xref: /netbsd-src/sys/kern/vfs_cache.c (revision b1c86f5f087524e68db12794ee9c3e3da1ab17a0)
1 /*	$NetBSD: vfs_cache.c,v 1.87 2010/07/21 17:52:12 hannken Exp $	*/
2 
3 /*-
4  * Copyright (c) 2008 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
17  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
18  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
19  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
20  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
21  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
22  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
23  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
24  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
25  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
26  * POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 /*
30  * Copyright (c) 1989, 1993
31  *	The Regents of the University of California.  All rights reserved.
32  *
33  * Redistribution and use in source and binary forms, with or without
34  * modification, are permitted provided that the following conditions
35  * are met:
36  * 1. Redistributions of source code must retain the above copyright
37  *    notice, this list of conditions and the following disclaimer.
38  * 2. Redistributions in binary form must reproduce the above copyright
39  *    notice, this list of conditions and the following disclaimer in the
40  *    documentation and/or other materials provided with the distribution.
41  * 3. Neither the name of the University nor the names of its contributors
42  *    may be used to endorse or promote products derived from this software
43  *    without specific prior written permission.
44  *
45  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
46  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
47  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
48  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
49  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
50  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
51  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
52  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
53  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
54  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
55  * SUCH DAMAGE.
56  *
57  *	@(#)vfs_cache.c	8.3 (Berkeley) 8/22/94
58  */
59 
60 #include <sys/cdefs.h>
61 __KERNEL_RCSID(0, "$NetBSD: vfs_cache.c,v 1.87 2010/07/21 17:52:12 hannken Exp $");
62 
63 #include "opt_ddb.h"
64 #include "opt_revcache.h"
65 
66 #include <sys/param.h>
67 #include <sys/systm.h>
68 #include <sys/time.h>
69 #include <sys/mount.h>
70 #include <sys/vnode.h>
71 #include <sys/namei.h>
72 #include <sys/errno.h>
73 #include <sys/pool.h>
74 #include <sys/mutex.h>
75 #include <sys/atomic.h>
76 #include <sys/kthread.h>
77 #include <sys/kernel.h>
78 #include <sys/cpu.h>
79 #include <sys/evcnt.h>
80 
81 #define NAMECACHE_ENTER_REVERSE
82 /*
83  * Name caching works as follows:
84  *
85  * Names found by directory scans are retained in a cache
86  * for future reference.  It is managed LRU, so frequently
87  * used names will hang around.  Cache is indexed by hash value
88  * obtained from (dvp, name) where dvp refers to the directory
89  * containing name.
90  *
91  * For simplicity (and economy of storage), names longer than
92  * a maximum length of NCHNAMLEN are not cached; they occur
93  * infrequently in any case, and are almost never of interest.
94  *
95  * Upon reaching the last segment of a path, if the reference
96  * is for DELETE, or NOCACHE is set (rewrite), and the
97  * name is located in the cache, it will be dropped.
98  * The entry is dropped also when it was not possible to lock
99  * the cached vnode, either because vget() failed or the generation
100  * number has changed while waiting for the lock.
101  */
102 
103 /*
104  * Per-cpu namecache data.
105  */
106 struct nchcpu {
107 	kmutex_t	cpu_lock;
108 	struct nchstats	cpu_stats;
109 };
110 
111 /*
112  * Structures associated with name cacheing.
113  */
114 LIST_HEAD(nchashhead, namecache) *nchashtbl;
115 u_long	nchash;				/* size of hash table - 1 */
116 #define	NCHASH(cnp, dvp)	\
117 	(((cnp)->cn_hash ^ ((uintptr_t)(dvp) >> 3)) & nchash)
118 
119 LIST_HEAD(ncvhashhead, namecache) *ncvhashtbl;
120 u_long	ncvhash;			/* size of hash table - 1 */
121 #define	NCVHASH(vp)		(((uintptr_t)(vp) >> 3) & ncvhash)
122 
123 long	numcache;			/* number of cache entries allocated */
124 static u_int	cache_gcpend;		/* number of entries pending GC */
125 static void	*cache_gcqueue;		/* garbage collection queue */
126 
127 TAILQ_HEAD(, namecache) nclruhead =		/* LRU chain */
128 	TAILQ_HEAD_INITIALIZER(nclruhead);
129 #define	COUNT(c,x)	(c.x++)
130 struct	nchstats nchstats;		/* cache effectiveness statistics */
131 
132 static pool_cache_t namecache_cache;
133 
134 int cache_lowat = 95;
135 int cache_hiwat = 98;
136 int cache_hottime = 5;			/* number of seconds */
137 int doingcache = 1;			/* 1 => enable the cache */
138 
139 static struct evcnt cache_ev_scan;
140 static struct evcnt cache_ev_gc;
141 static struct evcnt cache_ev_over;
142 static struct evcnt cache_ev_under;
143 static struct evcnt cache_ev_forced;
144 
145 /* A single lock to serialize modifications. */
146 static kmutex_t *namecache_lock;
147 
148 static void cache_invalidate(struct namecache *);
149 static inline struct namecache *cache_lookup_entry(
150     const struct vnode *, const struct componentname *);
151 static void cache_thread(void *);
152 static void cache_invalidate(struct namecache *);
153 static void cache_disassociate(struct namecache *);
154 static void cache_reclaim(void);
155 static int cache_ctor(void *, void *, int);
156 static void cache_dtor(void *, void *);
157 
158 /*
159  * Invalidate a cache entry and enqueue it for garbage collection.
160  */
161 static void
162 cache_invalidate(struct namecache *ncp)
163 {
164 	void *head;
165 
166 	KASSERT(mutex_owned(&ncp->nc_lock));
167 
168 	if (ncp->nc_dvp != NULL) {
169 		ncp->nc_vp = NULL;
170 		ncp->nc_dvp = NULL;
171 		do {
172 			head = cache_gcqueue;
173 			ncp->nc_gcqueue = head;
174 		} while (atomic_cas_ptr(&cache_gcqueue, head, ncp) != head);
175 		atomic_inc_uint(&cache_gcpend);
176 	}
177 }
178 
179 /*
180  * Disassociate a namecache entry from any vnodes it is attached to,
181  * and remove from the global LRU list.
182  */
183 static void
184 cache_disassociate(struct namecache *ncp)
185 {
186 
187 	KASSERT(mutex_owned(namecache_lock));
188 	KASSERT(ncp->nc_dvp == NULL);
189 
190 	if (ncp->nc_lru.tqe_prev != NULL) {
191 		TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
192 		ncp->nc_lru.tqe_prev = NULL;
193 	}
194 	if (ncp->nc_vhash.le_prev != NULL) {
195 		LIST_REMOVE(ncp, nc_vhash);
196 		ncp->nc_vhash.le_prev = NULL;
197 	}
198 	if (ncp->nc_vlist.le_prev != NULL) {
199 		LIST_REMOVE(ncp, nc_vlist);
200 		ncp->nc_vlist.le_prev = NULL;
201 	}
202 	if (ncp->nc_dvlist.le_prev != NULL) {
203 		LIST_REMOVE(ncp, nc_dvlist);
204 		ncp->nc_dvlist.le_prev = NULL;
205 	}
206 }
207 
208 /*
209  * Lock all CPUs to prevent any cache lookup activity.  Conceptually,
210  * this locks out all "readers".
211  */
212 static void
213 cache_lock_cpus(void)
214 {
215 	CPU_INFO_ITERATOR cii;
216 	struct cpu_info *ci;
217 	struct nchcpu *cpup;
218 	long *s, *d, *m;
219 
220 	for (CPU_INFO_FOREACH(cii, ci)) {
221 		cpup = ci->ci_data.cpu_nch;
222 		mutex_enter(&cpup->cpu_lock);
223 
224 		/* Collate statistics. */
225 		d = (long *)&nchstats;
226 		s = (long *)&cpup->cpu_stats;
227 		m = s + sizeof(nchstats) / sizeof(long);
228 		for (; s < m; s++, d++) {
229 			*d += *s;
230 			*s = 0;
231 		}
232 	}
233 }
234 
235 /*
236  * Release all CPU locks.
237  */
238 static void
239 cache_unlock_cpus(void)
240 {
241 	CPU_INFO_ITERATOR cii;
242 	struct cpu_info *ci;
243 	struct nchcpu *cpup;
244 
245 	for (CPU_INFO_FOREACH(cii, ci)) {
246 		cpup = ci->ci_data.cpu_nch;
247 		mutex_exit(&cpup->cpu_lock);
248 	}
249 }
250 
251 /*
252  * Find a single cache entry and return it locked.  'namecache_lock' or
253  * at least one of the per-CPU locks must be held.
254  */
255 static struct namecache *
256 cache_lookup_entry(const struct vnode *dvp, const struct componentname *cnp)
257 {
258 	struct nchashhead *ncpp;
259 	struct namecache *ncp;
260 
261 	KASSERT(dvp != NULL);
262 	ncpp = &nchashtbl[NCHASH(cnp, dvp)];
263 
264 	LIST_FOREACH(ncp, ncpp, nc_hash) {
265 		if (ncp->nc_dvp != dvp ||
266 		    ncp->nc_nlen != cnp->cn_namelen ||
267 		    memcmp(ncp->nc_name, cnp->cn_nameptr, (u_int)ncp->nc_nlen))
268 		    	continue;
269 	    	mutex_enter(&ncp->nc_lock);
270 		if (__predict_true(ncp->nc_dvp == dvp)) {
271 			ncp->nc_hittime = hardclock_ticks;
272 			return ncp;
273 		}
274 		/* Raced: entry has been nullified. */
275 		mutex_exit(&ncp->nc_lock);
276 	}
277 
278 	return NULL;
279 }
280 
281 /*
282  * Look for a the name in the cache. We don't do this
283  * if the segment name is long, simply so the cache can avoid
284  * holding long names (which would either waste space, or
285  * add greatly to the complexity).
286  *
287  * Lookup is called with ni_dvp pointing to the directory to search,
288  * ni_ptr pointing to the name of the entry being sought, ni_namelen
289  * tells the length of the name, and ni_hash contains a hash of
290  * the name. If the lookup succeeds, the vnode is locked, stored in ni_vp
291  * and a status of zero is returned. If the locking fails for whatever
292  * reason, the vnode is unlocked and the error is returned to caller.
293  * If the lookup determines that the name does not exist (negative cacheing),
294  * a status of ENOENT is returned. If the lookup fails, a status of -1
295  * is returned.
296  */
297 int
298 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
299 {
300 	struct namecache *ncp;
301 	struct vnode *vp;
302 	struct nchcpu *cpup;
303 	int error;
304 
305 	if (__predict_false(!doingcache)) {
306 		cnp->cn_flags &= ~MAKEENTRY;
307 		*vpp = NULL;
308 		return -1;
309 	}
310 
311 	cpup = curcpu()->ci_data.cpu_nch;
312 	mutex_enter(&cpup->cpu_lock);
313 	if (__predict_false(cnp->cn_namelen > NCHNAMLEN)) {
314 		COUNT(cpup->cpu_stats, ncs_long);
315 		cnp->cn_flags &= ~MAKEENTRY;
316 		mutex_exit(&cpup->cpu_lock);
317 		*vpp = NULL;
318 		return -1;
319 	}
320 	ncp = cache_lookup_entry(dvp, cnp);
321 	if (__predict_false(ncp == NULL)) {
322 		COUNT(cpup->cpu_stats, ncs_miss);
323 		mutex_exit(&cpup->cpu_lock);
324 		*vpp = NULL;
325 		return -1;
326 	}
327 	if ((cnp->cn_flags & MAKEENTRY) == 0) {
328 		COUNT(cpup->cpu_stats, ncs_badhits);
329 		/*
330 		 * Last component and we are renaming or deleting,
331 		 * the cache entry is invalid, or otherwise don't
332 		 * want cache entry to exist.
333 		 */
334 		cache_invalidate(ncp);
335 		mutex_exit(&ncp->nc_lock);
336 		mutex_exit(&cpup->cpu_lock);
337 		*vpp = NULL;
338 		return -1;
339 	} else if (ncp->nc_vp == NULL) {
340 		/*
341 		 * Restore the ISWHITEOUT flag saved earlier.
342 		 */
343 		KASSERT((ncp->nc_flags & ~ISWHITEOUT) == 0);
344 		cnp->cn_flags |= ncp->nc_flags;
345 		if (__predict_true(cnp->cn_nameiop != CREATE ||
346 		    (cnp->cn_flags & ISLASTCN) == 0)) {
347 			COUNT(cpup->cpu_stats, ncs_neghits);
348 			mutex_exit(&ncp->nc_lock);
349 			mutex_exit(&cpup->cpu_lock);
350 			return ENOENT;
351 		} else {
352 			COUNT(cpup->cpu_stats, ncs_badhits);
353 			/*
354 			 * Last component and we are renaming or
355 			 * deleting, the cache entry is invalid,
356 			 * or otherwise don't want cache entry to
357 			 * exist.
358 			 */
359 			cache_invalidate(ncp);
360 			mutex_exit(&ncp->nc_lock);
361 			mutex_exit(&cpup->cpu_lock);
362 			*vpp = NULL;
363 			return -1;
364 		}
365 	}
366 
367 	vp = ncp->nc_vp;
368 	if (vtryget(vp)) {
369 		mutex_exit(&ncp->nc_lock);
370 		mutex_exit(&cpup->cpu_lock);
371 	} else {
372 		mutex_enter(&vp->v_interlock);
373 		mutex_exit(&ncp->nc_lock);
374 		mutex_exit(&cpup->cpu_lock);
375 		error = vget(vp, LK_NOWAIT);
376 		if (error) {
377 			KASSERT(error == EBUSY);
378 			/*
379 			 * This vnode is being cleaned out.
380 			 * XXX badhits?
381 			 */
382 			COUNT(cpup->cpu_stats, ncs_falsehits);
383 			*vpp = NULL;
384 			return -1;
385 		}
386 	}
387 
388 #ifdef DEBUG
389 	/*
390 	 * since we released nb->nb_lock,
391 	 * we can't use this pointer any more.
392 	 */
393 	ncp = NULL;
394 #endif /* DEBUG */
395 
396 	if (vp == dvp) {	/* lookup on "." */
397 		error = 0;
398 	} else if (cnp->cn_flags & ISDOTDOT) {
399 		VOP_UNLOCK(dvp);
400 		error = vn_lock(vp, LK_EXCLUSIVE);
401 		vn_lock(dvp, LK_EXCLUSIVE | LK_RETRY);
402 	} else {
403 		error = vn_lock(vp, LK_EXCLUSIVE);
404 	}
405 
406 	/*
407 	 * Check that the lock succeeded.
408 	 */
409 	if (error) {
410 		/* Unlocked, but only for stats. */
411 		COUNT(cpup->cpu_stats, ncs_badhits);
412 		vrele(vp);
413 		*vpp = NULL;
414 		return -1;
415 	}
416 
417 	/* Unlocked, but only for stats. */
418 	COUNT(cpup->cpu_stats, ncs_goodhits);
419 	*vpp = vp;
420 	return 0;
421 }
422 
423 int
424 cache_lookup_raw(struct vnode *dvp, struct vnode **vpp,
425     struct componentname *cnp)
426 {
427 	struct namecache *ncp;
428 	struct vnode *vp;
429 	struct nchcpu *cpup;
430 	int error;
431 
432 	if (__predict_false(!doingcache)) {
433 		cnp->cn_flags &= ~MAKEENTRY;
434 		*vpp = NULL;
435 		return (-1);
436 	}
437 
438 	cpup = curcpu()->ci_data.cpu_nch;
439 	mutex_enter(&cpup->cpu_lock);
440 	if (__predict_false(cnp->cn_namelen > NCHNAMLEN)) {
441 		COUNT(cpup->cpu_stats, ncs_long);
442 		cnp->cn_flags &= ~MAKEENTRY;
443 		mutex_exit(&cpup->cpu_lock);
444 		*vpp = NULL;
445 		return -1;
446 	}
447 	ncp = cache_lookup_entry(dvp, cnp);
448 	if (__predict_false(ncp == NULL)) {
449 		COUNT(cpup->cpu_stats, ncs_miss);
450 		mutex_exit(&cpup->cpu_lock);
451 		*vpp = NULL;
452 		return -1;
453 	}
454 	vp = ncp->nc_vp;
455 	if (vp == NULL) {
456 		/*
457 		 * Restore the ISWHITEOUT flag saved earlier.
458 		 */
459 		KASSERT((ncp->nc_flags & ~ISWHITEOUT) == 0);
460 		cnp->cn_flags |= ncp->nc_flags;
461 		COUNT(cpup->cpu_stats, ncs_neghits);
462 		mutex_exit(&ncp->nc_lock);
463 		mutex_exit(&cpup->cpu_lock);
464 		return ENOENT;
465 	}
466 	if (vtryget(vp)) {
467 		mutex_exit(&ncp->nc_lock);
468 		mutex_exit(&cpup->cpu_lock);
469 	} else {
470 		mutex_enter(&vp->v_interlock);
471 		mutex_exit(&ncp->nc_lock);
472 		mutex_exit(&cpup->cpu_lock);
473 		error = vget(vp, LK_NOWAIT);
474 		if (error) {
475 			KASSERT(error == EBUSY);
476 			/*
477 			 * This vnode is being cleaned out.
478 			 * XXX badhits?
479 			 */
480 			COUNT(cpup->cpu_stats, ncs_falsehits);
481 			*vpp = NULL;
482 			return -1;
483 		}
484 	}
485 
486 	/* Unlocked, but only for stats. */
487 	COUNT(cpup->cpu_stats, ncs_goodhits); /* XXX can be "badhits" */
488 	*vpp = vp;
489 	return 0;
490 }
491 
492 /*
493  * Scan cache looking for name of directory entry pointing at vp.
494  *
495  * If the lookup succeeds the vnode is referenced and stored in dvpp.
496  *
497  * If bufp is non-NULL, also place the name in the buffer which starts
498  * at bufp, immediately before *bpp, and move bpp backwards to point
499  * at the start of it.  (Yes, this is a little baroque, but it's done
500  * this way to cater to the whims of getcwd).
501  *
502  * Returns 0 on success, -1 on cache miss, positive errno on failure.
503  */
504 int
505 cache_revlookup(struct vnode *vp, struct vnode **dvpp, char **bpp, char *bufp)
506 {
507 	struct namecache *ncp;
508 	struct vnode *dvp;
509 	struct ncvhashhead *nvcpp;
510 	char *bp;
511 	int error, nlen;
512 
513 	if (!doingcache)
514 		goto out;
515 
516 	nvcpp = &ncvhashtbl[NCVHASH(vp)];
517 
518 	mutex_enter(namecache_lock);
519 	LIST_FOREACH(ncp, nvcpp, nc_vhash) {
520 		mutex_enter(&ncp->nc_lock);
521 		if (ncp->nc_vp == vp &&
522 		    (dvp = ncp->nc_dvp) != NULL &&
523 		    dvp != vp) { 		/* avoid pesky . entries.. */
524 
525 #ifdef DIAGNOSTIC
526 			if (ncp->nc_nlen == 1 &&
527 			    ncp->nc_name[0] == '.')
528 				panic("cache_revlookup: found entry for .");
529 
530 			if (ncp->nc_nlen == 2 &&
531 			    ncp->nc_name[0] == '.' &&
532 			    ncp->nc_name[1] == '.')
533 				panic("cache_revlookup: found entry for ..");
534 #endif
535 			COUNT(nchstats, ncs_revhits);
536 			nlen = ncp->nc_nlen;
537 
538 			if (bufp) {
539 				bp = *bpp;
540 				bp -= nlen;
541 				if (bp <= bufp) {
542 					*dvpp = NULL;
543 					mutex_exit(&ncp->nc_lock);
544 					mutex_exit(namecache_lock);
545 					return (ERANGE);
546 				}
547 				memcpy(bp, ncp->nc_name, nlen);
548 				*bpp = bp;
549 			}
550 
551 			if (vtryget(dvp)) {
552 				mutex_exit(&ncp->nc_lock);
553 				mutex_exit(namecache_lock);
554 			} else {
555 				mutex_enter(&dvp->v_interlock);
556 				mutex_exit(&ncp->nc_lock);
557 				mutex_exit(namecache_lock);
558 				error = vget(dvp, LK_NOWAIT);
559 				if (error) {
560 					KASSERT(error == EBUSY);
561 					if (bufp)
562 						(*bpp) += nlen;
563 					*dvpp = NULL;
564 					return -1;
565 				}
566 			}
567 			*dvpp = dvp;
568 			return (0);
569 		}
570 		mutex_exit(&ncp->nc_lock);
571 	}
572 	COUNT(nchstats, ncs_revmiss);
573 	mutex_exit(namecache_lock);
574  out:
575 	*dvpp = NULL;
576 	return (-1);
577 }
578 
579 /*
580  * Add an entry to the cache
581  */
582 void
583 cache_enter(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
584 {
585 	struct namecache *ncp;
586 	struct namecache *oncp;
587 	struct nchashhead *ncpp;
588 	struct ncvhashhead *nvcpp;
589 
590 #ifdef DIAGNOSTIC
591 	if (cnp->cn_namelen > NCHNAMLEN)
592 		panic("cache_enter: name too long");
593 #endif
594 	if (!doingcache)
595 		return;
596 
597 	if (numcache > desiredvnodes) {
598 		mutex_enter(namecache_lock);
599 		cache_ev_forced.ev_count++;
600 		cache_reclaim();
601 		mutex_exit(namecache_lock);
602 	}
603 
604 	ncp = pool_cache_get(namecache_cache, PR_WAITOK);
605 	mutex_enter(namecache_lock);
606 	numcache++;
607 
608 	/*
609 	 * Concurrent lookups in the same directory may race for a
610 	 * cache entry.  if there's a duplicated entry, free it.
611 	 */
612 	oncp = cache_lookup_entry(dvp, cnp);
613 	if (oncp) {
614 		cache_invalidate(oncp);
615 		mutex_exit(&oncp->nc_lock);
616 	}
617 
618 	/* Grab the vnode we just found. */
619 	mutex_enter(&ncp->nc_lock);
620 	ncp->nc_vp = vp;
621 	ncp->nc_flags = 0;
622 	ncp->nc_hittime = 0;
623 	ncp->nc_gcqueue = NULL;
624 	if (vp == NULL) {
625 		/*
626 		 * For negative hits, save the ISWHITEOUT flag so we can
627 		 * restore it later when the cache entry is used again.
628 		 */
629 		ncp->nc_flags = cnp->cn_flags & ISWHITEOUT;
630 	}
631 	/* Fill in cache info. */
632 	ncp->nc_dvp = dvp;
633 	LIST_INSERT_HEAD(&dvp->v_dnclist, ncp, nc_dvlist);
634 	if (vp)
635 		LIST_INSERT_HEAD(&vp->v_nclist, ncp, nc_vlist);
636 	else {
637 		ncp->nc_vlist.le_prev = NULL;
638 		ncp->nc_vlist.le_next = NULL;
639 	}
640 	ncp->nc_nlen = cnp->cn_namelen;
641 	TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
642 	memcpy(ncp->nc_name, cnp->cn_nameptr, (unsigned)ncp->nc_nlen);
643 	ncpp = &nchashtbl[NCHASH(cnp, dvp)];
644 
645 	/*
646 	 * Flush updates before making visible in table.  No need for a
647 	 * memory barrier on the other side: to see modifications the
648 	 * list must be followed, meaning a dependent pointer load.
649 	 * The below is LIST_INSERT_HEAD() inlined, with the memory
650 	 * barrier included in the correct place.
651 	 */
652 	if ((ncp->nc_hash.le_next = ncpp->lh_first) != NULL)
653 		ncpp->lh_first->nc_hash.le_prev = &ncp->nc_hash.le_next;
654 	ncp->nc_hash.le_prev = &ncpp->lh_first;
655 	membar_producer();
656 	ncpp->lh_first = ncp;
657 
658 	ncp->nc_vhash.le_prev = NULL;
659 	ncp->nc_vhash.le_next = NULL;
660 
661 	/*
662 	 * Create reverse-cache entries (used in getcwd) for directories.
663 	 * (and in linux procfs exe node)
664 	 */
665 	if (vp != NULL &&
666 	    vp != dvp &&
667 #ifndef NAMECACHE_ENTER_REVERSE
668 	    vp->v_type == VDIR &&
669 #endif
670 	    (ncp->nc_nlen > 2 ||
671 	    (ncp->nc_nlen > 1 && ncp->nc_name[1] != '.') ||
672 	    (/* ncp->nc_nlen > 0 && */ ncp->nc_name[0] != '.'))) {
673 		nvcpp = &ncvhashtbl[NCVHASH(vp)];
674 		LIST_INSERT_HEAD(nvcpp, ncp, nc_vhash);
675 	}
676 	mutex_exit(&ncp->nc_lock);
677 	mutex_exit(namecache_lock);
678 }
679 
680 /*
681  * Name cache initialization, from vfs_init() when we are booting
682  */
683 void
684 nchinit(void)
685 {
686 	int error;
687 
688 	namecache_cache = pool_cache_init(sizeof(struct namecache),
689 	    coherency_unit, 0, 0, "ncache", NULL, IPL_NONE, cache_ctor,
690 	    cache_dtor, NULL);
691 	KASSERT(namecache_cache != NULL);
692 
693 	namecache_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_NONE);
694 
695 	nchashtbl = hashinit(desiredvnodes, HASH_LIST, true, &nchash);
696 	ncvhashtbl =
697 #ifdef NAMECACHE_ENTER_REVERSE
698 	    hashinit(desiredvnodes, HASH_LIST, true, &ncvhash);
699 #else
700 	    hashinit(desiredvnodes/8, HASH_LIST, true, &ncvhash);
701 #endif
702 
703 	error = kthread_create(PRI_VM, KTHREAD_MPSAFE, NULL, cache_thread,
704 	    NULL, NULL, "cachegc");
705 	if (error != 0)
706 		panic("nchinit %d", error);
707 
708 	evcnt_attach_dynamic(&cache_ev_scan, EVCNT_TYPE_MISC, NULL,
709 	   "namecache", "entries scanned");
710 	evcnt_attach_dynamic(&cache_ev_gc, EVCNT_TYPE_MISC, NULL,
711 	   "namecache", "entries collected");
712 	evcnt_attach_dynamic(&cache_ev_over, EVCNT_TYPE_MISC, NULL,
713 	   "namecache", "over scan target");
714 	evcnt_attach_dynamic(&cache_ev_under, EVCNT_TYPE_MISC, NULL,
715 	   "namecache", "under scan target");
716 	evcnt_attach_dynamic(&cache_ev_forced, EVCNT_TYPE_MISC, NULL,
717 	   "namecache", "forced reclaims");
718 }
719 
720 static int
721 cache_ctor(void *arg, void *obj, int flag)
722 {
723 	struct namecache *ncp;
724 
725 	ncp = obj;
726 	mutex_init(&ncp->nc_lock, MUTEX_DEFAULT, IPL_NONE);
727 
728 	return 0;
729 }
730 
731 static void
732 cache_dtor(void *arg, void *obj)
733 {
734 	struct namecache *ncp;
735 
736 	ncp = obj;
737 	mutex_destroy(&ncp->nc_lock);
738 }
739 
740 /*
741  * Called once for each CPU in the system as attached.
742  */
743 void
744 cache_cpu_init(struct cpu_info *ci)
745 {
746 	struct nchcpu *cpup;
747 	size_t sz;
748 
749 	sz = roundup2(sizeof(*cpup), coherency_unit) + coherency_unit;
750 	cpup = kmem_zalloc(sz, KM_SLEEP);
751 	cpup = (void *)roundup2((uintptr_t)cpup, coherency_unit);
752 	mutex_init(&cpup->cpu_lock, MUTEX_DEFAULT, IPL_NONE);
753 	ci->ci_data.cpu_nch = cpup;
754 }
755 
756 /*
757  * Name cache reinitialization, for when the maximum number of vnodes increases.
758  */
759 void
760 nchreinit(void)
761 {
762 	struct namecache *ncp;
763 	struct nchashhead *oldhash1, *hash1;
764 	struct ncvhashhead *oldhash2, *hash2;
765 	u_long i, oldmask1, oldmask2, mask1, mask2;
766 
767 	hash1 = hashinit(desiredvnodes, HASH_LIST, true, &mask1);
768 	hash2 =
769 #ifdef NAMECACHE_ENTER_REVERSE
770 	    hashinit(desiredvnodes, HASH_LIST, true, &mask2);
771 #else
772 	    hashinit(desiredvnodes/8, HASH_LIST, true, &mask2);
773 #endif
774 	mutex_enter(namecache_lock);
775 	cache_lock_cpus();
776 	oldhash1 = nchashtbl;
777 	oldmask1 = nchash;
778 	nchashtbl = hash1;
779 	nchash = mask1;
780 	oldhash2 = ncvhashtbl;
781 	oldmask2 = ncvhash;
782 	ncvhashtbl = hash2;
783 	ncvhash = mask2;
784 	for (i = 0; i <= oldmask1; i++) {
785 		while ((ncp = LIST_FIRST(&oldhash1[i])) != NULL) {
786 			LIST_REMOVE(ncp, nc_hash);
787 			ncp->nc_hash.le_prev = NULL;
788 		}
789 	}
790 	for (i = 0; i <= oldmask2; i++) {
791 		while ((ncp = LIST_FIRST(&oldhash2[i])) != NULL) {
792 			LIST_REMOVE(ncp, nc_vhash);
793 			ncp->nc_vhash.le_prev = NULL;
794 		}
795 	}
796 	cache_unlock_cpus();
797 	mutex_exit(namecache_lock);
798 	hashdone(oldhash1, HASH_LIST, oldmask1);
799 	hashdone(oldhash2, HASH_LIST, oldmask2);
800 }
801 
802 /*
803  * Cache flush, a particular vnode; called when a vnode is renamed to
804  * hide entries that would now be invalid
805  */
806 void
807 cache_purge1(struct vnode *vp, const struct componentname *cnp, int flags)
808 {
809 	struct namecache *ncp, *ncnext;
810 
811 	mutex_enter(namecache_lock);
812 	if (flags & PURGE_PARENTS) {
813 		for (ncp = LIST_FIRST(&vp->v_nclist); ncp != NULL;
814 		    ncp = ncnext) {
815 			ncnext = LIST_NEXT(ncp, nc_vlist);
816 			mutex_enter(&ncp->nc_lock);
817 			cache_invalidate(ncp);
818 			mutex_exit(&ncp->nc_lock);
819 			cache_disassociate(ncp);
820 		}
821 	}
822 	if (flags & PURGE_CHILDREN) {
823 		for (ncp = LIST_FIRST(&vp->v_dnclist); ncp != NULL;
824 		    ncp = ncnext) {
825 			ncnext = LIST_NEXT(ncp, nc_dvlist);
826 			mutex_enter(&ncp->nc_lock);
827 			cache_invalidate(ncp);
828 			mutex_exit(&ncp->nc_lock);
829 			cache_disassociate(ncp);
830 		}
831 	}
832 	if (cnp != NULL) {
833 		ncp = cache_lookup_entry(vp, cnp);
834 		if (ncp) {
835 			cache_invalidate(ncp);
836 			mutex_exit(&ncp->nc_lock);
837 			cache_disassociate(ncp);
838 		}
839 	}
840 	mutex_exit(namecache_lock);
841 }
842 
843 /*
844  * Cache flush, a whole filesystem; called when filesys is umounted to
845  * remove entries that would now be invalid.
846  */
847 void
848 cache_purgevfs(struct mount *mp)
849 {
850 	struct namecache *ncp, *nxtcp;
851 
852 	mutex_enter(namecache_lock);
853 	for (ncp = TAILQ_FIRST(&nclruhead); ncp != NULL; ncp = nxtcp) {
854 		nxtcp = TAILQ_NEXT(ncp, nc_lru);
855 		mutex_enter(&ncp->nc_lock);
856 		if (ncp->nc_dvp != NULL && ncp->nc_dvp->v_mount == mp) {
857 			/* Free the resources we had. */
858 			cache_invalidate(ncp);
859 			cache_disassociate(ncp);
860 		}
861 		mutex_exit(&ncp->nc_lock);
862 	}
863 	cache_reclaim();
864 	mutex_exit(namecache_lock);
865 }
866 
867 /*
868  * Scan global list invalidating entries until we meet a preset target.
869  * Prefer to invalidate entries that have not scored a hit within
870  * cache_hottime seconds.  We sort the LRU list only for this routine's
871  * benefit.
872  */
873 static void
874 cache_prune(int incache, int target)
875 {
876 	struct namecache *ncp, *nxtcp, *sentinel;
877 	int items, recent, tryharder;
878 
879 	KASSERT(mutex_owned(namecache_lock));
880 
881 	items = 0;
882 	tryharder = 0;
883 	recent = hardclock_ticks - hz * cache_hottime;
884 	sentinel = NULL;
885 	for (ncp = TAILQ_FIRST(&nclruhead); ncp != NULL; ncp = nxtcp) {
886 		if (incache <= target)
887 			break;
888 		items++;
889 		nxtcp = TAILQ_NEXT(ncp, nc_lru);
890 		if (ncp->nc_dvp == NULL)
891 			continue;
892 		if (ncp == sentinel) {
893 			/*
894 			 * If we looped back on ourself, then ignore
895 			 * recent entries and purge whatever we find.
896 			 */
897 			tryharder = 1;
898 		}
899 		if (!tryharder && (ncp->nc_hittime - recent) > 0) {
900 			if (sentinel == NULL)
901 				sentinel = ncp;
902 			TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
903 			TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
904 			continue;
905 		}
906 		mutex_enter(&ncp->nc_lock);
907 		if (ncp->nc_dvp != NULL) {
908 			cache_invalidate(ncp);
909 			cache_disassociate(ncp);
910 			incache--;
911 		}
912 		mutex_exit(&ncp->nc_lock);
913 	}
914 	cache_ev_scan.ev_count += items;
915 }
916 
917 /*
918  * Collect dead cache entries from all CPUs and garbage collect.
919  */
920 static void
921 cache_reclaim(void)
922 {
923 	struct namecache *ncp, *next;
924 	int items;
925 
926 	KASSERT(mutex_owned(namecache_lock));
927 
928 	/*
929 	 * If the number of extant entries not awaiting garbage collection
930 	 * exceeds the high water mark, then reclaim stale entries until we
931 	 * reach our low water mark.
932 	 */
933 	items = numcache - cache_gcpend;
934 	if (items > (uint64_t)desiredvnodes * cache_hiwat / 100) {
935 		cache_prune(items, (int)((uint64_t)desiredvnodes *
936 		    cache_lowat / 100));
937 		cache_ev_over.ev_count++;
938 	} else
939 		cache_ev_under.ev_count++;
940 
941 	/*
942 	 * Stop forward lookup activity on all CPUs and garbage collect dead
943 	 * entries.
944 	 */
945 	cache_lock_cpus();
946 	ncp = cache_gcqueue;
947 	cache_gcqueue = NULL;
948 	items = cache_gcpend;
949 	cache_gcpend = 0;
950 	while (ncp != NULL) {
951 		next = ncp->nc_gcqueue;
952 		cache_disassociate(ncp);
953 		KASSERT(ncp->nc_dvp == NULL);
954 		if (ncp->nc_hash.le_prev != NULL) {
955 			LIST_REMOVE(ncp, nc_hash);
956 			ncp->nc_hash.le_prev = NULL;
957 		}
958 		pool_cache_put(namecache_cache, ncp);
959 		ncp = next;
960 	}
961 	cache_unlock_cpus();
962 	numcache -= items;
963 	cache_ev_gc.ev_count += items;
964 }
965 
966 /*
967  * Cache maintainence thread, awakening once per second to:
968  *
969  * => keep number of entries below the high water mark
970  * => sort pseudo-LRU list
971  * => garbage collect dead entries
972  */
973 static void
974 cache_thread(void *arg)
975 {
976 
977 	mutex_enter(namecache_lock);
978 	for (;;) {
979 		cache_reclaim();
980 		kpause("cachegc", false, hz, namecache_lock);
981 	}
982 }
983 
984 #ifdef DDB
985 void
986 namecache_print(struct vnode *vp, void (*pr)(const char *, ...))
987 {
988 	struct vnode *dvp = NULL;
989 	struct namecache *ncp;
990 
991 	TAILQ_FOREACH(ncp, &nclruhead, nc_lru) {
992 		if (ncp->nc_vp == vp && ncp->nc_dvp != NULL) {
993 			(*pr)("name %.*s\n", ncp->nc_nlen, ncp->nc_name);
994 			dvp = ncp->nc_dvp;
995 		}
996 	}
997 	if (dvp == NULL) {
998 		(*pr)("name not found\n");
999 		return;
1000 	}
1001 	vp = dvp;
1002 	TAILQ_FOREACH(ncp, &nclruhead, nc_lru) {
1003 		if (ncp->nc_vp == vp) {
1004 			(*pr)("parent %.*s\n", ncp->nc_nlen, ncp->nc_name);
1005 		}
1006 	}
1007 }
1008 #endif
1009