xref: /dflybsd-src/sys/kern/vfs_cache.c (revision 21d8bc42f570abe46007647277c2c8d68cf8b55a)
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 1.86 2008/01/18 19:13:15 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 #include <sys/sysref2.h>
92 
93 #define MAX_RECURSION_DEPTH	64
94 
95 /*
96  * Random lookups in the cache are accomplished with a hash table using
97  * a hash key of (nc_src_vp, name).
98  *
99  * Negative entries may exist and correspond to structures where nc_vp
100  * is NULL.  In a negative entry, NCF_WHITEOUT will be set if the entry
101  * corresponds to a whited-out directory entry (verses simply not finding the
102  * entry at all).
103  *
104  * Upon reaching the last segment of a path, if the reference is for DELETE,
105  * or NOCACHE is set (rewrite), and the name is located in the cache, it
106  * will be dropped.
107  */
108 
109 /*
110  * Structures associated with name cacheing.
111  */
112 #define NCHHASH(hash)	(&nchashtbl[(hash) & nchash])
113 #define MINNEG		1024
114 
115 MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
116 
117 static LIST_HEAD(nchashhead, namecache) *nchashtbl;	/* Hash Table */
118 static struct namecache_list	ncneglist;		/* instead of vnode */
119 
120 /*
121  * ncvp_debug - debug cache_fromvp().  This is used by the NFS server
122  * to create the namecache infrastructure leading to a dangling vnode.
123  *
124  * 0	Only errors are reported
125  * 1	Successes are reported
126  * 2	Successes + the whole directory scan is reported
127  * 3	Force the directory scan code run as if the parent vnode did not
128  *	have a namecache record, even if it does have one.
129  */
130 static int	ncvp_debug;
131 SYSCTL_INT(_debug, OID_AUTO, ncvp_debug, CTLFLAG_RW, &ncvp_debug, 0, "");
132 
133 static u_long	nchash;			/* size of hash table */
134 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0, "");
135 
136 static u_long	ncnegfactor = 16;	/* ratio of negative entries */
137 SYSCTL_ULONG(_debug, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0, "");
138 
139 static int	nclockwarn;		/* warn on locked entries in ticks */
140 SYSCTL_INT(_debug, OID_AUTO, nclockwarn, CTLFLAG_RW, &nclockwarn, 0, "");
141 
142 static u_long	numneg;		/* number of cache entries allocated */
143 SYSCTL_ULONG(_debug, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0, "");
144 
145 static u_long	numcache;		/* number of cache entries allocated */
146 SYSCTL_ULONG(_debug, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0, "");
147 
148 static u_long	numunres;		/* number of unresolved entries */
149 SYSCTL_ULONG(_debug, OID_AUTO, numunres, CTLFLAG_RD, &numunres, 0, "");
150 
151 SYSCTL_INT(_debug, OID_AUTO, vnsize, CTLFLAG_RD, 0, sizeof(struct vnode), "");
152 SYSCTL_INT(_debug, OID_AUTO, ncsize, CTLFLAG_RD, 0, sizeof(struct namecache), "");
153 
154 static int cache_resolve_mp(struct mount *mp);
155 static void _cache_rehash(struct namecache *ncp);
156 static void _cache_lock(struct namecache *ncp);
157 static void _cache_setunresolved(struct namecache *ncp);
158 
159 /*
160  * The new name cache statistics
161  */
162 SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics");
163 #define STATNODE(mode, name, var) \
164 	SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, mode, var, 0, "");
165 STATNODE(CTLFLAG_RD, numneg, &numneg);
166 STATNODE(CTLFLAG_RD, numcache, &numcache);
167 static u_long numcalls; STATNODE(CTLFLAG_RD, numcalls, &numcalls);
168 static u_long dothits; STATNODE(CTLFLAG_RD, dothits, &dothits);
169 static u_long dotdothits; STATNODE(CTLFLAG_RD, dotdothits, &dotdothits);
170 static u_long numchecks; STATNODE(CTLFLAG_RD, numchecks, &numchecks);
171 static u_long nummiss; STATNODE(CTLFLAG_RD, nummiss, &nummiss);
172 static u_long nummisszap; STATNODE(CTLFLAG_RD, nummisszap, &nummisszap);
173 static u_long numposzaps; STATNODE(CTLFLAG_RD, numposzaps, &numposzaps);
174 static u_long numposhits; STATNODE(CTLFLAG_RD, numposhits, &numposhits);
175 static u_long numnegzaps; STATNODE(CTLFLAG_RD, numnegzaps, &numnegzaps);
176 static u_long numneghits; STATNODE(CTLFLAG_RD, numneghits, &numneghits);
177 
178 struct nchstats nchstats[SMP_MAXCPU];
179 /*
180  * Export VFS cache effectiveness statistics to user-land.
181  *
182  * The statistics are left for aggregation to user-land so
183  * neat things can be achieved, like observing per-CPU cache
184  * distribution.
185  */
186 static int
187 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
188 {
189 	struct globaldata *gd;
190 	int i, error;
191 
192 	error = 0;
193 	for (i = 0; i < ncpus; ++i) {
194 		gd = globaldata_find(i);
195 		if ((error = SYSCTL_OUT(req, (void *)&(*gd->gd_nchstats),
196 			sizeof(struct nchstats))))
197 			break;
198 	}
199 
200 	return (error);
201 }
202 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE|CTLFLAG_RD,
203   0, 0, sysctl_nchstats, "S,nchstats", "VFS cache effectiveness statistics");
204 
205 static void cache_zap(struct namecache *ncp);
206 
207 /*
208  * cache_hold() and cache_drop() prevent the premature deletion of a
209  * namecache entry but do not prevent operations (such as zapping) on
210  * that namecache entry.
211  *
212  * This routine may only be called from outside this source module if
213  * nc_refs is already at least 1.
214  *
215  * This is a rare case where callers are allowed to hold a spinlock,
216  * so we can't ourselves.
217  */
218 static __inline
219 struct namecache *
220 _cache_hold(struct namecache *ncp)
221 {
222 	atomic_add_int(&ncp->nc_refs, 1);
223 	return(ncp);
224 }
225 
226 /*
227  * When dropping an entry, if only one ref remains and the entry has not
228  * been resolved, zap it.  Since the one reference is being dropped the
229  * entry had better not be locked.
230  */
231 static __inline
232 void
233 _cache_drop(struct namecache *ncp)
234 {
235 	KKASSERT(ncp->nc_refs > 0);
236 	if (ncp->nc_refs == 1 &&
237 	    (ncp->nc_flag & NCF_UNRESOLVED) &&
238 	    TAILQ_EMPTY(&ncp->nc_list)
239 	) {
240 		KKASSERT(ncp->nc_exlocks == 0);
241 		_cache_lock(ncp);
242 		cache_zap(ncp);
243 	} else {
244 		atomic_subtract_int(&ncp->nc_refs, 1);
245 	}
246 }
247 
248 /*
249  * Link a new namecache entry to its parent.  Be careful to avoid races
250  * if vhold() blocks in the future.
251  */
252 static void
253 cache_link_parent(struct namecache *ncp, struct namecache *par)
254 {
255 	KKASSERT(ncp->nc_parent == NULL);
256 	ncp->nc_parent = par;
257 	if (TAILQ_EMPTY(&par->nc_list)) {
258 		TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
259 		/*
260 		 * Any vp associated with an ncp which has children must
261 		 * be held to prevent it from being recycled.
262 		 */
263 		if (par->nc_vp)
264 			vhold(par->nc_vp);
265 	} else {
266 		TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
267 	}
268 }
269 
270 /*
271  * Remove the parent association from a namecache structure.  If this is
272  * the last child of the parent the cache_drop(par) will attempt to
273  * recursively zap the parent.
274  */
275 static void
276 cache_unlink_parent(struct namecache *ncp)
277 {
278 	struct namecache *par;
279 
280 	if ((par = ncp->nc_parent) != NULL) {
281 		ncp->nc_parent = NULL;
282 		par = _cache_hold(par);
283 		TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
284 		if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
285 			vdrop(par->nc_vp);
286 		_cache_drop(par);
287 	}
288 }
289 
290 /*
291  * Allocate a new namecache structure.  Most of the code does not require
292  * zero-termination of the string but it makes vop_compat_ncreate() easier.
293  */
294 static struct namecache *
295 cache_alloc(int nlen)
296 {
297 	struct namecache *ncp;
298 
299 	ncp = kmalloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO);
300 	if (nlen)
301 		ncp->nc_name = kmalloc(nlen + 1, M_VFSCACHE, M_WAITOK);
302 	ncp->nc_nlen = nlen;
303 	ncp->nc_flag = NCF_UNRESOLVED;
304 	ncp->nc_error = ENOTCONN;	/* needs to be resolved */
305 	ncp->nc_refs = 1;
306 
307 	/*
308 	 * Construct a fake FSMID based on the time of day and a 32 bit
309 	 * roller for uniqueness.  This is used to generate a useful
310 	 * FSMID for filesystems which do not support it.
311 	 */
312 	ncp->nc_fsmid = cache_getnewfsmid();
313 	TAILQ_INIT(&ncp->nc_list);
314 	_cache_lock(ncp);
315 	return(ncp);
316 }
317 
318 static void
319 _cache_free(struct namecache *ncp)
320 {
321 	KKASSERT(ncp->nc_refs == 1 && ncp->nc_exlocks == 1);
322 	if (ncp->nc_name)
323 		kfree(ncp->nc_name, M_VFSCACHE);
324 	kfree(ncp, M_VFSCACHE);
325 }
326 
327 void
328 cache_zero(struct nchandle *nch)
329 {
330 	nch->ncp = NULL;
331 	nch->mount = NULL;
332 }
333 
334 /*
335  * Ref and deref a namecache structure.
336  *
337  * Warning: caller may hold an unrelated read spinlock, which means we can't
338  * use read spinlocks here.
339  */
340 struct nchandle *
341 cache_hold(struct nchandle *nch)
342 {
343 	_cache_hold(nch->ncp);
344 	++nch->mount->mnt_refs;
345 	return(nch);
346 }
347 
348 void
349 cache_copy(struct nchandle *nch, struct nchandle *target)
350 {
351 	*target = *nch;
352 	_cache_hold(target->ncp);
353 	++nch->mount->mnt_refs;
354 }
355 
356 void
357 cache_changemount(struct nchandle *nch, struct mount *mp)
358 {
359 	--nch->mount->mnt_refs;
360 	nch->mount = mp;
361 	++nch->mount->mnt_refs;
362 }
363 
364 void
365 cache_drop(struct nchandle *nch)
366 {
367 	--nch->mount->mnt_refs;
368 	_cache_drop(nch->ncp);
369 	nch->ncp = NULL;
370 	nch->mount = NULL;
371 }
372 
373 /*
374  * Namespace locking.  The caller must already hold a reference to the
375  * namecache structure in order to lock/unlock it.  This function prevents
376  * the namespace from being created or destroyed by accessors other then
377  * the lock holder.
378  *
379  * Note that holding a locked namecache structure prevents other threads
380  * from making namespace changes (e.g. deleting or creating), prevents
381  * vnode association state changes by other threads, and prevents the
382  * namecache entry from being resolved or unresolved by other threads.
383  *
384  * The lock owner has full authority to associate/disassociate vnodes
385  * and resolve/unresolve the locked ncp.
386  *
387  * WARNING!  Holding a locked ncp will prevent a vnode from being destroyed
388  * or recycled, but it does NOT help you if the vnode had already initiated
389  * a recyclement.  If this is important, use cache_get() rather then
390  * cache_lock() (and deal with the differences in the way the refs counter
391  * is handled).  Or, alternatively, make an unconditional call to
392  * cache_validate() or cache_resolve() after cache_lock() returns.
393  */
394 static
395 void
396 _cache_lock(struct namecache *ncp)
397 {
398 	thread_t td;
399 	int didwarn;
400 
401 	KKASSERT(ncp->nc_refs != 0);
402 	didwarn = 0;
403 	td = curthread;
404 
405 	for (;;) {
406 		if (ncp->nc_exlocks == 0) {
407 			ncp->nc_exlocks = 1;
408 			ncp->nc_locktd = td;
409 			/*
410 			 * The vp associated with a locked ncp must be held
411 			 * to prevent it from being recycled (which would
412 			 * cause the ncp to become unresolved).
413 			 *
414 			 * WARNING!  If VRECLAIMED is set the vnode could
415 			 * already be in the middle of a recycle.  Callers
416 			 * should not assume that nc_vp is usable when
417 			 * not NULL.  cache_vref() or cache_vget() must be
418 			 * called.
419 			 *
420 			 * XXX loop on race for later MPSAFE work.
421 			 */
422 			if (ncp->nc_vp)
423 				vhold(ncp->nc_vp);
424 			break;
425 		}
426 		if (ncp->nc_locktd == td) {
427 			++ncp->nc_exlocks;
428 			break;
429 		}
430 		ncp->nc_flag |= NCF_LOCKREQ;
431 		if (tsleep(ncp, 0, "clock", nclockwarn) == EWOULDBLOCK) {
432 			if (didwarn)
433 				continue;
434 			didwarn = 1;
435 			kprintf("[diagnostic] cache_lock: blocked on %p", ncp);
436 			kprintf(" \"%*.*s\"\n",
437 				ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
438 		}
439 	}
440 
441 	if (didwarn == 1) {
442 		kprintf("[diagnostic] cache_lock: unblocked %*.*s\n",
443 			ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
444 	}
445 }
446 
447 void
448 cache_lock(struct nchandle *nch)
449 {
450 	_cache_lock(nch->ncp);
451 }
452 
453 static
454 int
455 _cache_lock_nonblock(struct namecache *ncp)
456 {
457 	thread_t td;
458 
459 	KKASSERT(ncp->nc_refs != 0);
460 	td = curthread;
461 	if (ncp->nc_exlocks == 0) {
462 		ncp->nc_exlocks = 1;
463 		ncp->nc_locktd = td;
464 		/*
465 		 * The vp associated with a locked ncp must be held
466 		 * to prevent it from being recycled (which would
467 		 * cause the ncp to become unresolved).
468 		 *
469 		 * WARNING!  If VRECLAIMED is set the vnode could
470 		 * already be in the middle of a recycle.  Callers
471 		 * should not assume that nc_vp is usable when
472 		 * not NULL.  cache_vref() or cache_vget() must be
473 		 * called.
474 		 *
475 		 * XXX loop on race for later MPSAFE work.
476 		 */
477 		if (ncp->nc_vp)
478 			vhold(ncp->nc_vp);
479 		return(0);
480 	} else {
481 		return(EWOULDBLOCK);
482 	}
483 }
484 
485 int
486 cache_lock_nonblock(struct nchandle *nch)
487 {
488 	return(_cache_lock_nonblock(nch->ncp));
489 }
490 
491 static
492 void
493 _cache_unlock(struct namecache *ncp)
494 {
495 	thread_t td = curthread;
496 
497 	KKASSERT(ncp->nc_refs > 0);
498 	KKASSERT(ncp->nc_exlocks > 0);
499 	KKASSERT(ncp->nc_locktd == td);
500 	if (--ncp->nc_exlocks == 0) {
501 		if (ncp->nc_vp)
502 			vdrop(ncp->nc_vp);
503 		ncp->nc_locktd = NULL;
504 		if (ncp->nc_flag & NCF_LOCKREQ) {
505 			ncp->nc_flag &= ~NCF_LOCKREQ;
506 			wakeup(ncp);
507 		}
508 	}
509 }
510 
511 void
512 cache_unlock(struct nchandle *nch)
513 {
514 	_cache_unlock(nch->ncp);
515 }
516 
517 /*
518  * ref-and-lock, unlock-and-deref functions.
519  *
520  * This function is primarily used by nlookup.  Even though cache_lock
521  * holds the vnode, it is possible that the vnode may have already
522  * initiated a recyclement.  We want cache_get() to return a definitively
523  * usable vnode or a definitively unresolved ncp.
524  */
525 static
526 struct namecache *
527 _cache_get(struct namecache *ncp)
528 {
529 	_cache_hold(ncp);
530 	_cache_lock(ncp);
531 	if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
532 		_cache_setunresolved(ncp);
533 	return(ncp);
534 }
535 
536 /*
537  * note: the same nchandle can be passed for both arguments.
538  */
539 void
540 cache_get(struct nchandle *nch, struct nchandle *target)
541 {
542 	target->mount = nch->mount;
543 	target->ncp = _cache_get(nch->ncp);
544 	++target->mount->mnt_refs;
545 }
546 
547 static int
548 _cache_get_nonblock(struct namecache *ncp)
549 {
550 	/* XXX MP */
551 	if (ncp->nc_exlocks == 0 || ncp->nc_locktd == curthread) {
552 		_cache_hold(ncp);
553 		_cache_lock(ncp);
554 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
555 			_cache_setunresolved(ncp);
556 		return(0);
557 	}
558 	return(EWOULDBLOCK);
559 }
560 
561 int
562 cache_get_nonblock(struct nchandle *nch)
563 {
564 	return(_cache_get_nonblock(nch->ncp));
565 }
566 
567 static __inline
568 void
569 _cache_put(struct namecache *ncp)
570 {
571 	_cache_unlock(ncp);
572 	_cache_drop(ncp);
573 }
574 
575 void
576 cache_put(struct nchandle *nch)
577 {
578 	--nch->mount->mnt_refs;
579 	_cache_put(nch->ncp);
580 	nch->ncp = NULL;
581 	nch->mount = NULL;
582 }
583 
584 /*
585  * Resolve an unresolved ncp by associating a vnode with it.  If the
586  * vnode is NULL, a negative cache entry is created.
587  *
588  * The ncp should be locked on entry and will remain locked on return.
589  */
590 static
591 void
592 _cache_setvp(struct namecache *ncp, struct vnode *vp)
593 {
594 	KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
595 	ncp->nc_vp = vp;
596 	if (vp != NULL) {
597 		/*
598 		 * Any vp associated with an ncp which has children must
599 		 * be held.  Any vp associated with a locked ncp must be held.
600 		 */
601 		if (!TAILQ_EMPTY(&ncp->nc_list))
602 			vhold(vp);
603 		TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode);
604 		if (ncp->nc_exlocks)
605 			vhold(vp);
606 
607 		/*
608 		 * Set auxiliary flags
609 		 */
610 		switch(vp->v_type) {
611 		case VDIR:
612 			ncp->nc_flag |= NCF_ISDIR;
613 			break;
614 		case VLNK:
615 			ncp->nc_flag |= NCF_ISSYMLINK;
616 			/* XXX cache the contents of the symlink */
617 			break;
618 		default:
619 			break;
620 		}
621 		++numcache;
622 		ncp->nc_error = 0;
623 	} else {
624 		TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
625 		++numneg;
626 		ncp->nc_error = ENOENT;
627 	}
628 	ncp->nc_flag &= ~NCF_UNRESOLVED;
629 }
630 
631 void
632 cache_setvp(struct nchandle *nch, struct vnode *vp)
633 {
634 	_cache_setvp(nch->ncp, vp);
635 }
636 
637 void
638 cache_settimeout(struct nchandle *nch, int nticks)
639 {
640 	struct namecache *ncp = nch->ncp;
641 
642 	if ((ncp->nc_timeout = ticks + nticks) == 0)
643 		ncp->nc_timeout = 1;
644 }
645 
646 /*
647  * Disassociate the vnode or negative-cache association and mark a
648  * namecache entry as unresolved again.  Note that the ncp is still
649  * left in the hash table and still linked to its parent.
650  *
651  * The ncp should be locked and refd on entry and will remain locked and refd
652  * on return.
653  *
654  * This routine is normally never called on a directory containing children.
655  * However, NFS often does just that in its rename() code as a cop-out to
656  * avoid complex namespace operations.  This disconnects a directory vnode
657  * from its namecache and can cause the OLDAPI and NEWAPI to get out of
658  * sync.
659  *
660  * NOTE: NCF_FSMID must be cleared so a refurbishment of the ncp, such as
661  * in a create, properly propogates flag up the chain.
662  */
663 static
664 void
665 _cache_setunresolved(struct namecache *ncp)
666 {
667 	struct vnode *vp;
668 
669 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
670 		ncp->nc_flag |= NCF_UNRESOLVED;
671 		ncp->nc_timeout = 0;
672 		ncp->nc_error = ENOTCONN;
673 		++numunres;
674 		if ((vp = ncp->nc_vp) != NULL) {
675 			--numcache;
676 			ncp->nc_vp = NULL;
677 			TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode);
678 
679 			/*
680 			 * Any vp associated with an ncp with children is
681 			 * held by that ncp.  Any vp associated with a locked
682 			 * ncp is held by that ncp.  These conditions must be
683 			 * undone when the vp is cleared out from the ncp.
684 			 */
685 			if (ncp->nc_flag & NCF_FSMID)
686 				vupdatefsmid(vp);
687 			if (!TAILQ_EMPTY(&ncp->nc_list))
688 				vdrop(vp);
689 			if (ncp->nc_exlocks)
690 				vdrop(vp);
691 		} else {
692 			TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
693 			--numneg;
694 		}
695 		ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK|
696 				  NCF_FSMID);
697 	}
698 }
699 
700 void
701 cache_setunresolved(struct nchandle *nch)
702 {
703 	_cache_setunresolved(nch->ncp);
704 }
705 
706 /*
707  * Determine if we can clear NCF_ISMOUNTPT by scanning the mountlist
708  * looking for matches.  This flag tells the lookup code when it must
709  * check for a mount linkage and also prevents the directories in question
710  * from being deleted or renamed.
711  */
712 static
713 int
714 cache_clrmountpt_callback(struct mount *mp, void *data)
715 {
716 	struct nchandle *nch = data;
717 
718 	if (mp->mnt_ncmounton.ncp == nch->ncp)
719 		return(1);
720 	if (mp->mnt_ncmountpt.ncp == nch->ncp)
721 		return(1);
722 	return(0);
723 }
724 
725 void
726 cache_clrmountpt(struct nchandle *nch)
727 {
728 	int count;
729 
730 	count = mountlist_scan(cache_clrmountpt_callback, nch,
731 			       MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
732 	if (count == 0)
733 		nch->ncp->nc_flag &= ~NCF_ISMOUNTPT;
734 }
735 
736 /*
737  * Invalidate portions of the namecache topology given a starting entry.
738  * The passed ncp is set to an unresolved state and:
739  *
740  * The passed ncp must be locked.
741  *
742  * CINV_DESTROY		- Set a flag in the passed ncp entry indicating
743  *			  that the physical underlying nodes have been
744  *			  destroyed... as in deleted.  For example, when
745  *			  a directory is removed.  This will cause record
746  *			  lookups on the name to no longer be able to find
747  *			  the record and tells the resolver to return failure
748  *			  rather then trying to resolve through the parent.
749  *
750  *			  The topology itself, including ncp->nc_name,
751  *			  remains intact.
752  *
753  *			  This only applies to the passed ncp, if CINV_CHILDREN
754  *			  is specified the children are not flagged.
755  *
756  * CINV_CHILDREN	- Set all children (recursively) to an unresolved
757  *			  state as well.
758  *
759  *			  Note that this will also have the side effect of
760  *			  cleaning out any unreferenced nodes in the topology
761  *			  from the leaves up as the recursion backs out.
762  *
763  * Note that the topology for any referenced nodes remains intact.
764  *
765  * It is possible for cache_inval() to race a cache_resolve(), meaning that
766  * the namecache entry may not actually be invalidated on return if it was
767  * revalidated while recursing down into its children.  This code guarentees
768  * that the node(s) will go through an invalidation cycle, but does not
769  * guarentee that they will remain in an invalidated state.
770  *
771  * Returns non-zero if a revalidation was detected during the invalidation
772  * recursion, zero otherwise.  Note that since only the original ncp is
773  * locked the revalidation ultimately can only indicate that the original ncp
774  * *MIGHT* no have been reresolved.
775  *
776  * DEEP RECURSION HANDLING - If a recursive invalidation recurses deeply we
777  * have to avoid blowing out the kernel stack.  We do this by saving the
778  * deep namecache node and aborting the recursion, then re-recursing at that
779  * node using a depth-first algorithm in order to allow multiple deep
780  * recursions to chain through each other, then we restart the invalidation
781  * from scratch.
782  */
783 
784 struct cinvtrack {
785 	struct namecache *resume_ncp;
786 	int depth;
787 };
788 
789 static int _cache_inval_internal(struct namecache *, int, struct cinvtrack *);
790 
791 static
792 int
793 _cache_inval(struct namecache *ncp, int flags)
794 {
795 	struct cinvtrack track;
796 	struct namecache *ncp2;
797 	int r;
798 
799 	track.depth = 0;
800 	track.resume_ncp = NULL;
801 
802 	for (;;) {
803 		r = _cache_inval_internal(ncp, flags, &track);
804 		if (track.resume_ncp == NULL)
805 			break;
806 		kprintf("Warning: deep namecache recursion at %s\n",
807 			ncp->nc_name);
808 		_cache_unlock(ncp);
809 		while ((ncp2 = track.resume_ncp) != NULL) {
810 			track.resume_ncp = NULL;
811 			_cache_lock(ncp2);
812 			_cache_inval_internal(ncp2, flags & ~CINV_DESTROY,
813 					     &track);
814 			_cache_put(ncp2);
815 		}
816 		_cache_lock(ncp);
817 	}
818 	return(r);
819 }
820 
821 int
822 cache_inval(struct nchandle *nch, int flags)
823 {
824 	return(_cache_inval(nch->ncp, flags));
825 }
826 
827 static int
828 _cache_inval_internal(struct namecache *ncp, int flags, struct cinvtrack *track)
829 {
830 	struct namecache *kid;
831 	struct namecache *nextkid;
832 	int rcnt = 0;
833 
834 	KKASSERT(ncp->nc_exlocks);
835 
836 	_cache_setunresolved(ncp);
837 	if (flags & CINV_DESTROY)
838 		ncp->nc_flag |= NCF_DESTROYED;
839 
840 	if ((flags & CINV_CHILDREN) &&
841 	    (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL
842 	) {
843 		if (++track->depth > MAX_RECURSION_DEPTH) {
844 			track->resume_ncp = ncp;
845 			_cache_hold(ncp);
846 			++rcnt;
847 		}
848 		_cache_hold(kid);
849 		_cache_unlock(ncp);
850 		while (kid) {
851 			if (track->resume_ncp) {
852 				_cache_drop(kid);
853 				break;
854 			}
855 			if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL)
856 				_cache_hold(nextkid);
857 			if ((kid->nc_flag & NCF_UNRESOLVED) == 0 ||
858 			    TAILQ_FIRST(&kid->nc_list)
859 			) {
860 				_cache_lock(kid);
861 				rcnt += _cache_inval_internal(kid, flags & ~CINV_DESTROY, track);
862 				_cache_unlock(kid);
863 			}
864 			_cache_drop(kid);
865 			kid = nextkid;
866 		}
867 		--track->depth;
868 		_cache_lock(ncp);
869 	}
870 
871 	/*
872 	 * Someone could have gotten in there while ncp was unlocked,
873 	 * retry if so.
874 	 */
875 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
876 		++rcnt;
877 	return (rcnt);
878 }
879 
880 /*
881  * Invalidate a vnode's namecache associations.  To avoid races against
882  * the resolver we do not invalidate a node which we previously invalidated
883  * but which was then re-resolved while we were in the invalidation loop.
884  *
885  * Returns non-zero if any namecache entries remain after the invalidation
886  * loop completed.
887  *
888  * NOTE: unlike the namecache topology which guarentees that ncp's will not
889  * be ripped out of the topology while held, the vnode's v_namecache list
890  * has no such restriction.  NCP's can be ripped out of the list at virtually
891  * any time if not locked, even if held.
892  */
893 int
894 cache_inval_vp(struct vnode *vp, int flags)
895 {
896 	struct namecache *ncp;
897 	struct namecache *next;
898 
899 restart:
900 	ncp = TAILQ_FIRST(&vp->v_namecache);
901 	if (ncp)
902 		_cache_hold(ncp);
903 	while (ncp) {
904 		/* loop entered with ncp held */
905 		if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
906 			_cache_hold(next);
907 		_cache_lock(ncp);
908 		if (ncp->nc_vp != vp) {
909 			kprintf("Warning: cache_inval_vp: race-A detected on "
910 				"%s\n", ncp->nc_name);
911 			_cache_put(ncp);
912 			if (next)
913 				_cache_drop(next);
914 			goto restart;
915 		}
916 		_cache_inval(ncp, flags);
917 		_cache_put(ncp);		/* also releases reference */
918 		ncp = next;
919 		if (ncp && ncp->nc_vp != vp) {
920 			kprintf("Warning: cache_inval_vp: race-B detected on "
921 				"%s\n", ncp->nc_name);
922 			_cache_drop(ncp);
923 			goto restart;
924 		}
925 	}
926 	return(TAILQ_FIRST(&vp->v_namecache) != NULL);
927 }
928 
929 /*
930  * This routine is used instead of the normal cache_inval_vp() when we
931  * are trying to recycle otherwise good vnodes.
932  *
933  * Return 0 on success, non-zero if not all namecache records could be
934  * disassociated from the vnode (for various reasons).
935  */
936 int
937 cache_inval_vp_nonblock(struct vnode *vp)
938 {
939 	struct namecache *ncp;
940 	struct namecache *next;
941 
942 	ncp = TAILQ_FIRST(&vp->v_namecache);
943 	if (ncp)
944 		_cache_hold(ncp);
945 	while (ncp) {
946 		/* loop entered with ncp held */
947 		if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
948 			_cache_hold(next);
949 		if (_cache_lock_nonblock(ncp)) {
950 			_cache_drop(ncp);
951 			if (next)
952 				_cache_drop(next);
953 			break;
954 		}
955 		if (ncp->nc_vp != vp) {
956 			kprintf("Warning: cache_inval_vp: race-A detected on "
957 				"%s\n", ncp->nc_name);
958 			_cache_put(ncp);
959 			if (next)
960 				_cache_drop(next);
961 			break;
962 		}
963 		_cache_inval(ncp, 0);
964 		_cache_put(ncp);		/* also releases reference */
965 		ncp = next;
966 		if (ncp && ncp->nc_vp != vp) {
967 			kprintf("Warning: cache_inval_vp: race-B detected on "
968 				"%s\n", ncp->nc_name);
969 			_cache_drop(ncp);
970 			break;
971 		}
972 	}
973 	return(TAILQ_FIRST(&vp->v_namecache) != NULL);
974 }
975 
976 /*
977  * The source ncp has been renamed to the target ncp.  Both fncp and tncp
978  * must be locked.  The target ncp is destroyed (as a normal rename-over
979  * would destroy the target file or directory).
980  *
981  * Because there may be references to the source ncp we cannot copy its
982  * contents to the target.  Instead the source ncp is relinked as the target
983  * and the target ncp is removed from the namecache topology.
984  */
985 void
986 cache_rename(struct nchandle *fnch, struct nchandle *tnch)
987 {
988 	struct namecache *fncp = fnch->ncp;
989 	struct namecache *tncp = tnch->ncp;
990 	char *oname;
991 
992 	_cache_setunresolved(tncp);
993 	cache_unlink_parent(fncp);
994 	cache_link_parent(fncp, tncp->nc_parent);
995 	cache_unlink_parent(tncp);
996 	oname = fncp->nc_name;
997 	fncp->nc_name = tncp->nc_name;
998 	fncp->nc_nlen = tncp->nc_nlen;
999 	tncp->nc_name = NULL;
1000 	tncp->nc_nlen = 0;
1001 	if (fncp->nc_flag & NCF_HASHED)
1002 		_cache_rehash(fncp);
1003 	if (tncp->nc_flag & NCF_HASHED)
1004 		_cache_rehash(tncp);
1005 	if (oname)
1006 		kfree(oname, M_VFSCACHE);
1007 }
1008 
1009 /*
1010  * vget the vnode associated with the namecache entry.  Resolve the namecache
1011  * entry if necessary and deal with namecache/vp races.  The passed ncp must
1012  * be referenced and may be locked.  The ncp's ref/locking state is not
1013  * effected by this call.
1014  *
1015  * lk_type may be LK_SHARED, LK_EXCLUSIVE.  A ref'd, possibly locked
1016  * (depending on the passed lk_type) will be returned in *vpp with an error
1017  * of 0, or NULL will be returned in *vpp with a non-0 error code.  The
1018  * most typical error is ENOENT, meaning that the ncp represents a negative
1019  * cache hit and there is no vnode to retrieve, but other errors can occur
1020  * too.
1021  *
1022  * The main race we have to deal with are namecache zaps.  The ncp itself
1023  * will not disappear since it is referenced, and it turns out that the
1024  * validity of the vp pointer can be checked simply by rechecking the
1025  * contents of ncp->nc_vp.
1026  */
1027 int
1028 cache_vget(struct nchandle *nch, struct ucred *cred,
1029 	   int lk_type, struct vnode **vpp)
1030 {
1031 	struct namecache *ncp;
1032 	struct vnode *vp;
1033 	int error;
1034 
1035 	ncp = nch->ncp;
1036 again:
1037 	vp = NULL;
1038 	if (ncp->nc_flag & NCF_UNRESOLVED) {
1039 		_cache_lock(ncp);
1040 		error = cache_resolve(nch, cred);
1041 		_cache_unlock(ncp);
1042 	} else {
1043 		error = 0;
1044 	}
1045 	if (error == 0 && (vp = ncp->nc_vp) != NULL) {
1046 		/*
1047 		 * Accessing the vnode from the namecache is a bit
1048 		 * dangerous.  Because there are no refs on the vnode, it
1049 		 * could be in the middle of a reclaim.
1050 		 */
1051 		if (vp->v_flag & VRECLAIMED) {
1052 			kprintf("Warning: vnode reclaim race detected in cache_vget on %p (%s)\n", vp, ncp->nc_name);
1053 			_cache_lock(ncp);
1054 			_cache_setunresolved(ncp);
1055 			_cache_unlock(ncp);
1056 			goto again;
1057 		}
1058 		error = vget(vp, lk_type);
1059 		if (error) {
1060 			if (vp != ncp->nc_vp)
1061 				goto again;
1062 			vp = NULL;
1063 		} else if (vp != ncp->nc_vp) {
1064 			vput(vp);
1065 			goto again;
1066 		} else if (vp->v_flag & VRECLAIMED) {
1067 			panic("vget succeeded on a VRECLAIMED node! vp %p", vp);
1068 		}
1069 	}
1070 	if (error == 0 && vp == NULL)
1071 		error = ENOENT;
1072 	*vpp = vp;
1073 	return(error);
1074 }
1075 
1076 int
1077 cache_vref(struct nchandle *nch, struct ucred *cred, struct vnode **vpp)
1078 {
1079 	struct namecache *ncp;
1080 	struct vnode *vp;
1081 	int error;
1082 
1083 	ncp = nch->ncp;
1084 
1085 again:
1086 	vp = NULL;
1087 	if (ncp->nc_flag & NCF_UNRESOLVED) {
1088 		_cache_lock(ncp);
1089 		error = cache_resolve(nch, cred);
1090 		_cache_unlock(ncp);
1091 	} else {
1092 		error = 0;
1093 	}
1094 	if (error == 0 && (vp = ncp->nc_vp) != NULL) {
1095 		/*
1096 		 * Since we did not obtain any locks, a cache zap
1097 		 * race can occur here if the vnode is in the middle
1098 		 * of being reclaimed and has not yet been able to
1099 		 * clean out its cache node.  If that case occurs,
1100 		 * we must lock and unresolve the cache, then loop
1101 		 * to retry.
1102 		 */
1103 		if ((error = vget(vp, LK_SHARED)) != 0) {
1104 			if (error == ENOENT) {
1105 				kprintf("Warning: vnode reclaim race detected on cache_vref %p (%s)\n", vp, ncp->nc_name);
1106 				_cache_lock(ncp);
1107 				_cache_setunresolved(ncp);
1108 				_cache_unlock(ncp);
1109 				goto again;
1110 			}
1111 			/* fatal error */
1112 		} else {
1113 			/* caller does not want a lock */
1114 			vn_unlock(vp);
1115 		}
1116 	}
1117 	if (error == 0 && vp == NULL)
1118 		error = ENOENT;
1119 	*vpp = vp;
1120 	return(error);
1121 }
1122 
1123 /*
1124  * Recursively set the FSMID update flag for namecache nodes leading
1125  * to root.  This will cause the next getattr or reclaim to increment the
1126  * fsmid and mark the inode for lazy updating.
1127  *
1128  * Stop recursing when we hit a node whos NCF_FSMID flag is already set.
1129  * This makes FSMIDs work in an Einsteinian fashion - where the observation
1130  * effects the result.  In this case a program monitoring a higher level
1131  * node will have detected some prior change and started its scan (clearing
1132  * NCF_FSMID in higher level nodes), but since it has not yet observed the
1133  * node where we find NCF_FSMID still set, we can safely make the related
1134  * modification without interfering with the theorized program.
1135  *
1136  * This also means that FSMIDs cannot represent time-domain quantities
1137  * in a hierarchical sense.  But the main reason for doing it this way
1138  * is to reduce the amount of recursion that occurs in the critical path
1139  * when e.g. a program is writing to a file that sits deep in a directory
1140  * hierarchy.
1141  */
1142 void
1143 cache_update_fsmid(struct nchandle *nch)
1144 {
1145 	struct namecache *ncp;
1146 	struct namecache *scan;
1147 	struct vnode *vp;
1148 
1149 	ncp = nch->ncp;
1150 
1151 	/*
1152 	 * Warning: even if we get a non-NULL vp it could still be in the
1153 	 * middle of a recyclement.  Don't do anything fancy, just set
1154 	 * NCF_FSMID.
1155 	 */
1156 	if ((vp = ncp->nc_vp) != NULL) {
1157 		TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
1158 			for (scan = ncp; scan; scan = scan->nc_parent) {
1159 				if (scan->nc_flag & NCF_FSMID)
1160 					break;
1161 				scan->nc_flag |= NCF_FSMID;
1162 			}
1163 		}
1164 	} else {
1165 		while (ncp && (ncp->nc_flag & NCF_FSMID) == 0) {
1166 			ncp->nc_flag |= NCF_FSMID;
1167 			ncp = ncp->nc_parent;
1168 		}
1169 	}
1170 }
1171 
1172 void
1173 cache_update_fsmid_vp(struct vnode *vp)
1174 {
1175 	struct namecache *ncp;
1176 	struct namecache *scan;
1177 
1178 	TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
1179 		for (scan = ncp; scan; scan = scan->nc_parent) {
1180 			if (scan->nc_flag & NCF_FSMID)
1181 				break;
1182 			scan->nc_flag |= NCF_FSMID;
1183 		}
1184 	}
1185 }
1186 
1187 /*
1188  * If getattr is called on a vnode (e.g. a stat call), the filesystem
1189  * may call this routine to determine if the namecache has the hierarchical
1190  * change flag set, requiring the fsmid to be updated.
1191  *
1192  * Since 0 indicates no support, make sure the filesystem fsmid is at least
1193  * 1.
1194  */
1195 int
1196 cache_check_fsmid_vp(struct vnode *vp, int64_t *fsmid)
1197 {
1198 	struct namecache *ncp;
1199 	int changed = 0;
1200 
1201 	TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
1202 		if (ncp->nc_flag & NCF_FSMID) {
1203 			ncp->nc_flag &= ~NCF_FSMID;
1204 			changed = 1;
1205 		}
1206 	}
1207 	if (*fsmid == 0)
1208 		++*fsmid;
1209 	if (changed)
1210 		++*fsmid;
1211 	return(changed);
1212 }
1213 
1214 /*
1215  * Obtain the FSMID for a vnode for filesystems which do not support
1216  * a built-in FSMID.
1217  */
1218 int64_t
1219 cache_sync_fsmid_vp(struct vnode *vp)
1220 {
1221 	struct namecache *ncp;
1222 
1223 	if ((ncp = TAILQ_FIRST(&vp->v_namecache)) != NULL) {
1224 		if (ncp->nc_flag & NCF_FSMID) {
1225 			ncp->nc_flag &= ~NCF_FSMID;
1226 			++ncp->nc_fsmid;
1227 		}
1228 		return(ncp->nc_fsmid);
1229 	}
1230 	return(VNOVAL);
1231 }
1232 
1233 /*
1234  * Convert a directory vnode to a namecache record without any other
1235  * knowledge of the topology.  This ONLY works with directory vnodes and
1236  * is ONLY used by the NFS server.  dvp must be refd but unlocked, and the
1237  * returned ncp (if not NULL) will be held and unlocked.
1238  *
1239  * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned.
1240  * If 'makeit' is 1 we attempt to track-down and create the namecache topology
1241  * for dvp.  This will fail only if the directory has been deleted out from
1242  * under the caller.
1243  *
1244  * Callers must always check for a NULL return no matter the value of 'makeit'.
1245  *
1246  * To avoid underflowing the kernel stack each recursive call increments
1247  * the makeit variable.
1248  */
1249 
1250 static int cache_inefficient_scan(struct nchandle *nch, struct ucred *cred,
1251 				  struct vnode *dvp);
1252 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1253 				  struct vnode **saved_dvp);
1254 
1255 int
1256 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit,
1257 	      struct nchandle *nch)
1258 {
1259 	struct vnode *saved_dvp;
1260 	struct vnode *pvp;
1261 	int error;
1262 
1263 	nch->ncp = NULL;
1264 	nch->mount = dvp->v_mount;
1265 	saved_dvp = NULL;
1266 
1267 	/*
1268 	 * Temporary debugging code to force the directory scanning code
1269 	 * to be exercised.
1270 	 */
1271 	if (ncvp_debug >= 3 && makeit && TAILQ_FIRST(&dvp->v_namecache)) {
1272 		nch->ncp = TAILQ_FIRST(&dvp->v_namecache);
1273 		kprintf("cache_fromdvp: forcing %s\n", nch->ncp->nc_name);
1274 		goto force;
1275 	}
1276 
1277 	/*
1278 	 * Loop until resolution, inside code will break out on error.
1279 	 */
1280 	while ((nch->ncp = TAILQ_FIRST(&dvp->v_namecache)) == NULL && makeit) {
1281 force:
1282 		/*
1283 		 * If dvp is the root of its filesystem it should already
1284 		 * have a namecache pointer associated with it as a side
1285 		 * effect of the mount, but it may have been disassociated.
1286 		 */
1287 		if (dvp->v_flag & VROOT) {
1288 			nch->ncp = _cache_get(nch->mount->mnt_ncmountpt.ncp);
1289 			error = cache_resolve_mp(nch->mount);
1290 			_cache_put(nch->ncp);
1291 			if (ncvp_debug) {
1292 				kprintf("cache_fromdvp: resolve root of mount %p error %d",
1293 					dvp->v_mount, error);
1294 			}
1295 			if (error) {
1296 				if (ncvp_debug)
1297 					kprintf(" failed\n");
1298 				nch->ncp = NULL;
1299 				break;
1300 			}
1301 			if (ncvp_debug)
1302 				kprintf(" succeeded\n");
1303 			continue;
1304 		}
1305 
1306 		/*
1307 		 * If we are recursed too deeply resort to an O(n^2)
1308 		 * algorithm to resolve the namecache topology.  The
1309 		 * resolved pvp is left referenced in saved_dvp to
1310 		 * prevent the tree from being destroyed while we loop.
1311 		 */
1312 		if (makeit > 20) {
1313 			error = cache_fromdvp_try(dvp, cred, &saved_dvp);
1314 			if (error) {
1315 				kprintf("lookupdotdot(longpath) failed %d "
1316 				       "dvp %p\n", error, dvp);
1317 				break;
1318 			}
1319 			continue;
1320 		}
1321 
1322 		/*
1323 		 * Get the parent directory and resolve its ncp.
1324 		 */
1325 		error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred);
1326 		if (error) {
1327 			kprintf("lookupdotdot failed %d dvp %p\n", error, dvp);
1328 			break;
1329 		}
1330 		vn_unlock(pvp);
1331 
1332 		/*
1333 		 * Reuse makeit as a recursion depth counter.
1334 		 */
1335 		cache_fromdvp(pvp, cred, makeit + 1, nch);
1336 		vrele(pvp);
1337 		if (nch->ncp == NULL)
1338 			break;
1339 
1340 		/*
1341 		 * Do an inefficient scan of pvp (embodied by ncp) to look
1342 		 * for dvp.  This will create a namecache record for dvp on
1343 		 * success.  We loop up to recheck on success.
1344 		 *
1345 		 * ncp and dvp are both held but not locked.
1346 		 */
1347 		error = cache_inefficient_scan(nch, cred, dvp);
1348 		_cache_drop(nch->ncp);
1349 		if (error) {
1350 			kprintf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n",
1351 				pvp, nch->ncp->nc_name, dvp);
1352 			nch->ncp = NULL;
1353 			break;
1354 		}
1355 		if (ncvp_debug) {
1356 			kprintf("cache_fromdvp: scan %p (%s) succeeded\n",
1357 				pvp, nch->ncp->nc_name);
1358 		}
1359 	}
1360 
1361 	/*
1362 	 * hold it for real so the mount gets a ref
1363 	 */
1364 	if (nch->ncp)
1365 		cache_hold(nch);
1366 	if (saved_dvp)
1367 		vrele(saved_dvp);
1368 	if (nch->ncp)
1369 		return (0);
1370 	return (EINVAL);
1371 }
1372 
1373 /*
1374  * Go up the chain of parent directories until we find something
1375  * we can resolve into the namecache.  This is very inefficient.
1376  */
1377 static
1378 int
1379 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1380 		  struct vnode **saved_dvp)
1381 {
1382 	struct nchandle nch;
1383 	struct vnode *pvp;
1384 	int error;
1385 	static time_t last_fromdvp_report;
1386 
1387 	/*
1388 	 * Loop getting the parent directory vnode until we get something we
1389 	 * can resolve in the namecache.
1390 	 */
1391 	vref(dvp);
1392 	nch.mount = dvp->v_mount;
1393 
1394 	for (;;) {
1395 		error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred);
1396 		if (error) {
1397 			vrele(dvp);
1398 			return (error);
1399 		}
1400 		vn_unlock(pvp);
1401 		if ((nch.ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) {
1402 			_cache_hold(nch.ncp);
1403 			vrele(pvp);
1404 			break;
1405 		}
1406 		if (pvp->v_flag & VROOT) {
1407 			nch.ncp = _cache_get(pvp->v_mount->mnt_ncmountpt.ncp);
1408 			error = cache_resolve_mp(nch.mount);
1409 			_cache_unlock(nch.ncp);
1410 			vrele(pvp);
1411 			if (error) {
1412 				_cache_drop(nch.ncp);
1413 				vrele(dvp);
1414 				return (error);
1415 			}
1416 			break;
1417 		}
1418 		vrele(dvp);
1419 		dvp = pvp;
1420 	}
1421 	if (last_fromdvp_report != time_second) {
1422 		last_fromdvp_report = time_second;
1423 		kprintf("Warning: extremely inefficient path resolution on %s\n",
1424 			nch.ncp->nc_name);
1425 	}
1426 	error = cache_inefficient_scan(&nch, cred, dvp);
1427 
1428 	/*
1429 	 * Hopefully dvp now has a namecache record associated with it.
1430 	 * Leave it referenced to prevent the kernel from recycling the
1431 	 * vnode.  Otherwise extremely long directory paths could result
1432 	 * in endless recycling.
1433 	 */
1434 	if (*saved_dvp)
1435 	    vrele(*saved_dvp);
1436 	*saved_dvp = dvp;
1437 	return (error);
1438 }
1439 
1440 
1441 /*
1442  * Do an inefficient scan of the directory represented by ncp looking for
1443  * the directory vnode dvp.  ncp must be held but not locked on entry and
1444  * will be held on return.  dvp must be refd but not locked on entry and
1445  * will remain refd on return.
1446  *
1447  * Why do this at all?  Well, due to its stateless nature the NFS server
1448  * converts file handles directly to vnodes without necessarily going through
1449  * the namecache ops that would otherwise create the namecache topology
1450  * leading to the vnode.  We could either (1) Change the namecache algorithms
1451  * to allow disconnect namecache records that are re-merged opportunistically,
1452  * or (2) Make the NFS server backtrack and scan to recover a connected
1453  * namecache topology in order to then be able to issue new API lookups.
1454  *
1455  * It turns out that (1) is a huge mess.  It takes a nice clean set of
1456  * namecache algorithms and introduces a lot of complication in every subsystem
1457  * that calls into the namecache to deal with the re-merge case, especially
1458  * since we are using the namecache to placehold negative lookups and the
1459  * vnode might not be immediately assigned. (2) is certainly far less
1460  * efficient then (1), but since we are only talking about directories here
1461  * (which are likely to remain cached), the case does not actually run all
1462  * that often and has the supreme advantage of not polluting the namecache
1463  * algorithms.
1464  */
1465 static int
1466 cache_inefficient_scan(struct nchandle *nch, struct ucred *cred,
1467 		       struct vnode *dvp)
1468 {
1469 	struct nlcomponent nlc;
1470 	struct nchandle rncp;
1471 	struct dirent *den;
1472 	struct vnode *pvp;
1473 	struct vattr vat;
1474 	struct iovec iov;
1475 	struct uio uio;
1476 	int blksize;
1477 	int eofflag;
1478 	int bytes;
1479 	char *rbuf;
1480 	int error;
1481 
1482 	vat.va_blocksize = 0;
1483 	if ((error = VOP_GETATTR(dvp, &vat)) != 0)
1484 		return (error);
1485 	if ((error = cache_vref(nch, cred, &pvp)) != 0)
1486 		return (error);
1487 	if (ncvp_debug)
1488 		kprintf("inefficient_scan: directory iosize %ld vattr fileid = %lld\n", vat.va_blocksize, vat.va_fileid);
1489 	if ((blksize = vat.va_blocksize) == 0)
1490 		blksize = DEV_BSIZE;
1491 	rbuf = kmalloc(blksize, M_TEMP, M_WAITOK);
1492 	rncp.ncp = NULL;
1493 
1494 	eofflag = 0;
1495 	uio.uio_offset = 0;
1496 again:
1497 	iov.iov_base = rbuf;
1498 	iov.iov_len = blksize;
1499 	uio.uio_iov = &iov;
1500 	uio.uio_iovcnt = 1;
1501 	uio.uio_resid = blksize;
1502 	uio.uio_segflg = UIO_SYSSPACE;
1503 	uio.uio_rw = UIO_READ;
1504 	uio.uio_td = curthread;
1505 
1506 	if (ncvp_debug >= 2)
1507 		kprintf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset);
1508 	error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL);
1509 	if (error == 0) {
1510 		den = (struct dirent *)rbuf;
1511 		bytes = blksize - uio.uio_resid;
1512 
1513 		while (bytes > 0) {
1514 			if (ncvp_debug >= 2) {
1515 				kprintf("cache_inefficient_scan: %*.*s\n",
1516 					den->d_namlen, den->d_namlen,
1517 					den->d_name);
1518 			}
1519 			if (den->d_type != DT_WHT &&
1520 			    den->d_ino == vat.va_fileid) {
1521 				if (ncvp_debug) {
1522 					kprintf("cache_inefficient_scan: "
1523 					       "MATCHED inode %lld path %s/%*.*s\n",
1524 					       vat.va_fileid, nch->ncp->nc_name,
1525 					       den->d_namlen, den->d_namlen,
1526 					       den->d_name);
1527 				}
1528 				nlc.nlc_nameptr = den->d_name;
1529 				nlc.nlc_namelen = den->d_namlen;
1530 				rncp = cache_nlookup(nch, &nlc);
1531 				KKASSERT(rncp.ncp != NULL);
1532 				break;
1533 			}
1534 			bytes -= _DIRENT_DIRSIZ(den);
1535 			den = _DIRENT_NEXT(den);
1536 		}
1537 		if (rncp.ncp == NULL && eofflag == 0 && uio.uio_resid != blksize)
1538 			goto again;
1539 	}
1540 	vrele(pvp);
1541 	if (rncp.ncp) {
1542 		if (rncp.ncp->nc_flag & NCF_UNRESOLVED) {
1543 			_cache_setvp(rncp.ncp, dvp);
1544 			if (ncvp_debug >= 2) {
1545 				kprintf("cache_inefficient_scan: setvp %s/%s = %p\n",
1546 					nch->ncp->nc_name, rncp.ncp->nc_name, dvp);
1547 			}
1548 		} else {
1549 			if (ncvp_debug >= 2) {
1550 				kprintf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n",
1551 					nch->ncp->nc_name, rncp.ncp->nc_name, dvp,
1552 					rncp.ncp->nc_vp);
1553 			}
1554 		}
1555 		if (rncp.ncp->nc_vp == NULL)
1556 			error = rncp.ncp->nc_error;
1557 		_cache_put(rncp.ncp);
1558 	} else {
1559 		kprintf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n",
1560 			dvp, nch->ncp->nc_name);
1561 		error = ENOENT;
1562 	}
1563 	kfree(rbuf, M_TEMP);
1564 	return (error);
1565 }
1566 
1567 /*
1568  * Zap a namecache entry.  The ncp is unconditionally set to an unresolved
1569  * state, which disassociates it from its vnode or ncneglist.
1570  *
1571  * Then, if there are no additional references to the ncp and no children,
1572  * the ncp is removed from the topology and destroyed.  This function will
1573  * also run through the nc_parent chain and destroy parent ncps if possible.
1574  * As a side benefit, it turns out the only conditions that allow running
1575  * up the chain are also the conditions to ensure no deadlock will occur.
1576  *
1577  * References and/or children may exist if the ncp is in the middle of the
1578  * topology, preventing the ncp from being destroyed.
1579  *
1580  * This function must be called with the ncp held and locked and will unlock
1581  * and drop it during zapping.
1582  */
1583 static void
1584 cache_zap(struct namecache *ncp)
1585 {
1586 	struct namecache *par;
1587 
1588 	/*
1589 	 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED.
1590 	 */
1591 	_cache_setunresolved(ncp);
1592 
1593 	/*
1594 	 * Try to scrap the entry and possibly tail-recurse on its parent.
1595 	 * We only scrap unref'd (other then our ref) unresolved entries,
1596 	 * we do not scrap 'live' entries.
1597 	 */
1598 	while (ncp->nc_flag & NCF_UNRESOLVED) {
1599 		/*
1600 		 * Someone other then us has a ref, stop.
1601 		 */
1602 		if (ncp->nc_refs > 1)
1603 			goto done;
1604 
1605 		/*
1606 		 * We have children, stop.
1607 		 */
1608 		if (!TAILQ_EMPTY(&ncp->nc_list))
1609 			goto done;
1610 
1611 		/*
1612 		 * Remove ncp from the topology: hash table and parent linkage.
1613 		 */
1614 		if (ncp->nc_flag & NCF_HASHED) {
1615 			ncp->nc_flag &= ~NCF_HASHED;
1616 			LIST_REMOVE(ncp, nc_hash);
1617 		}
1618 		if ((par = ncp->nc_parent) != NULL) {
1619 			par = _cache_hold(par);
1620 			TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
1621 			ncp->nc_parent = NULL;
1622 			if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
1623 				vdrop(par->nc_vp);
1624 		}
1625 
1626 		/*
1627 		 * ncp should not have picked up any refs.  Physically
1628 		 * destroy the ncp.
1629 		 */
1630 		KKASSERT(ncp->nc_refs == 1);
1631 		--numunres;
1632 		/* _cache_unlock(ncp) not required */
1633 		ncp->nc_refs = -1;	/* safety */
1634 		if (ncp->nc_name)
1635 			kfree(ncp->nc_name, M_VFSCACHE);
1636 		kfree(ncp, M_VFSCACHE);
1637 
1638 		/*
1639 		 * Loop on the parent (it may be NULL).  Only bother looping
1640 		 * if the parent has a single ref (ours), which also means
1641 		 * we can lock it trivially.
1642 		 */
1643 		ncp = par;
1644 		if (ncp == NULL)
1645 			return;
1646 		if (ncp->nc_refs != 1) {
1647 			_cache_drop(ncp);
1648 			return;
1649 		}
1650 		KKASSERT(par->nc_exlocks == 0);
1651 		_cache_lock(ncp);
1652 	}
1653 done:
1654 	_cache_unlock(ncp);
1655 	atomic_subtract_int(&ncp->nc_refs, 1);
1656 }
1657 
1658 static enum { CHI_LOW, CHI_HIGH } cache_hysteresis_state = CHI_LOW;
1659 
1660 static __inline
1661 void
1662 cache_hysteresis(void)
1663 {
1664 	/*
1665 	 * Don't cache too many negative hits.  We use hysteresis to reduce
1666 	 * the impact on the critical path.
1667 	 */
1668 	switch(cache_hysteresis_state) {
1669 	case CHI_LOW:
1670 		if (numneg > MINNEG && numneg * ncnegfactor > numcache) {
1671 			cache_cleanneg(10);
1672 			cache_hysteresis_state = CHI_HIGH;
1673 		}
1674 		break;
1675 	case CHI_HIGH:
1676 		if (numneg > MINNEG * 9 / 10 &&
1677 		    numneg * ncnegfactor * 9 / 10 > numcache
1678 		) {
1679 			cache_cleanneg(10);
1680 		} else {
1681 			cache_hysteresis_state = CHI_LOW;
1682 		}
1683 		break;
1684 	}
1685 }
1686 
1687 /*
1688  * NEW NAMECACHE LOOKUP API
1689  *
1690  * Lookup an entry in the cache.  A locked, referenced, non-NULL
1691  * entry is *always* returned, even if the supplied component is illegal.
1692  * The resulting namecache entry should be returned to the system with
1693  * cache_put() or _cache_unlock() + cache_drop().
1694  *
1695  * namecache locks are recursive but care must be taken to avoid lock order
1696  * reversals.
1697  *
1698  * Nobody else will be able to manipulate the associated namespace (e.g.
1699  * create, delete, rename, rename-target) until the caller unlocks the
1700  * entry.
1701  *
1702  * The returned entry will be in one of three states:  positive hit (non-null
1703  * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set).
1704  * Unresolved entries must be resolved through the filesystem to associate the
1705  * vnode and/or determine whether a positive or negative hit has occured.
1706  *
1707  * It is not necessary to lock a directory in order to lock namespace under
1708  * that directory.  In fact, it is explicitly not allowed to do that.  A
1709  * directory is typically only locked when being created, renamed, or
1710  * destroyed.
1711  *
1712  * The directory (par) may be unresolved, in which case any returned child
1713  * will likely also be marked unresolved.  Likely but not guarenteed.  Since
1714  * the filesystem lookup requires a resolved directory vnode the caller is
1715  * responsible for resolving the namecache chain top-down.  This API
1716  * specifically allows whole chains to be created in an unresolved state.
1717  */
1718 struct nchandle
1719 cache_nlookup(struct nchandle *par_nch, struct nlcomponent *nlc)
1720 {
1721 	struct nchandle nch;
1722 	struct namecache *ncp;
1723 	struct namecache *new_ncp;
1724 	struct nchashhead *nchpp;
1725 	u_int32_t hash;
1726 	globaldata_t gd;
1727 
1728 	numcalls++;
1729 	gd = mycpu;
1730 
1731 	/*
1732 	 * Try to locate an existing entry
1733 	 */
1734 	hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
1735 	hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
1736 	new_ncp = NULL;
1737 restart:
1738 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1739 		numchecks++;
1740 
1741 		/*
1742 		 * Try to zap entries that have timed out.  We have
1743 		 * to be careful here because locked leafs may depend
1744 		 * on the vnode remaining intact in a parent, so only
1745 		 * do this under very specific conditions.
1746 		 */
1747 		if (ncp->nc_timeout &&
1748 		    (int)(ncp->nc_timeout - ticks) < 0 &&
1749 		    (ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
1750 		    ncp->nc_exlocks == 0 &&
1751 		    TAILQ_EMPTY(&ncp->nc_list)
1752 		) {
1753 			cache_zap(_cache_get(ncp));
1754 			goto restart;
1755 		}
1756 
1757 		/*
1758 		 * Break out if we find a matching entry.  Note that
1759 		 * UNRESOLVED entries may match, but DESTROYED entries
1760 		 * do not.
1761 		 */
1762 		if (ncp->nc_parent == par_nch->ncp &&
1763 		    ncp->nc_nlen == nlc->nlc_namelen &&
1764 		    bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
1765 		    (ncp->nc_flag & NCF_DESTROYED) == 0
1766 		) {
1767 			if (_cache_get_nonblock(ncp) == 0) {
1768 				if (new_ncp)
1769 					_cache_free(new_ncp);
1770 				goto found;
1771 			}
1772 			_cache_get(ncp);
1773 			_cache_put(ncp);
1774 			goto restart;
1775 		}
1776 	}
1777 
1778 	/*
1779 	 * We failed to locate an entry, create a new entry and add it to
1780 	 * the cache.  We have to relookup after possibly blocking in
1781 	 * malloc.
1782 	 */
1783 	if (new_ncp == NULL) {
1784 		new_ncp = cache_alloc(nlc->nlc_namelen);
1785 		goto restart;
1786 	}
1787 
1788 	ncp = new_ncp;
1789 
1790 	/*
1791 	 * Initialize as a new UNRESOLVED entry, lock (non-blocking),
1792 	 * and link to the parent.  The mount point is usually inherited
1793 	 * from the parent unless this is a special case such as a mount
1794 	 * point where nlc_namelen is 0.   If nlc_namelen is 0 nc_name will
1795 	 * be NULL.
1796 	 */
1797 	if (nlc->nlc_namelen) {
1798 		bcopy(nlc->nlc_nameptr, ncp->nc_name, nlc->nlc_namelen);
1799 		ncp->nc_name[nlc->nlc_namelen] = 0;
1800 	}
1801 	nchpp = NCHHASH(hash);
1802 	LIST_INSERT_HEAD(nchpp, ncp, nc_hash);
1803 	ncp->nc_flag |= NCF_HASHED;
1804 	cache_link_parent(ncp, par_nch->ncp);
1805 found:
1806 	/*
1807 	 * stats and namecache size management
1808 	 */
1809 	if (ncp->nc_flag & NCF_UNRESOLVED)
1810 		++gd->gd_nchstats->ncs_miss;
1811 	else if (ncp->nc_vp)
1812 		++gd->gd_nchstats->ncs_goodhits;
1813 	else
1814 		++gd->gd_nchstats->ncs_neghits;
1815 	cache_hysteresis();
1816 	nch.mount = par_nch->mount;
1817 	nch.ncp = ncp;
1818 	++nch.mount->mnt_refs;
1819 	return(nch);
1820 }
1821 
1822 /*
1823  * The namecache entry is marked as being used as a mount point.
1824  * Locate the mount if it is visible to the caller.
1825  */
1826 struct findmount_info {
1827 	struct mount *result;
1828 	struct mount *nch_mount;
1829 	struct namecache *nch_ncp;
1830 };
1831 
1832 static
1833 int
1834 cache_findmount_callback(struct mount *mp, void *data)
1835 {
1836 	struct findmount_info *info = data;
1837 
1838 	/*
1839 	 * Check the mount's mounted-on point against the passed nch.
1840 	 */
1841 	if (mp->mnt_ncmounton.mount == info->nch_mount &&
1842 	    mp->mnt_ncmounton.ncp == info->nch_ncp
1843 	) {
1844 	    info->result = mp;
1845 	    return(-1);
1846 	}
1847 	return(0);
1848 }
1849 
1850 struct mount *
1851 cache_findmount(struct nchandle *nch)
1852 {
1853 	struct findmount_info info;
1854 
1855 	info.result = NULL;
1856 	info.nch_mount = nch->mount;
1857 	info.nch_ncp = nch->ncp;
1858 	mountlist_scan(cache_findmount_callback, &info,
1859 			       MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
1860 	return(info.result);
1861 }
1862 
1863 /*
1864  * Resolve an unresolved namecache entry, generally by looking it up.
1865  * The passed ncp must be locked and refd.
1866  *
1867  * Theoretically since a vnode cannot be recycled while held, and since
1868  * the nc_parent chain holds its vnode as long as children exist, the
1869  * direct parent of the cache entry we are trying to resolve should
1870  * have a valid vnode.  If not then generate an error that we can
1871  * determine is related to a resolver bug.
1872  *
1873  * However, if a vnode was in the middle of a recyclement when the NCP
1874  * got locked, ncp->nc_vp might point to a vnode that is about to become
1875  * invalid.  cache_resolve() handles this case by unresolving the entry
1876  * and then re-resolving it.
1877  *
1878  * Note that successful resolution does not necessarily return an error
1879  * code of 0.  If the ncp resolves to a negative cache hit then ENOENT
1880  * will be returned.
1881  */
1882 int
1883 cache_resolve(struct nchandle *nch, struct ucred *cred)
1884 {
1885 	struct namecache *par;
1886 	struct namecache *ncp;
1887 	struct nchandle nctmp;
1888 	struct mount *mp;
1889 	struct vnode *dvp;
1890 	int error;
1891 
1892 	ncp = nch->ncp;
1893 	mp = nch->mount;
1894 restart:
1895 	/*
1896 	 * If the ncp is already resolved we have nothing to do.  However,
1897 	 * we do want to guarentee that a usable vnode is returned when
1898 	 * a vnode is present, so make sure it hasn't been reclaimed.
1899 	 */
1900 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1901 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1902 			_cache_setunresolved(ncp);
1903 		if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
1904 			return (ncp->nc_error);
1905 	}
1906 
1907 	/*
1908 	 * Mount points need special handling because the parent does not
1909 	 * belong to the same filesystem as the ncp.
1910 	 */
1911 	if (ncp == mp->mnt_ncmountpt.ncp)
1912 		return (cache_resolve_mp(mp));
1913 
1914 	/*
1915 	 * We expect an unbroken chain of ncps to at least the mount point,
1916 	 * and even all the way to root (but this code doesn't have to go
1917 	 * past the mount point).
1918 	 */
1919 	if (ncp->nc_parent == NULL) {
1920 		kprintf("EXDEV case 1 %p %*.*s\n", ncp,
1921 			ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
1922 		ncp->nc_error = EXDEV;
1923 		return(ncp->nc_error);
1924 	}
1925 
1926 	/*
1927 	 * The vp's of the parent directories in the chain are held via vhold()
1928 	 * due to the existance of the child, and should not disappear.
1929 	 * However, there are cases where they can disappear:
1930 	 *
1931 	 *	- due to filesystem I/O errors.
1932 	 *	- due to NFS being stupid about tracking the namespace and
1933 	 *	  destroys the namespace for entire directories quite often.
1934 	 *	- due to forced unmounts.
1935 	 *	- due to an rmdir (parent will be marked DESTROYED)
1936 	 *
1937 	 * When this occurs we have to track the chain backwards and resolve
1938 	 * it, looping until the resolver catches up to the current node.  We
1939 	 * could recurse here but we might run ourselves out of kernel stack
1940 	 * so we do it in a more painful manner.  This situation really should
1941 	 * not occur all that often, or if it does not have to go back too
1942 	 * many nodes to resolve the ncp.
1943 	 */
1944 	while (ncp->nc_parent->nc_vp == NULL) {
1945 		/*
1946 		 * This case can occur if a process is CD'd into a
1947 		 * directory which is then rmdir'd.  If the parent is marked
1948 		 * destroyed there is no point trying to resolve it.
1949 		 */
1950 		if (ncp->nc_parent->nc_flag & NCF_DESTROYED)
1951 			return(ENOENT);
1952 
1953 		par = ncp->nc_parent;
1954 		while (par->nc_parent && par->nc_parent->nc_vp == NULL)
1955 			par = par->nc_parent;
1956 		if (par->nc_parent == NULL) {
1957 			kprintf("EXDEV case 2 %*.*s\n",
1958 				par->nc_nlen, par->nc_nlen, par->nc_name);
1959 			return (EXDEV);
1960 		}
1961 		kprintf("[diagnostic] cache_resolve: had to recurse on %*.*s\n",
1962 			par->nc_nlen, par->nc_nlen, par->nc_name);
1963 		/*
1964 		 * The parent is not set in stone, ref and lock it to prevent
1965 		 * it from disappearing.  Also note that due to renames it
1966 		 * is possible for our ncp to move and for par to no longer
1967 		 * be one of its parents.  We resolve it anyway, the loop
1968 		 * will handle any moves.
1969 		 */
1970 		_cache_get(par);
1971 		if (par == nch->mount->mnt_ncmountpt.ncp) {
1972 			cache_resolve_mp(nch->mount);
1973 		} else if ((dvp = par->nc_parent->nc_vp) == NULL) {
1974 			kprintf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name);
1975 			_cache_put(par);
1976 			continue;
1977 		} else if (par->nc_flag & NCF_UNRESOLVED) {
1978 			/* vhold(dvp); - DVP can't go away */
1979 			nctmp.mount = mp;
1980 			nctmp.ncp = par;
1981 			par->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred);
1982 			/* vdrop(dvp); */
1983 		}
1984 		if ((error = par->nc_error) != 0) {
1985 			if (par->nc_error != EAGAIN) {
1986 				kprintf("EXDEV case 3 %*.*s error %d\n",
1987 				    par->nc_nlen, par->nc_nlen, par->nc_name,
1988 				    par->nc_error);
1989 				_cache_put(par);
1990 				return(error);
1991 			}
1992 			kprintf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n",
1993 				par, par->nc_nlen, par->nc_nlen, par->nc_name);
1994 		}
1995 		_cache_put(par);
1996 		/* loop */
1997 	}
1998 
1999 	/*
2000 	 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected
2001 	 * ncp's and reattach them.  If this occurs the original ncp is marked
2002 	 * EAGAIN to force a relookup.
2003 	 *
2004 	 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed
2005 	 * ncp must already be resolved.
2006 	 */
2007 	dvp = ncp->nc_parent->nc_vp;
2008 	/* vhold(dvp); - dvp can't go away */
2009 	nctmp.mount = mp;
2010 	nctmp.ncp = ncp;
2011 	ncp->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred);
2012 	/* vdrop(dvp); */
2013 	if (ncp->nc_error == EAGAIN) {
2014 		kprintf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n",
2015 			ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
2016 		goto restart;
2017 	}
2018 	return(ncp->nc_error);
2019 }
2020 
2021 /*
2022  * Resolve the ncp associated with a mount point.  Such ncp's almost always
2023  * remain resolved and this routine is rarely called.  NFS MPs tends to force
2024  * re-resolution more often due to its mac-truck-smash-the-namecache
2025  * method of tracking namespace changes.
2026  *
2027  * The semantics for this call is that the passed ncp must be locked on
2028  * entry and will be locked on return.  However, if we actually have to
2029  * resolve the mount point we temporarily unlock the entry in order to
2030  * avoid race-to-root deadlocks due to e.g. dead NFS mounts.  Because of
2031  * the unlock we have to recheck the flags after we relock.
2032  */
2033 static int
2034 cache_resolve_mp(struct mount *mp)
2035 {
2036 	struct namecache *ncp = mp->mnt_ncmountpt.ncp;
2037 	struct vnode *vp;
2038 	int error;
2039 
2040 	KKASSERT(mp != NULL);
2041 
2042 	/*
2043 	 * If the ncp is already resolved we have nothing to do.  However,
2044 	 * we do want to guarentee that a usable vnode is returned when
2045 	 * a vnode is present, so make sure it hasn't been reclaimed.
2046 	 */
2047 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
2048 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
2049 			_cache_setunresolved(ncp);
2050 	}
2051 
2052 	if (ncp->nc_flag & NCF_UNRESOLVED) {
2053 		_cache_unlock(ncp);
2054 		while (vfs_busy(mp, 0))
2055 			;
2056 		error = VFS_ROOT(mp, &vp);
2057 		_cache_lock(ncp);
2058 
2059 		/*
2060 		 * recheck the ncp state after relocking.
2061 		 */
2062 		if (ncp->nc_flag & NCF_UNRESOLVED) {
2063 			ncp->nc_error = error;
2064 			if (error == 0) {
2065 				_cache_setvp(ncp, vp);
2066 				vput(vp);
2067 			} else {
2068 				kprintf("[diagnostic] cache_resolve_mp: failed to resolve mount %p\n", mp);
2069 				_cache_setvp(ncp, NULL);
2070 			}
2071 		} else if (error == 0) {
2072 			vput(vp);
2073 		}
2074 		vfs_unbusy(mp);
2075 	}
2076 	return(ncp->nc_error);
2077 }
2078 
2079 void
2080 cache_cleanneg(int count)
2081 {
2082 	struct namecache *ncp;
2083 
2084 	/*
2085 	 * Automode from the vnlru proc - clean out 10% of the negative cache
2086 	 * entries.
2087 	 */
2088 	if (count == 0)
2089 		count = numneg / 10 + 1;
2090 
2091 	/*
2092 	 * Attempt to clean out the specified number of negative cache
2093 	 * entries.
2094 	 */
2095 	while (count) {
2096 		ncp = TAILQ_FIRST(&ncneglist);
2097 		if (ncp == NULL) {
2098 			KKASSERT(numneg == 0);
2099 			break;
2100 		}
2101 		TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
2102 		TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
2103 		if (_cache_get_nonblock(ncp) == 0)
2104 			cache_zap(ncp);
2105 		--count;
2106 	}
2107 }
2108 
2109 /*
2110  * Rehash a ncp.  Rehashing is typically required if the name changes (should
2111  * not generally occur) or the parent link changes.  This function will
2112  * unhash the ncp if the ncp is no longer hashable.
2113  */
2114 static void
2115 _cache_rehash(struct namecache *ncp)
2116 {
2117 	struct nchashhead *nchpp;
2118 	u_int32_t hash;
2119 
2120 	if (ncp->nc_flag & NCF_HASHED) {
2121 		ncp->nc_flag &= ~NCF_HASHED;
2122 		LIST_REMOVE(ncp, nc_hash);
2123 	}
2124 	if (ncp->nc_nlen && ncp->nc_parent) {
2125 		hash = fnv_32_buf(ncp->nc_name, ncp->nc_nlen, FNV1_32_INIT);
2126 		hash = fnv_32_buf(&ncp->nc_parent,
2127 					sizeof(ncp->nc_parent), hash);
2128 		nchpp = NCHHASH(hash);
2129 		LIST_INSERT_HEAD(nchpp, ncp, nc_hash);
2130 		ncp->nc_flag |= NCF_HASHED;
2131 	}
2132 }
2133 
2134 /*
2135  * Name cache initialization, from vfsinit() when we are booting
2136  */
2137 void
2138 nchinit(void)
2139 {
2140 	int i;
2141 	globaldata_t gd;
2142 
2143 	/* initialise per-cpu namecache effectiveness statistics. */
2144 	for (i = 0; i < ncpus; ++i) {
2145 		gd = globaldata_find(i);
2146 		gd->gd_nchstats = &nchstats[i];
2147 	}
2148 	TAILQ_INIT(&ncneglist);
2149 	nchashtbl = hashinit(desiredvnodes*2, M_VFSCACHE, &nchash);
2150 	nclockwarn = 1 * hz;
2151 }
2152 
2153 /*
2154  * Called from start_init() to bootstrap the root filesystem.  Returns
2155  * a referenced, unlocked namecache record.
2156  */
2157 void
2158 cache_allocroot(struct nchandle *nch, struct mount *mp, struct vnode *vp)
2159 {
2160 	nch->ncp = cache_alloc(0);
2161 	nch->mount = mp;
2162 	++mp->mnt_refs;
2163 	if (vp)
2164 		_cache_setvp(nch->ncp, vp);
2165 }
2166 
2167 /*
2168  * vfs_cache_setroot()
2169  *
2170  *	Create an association between the root of our namecache and
2171  *	the root vnode.  This routine may be called several times during
2172  *	booting.
2173  *
2174  *	If the caller intends to save the returned namecache pointer somewhere
2175  *	it must cache_hold() it.
2176  */
2177 void
2178 vfs_cache_setroot(struct vnode *nvp, struct nchandle *nch)
2179 {
2180 	struct vnode *ovp;
2181 	struct nchandle onch;
2182 
2183 	ovp = rootvnode;
2184 	onch = rootnch;
2185 	rootvnode = nvp;
2186 	if (nch)
2187 		rootnch = *nch;
2188 	else
2189 		cache_zero(&rootnch);
2190 	if (ovp)
2191 		vrele(ovp);
2192 	if (onch.ncp)
2193 		cache_drop(&onch);
2194 }
2195 
2196 /*
2197  * XXX OLD API COMPAT FUNCTION.  This really messes up the new namecache
2198  * topology and is being removed as quickly as possible.  The new VOP_N*()
2199  * API calls are required to make specific adjustments using the supplied
2200  * ncp pointers rather then just bogusly purging random vnodes.
2201  *
2202  * Invalidate all namecache entries to a particular vnode as well as
2203  * any direct children of that vnode in the namecache.  This is a
2204  * 'catch all' purge used by filesystems that do not know any better.
2205  *
2206  * Note that the linkage between the vnode and its namecache entries will
2207  * be removed, but the namecache entries themselves might stay put due to
2208  * active references from elsewhere in the system or due to the existance of
2209  * the children.   The namecache topology is left intact even if we do not
2210  * know what the vnode association is.  Such entries will be marked
2211  * NCF_UNRESOLVED.
2212  */
2213 void
2214 cache_purge(struct vnode *vp)
2215 {
2216 	cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN);
2217 }
2218 
2219 /*
2220  * Flush all entries referencing a particular filesystem.
2221  *
2222  * Since we need to check it anyway, we will flush all the invalid
2223  * entries at the same time.
2224  */
2225 #if 0
2226 
2227 void
2228 cache_purgevfs(struct mount *mp)
2229 {
2230 	struct nchashhead *nchpp;
2231 	struct namecache *ncp, *nnp;
2232 
2233 	/*
2234 	 * Scan hash tables for applicable entries.
2235 	 */
2236 	for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) {
2237 		ncp = LIST_FIRST(nchpp);
2238 		if (ncp)
2239 			_cache_hold(ncp);
2240 		while (ncp) {
2241 			nnp = LIST_NEXT(ncp, nc_hash);
2242 			if (nnp)
2243 				_cache_hold(nnp);
2244 			if (ncp->nc_mount == mp) {
2245 				_cache_lock(ncp);
2246 				cache_zap(ncp);
2247 			} else {
2248 				_cache_drop(ncp);
2249 			}
2250 			ncp = nnp;
2251 		}
2252 	}
2253 }
2254 
2255 #endif
2256 
2257 /*
2258  * Create a new (theoretically) unique fsmid
2259  */
2260 int64_t
2261 cache_getnewfsmid(void)
2262 {
2263 	static int fsmid_roller;
2264 	int64_t fsmid;
2265 
2266 	++fsmid_roller;
2267 	fsmid = ((int64_t)time_second << 32) |
2268 			(fsmid_roller & 0x7FFFFFFF);
2269 	return (fsmid);
2270 }
2271 
2272 
2273 static int disablecwd;
2274 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, "");
2275 
2276 static u_long numcwdcalls; STATNODE(CTLFLAG_RD, numcwdcalls, &numcwdcalls);
2277 static u_long numcwdfail1; STATNODE(CTLFLAG_RD, numcwdfail1, &numcwdfail1);
2278 static u_long numcwdfail2; STATNODE(CTLFLAG_RD, numcwdfail2, &numcwdfail2);
2279 static u_long numcwdfail3; STATNODE(CTLFLAG_RD, numcwdfail3, &numcwdfail3);
2280 static u_long numcwdfail4; STATNODE(CTLFLAG_RD, numcwdfail4, &numcwdfail4);
2281 static u_long numcwdfound; STATNODE(CTLFLAG_RD, numcwdfound, &numcwdfound);
2282 
2283 int
2284 sys___getcwd(struct __getcwd_args *uap)
2285 {
2286 	int buflen;
2287 	int error;
2288 	char *buf;
2289 	char *bp;
2290 
2291 	if (disablecwd)
2292 		return (ENODEV);
2293 
2294 	buflen = uap->buflen;
2295 	if (buflen < 2)
2296 		return (EINVAL);
2297 	if (buflen > MAXPATHLEN)
2298 		buflen = MAXPATHLEN;
2299 
2300 	buf = kmalloc(buflen, M_TEMP, M_WAITOK);
2301 	bp = kern_getcwd(buf, buflen, &error);
2302 	if (error == 0)
2303 		error = copyout(bp, uap->buf, strlen(bp) + 1);
2304 	kfree(buf, M_TEMP);
2305 	return (error);
2306 }
2307 
2308 char *
2309 kern_getcwd(char *buf, size_t buflen, int *error)
2310 {
2311 	struct proc *p = curproc;
2312 	char *bp;
2313 	int i, slash_prefixed;
2314 	struct filedesc *fdp;
2315 	struct nchandle nch;
2316 
2317 	numcwdcalls++;
2318 	bp = buf;
2319 	bp += buflen - 1;
2320 	*bp = '\0';
2321 	fdp = p->p_fd;
2322 	slash_prefixed = 0;
2323 
2324 	nch = fdp->fd_ncdir;
2325 	while (nch.ncp && (nch.ncp != fdp->fd_nrdir.ncp ||
2326 	       nch.mount != fdp->fd_nrdir.mount)
2327 	) {
2328 		/*
2329 		 * While traversing upwards if we encounter the root
2330 		 * of the current mount we have to skip to the mount point
2331 		 * in the underlying filesystem.
2332 		 */
2333 		if (nch.ncp == nch.mount->mnt_ncmountpt.ncp) {
2334 			nch = nch.mount->mnt_ncmounton;
2335 			continue;
2336 		}
2337 
2338 		/*
2339 		 * Prepend the path segment
2340 		 */
2341 		for (i = nch.ncp->nc_nlen - 1; i >= 0; i--) {
2342 			if (bp == buf) {
2343 				numcwdfail4++;
2344 				*error = ENOMEM;
2345 				return(NULL);
2346 			}
2347 			*--bp = nch.ncp->nc_name[i];
2348 		}
2349 		if (bp == buf) {
2350 			numcwdfail4++;
2351 			*error = ENOMEM;
2352 			return(NULL);
2353 		}
2354 		*--bp = '/';
2355 		slash_prefixed = 1;
2356 
2357 		/*
2358 		 * Go up a directory.  This isn't a mount point so we don't
2359 		 * have to check again.
2360 		 */
2361 		nch.ncp = nch.ncp->nc_parent;
2362 	}
2363 	if (nch.ncp == NULL) {
2364 		numcwdfail2++;
2365 		*error = ENOENT;
2366 		return(NULL);
2367 	}
2368 	if (!slash_prefixed) {
2369 		if (bp == buf) {
2370 			numcwdfail4++;
2371 			*error = ENOMEM;
2372 			return(NULL);
2373 		}
2374 		*--bp = '/';
2375 	}
2376 	numcwdfound++;
2377 	*error = 0;
2378 	return (bp);
2379 }
2380 
2381 /*
2382  * Thus begins the fullpath magic.
2383  */
2384 
2385 #undef STATNODE
2386 #define STATNODE(name)							\
2387 	static u_int name;						\
2388 	SYSCTL_UINT(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, "")
2389 
2390 static int disablefullpath;
2391 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW,
2392     &disablefullpath, 0, "");
2393 
2394 STATNODE(numfullpathcalls);
2395 STATNODE(numfullpathfail1);
2396 STATNODE(numfullpathfail2);
2397 STATNODE(numfullpathfail3);
2398 STATNODE(numfullpathfail4);
2399 STATNODE(numfullpathfound);
2400 
2401 int
2402 cache_fullpath(struct proc *p, struct nchandle *nchp, char **retbuf, char **freebuf)
2403 {
2404 	char *bp, *buf;
2405 	int i, slash_prefixed;
2406 	struct nchandle fd_nrdir;
2407 	struct nchandle nch;
2408 
2409 	numfullpathcalls--;
2410 
2411 	*retbuf = NULL;
2412 	*freebuf = NULL;
2413 
2414 	buf = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2415 	bp = buf + MAXPATHLEN - 1;
2416 	*bp = '\0';
2417 	if (p != NULL)
2418 		fd_nrdir = p->p_fd->fd_nrdir;
2419 	else
2420 		fd_nrdir = rootnch;
2421 	slash_prefixed = 0;
2422 	nch = *nchp;
2423 
2424 	while (nch.ncp &&
2425 	       (nch.ncp != fd_nrdir.ncp || nch.mount != fd_nrdir.mount)
2426 	) {
2427 		/*
2428 		 * While traversing upwards if we encounter the root
2429 		 * of the current mount we have to skip to the mount point.
2430 		 */
2431 		if (nch.ncp == nch.mount->mnt_ncmountpt.ncp) {
2432 			nch = nch.mount->mnt_ncmounton;
2433 			continue;
2434 		}
2435 
2436 		/*
2437 		 * Prepend the path segment
2438 		 */
2439 		for (i = nch.ncp->nc_nlen - 1; i >= 0; i--) {
2440 			if (bp == buf) {
2441 				numfullpathfail4++;
2442 				kfree(buf, M_TEMP);
2443 				return(ENOMEM);
2444 			}
2445 			*--bp = nch.ncp->nc_name[i];
2446 		}
2447 		if (bp == buf) {
2448 			numfullpathfail4++;
2449 			kfree(buf, M_TEMP);
2450 			return(ENOMEM);
2451 		}
2452 		*--bp = '/';
2453 		slash_prefixed = 1;
2454 
2455 		/*
2456 		 * Go up a directory.  This isn't a mount point so we don't
2457 		 * have to check again.
2458 		 */
2459 		nch.ncp = nch.ncp->nc_parent;
2460 	}
2461 	if (nch.ncp == NULL) {
2462 		numfullpathfail2++;
2463 		kfree(buf, M_TEMP);
2464 		return(ENOENT);
2465 	}
2466 
2467 	if (!slash_prefixed) {
2468 		if (bp == buf) {
2469 			numfullpathfail4++;
2470 			kfree(buf, M_TEMP);
2471 			return(ENOMEM);
2472 		}
2473 		*--bp = '/';
2474 	}
2475 	numfullpathfound++;
2476 	*retbuf = bp;
2477 	*freebuf = buf;
2478 
2479 	return(0);
2480 }
2481 
2482 int
2483 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf)
2484 {
2485 	struct namecache *ncp;
2486 	struct nchandle nch;
2487 
2488 	numfullpathcalls++;
2489 	if (disablefullpath)
2490 		return (ENODEV);
2491 
2492 	if (p == NULL)
2493 		return (EINVAL);
2494 
2495 	/* vn is NULL, client wants us to use p->p_textvp */
2496 	if (vn == NULL) {
2497 		if ((vn = p->p_textvp) == NULL)
2498 			return (EINVAL);
2499 	}
2500 	TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) {
2501 		if (ncp->nc_nlen)
2502 			break;
2503 	}
2504 	if (ncp == NULL)
2505 		return (EINVAL);
2506 
2507 	numfullpathcalls--;
2508 	nch.ncp = ncp;;
2509 	nch.mount = vn->v_mount;
2510 	return(cache_fullpath(p, &nch, retbuf, freebuf));
2511 }
2512