xref: /dflybsd-src/sys/kern/vfs_cache.c (revision efda3bd00c039d6845508b47bb18d1687c72154e)
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 1.76 2006/09/05 00:55:45 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  * This routine may only be called from outside this source module if
207  * nc_refs is already at least 1.
208  *
209  * This is a rare case where callers are allowed to hold a spinlock,
210  * so we can't ourselves.
211  */
212 static __inline
213 struct namecache *
214 _cache_hold(struct namecache *ncp)
215 {
216 	atomic_add_int(&ncp->nc_refs, 1);
217 	return(ncp);
218 }
219 
220 /*
221  * When dropping an entry, if only one ref remains and the entry has not
222  * been resolved, zap it.  Since the one reference is being dropped the
223  * entry had better not be locked.
224  */
225 static __inline
226 void
227 _cache_drop(struct namecache *ncp)
228 {
229 	KKASSERT(ncp->nc_refs > 0);
230 	if (ncp->nc_refs == 1 &&
231 	    (ncp->nc_flag & NCF_UNRESOLVED) &&
232 	    TAILQ_EMPTY(&ncp->nc_list)
233 	) {
234 		KKASSERT(ncp->nc_exlocks == 0);
235 		cache_lock(ncp);
236 		cache_zap(ncp);
237 	} else {
238 		atomic_subtract_int(&ncp->nc_refs, 1);
239 	}
240 }
241 
242 /*
243  * Link a new namecache entry to its parent.  Be careful to avoid races
244  * if vhold() blocks in the future.
245  */
246 static void
247 cache_link_parent(struct namecache *ncp, struct namecache *par)
248 {
249 	KKASSERT(ncp->nc_parent == NULL);
250 	ncp->nc_parent = par;
251 	if (TAILQ_EMPTY(&par->nc_list)) {
252 		TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
253 		/*
254 		 * Any vp associated with an ncp which has children must
255 		 * be held to prevent it from being recycled.
256 		 */
257 		if (par->nc_vp)
258 			vhold(par->nc_vp);
259 	} else {
260 		TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
261 	}
262 }
263 
264 /*
265  * Remove the parent association from a namecache structure.  If this is
266  * the last child of the parent the cache_drop(par) will attempt to
267  * recursively zap the parent.
268  */
269 static void
270 cache_unlink_parent(struct namecache *ncp)
271 {
272 	struct namecache *par;
273 
274 	if ((par = ncp->nc_parent) != NULL) {
275 		ncp->nc_parent = NULL;
276 		par = cache_hold(par);
277 		TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
278 		if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
279 			vdrop(par->nc_vp);
280 		cache_drop(par);
281 	}
282 }
283 
284 /*
285  * Allocate a new namecache structure.  Most of the code does not require
286  * zero-termination of the string but it makes vop_compat_ncreate() easier.
287  */
288 static struct namecache *
289 cache_alloc(int nlen)
290 {
291 	struct namecache *ncp;
292 
293 	ncp = kmalloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO);
294 	if (nlen)
295 		ncp->nc_name = kmalloc(nlen + 1, M_VFSCACHE, M_WAITOK);
296 	ncp->nc_nlen = nlen;
297 	ncp->nc_flag = NCF_UNRESOLVED;
298 	ncp->nc_error = ENOTCONN;	/* needs to be resolved */
299 	ncp->nc_refs = 1;
300 
301 	/*
302 	 * Construct a fake FSMID based on the time of day and a 32 bit
303 	 * roller for uniqueness.  This is used to generate a useful
304 	 * FSMID for filesystems which do not support it.
305 	 */
306 	ncp->nc_fsmid = cache_getnewfsmid();
307 	TAILQ_INIT(&ncp->nc_list);
308 	cache_lock(ncp);
309 	return(ncp);
310 }
311 
312 static void
313 cache_free(struct namecache *ncp)
314 {
315 	KKASSERT(ncp->nc_refs == 1 && ncp->nc_exlocks == 1);
316 	if (ncp->nc_name)
317 		kfree(ncp->nc_name, M_VFSCACHE);
318 	kfree(ncp, M_VFSCACHE);
319 }
320 
321 /*
322  * Ref and deref a namecache structure.
323  *
324  * Warning: caller may hold an unrelated read spinlock, which means we can't
325  * use read spinlocks here.
326  */
327 struct namecache *
328 cache_hold(struct namecache *ncp)
329 {
330 	return(_cache_hold(ncp));
331 }
332 
333 void
334 cache_drop(struct namecache *ncp)
335 {
336 	_cache_drop(ncp);
337 }
338 
339 /*
340  * Namespace locking.  The caller must already hold a reference to the
341  * namecache structure in order to lock/unlock it.  This function prevents
342  * the namespace from being created or destroyed by accessors other then
343  * the lock holder.
344  *
345  * Note that holding a locked namecache structure prevents other threads
346  * from making namespace changes (e.g. deleting or creating), prevents
347  * vnode association state changes by other threads, and prevents the
348  * namecache entry from being resolved or unresolved by other threads.
349  *
350  * The lock owner has full authority to associate/disassociate vnodes
351  * and resolve/unresolve the locked ncp.
352  *
353  * WARNING!  Holding a locked ncp will prevent a vnode from being destroyed
354  * or recycled, but it does NOT help you if the vnode had already initiated
355  * a recyclement.  If this is important, use cache_get() rather then
356  * cache_lock() (and deal with the differences in the way the refs counter
357  * is handled).  Or, alternatively, make an unconditional call to
358  * cache_validate() or cache_resolve() after cache_lock() returns.
359  */
360 void
361 cache_lock(struct namecache *ncp)
362 {
363 	thread_t td;
364 	int didwarn;
365 
366 	KKASSERT(ncp->nc_refs != 0);
367 	didwarn = 0;
368 	td = curthread;
369 
370 	for (;;) {
371 		if (ncp->nc_exlocks == 0) {
372 			ncp->nc_exlocks = 1;
373 			ncp->nc_locktd = td;
374 			/*
375 			 * The vp associated with a locked ncp must be held
376 			 * to prevent it from being recycled (which would
377 			 * cause the ncp to become unresolved).
378 			 *
379 			 * WARNING!  If VRECLAIMED is set the vnode could
380 			 * already be in the middle of a recycle.  Callers
381 			 * should not assume that nc_vp is usable when
382 			 * not NULL.  cache_vref() or cache_vget() must be
383 			 * called.
384 			 *
385 			 * XXX loop on race for later MPSAFE work.
386 			 */
387 			if (ncp->nc_vp)
388 				vhold(ncp->nc_vp);
389 			break;
390 		}
391 		if (ncp->nc_locktd == td) {
392 			++ncp->nc_exlocks;
393 			break;
394 		}
395 		ncp->nc_flag |= NCF_LOCKREQ;
396 		if (tsleep(ncp, 0, "clock", nclockwarn) == EWOULDBLOCK) {
397 			if (didwarn)
398 				continue;
399 			didwarn = 1;
400 			printf("[diagnostic] cache_lock: blocked on %p", ncp);
401 			if ((ncp->nc_flag & NCF_MOUNTPT) && ncp->nc_mount)
402 			    printf(" [MOUNTFROM %s]\n", ncp->nc_mount->mnt_stat.f_mntfromname);
403 			else
404 			    printf(" \"%*.*s\"\n",
405 				ncp->nc_nlen, ncp->nc_nlen,
406 				ncp->nc_name);
407 		}
408 	}
409 
410 	if (didwarn == 1) {
411 		printf("[diagnostic] cache_lock: unblocked %*.*s\n",
412 			ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
413 	}
414 }
415 
416 int
417 cache_lock_nonblock(struct namecache *ncp)
418 {
419 	thread_t td;
420 
421 	KKASSERT(ncp->nc_refs != 0);
422 	td = curthread;
423 	if (ncp->nc_exlocks == 0) {
424 		ncp->nc_exlocks = 1;
425 		ncp->nc_locktd = td;
426 		/*
427 		 * The vp associated with a locked ncp must be held
428 		 * to prevent it from being recycled (which would
429 		 * cause the ncp to become unresolved).
430 		 *
431 		 * WARNING!  If VRECLAIMED is set the vnode could
432 		 * already be in the middle of a recycle.  Callers
433 		 * should not assume that nc_vp is usable when
434 		 * not NULL.  cache_vref() or cache_vget() must be
435 		 * called.
436 		 *
437 		 * XXX loop on race for later MPSAFE work.
438 		 */
439 		if (ncp->nc_vp)
440 			vhold(ncp->nc_vp);
441 		return(0);
442 	} else {
443 		return(EWOULDBLOCK);
444 	}
445 }
446 
447 void
448 cache_unlock(struct namecache *ncp)
449 {
450 	thread_t td = curthread;
451 
452 	KKASSERT(ncp->nc_refs > 0);
453 	KKASSERT(ncp->nc_exlocks > 0);
454 	KKASSERT(ncp->nc_locktd == td);
455 	if (--ncp->nc_exlocks == 0) {
456 		if (ncp->nc_vp)
457 			vdrop(ncp->nc_vp);
458 		ncp->nc_locktd = NULL;
459 		if (ncp->nc_flag & NCF_LOCKREQ) {
460 			ncp->nc_flag &= ~NCF_LOCKREQ;
461 			wakeup(ncp);
462 		}
463 	}
464 }
465 
466 /*
467  * ref-and-lock, unlock-and-deref functions.
468  *
469  * This function is primarily used by nlookup.  Even though cache_lock
470  * holds the vnode, it is possible that the vnode may have already
471  * initiated a recyclement.  We want cache_get() to return a definitively
472  * usable vnode or a definitively unresolved ncp.
473  */
474 struct namecache *
475 cache_get(struct namecache *ncp)
476 {
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(ncp);
482 }
483 
484 int
485 cache_get_nonblock(struct namecache *ncp)
486 {
487 	/* XXX MP */
488 	if (ncp->nc_exlocks == 0 || ncp->nc_locktd == curthread) {
489 		_cache_hold(ncp);
490 		cache_lock(ncp);
491 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
492 			cache_setunresolved(ncp);
493 		return(0);
494 	}
495 	return(EWOULDBLOCK);
496 }
497 
498 void
499 cache_put(struct namecache *ncp)
500 {
501 	cache_unlock(ncp);
502 	_cache_drop(ncp);
503 }
504 
505 /*
506  * Resolve an unresolved ncp by associating a vnode with it.  If the
507  * vnode is NULL, a negative cache entry is created.
508  *
509  * The ncp should be locked on entry and will remain locked on return.
510  */
511 void
512 cache_setvp(struct namecache *ncp, struct vnode *vp)
513 {
514 	KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
515 	ncp->nc_vp = vp;
516 	if (vp != NULL) {
517 		/*
518 		 * Any vp associated with an ncp which has children must
519 		 * be held.  Any vp associated with a locked ncp must be held.
520 		 */
521 		if (!TAILQ_EMPTY(&ncp->nc_list))
522 			vhold(vp);
523 		TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode);
524 		if (ncp->nc_exlocks)
525 			vhold(vp);
526 
527 		/*
528 		 * Set auxillary flags
529 		 */
530 		switch(vp->v_type) {
531 		case VDIR:
532 			ncp->nc_flag |= NCF_ISDIR;
533 			break;
534 		case VLNK:
535 			ncp->nc_flag |= NCF_ISSYMLINK;
536 			/* XXX cache the contents of the symlink */
537 			break;
538 		default:
539 			break;
540 		}
541 		++numcache;
542 		ncp->nc_error = 0;
543 	} else {
544 		TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
545 		++numneg;
546 		ncp->nc_error = ENOENT;
547 	}
548 	ncp->nc_flag &= ~NCF_UNRESOLVED;
549 }
550 
551 void
552 cache_settimeout(struct namecache *ncp, int nticks)
553 {
554 	if ((ncp->nc_timeout = ticks + nticks) == 0)
555 		ncp->nc_timeout = 1;
556 }
557 
558 /*
559  * Disassociate the vnode or negative-cache association and mark a
560  * namecache entry as unresolved again.  Note that the ncp is still
561  * left in the hash table and still linked to its parent.
562  *
563  * The ncp should be locked and refd on entry and will remain locked and refd
564  * on return.
565  *
566  * This routine is normally never called on a directory containing children.
567  * However, NFS often does just that in its rename() code as a cop-out to
568  * avoid complex namespace operations.  This disconnects a directory vnode
569  * from its namecache and can cause the OLDAPI and NEWAPI to get out of
570  * sync.
571  *
572  * NOTE: NCF_FSMID must be cleared so a refurbishment of the ncp, such as
573  * in a create, properly propogates flag up the chain.
574  */
575 void
576 cache_setunresolved(struct namecache *ncp)
577 {
578 	struct vnode *vp;
579 
580 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
581 		ncp->nc_flag |= NCF_UNRESOLVED;
582 		ncp->nc_timeout = 0;
583 		ncp->nc_error = ENOTCONN;
584 		++numunres;
585 		if ((vp = ncp->nc_vp) != NULL) {
586 			--numcache;
587 			ncp->nc_vp = NULL;
588 			TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode);
589 
590 			/*
591 			 * Any vp associated with an ncp with children is
592 			 * held by that ncp.  Any vp associated with a locked
593 			 * ncp is held by that ncp.  These conditions must be
594 			 * undone when the vp is cleared out from the ncp.
595 			 */
596 			if (ncp->nc_flag & NCF_FSMID)
597 				vupdatefsmid(vp);
598 			if (!TAILQ_EMPTY(&ncp->nc_list))
599 				vdrop(vp);
600 			if (ncp->nc_exlocks)
601 				vdrop(vp);
602 		} else {
603 			TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
604 			--numneg;
605 		}
606 		ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK|
607 				  NCF_FSMID);
608 	}
609 }
610 
611 /*
612  * Invalidate portions of the namecache topology given a starting entry.
613  * The passed ncp is set to an unresolved state and:
614  *
615  * The passed ncp must be locked.
616  *
617  * CINV_DESTROY		- Set a flag in the passed ncp entry indicating
618  *			  that the physical underlying nodes have been
619  *			  destroyed... as in deleted.  For example, when
620  *			  a directory is removed.  This will cause record
621  *			  lookups on the name to no longer be able to find
622  *			  the record and tells the resolver to return failure
623  *			  rather then trying to resolve through the parent.
624  *
625  *			  The topology itself, including ncp->nc_name,
626  *			  remains intact.
627  *
628  *			  This only applies to the passed ncp, if CINV_CHILDREN
629  *			  is specified the children are not flagged.
630  *
631  * CINV_CHILDREN	- Set all children (recursively) to an unresolved
632  *			  state as well.
633  *
634  *			  Note that this will also have the side effect of
635  *			  cleaning out any unreferenced nodes in the topology
636  *			  from the leaves up as the recursion backs out.
637  *
638  * Note that the topology for any referenced nodes remains intact.
639  *
640  * It is possible for cache_inval() to race a cache_resolve(), meaning that
641  * the namecache entry may not actually be invalidated on return if it was
642  * revalidated while recursing down into its children.  This code guarentees
643  * that the node(s) will go through an invalidation cycle, but does not
644  * guarentee that they will remain in an invalidated state.
645  *
646  * Returns non-zero if a revalidation was detected during the invalidation
647  * recursion, zero otherwise.  Note that since only the original ncp is
648  * locked the revalidation ultimately can only indicate that the original ncp
649  * *MIGHT* no have been reresolved.
650  */
651 int
652 cache_inval(struct namecache *ncp, int flags)
653 {
654 	struct namecache *kid;
655 	struct namecache *nextkid;
656 	int rcnt = 0;
657 
658 	KKASSERT(ncp->nc_exlocks);
659 
660 	cache_setunresolved(ncp);
661 	if (flags & CINV_DESTROY)
662 		ncp->nc_flag |= NCF_DESTROYED;
663 
664 	if ((flags & CINV_CHILDREN) &&
665 	    (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL
666 	) {
667 		cache_hold(kid);
668 		cache_unlock(ncp);
669 		while (kid) {
670 			if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL)
671 				cache_hold(nextkid);
672 			if ((kid->nc_flag & NCF_UNRESOLVED) == 0 ||
673 			    TAILQ_FIRST(&kid->nc_list)
674 			) {
675 				cache_lock(kid);
676 				rcnt += cache_inval(kid, flags & ~CINV_DESTROY);
677 				cache_unlock(kid);
678 			}
679 			cache_drop(kid);
680 			kid = nextkid;
681 		}
682 		cache_lock(ncp);
683 	}
684 
685 	/*
686 	 * Someone could have gotten in there while ncp was unlocked,
687 	 * retry if so.
688 	 */
689 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
690 		++rcnt;
691 	return (rcnt);
692 }
693 
694 /*
695  * Invalidate a vnode's namecache associations.  To avoid races against
696  * the resolver we do not invalidate a node which we previously invalidated
697  * but which was then re-resolved while we were in the invalidation loop.
698  *
699  * Returns non-zero if any namecache entries remain after the invalidation
700  * loop completed.
701  *
702  * NOTE: unlike the namecache topology which guarentees that ncp's will not
703  * be ripped out of the topology while held, the vnode's v_namecache list
704  * has no such restriction.  NCP's can be ripped out of the list at virtually
705  * any time if not locked, even if held.
706  */
707 int
708 cache_inval_vp(struct vnode *vp, int flags)
709 {
710 	struct namecache *ncp;
711 	struct namecache *next;
712 
713 restart:
714 	ncp = TAILQ_FIRST(&vp->v_namecache);
715 	if (ncp)
716 		cache_hold(ncp);
717 	while (ncp) {
718 		/* loop entered with ncp held */
719 		if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
720 			cache_hold(next);
721 		cache_lock(ncp);
722 		if (ncp->nc_vp != vp) {
723 			printf("Warning: cache_inval_vp: race-A detected on "
724 				"%s\n", ncp->nc_name);
725 			cache_put(ncp);
726 			if (next)
727 				cache_drop(next);
728 			goto restart;
729 		}
730 		cache_inval(ncp, flags);
731 		cache_put(ncp);		/* also releases reference */
732 		ncp = next;
733 		if (ncp && ncp->nc_vp != vp) {
734 			printf("Warning: cache_inval_vp: race-B detected on "
735 				"%s\n", ncp->nc_name);
736 			cache_drop(ncp);
737 			goto restart;
738 		}
739 	}
740 	return(TAILQ_FIRST(&vp->v_namecache) != NULL);
741 }
742 
743 /*
744  * The source ncp has been renamed to the target ncp.  Both fncp and tncp
745  * must be locked.  Both will be set to unresolved, any children of tncp
746  * will be disconnected (the prior contents of the target is assumed to be
747  * destroyed by the rename operation, e.g. renaming over an empty directory),
748  * and all children of fncp will be moved to tncp.
749  *
750  * XXX the disconnection could pose a problem, check code paths to make
751  * sure any code that blocks can handle the parent being changed out from
752  * under it.  Maybe we should lock the children (watch out for deadlocks) ?
753  *
754  * After we return the caller has the option of calling cache_setvp() if
755  * the vnode of the new target ncp is known.
756  *
757  * Any process CD'd into any of the children will no longer be able to ".."
758  * back out.  An rm -rf can cause this situation to occur.
759  */
760 void
761 cache_rename(struct namecache *fncp, struct namecache *tncp)
762 {
763 	struct namecache *scan;
764 	int didwarn = 0;
765 
766 	cache_setunresolved(fncp);
767 	cache_setunresolved(tncp);
768 	while (cache_inval(tncp, CINV_CHILDREN) != 0) {
769 		if (didwarn++ % 10 == 0) {
770 			printf("Warning: cache_rename: race during "
771 				"rename %s->%s\n",
772 				fncp->nc_name, tncp->nc_name);
773 		}
774 		tsleep(tncp, 0, "mvrace", hz / 10);
775 		cache_setunresolved(tncp);
776 	}
777 	while ((scan = TAILQ_FIRST(&fncp->nc_list)) != NULL) {
778 		cache_hold(scan);
779 		cache_unlink_parent(scan);
780 		cache_link_parent(scan, tncp);
781 		if (scan->nc_flag & NCF_HASHED)
782 			cache_rehash(scan);
783 		cache_drop(scan);
784 	}
785 }
786 
787 /*
788  * vget the vnode associated with the namecache entry.  Resolve the namecache
789  * entry if necessary and deal with namecache/vp races.  The passed ncp must
790  * be referenced and may be locked.  The ncp's ref/locking state is not
791  * effected by this call.
792  *
793  * lk_type may be LK_SHARED, LK_EXCLUSIVE.  A ref'd, possibly locked
794  * (depending on the passed lk_type) will be returned in *vpp with an error
795  * of 0, or NULL will be returned in *vpp with a non-0 error code.  The
796  * most typical error is ENOENT, meaning that the ncp represents a negative
797  * cache hit and there is no vnode to retrieve, but other errors can occur
798  * too.
799  *
800  * The main race we have to deal with are namecache zaps.  The ncp itself
801  * will not disappear since it is referenced, and it turns out that the
802  * validity of the vp pointer can be checked simply by rechecking the
803  * contents of ncp->nc_vp.
804  */
805 int
806 cache_vget(struct namecache *ncp, struct ucred *cred,
807 	   int lk_type, struct vnode **vpp)
808 {
809 	struct vnode *vp;
810 	int error;
811 
812 again:
813 	vp = NULL;
814 	if (ncp->nc_flag & NCF_UNRESOLVED) {
815 		cache_lock(ncp);
816 		error = cache_resolve(ncp, cred);
817 		cache_unlock(ncp);
818 	} else {
819 		error = 0;
820 	}
821 	if (error == 0 && (vp = ncp->nc_vp) != NULL) {
822 		/*
823 		 * Accessing the vnode from the namecache is a bit
824 		 * dangerous.  Because there are no refs on the vnode, it
825 		 * could be in the middle of a reclaim.
826 		 */
827 		if (vp->v_flag & VRECLAIMED) {
828 			printf("Warning: vnode reclaim race detected in cache_vget on %p (%s)\n", vp, ncp->nc_name);
829 			cache_lock(ncp);
830 			cache_setunresolved(ncp);
831 			cache_unlock(ncp);
832 			goto again;
833 		}
834 		error = vget(vp, lk_type);
835 		if (error) {
836 			if (vp != ncp->nc_vp)
837 				goto again;
838 			vp = NULL;
839 		} else if (vp != ncp->nc_vp) {
840 			vput(vp);
841 			goto again;
842 		} else if (vp->v_flag & VRECLAIMED) {
843 			panic("vget succeeded on a VRECLAIMED node! vp %p", vp);
844 		}
845 	}
846 	if (error == 0 && vp == NULL)
847 		error = ENOENT;
848 	*vpp = vp;
849 	return(error);
850 }
851 
852 int
853 cache_vref(struct namecache *ncp, struct ucred *cred, struct vnode **vpp)
854 {
855 	struct vnode *vp;
856 	int error;
857 
858 again:
859 	vp = NULL;
860 	if (ncp->nc_flag & NCF_UNRESOLVED) {
861 		cache_lock(ncp);
862 		error = cache_resolve(ncp, cred);
863 		cache_unlock(ncp);
864 	} else {
865 		error = 0;
866 	}
867 	if (error == 0 && (vp = ncp->nc_vp) != NULL) {
868 		/*
869 		 * Since we did not obtain any locks, a cache zap
870 		 * race can occur here if the vnode is in the middle
871 		 * of being reclaimed and has not yet been able to
872 		 * clean out its cache node.  If that case occurs,
873 		 * we must lock and unresolve the cache, then loop
874 		 * to retry.
875 		 */
876 		if (vp->v_flag & VRECLAIMED) {
877 			printf("Warning: vnode reclaim race detected on cache_vref %p (%s)\n", vp, ncp->nc_name);
878 			cache_lock(ncp);
879 			cache_setunresolved(ncp);
880 			cache_unlock(ncp);
881 			goto again;
882 		}
883 		vref_initial(vp, 1);
884 	}
885 	if (error == 0 && vp == NULL)
886 		error = ENOENT;
887 	*vpp = vp;
888 	return(error);
889 }
890 
891 /*
892  * Recursively set the FSMID update flag for namecache nodes leading
893  * to root.  This will cause the next getattr or reclaim to increment the
894  * fsmid and mark the inode for lazy updating.
895  *
896  * Stop recursing when we hit a node whos NCF_FSMID flag is already set.
897  * This makes FSMIDs work in an Einsteinian fashion - where the observation
898  * effects the result.  In this case a program monitoring a higher level
899  * node will have detected some prior change and started its scan (clearing
900  * NCF_FSMID in higher level nodes), but since it has not yet observed the
901  * node where we find NCF_FSMID still set, we can safely make the related
902  * modification without interfering with the theorized program.
903  *
904  * This also means that FSMIDs cannot represent time-domain quantities
905  * in a hierarchical sense.  But the main reason for doing it this way
906  * is to reduce the amount of recursion that occurs in the critical path
907  * when e.g. a program is writing to a file that sits deep in a directory
908  * hierarchy.
909  */
910 void
911 cache_update_fsmid(struct namecache *ncp)
912 {
913 	struct vnode *vp;
914 	struct namecache *scan;
915 
916 	/*
917 	 * Warning: even if we get a non-NULL vp it could still be in the
918 	 * middle of a recyclement.  Don't do anything fancy, just set
919 	 * NCF_FSMID.
920 	 */
921 	if ((vp = ncp->nc_vp) != NULL) {
922 		TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
923 			for (scan = ncp; scan; scan = scan->nc_parent) {
924 				if (scan->nc_flag & NCF_FSMID)
925 					break;
926 				scan->nc_flag |= NCF_FSMID;
927 			}
928 		}
929 	} else {
930 		while (ncp && (ncp->nc_flag & NCF_FSMID) == 0) {
931 			ncp->nc_flag |= NCF_FSMID;
932 			ncp = ncp->nc_parent;
933 		}
934 	}
935 }
936 
937 void
938 cache_update_fsmid_vp(struct vnode *vp)
939 {
940 	struct namecache *ncp;
941 	struct namecache *scan;
942 
943 	TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
944 		for (scan = ncp; scan; scan = scan->nc_parent) {
945 			if (scan->nc_flag & NCF_FSMID)
946 				break;
947 			scan->nc_flag |= NCF_FSMID;
948 		}
949 	}
950 }
951 
952 /*
953  * If getattr is called on a vnode (e.g. a stat call), the filesystem
954  * may call this routine to determine if the namecache has the hierarchical
955  * change flag set, requiring the fsmid to be updated.
956  *
957  * Since 0 indicates no support, make sure the filesystem fsmid is at least
958  * 1.
959  */
960 int
961 cache_check_fsmid_vp(struct vnode *vp, int64_t *fsmid)
962 {
963 	struct namecache *ncp;
964 	int changed = 0;
965 
966 	TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
967 		if (ncp->nc_flag & NCF_FSMID) {
968 			ncp->nc_flag &= ~NCF_FSMID;
969 			changed = 1;
970 		}
971 	}
972 	if (*fsmid == 0)
973 		++*fsmid;
974 	if (changed)
975 		++*fsmid;
976 	return(changed);
977 }
978 
979 /*
980  * Obtain the FSMID for a vnode for filesystems which do not support
981  * a built-in FSMID.
982  */
983 int64_t
984 cache_sync_fsmid_vp(struct vnode *vp)
985 {
986 	struct namecache *ncp;
987 
988 	if ((ncp = TAILQ_FIRST(&vp->v_namecache)) != NULL) {
989 		if (ncp->nc_flag & NCF_FSMID) {
990 			ncp->nc_flag &= ~NCF_FSMID;
991 			++ncp->nc_fsmid;
992 		}
993 		return(ncp->nc_fsmid);
994 	}
995 	return(VNOVAL);
996 }
997 
998 /*
999  * Convert a directory vnode to a namecache record without any other
1000  * knowledge of the topology.  This ONLY works with directory vnodes and
1001  * is ONLY used by the NFS server.  dvp must be refd but unlocked, and the
1002  * returned ncp (if not NULL) will be held and unlocked.
1003  *
1004  * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned.
1005  * If 'makeit' is 1 we attempt to track-down and create the namecache topology
1006  * for dvp.  This will fail only if the directory has been deleted out from
1007  * under the caller.
1008  *
1009  * Callers must always check for a NULL return no matter the value of 'makeit'.
1010  *
1011  * To avoid underflowing the kernel stack each recursive call increments
1012  * the makeit variable.
1013  */
1014 
1015 static int cache_inefficient_scan(struct namecache *ncp, struct ucred *cred,
1016 				  struct vnode *dvp);
1017 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1018 				  struct vnode **saved_dvp);
1019 
1020 struct namecache *
1021 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit)
1022 {
1023 	struct namecache *ncp;
1024 	struct vnode *saved_dvp;
1025 	struct vnode *pvp;
1026 	int error;
1027 
1028 	ncp = NULL;
1029 	saved_dvp = NULL;
1030 
1031 	/*
1032 	 * Temporary debugging code to force the directory scanning code
1033 	 * to be exercised.
1034 	 */
1035 	if (ncvp_debug >= 3 && makeit && TAILQ_FIRST(&dvp->v_namecache)) {
1036 		ncp = TAILQ_FIRST(&dvp->v_namecache);
1037 		printf("cache_fromdvp: forcing %s\n", ncp->nc_name);
1038 		goto force;
1039 	}
1040 
1041 	/*
1042 	 * Loop until resolution, inside code will break out on error.
1043 	 */
1044 	while ((ncp = TAILQ_FIRST(&dvp->v_namecache)) == NULL && makeit) {
1045 force:
1046 		/*
1047 		 * If dvp is the root of its filesystem it should already
1048 		 * have a namecache pointer associated with it as a side
1049 		 * effect of the mount, but it may have been disassociated.
1050 		 */
1051 		if (dvp->v_flag & VROOT) {
1052 			ncp = cache_get(dvp->v_mount->mnt_ncp);
1053 			error = cache_resolve_mp(ncp);
1054 			cache_put(ncp);
1055 			if (ncvp_debug) {
1056 				printf("cache_fromdvp: resolve root of mount %p error %d",
1057 					dvp->v_mount, error);
1058 			}
1059 			if (error) {
1060 				if (ncvp_debug)
1061 					printf(" failed\n");
1062 				ncp = NULL;
1063 				break;
1064 			}
1065 			if (ncvp_debug)
1066 				printf(" succeeded\n");
1067 			continue;
1068 		}
1069 
1070 		/*
1071 		 * If we are recursed too deeply resort to an O(n^2)
1072 		 * algorithm to resolve the namecache topology.  The
1073 		 * resolved pvp is left referenced in saved_dvp to
1074 		 * prevent the tree from being destroyed while we loop.
1075 		 */
1076 		if (makeit > 20) {
1077 			error = cache_fromdvp_try(dvp, cred, &saved_dvp);
1078 			if (error) {
1079 				printf("lookupdotdot(longpath) failed %d "
1080 				       "dvp %p\n", error, dvp);
1081 				break;
1082 			}
1083 			continue;
1084 		}
1085 
1086 		/*
1087 		 * Get the parent directory and resolve its ncp.
1088 		 */
1089 		error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred);
1090 		if (error) {
1091 			printf("lookupdotdot failed %d dvp %p\n", error, dvp);
1092 			break;
1093 		}
1094 		vn_unlock(pvp);
1095 
1096 		/*
1097 		 * Reuse makeit as a recursion depth counter.
1098 		 */
1099 		ncp = cache_fromdvp(pvp, cred, makeit + 1);
1100 		vrele(pvp);
1101 		if (ncp == NULL)
1102 			break;
1103 
1104 		/*
1105 		 * Do an inefficient scan of pvp (embodied by ncp) to look
1106 		 * for dvp.  This will create a namecache record for dvp on
1107 		 * success.  We loop up to recheck on success.
1108 		 *
1109 		 * ncp and dvp are both held but not locked.
1110 		 */
1111 		error = cache_inefficient_scan(ncp, cred, dvp);
1112 		cache_drop(ncp);
1113 		if (error) {
1114 			printf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n",
1115 				pvp, ncp->nc_name, dvp);
1116 			ncp = NULL;
1117 			break;
1118 		}
1119 		if (ncvp_debug) {
1120 			printf("cache_fromdvp: scan %p (%s) succeeded\n",
1121 				pvp, ncp->nc_name);
1122 		}
1123 	}
1124 	if (ncp)
1125 		cache_hold(ncp);
1126 	if (saved_dvp)
1127 		vrele(saved_dvp);
1128 	return (ncp);
1129 }
1130 
1131 /*
1132  * Go up the chain of parent directories until we find something
1133  * we can resolve into the namecache.  This is very inefficient.
1134  */
1135 static
1136 int
1137 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1138 		  struct vnode **saved_dvp)
1139 {
1140 	struct namecache *ncp;
1141 	struct vnode *pvp;
1142 	int error;
1143 	static time_t last_fromdvp_report;
1144 
1145 	/*
1146 	 * Loop getting the parent directory vnode until we get something we
1147 	 * can resolve in the namecache.
1148 	 */
1149 	vref(dvp);
1150 	for (;;) {
1151 		error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred);
1152 		if (error) {
1153 			vrele(dvp);
1154 			return (error);
1155 		}
1156 		vn_unlock(pvp);
1157 		if ((ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) {
1158 			cache_hold(ncp);
1159 			vrele(pvp);
1160 			break;
1161 		}
1162 		if (pvp->v_flag & VROOT) {
1163 			ncp = cache_get(pvp->v_mount->mnt_ncp);
1164 			error = cache_resolve_mp(ncp);
1165 			cache_unlock(ncp);
1166 			vrele(pvp);
1167 			if (error) {
1168 				cache_drop(ncp);
1169 				vrele(dvp);
1170 				return (error);
1171 			}
1172 			break;
1173 		}
1174 		vrele(dvp);
1175 		dvp = pvp;
1176 	}
1177 	if (last_fromdvp_report != time_second) {
1178 		last_fromdvp_report = time_second;
1179 		printf("Warning: extremely inefficient path resolution on %s\n",
1180 			ncp->nc_name);
1181 	}
1182 	error = cache_inefficient_scan(ncp, cred, dvp);
1183 
1184 	/*
1185 	 * Hopefully dvp now has a namecache record associated with it.
1186 	 * Leave it referenced to prevent the kernel from recycling the
1187 	 * vnode.  Otherwise extremely long directory paths could result
1188 	 * in endless recycling.
1189 	 */
1190 	if (*saved_dvp)
1191 	    vrele(*saved_dvp);
1192 	*saved_dvp = dvp;
1193 	return (error);
1194 }
1195 
1196 
1197 /*
1198  * Do an inefficient scan of the directory represented by ncp looking for
1199  * the directory vnode dvp.  ncp must be held but not locked on entry and
1200  * will be held on return.  dvp must be refd but not locked on entry and
1201  * will remain refd on return.
1202  *
1203  * Why do this at all?  Well, due to its stateless nature the NFS server
1204  * converts file handles directly to vnodes without necessarily going through
1205  * the namecache ops that would otherwise create the namecache topology
1206  * leading to the vnode.  We could either (1) Change the namecache algorithms
1207  * to allow disconnect namecache records that are re-merged opportunistically,
1208  * or (2) Make the NFS server backtrack and scan to recover a connected
1209  * namecache topology in order to then be able to issue new API lookups.
1210  *
1211  * It turns out that (1) is a huge mess.  It takes a nice clean set of
1212  * namecache algorithms and introduces a lot of complication in every subsystem
1213  * that calls into the namecache to deal with the re-merge case, especially
1214  * since we are using the namecache to placehold negative lookups and the
1215  * vnode might not be immediately assigned. (2) is certainly far less
1216  * efficient then (1), but since we are only talking about directories here
1217  * (which are likely to remain cached), the case does not actually run all
1218  * that often and has the supreme advantage of not polluting the namecache
1219  * algorithms.
1220  */
1221 static int
1222 cache_inefficient_scan(struct namecache *ncp, struct ucred *cred,
1223 		       struct vnode *dvp)
1224 {
1225 	struct nlcomponent nlc;
1226 	struct namecache *rncp;
1227 	struct dirent *den;
1228 	struct vnode *pvp;
1229 	struct vattr vat;
1230 	struct iovec iov;
1231 	struct uio uio;
1232 	int blksize;
1233 	int eofflag;
1234 	int bytes;
1235 	char *rbuf;
1236 	int error;
1237 
1238 	vat.va_blocksize = 0;
1239 	if ((error = VOP_GETATTR(dvp, &vat)) != 0)
1240 		return (error);
1241 	if ((error = cache_vref(ncp, cred, &pvp)) != 0)
1242 		return (error);
1243 	if (ncvp_debug)
1244 		printf("inefficient_scan: directory iosize %ld vattr fileid = %ld\n", vat.va_blocksize, (long)vat.va_fileid);
1245 	if ((blksize = vat.va_blocksize) == 0)
1246 		blksize = DEV_BSIZE;
1247 	rbuf = kmalloc(blksize, M_TEMP, M_WAITOK);
1248 	rncp = NULL;
1249 
1250 	eofflag = 0;
1251 	uio.uio_offset = 0;
1252 again:
1253 	iov.iov_base = rbuf;
1254 	iov.iov_len = blksize;
1255 	uio.uio_iov = &iov;
1256 	uio.uio_iovcnt = 1;
1257 	uio.uio_resid = blksize;
1258 	uio.uio_segflg = UIO_SYSSPACE;
1259 	uio.uio_rw = UIO_READ;
1260 	uio.uio_td = curthread;
1261 
1262 	if (ncvp_debug >= 2)
1263 		printf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset);
1264 	error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL);
1265 	if (error == 0) {
1266 		den = (struct dirent *)rbuf;
1267 		bytes = blksize - uio.uio_resid;
1268 
1269 		while (bytes > 0) {
1270 			if (ncvp_debug >= 2) {
1271 				printf("cache_inefficient_scan: %*.*s\n",
1272 					den->d_namlen, den->d_namlen,
1273 					den->d_name);
1274 			}
1275 			if (den->d_type != DT_WHT &&
1276 			    den->d_ino == vat.va_fileid) {
1277 				if (ncvp_debug) {
1278 					printf("cache_inefficient_scan: "
1279 					       "MATCHED inode %ld path %s/%*.*s\n",
1280 					       vat.va_fileid, ncp->nc_name,
1281 					       den->d_namlen, den->d_namlen,
1282 					       den->d_name);
1283 				}
1284 				nlc.nlc_nameptr = den->d_name;
1285 				nlc.nlc_namelen = den->d_namlen;
1286 				rncp = cache_nlookup(ncp, &nlc);
1287 				KKASSERT(rncp != NULL);
1288 				break;
1289 			}
1290 			bytes -= _DIRENT_DIRSIZ(den);
1291 			den = _DIRENT_NEXT(den);
1292 		}
1293 		if (rncp == NULL && eofflag == 0 && uio.uio_resid != blksize)
1294 			goto again;
1295 	}
1296 	vrele(pvp);
1297 	if (rncp) {
1298 		if (rncp->nc_flag & NCF_UNRESOLVED) {
1299 			cache_setvp(rncp, dvp);
1300 			if (ncvp_debug >= 2) {
1301 				printf("cache_inefficient_scan: setvp %s/%s = %p\n",
1302 					ncp->nc_name, rncp->nc_name, dvp);
1303 			}
1304 		} else {
1305 			if (ncvp_debug >= 2) {
1306 				printf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n",
1307 					ncp->nc_name, rncp->nc_name, dvp,
1308 					rncp->nc_vp);
1309 			}
1310 		}
1311 		if (rncp->nc_vp == NULL)
1312 			error = rncp->nc_error;
1313 		cache_put(rncp);
1314 	} else {
1315 		printf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n",
1316 			dvp, ncp->nc_name);
1317 		error = ENOENT;
1318 	}
1319 	kfree(rbuf, M_TEMP);
1320 	return (error);
1321 }
1322 
1323 /*
1324  * Zap a namecache entry.  The ncp is unconditionally set to an unresolved
1325  * state, which disassociates it from its vnode or ncneglist.
1326  *
1327  * Then, if there are no additional references to the ncp and no children,
1328  * the ncp is removed from the topology and destroyed.  This function will
1329  * also run through the nc_parent chain and destroy parent ncps if possible.
1330  * As a side benefit, it turns out the only conditions that allow running
1331  * up the chain are also the conditions to ensure no deadlock will occur.
1332  *
1333  * References and/or children may exist if the ncp is in the middle of the
1334  * topology, preventing the ncp from being destroyed.
1335  *
1336  * This function must be called with the ncp held and locked and will unlock
1337  * and drop it during zapping.
1338  */
1339 static void
1340 cache_zap(struct namecache *ncp)
1341 {
1342 	struct namecache *par;
1343 
1344 	/*
1345 	 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED.
1346 	 */
1347 	cache_setunresolved(ncp);
1348 
1349 	/*
1350 	 * Try to scrap the entry and possibly tail-recurse on its parent.
1351 	 * We only scrap unref'd (other then our ref) unresolved entries,
1352 	 * we do not scrap 'live' entries.
1353 	 */
1354 	while (ncp->nc_flag & NCF_UNRESOLVED) {
1355 		/*
1356 		 * Someone other then us has a ref, stop.
1357 		 */
1358 		if (ncp->nc_refs > 1)
1359 			goto done;
1360 
1361 		/*
1362 		 * We have children, stop.
1363 		 */
1364 		if (!TAILQ_EMPTY(&ncp->nc_list))
1365 			goto done;
1366 
1367 		/*
1368 		 * Remove ncp from the topology: hash table and parent linkage.
1369 		 */
1370 		if (ncp->nc_flag & NCF_HASHED) {
1371 			ncp->nc_flag &= ~NCF_HASHED;
1372 			LIST_REMOVE(ncp, nc_hash);
1373 		}
1374 		if ((par = ncp->nc_parent) != NULL) {
1375 			par = cache_hold(par);
1376 			TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
1377 			ncp->nc_parent = NULL;
1378 			if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
1379 				vdrop(par->nc_vp);
1380 		}
1381 
1382 		/*
1383 		 * ncp should not have picked up any refs.  Physically
1384 		 * destroy the ncp.
1385 		 */
1386 		KKASSERT(ncp->nc_refs == 1);
1387 		--numunres;
1388 		/* cache_unlock(ncp) not required */
1389 		ncp->nc_refs = -1;	/* safety */
1390 		if (ncp->nc_name)
1391 			kfree(ncp->nc_name, M_VFSCACHE);
1392 		kfree(ncp, M_VFSCACHE);
1393 
1394 		/*
1395 		 * Loop on the parent (it may be NULL).  Only bother looping
1396 		 * if the parent has a single ref (ours), which also means
1397 		 * we can lock it trivially.
1398 		 */
1399 		ncp = par;
1400 		if (ncp == NULL)
1401 			return;
1402 		if (ncp->nc_refs != 1) {
1403 			cache_drop(ncp);
1404 			return;
1405 		}
1406 		KKASSERT(par->nc_exlocks == 0);
1407 		cache_lock(ncp);
1408 	}
1409 done:
1410 	cache_unlock(ncp);
1411 	atomic_subtract_int(&ncp->nc_refs, 1);
1412 }
1413 
1414 static enum { CHI_LOW, CHI_HIGH } cache_hysteresis_state = CHI_LOW;
1415 
1416 static __inline
1417 void
1418 cache_hysteresis(void)
1419 {
1420 	/*
1421 	 * Don't cache too many negative hits.  We use hysteresis to reduce
1422 	 * the impact on the critical path.
1423 	 */
1424 	switch(cache_hysteresis_state) {
1425 	case CHI_LOW:
1426 		if (numneg > MINNEG && numneg * ncnegfactor > numcache) {
1427 			cache_cleanneg(10);
1428 			cache_hysteresis_state = CHI_HIGH;
1429 		}
1430 		break;
1431 	case CHI_HIGH:
1432 		if (numneg > MINNEG * 9 / 10 &&
1433 		    numneg * ncnegfactor * 9 / 10 > numcache
1434 		) {
1435 			cache_cleanneg(10);
1436 		} else {
1437 			cache_hysteresis_state = CHI_LOW;
1438 		}
1439 		break;
1440 	}
1441 }
1442 
1443 /*
1444  * NEW NAMECACHE LOOKUP API
1445  *
1446  * Lookup an entry in the cache.  A locked, referenced, non-NULL
1447  * entry is *always* returned, even if the supplied component is illegal.
1448  * The resulting namecache entry should be returned to the system with
1449  * cache_put() or cache_unlock() + cache_drop().
1450  *
1451  * namecache locks are recursive but care must be taken to avoid lock order
1452  * reversals.
1453  *
1454  * Nobody else will be able to manipulate the associated namespace (e.g.
1455  * create, delete, rename, rename-target) until the caller unlocks the
1456  * entry.
1457  *
1458  * The returned entry will be in one of three states:  positive hit (non-null
1459  * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set).
1460  * Unresolved entries must be resolved through the filesystem to associate the
1461  * vnode and/or determine whether a positive or negative hit has occured.
1462  *
1463  * It is not necessary to lock a directory in order to lock namespace under
1464  * that directory.  In fact, it is explicitly not allowed to do that.  A
1465  * directory is typically only locked when being created, renamed, or
1466  * destroyed.
1467  *
1468  * The directory (par) may be unresolved, in which case any returned child
1469  * will likely also be marked unresolved.  Likely but not guarenteed.  Since
1470  * the filesystem lookup requires a resolved directory vnode the caller is
1471  * responsible for resolving the namecache chain top-down.  This API
1472  * specifically allows whole chains to be created in an unresolved state.
1473  */
1474 struct namecache *
1475 cache_nlookup(struct namecache *par, struct nlcomponent *nlc)
1476 {
1477 	struct namecache *ncp;
1478 	struct namecache *new_ncp;
1479 	struct nchashhead *nchpp;
1480 	u_int32_t hash;
1481 	globaldata_t gd;
1482 
1483 	numcalls++;
1484 	gd = mycpu;
1485 
1486 	/*
1487 	 * Try to locate an existing entry
1488 	 */
1489 	hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
1490 	hash = fnv_32_buf(&par, sizeof(par), hash);
1491 	new_ncp = NULL;
1492 restart:
1493 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1494 		numchecks++;
1495 
1496 		/*
1497 		 * Zap entries that have timed out.
1498 		 */
1499 		if (ncp->nc_timeout &&
1500 		    (int)(ncp->nc_timeout - ticks) < 0 &&
1501 		    (ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
1502 		    ncp->nc_exlocks == 0
1503 		) {
1504 			cache_zap(cache_get(ncp));
1505 			goto restart;
1506 		}
1507 
1508 		/*
1509 		 * Break out if we find a matching entry.  Note that
1510 		 * UNRESOLVED entries may match, but DESTROYED entries
1511 		 * do not.
1512 		 */
1513 		if (ncp->nc_parent == par &&
1514 		    ncp->nc_nlen == nlc->nlc_namelen &&
1515 		    bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
1516 		    (ncp->nc_flag & NCF_DESTROYED) == 0
1517 		) {
1518 			if (cache_get_nonblock(ncp) == 0) {
1519 				if (new_ncp)
1520 					cache_free(new_ncp);
1521 				goto found;
1522 			}
1523 			cache_get(ncp);
1524 			cache_put(ncp);
1525 			goto restart;
1526 		}
1527 	}
1528 
1529 	/*
1530 	 * We failed to locate an entry, create a new entry and add it to
1531 	 * the cache.  We have to relookup after possibly blocking in
1532 	 * malloc.
1533 	 */
1534 	if (new_ncp == NULL) {
1535 		new_ncp = cache_alloc(nlc->nlc_namelen);
1536 		goto restart;
1537 	}
1538 
1539 	ncp = new_ncp;
1540 
1541 	/*
1542 	 * Initialize as a new UNRESOLVED entry, lock (non-blocking),
1543 	 * and link to the parent.  The mount point is usually inherited
1544 	 * from the parent unless this is a special case such as a mount
1545 	 * point where nlc_namelen is 0.  The caller is responsible for
1546 	 * setting nc_mount in that case.  If nlc_namelen is 0 nc_name will
1547 	 * be NULL.
1548 	 */
1549 	if (nlc->nlc_namelen) {
1550 		bcopy(nlc->nlc_nameptr, ncp->nc_name, nlc->nlc_namelen);
1551 		ncp->nc_name[nlc->nlc_namelen] = 0;
1552 		ncp->nc_mount = par->nc_mount;
1553 	}
1554 	nchpp = NCHHASH(hash);
1555 	LIST_INSERT_HEAD(nchpp, ncp, nc_hash);
1556 	ncp->nc_flag |= NCF_HASHED;
1557 	cache_link_parent(ncp, par);
1558 found:
1559 	/*
1560 	 * stats and namecache size management
1561 	 */
1562 	if (ncp->nc_flag & NCF_UNRESOLVED)
1563 		++gd->gd_nchstats->ncs_miss;
1564 	else if (ncp->nc_vp)
1565 		++gd->gd_nchstats->ncs_goodhits;
1566 	else
1567 		++gd->gd_nchstats->ncs_neghits;
1568 	cache_hysteresis();
1569 	return(ncp);
1570 }
1571 
1572 /*
1573  * Given a locked ncp, validate that the vnode, if present, is actually
1574  * usable.  If it is not usable set the ncp to an unresolved state.
1575  */
1576 void
1577 cache_validate(struct namecache *ncp)
1578 {
1579 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1580 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1581 			cache_setunresolved(ncp);
1582 	}
1583 }
1584 
1585 /*
1586  * Resolve an unresolved namecache entry, generally by looking it up.
1587  * The passed ncp must be locked and refd.
1588  *
1589  * Theoretically since a vnode cannot be recycled while held, and since
1590  * the nc_parent chain holds its vnode as long as children exist, the
1591  * direct parent of the cache entry we are trying to resolve should
1592  * have a valid vnode.  If not then generate an error that we can
1593  * determine is related to a resolver bug.
1594  *
1595  * However, if a vnode was in the middle of a recyclement when the NCP
1596  * got locked, ncp->nc_vp might point to a vnode that is about to become
1597  * invalid.  cache_resolve() handles this case by unresolving the entry
1598  * and then re-resolving it.
1599  *
1600  * Note that successful resolution does not necessarily return an error
1601  * code of 0.  If the ncp resolves to a negative cache hit then ENOENT
1602  * will be returned.
1603  */
1604 int
1605 cache_resolve(struct namecache *ncp, struct ucred *cred)
1606 {
1607 	struct namecache *par;
1608 	int error;
1609 
1610 restart:
1611 	/*
1612 	 * If the ncp is already resolved we have nothing to do.  However,
1613 	 * we do want to guarentee that a usable vnode is returned when
1614 	 * a vnode is present, so make sure it hasn't been reclaimed.
1615 	 */
1616 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1617 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1618 			cache_setunresolved(ncp);
1619 		if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
1620 			return (ncp->nc_error);
1621 	}
1622 
1623 	/*
1624 	 * Mount points need special handling because the parent does not
1625 	 * belong to the same filesystem as the ncp.
1626 	 */
1627 	if (ncp->nc_flag & NCF_MOUNTPT)
1628 		return (cache_resolve_mp(ncp));
1629 
1630 	/*
1631 	 * We expect an unbroken chain of ncps to at least the mount point,
1632 	 * and even all the way to root (but this code doesn't have to go
1633 	 * past the mount point).
1634 	 */
1635 	if (ncp->nc_parent == NULL) {
1636 		printf("EXDEV case 1 %p %*.*s\n", ncp,
1637 			ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
1638 		ncp->nc_error = EXDEV;
1639 		return(ncp->nc_error);
1640 	}
1641 
1642 	/*
1643 	 * The vp's of the parent directories in the chain are held via vhold()
1644 	 * due to the existance of the child, and should not disappear.
1645 	 * However, there are cases where they can disappear:
1646 	 *
1647 	 *	- due to filesystem I/O errors.
1648 	 *	- due to NFS being stupid about tracking the namespace and
1649 	 *	  destroys the namespace for entire directories quite often.
1650 	 *	- due to forced unmounts.
1651 	 *	- due to an rmdir (parent will be marked DESTROYED)
1652 	 *
1653 	 * When this occurs we have to track the chain backwards and resolve
1654 	 * it, looping until the resolver catches up to the current node.  We
1655 	 * could recurse here but we might run ourselves out of kernel stack
1656 	 * so we do it in a more painful manner.  This situation really should
1657 	 * not occur all that often, or if it does not have to go back too
1658 	 * many nodes to resolve the ncp.
1659 	 */
1660 	while (ncp->nc_parent->nc_vp == NULL) {
1661 		/*
1662 		 * This case can occur if a process is CD'd into a
1663 		 * directory which is then rmdir'd.  If the parent is marked
1664 		 * destroyed there is no point trying to resolve it.
1665 		 */
1666 		if (ncp->nc_parent->nc_flag & NCF_DESTROYED)
1667 			return(ENOENT);
1668 
1669 		par = ncp->nc_parent;
1670 		while (par->nc_parent && par->nc_parent->nc_vp == NULL)
1671 			par = par->nc_parent;
1672 		if (par->nc_parent == NULL) {
1673 			printf("EXDEV case 2 %*.*s\n",
1674 				par->nc_nlen, par->nc_nlen, par->nc_name);
1675 			return (EXDEV);
1676 		}
1677 		printf("[diagnostic] cache_resolve: had to recurse on %*.*s\n",
1678 			par->nc_nlen, par->nc_nlen, par->nc_name);
1679 		/*
1680 		 * The parent is not set in stone, ref and lock it to prevent
1681 		 * it from disappearing.  Also note that due to renames it
1682 		 * is possible for our ncp to move and for par to no longer
1683 		 * be one of its parents.  We resolve it anyway, the loop
1684 		 * will handle any moves.
1685 		 */
1686 		cache_get(par);
1687 		if (par->nc_flag & NCF_MOUNTPT) {
1688 			cache_resolve_mp(par);
1689 		} else if (par->nc_parent->nc_vp == NULL) {
1690 			printf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name);
1691 			cache_put(par);
1692 			continue;
1693 		} else if (par->nc_flag & NCF_UNRESOLVED) {
1694 			par->nc_error = VOP_NRESOLVE(par, cred);
1695 		}
1696 		if ((error = par->nc_error) != 0) {
1697 			if (par->nc_error != EAGAIN) {
1698 				printf("EXDEV case 3 %*.*s error %d\n",
1699 				    par->nc_nlen, par->nc_nlen, par->nc_name,
1700 				    par->nc_error);
1701 				cache_put(par);
1702 				return(error);
1703 			}
1704 			printf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n",
1705 				par, par->nc_nlen, par->nc_nlen, par->nc_name);
1706 		}
1707 		cache_put(par);
1708 		/* loop */
1709 	}
1710 
1711 	/*
1712 	 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected
1713 	 * ncp's and reattach them.  If this occurs the original ncp is marked
1714 	 * EAGAIN to force a relookup.
1715 	 *
1716 	 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed
1717 	 * ncp must already be resolved.
1718 	 */
1719 	KKASSERT((ncp->nc_flag & NCF_MOUNTPT) == 0);
1720 	ncp->nc_error = VOP_NRESOLVE(ncp, cred);
1721 	/*vop_nresolve(*ncp->nc_parent->nc_vp->v_ops, ncp, cred);*/
1722 	if (ncp->nc_error == EAGAIN) {
1723 		printf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n",
1724 			ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
1725 		goto restart;
1726 	}
1727 	return(ncp->nc_error);
1728 }
1729 
1730 /*
1731  * Resolve the ncp associated with a mount point.  Such ncp's almost always
1732  * remain resolved and this routine is rarely called.  NFS MPs tends to force
1733  * re-resolution more often due to its mac-truck-smash-the-namecache
1734  * method of tracking namespace changes.
1735  *
1736  * The semantics for this call is that the passed ncp must be locked on
1737  * entry and will be locked on return.  However, if we actually have to
1738  * resolve the mount point we temporarily unlock the entry in order to
1739  * avoid race-to-root deadlocks due to e.g. dead NFS mounts.  Because of
1740  * the unlock we have to recheck the flags after we relock.
1741  */
1742 static int
1743 cache_resolve_mp(struct namecache *ncp)
1744 {
1745 	struct vnode *vp;
1746 	struct mount *mp = ncp->nc_mount;
1747 	int error;
1748 
1749 	KKASSERT(mp != NULL);
1750 
1751 	/*
1752 	 * If the ncp is already resolved we have nothing to do.  However,
1753 	 * we do want to guarentee that a usable vnode is returned when
1754 	 * a vnode is present, so make sure it hasn't been reclaimed.
1755 	 */
1756 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1757 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1758 			cache_setunresolved(ncp);
1759 	}
1760 
1761 	if (ncp->nc_flag & NCF_UNRESOLVED) {
1762 		cache_unlock(ncp);
1763 		while (vfs_busy(mp, 0))
1764 			;
1765 		error = VFS_ROOT(mp, &vp);
1766 		cache_lock(ncp);
1767 
1768 		/*
1769 		 * recheck the ncp state after relocking.
1770 		 */
1771 		if (ncp->nc_flag & NCF_UNRESOLVED) {
1772 			ncp->nc_error = error;
1773 			if (error == 0) {
1774 				cache_setvp(ncp, vp);
1775 				vput(vp);
1776 			} else {
1777 				printf("[diagnostic] cache_resolve_mp: failed to resolve mount %p\n", mp);
1778 				cache_setvp(ncp, NULL);
1779 			}
1780 		} else if (error == 0) {
1781 			vput(vp);
1782 		}
1783 		vfs_unbusy(mp);
1784 	}
1785 	return(ncp->nc_error);
1786 }
1787 
1788 void
1789 cache_cleanneg(int count)
1790 {
1791 	struct namecache *ncp;
1792 
1793 	/*
1794 	 * Automode from the vnlru proc - clean out 10% of the negative cache
1795 	 * entries.
1796 	 */
1797 	if (count == 0)
1798 		count = numneg / 10 + 1;
1799 
1800 	/*
1801 	 * Attempt to clean out the specified number of negative cache
1802 	 * entries.
1803 	 */
1804 	while (count) {
1805 		ncp = TAILQ_FIRST(&ncneglist);
1806 		if (ncp == NULL) {
1807 			KKASSERT(numneg == 0);
1808 			break;
1809 		}
1810 		TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
1811 		TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
1812 		if (cache_get_nonblock(ncp) == 0)
1813 			cache_zap(ncp);
1814 		--count;
1815 	}
1816 }
1817 
1818 /*
1819  * Rehash a ncp.  Rehashing is typically required if the name changes (should
1820  * not generally occur) or the parent link changes.  This function will
1821  * unhash the ncp if the ncp is no longer hashable.
1822  */
1823 static void
1824 cache_rehash(struct namecache *ncp)
1825 {
1826 	struct nchashhead *nchpp;
1827 	u_int32_t hash;
1828 
1829 	if (ncp->nc_flag & NCF_HASHED) {
1830 		ncp->nc_flag &= ~NCF_HASHED;
1831 		LIST_REMOVE(ncp, nc_hash);
1832 	}
1833 	if (ncp->nc_nlen && ncp->nc_parent) {
1834 		hash = fnv_32_buf(ncp->nc_name, ncp->nc_nlen, FNV1_32_INIT);
1835 		hash = fnv_32_buf(&ncp->nc_parent,
1836 					sizeof(ncp->nc_parent), hash);
1837 		nchpp = NCHHASH(hash);
1838 		LIST_INSERT_HEAD(nchpp, ncp, nc_hash);
1839 		ncp->nc_flag |= NCF_HASHED;
1840 	}
1841 }
1842 
1843 /*
1844  * Name cache initialization, from vfsinit() when we are booting
1845  */
1846 void
1847 nchinit(void)
1848 {
1849 	int i;
1850 	globaldata_t gd;
1851 
1852 	/* initialise per-cpu namecache effectiveness statistics. */
1853 	for (i = 0; i < ncpus; ++i) {
1854 		gd = globaldata_find(i);
1855 		gd->gd_nchstats = &nchstats[i];
1856 	}
1857 	TAILQ_INIT(&ncneglist);
1858 	nchashtbl = hashinit(desiredvnodes*2, M_VFSCACHE, &nchash);
1859 	nclockwarn = 1 * hz;
1860 }
1861 
1862 /*
1863  * Called from start_init() to bootstrap the root filesystem.  Returns
1864  * a referenced, unlocked namecache record.
1865  */
1866 struct namecache *
1867 cache_allocroot(struct mount *mp, struct vnode *vp)
1868 {
1869 	struct namecache *ncp = cache_alloc(0);
1870 
1871 	ncp->nc_flag |= NCF_MOUNTPT | NCF_ROOT;
1872 	ncp->nc_mount = mp;
1873 	cache_setvp(ncp, vp);
1874 	return(ncp);
1875 }
1876 
1877 /*
1878  * vfs_cache_setroot()
1879  *
1880  *	Create an association between the root of our namecache and
1881  *	the root vnode.  This routine may be called several times during
1882  *	booting.
1883  *
1884  *	If the caller intends to save the returned namecache pointer somewhere
1885  *	it must cache_hold() it.
1886  */
1887 void
1888 vfs_cache_setroot(struct vnode *nvp, struct namecache *ncp)
1889 {
1890 	struct vnode *ovp;
1891 	struct namecache *oncp;
1892 
1893 	ovp = rootvnode;
1894 	oncp = rootncp;
1895 	rootvnode = nvp;
1896 	rootncp = ncp;
1897 
1898 	if (ovp)
1899 		vrele(ovp);
1900 	if (oncp)
1901 		cache_drop(oncp);
1902 }
1903 
1904 /*
1905  * XXX OLD API COMPAT FUNCTION.  This really messes up the new namecache
1906  * topology and is being removed as quickly as possible.  The new VOP_N*()
1907  * API calls are required to make specific adjustments using the supplied
1908  * ncp pointers rather then just bogusly purging random vnodes.
1909  *
1910  * Invalidate all namecache entries to a particular vnode as well as
1911  * any direct children of that vnode in the namecache.  This is a
1912  * 'catch all' purge used by filesystems that do not know any better.
1913  *
1914  * Note that the linkage between the vnode and its namecache entries will
1915  * be removed, but the namecache entries themselves might stay put due to
1916  * active references from elsewhere in the system or due to the existance of
1917  * the children.   The namecache topology is left intact even if we do not
1918  * know what the vnode association is.  Such entries will be marked
1919  * NCF_UNRESOLVED.
1920  */
1921 void
1922 cache_purge(struct vnode *vp)
1923 {
1924 	cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN);
1925 }
1926 
1927 /*
1928  * Flush all entries referencing a particular filesystem.
1929  *
1930  * Since we need to check it anyway, we will flush all the invalid
1931  * entries at the same time.
1932  */
1933 void
1934 cache_purgevfs(struct mount *mp)
1935 {
1936 	struct nchashhead *nchpp;
1937 	struct namecache *ncp, *nnp;
1938 
1939 	/*
1940 	 * Scan hash tables for applicable entries.
1941 	 */
1942 	for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) {
1943 		ncp = LIST_FIRST(nchpp);
1944 		if (ncp)
1945 			cache_hold(ncp);
1946 		while (ncp) {
1947 			nnp = LIST_NEXT(ncp, nc_hash);
1948 			if (nnp)
1949 				cache_hold(nnp);
1950 			if (ncp->nc_mount == mp) {
1951 				cache_lock(ncp);
1952 				cache_zap(ncp);
1953 			} else {
1954 				cache_drop(ncp);
1955 			}
1956 			ncp = nnp;
1957 		}
1958 	}
1959 }
1960 
1961 /*
1962  * Create a new (theoretically) unique fsmid
1963  */
1964 int64_t
1965 cache_getnewfsmid(void)
1966 {
1967 	static int fsmid_roller;
1968 	int64_t fsmid;
1969 
1970 	++fsmid_roller;
1971 	fsmid = ((int64_t)time_second << 32) |
1972 			(fsmid_roller & 0x7FFFFFFF);
1973 	return (fsmid);
1974 }
1975 
1976 
1977 static int disablecwd;
1978 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, "");
1979 
1980 static u_long numcwdcalls; STATNODE(CTLFLAG_RD, numcwdcalls, &numcwdcalls);
1981 static u_long numcwdfail1; STATNODE(CTLFLAG_RD, numcwdfail1, &numcwdfail1);
1982 static u_long numcwdfail2; STATNODE(CTLFLAG_RD, numcwdfail2, &numcwdfail2);
1983 static u_long numcwdfail3; STATNODE(CTLFLAG_RD, numcwdfail3, &numcwdfail3);
1984 static u_long numcwdfail4; STATNODE(CTLFLAG_RD, numcwdfail4, &numcwdfail4);
1985 static u_long numcwdfound; STATNODE(CTLFLAG_RD, numcwdfound, &numcwdfound);
1986 
1987 int
1988 sys___getcwd(struct __getcwd_args *uap)
1989 {
1990 	int buflen;
1991 	int error;
1992 	char *buf;
1993 	char *bp;
1994 
1995 	if (disablecwd)
1996 		return (ENODEV);
1997 
1998 	buflen = uap->buflen;
1999 	if (buflen < 2)
2000 		return (EINVAL);
2001 	if (buflen > MAXPATHLEN)
2002 		buflen = MAXPATHLEN;
2003 
2004 	buf = kmalloc(buflen, M_TEMP, M_WAITOK);
2005 	bp = kern_getcwd(buf, buflen, &error);
2006 	if (error == 0)
2007 		error = copyout(bp, uap->buf, strlen(bp) + 1);
2008 	kfree(buf, M_TEMP);
2009 	return (error);
2010 }
2011 
2012 char *
2013 kern_getcwd(char *buf, size_t buflen, int *error)
2014 {
2015 	struct proc *p = curproc;
2016 	char *bp;
2017 	int i, slash_prefixed;
2018 	struct filedesc *fdp;
2019 	struct namecache *ncp;
2020 
2021 	numcwdcalls++;
2022 	bp = buf;
2023 	bp += buflen - 1;
2024 	*bp = '\0';
2025 	fdp = p->p_fd;
2026 	slash_prefixed = 0;
2027 
2028 	ncp = fdp->fd_ncdir;
2029 	while (ncp && ncp != fdp->fd_nrdir && (ncp->nc_flag & NCF_ROOT) == 0) {
2030 		if (ncp->nc_flag & NCF_MOUNTPT) {
2031 			if (ncp->nc_mount == NULL) {
2032 				*error = EBADF;		/* forced unmount? */
2033 				return(NULL);
2034 			}
2035 			ncp = ncp->nc_parent;
2036 			continue;
2037 		}
2038 		for (i = ncp->nc_nlen - 1; i >= 0; i--) {
2039 			if (bp == buf) {
2040 				numcwdfail4++;
2041 				*error = ENOMEM;
2042 				return(NULL);
2043 			}
2044 			*--bp = ncp->nc_name[i];
2045 		}
2046 		if (bp == buf) {
2047 			numcwdfail4++;
2048 			*error = ENOMEM;
2049 			return(NULL);
2050 		}
2051 		*--bp = '/';
2052 		slash_prefixed = 1;
2053 		ncp = ncp->nc_parent;
2054 	}
2055 	if (ncp == NULL) {
2056 		numcwdfail2++;
2057 		*error = ENOENT;
2058 		return(NULL);
2059 	}
2060 	if (!slash_prefixed) {
2061 		if (bp == buf) {
2062 			numcwdfail4++;
2063 			*error = ENOMEM;
2064 			return(NULL);
2065 		}
2066 		*--bp = '/';
2067 	}
2068 	numcwdfound++;
2069 	*error = 0;
2070 	return (bp);
2071 }
2072 
2073 /*
2074  * Thus begins the fullpath magic.
2075  */
2076 
2077 #undef STATNODE
2078 #define STATNODE(name)							\
2079 	static u_int name;						\
2080 	SYSCTL_UINT(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, "")
2081 
2082 static int disablefullpath;
2083 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW,
2084     &disablefullpath, 0, "");
2085 
2086 STATNODE(numfullpathcalls);
2087 STATNODE(numfullpathfail1);
2088 STATNODE(numfullpathfail2);
2089 STATNODE(numfullpathfail3);
2090 STATNODE(numfullpathfail4);
2091 STATNODE(numfullpathfound);
2092 
2093 int
2094 cache_fullpath(struct proc *p, struct namecache *ncp, char **retbuf, char **freebuf)
2095 {
2096 	char *bp, *buf;
2097 	int i, slash_prefixed;
2098 	struct namecache *fd_nrdir;
2099 
2100 	numfullpathcalls--;
2101 
2102 	buf = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2103 	bp = buf + MAXPATHLEN - 1;
2104 	*bp = '\0';
2105 	if (p != NULL)
2106 		fd_nrdir = p->p_fd->fd_nrdir;
2107 	else
2108 		fd_nrdir = NULL;
2109 	slash_prefixed = 0;
2110 	while (ncp && ncp != fd_nrdir && (ncp->nc_flag & NCF_ROOT) == 0) {
2111 		if (ncp->nc_flag & NCF_MOUNTPT) {
2112 			if (ncp->nc_mount == NULL) {
2113 				kfree(buf, M_TEMP);
2114 				return(EBADF);
2115 			}
2116 			ncp = ncp->nc_parent;
2117 			continue;
2118 		}
2119 		for (i = ncp->nc_nlen - 1; i >= 0; i--) {
2120 			if (bp == buf) {
2121 				numfullpathfail4++;
2122 				kfree(buf, M_TEMP);
2123 				return(ENOMEM);
2124 			}
2125 			*--bp = ncp->nc_name[i];
2126 		}
2127 		if (bp == buf) {
2128 			numfullpathfail4++;
2129 			kfree(buf, M_TEMP);
2130 			return(ENOMEM);
2131 		}
2132 		*--bp = '/';
2133 		slash_prefixed = 1;
2134 		ncp = ncp->nc_parent;
2135 	}
2136 	if (ncp == NULL) {
2137 		numfullpathfail2++;
2138 		kfree(buf, M_TEMP);
2139 		return(ENOENT);
2140 	}
2141 	if (p != NULL && (ncp->nc_flag & NCF_ROOT) && ncp != fd_nrdir) {
2142 		bp = buf + MAXPATHLEN - 1;
2143 		*bp = '\0';
2144 		slash_prefixed = 0;
2145 	}
2146 	if (!slash_prefixed) {
2147 		if (bp == buf) {
2148 			numfullpathfail4++;
2149 			kfree(buf, M_TEMP);
2150 			return(ENOMEM);
2151 		}
2152 		*--bp = '/';
2153 	}
2154 	numfullpathfound++;
2155 	*retbuf = bp;
2156 	*freebuf = buf;
2157 
2158 	return(0);
2159 }
2160 
2161 int
2162 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf)
2163 {
2164 	struct namecache *ncp;
2165 
2166 	numfullpathcalls++;
2167 	if (disablefullpath)
2168 		return (ENODEV);
2169 
2170 	if (p == NULL)
2171 		return (EINVAL);
2172 
2173 	/* vn is NULL, client wants us to use p->p_textvp */
2174 	if (vn == NULL) {
2175 		if ((vn = p->p_textvp) == NULL)
2176 			return (EINVAL);
2177 	}
2178 	TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) {
2179 		if (ncp->nc_nlen)
2180 			break;
2181 	}
2182 	if (ncp == NULL)
2183 		return (EINVAL);
2184 
2185 	numfullpathcalls--;
2186 	return(cache_fullpath(p, ncp, retbuf, freebuf));
2187 }
2188