache hit, vpp will be returned locked and ref'd.  If we're looking up
1288  * .., dvp is unlocked.  If we're looking up . an extra ref is taken, but the
1289  * lock is not recursively acquired.
1290  */
1291 int
1292 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1293     struct timespec *tsp, int *ticksp)
1294 {
1295 	struct namecache_ts *ncp_ts;
1296 	struct namecache *ncp;
1297 	struct rwlock *blp;
1298 	struct mtx *dvlp;
1299 	uint32_t hash;
1300 	enum vgetstate vs;
1301 	int error, ltype;
1302 
1303 #ifdef DEBUG_CACHE
1304 	if (__predict_false(!doingcache)) {
1305 		cnp->cn_flags &= ~MAKEENTRY;
1306 		return (0);
1307 	}
1308 #endif
1309 
1310 	counter_u64_add(numcalls, 1);
1311 
1312 	if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.'))
1313 		return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1314 
1315 	if ((cnp->cn_flags & MAKEENTRY) == 0)
1316 		return (cache_lookup_nomakeentry(dvp, vpp, cnp, tsp, ticksp));
1317 
1318 retry:
1319 	blp = NULL;
1320 	dvlp = NULL;
1321 	error = 0;
1322 	if (cnp->cn_namelen == 2 &&
1323 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1324 		counter_u64_add(dotdothits, 1);
1325 		dvlp = VP2VNODELOCK(dvp);
1326 		mtx_lock(dvlp);
1327 		ncp = dvp->v_cache_dd;
1328 		if (ncp == NULL) {
1329 			SDT_PROBE3(vfs, namecache, lookup, miss, dvp,
1330 			    "..", NULL);
1331 			mtx_unlock(dvlp);
1332 			return (0);
1333 		}
1334 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1335 			if (ncp->nc_flag & NCF_NEGATIVE)
1336 				*vpp = NULL;
1337 			else
1338 				*vpp = ncp->nc_vp;
1339 		} else
1340 			*vpp = ncp->nc_dvp;
1341 		/* Return failure if negative entry was found. */
1342 		if (*vpp == NULL)
1343 			goto negative_success;
1344 		CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..",
1345 		    dvp, cnp->cn_nameptr, *vpp);
1346 		SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..",
1347 		    *vpp);
1348 		cache_out_ts(ncp, tsp, ticksp);
1349 		if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1350 		    NCF_DTS && tsp != NULL) {
1351 			ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1352 			*tsp = ncp_ts->nc_dotdottime;
1353 		}
1354 		goto success;
1355 	}
1356 
1357 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1358 	blp = HASH2BUCKETLOCK(hash);
1359 	rw_rlock(blp);
1360 
1361 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1362 		counter_u64_add(numchecks, 1);
1363 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1364 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1365 			break;
1366 	}
1367 
1368 	/* We failed to find an entry */
1369 	if (__predict_false(ncp == NULL)) {
1370 		rw_runlock(blp);
1371 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1372 		    NULL);
1373 		counter_u64_add(nummiss, 1);
1374 		return (0);
1375 	}
1376 
1377 	if (ncp->nc_flag & NCF_NEGATIVE)
1378 		goto negative_success;
1379 
1380 	/* We found a "positive" match, return the vnode */
1381 	counter_u64_add(numposhits, 1);
1382 	*vpp = ncp->nc_vp;
1383 	CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p",
1384 	    dvp, cnp->cn_nameptr, *vpp, ncp);
1385 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name,
1386 	    *vpp);
1387 	cache_out_ts(ncp, tsp, ticksp);
1388 success:
1389 	/*
1390 	 * On success we return a locked and ref'd vnode as per the lookup
1391 	 * protocol.
1392 	 */
1393 	MPASS(dvp != *vpp);
1394 	ltype = 0;	/* silence gcc warning */
1395 	if (cnp->cn_flags & ISDOTDOT) {
1396 		ltype = VOP_ISLOCKED(dvp);
1397 		VOP_UNLOCK(dvp);
1398 	}
1399 	vs = vget_prep(*vpp);
1400 	cache_lookup_unlock(blp, dvlp);
1401 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1402 	if (cnp->cn_flags & ISDOTDOT) {
1403 		vn_lock(dvp, ltype | LK_RETRY);
1404 		if (VN_IS_DOOMED(dvp)) {
1405 			if (error == 0)
1406 				vput(*vpp);
1407 			*vpp = NULL;
1408 			return (ENOENT);
1409 		}
1410 	}
1411 	if (error) {
1412 		*vpp = NULL;
1413 		goto retry;
1414 	}
1415 	if ((cnp->cn_flags & ISLASTCN) &&
1416 	    (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) {
1417 		ASSERT_VOP_ELOCKED(*vpp, "cache_lookup");
1418 	}
1419 	return (-1);
1420 
1421 negative_success:
1422 	/* We found a negative match, and want to create it, so purge */
1423 	if (cnp->cn_nameiop == CREATE) {
1424 		counter_u64_add(numnegzaps, 1);
1425 		goto zap_and_exit;
1426 	}
1427 
1428 	counter_u64_add(numneghits, 1);
1429 	cache_negative_hit(ncp);
1430 	if (ncp->nc_flag & NCF_WHITE)
1431 		cnp->cn_flags |= ISWHITEOUT;
1432 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp,
1433 	    ncp->nc_name);
1434 	cache_out_ts(ncp, tsp, ticksp);
1435 	cache_lookup_unlock(blp, dvlp);
1436 	return (ENOENT);
1437 
1438 zap_and_exit:
1439 	if (blp != NULL)
1440 		error = cache_zap_rlocked_bucket(ncp, cnp, hash, blp);
1441 	else
1442 		error = cache_zap_locked_vnode(ncp, dvp);
1443 	if (__predict_false(error != 0)) {
1444 		zap_and_exit_bucket_fail2++;
1445 		cache_maybe_yield();
1446 		goto retry;
1447 	}
1448 	cache_free(ncp);
1449 	return (0);
1450 }
1451 
1452 struct celockstate {
1453 	struct mtx *vlp[3];
1454 	struct rwlock *blp[2];
1455 };
1456 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1457 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1458 
1459 static inline void
1460 cache_celockstate_init(struct celockstate *cel)
1461 {
1462 
1463 	bzero(cel, sizeof(*cel));
1464 }
1465 
1466 static void
1467 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1468     struct vnode *dvp)
1469 {
1470 	struct mtx *vlp1, *vlp2;
1471 
1472 	MPASS(cel->vlp[0] == NULL);
1473 	MPASS(cel->vlp[1] == NULL);
1474 	MPASS(cel->vlp[2] == NULL);
1475 
1476 	MPASS(vp != NULL || dvp != NULL);
1477 
1478 	vlp1 = VP2VNODELOCK(vp);
1479 	vlp2 = VP2VNODELOCK(dvp);
1480 	cache_sort_vnodes(&vlp1, &vlp2);
1481 
1482 	if (vlp1 != NULL) {
1483 		mtx_lock(vlp1);
1484 		cel->vlp[0] = vlp1;
1485 	}
1486 	mtx_lock(vlp2);
1487 	cel->vlp[1] = vlp2;
1488 }
1489 
1490 static void
1491 cache_unlock_vnodes_cel(struct celockstate *cel)
1492 {
1493 
1494 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
1495 
1496 	if (cel->vlp[0] != NULL)
1497 		mtx_unlock(cel->vlp[0]);
1498 	if (cel->vlp[1] != NULL)
1499 		mtx_unlock(cel->vlp[1]);
1500 	if (cel->vlp[2] != NULL)
1501 		mtx_unlock(cel->vlp[2]);
1502 }
1503 
1504 static bool
1505 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
1506 {
1507 	struct mtx *vlp;
1508 	bool ret;
1509 
1510 	cache_assert_vlp_locked(cel->vlp[0]);
1511 	cache_assert_vlp_locked(cel->vlp[1]);
1512 	MPASS(cel->vlp[2] == NULL);
1513 
1514 	MPASS(vp != NULL);
1515 	vlp = VP2VNODELOCK(vp);
1516 
1517 	ret = true;
1518 	if (vlp >= cel->vlp[1]) {
1519 		mtx_lock(vlp);
1520 	} else {
1521 		if (mtx_trylock(vlp))
1522 			goto out;
1523 		cache_lock_vnodes_cel_3_failures++;
1524 		cache_unlock_vnodes_cel(cel);
1525 		if (vlp < cel->vlp[0]) {
1526 			mtx_lock(vlp);
1527 			mtx_lock(cel->vlp[0]);
1528 			mtx_lock(cel->vlp[1]);
1529 		} else {
1530 			if (cel->vlp[0] != NULL)
1531 				mtx_lock(cel->vlp[0]);
1532 			mtx_lock(vlp);
1533 			mtx_lock(cel->vlp[1]);
1534 		}
1535 		ret = false;
1536 	}
1537 out:
1538 	cel->vlp[2] = vlp;
1539 	return (ret);
1540 }
1541 
1542 static void
1543 cache_lock_buckets_cel(struct celockstate *cel, struct rwlock *blp1,
1544     struct rwlock *blp2)
1545 {
1546 
1547 	MPASS(cel->blp[0] == NULL);
1548 	MPASS(cel->blp[1] == NULL);
1549 
1550 	cache_sort_vnodes(&blp1, &blp2);
1551 
1552 	if (blp1 != NULL) {
1553 		rw_wlock(blp1);
1554 		cel->blp[0] = blp1;
1555 	}
1556 	rw_wlock(blp2);
1557 	cel->blp[1] = blp2;
1558 }
1559 
1560 static void
1561 cache_unlock_buckets_cel(struct celockstate *cel)
1562 {
1563 
1564 	if (cel->blp[0] != NULL)
1565 		rw_wunlock(cel->blp[0]);
1566 	rw_wunlock(cel->blp[1]);
1567 }
1568 
1569 /*
1570  * Lock part of the cache affected by the insertion.
1571  *
1572  * This means vnodelocks for dvp, vp and the relevant bucketlock.
1573  * However, insertion can result in removal of an old entry. In this
1574  * case we have an additional vnode and bucketlock pair to lock. If the
1575  * entry is negative, ncelock is locked instead of the vnode.
1576  *
1577  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
1578  * preserving the locking order (smaller address first).
1579  */
1580 static void
1581 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1582     uint32_t hash)
1583 {
1584 	struct namecache *ncp;
1585 	struct rwlock *blps[2];
1586 
1587 	blps[0] = HASH2BUCKETLOCK(hash);
1588 	for (;;) {
1589 		blps[1] = NULL;
1590 		cache_lock_vnodes_cel(cel, dvp, vp);
1591 		if (vp == NULL || vp->v_type != VDIR)
1592 			break;
1593 		ncp = vp->v_cache_dd;
1594 		if (ncp == NULL)
1595 			break;
1596 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1597 			break;
1598 		MPASS(ncp->nc_dvp == vp);
1599 		blps[1] = NCP2BUCKETLOCK(ncp);
1600 		if (ncp->nc_flag & NCF_NEGATIVE)
1601 			break;
1602 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1603 			break;
1604 		/*
1605 		 * All vnodes got re-locked. Re-validate the state and if
1606 		 * nothing changed we are done. Otherwise restart.
1607 		 */
1608 		if (ncp == vp->v_cache_dd &&
1609 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1610 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1611 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1612 			break;
1613 		cache_unlock_vnodes_cel(cel);
1614 		cel->vlp[0] = NULL;
1615 		cel->vlp[1] = NULL;
1616 		cel->vlp[2] = NULL;
1617 	}
1618 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1619 }
1620 
1621 static void
1622 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
1623     uint32_t hash)
1624 {
1625 	struct namecache *ncp;
1626 	struct rwlock *blps[2];
1627 
1628 	blps[0] = HASH2BUCKETLOCK(hash);
1629 	for (;;) {
1630 		blps[1] = NULL;
1631 		cache_lock_vnodes_cel(cel, dvp, vp);
1632 		ncp = dvp->v_cache_dd;
1633 		if (ncp == NULL)
1634 			break;
1635 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
1636 			break;
1637 		MPASS(ncp->nc_dvp == dvp);
1638 		blps[1] = NCP2BUCKETLOCK(ncp);
1639 		if (ncp->nc_flag & NCF_NEGATIVE)
1640 			break;
1641 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
1642 			break;
1643 		if (ncp == dvp->v_cache_dd &&
1644 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
1645 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
1646 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
1647 			break;
1648 		cache_unlock_vnodes_cel(cel);
1649 		cel->vlp[0] = NULL;
1650 		cel->vlp[1] = NULL;
1651 		cel->vlp[2] = NULL;
1652 	}
1653 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
1654 }
1655 
1656 static void
1657 cache_enter_unlock(struct celockstate *cel)
1658 {
1659 
1660 	cache_unlock_buckets_cel(cel);
1661 	cache_unlock_vnodes_cel(cel);
1662 }
1663 
1664 static void __noinline
1665 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp,
1666     struct componentname *cnp)
1667 {
1668 	struct celockstate cel;
1669 	struct namecache *ncp;
1670 	uint32_t hash;
1671 	int len;
1672 
1673 	if (dvp->v_cache_dd == NULL)
1674 		return;
1675 	len = cnp->cn_namelen;
1676 	cache_celockstate_init(&cel);
1677 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1678 	cache_enter_lock_dd(&cel, dvp, vp, hash);
1679 	ncp = dvp->v_cache_dd;
1680 	if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) {
1681 		KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent"));
1682 		cache_zap_locked(ncp, false);
1683 	} else {
1684 		ncp = NULL;
1685 	}
1686 	dvp->v_cache_dd = NULL;
1687 	cache_enter_unlock(&cel);
1688 	cache_free(ncp);
1689 }
1690 
1691 /*
1692  * Add an entry to the cache.
1693  */
1694 void
1695 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
1696     struct timespec *tsp, struct timespec *dtsp)
1697 {
1698 	struct celockstate cel;
1699 	struct namecache *ncp, *n2, *ndd;
1700 	struct namecache_ts *ncp_ts, *n2_ts;
1701 	struct nchashhead *ncpp;
1702 	uint32_t hash;
1703 	int flag;
1704 	int len;
1705 	bool held_dvp;
1706 	u_long lnumcache;
1707 
1708 	CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr);
1709 	VNASSERT(vp == NULL || !VN_IS_DOOMED(vp), vp,
1710 	    ("cache_enter: Adding a doomed vnode"));
1711 	VNASSERT(dvp == NULL || !VN_IS_DOOMED(dvp), dvp,
1712 	    ("cache_enter: Doomed vnode used as src"));
1713 
1714 #ifdef DEBUG_CACHE
1715 	if (__predict_false(!doingcache))
1716 		return;
1717 #endif
1718 
1719 	flag = 0;
1720 	if (__predict_false(cnp->cn_nameptr[0] == '.')) {
1721 		if (cnp->cn_namelen == 1)
1722 			return;
1723 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
1724 			cache_enter_dotdot_prep(dvp, vp, cnp);
1725 			flag = NCF_ISDOTDOT;
1726 		}
1727 	}
1728 
1729 	/*
1730 	 * Avoid blowout in namecache entries.
1731 	 */
1732 	lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
1733 	if (__predict_false(lnumcache >= ncsize)) {
1734 		atomic_add_long(&numcache, -1);
1735 		return;
1736 	}
1737 
1738 	cache_celockstate_init(&cel);
1739 	ndd = NULL;
1740 	ncp_ts = NULL;
1741 
1742 	held_dvp = false;
1743 	if (LIST_EMPTY(&dvp->v_cache_src) && flag != NCF_ISDOTDOT) {
1744 		vhold(dvp);
1745 		atomic_add_long(&numcachehv, 1);
1746 		held_dvp = true;
1747 	}
1748 
1749 	/*
1750 	 * Calculate the hash key and setup as much of the new
1751 	 * namecache entry as possible before acquiring the lock.
1752 	 */
1753 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
1754 	ncp->nc_flag = flag;
1755 	ncp->nc_vp = vp;
1756 	if (vp == NULL)
1757 		ncp->nc_flag |= NCF_NEGATIVE;
1758 	ncp->nc_dvp = dvp;
1759 	if (tsp != NULL) {
1760 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1761 		ncp_ts->nc_time = *tsp;
1762 		ncp_ts->nc_ticks = ticks;
1763 		ncp_ts->nc_nc.nc_flag |= NCF_TS;
1764 		if (dtsp != NULL) {
1765 			ncp_ts->nc_dotdottime = *dtsp;
1766 			ncp_ts->nc_nc.nc_flag |= NCF_DTS;
1767 		}
1768 	}
1769 	len = ncp->nc_nlen = cnp->cn_namelen;
1770 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
1771 	strlcpy(ncp->nc_name, cnp->cn_nameptr, len + 1);
1772 	cache_enter_lock(&cel, dvp, vp, hash);
1773 
1774 	/*
1775 	 * See if this vnode or negative entry is already in the cache
1776 	 * with this name.  This can happen with concurrent lookups of
1777 	 * the same path name.
1778 	 */
1779 	ncpp = NCHHASH(hash);
1780 	LIST_FOREACH(n2, ncpp, nc_hash) {
1781 		if (n2->nc_dvp == dvp &&
1782 		    n2->nc_nlen == cnp->cn_namelen &&
1783 		    !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
1784 			if (tsp != NULL) {
1785 				KASSERT((n2->nc_flag & NCF_TS) != 0,
1786 				    ("no NCF_TS"));
1787 				n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
1788 				n2_ts->nc_time = ncp_ts->nc_time;
1789 				n2_ts->nc_ticks = ncp_ts->nc_ticks;
1790 				if (dtsp != NULL) {
1791 					n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
1792 					if (ncp->nc_flag & NCF_NEGATIVE)
1793 						mtx_lock(&ncneg_hot.nl_lock);
1794 					n2_ts->nc_nc.nc_flag |= NCF_DTS;
1795 					if (ncp->nc_flag & NCF_NEGATIVE)
1796 						mtx_unlock(&ncneg_hot.nl_lock);
1797 				}
1798 			}
1799 			goto out_unlock_free;
1800 		}
1801 	}
1802 
1803 	if (flag == NCF_ISDOTDOT) {
1804 		/*
1805 		 * See if we are trying to add .. entry, but some other lookup
1806 		 * has populated v_cache_dd pointer already.
1807 		 */
1808 		if (dvp->v_cache_dd != NULL)
1809 			goto out_unlock_free;
1810 		KASSERT(vp == NULL || vp->v_type == VDIR,
1811 		    ("wrong vnode type %p", vp));
1812 		dvp->v_cache_dd = ncp;
1813 	}
1814 
1815 	if (vp != NULL) {
1816 		if (vp->v_type == VDIR) {
1817 			if (flag != NCF_ISDOTDOT) {
1818 				/*
1819 				 * For this case, the cache entry maps both the
1820 				 * directory name in it and the name ".." for the
1821 				 * directory's parent.
1822 				 */
1823 				if ((ndd = vp->v_cache_dd) != NULL) {
1824 					if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
1825 						cache_zap_locked(ndd, false);
1826 					else
1827 						ndd = NULL;
1828 				}
1829 				vp->v_cache_dd = ncp;
1830 			}
1831 		} else {
1832 			vp->v_cache_dd = NULL;
1833 		}
1834 	}
1835 
1836 	if (flag != NCF_ISDOTDOT) {
1837 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1838 			if (!held_dvp) {
1839 				vhold(dvp);
1840 				atomic_add_long(&numcachehv, 1);
1841 			}
1842 		} else {
1843 			if (held_dvp) {
1844 				/*
1845 				 * This will not take the interlock as someone
1846 				 * else already holds the vnode on account of
1847 				 * the namecache and we hold locks preventing
1848 				 * this from changing.
1849 				 */
1850 				vdrop(dvp);
1851 				atomic_subtract_long(&numcachehv, 1);
1852 			}
1853 		}
1854 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
1855 	}
1856 
1857 	/*
1858 	 * Insert the new namecache entry into the appropriate chain
1859 	 * within the cache entries table.
1860 	 */
1861 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
1862 
1863 	/*
1864 	 * If the entry is "negative", we place it into the
1865 	 * "negative" cache queue, otherwise, we place it into the
1866 	 * destination vnode's cache entries queue.
1867 	 */
1868 	if (vp != NULL) {
1869 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
1870 		SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
1871 		    vp);
1872 	} else {
1873 		if (cnp->cn_flags & ISWHITEOUT)
1874 			ncp->nc_flag |= NCF_WHITE;
1875 		cache_negative_insert(ncp, false);
1876 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
1877 		    ncp->nc_name);
1878 	}
1879 	cache_enter_unlock(&cel);
1880 	if (numneg * ncnegfactor > lnumcache)
1881 		cache_negative_zap_one();
1882 	cache_free(ndd);
1883 	return;
1884 out_unlock_free:
1885 	cache_enter_unlock(&cel);
1886 	cache_free(ncp);
1887 	if (held_dvp) {
1888 		vdrop(dvp);
1889 		atomic_subtract_long(&numcachehv, 1);
1890 	}
1891 	return;
1892 }
1893 
1894 static u_int
1895 cache_roundup_2(u_int val)
1896 {
1897 	u_int res;
1898 
1899 	for (res = 1; res <= val; res <<= 1)
1900 		continue;
1901 
1902 	return (res);
1903 }
1904 
1905 /*
1906  * Name cache initialization, from vfs_init() when we are booting
1907  */
1908 static void
1909 nchinit(void *dummy __unused)
1910 {
1911 	u_int i;
1912 
1913 	cache_zone_small = uma_zcreate("S VFS Cache",
1914 	    sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1,
1915 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1916 	    UMA_ZONE_ZINIT);
1917 	cache_zone_small_ts = uma_zcreate("STS VFS Cache",
1918 	    sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1,
1919 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1920 	    UMA_ZONE_ZINIT);
1921 	cache_zone_large = uma_zcreate("L VFS Cache",
1922 	    sizeof(struct namecache) + NAME_MAX + 1,
1923 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache),
1924 	    UMA_ZONE_ZINIT);
1925 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache",
1926 	    sizeof(struct namecache_ts) + NAME_MAX + 1,
1927 	    NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct namecache_ts),
1928 	    UMA_ZONE_ZINIT);
1929 
1930 	ncsize = desiredvnodes * ncsizefactor;
1931 	nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash);
1932 	ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
1933 	if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
1934 		ncbuckethash = 7;
1935 	if (ncbuckethash > nchash)
1936 		ncbuckethash = nchash;
1937 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
1938 	    M_WAITOK | M_ZERO);
1939 	for (i = 0; i < numbucketlocks; i++)
1940 		rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK | RW_RECURSE);
1941 	ncvnodehash = ncbuckethash;
1942 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
1943 	    M_WAITOK | M_ZERO);
1944 	for (i = 0; i < numvnodelocks; i++)
1945 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
1946 	ncpurgeminvnodes = numbucketlocks * 2;
1947 
1948 	ncneghash = 3;
1949 	neglists = malloc(sizeof(*neglists) * numneglists, M_VFSCACHE,
1950 	    M_WAITOK | M_ZERO);
1951 	for (i = 0; i < numneglists; i++) {
1952 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
1953 		TAILQ_INIT(&neglists[i].nl_list);
1954 	}
1955 	mtx_init(&ncneg_hot.nl_lock, "ncneglh", NULL, MTX_DEF);
1956 	TAILQ_INIT(&ncneg_hot.nl_list);
1957 
1958 	mtx_init(&ncneg_shrink_lock, "ncnegs", NULL, MTX_DEF);
1959 
1960 	numcalls = counter_u64_alloc(M_WAITOK);
1961 	dothits = counter_u64_alloc(M_WAITOK);
1962 	dotdothits = counter_u64_alloc(M_WAITOK);
1963 	numchecks = counter_u64_alloc(M_WAITOK);
1964 	nummiss = counter_u64_alloc(M_WAITOK);
1965 	nummisszap = counter_u64_alloc(M_WAITOK);
1966 	numposzaps = counter_u64_alloc(M_WAITOK);
1967 	numposhits = counter_u64_alloc(M_WAITOK);
1968 	numnegzaps = counter_u64_alloc(M_WAITOK);
1969 	numneghits = counter_u64_alloc(M_WAITOK);
1970 	numfullpathcalls = counter_u64_alloc(M_WAITOK);
1971 	numfullpathfail1 = counter_u64_alloc(M_WAITOK);
1972 	numfullpathfail2 = counter_u64_alloc(M_WAITOK);
1973 	numfullpathfail4 = counter_u64_alloc(M_WAITOK);
1974 	numfullpathfound = counter_u64_alloc(M_WAITOK);
1975 	zap_and_exit_bucket_relock_success = counter_u64_alloc(M_WAITOK);
1976 	numneg_evicted = counter_u64_alloc(M_WAITOK);
1977 	shrinking_skipped = counter_u64_alloc(M_WAITOK);
1978 }
1979 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
1980 
1981 void
1982 cache_changesize(u_long newmaxvnodes)
1983 {
1984 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
1985 	u_long new_nchash, old_nchash;
1986 	struct namecache *ncp;
1987 	uint32_t hash;
1988 	u_long newncsize;
1989 	int i;
1990 
1991 	newncsize = newmaxvnodes * ncsizefactor;
1992 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
1993 	if (newmaxvnodes < numbucketlocks)
1994 		newmaxvnodes = numbucketlocks;
1995 
1996 	new_nchashtbl = hashinit(newmaxvnodes, M_VFSCACHE, &new_nchash);
1997 	/* If same hash table size, nothing to do */
1998 	if (nchash == new_nchash) {
1999 		free(new_nchashtbl, M_VFSCACHE);
2000 		return;
2001 	}
2002 	/*
2003 	 * Move everything from the old hash table to the new table.
2004 	 * None of the namecache entries in the table can be removed
2005 	 * because to do so, they have to be removed from the hash table.
2006 	 */
2007 	cache_lock_all_vnodes();
2008 	cache_lock_all_buckets();
2009 	old_nchashtbl = nchashtbl;
2010 	old_nchash = nchash;
2011 	nchashtbl = new_nchashtbl;
2012 	nchash = new_nchash;
2013 	for (i = 0; i <= old_nchash; i++) {
2014 		while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) {
2015 			hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
2016 			    ncp->nc_dvp);
2017 			LIST_REMOVE(ncp, nc_hash);
2018 			LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
2019 		}
2020 	}
2021 	ncsize = newncsize;
2022 	cache_unlock_all_buckets();
2023 	cache_unlock_all_vnodes();
2024 	free(old_nchashtbl, M_VFSCACHE);
2025 }
2026 
2027 /*
2028  * Invalidate all entries from and to a particular vnode.
2029  */
2030 void
2031 cache_purge(struct vnode *vp)
2032 {
2033 	TAILQ_HEAD(, namecache) ncps;
2034 	struct namecache *ncp, *nnp;
2035 	struct mtx *vlp, *vlp2;
2036 
2037 	CTR1(KTR_VFS, "cache_purge(%p)", vp);
2038 	SDT_PROBE1(vfs, namecache, purge, done, vp);
2039 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2040 	    vp->v_cache_dd == NULL)
2041 		return;
2042 	TAILQ_INIT(&ncps);
2043 	vlp = VP2VNODELOCK(vp);
2044 	vlp2 = NULL;
2045 	mtx_lock(vlp);
2046 retry:
2047 	while (!LIST_EMPTY(&vp->v_cache_src)) {
2048 		ncp = LIST_FIRST(&vp->v_cache_src);
2049 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2050 			goto retry;
2051 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2052 	}
2053 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2054 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
2055 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2056 			goto retry;
2057 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2058 	}
2059 	ncp = vp->v_cache_dd;
2060 	if (ncp != NULL) {
2061 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2062 		   ("lost dotdot link"));
2063 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2064 			goto retry;
2065 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2066 	}
2067 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2068 	mtx_unlock(vlp);
2069 	if (vlp2 != NULL)
2070 		mtx_unlock(vlp2);
2071 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2072 		cache_free(ncp);
2073 	}
2074 }
2075 
2076 /*
2077  * Invalidate all negative entries for a particular directory vnode.
2078  */
2079 void
2080 cache_purge_negative(struct vnode *vp)
2081 {
2082 	TAILQ_HEAD(, namecache) ncps;
2083 	struct namecache *ncp, *nnp;
2084 	struct mtx *vlp;
2085 
2086 	CTR1(KTR_VFS, "cache_purge_negative(%p)", vp);
2087 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2088 	if (LIST_EMPTY(&vp->v_cache_src))
2089 		return;
2090 	TAILQ_INIT(&ncps);
2091 	vlp = VP2VNODELOCK(vp);
2092 	mtx_lock(vlp);
2093 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2094 		if (!(ncp->nc_flag & NCF_NEGATIVE))
2095 			continue;
2096 		cache_zap_negative_locked_vnode_kl(ncp, vp);
2097 		TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst);
2098 	}
2099 	mtx_unlock(vlp);
2100 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2101 		cache_free(ncp);
2102 	}
2103 }
2104 
2105 /*
2106  * Flush all entries referencing a particular filesystem.
2107  */
2108 void
2109 cache_purgevfs(struct mount *mp, bool force)
2110 {
2111 	TAILQ_HEAD(, namecache) ncps;
2112 	struct mtx *vlp1, *vlp2;
2113 	struct rwlock *blp;
2114 	struct nchashhead *bucket;
2115 	struct namecache *ncp, *nnp;
2116 	u_long i, j, n_nchash;
2117 	int error;
2118 
2119 	/* Scan hash tables for applicable entries */
2120 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2121 	if (!force && mp->mnt_nvnodelistsize <= ncpurgeminvnodes)
2122 		return;
2123 	TAILQ_INIT(&ncps);
2124 	n_nchash = nchash + 1;
2125 	vlp1 = vlp2 = NULL;
2126 	for (i = 0; i < numbucketlocks; i++) {
2127 		blp = (struct rwlock *)&bucketlocks[i];
2128 		rw_wlock(blp);
2129 		for (j = i; j < n_nchash; j += numbucketlocks) {
2130 retry:
2131 			bucket = &nchashtbl[j];
2132 			LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) {
2133 				cache_assert_bucket_locked(ncp, RA_WLOCKED);
2134 				if (ncp->nc_dvp->v_mount != mp)
2135 					continue;
2136 				error = cache_zap_wlocked_bucket_kl(ncp, blp,
2137 				    &vlp1, &vlp2);
2138 				if (error != 0)
2139 					goto retry;
2140 				TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst);
2141 			}
2142 		}
2143 		rw_wunlock(blp);
2144 		if (vlp1 == NULL && vlp2 == NULL)
2145 			cache_maybe_yield();
2146 	}
2147 	if (vlp1 != NULL)
2148 		mtx_unlock(vlp1);
2149 	if (vlp2 != NULL)
2150 		mtx_unlock(vlp2);
2151 
2152 	TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) {
2153 		cache_free(ncp);
2154 	}
2155 }
2156 
2157 /*
2158  * Perform canonical checks and cache lookup and pass on to filesystem
2159  * through the vop_cachedlookup only if needed.
2160  */
2161 
2162 int
2163 vfs_cache_lookup(struct vop_lookup_args *ap)
2164 {
2165 	struct vnode *dvp;
2166 	int error;
2167 	struct vnode **vpp = ap->a_vpp;
2168 	struct componentname *cnp = ap->a_cnp;
2169 	struct ucred *cred = cnp->cn_cred;
2170 	int flags = cnp->cn_flags;
2171 	struct thread *td = cnp->cn_thread;
2172 
2173 	*vpp = NULL;
2174 	dvp = ap->a_dvp;
2175 
2176 	if (dvp->v_type != VDIR)
2177 		return (ENOTDIR);
2178 
2179 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2180 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2181 		return (EROFS);
2182 
2183 	error = VOP_ACCESS(dvp, VEXEC, cred, td);
2184 	if (error)
2185 		return (error);
2186 
2187 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2188 	if (error == 0)
2189 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2190 	if (error == -1)
2191 		return (0);
2192 	return (error);
2193 }
2194 
2195 /*
2196  * XXX All of these sysctls would probably be more productive dead.
2197  */
2198 static int __read_mostly disablecwd;
2199 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
2200    "Disable the getcwd syscall");
2201 
2202 /* Implementation of the getcwd syscall. */
2203 int
2204 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2205 {
2206 
2207 	return (kern___getcwd(td, uap->buf, UIO_USERSPACE, uap->buflen,
2208 	    MAXPATHLEN));
2209 }
2210 
2211 int
2212 kern___getcwd(struct thread *td, char *buf, enum uio_seg bufseg, size_t buflen,
2213     size_t path_max)
2214 {
2215 	char *bp, *tmpbuf;
2216 	struct filedesc *fdp;
2217 	struct vnode *cdir, *rdir;
2218 	int error;
2219 
2220 	if (__predict_false(disablecwd))
2221 		return (ENODEV);
2222 	if (__predict_false(buflen < 2))
2223 		return (EINVAL);
2224 	if (buflen > path_max)
2225 		buflen = path_max;
2226 
2227 	tmpbuf = malloc(buflen, M_TEMP, M_WAITOK);
2228 	fdp = td->td_proc->p_fd;
2229 	FILEDESC_SLOCK(fdp);
2230 	cdir = fdp->fd_cdir;
2231 	vrefact(cdir);
2232 	rdir = fdp->fd_rdir;
2233 	vrefact(rdir);
2234 	FILEDESC_SUNLOCK(fdp);
2235 	error = vn_fullpath1(td, cdir, rdir, tmpbuf, &bp, buflen);
2236 	vrele(rdir);
2237 	vrele(cdir);
2238 
2239 	if (!error) {
2240 		if (bufseg == UIO_SYSSPACE)
2241 			bcopy(bp, buf, strlen(bp) + 1);
2242 		else
2243 			error = copyout(bp, buf, strlen(bp) + 1);
2244 #ifdef KTRACE
2245 	if (KTRPOINT(curthread, KTR_NAMEI))
2246 		ktrnamei(bp);
2247 #endif
2248 	}
2249 	free(tmpbuf, M_TEMP);
2250 	return (error);
2251 }
2252 
2253 /*
2254  * Thus begins the fullpath magic.
2255  */
2256 
2257 static int __read_mostly disablefullpath;
2258 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, &disablefullpath, 0,
2259     "Disable the vn_fullpath function");
2260 
2261 /*
2262  * Retrieve the full filesystem path that correspond to a vnode from the name
2263  * cache (if available)
2264  */
2265 int
2266 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf)
2267 {
2268 	char *buf;
2269 	struct filedesc *fdp;
2270 	struct vnode *rdir;
2271 	int error;
2272 
2273 	if (__predict_false(disablefullpath))
2274 		return (ENODEV);
2275 	if (__predict_false(vn == NULL))
2276 		return (EINVAL);
2277 
2278 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2279 	fdp = td->td_proc->p_fd;
2280 	FILEDESC_SLOCK(fdp);
2281 	rdir = fdp->fd_rdir;
2282 	vrefact(rdir);
2283 	FILEDESC_SUNLOCK(fdp);
2284 	error = vn_fullpath1(td, vn, rdir, buf, retbuf, MAXPATHLEN);
2285 	vrele(rdir);
2286 
2287 	if (!error)
2288 		*freebuf = buf;
2289 	else
2290 		free(buf, M_TEMP);
2291 	return (error);
2292 }
2293 
2294 /*
2295  * This function is similar to vn_fullpath, but it attempts to lookup the
2296  * pathname relative to the global root mount point.  This is required for the
2297  * auditing sub-system, as audited pathnames must be absolute, relative to the
2298  * global root mount point.
2299  */
2300 int
2301 vn_fullpath_global(struct thread *td, struct vnode *vn,
2302     char **retbuf, char **freebuf)
2303 {
2304 	char *buf;
2305 	int error;
2306 
2307 	if (__predict_false(disablefullpath))
2308 		return (ENODEV);
2309 	if (__predict_false(vn == NULL))
2310 		return (EINVAL);
2311 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2312 	error = vn_fullpath1(td, vn, rootvnode, buf, retbuf, MAXPATHLEN);
2313 	if (!error)
2314 		*freebuf = buf;
2315 	else
2316 		free(buf, M_TEMP);
2317 	return (error);
2318 }
2319 
2320 int
2321 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen)
2322 {
2323 	struct vnode *dvp;
2324 	struct namecache *ncp;
2325 	struct mtx *vlp;
2326 	int error;
2327 
2328 	vlp = VP2VNODELOCK(*vp);
2329 	mtx_lock(vlp);
2330 	TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) {
2331 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2332 			break;
2333 	}
2334 	if (ncp != NULL) {
2335 		if (*buflen < ncp->nc_nlen) {
2336 			mtx_unlock(vlp);
2337 			vrele(*vp);
2338 			counter_u64_add(numfullpathfail4, 1);
2339 			error = ENOMEM;
2340 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2341 			    vp, NULL);
2342 			return (error);
2343 		}
2344 		*buflen -= ncp->nc_nlen;
2345 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2346 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2347 		    ncp->nc_name, vp);
2348 		dvp = *vp;
2349 		*vp = ncp->nc_dvp;
2350 		vref(*vp);
2351 		mtx_unlock(vlp);
2352 		vrele(dvp);
2353 		return (0);
2354 	}
2355 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2356 
2357 	mtx_unlock(vlp);
2358 	vn_lock(*vp, LK_SHARED | LK_RETRY);
2359 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
2360 	vput(*vp);
2361 	if (error) {
2362 		counter_u64_add(numfullpathfail2, 1);
2363 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
2364 		return (error);
2365 	}
2366 
2367 	*vp = dvp;
2368 	if (VN_IS_DOOMED(dvp)) {
2369 		/* forced unmount */
2370 		vrele(dvp);
2371 		error = ENOENT;
2372 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
2373 		return (error);
2374 	}
2375 	/*
2376 	 * *vp has its use count incremented still.
2377 	 */
2378 
2379 	return (0);
2380 }
2381 
2382 /*
2383  * The magic behind kern___getcwd() and vn_fullpath().
2384  */
2385 static int
2386 vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
2387     char *buf, char **retbuf, u_int buflen)
2388 {
2389 	int error, slash_prefixed;
2390 #ifdef KDTRACE_HOOKS
2391 	struct vnode *startvp = vp;
2392 #endif
2393 	struct vnode *vp1;
2394 
2395 	buflen--;
2396 	buf[buflen] = '\0';
2397 	error = 0;
2398 	slash_prefixed = 0;
2399 
2400 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
2401 	counter_u64_add(numfullpathcalls, 1);
2402 	vref(vp);
2403 	if (vp->v_type != VDIR) {
2404 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2405 		if (error)
2406 			return (error);
2407 		if (buflen == 0) {
2408 			vrele(vp);
2409 			return (ENOMEM);
2410 		}
2411 		buf[--buflen] = '/';
2412 		slash_prefixed = 1;
2413 	}
2414 	while (vp != rdir && vp != rootvnode) {
2415 		/*
2416 		 * The vp vnode must be already fully constructed,
2417 		 * since it is either found in namecache or obtained
2418 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
2419 		 * without obtaining the vnode lock.
2420 		 */
2421 		if ((vp->v_vflag & VV_ROOT) != 0) {
2422 			vn_lock(vp, LK_RETRY | LK_SHARED);
2423 
2424 			/*
2425 			 * With the vnode locked, check for races with
2426 			 * unmount, forced or not.  Note that we
2427 			 * already verified that vp is not equal to
2428 			 * the root vnode, which means that
2429 			 * mnt_vnodecovered can be NULL only for the
2430 			 * case of unmount.
2431 			 */
2432 			if (VN_IS_DOOMED(vp) ||
2433 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
2434 			    vp1->v_mountedhere != vp->v_mount) {
2435 				vput(vp);
2436 				error = ENOENT;
2437 				SDT_PROBE3(vfs, namecache, fullpath, return,
2438 				    error, vp, NULL);
2439 				break;
2440 			}
2441 
2442 			vref(vp1);
2443 			vput(vp);
2444 			vp = vp1;
2445 			continue;
2446 		}
2447 		if (vp->v_type != VDIR) {
2448 			vrele(vp);
2449 			counter_u64_add(numfullpathfail1, 1);
2450 			error = ENOTDIR;
2451 			SDT_PROBE3(vfs, namecache, fullpath, return,
2452 			    error, vp, NULL);
2453 			break;
2454 		}
2455 		error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen);
2456 		if (error)
2457 			break;
2458 		if (buflen == 0) {
2459 			vrele(vp);
2460 			error = ENOMEM;
2461 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2462 			    startvp, NULL);
2463 			break;
2464 		}
2465 		buf[--buflen] = '/';
2466 		slash_prefixed = 1;
2467 	}
2468 	if (error)
2469 		return (error);
2470 	if (!slash_prefixed) {
2471 		if (buflen == 0) {
2472 			vrele(vp);
2473 			counter_u64_add(numfullpathfail4, 1);
2474 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
2475 			    startvp, NULL);
2476 			return (ENOMEM);
2477 		}
2478 		buf[--buflen] = '/';
2479 	}
2480 	counter_u64_add(numfullpathfound, 1);
2481 	vrele(vp);
2482 
2483 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, buf + buflen);
2484 	*retbuf = buf + buflen;
2485 	return (0);
2486 }
2487 
2488 struct vnode *
2489 vn_dir_dd_ino(struct vnode *vp)
2490 {
2491 	struct namecache *ncp;
2492 	struct vnode *ddvp;
2493 	struct mtx *vlp;
2494 	enum vgetstate vs;
2495 
2496 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
2497 	vlp = VP2VNODELOCK(vp);
2498 	mtx_lock(vlp);
2499 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
2500 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
2501 			continue;
2502 		ddvp = ncp->nc_dvp;
2503 		vs = vget_prep(ddvp);
2504 		mtx_unlock(vlp);
2505 		if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
2506 			return (NULL);
2507 		return (ddvp);
2508 	}
2509 	mtx_unlock(vlp);
2510 	return (NULL);
2511 }
2512 
2513 int
2514 vn_commname(struct vnode *vp, char *buf, u_int buflen)
2515 {
2516 	struct namecache *ncp;
2517 	struct mtx *vlp;
2518 	int l;
2519 
2520 	vlp = VP2VNODELOCK(vp);
2521 	mtx_lock(vlp);
2522 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
2523 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2524 			break;
2525 	if (ncp == NULL) {
2526 		mtx_unlock(vlp);
2527 		return (ENOENT);
2528 	}
2529 	l = min(ncp->nc_nlen, buflen - 1);
2530 	memcpy(buf, ncp->nc_name, l);
2531 	mtx_unlock(vlp);
2532 	buf[l] = '\0';
2533 	return (0);
2534 }
2535 
2536 /*
2537  * This function updates path string to vnode's full global path
2538  * and checks the size of the new path string against the pathlen argument.
2539  *
2540  * Requires a locked, referenced vnode.
2541  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
2542  *
2543  * If sysctl debug.disablefullpath is set, ENODEV is returned,
2544  * vnode is left locked and path remain untouched.
2545  *
2546  * If vp is a directory, the call to vn_fullpath_global() always succeeds
2547  * because it falls back to the ".." lookup if the namecache lookup fails.
2548  */
2549 int
2550 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
2551     u_int pathlen)
2552 {
2553 	struct nameidata nd;
2554 	struct vnode *vp1;
2555 	char *rpath, *fbuf;
2556 	int error;
2557 
2558 	ASSERT_VOP_ELOCKED(vp, __func__);
2559 
2560 	/* Return ENODEV if sysctl debug.disablefullpath==1 */
2561 	if (__predict_false(disablefullpath))
2562 		return (ENODEV);
2563 
2564 	/* Construct global filesystem path from vp. */
2565 	VOP_UNLOCK(vp);
2566 	error = vn_fullpath_global(td, vp, &rpath, &fbuf);
2567 
2568 	if (error != 0) {
2569 		vrele(vp);
2570 		return (error);
2571 	}
2572 
2573 	if (strlen(rpath) >= pathlen) {
2574 		vrele(vp);
2575 		error = ENAMETOOLONG;
2576 		goto out;
2577 	}
2578 
2579 	/*
2580 	 * Re-lookup the vnode by path to detect a possible rename.
2581 	 * As a side effect, the vnode is relocked.
2582 	 * If vnode was renamed, return ENOENT.
2583 	 */
2584 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
2585 	    UIO_SYSSPACE, path, td);
2586 	error = namei(&nd);
2587 	if (error != 0) {
2588 		vrele(vp);
2589 		goto out;
2590 	}
2591 	NDFREE(&nd, NDF_ONLY_PNBUF);
2592 	vp1 = nd.ni_vp;
2593 	vrele(vp);
2594 	if (vp1 == vp)
2595 		strcpy(path, rpath);
2596 	else {
2597 		vput(vp1);
2598 		error = ENOENT;
2599 	}
2600 
2601 out:
2602 	free(fbuf, M_TEMP);
2603 	return (error);
2604 }
2605 
2606 #ifdef DDB
2607 static void
2608 db_print_vpath(struct vnode *vp)
2609 {
2610 
2611 	while (vp != NULL) {
2612 		db_printf("%p: ", vp);
2613 		if (vp == rootvnode) {
2614 			db_printf("/");
2615 			vp = NULL;
2616 		} else {
2617 			if (vp->v_vflag & VV_ROOT) {
2618 				db_printf("<mount point>");
2619 				vp = vp->v_mount->mnt_vnodecovered;
2620 			} else {
2621 				struct namecache *ncp;
2622 				char *ncn;
2623 				int i;
2624 
2625 				ncp = TAILQ_FIRST(&vp->v_cache_dst);
2626 				if (ncp != NULL) {
2627 					ncn = ncp->nc_name;
2628 					for (i = 0; i < ncp->nc_nlen; i++)
2629 						db_printf("%c", *ncn++);
2630 					vp = ncp->nc_dvp;
2631 				} else {
2632 					vp = NULL;
2633 				}
2634 			}
2635 		}
2636 		db_printf("\n");
2637 	}
2638 
2639 	return;
2640 }
2641 
2642 DB_SHOW_COMMAND(vpath, db_show_vpath)
2643 {
2644 	struct vnode *vp;
2645 
2646 	if (!have_addr) {
2647 		db_printf("usage: show vpath <struct vnode *>\n");
2648 		return;
2649 	}
2650 
2651 	vp = (struct vnode *)addr;
2652 	db_print_vpath(vp);
2653 }
2654 
2655 #endif
2656