xref: /dflybsd-src/sys/kern/vfs_cache.c (revision 120e5d0e14888b9ac5a09c87429292256e21c164)
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 1.72 2006/06/05 07:26:10 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  * If we are creating a child under an oldapi parent we must mark the
247  * child as being an oldapi entry as well.
248  */
249 static void
250 cache_link_parent(struct namecache *ncp, struct namecache *par)
251 {
252 	KKASSERT(ncp->nc_parent == NULL);
253 	ncp->nc_parent = par;
254 	if (TAILQ_EMPTY(&par->nc_list)) {
255 		TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
256 		/*
257 		 * Any vp associated with an ncp which has children must
258 		 * be held to prevent it from being recycled.
259 		 */
260 		if (par->nc_vp)
261 			vhold(par->nc_vp);
262 	} else {
263 		TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
264 	}
265 }
266 
267 /*
268  * Remove the parent association from a namecache structure.  If this is
269  * the last child of the parent the cache_drop(par) will attempt to
270  * recursively zap the parent.
271  */
272 static void
273 cache_unlink_parent(struct namecache *ncp)
274 {
275 	struct namecache *par;
276 
277 	if ((par = ncp->nc_parent) != NULL) {
278 		ncp->nc_parent = NULL;
279 		par = cache_hold(par);
280 		TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
281 		if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
282 			vdrop(par->nc_vp);
283 		cache_drop(par);
284 	}
285 }
286 
287 /*
288  * Allocate a new namecache structure.  Most of the code does not require
289  * zero-termination of the string but it makes vop_compat_ncreate() easier.
290  */
291 static struct namecache *
292 cache_alloc(int nlen)
293 {
294 	struct namecache *ncp;
295 
296 	ncp = malloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO);
297 	if (nlen)
298 		ncp->nc_name = malloc(nlen + 1, M_VFSCACHE, M_WAITOK);
299 	ncp->nc_nlen = nlen;
300 	ncp->nc_flag = NCF_UNRESOLVED;
301 	ncp->nc_error = ENOTCONN;	/* needs to be resolved */
302 	ncp->nc_refs = 1;
303 
304 	/*
305 	 * Construct a fake FSMID based on the time of day and a 32 bit
306 	 * roller for uniqueness.  This is used to generate a useful
307 	 * FSMID for filesystems which do not support it.
308 	 */
309 	ncp->nc_fsmid = cache_getnewfsmid();
310 	TAILQ_INIT(&ncp->nc_list);
311 	cache_lock(ncp);
312 	return(ncp);
313 }
314 
315 static void
316 cache_free(struct namecache *ncp)
317 {
318 	KKASSERT(ncp->nc_refs == 1 && ncp->nc_exlocks == 1);
319 	if (ncp->nc_name)
320 		free(ncp->nc_name, M_VFSCACHE);
321 	free(ncp, M_VFSCACHE);
322 }
323 
324 /*
325  * Ref and deref a namecache structure.
326  *
327  * Warning: caller may hold an unrelated read spinlock, which means we can't
328  * use read spinlocks here.
329  */
330 struct namecache *
331 cache_hold(struct namecache *ncp)
332 {
333 	return(_cache_hold(ncp));
334 }
335 
336 void
337 cache_drop(struct namecache *ncp)
338 {
339 	_cache_drop(ncp);
340 }
341 
342 /*
343  * Namespace locking.  The caller must already hold a reference to the
344  * namecache structure in order to lock/unlock it.  This function prevents
345  * the namespace from being created or destroyed by accessors other then
346  * the lock holder.
347  *
348  * Note that holding a locked namecache structure prevents other threads
349  * from making namespace changes (e.g. deleting or creating), prevents
350  * vnode association state changes by other threads, and prevents the
351  * namecache entry from being resolved or unresolved by other threads.
352  *
353  * The lock owner has full authority to associate/disassociate vnodes
354  * and resolve/unresolve the locked ncp.
355  *
356  * WARNING!  Holding a locked ncp will prevent a vnode from being destroyed
357  * or recycled, but it does NOT help you if the vnode had already initiated
358  * a recyclement.  If this is important, use cache_get() rather then
359  * cache_lock() (and deal with the differences in the way the refs counter
360  * is handled).  Or, alternatively, make an unconditional call to
361  * cache_validate() or cache_resolve() after cache_lock() returns.
362  */
363 void
364 cache_lock(struct namecache *ncp)
365 {
366 	thread_t td;
367 	int didwarn;
368 
369 	KKASSERT(ncp->nc_refs != 0);
370 	didwarn = 0;
371 	td = curthread;
372 
373 	for (;;) {
374 		if (ncp->nc_exlocks == 0) {
375 			ncp->nc_exlocks = 1;
376 			ncp->nc_locktd = td;
377 			/*
378 			 * The vp associated with a locked ncp must be held
379 			 * to prevent it from being recycled (which would
380 			 * cause the ncp to become unresolved).
381 			 *
382 			 * WARNING!  If VRECLAIMED is set the vnode could
383 			 * already be in the middle of a recycle.  Callers
384 			 * should not assume that nc_vp is usable when
385 			 * not NULL.  cache_vref() or cache_vget() must be
386 			 * called.
387 			 *
388 			 * XXX loop on race for later MPSAFE work.
389 			 */
390 			if (ncp->nc_vp)
391 				vhold(ncp->nc_vp);
392 			break;
393 		}
394 		if (ncp->nc_locktd == td) {
395 			++ncp->nc_exlocks;
396 			break;
397 		}
398 		ncp->nc_flag |= NCF_LOCKREQ;
399 		if (tsleep(ncp, 0, "clock", nclockwarn) == EWOULDBLOCK) {
400 			if (didwarn)
401 				continue;
402 			didwarn = 1;
403 			printf("[diagnostic] cache_lock: blocked on %p", ncp);
404 			if ((ncp->nc_flag & NCF_MOUNTPT) && ncp->nc_mount)
405 			    printf(" [MOUNTFROM %s]\n", ncp->nc_mount->mnt_stat.f_mntfromname);
406 			else
407 			    printf(" \"%*.*s\"\n",
408 				ncp->nc_nlen, ncp->nc_nlen,
409 				ncp->nc_name);
410 		}
411 	}
412 
413 	if (didwarn == 1) {
414 		printf("[diagnostic] cache_lock: unblocked %*.*s\n",
415 			ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
416 	}
417 }
418 
419 int
420 cache_lock_nonblock(struct namecache *ncp)
421 {
422 	thread_t td;
423 
424 	KKASSERT(ncp->nc_refs != 0);
425 	td = curthread;
426 	if (ncp->nc_exlocks == 0) {
427 		ncp->nc_exlocks = 1;
428 		ncp->nc_locktd = td;
429 		/*
430 		 * The vp associated with a locked ncp must be held
431 		 * to prevent it from being recycled (which would
432 		 * cause the ncp to become unresolved).
433 		 *
434 		 * WARNING!  If VRECLAIMED is set the vnode could
435 		 * already be in the middle of a recycle.  Callers
436 		 * should not assume that nc_vp is usable when
437 		 * not NULL.  cache_vref() or cache_vget() must be
438 		 * called.
439 		 *
440 		 * XXX loop on race for later MPSAFE work.
441 		 */
442 		if (ncp->nc_vp)
443 			vhold(ncp->nc_vp);
444 		return(0);
445 	} else {
446 		return(EWOULDBLOCK);
447 	}
448 }
449 
450 void
451 cache_unlock(struct namecache *ncp)
452 {
453 	thread_t td = curthread;
454 
455 	KKASSERT(ncp->nc_refs > 0);
456 	KKASSERT(ncp->nc_exlocks > 0);
457 	KKASSERT(ncp->nc_locktd == td);
458 	if (--ncp->nc_exlocks == 0) {
459 		if (ncp->nc_vp)
460 			vdrop(ncp->nc_vp);
461 		ncp->nc_locktd = NULL;
462 		if (ncp->nc_flag & NCF_LOCKREQ) {
463 			ncp->nc_flag &= ~NCF_LOCKREQ;
464 			wakeup(ncp);
465 		}
466 	}
467 }
468 
469 /*
470  * ref-and-lock, unlock-and-deref functions.
471  *
472  * This function is primarily used by nlookup.  Even though cache_lock
473  * holds the vnode, it is possible that the vnode may have already
474  * initiated a recyclement.  We want cache_get() to return a definitively
475  * usable vnode or a definitively unresolved ncp.
476  */
477 struct namecache *
478 cache_get(struct namecache *ncp)
479 {
480 	_cache_hold(ncp);
481 	cache_lock(ncp);
482 	if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
483 		cache_setunresolved(ncp);
484 	return(ncp);
485 }
486 
487 int
488 cache_get_nonblock(struct namecache *ncp)
489 {
490 	/* XXX MP */
491 	if (ncp->nc_exlocks == 0 || ncp->nc_locktd == curthread) {
492 		_cache_hold(ncp);
493 		cache_lock(ncp);
494 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
495 			cache_setunresolved(ncp);
496 		return(0);
497 	}
498 	return(EWOULDBLOCK);
499 }
500 
501 void
502 cache_put(struct namecache *ncp)
503 {
504 	cache_unlock(ncp);
505 	_cache_drop(ncp);
506 }
507 
508 /*
509  * Resolve an unresolved ncp by associating a vnode with it.  If the
510  * vnode is NULL, a negative cache entry is created.
511  *
512  * The ncp should be locked on entry and will remain locked on return.
513  */
514 void
515 cache_setvp(struct namecache *ncp, struct vnode *vp)
516 {
517 	KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
518 	ncp->nc_vp = vp;
519 	if (vp != NULL) {
520 		/*
521 		 * Any vp associated with an ncp which has children must
522 		 * be held.  Any vp associated with a locked ncp must be held.
523 		 */
524 		if (!TAILQ_EMPTY(&ncp->nc_list))
525 			vhold(vp);
526 		TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode);
527 		if (ncp->nc_exlocks)
528 			vhold(vp);
529 
530 		/*
531 		 * Set auxillary flags
532 		 */
533 		switch(vp->v_type) {
534 		case VDIR:
535 			ncp->nc_flag |= NCF_ISDIR;
536 			break;
537 		case VLNK:
538 			ncp->nc_flag |= NCF_ISSYMLINK;
539 			/* XXX cache the contents of the symlink */
540 			break;
541 		default:
542 			break;
543 		}
544 		++numcache;
545 		ncp->nc_error = 0;
546 	} else {
547 		TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
548 		++numneg;
549 		ncp->nc_error = ENOENT;
550 	}
551 	ncp->nc_flag &= ~NCF_UNRESOLVED;
552 }
553 
554 void
555 cache_settimeout(struct namecache *ncp, int nticks)
556 {
557 	if ((ncp->nc_timeout = ticks + nticks) == 0)
558 		ncp->nc_timeout = 1;
559 }
560 
561 /*
562  * Disassociate the vnode or negative-cache association and mark a
563  * namecache entry as unresolved again.  Note that the ncp is still
564  * left in the hash table and still linked to its parent.
565  *
566  * The ncp should be locked and refd on entry and will remain locked and refd
567  * on return.
568  *
569  * This routine is normally never called on a directory containing children.
570  * However, NFS often does just that in its rename() code as a cop-out to
571  * avoid complex namespace operations.  This disconnects a directory vnode
572  * from its namecache and can cause the OLDAPI and NEWAPI to get out of
573  * sync.
574  *
575  * NOTE: NCF_FSMID must be cleared so a refurbishment of the ncp, such as
576  * in a create, properly propogates flag up the chain.
577  */
578 void
579 cache_setunresolved(struct namecache *ncp)
580 {
581 	struct vnode *vp;
582 
583 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
584 		ncp->nc_flag |= NCF_UNRESOLVED;
585 		ncp->nc_timeout = 0;
586 		ncp->nc_error = ENOTCONN;
587 		++numunres;
588 		if ((vp = ncp->nc_vp) != NULL) {
589 			--numcache;
590 			ncp->nc_vp = NULL;
591 			TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode);
592 
593 			/*
594 			 * Any vp associated with an ncp with children is
595 			 * held by that ncp.  Any vp associated with a locked
596 			 * ncp is held by that ncp.  These conditions must be
597 			 * undone when the vp is cleared out from the ncp.
598 			 */
599 			if (ncp->nc_flag & NCF_FSMID)
600 				vupdatefsmid(vp);
601 			if (!TAILQ_EMPTY(&ncp->nc_list))
602 				vdrop(vp);
603 			if (ncp->nc_exlocks)
604 				vdrop(vp);
605 		} else {
606 			TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
607 			--numneg;
608 		}
609 		ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK|
610 				  NCF_FSMID);
611 	}
612 }
613 
614 /*
615  * Invalidate portions of the namecache topology given a starting entry.
616  * The passed ncp is set to an unresolved state and:
617  *
618  * The passed ncp must be locked.
619  *
620  * CINV_DESTROY		- Set a flag in the passed ncp entry indicating
621  *			  that the physical underlying nodes have been
622  *			  destroyed... as in deleted.  For example, when
623  *			  a directory is removed.  This will cause record
624  *			  lookups on the name to no longer be able to find
625  *			  the record and tells the resolver to return failure
626  *			  rather then trying to resolve through the parent.
627  *
628  *			  The topology itself, including ncp->nc_name,
629  *			  remains intact.
630  *
631  *			  This only applies to the passed ncp, if CINV_CHILDREN
632  *			  is specified the children are not flagged.
633  *
634  * CINV_CHILDREN	- Set all children (recursively) to an unresolved
635  *			  state as well.
636  *
637  *			  Note that this will also have the side effect of
638  *			  cleaning out any unreferenced nodes in the topology
639  *			  from the leaves up as the recursion backs out.
640  *
641  * Note that the topology for any referenced nodes remains intact.
642  *
643  * It is possible for cache_inval() to race a cache_resolve(), meaning that
644  * the namecache entry may not actually be invalidated on return if it was
645  * revalidated while recursing down into its children.  This code guarentees
646  * that the node(s) will go through an invalidation cycle, but does not
647  * guarentee that they will remain in an invalidated state.
648  *
649  * Returns non-zero if a revalidation was detected during the invalidation
650  * recursion, zero otherwise.  Note that since only the original ncp is
651  * locked the revalidation ultimately can only indicate that the original ncp
652  * *MIGHT* no have been reresolved.
653  */
654 int
655 cache_inval(struct namecache *ncp, int flags)
656 {
657 	struct namecache *kid;
658 	struct namecache *nextkid;
659 	int rcnt = 0;
660 
661 	KKASSERT(ncp->nc_exlocks);
662 
663 	cache_setunresolved(ncp);
664 	if (flags & CINV_DESTROY)
665 		ncp->nc_flag |= NCF_DESTROYED;
666 
667 	if ((flags & CINV_CHILDREN) &&
668 	    (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL
669 	) {
670 		cache_hold(kid);
671 		cache_unlock(ncp);
672 		while (kid) {
673 			if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL)
674 				cache_hold(nextkid);
675 			if ((kid->nc_flag & NCF_UNRESOLVED) == 0 ||
676 			    TAILQ_FIRST(&kid->nc_list)
677 			) {
678 				cache_lock(kid);
679 				rcnt += cache_inval(kid, flags & ~CINV_DESTROY);
680 				cache_unlock(kid);
681 			}
682 			cache_drop(kid);
683 			kid = nextkid;
684 		}
685 		cache_lock(ncp);
686 	}
687 
688 	/*
689 	 * Someone could have gotten in there while ncp was unlocked,
690 	 * retry if so.
691 	 */
692 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
693 		++rcnt;
694 	return (rcnt);
695 }
696 
697 /*
698  * Invalidate a vnode's namecache associations.  To avoid races against
699  * the resolver we do not invalidate a node which we previously invalidated
700  * but which was then re-resolved while we were in the invalidation loop.
701  *
702  * Returns non-zero if any namecache entries remain after the invalidation
703  * loop completed.
704  *
705  * NOTE: unlike the namecache topology which guarentees that ncp's will not
706  * be ripped out of the topology while held, the vnode's v_namecache list
707  * has no such restriction.  NCP's can be ripped out of the list at virtually
708  * any time if not locked, even if held.
709  */
710 int
711 cache_inval_vp(struct vnode *vp, int flags)
712 {
713 	struct namecache *ncp;
714 	struct namecache *next;
715 
716 restart:
717 	ncp = TAILQ_FIRST(&vp->v_namecache);
718 	if (ncp)
719 		cache_hold(ncp);
720 	while (ncp) {
721 		/* loop entered with ncp held */
722 		if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
723 			cache_hold(next);
724 		cache_lock(ncp);
725 		if (ncp->nc_vp != vp) {
726 			printf("Warning: cache_inval_vp: race-A detected on "
727 				"%s\n", ncp->nc_name);
728 			cache_put(ncp);
729 			if (next)
730 				cache_drop(next);
731 			goto restart;
732 		}
733 		cache_inval(ncp, flags);
734 		cache_put(ncp);		/* also releases reference */
735 		ncp = next;
736 		if (ncp && ncp->nc_vp != vp) {
737 			printf("Warning: cache_inval_vp: race-B detected on "
738 				"%s\n", ncp->nc_name);
739 			cache_drop(ncp);
740 			goto restart;
741 		}
742 	}
743 	return(TAILQ_FIRST(&vp->v_namecache) != NULL);
744 }
745 
746 /*
747  * The source ncp has been renamed to the target ncp.  Both fncp and tncp
748  * must be locked.  Both will be set to unresolved, any children of tncp
749  * will be disconnected (the prior contents of the target is assumed to be
750  * destroyed by the rename operation, e.g. renaming over an empty directory),
751  * and all children of fncp will be moved to tncp.
752  *
753  * XXX the disconnection could pose a problem, check code paths to make
754  * sure any code that blocks can handle the parent being changed out from
755  * under it.  Maybe we should lock the children (watch out for deadlocks) ?
756  *
757  * After we return the caller has the option of calling cache_setvp() if
758  * the vnode of the new target ncp is known.
759  *
760  * Any process CD'd into any of the children will no longer be able to ".."
761  * back out.  An rm -rf can cause this situation to occur.
762  */
763 void
764 cache_rename(struct namecache *fncp, struct namecache *tncp)
765 {
766 	struct namecache *scan;
767 	int didwarn = 0;
768 
769 	cache_setunresolved(fncp);
770 	cache_setunresolved(tncp);
771 	while (cache_inval(tncp, CINV_CHILDREN) != 0) {
772 		if (didwarn++ % 10 == 0) {
773 			printf("Warning: cache_rename: race during "
774 				"rename %s->%s\n",
775 				fncp->nc_name, tncp->nc_name);
776 		}
777 		tsleep(tncp, 0, "mvrace", hz / 10);
778 		cache_setunresolved(tncp);
779 	}
780 	while ((scan = TAILQ_FIRST(&fncp->nc_list)) != NULL) {
781 		cache_hold(scan);
782 		cache_unlink_parent(scan);
783 		cache_link_parent(scan, tncp);
784 		if (scan->nc_flag & NCF_HASHED)
785 			cache_rehash(scan);
786 		cache_drop(scan);
787 	}
788 }
789 
790 /*
791  * vget the vnode associated with the namecache entry.  Resolve the namecache
792  * entry if necessary and deal with namecache/vp races.  The passed ncp must
793  * be referenced and may be locked.  The ncp's ref/locking state is not
794  * effected by this call.
795  *
796  * lk_type may be LK_SHARED, LK_EXCLUSIVE.  A ref'd, possibly locked
797  * (depending on the passed lk_type) will be returned in *vpp with an error
798  * of 0, or NULL will be returned in *vpp with a non-0 error code.  The
799  * most typical error is ENOENT, meaning that the ncp represents a negative
800  * cache hit and there is no vnode to retrieve, but other errors can occur
801  * too.
802  *
803  * The main race we have to deal with are namecache zaps.  The ncp itself
804  * will not disappear since it is referenced, and it turns out that the
805  * validity of the vp pointer can be checked simply by rechecking the
806  * contents of ncp->nc_vp.
807  */
808 int
809 cache_vget(struct namecache *ncp, struct ucred *cred,
810 	   int lk_type, struct vnode **vpp)
811 {
812 	struct vnode *vp;
813 	int error;
814 
815 again:
816 	vp = NULL;
817 	if (ncp->nc_flag & NCF_UNRESOLVED) {
818 		cache_lock(ncp);
819 		error = cache_resolve(ncp, cred);
820 		cache_unlock(ncp);
821 	} else {
822 		error = 0;
823 	}
824 	if (error == 0 && (vp = ncp->nc_vp) != NULL) {
825 		/*
826 		 * Accessing the vnode from the namecache is a bit
827 		 * dangerous.  Because there are no refs on the vnode, it
828 		 * could be in the middle of a reclaim.
829 		 */
830 		if (vp->v_flag & VRECLAIMED) {
831 			printf("Warning: vnode reclaim race detected in cache_vget on %p (%s)\n", vp, ncp->nc_name);
832 			cache_lock(ncp);
833 			cache_setunresolved(ncp);
834 			cache_unlock(ncp);
835 			goto again;
836 		}
837 		error = vget(vp, lk_type);
838 		if (error) {
839 			if (vp != ncp->nc_vp)
840 				goto again;
841 			vp = NULL;
842 		} else if (vp != ncp->nc_vp) {
843 			vput(vp);
844 			goto again;
845 		} else if (vp->v_flag & VRECLAIMED) {
846 			panic("vget succeeded on a VRECLAIMED node! vp %p", vp);
847 		}
848 	}
849 	if (error == 0 && vp == NULL)
850 		error = ENOENT;
851 	*vpp = vp;
852 	return(error);
853 }
854 
855 int
856 cache_vref(struct namecache *ncp, struct ucred *cred, struct vnode **vpp)
857 {
858 	struct vnode *vp;
859 	int error;
860 
861 again:
862 	vp = NULL;
863 	if (ncp->nc_flag & NCF_UNRESOLVED) {
864 		cache_lock(ncp);
865 		error = cache_resolve(ncp, cred);
866 		cache_unlock(ncp);
867 	} else {
868 		error = 0;
869 	}
870 	if (error == 0 && (vp = ncp->nc_vp) != NULL) {
871 		/*
872 		 * Since we did not obtain any locks, a cache zap
873 		 * race can occur here if the vnode is in the middle
874 		 * of being reclaimed and has not yet been able to
875 		 * clean out its cache node.  If that case occurs,
876 		 * we must lock and unresolve the cache, then loop
877 		 * to retry.
878 		 */
879 		if (vp->v_flag & VRECLAIMED) {
880 			printf("Warning: vnode reclaim race detected on cache_vref %p (%s)\n", vp, ncp->nc_name);
881 			cache_lock(ncp);
882 			cache_setunresolved(ncp);
883 			cache_unlock(ncp);
884 			goto again;
885 		}
886 		vref(vp);
887 	}
888 	if (error == 0 && vp == NULL)
889 		error = ENOENT;
890 	*vpp = vp;
891 	return(error);
892 }
893 
894 /*
895  * Recursively set the FSMID update flag for namecache nodes leading
896  * to root.  This will cause the next getattr or reclaim to increment the
897  * fsmid and mark the inode for lazy updating.
898  *
899  * Stop recursing when we hit a node whos NCF_FSMID flag is already set.
900  * This makes FSMIDs work in an Einsteinian fashion - where the observation
901  * effects the result.  In this case a program monitoring a higher level
902  * node will have detected some prior change and started its scan (clearing
903  * NCF_FSMID in higher level nodes), but since it has not yet observed the
904  * node where we find NCF_FSMID still set, we can safely make the related
905  * modification without interfering with the theorized program.
906  *
907  * This also means that FSMIDs cannot represent time-domain quantities
908  * in a hierarchical sense.  But the main reason for doing it this way
909  * is to reduce the amount of recursion that occurs in the critical path
910  * when e.g. a program is writing to a file that sits deep in a directory
911  * hierarchy.
912  */
913 void
914 cache_update_fsmid(struct namecache *ncp)
915 {
916 	struct vnode *vp;
917 	struct namecache *scan;
918 
919 	/*
920 	 * Warning: even if we get a non-NULL vp it could still be in the
921 	 * middle of a recyclement.  Don't do anything fancy, just set
922 	 * NCF_FSMID.
923 	 */
924 	if ((vp = ncp->nc_vp) != NULL) {
925 		TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
926 			for (scan = ncp; scan; scan = scan->nc_parent) {
927 				if (scan->nc_flag & NCF_FSMID)
928 					break;
929 				scan->nc_flag |= NCF_FSMID;
930 			}
931 		}
932 	} else {
933 		while (ncp && (ncp->nc_flag & NCF_FSMID) == 0) {
934 			ncp->nc_flag |= NCF_FSMID;
935 			ncp = ncp->nc_parent;
936 		}
937 	}
938 }
939 
940 void
941 cache_update_fsmid_vp(struct vnode *vp)
942 {
943 	struct namecache *ncp;
944 	struct namecache *scan;
945 
946 	TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
947 		for (scan = ncp; scan; scan = scan->nc_parent) {
948 			if (scan->nc_flag & NCF_FSMID)
949 				break;
950 			scan->nc_flag |= NCF_FSMID;
951 		}
952 	}
953 }
954 
955 /*
956  * If getattr is called on a vnode (e.g. a stat call), the filesystem
957  * may call this routine to determine if the namecache has the hierarchical
958  * change flag set, requiring the fsmid to be updated.
959  *
960  * Since 0 indicates no support, make sure the filesystem fsmid is at least
961  * 1.
962  */
963 int
964 cache_check_fsmid_vp(struct vnode *vp, int64_t *fsmid)
965 {
966 	struct namecache *ncp;
967 	int changed = 0;
968 
969 	TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
970 		if (ncp->nc_flag & NCF_FSMID) {
971 			ncp->nc_flag &= ~NCF_FSMID;
972 			changed = 1;
973 		}
974 	}
975 	if (*fsmid == 0)
976 		++*fsmid;
977 	if (changed)
978 		++*fsmid;
979 	return(changed);
980 }
981 
982 /*
983  * Obtain the FSMID for a vnode for filesystems which do not support
984  * a built-in FSMID.
985  */
986 int64_t
987 cache_sync_fsmid_vp(struct vnode *vp)
988 {
989 	struct namecache *ncp;
990 
991 	if ((ncp = TAILQ_FIRST(&vp->v_namecache)) != NULL) {
992 		if (ncp->nc_flag & NCF_FSMID) {
993 			ncp->nc_flag &= ~NCF_FSMID;
994 			++ncp->nc_fsmid;
995 		}
996 		return(ncp->nc_fsmid);
997 	}
998 	return(VNOVAL);
999 }
1000 
1001 /*
1002  * Convert a directory vnode to a namecache record without any other
1003  * knowledge of the topology.  This ONLY works with directory vnodes and
1004  * is ONLY used by the NFS server.  dvp must be refd but unlocked, and the
1005  * returned ncp (if not NULL) will be held and unlocked.
1006  *
1007  * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned.
1008  * If 'makeit' is 1 we attempt to track-down and create the namecache topology
1009  * for dvp.  This will fail only if the directory has been deleted out from
1010  * under the caller.
1011  *
1012  * Callers must always check for a NULL return no matter the value of 'makeit'.
1013  *
1014  * To avoid underflowing the kernel stack each recursive call increments
1015  * the makeit variable.
1016  */
1017 
1018 static int cache_inefficient_scan(struct namecache *ncp, struct ucred *cred,
1019 				  struct vnode *dvp);
1020 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1021 				  struct vnode **saved_dvp);
1022 
1023 struct namecache *
1024 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit)
1025 {
1026 	struct namecache *ncp;
1027 	struct vnode *saved_dvp;
1028 	struct vnode *pvp;
1029 	int error;
1030 
1031 	ncp = NULL;
1032 	saved_dvp = NULL;
1033 
1034 	/*
1035 	 * Temporary debugging code to force the directory scanning code
1036 	 * to be exercised.
1037 	 */
1038 	if (ncvp_debug >= 3 && makeit && TAILQ_FIRST(&dvp->v_namecache)) {
1039 		ncp = TAILQ_FIRST(&dvp->v_namecache);
1040 		printf("cache_fromdvp: forcing %s\n", ncp->nc_name);
1041 		goto force;
1042 	}
1043 
1044 	/*
1045 	 * Loop until resolution, inside code will break out on error.
1046 	 */
1047 	while ((ncp = TAILQ_FIRST(&dvp->v_namecache)) == NULL && makeit) {
1048 force:
1049 		/*
1050 		 * If dvp is the root of its filesystem it should already
1051 		 * have a namecache pointer associated with it as a side
1052 		 * effect of the mount, but it may have been disassociated.
1053 		 */
1054 		if (dvp->v_flag & VROOT) {
1055 			ncp = cache_get(dvp->v_mount->mnt_ncp);
1056 			error = cache_resolve_mp(ncp);
1057 			cache_put(ncp);
1058 			if (ncvp_debug) {
1059 				printf("cache_fromdvp: resolve root of mount %p error %d",
1060 					dvp->v_mount, error);
1061 			}
1062 			if (error) {
1063 				if (ncvp_debug)
1064 					printf(" failed\n");
1065 				ncp = NULL;
1066 				break;
1067 			}
1068 			if (ncvp_debug)
1069 				printf(" succeeded\n");
1070 			continue;
1071 		}
1072 
1073 		/*
1074 		 * If we are recursed too deeply resort to an O(n^2)
1075 		 * algorithm to resolve the namecache topology.  The
1076 		 * resolved pvp is left referenced in saved_dvp to
1077 		 * prevent the tree from being destroyed while we loop.
1078 		 */
1079 		if (makeit > 20) {
1080 			error = cache_fromdvp_try(dvp, cred, &saved_dvp);
1081 			if (error) {
1082 				printf("lookupdotdot(longpath) failed %d "
1083 				       "dvp %p\n", error, dvp);
1084 				break;
1085 			}
1086 			continue;
1087 		}
1088 
1089 		/*
1090 		 * Get the parent directory and resolve its ncp.
1091 		 */
1092 		error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred);
1093 		if (error) {
1094 			printf("lookupdotdot failed %d dvp %p\n", error, dvp);
1095 			break;
1096 		}
1097 		VOP_UNLOCK(pvp, 0);
1098 
1099 		/*
1100 		 * Reuse makeit as a recursion depth counter.
1101 		 */
1102 		ncp = cache_fromdvp(pvp, cred, makeit + 1);
1103 		vrele(pvp);
1104 		if (ncp == NULL)
1105 			break;
1106 
1107 		/*
1108 		 * Do an inefficient scan of pvp (embodied by ncp) to look
1109 		 * for dvp.  This will create a namecache record for dvp on
1110 		 * success.  We loop up to recheck on success.
1111 		 *
1112 		 * ncp and dvp are both held but not locked.
1113 		 */
1114 		error = cache_inefficient_scan(ncp, cred, dvp);
1115 		cache_drop(ncp);
1116 		if (error) {
1117 			printf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n",
1118 				pvp, ncp->nc_name, dvp);
1119 			ncp = NULL;
1120 			break;
1121 		}
1122 		if (ncvp_debug) {
1123 			printf("cache_fromdvp: scan %p (%s) succeeded\n",
1124 				pvp, ncp->nc_name);
1125 		}
1126 	}
1127 	if (ncp)
1128 		cache_hold(ncp);
1129 	if (saved_dvp)
1130 		vrele(saved_dvp);
1131 	return (ncp);
1132 }
1133 
1134 /*
1135  * Go up the chain of parent directories until we find something
1136  * we can resolve into the namecache.  This is very inefficient.
1137  */
1138 static
1139 int
1140 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1141 		  struct vnode **saved_dvp)
1142 {
1143 	struct namecache *ncp;
1144 	struct vnode *pvp;
1145 	int error;
1146 	static time_t last_fromdvp_report;
1147 
1148 	/*
1149 	 * Loop getting the parent directory vnode until we get something we
1150 	 * can resolve in the namecache.
1151 	 */
1152 	vref(dvp);
1153 	for (;;) {
1154 		error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred);
1155 		if (error) {
1156 			vrele(dvp);
1157 			return (error);
1158 		}
1159 		VOP_UNLOCK(pvp, 0);
1160 		if ((ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) {
1161 			cache_hold(ncp);
1162 			vrele(pvp);
1163 			break;
1164 		}
1165 		if (pvp->v_flag & VROOT) {
1166 			ncp = cache_get(pvp->v_mount->mnt_ncp);
1167 			error = cache_resolve_mp(ncp);
1168 			cache_unlock(ncp);
1169 			vrele(pvp);
1170 			if (error) {
1171 				cache_drop(ncp);
1172 				vrele(dvp);
1173 				return (error);
1174 			}
1175 			break;
1176 		}
1177 		vrele(dvp);
1178 		dvp = pvp;
1179 	}
1180 	if (last_fromdvp_report != time_second) {
1181 		last_fromdvp_report = time_second;
1182 		printf("Warning: extremely inefficient path resolution on %s\n",
1183 			ncp->nc_name);
1184 	}
1185 	error = cache_inefficient_scan(ncp, cred, dvp);
1186 
1187 	/*
1188 	 * Hopefully dvp now has a namecache record associated with it.
1189 	 * Leave it referenced to prevent the kernel from recycling the
1190 	 * vnode.  Otherwise extremely long directory paths could result
1191 	 * in endless recycling.
1192 	 */
1193 	if (*saved_dvp)
1194 	    vrele(*saved_dvp);
1195 	*saved_dvp = dvp;
1196 	return (error);
1197 }
1198 
1199 
1200 /*
1201  * Do an inefficient scan of the directory represented by ncp looking for
1202  * the directory vnode dvp.  ncp must be held but not locked on entry and
1203  * will be held on return.  dvp must be refd but not locked on entry and
1204  * will remain refd on return.
1205  *
1206  * Why do this at all?  Well, due to its stateless nature the NFS server
1207  * converts file handles directly to vnodes without necessarily going through
1208  * the namecache ops that would otherwise create the namecache topology
1209  * leading to the vnode.  We could either (1) Change the namecache algorithms
1210  * to allow disconnect namecache records that are re-merged opportunistically,
1211  * or (2) Make the NFS server backtrack and scan to recover a connected
1212  * namecache topology in order to then be able to issue new API lookups.
1213  *
1214  * It turns out that (1) is a huge mess.  It takes a nice clean set of
1215  * namecache algorithms and introduces a lot of complication in every subsystem
1216  * that calls into the namecache to deal with the re-merge case, especially
1217  * since we are using the namecache to placehold negative lookups and the
1218  * vnode might not be immediately assigned. (2) is certainly far less
1219  * efficient then (1), but since we are only talking about directories here
1220  * (which are likely to remain cached), the case does not actually run all
1221  * that often and has the supreme advantage of not polluting the namecache
1222  * algorithms.
1223  */
1224 static int
1225 cache_inefficient_scan(struct namecache *ncp, struct ucred *cred,
1226 		       struct vnode *dvp)
1227 {
1228 	struct nlcomponent nlc;
1229 	struct namecache *rncp;
1230 	struct dirent *den;
1231 	struct vnode *pvp;
1232 	struct vattr vat;
1233 	struct iovec iov;
1234 	struct uio uio;
1235 	int blksize;
1236 	int eofflag;
1237 	int bytes;
1238 	char *rbuf;
1239 	int error;
1240 
1241 	vat.va_blocksize = 0;
1242 	if ((error = VOP_GETATTR(dvp, &vat)) != 0)
1243 		return (error);
1244 	if ((error = cache_vget(ncp, cred, LK_SHARED, &pvp)) != 0)
1245 		return (error);
1246 	if (ncvp_debug)
1247 		printf("inefficient_scan: directory iosize %ld vattr fileid = %ld\n", vat.va_blocksize, (long)vat.va_fileid);
1248 	if ((blksize = vat.va_blocksize) == 0)
1249 		blksize = DEV_BSIZE;
1250 	rbuf = malloc(blksize, M_TEMP, M_WAITOK);
1251 	rncp = NULL;
1252 
1253 	eofflag = 0;
1254 	uio.uio_offset = 0;
1255 again:
1256 	iov.iov_base = rbuf;
1257 	iov.iov_len = blksize;
1258 	uio.uio_iov = &iov;
1259 	uio.uio_iovcnt = 1;
1260 	uio.uio_resid = blksize;
1261 	uio.uio_segflg = UIO_SYSSPACE;
1262 	uio.uio_rw = UIO_READ;
1263 	uio.uio_td = curthread;
1264 
1265 	if (ncvp_debug >= 2)
1266 		printf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset);
1267 	error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL);
1268 	if (error == 0) {
1269 		den = (struct dirent *)rbuf;
1270 		bytes = blksize - uio.uio_resid;
1271 
1272 		while (bytes > 0) {
1273 			if (ncvp_debug >= 2) {
1274 				printf("cache_inefficient_scan: %*.*s\n",
1275 					den->d_namlen, den->d_namlen,
1276 					den->d_name);
1277 			}
1278 			if (den->d_type != DT_WHT &&
1279 			    den->d_ino == vat.va_fileid) {
1280 				if (ncvp_debug) {
1281 					printf("cache_inefficient_scan: "
1282 					       "MATCHED inode %ld path %s/%*.*s\n",
1283 					       vat.va_fileid, ncp->nc_name,
1284 					       den->d_namlen, den->d_namlen,
1285 					       den->d_name);
1286 				}
1287 				nlc.nlc_nameptr = den->d_name;
1288 				nlc.nlc_namelen = den->d_namlen;
1289 				VOP_UNLOCK(pvp, 0);
1290 				rncp = cache_nlookup(ncp, &nlc);
1291 				KKASSERT(rncp != NULL);
1292 				break;
1293 			}
1294 			bytes -= _DIRENT_DIRSIZ(den);
1295 			den = _DIRENT_NEXT(den);
1296 		}
1297 		if (rncp == NULL && eofflag == 0 && uio.uio_resid != blksize)
1298 			goto again;
1299 	}
1300 	if (rncp) {
1301 		vrele(pvp);
1302 		if (rncp->nc_flag & NCF_UNRESOLVED) {
1303 			cache_setvp(rncp, dvp);
1304 			if (ncvp_debug >= 2) {
1305 				printf("cache_inefficient_scan: setvp %s/%s = %p\n",
1306 					ncp->nc_name, rncp->nc_name, dvp);
1307 			}
1308 		} else {
1309 			if (ncvp_debug >= 2) {
1310 				printf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n",
1311 					ncp->nc_name, rncp->nc_name, dvp,
1312 					rncp->nc_vp);
1313 			}
1314 		}
1315 		if (rncp->nc_vp == NULL)
1316 			error = rncp->nc_error;
1317 		cache_put(rncp);
1318 	} else {
1319 		printf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n",
1320 			dvp, ncp->nc_name);
1321 		vput(pvp);
1322 		error = ENOENT;
1323 	}
1324 	free(rbuf, M_TEMP);
1325 	return (error);
1326 }
1327 
1328 /*
1329  * Zap a namecache entry.  The ncp is unconditionally set to an unresolved
1330  * state, which disassociates it from its vnode or ncneglist.
1331  *
1332  * Then, if there are no additional references to the ncp and no children,
1333  * the ncp is removed from the topology and destroyed.  This function will
1334  * also run through the nc_parent chain and destroy parent ncps if possible.
1335  * As a side benefit, it turns out the only conditions that allow running
1336  * up the chain are also the conditions to ensure no deadlock will occur.
1337  *
1338  * References and/or children may exist if the ncp is in the middle of the
1339  * topology, preventing the ncp from being destroyed.
1340  *
1341  * This function must be called with the ncp held and locked and will unlock
1342  * and drop it during zapping.
1343  */
1344 static void
1345 cache_zap(struct namecache *ncp)
1346 {
1347 	struct namecache *par;
1348 
1349 	/*
1350 	 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED.
1351 	 */
1352 	cache_setunresolved(ncp);
1353 
1354 	/*
1355 	 * Try to scrap the entry and possibly tail-recurse on its parent.
1356 	 * We only scrap unref'd (other then our ref) unresolved entries,
1357 	 * we do not scrap 'live' entries.
1358 	 */
1359 	while (ncp->nc_flag & NCF_UNRESOLVED) {
1360 		/*
1361 		 * Someone other then us has a ref, stop.
1362 		 */
1363 		if (ncp->nc_refs > 1)
1364 			goto done;
1365 
1366 		/*
1367 		 * We have children, stop.
1368 		 */
1369 		if (!TAILQ_EMPTY(&ncp->nc_list))
1370 			goto done;
1371 
1372 		/*
1373 		 * Remove ncp from the topology: hash table and parent linkage.
1374 		 */
1375 		if (ncp->nc_flag & NCF_HASHED) {
1376 			ncp->nc_flag &= ~NCF_HASHED;
1377 			LIST_REMOVE(ncp, nc_hash);
1378 		}
1379 		if ((par = ncp->nc_parent) != NULL) {
1380 			par = cache_hold(par);
1381 			TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
1382 			ncp->nc_parent = NULL;
1383 			if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
1384 				vdrop(par->nc_vp);
1385 		}
1386 
1387 		/*
1388 		 * ncp should not have picked up any refs.  Physically
1389 		 * destroy the ncp.
1390 		 */
1391 		KKASSERT(ncp->nc_refs == 1);
1392 		--numunres;
1393 		/* cache_unlock(ncp) not required */
1394 		ncp->nc_refs = -1;	/* safety */
1395 		if (ncp->nc_name)
1396 			free(ncp->nc_name, M_VFSCACHE);
1397 		free(ncp, M_VFSCACHE);
1398 
1399 		/*
1400 		 * Loop on the parent (it may be NULL).  Only bother looping
1401 		 * if the parent has a single ref (ours), which also means
1402 		 * we can lock it trivially.
1403 		 */
1404 		ncp = par;
1405 		if (ncp == NULL)
1406 			return;
1407 		if (ncp->nc_refs != 1) {
1408 			cache_drop(ncp);
1409 			return;
1410 		}
1411 		KKASSERT(par->nc_exlocks == 0);
1412 		cache_lock(ncp);
1413 	}
1414 done:
1415 	cache_unlock(ncp);
1416 	atomic_subtract_int(&ncp->nc_refs, 1);
1417 }
1418 
1419 static enum { CHI_LOW, CHI_HIGH } cache_hysteresis_state = CHI_LOW;
1420 
1421 static __inline
1422 void
1423 cache_hysteresis(void)
1424 {
1425 	/*
1426 	 * Don't cache too many negative hits.  We use hysteresis to reduce
1427 	 * the impact on the critical path.
1428 	 */
1429 	switch(cache_hysteresis_state) {
1430 	case CHI_LOW:
1431 		if (numneg > MINNEG && numneg * ncnegfactor > numcache) {
1432 			cache_cleanneg(10);
1433 			cache_hysteresis_state = CHI_HIGH;
1434 		}
1435 		break;
1436 	case CHI_HIGH:
1437 		if (numneg > MINNEG * 9 / 10 &&
1438 		    numneg * ncnegfactor * 9 / 10 > numcache
1439 		) {
1440 			cache_cleanneg(10);
1441 		} else {
1442 			cache_hysteresis_state = CHI_LOW;
1443 		}
1444 		break;
1445 	}
1446 }
1447 
1448 /*
1449  * NEW NAMECACHE LOOKUP API
1450  *
1451  * Lookup an entry in the cache.  A locked, referenced, non-NULL
1452  * entry is *always* returned, even if the supplied component is illegal.
1453  * The resulting namecache entry should be returned to the system with
1454  * cache_put() or cache_unlock() + cache_drop().
1455  *
1456  * namecache locks are recursive but care must be taken to avoid lock order
1457  * reversals.
1458  *
1459  * Nobody else will be able to manipulate the associated namespace (e.g.
1460  * create, delete, rename, rename-target) until the caller unlocks the
1461  * entry.
1462  *
1463  * The returned entry will be in one of three states:  positive hit (non-null
1464  * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set).
1465  * Unresolved entries must be resolved through the filesystem to associate the
1466  * vnode and/or determine whether a positive or negative hit has occured.
1467  *
1468  * It is not necessary to lock a directory in order to lock namespace under
1469  * that directory.  In fact, it is explicitly not allowed to do that.  A
1470  * directory is typically only locked when being created, renamed, or
1471  * destroyed.
1472  *
1473  * The directory (par) may be unresolved, in which case any returned child
1474  * will likely also be marked unresolved.  Likely but not guarenteed.  Since
1475  * the filesystem lookup requires a resolved directory vnode the caller is
1476  * responsible for resolving the namecache chain top-down.  This API
1477  * specifically allows whole chains to be created in an unresolved state.
1478  */
1479 struct namecache *
1480 cache_nlookup(struct namecache *par, struct nlcomponent *nlc)
1481 {
1482 	struct namecache *ncp;
1483 	struct namecache *new_ncp;
1484 	struct nchashhead *nchpp;
1485 	u_int32_t hash;
1486 	globaldata_t gd;
1487 
1488 	numcalls++;
1489 	gd = mycpu;
1490 
1491 	/*
1492 	 * Try to locate an existing entry
1493 	 */
1494 	hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
1495 	hash = fnv_32_buf(&par, sizeof(par), hash);
1496 	new_ncp = NULL;
1497 restart:
1498 	LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1499 		numchecks++;
1500 
1501 		/*
1502 		 * Zap entries that have timed out.
1503 		 */
1504 		if (ncp->nc_timeout &&
1505 		    (int)(ncp->nc_timeout - ticks) < 0 &&
1506 		    (ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
1507 		    ncp->nc_exlocks == 0
1508 		) {
1509 			cache_zap(cache_get(ncp));
1510 			goto restart;
1511 		}
1512 
1513 		/*
1514 		 * Break out if we find a matching entry.  Note that
1515 		 * UNRESOLVED entries may match, but DESTROYED entries
1516 		 * do not.
1517 		 */
1518 		if (ncp->nc_parent == par &&
1519 		    ncp->nc_nlen == nlc->nlc_namelen &&
1520 		    bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
1521 		    (ncp->nc_flag & NCF_DESTROYED) == 0
1522 		) {
1523 			if (cache_get_nonblock(ncp) == 0) {
1524 				if (new_ncp)
1525 					cache_free(new_ncp);
1526 				goto found;
1527 			}
1528 			cache_get(ncp);
1529 			cache_put(ncp);
1530 			goto restart;
1531 		}
1532 	}
1533 
1534 	/*
1535 	 * We failed to locate an entry, create a new entry and add it to
1536 	 * the cache.  We have to relookup after possibly blocking in
1537 	 * malloc.
1538 	 */
1539 	if (new_ncp == NULL) {
1540 		new_ncp = cache_alloc(nlc->nlc_namelen);
1541 		goto restart;
1542 	}
1543 
1544 	ncp = new_ncp;
1545 
1546 	/*
1547 	 * Initialize as a new UNRESOLVED entry, lock (non-blocking),
1548 	 * and link to the parent.  The mount point is usually inherited
1549 	 * from the parent unless this is a special case such as a mount
1550 	 * point where nlc_namelen is 0.  The caller is responsible for
1551 	 * setting nc_mount in that case.  If nlc_namelen is 0 nc_name will
1552 	 * be NULL.
1553 	 */
1554 	if (nlc->nlc_namelen) {
1555 		bcopy(nlc->nlc_nameptr, ncp->nc_name, nlc->nlc_namelen);
1556 		ncp->nc_name[nlc->nlc_namelen] = 0;
1557 		ncp->nc_mount = par->nc_mount;
1558 	}
1559 	nchpp = NCHHASH(hash);
1560 	LIST_INSERT_HEAD(nchpp, ncp, nc_hash);
1561 	ncp->nc_flag |= NCF_HASHED;
1562 	cache_link_parent(ncp, par);
1563 found:
1564 	/*
1565 	 * stats and namecache size management
1566 	 */
1567 	if (ncp->nc_flag & NCF_UNRESOLVED)
1568 		++gd->gd_nchstats->ncs_miss;
1569 	else if (ncp->nc_vp)
1570 		++gd->gd_nchstats->ncs_goodhits;
1571 	else
1572 		++gd->gd_nchstats->ncs_neghits;
1573 	cache_hysteresis();
1574 	return(ncp);
1575 }
1576 
1577 /*
1578  * Given a locked ncp, validate that the vnode, if present, is actually
1579  * usable.  If it is not usable set the ncp to an unresolved state.
1580  */
1581 void
1582 cache_validate(struct namecache *ncp)
1583 {
1584 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1585 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1586 			cache_setunresolved(ncp);
1587 	}
1588 }
1589 
1590 /*
1591  * Resolve an unresolved namecache entry, generally by looking it up.
1592  * The passed ncp must be locked and refd.
1593  *
1594  * Theoretically since a vnode cannot be recycled while held, and since
1595  * the nc_parent chain holds its vnode as long as children exist, the
1596  * direct parent of the cache entry we are trying to resolve should
1597  * have a valid vnode.  If not then generate an error that we can
1598  * determine is related to a resolver bug.
1599  *
1600  * However, if a vnode was in the middle of a recyclement when the NCP
1601  * got locked, ncp->nc_vp might point to a vnode that is about to become
1602  * invalid.  cache_resolve() handles this case by unresolving the entry
1603  * and then re-resolving it.
1604  *
1605  * Note that successful resolution does not necessarily return an error
1606  * code of 0.  If the ncp resolves to a negative cache hit then ENOENT
1607  * will be returned.
1608  */
1609 int
1610 cache_resolve(struct namecache *ncp, struct ucred *cred)
1611 {
1612 	struct namecache *par;
1613 	int error;
1614 
1615 restart:
1616 	/*
1617 	 * If the ncp is already resolved we have nothing to do.  However,
1618 	 * we do want to guarentee that a usable vnode is returned when
1619 	 * a vnode is present, so make sure it hasn't been reclaimed.
1620 	 */
1621 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1622 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1623 			cache_setunresolved(ncp);
1624 		if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
1625 			return (ncp->nc_error);
1626 	}
1627 
1628 	/*
1629 	 * Mount points need special handling because the parent does not
1630 	 * belong to the same filesystem as the ncp.
1631 	 */
1632 	if (ncp->nc_flag & NCF_MOUNTPT)
1633 		return (cache_resolve_mp(ncp));
1634 
1635 	/*
1636 	 * We expect an unbroken chain of ncps to at least the mount point,
1637 	 * and even all the way to root (but this code doesn't have to go
1638 	 * past the mount point).
1639 	 */
1640 	if (ncp->nc_parent == NULL) {
1641 		printf("EXDEV case 1 %p %*.*s\n", ncp,
1642 			ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
1643 		ncp->nc_error = EXDEV;
1644 		return(ncp->nc_error);
1645 	}
1646 
1647 	/*
1648 	 * The vp's of the parent directories in the chain are held via vhold()
1649 	 * due to the existance of the child, and should not disappear.
1650 	 * However, there are cases where they can disappear:
1651 	 *
1652 	 *	- due to filesystem I/O errors.
1653 	 *	- due to NFS being stupid about tracking the namespace and
1654 	 *	  destroys the namespace for entire directories quite often.
1655 	 *	- due to forced unmounts.
1656 	 *	- due to an rmdir (parent will be marked DESTROYED)
1657 	 *
1658 	 * When this occurs we have to track the chain backwards and resolve
1659 	 * it, looping until the resolver catches up to the current node.  We
1660 	 * could recurse here but we might run ourselves out of kernel stack
1661 	 * so we do it in a more painful manner.  This situation really should
1662 	 * not occur all that often, or if it does not have to go back too
1663 	 * many nodes to resolve the ncp.
1664 	 */
1665 	while (ncp->nc_parent->nc_vp == NULL) {
1666 		/*
1667 		 * This case can occur if a process is CD'd into a
1668 		 * directory which is then rmdir'd.  If the parent is marked
1669 		 * destroyed there is no point trying to resolve it.
1670 		 */
1671 		if (ncp->nc_parent->nc_flag & NCF_DESTROYED)
1672 			return(ENOENT);
1673 
1674 		par = ncp->nc_parent;
1675 		while (par->nc_parent && par->nc_parent->nc_vp == NULL)
1676 			par = par->nc_parent;
1677 		if (par->nc_parent == NULL) {
1678 			printf("EXDEV case 2 %*.*s\n",
1679 				par->nc_nlen, par->nc_nlen, par->nc_name);
1680 			return (EXDEV);
1681 		}
1682 		printf("[diagnostic] cache_resolve: had to recurse on %*.*s\n",
1683 			par->nc_nlen, par->nc_nlen, par->nc_name);
1684 		/*
1685 		 * The parent is not set in stone, ref and lock it to prevent
1686 		 * it from disappearing.  Also note that due to renames it
1687 		 * is possible for our ncp to move and for par to no longer
1688 		 * be one of its parents.  We resolve it anyway, the loop
1689 		 * will handle any moves.
1690 		 */
1691 		cache_get(par);
1692 		if (par->nc_flag & NCF_MOUNTPT) {
1693 			cache_resolve_mp(par);
1694 		} else if (par->nc_parent->nc_vp == NULL) {
1695 			printf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name);
1696 			cache_put(par);
1697 			continue;
1698 		} else if (par->nc_flag & NCF_UNRESOLVED) {
1699 			par->nc_error = VOP_NRESOLVE(par, cred);
1700 		}
1701 		if ((error = par->nc_error) != 0) {
1702 			if (par->nc_error != EAGAIN) {
1703 				printf("EXDEV case 3 %*.*s error %d\n",
1704 				    par->nc_nlen, par->nc_nlen, par->nc_name,
1705 				    par->nc_error);
1706 				cache_put(par);
1707 				return(error);
1708 			}
1709 			printf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n",
1710 				par, par->nc_nlen, par->nc_nlen, par->nc_name);
1711 		}
1712 		cache_put(par);
1713 		/* loop */
1714 	}
1715 
1716 	/*
1717 	 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected
1718 	 * ncp's and reattach them.  If this occurs the original ncp is marked
1719 	 * EAGAIN to force a relookup.
1720 	 *
1721 	 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed
1722 	 * ncp must already be resolved.
1723 	 */
1724 	KKASSERT((ncp->nc_flag & NCF_MOUNTPT) == 0);
1725 	ncp->nc_error = VOP_NRESOLVE(ncp, cred);
1726 	/*vop_nresolve(*ncp->nc_parent->nc_vp->v_ops, ncp, cred);*/
1727 	if (ncp->nc_error == EAGAIN) {
1728 		printf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n",
1729 			ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
1730 		goto restart;
1731 	}
1732 	return(ncp->nc_error);
1733 }
1734 
1735 /*
1736  * Resolve the ncp associated with a mount point.  Such ncp's almost always
1737  * remain resolved and this routine is rarely called.  NFS MPs tends to force
1738  * re-resolution more often due to its mac-truck-smash-the-namecache
1739  * method of tracking namespace changes.
1740  *
1741  * The semantics for this call is that the passed ncp must be locked on
1742  * entry and will be locked on return.  However, if we actually have to
1743  * resolve the mount point we temporarily unlock the entry in order to
1744  * avoid race-to-root deadlocks due to e.g. dead NFS mounts.  Because of
1745  * the unlock we have to recheck the flags after we relock.
1746  */
1747 static int
1748 cache_resolve_mp(struct namecache *ncp)
1749 {
1750 	struct vnode *vp;
1751 	struct mount *mp = ncp->nc_mount;
1752 	int error;
1753 
1754 	KKASSERT(mp != NULL);
1755 
1756 	/*
1757 	 * If the ncp is already resolved we have nothing to do.  However,
1758 	 * we do want to guarentee that a usable vnode is returned when
1759 	 * a vnode is present, so make sure it hasn't been reclaimed.
1760 	 */
1761 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1762 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1763 			cache_setunresolved(ncp);
1764 	}
1765 
1766 	if (ncp->nc_flag & NCF_UNRESOLVED) {
1767 		cache_unlock(ncp);
1768 		while (vfs_busy(mp, 0))
1769 			;
1770 		error = VFS_ROOT(mp, &vp);
1771 		cache_lock(ncp);
1772 
1773 		/*
1774 		 * recheck the ncp state after relocking.
1775 		 */
1776 		if (ncp->nc_flag & NCF_UNRESOLVED) {
1777 			ncp->nc_error = error;
1778 			if (error == 0) {
1779 				cache_setvp(ncp, vp);
1780 				vput(vp);
1781 			} else {
1782 				printf("[diagnostic] cache_resolve_mp: failed to resolve mount %p\n", mp);
1783 				cache_setvp(ncp, NULL);
1784 			}
1785 		} else if (error == 0) {
1786 			vput(vp);
1787 		}
1788 		vfs_unbusy(mp);
1789 	}
1790 	return(ncp->nc_error);
1791 }
1792 
1793 void
1794 cache_cleanneg(int count)
1795 {
1796 	struct namecache *ncp;
1797 
1798 	/*
1799 	 * Automode from the vnlru proc - clean out 10% of the negative cache
1800 	 * entries.
1801 	 */
1802 	if (count == 0)
1803 		count = numneg / 10 + 1;
1804 
1805 	/*
1806 	 * Attempt to clean out the specified number of negative cache
1807 	 * entries.
1808 	 */
1809 	while (count) {
1810 		ncp = TAILQ_FIRST(&ncneglist);
1811 		if (ncp == NULL) {
1812 			KKASSERT(numneg == 0);
1813 			break;
1814 		}
1815 		TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
1816 		TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
1817 		if (cache_get_nonblock(ncp) == 0)
1818 			cache_zap(ncp);
1819 		--count;
1820 	}
1821 }
1822 
1823 /*
1824  * Rehash a ncp.  Rehashing is typically required if the name changes (should
1825  * not generally occur) or the parent link changes.  This function will
1826  * unhash the ncp if the ncp is no longer hashable.
1827  */
1828 static void
1829 cache_rehash(struct namecache *ncp)
1830 {
1831 	struct nchashhead *nchpp;
1832 	u_int32_t hash;
1833 
1834 	if (ncp->nc_flag & NCF_HASHED) {
1835 		ncp->nc_flag &= ~NCF_HASHED;
1836 		LIST_REMOVE(ncp, nc_hash);
1837 	}
1838 	if (ncp->nc_nlen && ncp->nc_parent) {
1839 		hash = fnv_32_buf(ncp->nc_name, ncp->nc_nlen, FNV1_32_INIT);
1840 		hash = fnv_32_buf(&ncp->nc_parent,
1841 					sizeof(ncp->nc_parent), hash);
1842 		nchpp = NCHHASH(hash);
1843 		LIST_INSERT_HEAD(nchpp, ncp, nc_hash);
1844 		ncp->nc_flag |= NCF_HASHED;
1845 	}
1846 }
1847 
1848 /*
1849  * Name cache initialization, from vfsinit() when we are booting
1850  */
1851 void
1852 nchinit(void)
1853 {
1854 	int i;
1855 	globaldata_t gd;
1856 
1857 	/* initialise per-cpu namecache effectiveness statistics. */
1858 	for (i = 0; i < ncpus; ++i) {
1859 		gd = globaldata_find(i);
1860 		gd->gd_nchstats = &nchstats[i];
1861 	}
1862 	TAILQ_INIT(&ncneglist);
1863 	nchashtbl = hashinit(desiredvnodes*2, M_VFSCACHE, &nchash);
1864 	nclockwarn = 1 * hz;
1865 }
1866 
1867 /*
1868  * Called from start_init() to bootstrap the root filesystem.  Returns
1869  * a referenced, unlocked namecache record.
1870  */
1871 struct namecache *
1872 cache_allocroot(struct mount *mp, struct vnode *vp)
1873 {
1874 	struct namecache *ncp = cache_alloc(0);
1875 
1876 	ncp->nc_flag |= NCF_MOUNTPT | NCF_ROOT;
1877 	ncp->nc_mount = mp;
1878 	cache_setvp(ncp, vp);
1879 	return(ncp);
1880 }
1881 
1882 /*
1883  * vfs_cache_setroot()
1884  *
1885  *	Create an association between the root of our namecache and
1886  *	the root vnode.  This routine may be called several times during
1887  *	booting.
1888  *
1889  *	If the caller intends to save the returned namecache pointer somewhere
1890  *	it must cache_hold() it.
1891  */
1892 void
1893 vfs_cache_setroot(struct vnode *nvp, struct namecache *ncp)
1894 {
1895 	struct vnode *ovp;
1896 	struct namecache *oncp;
1897 
1898 	ovp = rootvnode;
1899 	oncp = rootncp;
1900 	rootvnode = nvp;
1901 	rootncp = ncp;
1902 
1903 	if (ovp)
1904 		vrele(ovp);
1905 	if (oncp)
1906 		cache_drop(oncp);
1907 }
1908 
1909 /*
1910  * XXX OLD API COMPAT FUNCTION.  This really messes up the new namecache
1911  * topology and is being removed as quickly as possible.  The new VOP_N*()
1912  * API calls are required to make specific adjustments using the supplied
1913  * ncp pointers rather then just bogusly purging random vnodes.
1914  *
1915  * Invalidate all namecache entries to a particular vnode as well as
1916  * any direct children of that vnode in the namecache.  This is a
1917  * 'catch all' purge used by filesystems that do not know any better.
1918  *
1919  * Note that the linkage between the vnode and its namecache entries will
1920  * be removed, but the namecache entries themselves might stay put due to
1921  * active references from elsewhere in the system or due to the existance of
1922  * the children.   The namecache topology is left intact even if we do not
1923  * know what the vnode association is.  Such entries will be marked
1924  * NCF_UNRESOLVED.
1925  */
1926 void
1927 cache_purge(struct vnode *vp)
1928 {
1929 	cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN);
1930 }
1931 
1932 /*
1933  * Flush all entries referencing a particular filesystem.
1934  *
1935  * Since we need to check it anyway, we will flush all the invalid
1936  * entries at the same time.
1937  */
1938 void
1939 cache_purgevfs(struct mount *mp)
1940 {
1941 	struct nchashhead *nchpp;
1942 	struct namecache *ncp, *nnp;
1943 
1944 	/*
1945 	 * Scan hash tables for applicable entries.
1946 	 */
1947 	for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) {
1948 		ncp = LIST_FIRST(nchpp);
1949 		if (ncp)
1950 			cache_hold(ncp);
1951 		while (ncp) {
1952 			nnp = LIST_NEXT(ncp, nc_hash);
1953 			if (nnp)
1954 				cache_hold(nnp);
1955 			if (ncp->nc_mount == mp) {
1956 				cache_lock(ncp);
1957 				cache_zap(ncp);
1958 			} else {
1959 				cache_drop(ncp);
1960 			}
1961 			ncp = nnp;
1962 		}
1963 	}
1964 }
1965 
1966 /*
1967  * Create a new (theoretically) unique fsmid
1968  */
1969 int64_t
1970 cache_getnewfsmid(void)
1971 {
1972 	static int fsmid_roller;
1973 	int64_t fsmid;
1974 
1975 	++fsmid_roller;
1976 	fsmid = ((int64_t)time_second << 32) |
1977 			(fsmid_roller & 0x7FFFFFFF);
1978 	return (fsmid);
1979 }
1980 
1981 
1982 static int disablecwd;
1983 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, "");
1984 
1985 static u_long numcwdcalls; STATNODE(CTLFLAG_RD, numcwdcalls, &numcwdcalls);
1986 static u_long numcwdfail1; STATNODE(CTLFLAG_RD, numcwdfail1, &numcwdfail1);
1987 static u_long numcwdfail2; STATNODE(CTLFLAG_RD, numcwdfail2, &numcwdfail2);
1988 static u_long numcwdfail3; STATNODE(CTLFLAG_RD, numcwdfail3, &numcwdfail3);
1989 static u_long numcwdfail4; STATNODE(CTLFLAG_RD, numcwdfail4, &numcwdfail4);
1990 static u_long numcwdfound; STATNODE(CTLFLAG_RD, numcwdfound, &numcwdfound);
1991 
1992 int
1993 sys___getcwd(struct __getcwd_args *uap)
1994 {
1995 	int buflen;
1996 	int error;
1997 	char *buf;
1998 	char *bp;
1999 
2000 	if (disablecwd)
2001 		return (ENODEV);
2002 
2003 	buflen = uap->buflen;
2004 	if (buflen < 2)
2005 		return (EINVAL);
2006 	if (buflen > MAXPATHLEN)
2007 		buflen = MAXPATHLEN;
2008 
2009 	buf = malloc(buflen, M_TEMP, M_WAITOK);
2010 	bp = kern_getcwd(buf, buflen, &error);
2011 	if (error == 0)
2012 		error = copyout(bp, uap->buf, strlen(bp) + 1);
2013 	free(buf, M_TEMP);
2014 	return (error);
2015 }
2016 
2017 char *
2018 kern_getcwd(char *buf, size_t buflen, int *error)
2019 {
2020 	struct proc *p = curproc;
2021 	char *bp;
2022 	int i, slash_prefixed;
2023 	struct filedesc *fdp;
2024 	struct namecache *ncp;
2025 
2026 	numcwdcalls++;
2027 	bp = buf;
2028 	bp += buflen - 1;
2029 	*bp = '\0';
2030 	fdp = p->p_fd;
2031 	slash_prefixed = 0;
2032 
2033 	ncp = fdp->fd_ncdir;
2034 	while (ncp && ncp != fdp->fd_nrdir && (ncp->nc_flag & NCF_ROOT) == 0) {
2035 		if (ncp->nc_flag & NCF_MOUNTPT) {
2036 			if (ncp->nc_mount == NULL) {
2037 				*error = EBADF;		/* forced unmount? */
2038 				return(NULL);
2039 			}
2040 			ncp = ncp->nc_parent;
2041 			continue;
2042 		}
2043 		for (i = ncp->nc_nlen - 1; i >= 0; i--) {
2044 			if (bp == buf) {
2045 				numcwdfail4++;
2046 				*error = ENOMEM;
2047 				return(NULL);
2048 			}
2049 			*--bp = ncp->nc_name[i];
2050 		}
2051 		if (bp == buf) {
2052 			numcwdfail4++;
2053 			*error = ENOMEM;
2054 			return(NULL);
2055 		}
2056 		*--bp = '/';
2057 		slash_prefixed = 1;
2058 		ncp = ncp->nc_parent;
2059 	}
2060 	if (ncp == NULL) {
2061 		numcwdfail2++;
2062 		*error = ENOENT;
2063 		return(NULL);
2064 	}
2065 	if (!slash_prefixed) {
2066 		if (bp == buf) {
2067 			numcwdfail4++;
2068 			*error = ENOMEM;
2069 			return(NULL);
2070 		}
2071 		*--bp = '/';
2072 	}
2073 	numcwdfound++;
2074 	*error = 0;
2075 	return (bp);
2076 }
2077 
2078 /*
2079  * Thus begins the fullpath magic.
2080  */
2081 
2082 #undef STATNODE
2083 #define STATNODE(name)							\
2084 	static u_int name;						\
2085 	SYSCTL_UINT(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, "")
2086 
2087 static int disablefullpath;
2088 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW,
2089     &disablefullpath, 0, "");
2090 
2091 STATNODE(numfullpathcalls);
2092 STATNODE(numfullpathfail1);
2093 STATNODE(numfullpathfail2);
2094 STATNODE(numfullpathfail3);
2095 STATNODE(numfullpathfail4);
2096 STATNODE(numfullpathfound);
2097 
2098 int
2099 cache_fullpath(struct proc *p, struct namecache *ncp, char **retbuf, char **freebuf)
2100 {
2101 	char *bp, *buf;
2102 	int i, slash_prefixed;
2103 	struct namecache *fd_nrdir;
2104 
2105 	numfullpathcalls--;
2106 
2107 	buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
2108 	bp = buf + MAXPATHLEN - 1;
2109 	*bp = '\0';
2110 	if (p != NULL)
2111 		fd_nrdir = p->p_fd->fd_nrdir;
2112 	else
2113 		fd_nrdir = NULL;
2114 	slash_prefixed = 0;
2115 	while (ncp && ncp != fd_nrdir && (ncp->nc_flag & NCF_ROOT) == 0) {
2116 		if (ncp->nc_flag & NCF_MOUNTPT) {
2117 			if (ncp->nc_mount == NULL) {
2118 				free(buf, M_TEMP);
2119 				return(EBADF);
2120 			}
2121 			ncp = ncp->nc_parent;
2122 			continue;
2123 		}
2124 		for (i = ncp->nc_nlen - 1; i >= 0; i--) {
2125 			if (bp == buf) {
2126 				numfullpathfail4++;
2127 				free(buf, M_TEMP);
2128 				return(ENOMEM);
2129 			}
2130 			*--bp = ncp->nc_name[i];
2131 		}
2132 		if (bp == buf) {
2133 			numfullpathfail4++;
2134 			free(buf, M_TEMP);
2135 			return(ENOMEM);
2136 		}
2137 		*--bp = '/';
2138 		slash_prefixed = 1;
2139 		ncp = ncp->nc_parent;
2140 	}
2141 	if (ncp == NULL) {
2142 		numfullpathfail2++;
2143 		free(buf, M_TEMP);
2144 		return(ENOENT);
2145 	}
2146 	if (p != NULL && (ncp->nc_flag & NCF_ROOT) && ncp != fd_nrdir) {
2147 		bp = buf + MAXPATHLEN - 1;
2148 		*bp = '\0';
2149 		slash_prefixed = 0;
2150 	}
2151 	if (!slash_prefixed) {
2152 		if (bp == buf) {
2153 			numfullpathfail4++;
2154 			free(buf, M_TEMP);
2155 			return(ENOMEM);
2156 		}
2157 		*--bp = '/';
2158 	}
2159 	numfullpathfound++;
2160 	*retbuf = bp;
2161 	*freebuf = buf;
2162 
2163 	return(0);
2164 }
2165 
2166 int
2167 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf)
2168 {
2169 	struct namecache *ncp;
2170 
2171 	numfullpathcalls++;
2172 	if (disablefullpath)
2173 		return (ENODEV);
2174 
2175 	if (p == NULL)
2176 		return (EINVAL);
2177 
2178 	/* vn is NULL, client wants us to use p->p_textvp */
2179 	if (vn == NULL) {
2180 		if ((vn = p->p_textvp) == NULL)
2181 			return (EINVAL);
2182 	}
2183 	TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) {
2184 		if (ncp->nc_nlen)
2185 			break;
2186 	}
2187 	if (ncp == NULL)
2188 		return (EINVAL);
2189 
2190 	numfullpathcalls--;
2191 	return(cache_fullpath(p, ncp, retbuf, freebuf));
2192 }
2193