xref: /netbsd-src/sys/kern/vfs_cache.c (revision cac8e449158efc7261bebc8657cbb0125a2cfdde)
1 /*	$NetBSD: vfs_cache.c,v 1.77 2008/06/03 15:50:22 ad 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 1.77 2008/06/03 15:50:22 ad 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 MALLOC_DEFINE(M_CACHE, "namecache", "Dynamically allocated cache entries");
135 
136 int cache_lowat = 95;
137 int cache_hiwat = 98;
138 int cache_hottime = 5;			/* number of seconds */
139 int doingcache = 1;			/* 1 => enable the cache */
140 
141 static struct evcnt cache_ev_scan;
142 static struct evcnt cache_ev_gc;
143 static struct evcnt cache_ev_over;
144 static struct evcnt cache_ev_under;
145 static struct evcnt cache_ev_forced;
146 
147 /* A single lock to serialize modifications. */
148 static kmutex_t *namecache_lock;
149 
150 static void cache_invalidate(struct namecache *);
151 static inline struct namecache *cache_lookup_entry(
152     const struct vnode *, const struct componentname *);
153 static void cache_thread(void *);
154 static void cache_invalidate(struct namecache *);
155 static void cache_disassociate(struct namecache *);
156 static void cache_reclaim(void);
157 static int cache_ctor(void *, void *, int);
158 static void cache_dtor(void *, void *);
159 
160 /*
161  * Invalidate a cache entry and enqueue it for garbage collection.
162  */
163 static void
164 cache_invalidate(struct namecache *ncp)
165 {
166 	void *head;
167 
168 	KASSERT(mutex_owned(&ncp->nc_lock));
169 
170 	if (ncp->nc_dvp != NULL) {
171 		ncp->nc_vp = NULL;
172 		ncp->nc_dvp = NULL;
173 		do {
174 			head = cache_gcqueue;
175 			ncp->nc_gcqueue = head;
176 		} while (atomic_cas_ptr(&cache_gcqueue, head, ncp) != head);
177 		atomic_inc_uint(&cache_gcpend);
178 	}
179 }
180 
181 /*
182  * Disassociate a namecache entry from any vnodes it is attached to,
183  * and remove from the global LRU list.
184  */
185 static void
186 cache_disassociate(struct namecache *ncp)
187 {
188 
189 	KASSERT(mutex_owned(namecache_lock));
190 	KASSERT(ncp->nc_dvp == NULL);
191 
192 	if (ncp->nc_lru.tqe_prev != NULL) {
193 		TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
194 		ncp->nc_lru.tqe_prev = NULL;
195 	}
196 	if (ncp->nc_vhash.le_prev != NULL) {
197 		LIST_REMOVE(ncp, nc_vhash);
198 		ncp->nc_vhash.le_prev = NULL;
199 	}
200 	if (ncp->nc_vlist.le_prev != NULL) {
201 		LIST_REMOVE(ncp, nc_vlist);
202 		ncp->nc_vlist.le_prev = NULL;
203 	}
204 	if (ncp->nc_dvlist.le_prev != NULL) {
205 		LIST_REMOVE(ncp, nc_dvlist);
206 		ncp->nc_dvlist.le_prev = NULL;
207 	}
208 }
209 
210 /*
211  * Lock all CPUs to prevent any cache lookup activity.  Conceptually,
212  * this locks out all "readers".
213  */
214 static void
215 cache_lock_cpus(void)
216 {
217 	CPU_INFO_ITERATOR cii;
218 	struct cpu_info *ci;
219 	struct nchcpu *cpup;
220 	long *s, *d, *m;
221 
222 	for (CPU_INFO_FOREACH(cii, ci)) {
223 		cpup = ci->ci_data.cpu_nch;
224 		mutex_enter(&cpup->cpu_lock);
225 
226 		/* Collate statistics. */
227 		d = (long *)&nchstats;
228 		s = (long *)&cpup->cpu_stats;
229 		m = s + sizeof(nchstats) / sizeof(long);
230 		for (; s < m; s++, d++) {
231 			*d += *s;
232 			*s = 0;
233 		}
234 	}
235 }
236 
237 /*
238  * Release all CPU locks.
239  */
240 static void
241 cache_unlock_cpus(void)
242 {
243 	CPU_INFO_ITERATOR cii;
244 	struct cpu_info *ci;
245 	struct nchcpu *cpup;
246 
247 	for (CPU_INFO_FOREACH(cii, ci)) {
248 		cpup = ci->ci_data.cpu_nch;
249 		mutex_exit(&cpup->cpu_lock);
250 	}
251 }
252 
253 /*
254  * Find a single cache entry and return it locked.  'namecache_lock' or
255  * at least one of the per-CPU locks must be held.
256  */
257 static struct namecache *
258 cache_lookup_entry(const struct vnode *dvp, const struct componentname *cnp)
259 {
260 	struct nchashhead *ncpp;
261 	struct namecache *ncp;
262 
263 	ncpp = &nchashtbl[NCHASH(cnp, dvp)];
264 
265 	LIST_FOREACH(ncp, ncpp, nc_hash) {
266 		if (ncp->nc_dvp != dvp ||
267 		    ncp->nc_nlen != cnp->cn_namelen ||
268 		    memcmp(ncp->nc_name, cnp->cn_nameptr, (u_int)ncp->nc_nlen))
269 		    	continue;
270 	    	mutex_enter(&ncp->nc_lock);
271 		if (__predict_true(ncp->nc_dvp == dvp)) {
272 			ncp->nc_hittime = hardclock_ticks;
273 			return ncp;
274 		}
275 		/* Raced: entry has been nullified. */
276 		mutex_exit(&ncp->nc_lock);
277 	}
278 
279 	return NULL;
280 }
281 
282 /*
283  * Look for a the name in the cache. We don't do this
284  * if the segment name is long, simply so the cache can avoid
285  * holding long names (which would either waste space, or
286  * add greatly to the complexity).
287  *
288  * Lookup is called with ni_dvp pointing to the directory to search,
289  * ni_ptr pointing to the name of the entry being sought, ni_namelen
290  * tells the length of the name, and ni_hash contains a hash of
291  * the name. If the lookup succeeds, the vnode is locked, stored in ni_vp
292  * and a status of zero is returned. If the locking fails for whatever
293  * reason, the vnode is unlocked and the error is returned to caller.
294  * If the lookup determines that the name does not exist (negative cacheing),
295  * a status of ENOENT is returned. If the lookup fails, a status of -1
296  * is returned.
297  */
298 int
299 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
300 {
301 	struct namecache *ncp;
302 	struct vnode *vp;
303 	struct nchcpu *cpup;
304 	int error;
305 
306 	if (__predict_false(!doingcache)) {
307 		cnp->cn_flags &= ~MAKEENTRY;
308 		*vpp = NULL;
309 		return -1;
310 	}
311 
312 	cpup = curcpu()->ci_data.cpu_nch;
313 	mutex_enter(&cpup->cpu_lock);
314 	if (__predict_false(cnp->cn_namelen > NCHNAMLEN)) {
315 		COUNT(cpup->cpu_stats, ncs_long);
316 		cnp->cn_flags &= ~MAKEENTRY;
317 		mutex_exit(&cpup->cpu_lock);
318 		*vpp = NULL;
319 		return -1;
320 	}
321 	ncp = cache_lookup_entry(dvp, cnp);
322 	if (__predict_false(ncp == NULL)) {
323 		COUNT(cpup->cpu_stats, ncs_miss);
324 		mutex_exit(&cpup->cpu_lock);
325 		*vpp = NULL;
326 		return -1;
327 	}
328 	if ((cnp->cn_flags & MAKEENTRY) == 0) {
329 		COUNT(cpup->cpu_stats, ncs_badhits);
330 		/*
331 		 * Last component and we are renaming or deleting,
332 		 * the cache entry is invalid, or otherwise don't
333 		 * want cache entry to exist.
334 		 */
335 		cache_invalidate(ncp);
336 		mutex_exit(&ncp->nc_lock);
337 		mutex_exit(&cpup->cpu_lock);
338 		*vpp = NULL;
339 		return -1;
340 	} else if (ncp->nc_vp == NULL) {
341 		/*
342 		 * Restore the ISWHITEOUT flag saved earlier.
343 		 */
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 | LK_INTERLOCK);
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, 0);
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 		*vpp = NULL;
413 		return -1;
414 	}
415 
416 	/* Unlocked, but only for stats. */
417 	COUNT(cpup->cpu_stats, ncs_goodhits);
418 	*vpp = vp;
419 	return 0;
420 }
421 
422 int
423 cache_lookup_raw(struct vnode *dvp, struct vnode **vpp,
424     struct componentname *cnp)
425 {
426 	struct namecache *ncp;
427 	struct vnode *vp;
428 	struct nchcpu *cpup;
429 	int error;
430 
431 	if (__predict_false(!doingcache)) {
432 		cnp->cn_flags &= ~MAKEENTRY;
433 		*vpp = NULL;
434 		return (-1);
435 	}
436 
437 	cpup = curcpu()->ci_data.cpu_nch;
438 	mutex_enter(&cpup->cpu_lock);
439 	if (__predict_false(cnp->cn_namelen > NCHNAMLEN)) {
440 		COUNT(cpup->cpu_stats, ncs_long);
441 		cnp->cn_flags &= ~MAKEENTRY;
442 		mutex_exit(&cpup->cpu_lock);
443 		*vpp = NULL;
444 		return -1;
445 	}
446 	ncp = cache_lookup_entry(dvp, cnp);
447 	if (__predict_false(ncp == NULL)) {
448 		COUNT(cpup->cpu_stats, ncs_miss);
449 		mutex_exit(&cpup->cpu_lock);
450 		*vpp = NULL;
451 		return -1;
452 	}
453 	vp = ncp->nc_vp;
454 	if (vp == NULL) {
455 		/*
456 		 * Restore the ISWHITEOUT flag saved earlier.
457 		 */
458 		cnp->cn_flags |= ncp->nc_flags;
459 		COUNT(cpup->cpu_stats, ncs_neghits);
460 		mutex_exit(&ncp->nc_lock);
461 		mutex_exit(&cpup->cpu_lock);
462 		return ENOENT;
463 	}
464 	if (vtryget(vp)) {
465 		mutex_exit(&ncp->nc_lock);
466 		mutex_exit(&cpup->cpu_lock);
467 	} else {
468 		mutex_enter(&vp->v_interlock);
469 		mutex_exit(&ncp->nc_lock);
470 		mutex_exit(&cpup->cpu_lock);
471 		error = vget(vp, LK_NOWAIT | LK_INTERLOCK);
472 		if (error) {
473 			KASSERT(error == EBUSY);
474 			/*
475 			 * This vnode is being cleaned out.
476 			 * XXX badhits?
477 			 */
478 			COUNT(cpup->cpu_stats, ncs_falsehits);
479 			*vpp = NULL;
480 			return -1;
481 		}
482 	}
483 
484 	*vpp = vp;
485 	return 0;
486 }
487 
488 /*
489  * Scan cache looking for name of directory entry pointing at vp.
490  *
491  * Fill in dvpp.
492  *
493  * If bufp is non-NULL, also place the name in the buffer which starts
494  * at bufp, immediately before *bpp, and move bpp backwards to point
495  * at the start of it.  (Yes, this is a little baroque, but it's done
496  * this way to cater to the whims of getcwd).
497  *
498  * Returns 0 on success, -1 on cache miss, positive errno on failure.
499  */
500 int
501 cache_revlookup(struct vnode *vp, struct vnode **dvpp, char **bpp, char *bufp)
502 {
503 	struct namecache *ncp;
504 	struct vnode *dvp;
505 	struct ncvhashhead *nvcpp;
506 	char *bp;
507 
508 	if (!doingcache)
509 		goto out;
510 
511 	nvcpp = &ncvhashtbl[NCVHASH(vp)];
512 
513 	mutex_enter(namecache_lock);
514 	LIST_FOREACH(ncp, nvcpp, nc_vhash) {
515 		mutex_enter(&ncp->nc_lock);
516 		if (ncp->nc_vp == vp &&
517 		    (dvp = ncp->nc_dvp) != NULL &&
518 		    dvp != vp) { 		/* avoid pesky . entries.. */
519 
520 #ifdef DIAGNOSTIC
521 			if (ncp->nc_nlen == 1 &&
522 			    ncp->nc_name[0] == '.')
523 				panic("cache_revlookup: found entry for .");
524 
525 			if (ncp->nc_nlen == 2 &&
526 			    ncp->nc_name[0] == '.' &&
527 			    ncp->nc_name[1] == '.')
528 				panic("cache_revlookup: found entry for ..");
529 #endif
530 			COUNT(nchstats, ncs_revhits);
531 
532 			if (bufp) {
533 				bp = *bpp;
534 				bp -= ncp->nc_nlen;
535 				if (bp <= bufp) {
536 					*dvpp = NULL;
537 					mutex_exit(&ncp->nc_lock);
538 					mutex_exit(namecache_lock);
539 					return (ERANGE);
540 				}
541 				memcpy(bp, ncp->nc_name, ncp->nc_nlen);
542 				*bpp = bp;
543 			}
544 
545 			/* XXX MP: how do we know dvp won't evaporate? */
546 			*dvpp = dvp;
547 			mutex_exit(&ncp->nc_lock);
548 			mutex_exit(namecache_lock);
549 			return (0);
550 		}
551 		mutex_exit(&ncp->nc_lock);
552 	}
553 	COUNT(nchstats, ncs_revmiss);
554 	mutex_exit(namecache_lock);
555  out:
556 	*dvpp = NULL;
557 	return (-1);
558 }
559 
560 /*
561  * Add an entry to the cache
562  */
563 void
564 cache_enter(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
565 {
566 	struct namecache *ncp;
567 	struct namecache *oncp;
568 	struct nchashhead *ncpp;
569 	struct ncvhashhead *nvcpp;
570 
571 #ifdef DIAGNOSTIC
572 	if (cnp->cn_namelen > NCHNAMLEN)
573 		panic("cache_enter: name too long");
574 #endif
575 	if (!doingcache)
576 		return;
577 
578 	if (numcache > desiredvnodes) {
579 		mutex_enter(namecache_lock);
580 		cache_ev_forced.ev_count++;
581 		cache_reclaim();
582 		mutex_exit(namecache_lock);
583 	}
584 
585 	ncp = pool_cache_get(namecache_cache, PR_WAITOK);
586 	mutex_enter(namecache_lock);
587 	numcache++;
588 
589 	/*
590 	 * Concurrent lookups in the same directory may race for a
591 	 * cache entry.  if there's a duplicated entry, free it.
592 	 */
593 	oncp = cache_lookup_entry(dvp, cnp);
594 	if (oncp) {
595 		cache_invalidate(oncp);
596 		mutex_exit(&oncp->nc_lock);
597 	}
598 
599 	/* Grab the vnode we just found. */
600 	mutex_enter(&ncp->nc_lock);
601 	ncp->nc_vp = vp;
602 	ncp->nc_flags = 0;
603 	ncp->nc_hittime = 0;
604 	ncp->nc_gcqueue = NULL;
605 	if (vp == NULL) {
606 		/*
607 		 * For negative hits, save the ISWHITEOUT flag so we can
608 		 * restore it later when the cache entry is used again.
609 		 */
610 		ncp->nc_flags = cnp->cn_flags & ISWHITEOUT;
611 	}
612 	/* Fill in cache info. */
613 	ncp->nc_dvp = dvp;
614 	LIST_INSERT_HEAD(&dvp->v_dnclist, ncp, nc_dvlist);
615 	if (vp)
616 		LIST_INSERT_HEAD(&vp->v_nclist, ncp, nc_vlist);
617 	else {
618 		ncp->nc_vlist.le_prev = NULL;
619 		ncp->nc_vlist.le_next = NULL;
620 	}
621 	ncp->nc_nlen = cnp->cn_namelen;
622 	TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
623 	memcpy(ncp->nc_name, cnp->cn_nameptr, (unsigned)ncp->nc_nlen);
624 	ncpp = &nchashtbl[NCHASH(cnp, dvp)];
625 
626 	/*
627 	 * Flush updates before making visible in table.  No need for a
628 	 * memory barrier on the other side: to see modifications the
629 	 * list must be followed, meaning a dependent pointer load.
630 	 * The below is LIST_INSERT_HEAD() inlined, with the memory
631 	 * barrier included in the correct place.
632 	 */
633 	if ((ncp->nc_hash.le_next = ncpp->lh_first) != NULL)
634 		ncpp->lh_first->nc_hash.le_prev = &ncp->nc_hash.le_next;
635 	ncp->nc_hash.le_prev = &ncpp->lh_first;
636 	membar_producer();
637 	ncpp->lh_first = ncp;
638 
639 	ncp->nc_vhash.le_prev = NULL;
640 	ncp->nc_vhash.le_next = NULL;
641 
642 	/*
643 	 * Create reverse-cache entries (used in getcwd) for directories.
644 	 * (and in linux procfs exe node)
645 	 */
646 	if (vp != NULL &&
647 	    vp != dvp &&
648 #ifndef NAMECACHE_ENTER_REVERSE
649 	    vp->v_type == VDIR &&
650 #endif
651 	    (ncp->nc_nlen > 2 ||
652 	    (ncp->nc_nlen > 1 && ncp->nc_name[1] != '.') ||
653 	    (/* ncp->nc_nlen > 0 && */ ncp->nc_name[0] != '.'))) {
654 		nvcpp = &ncvhashtbl[NCVHASH(vp)];
655 		LIST_INSERT_HEAD(nvcpp, ncp, nc_vhash);
656 	}
657 	mutex_exit(&ncp->nc_lock);
658 	mutex_exit(namecache_lock);
659 }
660 
661 /*
662  * Name cache initialization, from vfs_init() when we are booting
663  */
664 void
665 nchinit(void)
666 {
667 	int error;
668 
669 	namecache_cache = pool_cache_init(sizeof(struct namecache),
670 	    coherency_unit, 0, 0, "ncache", NULL, IPL_NONE, cache_ctor,
671 	    cache_dtor, NULL);
672 	KASSERT(namecache_cache != NULL);
673 
674 	namecache_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_NONE);
675 
676 	nchashtbl = hashinit(desiredvnodes, HASH_LIST, true, &nchash);
677 	ncvhashtbl =
678 #ifdef NAMECACHE_ENTER_REVERSE
679 	    hashinit(desiredvnodes, HASH_LIST, true, &ncvhash);
680 #else
681 	    hashinit(desiredvnodes/8, HASH_LIST, true, &ncvhash);
682 #endif
683 
684 	error = kthread_create(PRI_VM, KTHREAD_MPSAFE, NULL, cache_thread,
685 	    NULL, NULL, "cachegc");
686 	if (error != 0)
687 		panic("nchinit %d", error);
688 
689 	evcnt_attach_dynamic(&cache_ev_scan, EVCNT_TYPE_MISC, NULL,
690 	   "namecache", "entries scanned");
691 	evcnt_attach_dynamic(&cache_ev_gc, EVCNT_TYPE_MISC, NULL,
692 	   "namecache", "entries collected");
693 	evcnt_attach_dynamic(&cache_ev_over, EVCNT_TYPE_MISC, NULL,
694 	   "namecache", "over scan target");
695 	evcnt_attach_dynamic(&cache_ev_under, EVCNT_TYPE_MISC, NULL,
696 	   "namecache", "under scan target");
697 	evcnt_attach_dynamic(&cache_ev_forced, EVCNT_TYPE_MISC, NULL,
698 	   "namecache", "forced reclaims");
699 }
700 
701 static int
702 cache_ctor(void *arg, void *obj, int flag)
703 {
704 	struct namecache *ncp;
705 
706 	ncp = obj;
707 	mutex_init(&ncp->nc_lock, MUTEX_DEFAULT, IPL_NONE);
708 
709 	return 0;
710 }
711 
712 static void
713 cache_dtor(void *arg, void *obj)
714 {
715 	struct namecache *ncp;
716 
717 	ncp = obj;
718 	mutex_destroy(&ncp->nc_lock);
719 }
720 
721 /*
722  * Called once for each CPU in the system as attached.
723  */
724 void
725 cache_cpu_init(struct cpu_info *ci)
726 {
727 	struct nchcpu *cpup;
728 	size_t sz;
729 
730 	sz = roundup2(sizeof(*cpup), coherency_unit) + coherency_unit;
731 	cpup = kmem_zalloc(sz, KM_SLEEP);
732 	cpup = (void *)roundup2((uintptr_t)cpup, coherency_unit);
733 	mutex_init(&cpup->cpu_lock, MUTEX_DEFAULT, IPL_NONE);
734 	ci->ci_data.cpu_nch = cpup;
735 }
736 
737 /*
738  * Name cache reinitialization, for when the maximum number of vnodes increases.
739  */
740 void
741 nchreinit(void)
742 {
743 	struct namecache *ncp;
744 	struct nchashhead *oldhash1, *hash1;
745 	struct ncvhashhead *oldhash2, *hash2;
746 	u_long i, oldmask1, oldmask2, mask1, mask2;
747 
748 	hash1 = hashinit(desiredvnodes, HASH_LIST, true, &mask1);
749 	hash2 =
750 #ifdef NAMECACHE_ENTER_REVERSE
751 	    hashinit(desiredvnodes, HASH_LIST, true, &mask2);
752 #else
753 	    hashinit(desiredvnodes/8, HASH_LIST, true, &mask2);
754 #endif
755 	mutex_enter(namecache_lock);
756 	cache_lock_cpus();
757 	oldhash1 = nchashtbl;
758 	oldmask1 = nchash;
759 	nchashtbl = hash1;
760 	nchash = mask1;
761 	oldhash2 = ncvhashtbl;
762 	oldmask2 = ncvhash;
763 	ncvhashtbl = hash2;
764 	ncvhash = mask2;
765 	for (i = 0; i <= oldmask1; i++) {
766 		while ((ncp = LIST_FIRST(&oldhash1[i])) != NULL) {
767 			LIST_REMOVE(ncp, nc_hash);
768 			ncp->nc_hash.le_prev = NULL;
769 		}
770 	}
771 	for (i = 0; i <= oldmask2; i++) {
772 		while ((ncp = LIST_FIRST(&oldhash2[i])) != NULL) {
773 			LIST_REMOVE(ncp, nc_vhash);
774 			ncp->nc_vhash.le_prev = NULL;
775 		}
776 	}
777 	cache_unlock_cpus();
778 	mutex_exit(namecache_lock);
779 	hashdone(oldhash1, HASH_LIST, oldmask1);
780 	hashdone(oldhash2, HASH_LIST, oldmask2);
781 }
782 
783 /*
784  * Cache flush, a particular vnode; called when a vnode is renamed to
785  * hide entries that would now be invalid
786  */
787 void
788 cache_purge1(struct vnode *vp, const struct componentname *cnp, int flags)
789 {
790 	struct namecache *ncp, *ncnext;
791 
792 	mutex_enter(namecache_lock);
793 	if (flags & PURGE_PARENTS) {
794 		for (ncp = LIST_FIRST(&vp->v_nclist); ncp != NULL;
795 		    ncp = ncnext) {
796 			ncnext = LIST_NEXT(ncp, nc_vlist);
797 			mutex_enter(&ncp->nc_lock);
798 			cache_invalidate(ncp);
799 			mutex_exit(&ncp->nc_lock);
800 			cache_disassociate(ncp);
801 		}
802 	}
803 	if (flags & PURGE_CHILDREN) {
804 		for (ncp = LIST_FIRST(&vp->v_dnclist); ncp != NULL;
805 		    ncp = ncnext) {
806 			ncnext = LIST_NEXT(ncp, nc_dvlist);
807 			mutex_enter(&ncp->nc_lock);
808 			cache_invalidate(ncp);
809 			mutex_exit(&ncp->nc_lock);
810 			cache_disassociate(ncp);
811 		}
812 	}
813 	if (cnp != NULL) {
814 		ncp = cache_lookup_entry(vp, cnp);
815 		if (ncp) {
816 			cache_invalidate(ncp);
817 			cache_disassociate(ncp);
818 			mutex_exit(&ncp->nc_lock);
819 		}
820 	}
821 	mutex_exit(namecache_lock);
822 }
823 
824 /*
825  * Cache flush, a whole filesystem; called when filesys is umounted to
826  * remove entries that would now be invalid.
827  */
828 void
829 cache_purgevfs(struct mount *mp)
830 {
831 	struct namecache *ncp, *nxtcp;
832 
833 	mutex_enter(namecache_lock);
834 	for (ncp = TAILQ_FIRST(&nclruhead); ncp != NULL; ncp = nxtcp) {
835 		nxtcp = TAILQ_NEXT(ncp, nc_lru);
836 		mutex_enter(&ncp->nc_lock);
837 		if (ncp->nc_dvp != NULL && ncp->nc_dvp->v_mount == mp) {
838 			/* Free the resources we had. */
839 			cache_invalidate(ncp);
840 			cache_disassociate(ncp);
841 		}
842 		mutex_exit(&ncp->nc_lock);
843 	}
844 	cache_reclaim();
845 	mutex_exit(namecache_lock);
846 }
847 
848 /*
849  * Scan global list invalidating entries until we meet a preset target.
850  * Prefer to invalidate entries that have not scored a hit within
851  * cache_hottime seconds.  We sort the LRU list only for this routine's
852  * benefit.
853  */
854 static void
855 cache_prune(int incache, int target)
856 {
857 	struct namecache *ncp, *nxtcp, *sentinel;
858 	int items, recent, tryharder;
859 
860 	KASSERT(mutex_owned(namecache_lock));
861 
862 	items = 0;
863 	tryharder = 0;
864 	recent = hardclock_ticks - hz * cache_hottime;
865 	sentinel = NULL;
866 	for (ncp = TAILQ_FIRST(&nclruhead); ncp != NULL; ncp = nxtcp) {
867 		if (incache <= target)
868 			break;
869 		items++;
870 		nxtcp = TAILQ_NEXT(ncp, nc_lru);
871 		if (ncp->nc_dvp == NULL)
872 			continue;
873 		if (ncp == sentinel) {
874 			/*
875 			 * If we looped back on ourself, then ignore
876 			 * recent entries and purge whatever we find.
877 			 */
878 			tryharder = 1;
879 		}
880 		if (!tryharder && ncp->nc_hittime > recent) {
881 			if (sentinel == NULL)
882 				sentinel = ncp;
883 			TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
884 			TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
885 			continue;
886 		}
887 		mutex_enter(&ncp->nc_lock);
888 		if (ncp->nc_dvp != NULL) {
889 			cache_invalidate(ncp);
890 			cache_disassociate(ncp);
891 			incache--;
892 		}
893 		mutex_exit(&ncp->nc_lock);
894 	}
895 	cache_ev_scan.ev_count += items;
896 }
897 
898 /*
899  * Collect dead cache entries from all CPUs and garbage collect.
900  */
901 static void
902 cache_reclaim(void)
903 {
904 	struct namecache *ncp, *next;
905 	int items;
906 
907 	KASSERT(mutex_owned(namecache_lock));
908 
909 	/*
910 	 * If the number of extant entries not awaiting garbage collection
911 	 * exceeds the high water mark, then reclaim stale entries until we
912 	 * reach our low water mark.
913 	 */
914 	items = numcache - cache_gcpend;
915 	if (items > (uint64_t)desiredvnodes * cache_hiwat / 100) {
916 		cache_prune(items, (int)((uint64_t)desiredvnodes *
917 		    cache_lowat / 100));
918 		cache_ev_over.ev_count++;
919 	} else
920 		cache_ev_under.ev_count++;
921 
922 	/*
923 	 * Stop forward lookup activity on all CPUs and garbage collect dead
924 	 * entries.
925 	 */
926 	cache_lock_cpus();
927 	ncp = cache_gcqueue;
928 	cache_gcqueue = NULL;
929 	items = cache_gcpend;
930 	cache_gcpend = 0;
931 	while (ncp != NULL) {
932 		next = ncp->nc_gcqueue;
933 		cache_disassociate(ncp);
934 		KASSERT(ncp->nc_dvp == NULL);
935 		if (ncp->nc_hash.le_prev != NULL) {
936 			LIST_REMOVE(ncp, nc_hash);
937 			ncp->nc_hash.le_prev = NULL;
938 		}
939 		pool_cache_put(namecache_cache, ncp);
940 		ncp = next;
941 	}
942 	cache_unlock_cpus();
943 	numcache -= items;
944 	cache_ev_gc.ev_count += items;
945 }
946 
947 /*
948  * Cache maintainence thread, awakening once per second to:
949  *
950  * => keep number of entries below the high water mark
951  * => sort pseudo-LRU list
952  * => garbage collect dead entries
953  */
954 static void
955 cache_thread(void *arg)
956 {
957 
958 	mutex_enter(namecache_lock);
959 	for (;;) {
960 		cache_reclaim();
961 		kpause("cachegc", false, hz, namecache_lock);
962 	}
963 }
964 
965 #ifdef DDB
966 void
967 namecache_print(struct vnode *vp, void (*pr)(const char *, ...))
968 {
969 	struct vnode *dvp = NULL;
970 	struct namecache *ncp;
971 
972 	TAILQ_FOREACH(ncp, &nclruhead, nc_lru) {
973 		if (ncp->nc_vp == vp && ncp->nc_dvp != NULL) {
974 			(*pr)("name %.*s\n", ncp->nc_nlen, ncp->nc_name);
975 			dvp = ncp->nc_dvp;
976 		}
977 	}
978 	if (dvp == NULL) {
979 		(*pr)("name not found\n");
980 		return;
981 	}
982 	vp = dvp;
983 	TAILQ_FOREACH(ncp, &nclruhead, nc_lru) {
984 		if (ncp->nc_vp == vp) {
985 			(*pr)("parent %.*s\n", ncp->nc_nlen, ncp->nc_name);
986 		}
987 	}
988 }
989 #endif
990