xref: /dflybsd-src/sys/kern/vfs_cache.c (revision 41871674d0079dec70d55eb824f39d07dc7b3310)
1 /*
2  * Copyright (c) 2003,2004 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
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  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  * Copyright (c) 1989, 1993, 1995
35  *	The Regents of the University of California.  All rights reserved.
36  *
37  * This code is derived from software contributed to Berkeley by
38  * Poul-Henning Kamp of the FreeBSD Project.
39  *
40  * Redistribution and use in source and binary forms, with or without
41  * modification, are permitted provided that the following conditions
42  * are met:
43  * 1. Redistributions of source code must retain the above copyright
44  *    notice, this list of conditions and the following disclaimer.
45  * 2. Redistributions in binary form must reproduce the above copyright
46  *    notice, this list of conditions and the following disclaimer in the
47  *    documentation and/or other materials provided with the distribution.
48  * 3. All advertising materials mentioning features or use of this software
49  *    must display the following acknowledgement:
50  *	This product includes software developed by the University of
51  *	California, Berkeley and its contributors.
52  * 4. Neither the name of the University nor the names of its contributors
53  *    may be used to endorse or promote products derived from this software
54  *    without specific prior written permission.
55  *
56  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
57  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
58  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
59  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
60  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
61  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
62  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
63  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
64  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
65  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
66  * SUCH DAMAGE.
67  *
68  *	@(#)vfs_cache.c	8.5 (Berkeley) 3/22/95
69  * $FreeBSD: src/sys/kern/vfs_cache.c,v 1.42.2.6 2001/10/05 20:07:03 dillon Exp $
70  * $DragonFly: src/sys/kern/vfs_cache.c,v 1.62 2006/03/30 02:39:46 dillon Exp $
71  */
72 
73 #include <sys/param.h>
74 #include <sys/systm.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/mount.h>
78 #include <sys/vnode.h>
79 #include <sys/malloc.h>
80 #include <sys/sysproto.h>
81 #include <sys/proc.h>
82 #include <sys/namei.h>
83 #include <sys/nlookup.h>
84 #include <sys/filedesc.h>
85 #include <sys/fnv_hash.h>
86 #include <sys/globaldata.h>
87 #include <sys/kern_syscall.h>
88 #include <sys/dirent.h>
89 #include <ddb/ddb.h>
90 
91 /*
92  * Random lookups in the cache are accomplished with a hash table using
93  * a hash key of (nc_src_vp, name).
94  *
95  * Negative entries may exist and correspond to structures where nc_vp
96  * is NULL.  In a negative entry, NCF_WHITEOUT will be set if the entry
97  * corresponds to a whited-out directory entry (verses simply not finding the
98  * entry at all).
99  *
100  * Upon reaching the last segment of a path, if the reference is for DELETE,
101  * or NOCACHE is set (rewrite), and the name is located in the cache, it
102  * will be dropped.
103  */
104 
105 /*
106  * Structures associated with name cacheing.
107  */
108 #define NCHHASH(hash)	(&nchashtbl[(hash) & nchash])
109 #define MINNEG		1024
110 
111 MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
112 
113 static LIST_HEAD(nchashhead, namecache) *nchashtbl;	/* Hash Table */
114 static struct namecache_list	ncneglist;		/* instead of vnode */
115 
116 /*
117  * ncvp_debug - debug cache_fromvp().  This is used by the NFS server
118  * to create the namecache infrastructure leading to a dangling vnode.
119  *
120  * 0	Only errors are reported
121  * 1	Successes are reported
122  * 2	Successes + the whole directory scan is reported
123  * 3	Force the directory scan code run as if the parent vnode did not
124  *	have a namecache record, even if it does have one.
125  */
126 static int	ncvp_debug;
127 SYSCTL_INT(_debug, OID_AUTO, ncvp_debug, CTLFLAG_RW, &ncvp_debug, 0, "");
128 
129 static u_long	nchash;			/* size of hash table */
130 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0, "");
131 
132 static u_long	ncnegfactor = 16;	/* ratio of negative entries */
133 SYSCTL_ULONG(_debug, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0, "");
134 
135 static int	nclockwarn;		/* warn on locked entries in ticks */
136 SYSCTL_INT(_debug, OID_AUTO, nclockwarn, CTLFLAG_RW, &nclockwarn, 0, "");
137 
138 static u_long	numneg;		/* number of cache entries allocated */
139 SYSCTL_ULONG(_debug, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0, "");
140 
141 static u_long	numcache;		/* number of cache entries allocated */
142 SYSCTL_ULONG(_debug, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0, "");
143 
144 static u_long	numunres;		/* number of unresolved entries */
145 SYSCTL_ULONG(_debug, OID_AUTO, numunres, CTLFLAG_RD, &numunres, 0, "");
146 
147 SYSCTL_INT(_debug, OID_AUTO, vnsize, CTLFLAG_RD, 0, sizeof(struct vnode), "");
148 SYSCTL_INT(_debug, OID_AUTO, ncsize, CTLFLAG_RD, 0, sizeof(struct namecache), "");
149 
150 static int cache_resolve_mp(struct namecache *ncp);
151 static void cache_rehash(struct namecache *ncp);
152 
153 /*
154  * The new name cache statistics
155  */
156 SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics");
157 #define STATNODE(mode, name, var) \
158 	SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, mode, var, 0, "");
159 STATNODE(CTLFLAG_RD, numneg, &numneg);
160 STATNODE(CTLFLAG_RD, numcache, &numcache);
161 static u_long numcalls; STATNODE(CTLFLAG_RD, numcalls, &numcalls);
162 static u_long dothits; STATNODE(CTLFLAG_RD, dothits, &dothits);
163 static u_long dotdothits; STATNODE(CTLFLAG_RD, dotdothits, &dotdothits);
164 static u_long numchecks; STATNODE(CTLFLAG_RD, numchecks, &numchecks);
165 static u_long nummiss; STATNODE(CTLFLAG_RD, nummiss, &nummiss);
166 static u_long nummisszap; STATNODE(CTLFLAG_RD, nummisszap, &nummisszap);
167 static u_long numposzaps; STATNODE(CTLFLAG_RD, numposzaps, &numposzaps);
168 static u_long numposhits; STATNODE(CTLFLAG_RD, numposhits, &numposhits);
169 static u_long numnegzaps; STATNODE(CTLFLAG_RD, numnegzaps, &numnegzaps);
170 static u_long numneghits; STATNODE(CTLFLAG_RD, numneghits, &numneghits);
171 
172 struct nchstats nchstats[SMP_MAXCPU];
173 /*
174  * Export VFS cache effectiveness statistics to user-land.
175  *
176  * The statistics are left for aggregation to user-land so
177  * neat things can be achieved, like observing per-CPU cache
178  * distribution.
179  */
180 static int
181 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
182 {
183 	struct globaldata *gd;
184 	int i, error;
185 
186 	error = 0;
187 	for (i = 0; i < ncpus; ++i) {
188 		gd = globaldata_find(i);
189 		if ((error = SYSCTL_OUT(req, (void *)&(*gd->gd_nchstats),
190 			sizeof(struct nchstats))))
191 			break;
192 	}
193 
194 	return (error);
195 }
196 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE|CTLFLAG_RD,
197   0, 0, sysctl_nchstats, "S,nchstats", "VFS cache effectiveness statistics");
198 
199 static void cache_zap(struct namecache *ncp);
200 
201 /*
202  * cache_hold() and cache_drop() prevent the premature deletion of a
203  * namecache entry but do not prevent operations (such as zapping) on
204  * that namecache entry.
205  */
206 static __inline
207 struct namecache *
208 _cache_hold(struct namecache *ncp)
209 {
210 	++ncp->nc_refs;
211 	return(ncp);
212 }
213 
214 /*
215  * When dropping an entry, if only one ref remains and the entry has not
216  * been resolved, zap it.  Since the one reference is being dropped the
217  * entry had better not be locked.
218  */
219 static __inline
220 void
221 _cache_drop(struct namecache *ncp)
222 {
223 	KKASSERT(ncp->nc_refs > 0);
224 	if (ncp->nc_refs == 1 &&
225 	    (ncp->nc_flag & NCF_UNRESOLVED) &&
226 	    TAILQ_EMPTY(&ncp->nc_list)
227 	) {
228 		KKASSERT(ncp->nc_exlocks == 0);
229 		cache_lock(ncp);
230 		cache_zap(ncp);
231 	} else {
232 		--ncp->nc_refs;
233 	}
234 }
235 
236 /*
237  * Link a new namecache entry to its parent.  Be careful to avoid races
238  * if vhold() blocks in the future.
239  *
240  * If we are creating a child under an oldapi parent we must mark the
241  * child as being an oldapi entry as well.
242  */
243 static void
244 cache_link_parent(struct namecache *ncp, struct namecache *par)
245 {
246 	KKASSERT(ncp->nc_parent == NULL);
247 	ncp->nc_parent = par;
248 	if (TAILQ_EMPTY(&par->nc_list)) {
249 		TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
250 		/*
251 		 * Any vp associated with an ncp which has children must
252 		 * be held to prevent it from being recycled.
253 		 */
254 		if (par->nc_vp)
255 			vhold(par->nc_vp);
256 	} else {
257 		TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
258 	}
259 }
260 
261 /*
262  * Remove the parent association from a namecache structure.  If this is
263  * the last child of the parent the cache_drop(par) will attempt to
264  * recursively zap the parent.
265  */
266 static void
267 cache_unlink_parent(struct namecache *ncp)
268 {
269 	struct namecache *par;
270 
271 	if ((par = ncp->nc_parent) != NULL) {
272 		ncp->nc_parent = NULL;
273 		par = cache_hold(par);
274 		TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
275 		if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
276 			vdrop(par->nc_vp);
277 		cache_drop(par);
278 	}
279 }
280 
281 /*
282  * Allocate a new namecache structure.  Most of the code does not require
283  * zero-termination of the string but it makes vop_compat_ncreate() easier.
284  */
285 static struct namecache *
286 cache_alloc(int nlen)
287 {
288 	struct namecache *ncp;
289 
290 	ncp = malloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO);
291 	if (nlen)
292 		ncp->nc_name = malloc(nlen + 1, M_VFSCACHE, M_WAITOK);
293 	ncp->nc_nlen = nlen;
294 	ncp->nc_flag = NCF_UNRESOLVED;
295 	ncp->nc_error = ENOTCONN;	/* needs to be resolved */
296 	ncp->nc_refs = 1;
297 	ncp->nc_fsmid = 1;
298 	TAILQ_INIT(&ncp->nc_list);
299 	cache_lock(ncp);
300 	return(ncp);
301 }
302 
303 static void
304 cache_free(struct namecache *ncp)
305 {
306 	KKASSERT(ncp->nc_refs == 1 && ncp->nc_exlocks == 1);
307 	if (ncp->nc_name)
308 		free(ncp->nc_name, M_VFSCACHE);
309 	free(ncp, M_VFSCACHE);
310 }
311 
312 /*
313  * Ref and deref a namecache structure.
314  */
315 struct namecache *
316 cache_hold(struct namecache *ncp)
317 {
318 	return(_cache_hold(ncp));
319 }
320 
321 void
322 cache_drop(struct namecache *ncp)
323 {
324 	_cache_drop(ncp);
325 }
326 
327 /*
328  * Namespace locking.  The caller must already hold a reference to the
329  * namecache structure in order to lock/unlock it.  This function prevents
330  * the namespace from being created or destroyed by accessors other then
331  * the lock holder.
332  *
333  * Note that holding a locked namecache structure prevents other threads
334  * from making namespace changes (e.g. deleting or creating), prevents
335  * vnode association state changes by other threads, and prevents the
336  * namecache entry from being resolved or unresolved by other threads.
337  *
338  * The lock owner has full authority to associate/disassociate vnodes
339  * and resolve/unresolve the locked ncp.
340  *
341  * WARNING!  Holding a locked ncp will prevent a vnode from being destroyed
342  * or recycled, but it does NOT help you if the vnode had already initiated
343  * a recyclement.  If this is important, use cache_get() rather then
344  * cache_lock() (and deal with the differences in the way the refs counter
345  * is handled).  Or, alternatively, make an unconditional call to
346  * cache_validate() or cache_resolve() after cache_lock() returns.
347  */
348 void
349 cache_lock(struct namecache *ncp)
350 {
351 	thread_t td;
352 	int didwarn;
353 
354 	KKASSERT(ncp->nc_refs != 0);
355 	didwarn = 0;
356 	td = curthread;
357 
358 	for (;;) {
359 		if (ncp->nc_exlocks == 0) {
360 			ncp->nc_exlocks = 1;
361 			ncp->nc_locktd = td;
362 			/*
363 			 * The vp associated with a locked ncp must be held
364 			 * to prevent it from being recycled (which would
365 			 * cause the ncp to become unresolved).
366 			 *
367 			 * WARNING!  If VRECLAIMED is set the vnode could
368 			 * already be in the middle of a recycle.  Callers
369 			 * should not assume that nc_vp is usable when
370 			 * not NULL.  cache_vref() or cache_vget() must be
371 			 * called.
372 			 *
373 			 * XXX loop on race for later MPSAFE work.
374 			 */
375 			if (ncp->nc_vp)
376 				vhold(ncp->nc_vp);
377 			break;
378 		}
379 		if (ncp->nc_locktd == td) {
380 			++ncp->nc_exlocks;
381 			break;
382 		}
383 		ncp->nc_flag |= NCF_LOCKREQ;
384 		if (tsleep(ncp, 0, "clock", nclockwarn) == EWOULDBLOCK) {
385 			if (didwarn)
386 				continue;
387 			didwarn = 1;
388 			printf("[diagnostic] cache_lock: blocked on %p", ncp);
389 			if ((ncp->nc_flag & NCF_MOUNTPT) && ncp->nc_mount)
390 			    printf(" [MOUNTFROM %s]\n", ncp->nc_mount->mnt_stat.f_mntfromname);
391 			else
392 			    printf(" \"%*.*s\"\n",
393 				ncp->nc_nlen, ncp->nc_nlen,
394 				ncp->nc_name);
395 		}
396 	}
397 
398 	if (didwarn == 1) {
399 		printf("[diagnostic] cache_lock: unblocked %*.*s\n",
400 			ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
401 	}
402 }
403 
404 int
405 cache_lock_nonblock(struct namecache *ncp)
406 {
407 	thread_t td;
408 
409 	KKASSERT(ncp->nc_refs != 0);
410 	td = curthread;
411 	if (ncp->nc_exlocks == 0) {
412 		ncp->nc_exlocks = 1;
413 		ncp->nc_locktd = td;
414 		/*
415 		 * The vp associated with a locked ncp must be held
416 		 * to prevent it from being recycled (which would
417 		 * cause the ncp to become unresolved).
418 		 *
419 		 * WARNING!  If VRECLAIMED is set the vnode could
420 		 * already be in the middle of a recycle.  Callers
421 		 * should not assume that nc_vp is usable when
422 		 * not NULL.  cache_vref() or cache_vget() must be
423 		 * called.
424 		 *
425 		 * XXX loop on race for later MPSAFE work.
426 		 */
427 		if (ncp->nc_vp)
428 			vhold(ncp->nc_vp);
429 		return(0);
430 	} else {
431 		return(EWOULDBLOCK);
432 	}
433 }
434 
435 void
436 cache_unlock(struct namecache *ncp)
437 {
438 	thread_t td = curthread;
439 
440 	KKASSERT(ncp->nc_refs > 0);
441 	KKASSERT(ncp->nc_exlocks > 0);
442 	KKASSERT(ncp->nc_locktd == td);
443 	if (--ncp->nc_exlocks == 0) {
444 		if (ncp->nc_vp)
445 			vdrop(ncp->nc_vp);
446 		ncp->nc_locktd = NULL;
447 		if (ncp->nc_flag & NCF_LOCKREQ) {
448 			ncp->nc_flag &= ~NCF_LOCKREQ;
449 			wakeup(ncp);
450 		}
451 	}
452 }
453 
454 /*
455  * ref-and-lock, unlock-and-deref functions.
456  *
457  * This function is primarily used by nlookup.  Even though cache_lock
458  * holds the vnode, it is possible that the vnode may have already
459  * initiated a recyclement.  We want cache_get() to return a definitively
460  * usable vnode or a definitively unresolved ncp.
461  */
462 struct namecache *
463 cache_get(struct namecache *ncp)
464 {
465 	_cache_hold(ncp);
466 	cache_lock(ncp);
467 	if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
468 		cache_setunresolved(ncp);
469 	return(ncp);
470 }
471 
472 int
473 cache_get_nonblock(struct namecache *ncp)
474 {
475 	/* XXX MP */
476 	if (ncp->nc_exlocks == 0 || ncp->nc_locktd == curthread) {
477 		_cache_hold(ncp);
478 		cache_lock(ncp);
479 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
480 			cache_setunresolved(ncp);
481 		return(0);
482 	}
483 	return(EWOULDBLOCK);
484 }
485 
486 void
487 cache_put(struct namecache *ncp)
488 {
489 	cache_unlock(ncp);
490 	_cache_drop(ncp);
491 }
492 
493 /*
494  * Resolve an unresolved ncp by associating a vnode with it.  If the
495  * vnode is NULL, a negative cache entry is created.
496  *
497  * The ncp should be locked on entry and will remain locked on return.
498  */
499 void
500 cache_setvp(struct namecache *ncp, struct vnode *vp)
501 {
502 	KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
503 	ncp->nc_vp = vp;
504 	if (vp != NULL) {
505 		/*
506 		 * Any vp associated with an ncp which has children must
507 		 * be held.  Any vp associated with a locked ncp must be held.
508 		 */
509 		if (!TAILQ_EMPTY(&ncp->nc_list))
510 			vhold(vp);
511 		TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode);
512 		if (ncp->nc_exlocks)
513 			vhold(vp);
514 
515 		/*
516 		 * Set auxillary flags
517 		 */
518 		switch(vp->v_type) {
519 		case VDIR:
520 			ncp->nc_flag |= NCF_ISDIR;
521 			break;
522 		case VLNK:
523 			ncp->nc_flag |= NCF_ISSYMLINK;
524 			/* XXX cache the contents of the symlink */
525 			break;
526 		default:
527 			break;
528 		}
529 		++numcache;
530 		ncp->nc_error = 0;
531 	} else {
532 		TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
533 		++numneg;
534 		ncp->nc_error = ENOENT;
535 	}
536 	ncp->nc_flag &= ~NCF_UNRESOLVED;
537 }
538 
539 void
540 cache_settimeout(struct namecache *ncp, int nticks)
541 {
542 	if ((ncp->nc_timeout = ticks + nticks) == 0)
543 		ncp->nc_timeout = 1;
544 }
545 
546 /*
547  * Disassociate the vnode or negative-cache association and mark a
548  * namecache entry as unresolved again.  Note that the ncp is still
549  * left in the hash table and still linked to its parent.
550  *
551  * The ncp should be locked and refd on entry and will remain locked and refd
552  * on return.
553  *
554  * This routine is normally never called on a directory containing children.
555  * However, NFS often does just that in its rename() code as a cop-out to
556  * avoid complex namespace operations.  This disconnects a directory vnode
557  * from its namecache and can cause the OLDAPI and NEWAPI to get out of
558  * sync.
559  *
560  * NOTE: NCF_FSMID must be cleared so a refurbishment of the ncp, such as
561  * in a create, properly propogates flag up the chain.
562  */
563 void
564 cache_setunresolved(struct namecache *ncp)
565 {
566 	struct vnode *vp;
567 
568 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
569 		ncp->nc_flag |= NCF_UNRESOLVED;
570 		ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK|
571 				  NCF_FSMID);
572 		ncp->nc_timeout = 0;
573 		ncp->nc_error = ENOTCONN;
574 		++numunres;
575 		if ((vp = ncp->nc_vp) != NULL) {
576 			--numcache;
577 			ncp->nc_vp = NULL;
578 			TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode);
579 
580 			/*
581 			 * Any vp associated with an ncp with children is
582 			 * held by that ncp.  Any vp associated with a locked
583 			 * ncp is held by that ncp.  These conditions must be
584 			 * undone when the vp is cleared out from the ncp.
585 			 */
586 			if (!TAILQ_EMPTY(&ncp->nc_list))
587 				vdrop(vp);
588 			if (ncp->nc_exlocks)
589 				vdrop(vp);
590 		} else {
591 			TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
592 			--numneg;
593 		}
594 	}
595 }
596 
597 /*
598  * Invalidate portions of the namecache topology given a starting entry.
599  * The passed ncp is set to an unresolved state and:
600  *
601  * The passed ncp must be locked.
602  *
603  * CINV_DESTROY		- Set a flag in the passed ncp entry indicating
604  *			  that the physical underlying nodes have been
605  *			  destroyed... as in deleted.  For example, when
606  *			  a directory is removed.  This will cause record
607  *			  lookups on the name to no longer be able to find
608  *			  the record and tells the resolver to return failure
609  *			  rather then trying to resolve through the parent.
610  *
611  *			  The topology itself, including ncp->nc_name,
612  *			  remains intact.
613  *
614  *			  This only applies to the passed ncp, if CINV_CHILDREN
615  *			  is specified the children are not flagged.
616  *
617  * CINV_CHILDREN	- Set all children (recursively) to an unresolved
618  *			  state as well.
619  *
620  *			  Note that this will also have the side effect of
621  *			  cleaning out any unreferenced nodes in the topology
622  *			  from the leaves up as the recursion backs out.
623  *
624  * Note that the topology for any referenced nodes remains intact.
625  *
626  * It is possible for cache_inval() to race a cache_resolve(), meaning that
627  * the namecache entry may not actually be invalidated on return if it was
628  * revalidated while recursing down into its children.  This code guarentees
629  * that the node(s) will go through an invalidation cycle, but does not
630  * guarentee that they will remain in an invalidated state.
631  *
632  * Returns non-zero if a revalidation was detected during the invalidation
633  * recursion, zero otherwise.  Note that since only the original ncp is
634  * locked the revalidation ultimately can only indicate that the original ncp
635  * *MIGHT* no have been reresolved.
636  */
637 int
638 cache_inval(struct namecache *ncp, int flags)
639 {
640 	struct namecache *kid;
641 	struct namecache *nextkid;
642 	int rcnt = 0;
643 
644 	KKASSERT(ncp->nc_exlocks);
645 
646 	cache_setunresolved(ncp);
647 	if (flags & CINV_DESTROY)
648 		ncp->nc_flag |= NCF_DESTROYED;
649 
650 	if ((flags & CINV_CHILDREN) &&
651 	    (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL
652 	) {
653 		cache_hold(kid);
654 		cache_unlock(ncp);
655 		while (kid) {
656 			if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL)
657 				cache_hold(nextkid);
658 			if ((kid->nc_flag & NCF_UNRESOLVED) == 0 ||
659 			    TAILQ_FIRST(&kid->nc_list)
660 			) {
661 				cache_lock(kid);
662 				rcnt += cache_inval(kid, flags & ~CINV_DESTROY);
663 				cache_unlock(kid);
664 			}
665 			cache_drop(kid);
666 			kid = nextkid;
667 		}
668 		cache_lock(ncp);
669 	}
670 
671 	/*
672 	 * Someone could have gotten in there while ncp was unlocked,
673 	 * retry if so.
674 	 */
675 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
676 		++rcnt;
677 	return (rcnt);
678 }
679 
680 /*
681  * Invalidate a vnode's namecache associations.  To avoid races against
682  * the resolver we do not invalidate a node which we previously invalidated
683  * but which was then re-resolved while we were in the invalidation loop.
684  *
685  * Returns non-zero if any namecache entries remain after the invalidation
686  * loop completed.
687  *
688  * NOTE: unlike the namecache topology which guarentees that ncp's will not
689  * be ripped out of the topology while held, the vnode's v_namecache list
690  * has no such restriction.  NCP's can be ripped out of the list at virtually
691  * any time if not locked, even if held.
692  */
693 int
694 cache_inval_vp(struct vnode *vp, int flags, int *retflags)
695 {
696 	struct namecache *ncp;
697 	struct namecache *next;
698 
699 restart:
700 	ncp = TAILQ_FIRST(&vp->v_namecache);
701 	if (ncp)
702 		cache_hold(ncp);
703 	while (ncp) {
704 		/* loop entered with ncp held */
705 		if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
706 			cache_hold(next);
707 		cache_lock(ncp);
708 		if (ncp->nc_vp != vp) {
709 			printf("Warning: cache_inval_vp: race-A detected on "
710 				"%s\n", ncp->nc_name);
711 			cache_put(ncp);
712 			if (next)
713 				cache_drop(next);
714 			goto restart;
715 		}
716 		*retflags |= ncp->nc_flag & NCF_FSMID;
717 		cache_inval(ncp, flags);
718 		cache_put(ncp);		/* also releases reference */
719 		ncp = next;
720 		if (ncp && ncp->nc_vp != vp) {
721 			printf("Warning: cache_inval_vp: race-B detected on "
722 				"%s\n", ncp->nc_name);
723 			cache_drop(ncp);
724 			goto restart;
725 		}
726 	}
727 	return(TAILQ_FIRST(&vp->v_namecache) != NULL);
728 }
729 
730 /*
731  * The source ncp has been renamed to the target ncp.  Both fncp and tncp
732  * must be locked.  Both will be set to unresolved, any children of tncp
733  * will be disconnected (the prior contents of the target is assumed to be
734  * destroyed by the rename operation, e.g. renaming over an empty directory),
735  * and all children of fncp will be moved to tncp.
736  *
737  * XXX the disconnection could pose a problem, check code paths to make
738  * sure any code that blocks can handle the parent being changed out from
739  * under it.  Maybe we should lock the children (watch out for deadlocks) ?
740  *
741  * After we return the caller has the option of calling cache_setvp() if
742  * the vnode of the new target ncp is known.
743  *
744  * Any process CD'd into any of the children will no longer be able to ".."
745  * back out.  An rm -rf can cause this situation to occur.
746  */
747 void
748 cache_rename(struct namecache *fncp, struct namecache *tncp)
749 {
750 	struct namecache *scan;
751 	int didwarn = 0;
752 
753 	cache_setunresolved(fncp);
754 	cache_setunresolved(tncp);
755 	while (cache_inval(tncp, CINV_CHILDREN) != 0) {
756 		if (didwarn++ % 10 == 0) {
757 			printf("Warning: cache_rename: race during "
758 				"rename %s->%s\n",
759 				fncp->nc_name, tncp->nc_name);
760 		}
761 		tsleep(tncp, 0, "mvrace", hz / 10);
762 		cache_setunresolved(tncp);
763 	}
764 	while ((scan = TAILQ_FIRST(&fncp->nc_list)) != NULL) {
765 		cache_hold(scan);
766 		cache_unlink_parent(scan);
767 		cache_link_parent(scan, tncp);
768 		if (scan->nc_flag & NCF_HASHED)
769 			cache_rehash(scan);
770 		cache_drop(scan);
771 	}
772 }
773 
774 /*
775  * vget the vnode associated with the namecache entry.  Resolve the namecache
776  * entry if necessary and deal with namecache/vp races.  The passed ncp must
777  * be referenced and may be locked.  The ncp's ref/locking state is not
778  * effected by this call.
779  *
780  * lk_type may be LK_SHARED, LK_EXCLUSIVE.  A ref'd, possibly locked
781  * (depending on the passed lk_type) will be returned in *vpp with an error
782  * of 0, or NULL will be returned in *vpp with a non-0 error code.  The
783  * most typical error is ENOENT, meaning that the ncp represents a negative
784  * cache hit and there is no vnode to retrieve, but other errors can occur
785  * too.
786  *
787  * The main race we have to deal with are namecache zaps.  The ncp itself
788  * will not disappear since it is referenced, and it turns out that the
789  * validity of the vp pointer can be checked simply by rechecking the
790  * contents of ncp->nc_vp.
791  */
792 int
793 cache_vget(struct namecache *ncp, struct ucred *cred,
794 	   int lk_type, struct vnode **vpp)
795 {
796 	struct vnode *vp;
797 	int error;
798 
799 again:
800 	vp = NULL;
801 	if (ncp->nc_flag & NCF_UNRESOLVED) {
802 		cache_lock(ncp);
803 		error = cache_resolve(ncp, cred);
804 		cache_unlock(ncp);
805 	} else {
806 		error = 0;
807 	}
808 	if (error == 0 && (vp = ncp->nc_vp) != NULL) {
809 		/*
810 		 * Accessing the vnode from the namecache is a bit
811 		 * dangerous.  Because there are no refs on the vnode, it
812 		 * could be in the middle of a reclaim.
813 		 */
814 		if (vp->v_flag & VRECLAIMED) {
815 			printf("Warning: vnode reclaim race detected in cache_vget on %p (%s)\n", vp, ncp->nc_name);
816 			cache_lock(ncp);
817 			cache_setunresolved(ncp);
818 			cache_unlock(ncp);
819 			goto again;
820 		}
821 		error = vget(vp, lk_type, curthread);
822 		if (error) {
823 			if (vp != ncp->nc_vp)
824 				goto again;
825 			vp = NULL;
826 		} else if (vp != ncp->nc_vp) {
827 			vput(vp);
828 			goto again;
829 		} else if (vp->v_flag & VRECLAIMED) {
830 			panic("vget succeeded on a VRECLAIMED node! vp %p", vp);
831 		}
832 	}
833 	if (error == 0 && vp == NULL)
834 		error = ENOENT;
835 	*vpp = vp;
836 	return(error);
837 }
838 
839 int
840 cache_vref(struct namecache *ncp, struct ucred *cred, struct vnode **vpp)
841 {
842 	struct vnode *vp;
843 	int error;
844 
845 again:
846 	vp = NULL;
847 	if (ncp->nc_flag & NCF_UNRESOLVED) {
848 		cache_lock(ncp);
849 		error = cache_resolve(ncp, cred);
850 		cache_unlock(ncp);
851 	} else {
852 		error = 0;
853 	}
854 	if (error == 0 && (vp = ncp->nc_vp) != NULL) {
855 		/*
856 		 * Since we did not obtain any locks, a cache zap
857 		 * race can occur here if the vnode is in the middle
858 		 * of being reclaimed and has not yet been able to
859 		 * clean out its cache node.  If that case occurs,
860 		 * we must lock and unresolve the cache, then loop
861 		 * to retry.
862 		 */
863 		if (vp->v_flag & VRECLAIMED) {
864 			printf("Warning: vnode reclaim race detected on cache_vref %p (%s)\n", vp, ncp->nc_name);
865 			cache_lock(ncp);
866 			cache_setunresolved(ncp);
867 			cache_unlock(ncp);
868 			goto again;
869 		}
870 		vref(vp);
871 	}
872 	if (error == 0 && vp == NULL)
873 		error = ENOENT;
874 	*vpp = vp;
875 	return(error);
876 }
877 
878 /*
879  * Recursively set the FSMID update flag for namecache nodes leading
880  * to root.  This will cause the next getattr or reclaim to increment the
881  * fsmid and mark the inode for lazy updating.
882  *
883  * Stop recursing when we hit a node whos NCF_FSMID flag is already set.
884  * This makes FSMIDs work in an Einsteinian fashion - where the observation
885  * effects the result.  In this case a program monitoring a higher level
886  * node will have detected some prior change and started its scan (clearing
887  * NCF_FSMID in higher level nodes), but since it has not yet observed the
888  * node where we find NCF_FSMID still set, we can safely make the related
889  * modification without interfering with the theorized program.
890  *
891  * This also means that FSMIDs cannot represent time-domain quantities
892  * in a hierarchical sense.  But the main reason for doing it this way
893  * is to reduce the amount of recursion that occurs in the critical path
894  * when e.g. a program is writing to a file that sits deep in a directory
895  * hierarchy.
896  */
897 void
898 cache_update_fsmid(struct namecache *ncp)
899 {
900 	struct vnode *vp;
901 	struct namecache *scan;
902 
903 	/*
904 	 * Warning: even if we get a non-NULL vp it could still be in the
905 	 * middle of a recyclement.  Don't do anything fancy, just set
906 	 * NCF_FSMID.
907 	 */
908 	if ((vp = ncp->nc_vp) != NULL) {
909 		TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
910 			for (scan = ncp; scan; scan = scan->nc_parent) {
911 				if (scan->nc_flag & NCF_FSMID)
912 					break;
913 				scan->nc_flag |= NCF_FSMID;
914 			}
915 		}
916 	} else {
917 		while (ncp && (ncp->nc_flag & NCF_FSMID) == 0) {
918 			ncp->nc_flag |= NCF_FSMID;
919 			ncp = ncp->nc_parent;
920 		}
921 	}
922 }
923 
924 void
925 cache_update_fsmid_vp(struct vnode *vp)
926 {
927 	struct namecache *ncp;
928 	struct namecache *scan;
929 
930 	TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
931 		for (scan = ncp; scan; scan = scan->nc_parent) {
932 			if (scan->nc_flag & NCF_FSMID)
933 				break;
934 			scan->nc_flag |= NCF_FSMID;
935 		}
936 	}
937 }
938 
939 /*
940  * If getattr is called on a vnode (e.g. a stat call), the filesystem
941  * may call this routine to determine if the namecache has the hierarchical
942  * change flag set, requiring the fsmid to be updated.
943  *
944  * Since 0 indicates no support, make sure the filesystem fsmid is at least
945  * 1.
946  */
947 int
948 cache_check_fsmid_vp(struct vnode *vp, int64_t *fsmid)
949 {
950 	struct namecache *ncp;
951 	int changed = 0;
952 
953 	TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
954 		if (ncp->nc_flag & NCF_FSMID) {
955 			ncp->nc_flag &= ~NCF_FSMID;
956 			changed = 1;
957 		}
958 	}
959 	if (*fsmid == 0)
960 		++*fsmid;
961 	if (changed)
962 		++*fsmid;
963 	return(changed);
964 }
965 
966 /*
967  * Convert a directory vnode to a namecache record without any other
968  * knowledge of the topology.  This ONLY works with directory vnodes and
969  * is ONLY used by the NFS server.  dvp must be refd but unlocked, and the
970  * returned ncp (if not NULL) will be held and unlocked.
971  *
972  * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned.
973  * If 'makeit' is 1 we attempt to track-down and create the namecache topology
974  * for dvp.  This will fail only if the directory has been deleted out from
975  * under the caller.
976  *
977  * Callers must always check for a NULL return no matter the value of 'makeit'.
978  *
979  * To avoid underflowing the kernel stack each recursive call increments
980  * the makeit variable.
981  */
982 
983 static int cache_inefficient_scan(struct namecache *ncp, struct ucred *cred,
984 				  struct vnode *dvp);
985 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
986 				  struct vnode **saved_dvp);
987 
988 struct namecache *
989 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit)
990 {
991 	struct namecache *ncp;
992 	struct vnode *saved_dvp;
993 	struct vnode *pvp;
994 	int error;
995 
996 	ncp = NULL;
997 	saved_dvp = NULL;
998 
999 	/*
1000 	 * Temporary debugging code to force the directory scanning code
1001 	 * to be exercised.
1002 	 */
1003 	if (ncvp_debug >= 3 && makeit && TAILQ_FIRST(&dvp->v_namecache)) {
1004 		ncp = TAILQ_FIRST(&dvp->v_namecache);
1005 		printf("cache_fromdvp: forcing %s\n", ncp->nc_name);
1006 		goto force;
1007 	}
1008 
1009 	/*
1010 	 * Loop until resolution, inside code will break out on error.
1011 	 */
1012 	while ((ncp = TAILQ_FIRST(&dvp->v_namecache)) == NULL && makeit) {
1013 force:
1014 		/*
1015 		 * If dvp is the root of its filesystem it should already
1016 		 * have a namecache pointer associated with it as a side
1017 		 * effect of the mount, but it may have been disassociated.
1018 		 */
1019 		if (dvp->v_flag & VROOT) {
1020 			ncp = cache_get(dvp->v_mount->mnt_ncp);
1021 			error = cache_resolve_mp(ncp);
1022 			cache_put(ncp);
1023 			if (ncvp_debug) {
1024 				printf("cache_fromdvp: resolve root of mount %p error %d",
1025 					dvp->v_mount, error);
1026 			}
1027 			if (error) {
1028 				if (ncvp_debug)
1029 					printf(" failed\n");
1030 				ncp = NULL;
1031 				break;
1032 			}
1033 			if (ncvp_debug)
1034 				printf(" succeeded\n");
1035 			continue;
1036 		}
1037 
1038 		/*
1039 		 * If we are recursed too deeply resort to an O(n^2)
1040 		 * algorithm to resolve the namecache topology.  The
1041 		 * resolved pvp is left referenced in saved_dvp to
1042 		 * prevent the tree from being destroyed while we loop.
1043 		 */
1044 		if (makeit > 20) {
1045 			error = cache_fromdvp_try(dvp, cred, &saved_dvp);
1046 			if (error) {
1047 				printf("lookupdotdot(longpath) failed %d "
1048 				       "dvp %p\n", error, dvp);
1049 				break;
1050 			}
1051 			continue;
1052 		}
1053 
1054 		/*
1055 		 * Get the parent directory and resolve its ncp.
1056 		 */
1057 		error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred);
1058 		if (error) {
1059 			printf("lookupdotdot failed %d dvp %p\n", error, dvp);
1060 			break;
1061 		}
1062 		VOP_UNLOCK(pvp, 0, curthread);
1063 
1064 		/*
1065 		 * Reuse makeit as a recursion depth counter.
1066 		 */
1067 		ncp = cache_fromdvp(pvp, cred, makeit + 1);
1068 		vrele(pvp);
1069 		if (ncp == NULL)
1070 			break;
1071 
1072 		/*
1073 		 * Do an inefficient scan of pvp (embodied by ncp) to look
1074 		 * for dvp.  This will create a namecache record for dvp on
1075 		 * success.  We loop up to recheck on success.
1076 		 *
1077 		 * ncp and dvp are both held but not locked.
1078 		 */
1079 		error = cache_inefficient_scan(ncp, cred, dvp);
1080 		cache_drop(ncp);
1081 		if (error) {
1082 			printf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n",
1083 				pvp, ncp->nc_name, dvp);
1084 			ncp = NULL;
1085 			break;
1086 		}
1087 		if (ncvp_debug) {
1088 			printf("cache_fromdvp: scan %p (%s) succeeded\n",
1089 				pvp, ncp->nc_name);
1090 		}
1091 	}
1092 	if (ncp)
1093 		cache_hold(ncp);
1094 	if (saved_dvp)
1095 		vrele(saved_dvp);
1096 	return (ncp);
1097 }
1098 
1099 /*
1100  * Go up the chain of parent directories until we find something
1101  * we can resolve into the namecache.  This is very inefficient.
1102  */
1103 static
1104 int
1105 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1106 		  struct vnode **saved_dvp)
1107 {
1108 	struct namecache *ncp;
1109 	struct vnode *pvp;
1110 	int error;
1111 	static time_t last_fromdvp_report;
1112 
1113 	/*
1114 	 * Loop getting the parent directory vnode until we get something we
1115 	 * can resolve in the namecache.
1116 	 */
1117 	vref(dvp);
1118 	for (;;) {
1119 		error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred);
1120 		if (error) {
1121 			vrele(dvp);
1122 			return (error);
1123 		}
1124 		VOP_UNLOCK(pvp, 0, curthread);
1125 		if ((ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) {
1126 			cache_hold(ncp);
1127 			vrele(pvp);
1128 			break;
1129 		}
1130 		if (pvp->v_flag & VROOT) {
1131 			ncp = cache_get(pvp->v_mount->mnt_ncp);
1132 			error = cache_resolve_mp(ncp);
1133 			cache_unlock(ncp);
1134 			vrele(pvp);
1135 			if (error) {
1136 				cache_drop(ncp);
1137 				vrele(dvp);
1138 				return (error);
1139 			}
1140 			break;
1141 		}
1142 		vrele(dvp);
1143 		dvp = pvp;
1144 	}
1145 	if (last_fromdvp_report != time_second) {
1146 		last_fromdvp_report = time_second;
1147 		printf("Warning: extremely inefficient path resolution on %s\n",
1148 			ncp->nc_name);
1149 	}
1150 	error = cache_inefficient_scan(ncp, cred, dvp);
1151 
1152 	/*
1153 	 * Hopefully dvp now has a namecache record associated with it.
1154 	 * Leave it referenced to prevent the kernel from recycling the
1155 	 * vnode.  Otherwise extremely long directory paths could result
1156 	 * in endless recycling.
1157 	 */
1158 	if (*saved_dvp)
1159 	    vrele(*saved_dvp);
1160 	*saved_dvp = dvp;
1161 	return (error);
1162 }
1163 
1164 
1165 /*
1166  * Do an inefficient scan of the directory represented by ncp looking for
1167  * the directory vnode dvp.  ncp must be held but not locked on entry and
1168  * will be held on return.  dvp must be refd but not locked on entry and
1169  * will remain refd on return.
1170  *
1171  * Why do this at all?  Well, due to its stateless nature the NFS server
1172  * converts file handles directly to vnodes without necessarily going through
1173  * the namecache ops that would otherwise create the namecache topology
1174  * leading to the vnode.  We could either (1) Change the namecache algorithms
1175  * to allow disconnect namecache records that are re-merged opportunistically,
1176  * or (2) Make the NFS server backtrack and scan to recover a connected
1177  * namecache topology in order to then be able to issue new API lookups.
1178  *
1179  * It turns out that (1) is a huge mess.  It takes a nice clean set of
1180  * namecache algorithms and introduces a lot of complication in every subsystem
1181  * that calls into the namecache to deal with the re-merge case, especially
1182  * since we are using the namecache to placehold negative lookups and the
1183  * vnode might not be immediately assigned. (2) is certainly far less
1184  * efficient then (1), but since we are only talking about directories here
1185  * (which are likely to remain cached), the case does not actually run all
1186  * that often and has the supreme advantage of not polluting the namecache
1187  * algorithms.
1188  */
1189 static int
1190 cache_inefficient_scan(struct namecache *ncp, struct ucred *cred,
1191 		       struct vnode *dvp)
1192 {
1193 	struct nlcomponent nlc;
1194 	struct namecache *rncp;
1195 	struct dirent *den;
1196 	struct vnode *pvp;
1197 	struct vattr vat;
1198 	struct iovec iov;
1199 	struct uio uio;
1200 	int blksize;
1201 	int eofflag;
1202 	int bytes;
1203 	char *rbuf;
1204 	int error;
1205 
1206 	vat.va_blocksize = 0;
1207 	if ((error = VOP_GETATTR(dvp, &vat, curthread)) != 0)
1208 		return (error);
1209 	if ((error = cache_vget(ncp, cred, LK_SHARED, &pvp)) != 0)
1210 		return (error);
1211 	if (ncvp_debug)
1212 		printf("inefficient_scan: directory iosize %ld vattr fileid = %ld\n", vat.va_blocksize, (long)vat.va_fileid);
1213 	if ((blksize = vat.va_blocksize) == 0)
1214 		blksize = DEV_BSIZE;
1215 	rbuf = malloc(blksize, M_TEMP, M_WAITOK);
1216 	rncp = NULL;
1217 
1218 	eofflag = 0;
1219 	uio.uio_offset = 0;
1220 again:
1221 	iov.iov_base = rbuf;
1222 	iov.iov_len = blksize;
1223 	uio.uio_iov = &iov;
1224 	uio.uio_iovcnt = 1;
1225 	uio.uio_resid = blksize;
1226 	uio.uio_segflg = UIO_SYSSPACE;
1227 	uio.uio_rw = UIO_READ;
1228 	uio.uio_td = curthread;
1229 
1230 	if (ncvp_debug >= 2)
1231 		printf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset);
1232 	error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL);
1233 	if (error == 0) {
1234 		den = (struct dirent *)rbuf;
1235 		bytes = blksize - uio.uio_resid;
1236 
1237 		while (bytes > 0) {
1238 			if (ncvp_debug >= 2) {
1239 				printf("cache_inefficient_scan: %*.*s\n",
1240 					den->d_namlen, den->d_namlen,
1241 					den->d_name);
1242 			}
1243 			if (den->d_type != DT_WHT &&
1244 			    den->d_ino == vat.va_fileid) {
1245 				if (ncvp_debug) {
1246 					printf("cache_inefficient_scan: "
1247 					       "MATCHED inode %ld path %s/%*.*s\n",
1248 					       vat.va_fileid, ncp->nc_name,
1249 					       den->d_namlen, den->d_namlen,
1250 					       den->d_name);
1251 				}
1252 				nlc.nlc_nameptr = den->d_name;
1253 				nlc.nlc_namelen = den->d_namlen;
1254 				VOP_UNLOCK(pvp, 0, curthread);
1255 				rncp = cache_nlookup(ncp, &nlc);
1256 				KKASSERT(rncp != NULL);
1257 				break;
1258 			}
1259 			bytes -= _DIRENT_DIRSIZ(den);
1260 			den = _DIRENT_NEXT(den);
1261 		}
1262 		if (rncp == NULL && eofflag == 0 && uio.uio_resid != blksize)
1263 			goto again;
1264 	}
1265 	if (rncp) {
1266 		vrele(pvp);
1267 		if (rncp->nc_flag & NCF_UNRESOLVED) {
1268 			cache_setvp(rncp, dvp);
1269 			if (ncvp_debug >= 2) {
1270 				printf("cache_inefficient_scan: setvp %s/%s = %p\n",
1271 					ncp->nc_name, rncp->nc_name, dvp);
1272 			}
1273 		} else {
1274 			if (ncvp_debug >= 2) {
1275 				printf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n",
1276 					ncp->nc_name, rncp->nc_name, dvp,
1277 					rncp->nc_vp);
1278 			}
1279 		}
1280 		if (rncp->nc_vp == NULL)
1281 			error = rncp->nc_error;
1282 		cache_put(rncp);
1283 	} else {
1284 		printf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n",
1285 			dvp, ncp->nc_name);
1286 		vput(pvp);
1287 		error = ENOENT;
1288 	}
1289 	free(rbuf, M_TEMP);
1290 	return (error);
1291 }
1292 
1293 /*
1294  * Zap a namecache entry.  The ncp is unconditionally set to an unresolved
1295  * state, which disassociates it from its vnode or ncneglist.
1296  *
1297  * Then, if there are no additional references to the ncp and no children,
1298  * the ncp is removed from the topology and destroyed.  This function will
1299  * also run through the nc_parent chain and destroy parent ncps if possible.
1300  * As a side benefit, it turns out the only conditions that allow running
1301  * up the chain are also the conditions to ensure no deadlock will occur.
1302  *
1303  * References and/or children may exist if the ncp is in the middle of the
1304  * topology, preventing the ncp from being destroyed.
1305  *
1306  * This function must be called with the ncp held and locked and will unlock
1307  * and drop it during zapping.
1308  */
1309 static void
1310 cache_zap(struct namecache *ncp)
1311 {
1312 	struct namecache *par;
1313 
1314 	/*
1315 	 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED.
1316 	 */
1317 	cache_setunresolved(ncp);
1318 
1319 	/*
1320 	 * Try to scrap the entry and possibly tail-recurse on its parent.
1321 	 * We only scrap unref'd (other then our ref) unresolved entries,
1322 	 * we do not scrap 'live' entries.
1323 	 */
1324 	while (ncp->nc_flag & NCF_UNRESOLVED) {
1325 		/*
1326 		 * Someone other then us has a ref, stop.
1327 		 */
1328 		if (ncp->nc_refs > 1)
1329 			goto done;
1330 
1331 		/*
1332 		 * We have children, stop.
1333 		 */
1334 		if (!TAILQ_EMPTY(&ncp->nc_list))
1335 			goto done;
1336 
1337 		/*
1338 		 * Remove ncp from the topology: hash table and parent linkage.
1339 		 */
1340 		if (ncp->nc_flag & NCF_HASHED) {
1341 			ncp->nc_flag &= ~NCF_HASHED;
1342 			LIST_REMOVE(ncp, nc_hash);
1343 		}
1344 		if ((par = ncp->nc_parent) != NULL) {
1345 			par = cache_hold(par);
1346 			TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
1347 			ncp->nc_parent = NULL;
1348 			if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
1349 				vdrop(par->nc_vp);
1350 		}
1351 
1352 		/*
1353 		 * ncp should not have picked up any refs.  Physically
1354 		 * destroy the ncp.
1355 		 */
1356 		KKASSERT(ncp->nc_refs == 1);
1357 		--numunres;
1358 		/* cache_unlock(ncp) not required */
1359 		ncp->nc_refs = -1;	/* safety */
1360 		if (ncp->nc_name)
1361 			free(ncp->nc_name, M_VFSCACHE);
1362 		free(ncp, M_VFSCACHE);
1363 
1364 		/*
1365 		 * Loop on the parent (it may be NULL).  Only bother looping
1366 		 * if the parent has a single ref (ours), which also means
1367 		 * we can lock it trivially.
1368 		 */
1369 		ncp = par;
1370 		if (ncp == NULL)
1371 			return;
1372 		if (ncp->nc_refs != 1) {
1373 			cache_drop(ncp);
1374 			return;
1375 		}
1376 		KKASSERT(par->nc_exlocks == 0);
1377 		cache_lock(ncp);
1378 	}
1379 done:
1380 	cache_unlock(ncp);
1381 	--ncp->nc_refs;
1382 }
1383 
1384 static enum { CHI_LOW, CHI_HIGH } cache_hysteresis_state = CHI_LOW;
1385 
1386 static __inline
1387 void
1388 cache_hysteresis(void)
1389 {
1390 	/*
1391 	 * Don't cache too many negative hits.  We use hysteresis to reduce
1392 	 * the impact on the critical path.
1393 	 */
1394 	switch(cache_hysteresis_state) {
1395 	case CHI_LOW:
1396 		if (numneg > MINNEG && numneg * ncnegfactor > numcache) {
1397 			cache_cleanneg(10);
1398 			cache_hysteresis_state = CHI_HIGH;
1399 		}
1400 		break;
1401 	case CHI_HIGH:
1402 		if (numneg > MINNEG * 9 / 10 &&
1403 		    numneg * ncnegfactor * 9 / 10 > numcache
1404 		) {
1405 			cache_cleanneg(10);
1406 		} else {
1407 			cache_hysteresis_state = CHI_LOW;
1408 		}
1409 		break;
1410 	}
1411 }
1412 
1413 /*
1414  * NEW NAMECACHE LOOKUP API
1415  *
1416  * Lookup an entry in the cache.  A locked, referenced, non-NULL
1417  * entry is *always* returned, even if the supplied component is illegal.
1418  * The resulting namecache entry should be returned to the system with
1419  * cache_put() or cache_unlock() + cache_drop().
1420  *
1421  * namecache locks are recursive but care must be taken to avoid lock order
1422  * reversals.
1423  *
1424  * Nobody else will be able to manipulate the associated namespace (e.g.
1425  * create, delete, rename, rename-target) until the caller unlocks the
1426  * entry.
1427  *
1428  * The returned entry will be in one of three states:  positive hit (non-null
1429  * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set).
1430  * Unresolved entries must be resolved through the filesystem to associate the
1431  * vnode and/or determine whether a positive or negative hit has occured.
1432  *
1433  * It is not necessary to lock a directory in order to lock namespace under
1434  * that directory.  In fact, it is explicitly not allowed to do that.  A
1435  * directory is typically only locked when being created, renamed, or
1436  * destroyed.
1437  *
1438  * The directory (par) may be unresolved, in which case any returned child
1439  * will likely also be marked unresolved.  Likely but not guarenteed.  Since
1440  * the filesystem lookup requires a resolved directory vnode the caller is
1441  * responsible for resolving the namecache chain top-down.  This API
1442  * specifically allows whole chains to be created in an unresolved state.
1443  */
1444 struct namecache *
1445 cache_nlookup(struct namecache *par, struct nlcomponent *nlc)
1446 {
1447 	struct namecache *ncp;
1448 	struct namecache *new_ncp;
1449 	struct nchashhead *nchpp;
1450 	u_int32_t hash;
1451 	globaldata_t gd;
1452 
1453 	numcalls++;
1454 	gd = mycpu;
1455 
1456 	/*
1457 	 * Try to locate an existing entry
1458 	 */
1459 	hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
1460 	hash = fnv_32_buf(&par, sizeof(par), hash);
1461 	new_ncp = NULL;
1462 restart:
1463 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1464 		numchecks++;
1465 
1466 		/*
1467 		 * Zap entries that have timed out.
1468 		 */
1469 		if (ncp->nc_timeout &&
1470 		    (int)(ncp->nc_timeout - ticks) < 0 &&
1471 		    (ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
1472 		    ncp->nc_exlocks == 0
1473 		) {
1474 			cache_zap(cache_get(ncp));
1475 			goto restart;
1476 		}
1477 
1478 		/*
1479 		 * Break out if we find a matching entry.  Note that
1480 		 * UNRESOLVED entries may match, but DESTROYED entries
1481 		 * do not.
1482 		 */
1483 		if (ncp->nc_parent == par &&
1484 		    ncp->nc_nlen == nlc->nlc_namelen &&
1485 		    bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
1486 		    (ncp->nc_flag & NCF_DESTROYED) == 0
1487 		) {
1488 			if (cache_get_nonblock(ncp) == 0) {
1489 				if (new_ncp)
1490 					cache_free(new_ncp);
1491 				goto found;
1492 			}
1493 			cache_get(ncp);
1494 			cache_put(ncp);
1495 			goto restart;
1496 		}
1497 	}
1498 
1499 	/*
1500 	 * We failed to locate an entry, create a new entry and add it to
1501 	 * the cache.  We have to relookup after possibly blocking in
1502 	 * malloc.
1503 	 */
1504 	if (new_ncp == NULL) {
1505 		new_ncp = cache_alloc(nlc->nlc_namelen);
1506 		goto restart;
1507 	}
1508 
1509 	ncp = new_ncp;
1510 
1511 	/*
1512 	 * Initialize as a new UNRESOLVED entry, lock (non-blocking),
1513 	 * and link to the parent.  The mount point is usually inherited
1514 	 * from the parent unless this is a special case such as a mount
1515 	 * point where nlc_namelen is 0.  The caller is responsible for
1516 	 * setting nc_mount in that case.  If nlc_namelen is 0 nc_name will
1517 	 * be NULL.
1518 	 */
1519 	if (nlc->nlc_namelen) {
1520 		bcopy(nlc->nlc_nameptr, ncp->nc_name, nlc->nlc_namelen);
1521 		ncp->nc_name[nlc->nlc_namelen] = 0;
1522 		ncp->nc_mount = par->nc_mount;
1523 	}
1524 	nchpp = NCHHASH(hash);
1525 	LIST_INSERT_HEAD(nchpp, ncp, nc_hash);
1526 	ncp->nc_flag |= NCF_HASHED;
1527 	cache_link_parent(ncp, par);
1528 found:
1529 	/*
1530 	 * stats and namecache size management
1531 	 */
1532 	if (ncp->nc_flag & NCF_UNRESOLVED)
1533 		++gd->gd_nchstats->ncs_miss;
1534 	else if (ncp->nc_vp)
1535 		++gd->gd_nchstats->ncs_goodhits;
1536 	else
1537 		++gd->gd_nchstats->ncs_neghits;
1538 	cache_hysteresis();
1539 	return(ncp);
1540 }
1541 
1542 /*
1543  * Given a locked ncp, validate that the vnode, if present, is actually
1544  * usable.  If it is not usable set the ncp to an unresolved state.
1545  */
1546 void
1547 cache_validate(struct namecache *ncp)
1548 {
1549 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1550 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1551 			cache_setunresolved(ncp);
1552 	}
1553 }
1554 
1555 /*
1556  * Resolve an unresolved namecache entry, generally by looking it up.
1557  * The passed ncp must be locked and refd.
1558  *
1559  * Theoretically since a vnode cannot be recycled while held, and since
1560  * the nc_parent chain holds its vnode as long as children exist, the
1561  * direct parent of the cache entry we are trying to resolve should
1562  * have a valid vnode.  If not then generate an error that we can
1563  * determine is related to a resolver bug.
1564  *
1565  * However, if a vnode was in the middle of a recyclement when the NCP
1566  * got locked, ncp->nc_vp might point to a vnode that is about to become
1567  * invalid.  cache_resolve() handles this case by unresolving the entry
1568  * and then re-resolving it.
1569  *
1570  * Note that successful resolution does not necessarily return an error
1571  * code of 0.  If the ncp resolves to a negative cache hit then ENOENT
1572  * will be returned.
1573  */
1574 int
1575 cache_resolve(struct namecache *ncp, struct ucred *cred)
1576 {
1577 	struct namecache *par;
1578 	int error;
1579 
1580 restart:
1581 	/*
1582 	 * If the ncp is already resolved we have nothing to do.  However,
1583 	 * we do want to guarentee that a usable vnode is returned when
1584 	 * a vnode is present, so make sure it hasn't been reclaimed.
1585 	 */
1586 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1587 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1588 			cache_setunresolved(ncp);
1589 		if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
1590 			return (ncp->nc_error);
1591 	}
1592 
1593 	/*
1594 	 * Mount points need special handling because the parent does not
1595 	 * belong to the same filesystem as the ncp.
1596 	 */
1597 	if (ncp->nc_flag & NCF_MOUNTPT)
1598 		return (cache_resolve_mp(ncp));
1599 
1600 	/*
1601 	 * We expect an unbroken chain of ncps to at least the mount point,
1602 	 * and even all the way to root (but this code doesn't have to go
1603 	 * past the mount point).
1604 	 */
1605 	if (ncp->nc_parent == NULL) {
1606 		printf("EXDEV case 1 %p %*.*s\n", ncp,
1607 			ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
1608 		ncp->nc_error = EXDEV;
1609 		return(ncp->nc_error);
1610 	}
1611 
1612 	/*
1613 	 * The vp's of the parent directories in the chain are held via vhold()
1614 	 * due to the existance of the child, and should not disappear.
1615 	 * However, there are cases where they can disappear:
1616 	 *
1617 	 *	- due to filesystem I/O errors.
1618 	 *	- due to NFS being stupid about tracking the namespace and
1619 	 *	  destroys the namespace for entire directories quite often.
1620 	 *	- due to forced unmounts.
1621 	 *	- due to an rmdir (parent will be marked DESTROYED)
1622 	 *
1623 	 * When this occurs we have to track the chain backwards and resolve
1624 	 * it, looping until the resolver catches up to the current node.  We
1625 	 * could recurse here but we might run ourselves out of kernel stack
1626 	 * so we do it in a more painful manner.  This situation really should
1627 	 * not occur all that often, or if it does not have to go back too
1628 	 * many nodes to resolve the ncp.
1629 	 */
1630 	while (ncp->nc_parent->nc_vp == NULL) {
1631 		/*
1632 		 * This case can occur if a process is CD'd into a
1633 		 * directory which is then rmdir'd.  If the parent is marked
1634 		 * destroyed there is no point trying to resolve it.
1635 		 */
1636 		if (ncp->nc_parent->nc_flag & NCF_DESTROYED)
1637 			return(ENOENT);
1638 
1639 		par = ncp->nc_parent;
1640 		while (par->nc_parent && par->nc_parent->nc_vp == NULL)
1641 			par = par->nc_parent;
1642 		if (par->nc_parent == NULL) {
1643 			printf("EXDEV case 2 %*.*s\n",
1644 				par->nc_nlen, par->nc_nlen, par->nc_name);
1645 			return (EXDEV);
1646 		}
1647 		printf("[diagnostic] cache_resolve: had to recurse on %*.*s\n",
1648 			par->nc_nlen, par->nc_nlen, par->nc_name);
1649 		/*
1650 		 * The parent is not set in stone, ref and lock it to prevent
1651 		 * it from disappearing.  Also note that due to renames it
1652 		 * is possible for our ncp to move and for par to no longer
1653 		 * be one of its parents.  We resolve it anyway, the loop
1654 		 * will handle any moves.
1655 		 */
1656 		cache_get(par);
1657 		if (par->nc_flag & NCF_MOUNTPT) {
1658 			cache_resolve_mp(par);
1659 		} else if (par->nc_parent->nc_vp == NULL) {
1660 			printf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name);
1661 			cache_put(par);
1662 			continue;
1663 		} else if (par->nc_flag & NCF_UNRESOLVED) {
1664 			par->nc_error = VOP_NRESOLVE(par, cred);
1665 		}
1666 		if ((error = par->nc_error) != 0) {
1667 			if (par->nc_error != EAGAIN) {
1668 				printf("EXDEV case 3 %*.*s error %d\n",
1669 				    par->nc_nlen, par->nc_nlen, par->nc_name,
1670 				    par->nc_error);
1671 				cache_put(par);
1672 				return(error);
1673 			}
1674 			printf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n",
1675 				par, par->nc_nlen, par->nc_nlen, par->nc_name);
1676 		}
1677 		cache_put(par);
1678 		/* loop */
1679 	}
1680 
1681 	/*
1682 	 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected
1683 	 * ncp's and reattach them.  If this occurs the original ncp is marked
1684 	 * EAGAIN to force a relookup.
1685 	 *
1686 	 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed
1687 	 * ncp must already be resolved.
1688 	 */
1689 	KKASSERT((ncp->nc_flag & NCF_MOUNTPT) == 0);
1690 	ncp->nc_error = VOP_NRESOLVE(ncp, cred);
1691 	/*vop_nresolve(*ncp->nc_parent->nc_vp->v_ops, ncp, cred);*/
1692 	if (ncp->nc_error == EAGAIN) {
1693 		printf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n",
1694 			ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
1695 		goto restart;
1696 	}
1697 	return(ncp->nc_error);
1698 }
1699 
1700 /*
1701  * Resolve the ncp associated with a mount point.  Such ncp's almost always
1702  * remain resolved and this routine is rarely called.  NFS MPs tends to force
1703  * re-resolution more often due to its mac-truck-smash-the-namecache
1704  * method of tracking namespace changes.
1705  *
1706  * The semantics for this call is that the passed ncp must be locked on
1707  * entry and will be locked on return.  However, if we actually have to
1708  * resolve the mount point we temporarily unlock the entry in order to
1709  * avoid race-to-root deadlocks due to e.g. dead NFS mounts.  Because of
1710  * the unlock we have to recheck the flags after we relock.
1711  */
1712 static int
1713 cache_resolve_mp(struct namecache *ncp)
1714 {
1715 	struct vnode *vp;
1716 	struct mount *mp = ncp->nc_mount;
1717 	int error;
1718 
1719 	KKASSERT(mp != NULL);
1720 
1721 	/*
1722 	 * If the ncp is already resolved we have nothing to do.  However,
1723 	 * we do want to guarentee that a usable vnode is returned when
1724 	 * a vnode is present, so make sure it hasn't been reclaimed.
1725 	 */
1726 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1727 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1728 			cache_setunresolved(ncp);
1729 	}
1730 
1731 	if (ncp->nc_flag & NCF_UNRESOLVED) {
1732 		cache_unlock(ncp);
1733 		while (vfs_busy(mp, 0, curthread))
1734 			;
1735 		error = VFS_ROOT(mp, &vp);
1736 		cache_lock(ncp);
1737 
1738 		/*
1739 		 * recheck the ncp state after relocking.
1740 		 */
1741 		if (ncp->nc_flag & NCF_UNRESOLVED) {
1742 			ncp->nc_error = error;
1743 			if (error == 0) {
1744 				cache_setvp(ncp, vp);
1745 				vput(vp);
1746 			} else {
1747 				printf("[diagnostic] cache_resolve_mp: failed to resolve mount %p\n", mp);
1748 				cache_setvp(ncp, NULL);
1749 			}
1750 		} else if (error == 0) {
1751 			vput(vp);
1752 		}
1753 		vfs_unbusy(mp, curthread);
1754 	}
1755 	return(ncp->nc_error);
1756 }
1757 
1758 void
1759 cache_cleanneg(int count)
1760 {
1761 	struct namecache *ncp;
1762 
1763 	/*
1764 	 * Automode from the vnlru proc - clean out 10% of the negative cache
1765 	 * entries.
1766 	 */
1767 	if (count == 0)
1768 		count = numneg / 10 + 1;
1769 
1770 	/*
1771 	 * Attempt to clean out the specified number of negative cache
1772 	 * entries.
1773 	 */
1774 	while (count) {
1775 		ncp = TAILQ_FIRST(&ncneglist);
1776 		if (ncp == NULL) {
1777 			KKASSERT(numneg == 0);
1778 			break;
1779 		}
1780 		TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
1781 		TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
1782 		if (cache_get_nonblock(ncp) == 0)
1783 			cache_zap(ncp);
1784 		--count;
1785 	}
1786 }
1787 
1788 /*
1789  * Rehash a ncp.  Rehashing is typically required if the name changes (should
1790  * not generally occur) or the parent link changes.  This function will
1791  * unhash the ncp if the ncp is no longer hashable.
1792  */
1793 static void
1794 cache_rehash(struct namecache *ncp)
1795 {
1796 	struct nchashhead *nchpp;
1797 	u_int32_t hash;
1798 
1799 	if (ncp->nc_flag & NCF_HASHED) {
1800 		ncp->nc_flag &= ~NCF_HASHED;
1801 		LIST_REMOVE(ncp, nc_hash);
1802 	}
1803 	if (ncp->nc_nlen && ncp->nc_parent) {
1804 		hash = fnv_32_buf(ncp->nc_name, ncp->nc_nlen, FNV1_32_INIT);
1805 		hash = fnv_32_buf(&ncp->nc_parent,
1806 					sizeof(ncp->nc_parent), hash);
1807 		nchpp = NCHHASH(hash);
1808 		LIST_INSERT_HEAD(nchpp, ncp, nc_hash);
1809 		ncp->nc_flag |= NCF_HASHED;
1810 	}
1811 }
1812 
1813 /*
1814  * Name cache initialization, from vfsinit() when we are booting
1815  */
1816 void
1817 nchinit(void)
1818 {
1819 	int i;
1820 	globaldata_t gd;
1821 
1822 	/* initialise per-cpu namecache effectiveness statistics. */
1823 	for (i = 0; i < ncpus; ++i) {
1824 		gd = globaldata_find(i);
1825 		gd->gd_nchstats = &nchstats[i];
1826 	}
1827 	TAILQ_INIT(&ncneglist);
1828 	nchashtbl = hashinit(desiredvnodes*2, M_VFSCACHE, &nchash);
1829 	nclockwarn = 1 * hz;
1830 }
1831 
1832 /*
1833  * Called from start_init() to bootstrap the root filesystem.  Returns
1834  * a referenced, unlocked namecache record.
1835  */
1836 struct namecache *
1837 cache_allocroot(struct mount *mp, struct vnode *vp)
1838 {
1839 	struct namecache *ncp = cache_alloc(0);
1840 
1841 	ncp->nc_flag |= NCF_MOUNTPT | NCF_ROOT;
1842 	ncp->nc_mount = mp;
1843 	cache_setvp(ncp, vp);
1844 	return(ncp);
1845 }
1846 
1847 /*
1848  * vfs_cache_setroot()
1849  *
1850  *	Create an association between the root of our namecache and
1851  *	the root vnode.  This routine may be called several times during
1852  *	booting.
1853  *
1854  *	If the caller intends to save the returned namecache pointer somewhere
1855  *	it must cache_hold() it.
1856  */
1857 void
1858 vfs_cache_setroot(struct vnode *nvp, struct namecache *ncp)
1859 {
1860 	struct vnode *ovp;
1861 	struct namecache *oncp;
1862 
1863 	ovp = rootvnode;
1864 	oncp = rootncp;
1865 	rootvnode = nvp;
1866 	rootncp = ncp;
1867 
1868 	if (ovp)
1869 		vrele(ovp);
1870 	if (oncp)
1871 		cache_drop(oncp);
1872 }
1873 
1874 /*
1875  * XXX OLD API COMPAT FUNCTION.  This really messes up the new namecache
1876  * topology and is being removed as quickly as possible.  The new VOP_N*()
1877  * API calls are required to make specific adjustments using the supplied
1878  * ncp pointers rather then just bogusly purging random vnodes.
1879  *
1880  * Invalidate all namecache entries to a particular vnode as well as
1881  * any direct children of that vnode in the namecache.  This is a
1882  * 'catch all' purge used by filesystems that do not know any better.
1883  *
1884  * A new vnode v_id is generated.  Note that no vnode will ever have a
1885  * v_id of 0.
1886  *
1887  * Note that the linkage between the vnode and its namecache entries will
1888  * be removed, but the namecache entries themselves might stay put due to
1889  * active references from elsewhere in the system or due to the existance of
1890  * the children.   The namecache topology is left intact even if we do not
1891  * know what the vnode association is.  Such entries will be marked
1892  * NCF_UNRESOLVED.
1893  *
1894  * XXX: Only time and the size of v_id prevents this from failing:
1895  * XXX: In theory we should hunt down all (struct vnode*, v_id)
1896  * XXX: soft references and nuke them, at least on the global
1897  * XXX: v_id wraparound.  The period of resistance can be extended
1898  * XXX: by incrementing each vnodes v_id individually instead of
1899  * XXX: using the global v_id.
1900  *
1901  * Does not support NCP_FSMID accumulation on invalidation (retflags is
1902  * not used).
1903  */
1904 void
1905 cache_purge(struct vnode *vp)
1906 {
1907 	static u_long nextid;
1908 	int retflags = 0;
1909 
1910 	cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN, &retflags);
1911 
1912 	/*
1913 	 * Calculate a new unique id for ".." handling
1914 	 */
1915 	do {
1916 		nextid++;
1917 	} while (nextid == vp->v_id || nextid == 0);
1918 	vp->v_id = nextid;
1919 }
1920 
1921 /*
1922  * Flush all entries referencing a particular filesystem.
1923  *
1924  * Since we need to check it anyway, we will flush all the invalid
1925  * entries at the same time.
1926  */
1927 void
1928 cache_purgevfs(struct mount *mp)
1929 {
1930 	struct nchashhead *nchpp;
1931 	struct namecache *ncp, *nnp;
1932 
1933 	/*
1934 	 * Scan hash tables for applicable entries.
1935 	 */
1936 	for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) {
1937 		ncp = LIST_FIRST(nchpp);
1938 		if (ncp)
1939 			cache_hold(ncp);
1940 		while (ncp) {
1941 			nnp = LIST_NEXT(ncp, nc_hash);
1942 			if (nnp)
1943 				cache_hold(nnp);
1944 			if (ncp->nc_mount == mp) {
1945 				cache_lock(ncp);
1946 				cache_zap(ncp);
1947 			} else {
1948 				cache_drop(ncp);
1949 			}
1950 			ncp = nnp;
1951 		}
1952 	}
1953 }
1954 
1955 static int disablecwd;
1956 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, "");
1957 
1958 static u_long numcwdcalls; STATNODE(CTLFLAG_RD, numcwdcalls, &numcwdcalls);
1959 static u_long numcwdfail1; STATNODE(CTLFLAG_RD, numcwdfail1, &numcwdfail1);
1960 static u_long numcwdfail2; STATNODE(CTLFLAG_RD, numcwdfail2, &numcwdfail2);
1961 static u_long numcwdfail3; STATNODE(CTLFLAG_RD, numcwdfail3, &numcwdfail3);
1962 static u_long numcwdfail4; STATNODE(CTLFLAG_RD, numcwdfail4, &numcwdfail4);
1963 static u_long numcwdfound; STATNODE(CTLFLAG_RD, numcwdfound, &numcwdfound);
1964 
1965 int
1966 __getcwd(struct __getcwd_args *uap)
1967 {
1968 	int buflen;
1969 	int error;
1970 	char *buf;
1971 	char *bp;
1972 
1973 	if (disablecwd)
1974 		return (ENODEV);
1975 
1976 	buflen = uap->buflen;
1977 	if (buflen < 2)
1978 		return (EINVAL);
1979 	if (buflen > MAXPATHLEN)
1980 		buflen = MAXPATHLEN;
1981 
1982 	buf = malloc(buflen, M_TEMP, M_WAITOK);
1983 	bp = kern_getcwd(buf, buflen, &error);
1984 	if (error == 0)
1985 		error = copyout(bp, uap->buf, strlen(bp) + 1);
1986 	free(buf, M_TEMP);
1987 	return (error);
1988 }
1989 
1990 char *
1991 kern_getcwd(char *buf, size_t buflen, int *error)
1992 {
1993 	struct proc *p = curproc;
1994 	char *bp;
1995 	int i, slash_prefixed;
1996 	struct filedesc *fdp;
1997 	struct namecache *ncp;
1998 
1999 	numcwdcalls++;
2000 	bp = buf;
2001 	bp += buflen - 1;
2002 	*bp = '\0';
2003 	fdp = p->p_fd;
2004 	slash_prefixed = 0;
2005 
2006 	ncp = fdp->fd_ncdir;
2007 	while (ncp && ncp != fdp->fd_nrdir && (ncp->nc_flag & NCF_ROOT) == 0) {
2008 		if (ncp->nc_flag & NCF_MOUNTPT) {
2009 			if (ncp->nc_mount == NULL) {
2010 				*error = EBADF;		/* forced unmount? */
2011 				return(NULL);
2012 			}
2013 			ncp = ncp->nc_parent;
2014 			continue;
2015 		}
2016 		for (i = ncp->nc_nlen - 1; i >= 0; i--) {
2017 			if (bp == buf) {
2018 				numcwdfail4++;
2019 				*error = ENOMEM;
2020 				return(NULL);
2021 			}
2022 			*--bp = ncp->nc_name[i];
2023 		}
2024 		if (bp == buf) {
2025 			numcwdfail4++;
2026 			*error = ENOMEM;
2027 			return(NULL);
2028 		}
2029 		*--bp = '/';
2030 		slash_prefixed = 1;
2031 		ncp = ncp->nc_parent;
2032 	}
2033 	if (ncp == NULL) {
2034 		numcwdfail2++;
2035 		*error = ENOENT;
2036 		return(NULL);
2037 	}
2038 	if (!slash_prefixed) {
2039 		if (bp == buf) {
2040 			numcwdfail4++;
2041 			*error = ENOMEM;
2042 			return(NULL);
2043 		}
2044 		*--bp = '/';
2045 	}
2046 	numcwdfound++;
2047 	*error = 0;
2048 	return (bp);
2049 }
2050 
2051 /*
2052  * Thus begins the fullpath magic.
2053  */
2054 
2055 #undef STATNODE
2056 #define STATNODE(name)							\
2057 	static u_int name;						\
2058 	SYSCTL_UINT(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, "")
2059 
2060 static int disablefullpath;
2061 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW,
2062     &disablefullpath, 0, "");
2063 
2064 STATNODE(numfullpathcalls);
2065 STATNODE(numfullpathfail1);
2066 STATNODE(numfullpathfail2);
2067 STATNODE(numfullpathfail3);
2068 STATNODE(numfullpathfail4);
2069 STATNODE(numfullpathfound);
2070 
2071 int
2072 cache_fullpath(struct proc *p, struct namecache *ncp, char **retbuf, char **freebuf)
2073 {
2074 	char *bp, *buf;
2075 	int i, slash_prefixed;
2076 	struct namecache *fd_nrdir;
2077 
2078 	numfullpathcalls--;
2079 
2080 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2081 	bp = buf + MAXPATHLEN - 1;
2082 	*bp = '\0';
2083 	if (p != NULL)
2084 		fd_nrdir = p->p_fd->fd_nrdir;
2085 	else
2086 		fd_nrdir = NULL;
2087 	slash_prefixed = 0;
2088 	while (ncp && ncp != fd_nrdir && (ncp->nc_flag & NCF_ROOT) == 0) {
2089 		if (ncp->nc_flag & NCF_MOUNTPT) {
2090 			if (ncp->nc_mount == NULL) {
2091 				free(buf, M_TEMP);
2092 				return(EBADF);
2093 			}
2094 			ncp = ncp->nc_parent;
2095 			continue;
2096 		}
2097 		for (i = ncp->nc_nlen - 1; i >= 0; i--) {
2098 			if (bp == buf) {
2099 				numfullpathfail4++;
2100 				free(buf, M_TEMP);
2101 				return(ENOMEM);
2102 			}
2103 			*--bp = ncp->nc_name[i];
2104 		}
2105 		if (bp == buf) {
2106 			numfullpathfail4++;
2107 			free(buf, M_TEMP);
2108 			return(ENOMEM);
2109 		}
2110 		*--bp = '/';
2111 		slash_prefixed = 1;
2112 		ncp = ncp->nc_parent;
2113 	}
2114 	if (ncp == NULL) {
2115 		numfullpathfail2++;
2116 		free(buf, M_TEMP);
2117 		return(ENOENT);
2118 	}
2119 	if (p != NULL && (ncp->nc_flag & NCF_ROOT) && ncp != fd_nrdir) {
2120 		bp = buf + MAXPATHLEN - 1;
2121 		*bp = '\0';
2122 		slash_prefixed = 0;
2123 	}
2124 	if (!slash_prefixed) {
2125 		if (bp == buf) {
2126 			numfullpathfail4++;
2127 			free(buf, M_TEMP);
2128 			return(ENOMEM);
2129 		}
2130 		*--bp = '/';
2131 	}
2132 	numfullpathfound++;
2133 	*retbuf = bp;
2134 	*freebuf = buf;
2135 
2136 	return(0);
2137 }
2138 
2139 int
2140 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf)
2141 {
2142 	struct namecache *ncp;
2143 
2144 	numfullpathcalls++;
2145 	if (disablefullpath)
2146 		return (ENODEV);
2147 
2148 	if (p == NULL)
2149 		return (EINVAL);
2150 
2151 	/* vn is NULL, client wants us to use p->p_textvp */
2152 	if (vn == NULL) {
2153 		if ((vn = p->p_textvp) == NULL)
2154 			return (EINVAL);
2155 	}
2156 	TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) {
2157 		if (ncp->nc_nlen)
2158 			break;
2159 	}
2160 	if (ncp == NULL)
2161 		return (EINVAL);
2162 
2163 	numfullpathcalls--;
2164 	return(cache_fullpath(p, ncp, retbuf, freebuf));
2165 }
2166