xref: /dflybsd-src/sys/kern/vfs_cache.c (revision 681e352e6cf7f30481ae49f284ab7dab4168ba5a)
1 /*
2  * Copyright (c) 2003,2004,2009 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 
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/kernel.h>
72 #include <sys/sysctl.h>
73 #include <sys/mount.h>
74 #include <sys/vnode.h>
75 #include <sys/malloc.h>
76 #include <sys/sysproto.h>
77 #include <sys/spinlock.h>
78 #include <sys/proc.h>
79 #include <sys/namei.h>
80 #include <sys/nlookup.h>
81 #include <sys/filedesc.h>
82 #include <sys/fnv_hash.h>
83 #include <sys/globaldata.h>
84 #include <sys/kern_syscall.h>
85 #include <sys/dirent.h>
86 #include <ddb/ddb.h>
87 
88 #include <sys/sysref2.h>
89 #include <sys/spinlock2.h>
90 #include <sys/mplock2.h>
91 
92 #define MAX_RECURSION_DEPTH	64
93 
94 /*
95  * Random lookups in the cache are accomplished with a hash table using
96  * a hash key of (nc_src_vp, name).  Each hash chain has its own spin lock.
97  *
98  * Negative entries may exist and correspond to resolved namecache
99  * structures where nc_vp is NULL.  In a negative entry, NCF_WHITEOUT
100  * will be set if the entry corresponds to a whited-out directory entry
101  * (verses simply not finding the entry at all).   ncneglist is locked
102  * with a global spinlock (ncspin).
103  *
104  * MPSAFE RULES:
105  *
106  * (1) A ncp must be referenced before it can be locked.
107  *
108  * (2) A ncp must be locked in order to modify it.
109  *
110  * (3) ncp locks are always ordered child -> parent.  That may seem
111  *     backwards but forward scans use the hash table and thus can hold
112  *     the parent unlocked when traversing downward.
113  *
114  *     This allows insert/rename/delete/dot-dot and other operations
115  *     to use ncp->nc_parent links.
116  *
117  *     This also prevents a locked up e.g. NFS node from creating a
118  *     chain reaction all the way back to the root vnode / namecache.
119  *
120  * (4) parent linkages require both the parent and child to be locked.
121  */
122 
123 /*
124  * Structures associated with name cacheing.
125  */
126 #define NCHHASH(hash)		(&nchashtbl[(hash) & nchash])
127 #define MINNEG			1024
128 #define MINPOS			1024
129 #define NCMOUNT_NUMCACHE	1009	/* prime number */
130 
131 MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
132 
133 LIST_HEAD(nchash_list, namecache);
134 
135 struct nchash_head {
136        struct nchash_list list;
137        struct spinlock	spin;
138 };
139 
140 struct ncmount_cache {
141 	struct spinlock	spin;
142 	struct namecache *ncp;
143 	struct mount *mp;
144 	int isneg;		/* if != 0 mp is originator and not target */
145 };
146 
147 static struct nchash_head	*nchashtbl;
148 static struct namecache_list	ncneglist;
149 static struct spinlock		ncspin;
150 static struct ncmount_cache	ncmount_cache[NCMOUNT_NUMCACHE];
151 
152 /*
153  * ncvp_debug - debug cache_fromvp().  This is used by the NFS server
154  * to create the namecache infrastructure leading to a dangling vnode.
155  *
156  * 0	Only errors are reported
157  * 1	Successes are reported
158  * 2	Successes + the whole directory scan is reported
159  * 3	Force the directory scan code run as if the parent vnode did not
160  *	have a namecache record, even if it does have one.
161  */
162 static int	ncvp_debug;
163 SYSCTL_INT(_debug, OID_AUTO, ncvp_debug, CTLFLAG_RW, &ncvp_debug, 0,
164     "Namecache debug level (0-3)");
165 
166 static u_long	nchash;			/* size of hash table */
167 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
168     "Size of namecache hash table");
169 
170 static int	ncnegfactor = 16;	/* ratio of negative entries */
171 SYSCTL_INT(_debug, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
172     "Ratio of namecache negative entries");
173 
174 static int	nclockwarn;		/* warn on locked entries in ticks */
175 SYSCTL_INT(_debug, OID_AUTO, nclockwarn, CTLFLAG_RW, &nclockwarn, 0,
176     "Warn on locked namecache entries in ticks");
177 
178 static int	numdefered;		/* number of cache entries allocated */
179 SYSCTL_INT(_debug, OID_AUTO, numdefered, CTLFLAG_RD, &numdefered, 0,
180     "Number of cache entries allocated");
181 
182 static int	ncposlimit;		/* number of cache entries allocated */
183 SYSCTL_INT(_debug, OID_AUTO, ncposlimit, CTLFLAG_RW, &ncposlimit, 0,
184     "Number of cache entries allocated");
185 
186 SYSCTL_INT(_debug, OID_AUTO, vnsize, CTLFLAG_RD, 0, sizeof(struct vnode),
187     "sizeof(struct vnode)");
188 SYSCTL_INT(_debug, OID_AUTO, ncsize, CTLFLAG_RD, 0, sizeof(struct namecache),
189     "sizeof(struct namecache)");
190 
191 static int	ncmount_cache_enable = 1;
192 SYSCTL_INT(_debug, OID_AUTO, ncmount_cache_enable, CTLFLAG_RW,
193 	   &ncmount_cache_enable, 0, "mount point cache");
194 static long	ncmount_cache_hit;
195 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_hit, CTLFLAG_RW,
196 	    &ncmount_cache_hit, 0, "mpcache hits");
197 static long	ncmount_cache_miss;
198 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_miss, CTLFLAG_RW,
199 	    &ncmount_cache_miss, 0, "mpcache misses");
200 static long	ncmount_cache_overwrite;
201 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_overwrite, CTLFLAG_RW,
202 	    &ncmount_cache_overwrite, 0, "mpcache entry overwrites");
203 
204 static int cache_resolve_mp(struct mount *mp);
205 static struct vnode *cache_dvpref(struct namecache *ncp);
206 static void _cache_lock(struct namecache *ncp);
207 static void _cache_setunresolved(struct namecache *ncp);
208 static void _cache_cleanneg(int count);
209 static void _cache_cleanpos(int count);
210 static void _cache_cleandefered(void);
211 static void _cache_unlink(struct namecache *ncp);
212 
213 /*
214  * The new name cache statistics
215  */
216 SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics");
217 static int numneg;
218 SYSCTL_INT(_vfs_cache, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0,
219     "Number of negative namecache entries");
220 static int numcache;
221 SYSCTL_INT(_vfs_cache, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0,
222     "Number of namecaches entries");
223 static u_long numcalls;
224 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcalls, CTLFLAG_RD, &numcalls, 0,
225     "Number of namecache lookups");
226 static u_long numchecks;
227 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numchecks, CTLFLAG_RD, &numchecks, 0,
228     "Number of checked entries in namecache lookups");
229 
230 struct nchstats nchstats[SMP_MAXCPU];
231 /*
232  * Export VFS cache effectiveness statistics to user-land.
233  *
234  * The statistics are left for aggregation to user-land so
235  * neat things can be achieved, like observing per-CPU cache
236  * distribution.
237  */
238 static int
239 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
240 {
241 	struct globaldata *gd;
242 	int i, error;
243 
244 	error = 0;
245 	for (i = 0; i < ncpus; ++i) {
246 		gd = globaldata_find(i);
247 		if ((error = SYSCTL_OUT(req, (void *)&(*gd->gd_nchstats),
248 			sizeof(struct nchstats))))
249 			break;
250 	}
251 
252 	return (error);
253 }
254 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE|CTLFLAG_RD,
255   0, 0, sysctl_nchstats, "S,nchstats", "VFS cache effectiveness statistics");
256 
257 static struct namecache *cache_zap(struct namecache *ncp, int nonblock);
258 
259 /*
260  * Namespace locking.  The caller must already hold a reference to the
261  * namecache structure in order to lock/unlock it.  This function prevents
262  * the namespace from being created or destroyed by accessors other then
263  * the lock holder.
264  *
265  * Note that holding a locked namecache structure prevents other threads
266  * from making namespace changes (e.g. deleting or creating), prevents
267  * vnode association state changes by other threads, and prevents the
268  * namecache entry from being resolved or unresolved by other threads.
269  *
270  * The lock owner has full authority to associate/disassociate vnodes
271  * and resolve/unresolve the locked ncp.
272  *
273  * The primary lock field is nc_exlocks.  nc_locktd is set after the
274  * fact (when locking) or cleared prior to unlocking.
275  *
276  * WARNING!  Holding a locked ncp will prevent a vnode from being destroyed
277  *	     or recycled, but it does NOT help you if the vnode had already
278  *	     initiated a recyclement.  If this is important, use cache_get()
279  *	     rather then cache_lock() (and deal with the differences in the
280  *	     way the refs counter is handled).  Or, alternatively, make an
281  *	     unconditional call to cache_validate() or cache_resolve()
282  *	     after cache_lock() returns.
283  *
284  * MPSAFE
285  */
286 static
287 void
288 _cache_lock(struct namecache *ncp)
289 {
290 	thread_t td;
291 	int didwarn;
292 	int error;
293 	u_int count;
294 
295 	KKASSERT(ncp->nc_refs != 0);
296 	didwarn = 0;
297 	td = curthread;
298 
299 	for (;;) {
300 		count = ncp->nc_exlocks;
301 
302 		if (count == 0) {
303 			if (atomic_cmpset_int(&ncp->nc_exlocks, 0, 1)) {
304 				/*
305 				 * The vp associated with a locked ncp must
306 				 * be held to prevent it from being recycled.
307 				 *
308 				 * WARNING!  If VRECLAIMED is set the vnode
309 				 * could already be in the middle of a recycle.
310 				 * Callers must use cache_vref() or
311 				 * cache_vget() on the locked ncp to
312 				 * validate the vp or set the cache entry
313 				 * to unresolved.
314 				 *
315 				 * NOTE! vhold() is allowed if we hold a
316 				 *	 lock on the ncp (which we do).
317 				 */
318 				ncp->nc_locktd = td;
319 				if (ncp->nc_vp)
320 					vhold(ncp->nc_vp);	/* MPSAFE */
321 				break;
322 			}
323 			/* cmpset failed */
324 			continue;
325 		}
326 		if (ncp->nc_locktd == td) {
327 			if (atomic_cmpset_int(&ncp->nc_exlocks, count,
328 					      count + 1)) {
329 				break;
330 			}
331 			/* cmpset failed */
332 			continue;
333 		}
334 		tsleep_interlock(ncp, 0);
335 		if (atomic_cmpset_int(&ncp->nc_exlocks, count,
336 				      count | NC_EXLOCK_REQ) == 0) {
337 			/* cmpset failed */
338 			continue;
339 		}
340 		error = tsleep(ncp, PINTERLOCKED, "clock", nclockwarn);
341 		if (error == EWOULDBLOCK) {
342 			if (didwarn == 0) {
343 				didwarn = ticks;
344 				kprintf("[diagnostic] cache_lock: blocked "
345 					"on %p",
346 					ncp);
347 				kprintf(" \"%*.*s\"\n",
348 					ncp->nc_nlen, ncp->nc_nlen,
349 					ncp->nc_name);
350 			}
351 		}
352 	}
353 	if (didwarn) {
354 		kprintf("[diagnostic] cache_lock: unblocked %*.*s after "
355 			"%d secs\n",
356 			ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name,
357 			(int)(ticks - didwarn) / hz);
358 	}
359 }
360 
361 /*
362  * NOTE: nc_refs may be zero if the ncp is interlocked by circumstance,
363  *	 such as the case where one of its children is locked.
364  *
365  * MPSAFE
366  */
367 static
368 int
369 _cache_lock_nonblock(struct namecache *ncp)
370 {
371 	thread_t td;
372 	u_int count;
373 
374 	td = curthread;
375 
376 	for (;;) {
377 		count = ncp->nc_exlocks;
378 
379 		if (count == 0) {
380 			if (atomic_cmpset_int(&ncp->nc_exlocks, 0, 1)) {
381 				/*
382 				 * The vp associated with a locked ncp must
383 				 * be held to prevent it from being recycled.
384 				 *
385 				 * WARNING!  If VRECLAIMED is set the vnode
386 				 * could already be in the middle of a recycle.
387 				 * Callers must use cache_vref() or
388 				 * cache_vget() on the locked ncp to
389 				 * validate the vp or set the cache entry
390 				 * to unresolved.
391 				 *
392 				 * NOTE! vhold() is allowed if we hold a
393 				 *	 lock on the ncp (which we do).
394 				 */
395 				ncp->nc_locktd = td;
396 				if (ncp->nc_vp)
397 					vhold(ncp->nc_vp);	/* MPSAFE */
398 				break;
399 			}
400 			/* cmpset failed */
401 			continue;
402 		}
403 		if (ncp->nc_locktd == td) {
404 			if (atomic_cmpset_int(&ncp->nc_exlocks, count,
405 					      count + 1)) {
406 				break;
407 			}
408 			/* cmpset failed */
409 			continue;
410 		}
411 		return(EWOULDBLOCK);
412 	}
413 	return(0);
414 }
415 
416 /*
417  * Helper function
418  *
419  * NOTE: nc_refs can be 0 (degenerate case during _cache_drop).
420  *
421  *	 nc_locktd must be NULLed out prior to nc_exlocks getting cleared.
422  *
423  * MPSAFE
424  */
425 static
426 void
427 _cache_unlock(struct namecache *ncp)
428 {
429 	thread_t td __debugvar = curthread;
430 	u_int count;
431 
432 	KKASSERT(ncp->nc_refs >= 0);
433 	KKASSERT(ncp->nc_exlocks > 0);
434 	KKASSERT(ncp->nc_locktd == td);
435 
436 	count = ncp->nc_exlocks;
437 	if ((count & ~NC_EXLOCK_REQ) == 1) {
438 		ncp->nc_locktd = NULL;
439 		if (ncp->nc_vp)
440 			vdrop(ncp->nc_vp);
441 	}
442 	for (;;) {
443 		if ((count & ~NC_EXLOCK_REQ) == 1) {
444 			if (atomic_cmpset_int(&ncp->nc_exlocks, count, 0)) {
445 				if (count & NC_EXLOCK_REQ)
446 					wakeup(ncp);
447 				break;
448 			}
449 		} else {
450 			if (atomic_cmpset_int(&ncp->nc_exlocks, count,
451 					      count - 1)) {
452 				break;
453 			}
454 		}
455 		count = ncp->nc_exlocks;
456 	}
457 }
458 
459 
460 /*
461  * cache_hold() and cache_drop() prevent the premature deletion of a
462  * namecache entry but do not prevent operations (such as zapping) on
463  * that namecache entry.
464  *
465  * This routine may only be called from outside this source module if
466  * nc_refs is already at least 1.
467  *
468  * This is a rare case where callers are allowed to hold a spinlock,
469  * so we can't ourselves.
470  *
471  * MPSAFE
472  */
473 static __inline
474 struct namecache *
475 _cache_hold(struct namecache *ncp)
476 {
477 	atomic_add_int(&ncp->nc_refs, 1);
478 	return(ncp);
479 }
480 
481 /*
482  * Drop a cache entry, taking care to deal with races.
483  *
484  * For potential 1->0 transitions we must hold the ncp lock to safely
485  * test its flags.  An unresolved entry with no children must be zapped
486  * to avoid leaks.
487  *
488  * The call to cache_zap() itself will handle all remaining races and
489  * will decrement the ncp's refs regardless.  If we are resolved or
490  * have children nc_refs can safely be dropped to 0 without having to
491  * zap the entry.
492  *
493  * NOTE: cache_zap() will re-check nc_refs and nc_list in a MPSAFE fashion.
494  *
495  * NOTE: cache_zap() may return a non-NULL referenced parent which must
496  *	 be dropped in a loop.
497  *
498  * MPSAFE
499  */
500 static __inline
501 void
502 _cache_drop(struct namecache *ncp)
503 {
504 	int refs;
505 
506 	while (ncp) {
507 		KKASSERT(ncp->nc_refs > 0);
508 		refs = ncp->nc_refs;
509 
510 		if (refs == 1) {
511 			if (_cache_lock_nonblock(ncp) == 0) {
512 				ncp->nc_flag &= ~NCF_DEFEREDZAP;
513 				if ((ncp->nc_flag & NCF_UNRESOLVED) &&
514 				    TAILQ_EMPTY(&ncp->nc_list)) {
515 					ncp = cache_zap(ncp, 1);
516 					continue;
517 				}
518 				if (atomic_cmpset_int(&ncp->nc_refs, 1, 0)) {
519 					_cache_unlock(ncp);
520 					break;
521 				}
522 				_cache_unlock(ncp);
523 			}
524 		} else {
525 			if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1))
526 				break;
527 		}
528 		cpu_pause();
529 	}
530 }
531 
532 /*
533  * Link a new namecache entry to its parent and to the hash table.  Be
534  * careful to avoid races if vhold() blocks in the future.
535  *
536  * Both ncp and par must be referenced and locked.
537  *
538  * NOTE: The hash table spinlock is likely held during this call, we
539  *	 can't do anything fancy.
540  *
541  * MPSAFE
542  */
543 static void
544 _cache_link_parent(struct namecache *ncp, struct namecache *par,
545 		   struct nchash_head *nchpp)
546 {
547 	KKASSERT(ncp->nc_parent == NULL);
548 	ncp->nc_parent = par;
549 	ncp->nc_head = nchpp;
550 
551 	/*
552 	 * Set inheritance flags.  Note that the parent flags may be
553 	 * stale due to getattr potentially not having been run yet
554 	 * (it gets run during nlookup()'s).
555 	 */
556 	ncp->nc_flag &= ~(NCF_SF_PNOCACHE | NCF_UF_PCACHE);
557 	if (par->nc_flag & (NCF_SF_NOCACHE | NCF_SF_PNOCACHE))
558 		ncp->nc_flag |= NCF_SF_PNOCACHE;
559 	if (par->nc_flag & (NCF_UF_CACHE | NCF_UF_PCACHE))
560 		ncp->nc_flag |= NCF_UF_PCACHE;
561 
562 	LIST_INSERT_HEAD(&nchpp->list, ncp, nc_hash);
563 
564 	if (TAILQ_EMPTY(&par->nc_list)) {
565 		TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
566 		/*
567 		 * Any vp associated with an ncp which has children must
568 		 * be held to prevent it from being recycled.
569 		 */
570 		if (par->nc_vp)
571 			vhold(par->nc_vp);
572 	} else {
573 		TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
574 	}
575 }
576 
577 /*
578  * Remove the parent and hash associations from a namecache structure.
579  * If this is the last child of the parent the cache_drop(par) will
580  * attempt to recursively zap the parent.
581  *
582  * ncp must be locked.  This routine will acquire a temporary lock on
583  * the parent as wlel as the appropriate hash chain.
584  *
585  * MPSAFE
586  */
587 static void
588 _cache_unlink_parent(struct namecache *ncp)
589 {
590 	struct namecache *par;
591 	struct vnode *dropvp;
592 
593 	if ((par = ncp->nc_parent) != NULL) {
594 		KKASSERT(ncp->nc_parent == par);
595 		_cache_hold(par);
596 		_cache_lock(par);
597 		spin_lock(&ncp->nc_head->spin);
598 		LIST_REMOVE(ncp, nc_hash);
599 		TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
600 		dropvp = NULL;
601 		if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
602 			dropvp = par->nc_vp;
603 		spin_unlock(&ncp->nc_head->spin);
604 		ncp->nc_parent = NULL;
605 		ncp->nc_head = NULL;
606 		_cache_unlock(par);
607 		_cache_drop(par);
608 
609 		/*
610 		 * We can only safely vdrop with no spinlocks held.
611 		 */
612 		if (dropvp)
613 			vdrop(dropvp);
614 	}
615 }
616 
617 /*
618  * Allocate a new namecache structure.  Most of the code does not require
619  * zero-termination of the string but it makes vop_compat_ncreate() easier.
620  *
621  * MPSAFE
622  */
623 static struct namecache *
624 cache_alloc(int nlen)
625 {
626 	struct namecache *ncp;
627 
628 	ncp = kmalloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO);
629 	if (nlen)
630 		ncp->nc_name = kmalloc(nlen + 1, M_VFSCACHE, M_WAITOK);
631 	ncp->nc_nlen = nlen;
632 	ncp->nc_flag = NCF_UNRESOLVED;
633 	ncp->nc_error = ENOTCONN;	/* needs to be resolved */
634 	ncp->nc_refs = 1;
635 
636 	TAILQ_INIT(&ncp->nc_list);
637 	_cache_lock(ncp);
638 	return(ncp);
639 }
640 
641 /*
642  * Can only be called for the case where the ncp has never been
643  * associated with anything (so no spinlocks are needed).
644  *
645  * MPSAFE
646  */
647 static void
648 _cache_free(struct namecache *ncp)
649 {
650 	KKASSERT(ncp->nc_refs == 1 && ncp->nc_exlocks == 1);
651 	if (ncp->nc_name)
652 		kfree(ncp->nc_name, M_VFSCACHE);
653 	kfree(ncp, M_VFSCACHE);
654 }
655 
656 /*
657  * MPSAFE
658  */
659 void
660 cache_zero(struct nchandle *nch)
661 {
662 	nch->ncp = NULL;
663 	nch->mount = NULL;
664 }
665 
666 /*
667  * Ref and deref a namecache structure.
668  *
669  * The caller must specify a stable ncp pointer, typically meaning the
670  * ncp is already referenced but this can also occur indirectly through
671  * e.g. holding a lock on a direct child.
672  *
673  * WARNING: Caller may hold an unrelated read spinlock, which means we can't
674  *	    use read spinlocks here.
675  *
676  * MPSAFE if nch is
677  */
678 struct nchandle *
679 cache_hold(struct nchandle *nch)
680 {
681 	_cache_hold(nch->ncp);
682 	atomic_add_int(&nch->mount->mnt_refs, 1);
683 	return(nch);
684 }
685 
686 /*
687  * Create a copy of a namecache handle for an already-referenced
688  * entry.
689  *
690  * MPSAFE if nch is
691  */
692 void
693 cache_copy(struct nchandle *nch, struct nchandle *target)
694 {
695 	*target = *nch;
696 	if (target->ncp)
697 		_cache_hold(target->ncp);
698 	atomic_add_int(&nch->mount->mnt_refs, 1);
699 }
700 
701 /*
702  * MPSAFE if nch is
703  */
704 void
705 cache_changemount(struct nchandle *nch, struct mount *mp)
706 {
707 	atomic_add_int(&nch->mount->mnt_refs, -1);
708 	nch->mount = mp;
709 	atomic_add_int(&nch->mount->mnt_refs, 1);
710 }
711 
712 /*
713  * MPSAFE
714  */
715 void
716 cache_drop(struct nchandle *nch)
717 {
718 	atomic_add_int(&nch->mount->mnt_refs, -1);
719 	_cache_drop(nch->ncp);
720 	nch->ncp = NULL;
721 	nch->mount = NULL;
722 }
723 
724 /*
725  * MPSAFE
726  */
727 void
728 cache_lock(struct nchandle *nch)
729 {
730 	_cache_lock(nch->ncp);
731 }
732 
733 /*
734  * Relock nch1 given an unlocked nch1 and a locked nch2.  The caller
735  * is responsible for checking both for validity on return as they
736  * may have become invalid.
737  *
738  * We have to deal with potential deadlocks here, just ping pong
739  * the lock until we get it (we will always block somewhere when
740  * looping so this is not cpu-intensive).
741  *
742  * which = 0	nch1 not locked, nch2 is locked
743  * which = 1	nch1 is locked, nch2 is not locked
744  */
745 void
746 cache_relock(struct nchandle *nch1, struct ucred *cred1,
747 	     struct nchandle *nch2, struct ucred *cred2)
748 {
749 	int which;
750 
751 	which = 0;
752 
753 	for (;;) {
754 		if (which == 0) {
755 			if (cache_lock_nonblock(nch1) == 0) {
756 				cache_resolve(nch1, cred1);
757 				break;
758 			}
759 			cache_unlock(nch2);
760 			cache_lock(nch1);
761 			cache_resolve(nch1, cred1);
762 			which = 1;
763 		} else {
764 			if (cache_lock_nonblock(nch2) == 0) {
765 				cache_resolve(nch2, cred2);
766 				break;
767 			}
768 			cache_unlock(nch1);
769 			cache_lock(nch2);
770 			cache_resolve(nch2, cred2);
771 			which = 0;
772 		}
773 	}
774 }
775 
776 /*
777  * MPSAFE
778  */
779 int
780 cache_lock_nonblock(struct nchandle *nch)
781 {
782 	return(_cache_lock_nonblock(nch->ncp));
783 }
784 
785 
786 /*
787  * MPSAFE
788  */
789 void
790 cache_unlock(struct nchandle *nch)
791 {
792 	_cache_unlock(nch->ncp);
793 }
794 
795 /*
796  * ref-and-lock, unlock-and-deref functions.
797  *
798  * This function is primarily used by nlookup.  Even though cache_lock
799  * holds the vnode, it is possible that the vnode may have already
800  * initiated a recyclement.
801  *
802  * We want cache_get() to return a definitively usable vnode or a
803  * definitively unresolved ncp.
804  *
805  * MPSAFE
806  */
807 static
808 struct namecache *
809 _cache_get(struct namecache *ncp)
810 {
811 	_cache_hold(ncp);
812 	_cache_lock(ncp);
813 	if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
814 		_cache_setunresolved(ncp);
815 	return(ncp);
816 }
817 
818 /*
819  * This is a special form of _cache_lock() which only succeeds if
820  * it can get a pristine, non-recursive lock.  The caller must have
821  * already ref'd the ncp.
822  *
823  * On success the ncp will be locked, on failure it will not.  The
824  * ref count does not change either way.
825  *
826  * We want _cache_lock_special() (on success) to return a definitively
827  * usable vnode or a definitively unresolved ncp.
828  *
829  * MPSAFE
830  */
831 static int
832 _cache_lock_special(struct namecache *ncp)
833 {
834 	if (_cache_lock_nonblock(ncp) == 0) {
835 		if ((ncp->nc_exlocks & ~NC_EXLOCK_REQ) == 1) {
836 			if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
837 				_cache_setunresolved(ncp);
838 			return(0);
839 		}
840 		_cache_unlock(ncp);
841 	}
842 	return(EWOULDBLOCK);
843 }
844 
845 
846 /*
847  * NOTE: The same nchandle can be passed for both arguments.
848  *
849  * MPSAFE
850  */
851 void
852 cache_get(struct nchandle *nch, struct nchandle *target)
853 {
854 	KKASSERT(nch->ncp->nc_refs > 0);
855 	target->mount = nch->mount;
856 	target->ncp = _cache_get(nch->ncp);
857 	atomic_add_int(&target->mount->mnt_refs, 1);
858 }
859 
860 /*
861  * MPSAFE
862  */
863 static __inline
864 void
865 _cache_put(struct namecache *ncp)
866 {
867 	_cache_unlock(ncp);
868 	_cache_drop(ncp);
869 }
870 
871 /*
872  * MPSAFE
873  */
874 void
875 cache_put(struct nchandle *nch)
876 {
877 	atomic_add_int(&nch->mount->mnt_refs, -1);
878 	_cache_put(nch->ncp);
879 	nch->ncp = NULL;
880 	nch->mount = NULL;
881 }
882 
883 /*
884  * Resolve an unresolved ncp by associating a vnode with it.  If the
885  * vnode is NULL, a negative cache entry is created.
886  *
887  * The ncp should be locked on entry and will remain locked on return.
888  *
889  * MPSAFE
890  */
891 static
892 void
893 _cache_setvp(struct mount *mp, struct namecache *ncp, struct vnode *vp)
894 {
895 	KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
896 
897 	if (vp != NULL) {
898 		/*
899 		 * Any vp associated with an ncp which has children must
900 		 * be held.  Any vp associated with a locked ncp must be held.
901 		 */
902 		if (!TAILQ_EMPTY(&ncp->nc_list))
903 			vhold(vp);
904 		spin_lock(&vp->v_spin);
905 		ncp->nc_vp = vp;
906 		TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode);
907 		spin_unlock(&vp->v_spin);
908 		if (ncp->nc_exlocks)
909 			vhold(vp);
910 
911 		/*
912 		 * Set auxiliary flags
913 		 */
914 		switch(vp->v_type) {
915 		case VDIR:
916 			ncp->nc_flag |= NCF_ISDIR;
917 			break;
918 		case VLNK:
919 			ncp->nc_flag |= NCF_ISSYMLINK;
920 			/* XXX cache the contents of the symlink */
921 			break;
922 		default:
923 			break;
924 		}
925 		atomic_add_int(&numcache, 1);
926 		ncp->nc_error = 0;
927 		/* XXX: this is a hack to work-around the lack of a real pfs vfs
928 		 * implementation*/
929 		if (mp != NULL)
930 			if (strncmp(mp->mnt_stat.f_fstypename, "null", 5) == 0)
931 				vp->v_pfsmp = mp;
932 	} else {
933 		/*
934 		 * When creating a negative cache hit we set the
935 		 * namecache_gen.  A later resolve will clean out the
936 		 * negative cache hit if the mount point's namecache_gen
937 		 * has changed.  Used by devfs, could also be used by
938 		 * other remote FSs.
939 		 */
940 		ncp->nc_vp = NULL;
941 		spin_lock(&ncspin);
942 		TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
943 		++numneg;
944 		spin_unlock(&ncspin);
945 		ncp->nc_error = ENOENT;
946 		if (mp)
947 			VFS_NCPGEN_SET(mp, ncp);
948 	}
949 	ncp->nc_flag &= ~(NCF_UNRESOLVED | NCF_DEFEREDZAP);
950 }
951 
952 /*
953  * MPSAFE
954  */
955 void
956 cache_setvp(struct nchandle *nch, struct vnode *vp)
957 {
958 	_cache_setvp(nch->mount, nch->ncp, vp);
959 }
960 
961 /*
962  * MPSAFE
963  */
964 void
965 cache_settimeout(struct nchandle *nch, int nticks)
966 {
967 	struct namecache *ncp = nch->ncp;
968 
969 	if ((ncp->nc_timeout = ticks + nticks) == 0)
970 		ncp->nc_timeout = 1;
971 }
972 
973 /*
974  * Disassociate the vnode or negative-cache association and mark a
975  * namecache entry as unresolved again.  Note that the ncp is still
976  * left in the hash table and still linked to its parent.
977  *
978  * The ncp should be locked and refd on entry and will remain locked and refd
979  * on return.
980  *
981  * This routine is normally never called on a directory containing children.
982  * However, NFS often does just that in its rename() code as a cop-out to
983  * avoid complex namespace operations.  This disconnects a directory vnode
984  * from its namecache and can cause the OLDAPI and NEWAPI to get out of
985  * sync.
986  *
987  * MPSAFE
988  */
989 static
990 void
991 _cache_setunresolved(struct namecache *ncp)
992 {
993 	struct vnode *vp;
994 
995 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
996 		ncp->nc_flag |= NCF_UNRESOLVED;
997 		ncp->nc_timeout = 0;
998 		ncp->nc_error = ENOTCONN;
999 		if ((vp = ncp->nc_vp) != NULL) {
1000 			atomic_add_int(&numcache, -1);
1001 			spin_lock(&vp->v_spin);
1002 			ncp->nc_vp = NULL;
1003 			TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode);
1004 			spin_unlock(&vp->v_spin);
1005 
1006 			/*
1007 			 * Any vp associated with an ncp with children is
1008 			 * held by that ncp.  Any vp associated with a locked
1009 			 * ncp is held by that ncp.  These conditions must be
1010 			 * undone when the vp is cleared out from the ncp.
1011 			 */
1012 			if (!TAILQ_EMPTY(&ncp->nc_list))
1013 				vdrop(vp);
1014 			if (ncp->nc_exlocks)
1015 				vdrop(vp);
1016 		} else {
1017 			spin_lock(&ncspin);
1018 			TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
1019 			--numneg;
1020 			spin_unlock(&ncspin);
1021 		}
1022 		ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK);
1023 	}
1024 }
1025 
1026 /*
1027  * The cache_nresolve() code calls this function to automatically
1028  * set a resolved cache element to unresolved if it has timed out
1029  * or if it is a negative cache hit and the mount point namecache_gen
1030  * has changed.
1031  *
1032  * MPSAFE
1033  */
1034 static __inline void
1035 _cache_auto_unresolve(struct mount *mp, struct namecache *ncp)
1036 {
1037 	/*
1038 	 * Already in an unresolved state, nothing to do.
1039 	 */
1040 	if (ncp->nc_flag & NCF_UNRESOLVED)
1041 		return;
1042 
1043 	/*
1044 	 * Try to zap entries that have timed out.  We have
1045 	 * to be careful here because locked leafs may depend
1046 	 * on the vnode remaining intact in a parent, so only
1047 	 * do this under very specific conditions.
1048 	 */
1049 	if (ncp->nc_timeout && (int)(ncp->nc_timeout - ticks) < 0 &&
1050 	    TAILQ_EMPTY(&ncp->nc_list)) {
1051 		_cache_setunresolved(ncp);
1052 		return;
1053 	}
1054 
1055 	/*
1056 	 * If a resolved negative cache hit is invalid due to
1057 	 * the mount's namecache generation being bumped, zap it.
1058 	 */
1059 	if (ncp->nc_vp == NULL && VFS_NCPGEN_TEST(mp, ncp)) {
1060 		_cache_setunresolved(ncp);
1061 		return;
1062 	}
1063 }
1064 
1065 /*
1066  * MPSAFE
1067  */
1068 void
1069 cache_setunresolved(struct nchandle *nch)
1070 {
1071 	_cache_setunresolved(nch->ncp);
1072 }
1073 
1074 /*
1075  * Determine if we can clear NCF_ISMOUNTPT by scanning the mountlist
1076  * looking for matches.  This flag tells the lookup code when it must
1077  * check for a mount linkage and also prevents the directories in question
1078  * from being deleted or renamed.
1079  *
1080  * MPSAFE
1081  */
1082 static
1083 int
1084 cache_clrmountpt_callback(struct mount *mp, void *data)
1085 {
1086 	struct nchandle *nch = data;
1087 
1088 	if (mp->mnt_ncmounton.ncp == nch->ncp)
1089 		return(1);
1090 	if (mp->mnt_ncmountpt.ncp == nch->ncp)
1091 		return(1);
1092 	return(0);
1093 }
1094 
1095 /*
1096  * MPSAFE
1097  */
1098 void
1099 cache_clrmountpt(struct nchandle *nch)
1100 {
1101 	int count;
1102 
1103 	count = mountlist_scan(cache_clrmountpt_callback, nch,
1104 			       MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
1105 	if (count == 0)
1106 		nch->ncp->nc_flag &= ~NCF_ISMOUNTPT;
1107 }
1108 
1109 /*
1110  * Invalidate portions of the namecache topology given a starting entry.
1111  * The passed ncp is set to an unresolved state and:
1112  *
1113  * The passed ncp must be referencxed and locked.  The routine may unlock
1114  * and relock ncp several times, and will recheck the children and loop
1115  * to catch races.  When done the passed ncp will be returned with the
1116  * reference and lock intact.
1117  *
1118  * CINV_DESTROY		- Set a flag in the passed ncp entry indicating
1119  *			  that the physical underlying nodes have been
1120  *			  destroyed... as in deleted.  For example, when
1121  *			  a directory is removed.  This will cause record
1122  *			  lookups on the name to no longer be able to find
1123  *			  the record and tells the resolver to return failure
1124  *			  rather then trying to resolve through the parent.
1125  *
1126  *			  The topology itself, including ncp->nc_name,
1127  *			  remains intact.
1128  *
1129  *			  This only applies to the passed ncp, if CINV_CHILDREN
1130  *			  is specified the children are not flagged.
1131  *
1132  * CINV_CHILDREN	- Set all children (recursively) to an unresolved
1133  *			  state as well.
1134  *
1135  *			  Note that this will also have the side effect of
1136  *			  cleaning out any unreferenced nodes in the topology
1137  *			  from the leaves up as the recursion backs out.
1138  *
1139  * Note that the topology for any referenced nodes remains intact, but
1140  * the nodes will be marked as having been destroyed and will be set
1141  * to an unresolved state.
1142  *
1143  * It is possible for cache_inval() to race a cache_resolve(), meaning that
1144  * the namecache entry may not actually be invalidated on return if it was
1145  * revalidated while recursing down into its children.  This code guarentees
1146  * that the node(s) will go through an invalidation cycle, but does not
1147  * guarentee that they will remain in an invalidated state.
1148  *
1149  * Returns non-zero if a revalidation was detected during the invalidation
1150  * recursion, zero otherwise.  Note that since only the original ncp is
1151  * locked the revalidation ultimately can only indicate that the original ncp
1152  * *MIGHT* no have been reresolved.
1153  *
1154  * DEEP RECURSION HANDLING - If a recursive invalidation recurses deeply we
1155  * have to avoid blowing out the kernel stack.  We do this by saving the
1156  * deep namecache node and aborting the recursion, then re-recursing at that
1157  * node using a depth-first algorithm in order to allow multiple deep
1158  * recursions to chain through each other, then we restart the invalidation
1159  * from scratch.
1160  *
1161  * MPSAFE
1162  */
1163 
1164 struct cinvtrack {
1165 	struct namecache *resume_ncp;
1166 	int depth;
1167 };
1168 
1169 static int _cache_inval_internal(struct namecache *, int, struct cinvtrack *);
1170 
1171 static
1172 int
1173 _cache_inval(struct namecache *ncp, int flags)
1174 {
1175 	struct cinvtrack track;
1176 	struct namecache *ncp2;
1177 	int r;
1178 
1179 	track.depth = 0;
1180 	track.resume_ncp = NULL;
1181 
1182 	for (;;) {
1183 		r = _cache_inval_internal(ncp, flags, &track);
1184 		if (track.resume_ncp == NULL)
1185 			break;
1186 		kprintf("Warning: deep namecache recursion at %s\n",
1187 			ncp->nc_name);
1188 		_cache_unlock(ncp);
1189 		while ((ncp2 = track.resume_ncp) != NULL) {
1190 			track.resume_ncp = NULL;
1191 			_cache_lock(ncp2);
1192 			_cache_inval_internal(ncp2, flags & ~CINV_DESTROY,
1193 					     &track);
1194 			_cache_put(ncp2);
1195 		}
1196 		_cache_lock(ncp);
1197 	}
1198 	return(r);
1199 }
1200 
1201 int
1202 cache_inval(struct nchandle *nch, int flags)
1203 {
1204 	return(_cache_inval(nch->ncp, flags));
1205 }
1206 
1207 /*
1208  * Helper for _cache_inval().  The passed ncp is refd and locked and
1209  * remains that way on return, but may be unlocked/relocked multiple
1210  * times by the routine.
1211  */
1212 static int
1213 _cache_inval_internal(struct namecache *ncp, int flags, struct cinvtrack *track)
1214 {
1215 	struct namecache *kid;
1216 	struct namecache *nextkid;
1217 	int rcnt = 0;
1218 
1219 	KKASSERT(ncp->nc_exlocks);
1220 
1221 	_cache_setunresolved(ncp);
1222 	if (flags & CINV_DESTROY)
1223 		ncp->nc_flag |= NCF_DESTROYED;
1224 	if ((flags & CINV_CHILDREN) &&
1225 	    (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL
1226 	) {
1227 		_cache_hold(kid);
1228 		if (++track->depth > MAX_RECURSION_DEPTH) {
1229 			track->resume_ncp = ncp;
1230 			_cache_hold(ncp);
1231 			++rcnt;
1232 		}
1233 		_cache_unlock(ncp);
1234 		while (kid) {
1235 			if (track->resume_ncp) {
1236 				_cache_drop(kid);
1237 				break;
1238 			}
1239 			if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL)
1240 				_cache_hold(nextkid);
1241 			if ((kid->nc_flag & NCF_UNRESOLVED) == 0 ||
1242 			    TAILQ_FIRST(&kid->nc_list)
1243 			) {
1244 				_cache_lock(kid);
1245 				rcnt += _cache_inval_internal(kid, flags & ~CINV_DESTROY, track);
1246 				_cache_unlock(kid);
1247 			}
1248 			_cache_drop(kid);
1249 			kid = nextkid;
1250 		}
1251 		--track->depth;
1252 		_cache_lock(ncp);
1253 	}
1254 
1255 	/*
1256 	 * Someone could have gotten in there while ncp was unlocked,
1257 	 * retry if so.
1258 	 */
1259 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
1260 		++rcnt;
1261 	return (rcnt);
1262 }
1263 
1264 /*
1265  * Invalidate a vnode's namecache associations.  To avoid races against
1266  * the resolver we do not invalidate a node which we previously invalidated
1267  * but which was then re-resolved while we were in the invalidation loop.
1268  *
1269  * Returns non-zero if any namecache entries remain after the invalidation
1270  * loop completed.
1271  *
1272  * NOTE: Unlike the namecache topology which guarentees that ncp's will not
1273  *	 be ripped out of the topology while held, the vnode's v_namecache
1274  *	 list has no such restriction.  NCP's can be ripped out of the list
1275  *	 at virtually any time if not locked, even if held.
1276  *
1277  *	 In addition, the v_namecache list itself must be locked via
1278  *	 the vnode's spinlock.
1279  *
1280  * MPSAFE
1281  */
1282 int
1283 cache_inval_vp(struct vnode *vp, int flags)
1284 {
1285 	struct namecache *ncp;
1286 	struct namecache *next;
1287 
1288 restart:
1289 	spin_lock(&vp->v_spin);
1290 	ncp = TAILQ_FIRST(&vp->v_namecache);
1291 	if (ncp)
1292 		_cache_hold(ncp);
1293 	while (ncp) {
1294 		/* loop entered with ncp held and vp spin-locked */
1295 		if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
1296 			_cache_hold(next);
1297 		spin_unlock(&vp->v_spin);
1298 		_cache_lock(ncp);
1299 		if (ncp->nc_vp != vp) {
1300 			kprintf("Warning: cache_inval_vp: race-A detected on "
1301 				"%s\n", ncp->nc_name);
1302 			_cache_put(ncp);
1303 			if (next)
1304 				_cache_drop(next);
1305 			goto restart;
1306 		}
1307 		_cache_inval(ncp, flags);
1308 		_cache_put(ncp);		/* also releases reference */
1309 		ncp = next;
1310 		spin_lock(&vp->v_spin);
1311 		if (ncp && ncp->nc_vp != vp) {
1312 			spin_unlock(&vp->v_spin);
1313 			kprintf("Warning: cache_inval_vp: race-B detected on "
1314 				"%s\n", ncp->nc_name);
1315 			_cache_drop(ncp);
1316 			goto restart;
1317 		}
1318 	}
1319 	spin_unlock(&vp->v_spin);
1320 	return(TAILQ_FIRST(&vp->v_namecache) != NULL);
1321 }
1322 
1323 /*
1324  * This routine is used instead of the normal cache_inval_vp() when we
1325  * are trying to recycle otherwise good vnodes.
1326  *
1327  * Return 0 on success, non-zero if not all namecache records could be
1328  * disassociated from the vnode (for various reasons).
1329  *
1330  * MPSAFE
1331  */
1332 int
1333 cache_inval_vp_nonblock(struct vnode *vp)
1334 {
1335 	struct namecache *ncp;
1336 	struct namecache *next;
1337 
1338 	spin_lock(&vp->v_spin);
1339 	ncp = TAILQ_FIRST(&vp->v_namecache);
1340 	if (ncp)
1341 		_cache_hold(ncp);
1342 	while (ncp) {
1343 		/* loop entered with ncp held */
1344 		if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
1345 			_cache_hold(next);
1346 		spin_unlock(&vp->v_spin);
1347 		if (_cache_lock_nonblock(ncp)) {
1348 			_cache_drop(ncp);
1349 			if (next)
1350 				_cache_drop(next);
1351 			goto done;
1352 		}
1353 		if (ncp->nc_vp != vp) {
1354 			kprintf("Warning: cache_inval_vp: race-A detected on "
1355 				"%s\n", ncp->nc_name);
1356 			_cache_put(ncp);
1357 			if (next)
1358 				_cache_drop(next);
1359 			goto done;
1360 		}
1361 		_cache_inval(ncp, 0);
1362 		_cache_put(ncp);		/* also releases reference */
1363 		ncp = next;
1364 		spin_lock(&vp->v_spin);
1365 		if (ncp && ncp->nc_vp != vp) {
1366 			spin_unlock(&vp->v_spin);
1367 			kprintf("Warning: cache_inval_vp: race-B detected on "
1368 				"%s\n", ncp->nc_name);
1369 			_cache_drop(ncp);
1370 			goto done;
1371 		}
1372 	}
1373 	spin_unlock(&vp->v_spin);
1374 done:
1375 	return(TAILQ_FIRST(&vp->v_namecache) != NULL);
1376 }
1377 
1378 /*
1379  * The source ncp has been renamed to the target ncp.  Both fncp and tncp
1380  * must be locked.  The target ncp is destroyed (as a normal rename-over
1381  * would destroy the target file or directory).
1382  *
1383  * Because there may be references to the source ncp we cannot copy its
1384  * contents to the target.  Instead the source ncp is relinked as the target
1385  * and the target ncp is removed from the namecache topology.
1386  *
1387  * MPSAFE
1388  */
1389 void
1390 cache_rename(struct nchandle *fnch, struct nchandle *tnch)
1391 {
1392 	struct namecache *fncp = fnch->ncp;
1393 	struct namecache *tncp = tnch->ncp;
1394 	struct namecache *tncp_par;
1395 	struct nchash_head *nchpp;
1396 	u_int32_t hash;
1397 	char *oname;
1398 	char *nname;
1399 
1400 	if (tncp->nc_nlen) {
1401 		nname = kmalloc(tncp->nc_nlen + 1, M_VFSCACHE, M_WAITOK);
1402 		bcopy(tncp->nc_name, nname, tncp->nc_nlen);
1403 		nname[tncp->nc_nlen] = 0;
1404 	} else {
1405 		nname = NULL;
1406 	}
1407 
1408 	/*
1409 	 * Rename fncp (unlink)
1410 	 */
1411 	_cache_unlink_parent(fncp);
1412 	oname = fncp->nc_name;
1413 	fncp->nc_name = nname;
1414 	fncp->nc_nlen = tncp->nc_nlen;
1415 	if (oname)
1416 		kfree(oname, M_VFSCACHE);
1417 
1418 	tncp_par = tncp->nc_parent;
1419 	_cache_hold(tncp_par);
1420 	_cache_lock(tncp_par);
1421 
1422 	/*
1423 	 * Rename fncp (relink)
1424 	 */
1425 	hash = fnv_32_buf(fncp->nc_name, fncp->nc_nlen, FNV1_32_INIT);
1426 	hash = fnv_32_buf(&tncp_par, sizeof(tncp_par), hash);
1427 	nchpp = NCHHASH(hash);
1428 
1429 	spin_lock(&nchpp->spin);
1430 	_cache_link_parent(fncp, tncp_par, nchpp);
1431 	spin_unlock(&nchpp->spin);
1432 
1433 	_cache_put(tncp_par);
1434 
1435 	/*
1436 	 * Get rid of the overwritten tncp (unlink)
1437 	 */
1438 	_cache_unlink(tncp);
1439 }
1440 
1441 /*
1442  * Perform actions consistent with unlinking a file.  The passed-in ncp
1443  * must be locked.
1444  *
1445  * The ncp is marked DESTROYED so it no longer shows up in searches,
1446  * and will be physically deleted when the vnode goes away.
1447  *
1448  * If the related vnode has no refs then we cycle it through vget()/vput()
1449  * to (possibly if we don't have a ref race) trigger a deactivation,
1450  * allowing the VFS to trivially detect and recycle the deleted vnode
1451  * via VOP_INACTIVE().
1452  *
1453  * NOTE: _cache_rename() will automatically call _cache_unlink() on the
1454  *	 target ncp.
1455  */
1456 void
1457 cache_unlink(struct nchandle *nch)
1458 {
1459 	_cache_unlink(nch->ncp);
1460 }
1461 
1462 static void
1463 _cache_unlink(struct namecache *ncp)
1464 {
1465 	struct vnode *vp;
1466 
1467 	/*
1468 	 * Causes lookups to fail and allows another ncp with the same
1469 	 * name to be created under ncp->nc_parent.
1470 	 */
1471 	ncp->nc_flag |= NCF_DESTROYED;
1472 
1473 	/*
1474 	 * Attempt to trigger a deactivation.
1475 	 */
1476 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
1477 	    (vp = ncp->nc_vp) != NULL &&
1478 	    !sysref_isactive(&vp->v_sysref)) {
1479 		if (vget(vp, LK_SHARED) == 0)
1480 			vput(vp);
1481 	}
1482 }
1483 
1484 /*
1485  * vget the vnode associated with the namecache entry.  Resolve the namecache
1486  * entry if necessary.  The passed ncp must be referenced and locked.
1487  *
1488  * lk_type may be LK_SHARED, LK_EXCLUSIVE.  A ref'd, possibly locked
1489  * (depending on the passed lk_type) will be returned in *vpp with an error
1490  * of 0, or NULL will be returned in *vpp with a non-0 error code.  The
1491  * most typical error is ENOENT, meaning that the ncp represents a negative
1492  * cache hit and there is no vnode to retrieve, but other errors can occur
1493  * too.
1494  *
1495  * The vget() can race a reclaim.  If this occurs we re-resolve the
1496  * namecache entry.
1497  *
1498  * There are numerous places in the kernel where vget() is called on a
1499  * vnode while one or more of its namecache entries is locked.  Releasing
1500  * a vnode never deadlocks against locked namecache entries (the vnode
1501  * will not get recycled while referenced ncp's exist).  This means we
1502  * can safely acquire the vnode.  In fact, we MUST NOT release the ncp
1503  * lock when acquiring the vp lock or we might cause a deadlock.
1504  *
1505  * MPSAFE
1506  */
1507 int
1508 cache_vget(struct nchandle *nch, struct ucred *cred,
1509 	   int lk_type, struct vnode **vpp)
1510 {
1511 	struct namecache *ncp;
1512 	struct vnode *vp;
1513 	int error;
1514 
1515 	ncp = nch->ncp;
1516 	KKASSERT(ncp->nc_locktd == curthread);
1517 again:
1518 	vp = NULL;
1519 	if (ncp->nc_flag & NCF_UNRESOLVED)
1520 		error = cache_resolve(nch, cred);
1521 	else
1522 		error = 0;
1523 
1524 	if (error == 0 && (vp = ncp->nc_vp) != NULL) {
1525 		error = vget(vp, lk_type);
1526 		if (error) {
1527 			/*
1528 			 * VRECLAIM race
1529 			 */
1530 			if (error == ENOENT) {
1531 				kprintf("Warning: vnode reclaim race detected "
1532 					"in cache_vget on %p (%s)\n",
1533 					vp, ncp->nc_name);
1534 				_cache_setunresolved(ncp);
1535 				goto again;
1536 			}
1537 
1538 			/*
1539 			 * Not a reclaim race, some other error.
1540 			 */
1541 			KKASSERT(ncp->nc_vp == vp);
1542 			vp = NULL;
1543 		} else {
1544 			KKASSERT(ncp->nc_vp == vp);
1545 			KKASSERT((vp->v_flag & VRECLAIMED) == 0);
1546 		}
1547 	}
1548 	if (error == 0 && vp == NULL)
1549 		error = ENOENT;
1550 	*vpp = vp;
1551 	return(error);
1552 }
1553 
1554 int
1555 cache_vref(struct nchandle *nch, struct ucred *cred, struct vnode **vpp)
1556 {
1557 	struct namecache *ncp;
1558 	struct vnode *vp;
1559 	int error;
1560 
1561 	ncp = nch->ncp;
1562 	KKASSERT(ncp->nc_locktd == curthread);
1563 again:
1564 	vp = NULL;
1565 	if (ncp->nc_flag & NCF_UNRESOLVED)
1566 		error = cache_resolve(nch, cred);
1567 	else
1568 		error = 0;
1569 
1570 	if (error == 0 && (vp = ncp->nc_vp) != NULL) {
1571 		error = vget(vp, LK_SHARED);
1572 		if (error) {
1573 			/*
1574 			 * VRECLAIM race
1575 			 */
1576 			if (error == ENOENT) {
1577 				kprintf("Warning: vnode reclaim race detected "
1578 					"in cache_vget on %p (%s)\n",
1579 					vp, ncp->nc_name);
1580 				_cache_setunresolved(ncp);
1581 				goto again;
1582 			}
1583 
1584 			/*
1585 			 * Not a reclaim race, some other error.
1586 			 */
1587 			KKASSERT(ncp->nc_vp == vp);
1588 			vp = NULL;
1589 		} else {
1590 			KKASSERT(ncp->nc_vp == vp);
1591 			KKASSERT((vp->v_flag & VRECLAIMED) == 0);
1592 			/* caller does not want a lock */
1593 			vn_unlock(vp);
1594 		}
1595 	}
1596 	if (error == 0 && vp == NULL)
1597 		error = ENOENT;
1598 	*vpp = vp;
1599 	return(error);
1600 }
1601 
1602 /*
1603  * Return a referenced vnode representing the parent directory of
1604  * ncp.
1605  *
1606  * Because the caller has locked the ncp it should not be possible for
1607  * the parent ncp to go away.  However, the parent can unresolve its
1608  * dvp at any time so we must be able to acquire a lock on the parent
1609  * to safely access nc_vp.
1610  *
1611  * We have to leave par unlocked when vget()ing dvp to avoid a deadlock,
1612  * so use vhold()/vdrop() while holding the lock to prevent dvp from
1613  * getting destroyed.
1614  *
1615  * MPSAFE - Note vhold() is allowed when dvp has 0 refs if we hold a
1616  *	    lock on the ncp in question..
1617  */
1618 static struct vnode *
1619 cache_dvpref(struct namecache *ncp)
1620 {
1621 	struct namecache *par;
1622 	struct vnode *dvp;
1623 
1624 	dvp = NULL;
1625 	if ((par = ncp->nc_parent) != NULL) {
1626 		_cache_hold(par);
1627 		_cache_lock(par);
1628 		if ((par->nc_flag & NCF_UNRESOLVED) == 0) {
1629 			if ((dvp = par->nc_vp) != NULL)
1630 				vhold(dvp);
1631 		}
1632 		_cache_unlock(par);
1633 		if (dvp) {
1634 			if (vget(dvp, LK_SHARED) == 0) {
1635 				vn_unlock(dvp);
1636 				vdrop(dvp);
1637 				/* return refd, unlocked dvp */
1638 			} else {
1639 				vdrop(dvp);
1640 				dvp = NULL;
1641 			}
1642 		}
1643 		_cache_drop(par);
1644 	}
1645 	return(dvp);
1646 }
1647 
1648 /*
1649  * Convert a directory vnode to a namecache record without any other
1650  * knowledge of the topology.  This ONLY works with directory vnodes and
1651  * is ONLY used by the NFS server.  dvp must be refd but unlocked, and the
1652  * returned ncp (if not NULL) will be held and unlocked.
1653  *
1654  * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned.
1655  * If 'makeit' is 1 we attempt to track-down and create the namecache topology
1656  * for dvp.  This will fail only if the directory has been deleted out from
1657  * under the caller.
1658  *
1659  * Callers must always check for a NULL return no matter the value of 'makeit'.
1660  *
1661  * To avoid underflowing the kernel stack each recursive call increments
1662  * the makeit variable.
1663  */
1664 
1665 static int cache_inefficient_scan(struct nchandle *nch, struct ucred *cred,
1666 				  struct vnode *dvp, char *fakename);
1667 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1668 				  struct vnode **saved_dvp);
1669 
1670 int
1671 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit,
1672 	      struct nchandle *nch)
1673 {
1674 	struct vnode *saved_dvp;
1675 	struct vnode *pvp;
1676 	char *fakename;
1677 	int error;
1678 
1679 	nch->ncp = NULL;
1680 	nch->mount = dvp->v_mount;
1681 	saved_dvp = NULL;
1682 	fakename = NULL;
1683 
1684 	/*
1685 	 * Handle the makeit == 0 degenerate case
1686 	 */
1687 	if (makeit == 0) {
1688 		spin_lock(&dvp->v_spin);
1689 		nch->ncp = TAILQ_FIRST(&dvp->v_namecache);
1690 		if (nch->ncp)
1691 			cache_hold(nch);
1692 		spin_unlock(&dvp->v_spin);
1693 	}
1694 
1695 	/*
1696 	 * Loop until resolution, inside code will break out on error.
1697 	 */
1698 	while (makeit) {
1699 		/*
1700 		 * Break out if we successfully acquire a working ncp.
1701 		 */
1702 		spin_lock(&dvp->v_spin);
1703 		nch->ncp = TAILQ_FIRST(&dvp->v_namecache);
1704 		if (nch->ncp) {
1705 			cache_hold(nch);
1706 			spin_unlock(&dvp->v_spin);
1707 			break;
1708 		}
1709 		spin_unlock(&dvp->v_spin);
1710 
1711 		/*
1712 		 * If dvp is the root of its filesystem it should already
1713 		 * have a namecache pointer associated with it as a side
1714 		 * effect of the mount, but it may have been disassociated.
1715 		 */
1716 		if (dvp->v_flag & VROOT) {
1717 			nch->ncp = _cache_get(nch->mount->mnt_ncmountpt.ncp);
1718 			error = cache_resolve_mp(nch->mount);
1719 			_cache_put(nch->ncp);
1720 			if (ncvp_debug) {
1721 				kprintf("cache_fromdvp: resolve root of mount %p error %d",
1722 					dvp->v_mount, error);
1723 			}
1724 			if (error) {
1725 				if (ncvp_debug)
1726 					kprintf(" failed\n");
1727 				nch->ncp = NULL;
1728 				break;
1729 			}
1730 			if (ncvp_debug)
1731 				kprintf(" succeeded\n");
1732 			continue;
1733 		}
1734 
1735 		/*
1736 		 * If we are recursed too deeply resort to an O(n^2)
1737 		 * algorithm to resolve the namecache topology.  The
1738 		 * resolved pvp is left referenced in saved_dvp to
1739 		 * prevent the tree from being destroyed while we loop.
1740 		 */
1741 		if (makeit > 20) {
1742 			error = cache_fromdvp_try(dvp, cred, &saved_dvp);
1743 			if (error) {
1744 				kprintf("lookupdotdot(longpath) failed %d "
1745 				       "dvp %p\n", error, dvp);
1746 				nch->ncp = NULL;
1747 				break;
1748 			}
1749 			continue;
1750 		}
1751 
1752 		/*
1753 		 * Get the parent directory and resolve its ncp.
1754 		 */
1755 		if (fakename) {
1756 			kfree(fakename, M_TEMP);
1757 			fakename = NULL;
1758 		}
1759 		error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred,
1760 					  &fakename);
1761 		if (error) {
1762 			kprintf("lookupdotdot failed %d dvp %p\n", error, dvp);
1763 			break;
1764 		}
1765 		vn_unlock(pvp);
1766 
1767 		/*
1768 		 * Reuse makeit as a recursion depth counter.  On success
1769 		 * nch will be fully referenced.
1770 		 */
1771 		cache_fromdvp(pvp, cred, makeit + 1, nch);
1772 		vrele(pvp);
1773 		if (nch->ncp == NULL)
1774 			break;
1775 
1776 		/*
1777 		 * Do an inefficient scan of pvp (embodied by ncp) to look
1778 		 * for dvp.  This will create a namecache record for dvp on
1779 		 * success.  We loop up to recheck on success.
1780 		 *
1781 		 * ncp and dvp are both held but not locked.
1782 		 */
1783 		error = cache_inefficient_scan(nch, cred, dvp, fakename);
1784 		if (error) {
1785 			kprintf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n",
1786 				pvp, nch->ncp->nc_name, dvp);
1787 			cache_drop(nch);
1788 			/* nch was NULLed out, reload mount */
1789 			nch->mount = dvp->v_mount;
1790 			break;
1791 		}
1792 		if (ncvp_debug) {
1793 			kprintf("cache_fromdvp: scan %p (%s) succeeded\n",
1794 				pvp, nch->ncp->nc_name);
1795 		}
1796 		cache_drop(nch);
1797 		/* nch was NULLed out, reload mount */
1798 		nch->mount = dvp->v_mount;
1799 	}
1800 
1801 	/*
1802 	 * If nch->ncp is non-NULL it will have been held already.
1803 	 */
1804 	if (fakename)
1805 		kfree(fakename, M_TEMP);
1806 	if (saved_dvp)
1807 		vrele(saved_dvp);
1808 	if (nch->ncp)
1809 		return (0);
1810 	return (EINVAL);
1811 }
1812 
1813 /*
1814  * Go up the chain of parent directories until we find something
1815  * we can resolve into the namecache.  This is very inefficient.
1816  */
1817 static
1818 int
1819 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
1820 		  struct vnode **saved_dvp)
1821 {
1822 	struct nchandle nch;
1823 	struct vnode *pvp;
1824 	int error;
1825 	static time_t last_fromdvp_report;
1826 	char *fakename;
1827 
1828 	/*
1829 	 * Loop getting the parent directory vnode until we get something we
1830 	 * can resolve in the namecache.
1831 	 */
1832 	vref(dvp);
1833 	nch.mount = dvp->v_mount;
1834 	nch.ncp = NULL;
1835 	fakename = NULL;
1836 
1837 	for (;;) {
1838 		if (fakename) {
1839 			kfree(fakename, M_TEMP);
1840 			fakename = NULL;
1841 		}
1842 		error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred,
1843 					  &fakename);
1844 		if (error) {
1845 			vrele(dvp);
1846 			break;
1847 		}
1848 		vn_unlock(pvp);
1849 		spin_lock(&pvp->v_spin);
1850 		if ((nch.ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) {
1851 			_cache_hold(nch.ncp);
1852 			spin_unlock(&pvp->v_spin);
1853 			vrele(pvp);
1854 			break;
1855 		}
1856 		spin_unlock(&pvp->v_spin);
1857 		if (pvp->v_flag & VROOT) {
1858 			nch.ncp = _cache_get(pvp->v_mount->mnt_ncmountpt.ncp);
1859 			error = cache_resolve_mp(nch.mount);
1860 			_cache_unlock(nch.ncp);
1861 			vrele(pvp);
1862 			if (error) {
1863 				_cache_drop(nch.ncp);
1864 				nch.ncp = NULL;
1865 				vrele(dvp);
1866 			}
1867 			break;
1868 		}
1869 		vrele(dvp);
1870 		dvp = pvp;
1871 	}
1872 	if (error == 0) {
1873 		if (last_fromdvp_report != time_second) {
1874 			last_fromdvp_report = time_second;
1875 			kprintf("Warning: extremely inefficient path "
1876 				"resolution on %s\n",
1877 				nch.ncp->nc_name);
1878 		}
1879 		error = cache_inefficient_scan(&nch, cred, dvp, fakename);
1880 
1881 		/*
1882 		 * Hopefully dvp now has a namecache record associated with
1883 		 * it.  Leave it referenced to prevent the kernel from
1884 		 * recycling the vnode.  Otherwise extremely long directory
1885 		 * paths could result in endless recycling.
1886 		 */
1887 		if (*saved_dvp)
1888 		    vrele(*saved_dvp);
1889 		*saved_dvp = dvp;
1890 		_cache_drop(nch.ncp);
1891 	}
1892 	if (fakename)
1893 		kfree(fakename, M_TEMP);
1894 	return (error);
1895 }
1896 
1897 /*
1898  * Do an inefficient scan of the directory represented by ncp looking for
1899  * the directory vnode dvp.  ncp must be held but not locked on entry and
1900  * will be held on return.  dvp must be refd but not locked on entry and
1901  * will remain refd on return.
1902  *
1903  * Why do this at all?  Well, due to its stateless nature the NFS server
1904  * converts file handles directly to vnodes without necessarily going through
1905  * the namecache ops that would otherwise create the namecache topology
1906  * leading to the vnode.  We could either (1) Change the namecache algorithms
1907  * to allow disconnect namecache records that are re-merged opportunistically,
1908  * or (2) Make the NFS server backtrack and scan to recover a connected
1909  * namecache topology in order to then be able to issue new API lookups.
1910  *
1911  * It turns out that (1) is a huge mess.  It takes a nice clean set of
1912  * namecache algorithms and introduces a lot of complication in every subsystem
1913  * that calls into the namecache to deal with the re-merge case, especially
1914  * since we are using the namecache to placehold negative lookups and the
1915  * vnode might not be immediately assigned. (2) is certainly far less
1916  * efficient then (1), but since we are only talking about directories here
1917  * (which are likely to remain cached), the case does not actually run all
1918  * that often and has the supreme advantage of not polluting the namecache
1919  * algorithms.
1920  *
1921  * If a fakename is supplied just construct a namecache entry using the
1922  * fake name.
1923  */
1924 static int
1925 cache_inefficient_scan(struct nchandle *nch, struct ucred *cred,
1926 		       struct vnode *dvp, char *fakename)
1927 {
1928 	struct nlcomponent nlc;
1929 	struct nchandle rncp;
1930 	struct dirent *den;
1931 	struct vnode *pvp;
1932 	struct vattr vat;
1933 	struct iovec iov;
1934 	struct uio uio;
1935 	int blksize;
1936 	int eofflag;
1937 	int bytes;
1938 	char *rbuf;
1939 	int error;
1940 
1941 	vat.va_blocksize = 0;
1942 	if ((error = VOP_GETATTR(dvp, &vat)) != 0)
1943 		return (error);
1944 	cache_lock(nch);
1945 	error = cache_vref(nch, cred, &pvp);
1946 	cache_unlock(nch);
1947 	if (error)
1948 		return (error);
1949 	if (ncvp_debug) {
1950 		kprintf("inefficient_scan: directory iosize %ld "
1951 			"vattr fileid = %lld\n",
1952 			vat.va_blocksize,
1953 			(long long)vat.va_fileid);
1954 	}
1955 
1956 	/*
1957 	 * Use the supplied fakename if not NULL.  Fake names are typically
1958 	 * not in the actual filesystem hierarchy.  This is used by HAMMER
1959 	 * to glue @@timestamp recursions together.
1960 	 */
1961 	if (fakename) {
1962 		nlc.nlc_nameptr = fakename;
1963 		nlc.nlc_namelen = strlen(fakename);
1964 		rncp = cache_nlookup(nch, &nlc);
1965 		goto done;
1966 	}
1967 
1968 	if ((blksize = vat.va_blocksize) == 0)
1969 		blksize = DEV_BSIZE;
1970 	rbuf = kmalloc(blksize, M_TEMP, M_WAITOK);
1971 	rncp.ncp = NULL;
1972 
1973 	eofflag = 0;
1974 	uio.uio_offset = 0;
1975 again:
1976 	iov.iov_base = rbuf;
1977 	iov.iov_len = blksize;
1978 	uio.uio_iov = &iov;
1979 	uio.uio_iovcnt = 1;
1980 	uio.uio_resid = blksize;
1981 	uio.uio_segflg = UIO_SYSSPACE;
1982 	uio.uio_rw = UIO_READ;
1983 	uio.uio_td = curthread;
1984 
1985 	if (ncvp_debug >= 2)
1986 		kprintf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset);
1987 	error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL);
1988 	if (error == 0) {
1989 		den = (struct dirent *)rbuf;
1990 		bytes = blksize - uio.uio_resid;
1991 
1992 		while (bytes > 0) {
1993 			if (ncvp_debug >= 2) {
1994 				kprintf("cache_inefficient_scan: %*.*s\n",
1995 					den->d_namlen, den->d_namlen,
1996 					den->d_name);
1997 			}
1998 			if (den->d_type != DT_WHT &&
1999 			    den->d_ino == vat.va_fileid) {
2000 				if (ncvp_debug) {
2001 					kprintf("cache_inefficient_scan: "
2002 					       "MATCHED inode %lld path %s/%*.*s\n",
2003 					       (long long)vat.va_fileid,
2004 					       nch->ncp->nc_name,
2005 					       den->d_namlen, den->d_namlen,
2006 					       den->d_name);
2007 				}
2008 				nlc.nlc_nameptr = den->d_name;
2009 				nlc.nlc_namelen = den->d_namlen;
2010 				rncp = cache_nlookup(nch, &nlc);
2011 				KKASSERT(rncp.ncp != NULL);
2012 				break;
2013 			}
2014 			bytes -= _DIRENT_DIRSIZ(den);
2015 			den = _DIRENT_NEXT(den);
2016 		}
2017 		if (rncp.ncp == NULL && eofflag == 0 && uio.uio_resid != blksize)
2018 			goto again;
2019 	}
2020 	kfree(rbuf, M_TEMP);
2021 done:
2022 	vrele(pvp);
2023 	if (rncp.ncp) {
2024 		if (rncp.ncp->nc_flag & NCF_UNRESOLVED) {
2025 			_cache_setvp(rncp.mount, rncp.ncp, dvp);
2026 			if (ncvp_debug >= 2) {
2027 				kprintf("cache_inefficient_scan: setvp %s/%s = %p\n",
2028 					nch->ncp->nc_name, rncp.ncp->nc_name, dvp);
2029 			}
2030 		} else {
2031 			if (ncvp_debug >= 2) {
2032 				kprintf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n",
2033 					nch->ncp->nc_name, rncp.ncp->nc_name, dvp,
2034 					rncp.ncp->nc_vp);
2035 			}
2036 		}
2037 		if (rncp.ncp->nc_vp == NULL)
2038 			error = rncp.ncp->nc_error;
2039 		/*
2040 		 * Release rncp after a successful nlookup.  rncp was fully
2041 		 * referenced.
2042 		 */
2043 		cache_put(&rncp);
2044 	} else {
2045 		kprintf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n",
2046 			dvp, nch->ncp->nc_name);
2047 		error = ENOENT;
2048 	}
2049 	return (error);
2050 }
2051 
2052 /*
2053  * Zap a namecache entry.  The ncp is unconditionally set to an unresolved
2054  * state, which disassociates it from its vnode or ncneglist.
2055  *
2056  * Then, if there are no additional references to the ncp and no children,
2057  * the ncp is removed from the topology and destroyed.
2058  *
2059  * References and/or children may exist if the ncp is in the middle of the
2060  * topology, preventing the ncp from being destroyed.
2061  *
2062  * This function must be called with the ncp held and locked and will unlock
2063  * and drop it during zapping.
2064  *
2065  * If nonblock is non-zero and the parent ncp cannot be locked we give up.
2066  * This case can occur in the cache_drop() path.
2067  *
2068  * This function may returned a held (but NOT locked) parent node which the
2069  * caller must drop.  We do this so _cache_drop() can loop, to avoid
2070  * blowing out the kernel stack.
2071  *
2072  * WARNING!  For MPSAFE operation this routine must acquire up to three
2073  *	     spin locks to be able to safely test nc_refs.  Lock order is
2074  *	     very important.
2075  *
2076  *	     hash spinlock if on hash list
2077  *	     parent spinlock if child of parent
2078  *	     (the ncp is unresolved so there is no vnode association)
2079  */
2080 static struct namecache *
2081 cache_zap(struct namecache *ncp, int nonblock)
2082 {
2083 	struct namecache *par;
2084 	struct vnode *dropvp;
2085 	int refs;
2086 
2087 	/*
2088 	 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED.
2089 	 */
2090 	_cache_setunresolved(ncp);
2091 
2092 	/*
2093 	 * Try to scrap the entry and possibly tail-recurse on its parent.
2094 	 * We only scrap unref'd (other then our ref) unresolved entries,
2095 	 * we do not scrap 'live' entries.
2096 	 *
2097 	 * Note that once the spinlocks are acquired if nc_refs == 1 no
2098 	 * other references are possible.  If it isn't, however, we have
2099 	 * to decrement but also be sure to avoid a 1->0 transition.
2100 	 */
2101 	KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
2102 	KKASSERT(ncp->nc_refs > 0);
2103 
2104 	/*
2105 	 * Acquire locks.  Note that the parent can't go away while we hold
2106 	 * a child locked.
2107 	 */
2108 	if ((par = ncp->nc_parent) != NULL) {
2109 		if (nonblock) {
2110 			for (;;) {
2111 				if (_cache_lock_nonblock(par) == 0)
2112 					break;
2113 				refs = ncp->nc_refs;
2114 				ncp->nc_flag |= NCF_DEFEREDZAP;
2115 				++numdefered;	/* MP race ok */
2116 				if (atomic_cmpset_int(&ncp->nc_refs,
2117 						      refs, refs - 1)) {
2118 					_cache_unlock(ncp);
2119 					return(NULL);
2120 				}
2121 				cpu_pause();
2122 			}
2123 			_cache_hold(par);
2124 		} else {
2125 			_cache_hold(par);
2126 			_cache_lock(par);
2127 		}
2128 		spin_lock(&ncp->nc_head->spin);
2129 	}
2130 
2131 	/*
2132 	 * If someone other then us has a ref or we have children
2133 	 * we cannot zap the entry.  The 1->0 transition and any
2134 	 * further list operation is protected by the spinlocks
2135 	 * we have acquired but other transitions are not.
2136 	 */
2137 	for (;;) {
2138 		refs = ncp->nc_refs;
2139 		if (refs == 1 && TAILQ_EMPTY(&ncp->nc_list))
2140 			break;
2141 		if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1)) {
2142 			if (par) {
2143 				spin_unlock(&ncp->nc_head->spin);
2144 				_cache_put(par);
2145 			}
2146 			_cache_unlock(ncp);
2147 			return(NULL);
2148 		}
2149 		cpu_pause();
2150 	}
2151 
2152 	/*
2153 	 * We are the only ref and with the spinlocks held no further
2154 	 * refs can be acquired by others.
2155 	 *
2156 	 * Remove us from the hash list and parent list.  We have to
2157 	 * drop a ref on the parent's vp if the parent's list becomes
2158 	 * empty.
2159 	 */
2160 	dropvp = NULL;
2161 	if (par) {
2162 		struct nchash_head *nchpp = ncp->nc_head;
2163 
2164 		KKASSERT(nchpp != NULL);
2165 		LIST_REMOVE(ncp, nc_hash);
2166 		TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
2167 		if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
2168 			dropvp = par->nc_vp;
2169 		ncp->nc_head = NULL;
2170 		ncp->nc_parent = NULL;
2171 		spin_unlock(&nchpp->spin);
2172 		_cache_unlock(par);
2173 	} else {
2174 		KKASSERT(ncp->nc_head == NULL);
2175 	}
2176 
2177 	/*
2178 	 * ncp should not have picked up any refs.  Physically
2179 	 * destroy the ncp.
2180 	 */
2181 	KKASSERT(ncp->nc_refs == 1);
2182 	/* _cache_unlock(ncp) not required */
2183 	ncp->nc_refs = -1;	/* safety */
2184 	if (ncp->nc_name)
2185 		kfree(ncp->nc_name, M_VFSCACHE);
2186 	kfree(ncp, M_VFSCACHE);
2187 
2188 	/*
2189 	 * Delayed drop (we had to release our spinlocks)
2190 	 *
2191 	 * The refed parent (if not  NULL) must be dropped.  The
2192 	 * caller is responsible for looping.
2193 	 */
2194 	if (dropvp)
2195 		vdrop(dropvp);
2196 	return(par);
2197 }
2198 
2199 /*
2200  * Clean up dangling negative cache and defered-drop entries in the
2201  * namecache.
2202  */
2203 typedef enum { CHI_LOW, CHI_HIGH } cache_hs_t;
2204 
2205 static cache_hs_t neg_cache_hysteresis_state = CHI_LOW;
2206 static cache_hs_t pos_cache_hysteresis_state = CHI_LOW;
2207 
2208 void
2209 cache_hysteresis(void)
2210 {
2211 	int poslimit;
2212 
2213 	/*
2214 	 * Don't cache too many negative hits.  We use hysteresis to reduce
2215 	 * the impact on the critical path.
2216 	 */
2217 	switch(neg_cache_hysteresis_state) {
2218 	case CHI_LOW:
2219 		if (numneg > MINNEG && numneg * ncnegfactor > numcache) {
2220 			_cache_cleanneg(10);
2221 			neg_cache_hysteresis_state = CHI_HIGH;
2222 		}
2223 		break;
2224 	case CHI_HIGH:
2225 		if (numneg > MINNEG * 9 / 10 &&
2226 		    numneg * ncnegfactor * 9 / 10 > numcache
2227 		) {
2228 			_cache_cleanneg(10);
2229 		} else {
2230 			neg_cache_hysteresis_state = CHI_LOW;
2231 		}
2232 		break;
2233 	}
2234 
2235 	/*
2236 	 * Don't cache too many positive hits.  We use hysteresis to reduce
2237 	 * the impact on the critical path.
2238 	 *
2239 	 * Excessive positive hits can accumulate due to large numbers of
2240 	 * hardlinks (the vnode cache will not prevent hl ncps from growing
2241 	 * into infinity).
2242 	 */
2243 	if ((poslimit = ncposlimit) == 0)
2244 		poslimit = desiredvnodes * 2;
2245 
2246 	switch(pos_cache_hysteresis_state) {
2247 	case CHI_LOW:
2248 		if (numcache > poslimit && numcache > MINPOS) {
2249 			_cache_cleanpos(10);
2250 			pos_cache_hysteresis_state = CHI_HIGH;
2251 		}
2252 		break;
2253 	case CHI_HIGH:
2254 		if (numcache > poslimit * 5 / 6 && numcache > MINPOS) {
2255 			_cache_cleanpos(10);
2256 		} else {
2257 			pos_cache_hysteresis_state = CHI_LOW;
2258 		}
2259 		break;
2260 	}
2261 
2262 	/*
2263 	 * Clean out dangling defered-zap ncps which could not
2264 	 * be cleanly dropped if too many build up.  Note
2265 	 * that numdefered is not an exact number as such ncps
2266 	 * can be reused and the counter is not handled in a MP
2267 	 * safe manner by design.
2268 	 */
2269 	if (numdefered * ncnegfactor > numcache) {
2270 		_cache_cleandefered();
2271 	}
2272 }
2273 
2274 /*
2275  * NEW NAMECACHE LOOKUP API
2276  *
2277  * Lookup an entry in the namecache.  The passed par_nch must be referenced
2278  * and unlocked.  A referenced and locked nchandle with a non-NULL nch.ncp
2279  * is ALWAYS returned, eve if the supplied component is illegal.
2280  *
2281  * The resulting namecache entry should be returned to the system with
2282  * cache_put() or cache_unlock() + cache_drop().
2283  *
2284  * namecache locks are recursive but care must be taken to avoid lock order
2285  * reversals (hence why the passed par_nch must be unlocked).  Locking
2286  * rules are to order for parent traversals, not for child traversals.
2287  *
2288  * Nobody else will be able to manipulate the associated namespace (e.g.
2289  * create, delete, rename, rename-target) until the caller unlocks the
2290  * entry.
2291  *
2292  * The returned entry will be in one of three states:  positive hit (non-null
2293  * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set).
2294  * Unresolved entries must be resolved through the filesystem to associate the
2295  * vnode and/or determine whether a positive or negative hit has occured.
2296  *
2297  * It is not necessary to lock a directory in order to lock namespace under
2298  * that directory.  In fact, it is explicitly not allowed to do that.  A
2299  * directory is typically only locked when being created, renamed, or
2300  * destroyed.
2301  *
2302  * The directory (par) may be unresolved, in which case any returned child
2303  * will likely also be marked unresolved.  Likely but not guarenteed.  Since
2304  * the filesystem lookup requires a resolved directory vnode the caller is
2305  * responsible for resolving the namecache chain top-down.  This API
2306  * specifically allows whole chains to be created in an unresolved state.
2307  */
2308 struct nchandle
2309 cache_nlookup(struct nchandle *par_nch, struct nlcomponent *nlc)
2310 {
2311 	struct nchandle nch;
2312 	struct namecache *ncp;
2313 	struct namecache *new_ncp;
2314 	struct nchash_head *nchpp;
2315 	struct mount *mp;
2316 	u_int32_t hash;
2317 	globaldata_t gd;
2318 	int par_locked;
2319 
2320 	numcalls++;
2321 	gd = mycpu;
2322 	mp = par_nch->mount;
2323 	par_locked = 0;
2324 
2325 	/*
2326 	 * This is a good time to call it, no ncp's are locked by
2327 	 * the caller or us.
2328 	 */
2329 	cache_hysteresis();
2330 
2331 	/*
2332 	 * Try to locate an existing entry
2333 	 */
2334 	hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
2335 	hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
2336 	new_ncp = NULL;
2337 	nchpp = NCHHASH(hash);
2338 restart:
2339 	spin_lock(&nchpp->spin);
2340 	LIST_FOREACH(ncp, &nchpp->list, nc_hash) {
2341 		numchecks++;
2342 
2343 		/*
2344 		 * Break out if we find a matching entry.  Note that
2345 		 * UNRESOLVED entries may match, but DESTROYED entries
2346 		 * do not.
2347 		 */
2348 		if (ncp->nc_parent == par_nch->ncp &&
2349 		    ncp->nc_nlen == nlc->nlc_namelen &&
2350 		    bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
2351 		    (ncp->nc_flag & NCF_DESTROYED) == 0
2352 		) {
2353 			_cache_hold(ncp);
2354 			spin_unlock(&nchpp->spin);
2355 			if (par_locked) {
2356 				_cache_unlock(par_nch->ncp);
2357 				par_locked = 0;
2358 			}
2359 			if (_cache_lock_special(ncp) == 0) {
2360 				_cache_auto_unresolve(mp, ncp);
2361 				if (new_ncp)
2362 					_cache_free(new_ncp);
2363 				goto found;
2364 			}
2365 			_cache_get(ncp);
2366 			_cache_put(ncp);
2367 			_cache_drop(ncp);
2368 			goto restart;
2369 		}
2370 	}
2371 
2372 	/*
2373 	 * We failed to locate an entry, create a new entry and add it to
2374 	 * the cache.  The parent ncp must also be locked so we
2375 	 * can link into it.
2376 	 *
2377 	 * We have to relookup after possibly blocking in kmalloc or
2378 	 * when locking par_nch.
2379 	 *
2380 	 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special
2381 	 *	 mount case, in which case nc_name will be NULL.
2382 	 */
2383 	if (new_ncp == NULL) {
2384 		spin_unlock(&nchpp->spin);
2385 		new_ncp = cache_alloc(nlc->nlc_namelen);
2386 		if (nlc->nlc_namelen) {
2387 			bcopy(nlc->nlc_nameptr, new_ncp->nc_name,
2388 			      nlc->nlc_namelen);
2389 			new_ncp->nc_name[nlc->nlc_namelen] = 0;
2390 		}
2391 		goto restart;
2392 	}
2393 	if (par_locked == 0) {
2394 		spin_unlock(&nchpp->spin);
2395 		_cache_lock(par_nch->ncp);
2396 		par_locked = 1;
2397 		goto restart;
2398 	}
2399 
2400 	/*
2401 	 * WARNING!  We still hold the spinlock.  We have to set the hash
2402 	 *	     table entry atomically.
2403 	 */
2404 	ncp = new_ncp;
2405 	_cache_link_parent(ncp, par_nch->ncp, nchpp);
2406 	spin_unlock(&nchpp->spin);
2407 	_cache_unlock(par_nch->ncp);
2408 	/* par_locked = 0 - not used */
2409 found:
2410 	/*
2411 	 * stats and namecache size management
2412 	 */
2413 	if (ncp->nc_flag & NCF_UNRESOLVED)
2414 		++gd->gd_nchstats->ncs_miss;
2415 	else if (ncp->nc_vp)
2416 		++gd->gd_nchstats->ncs_goodhits;
2417 	else
2418 		++gd->gd_nchstats->ncs_neghits;
2419 	nch.mount = mp;
2420 	nch.ncp = ncp;
2421 	atomic_add_int(&nch.mount->mnt_refs, 1);
2422 	return(nch);
2423 }
2424 
2425 /*
2426  * This is a non-blocking verison of cache_nlookup() used by
2427  * nfs_readdirplusrpc_uio().  It can fail for any reason and
2428  * will return nch.ncp == NULL in that case.
2429  */
2430 struct nchandle
2431 cache_nlookup_nonblock(struct nchandle *par_nch, struct nlcomponent *nlc)
2432 {
2433 	struct nchandle nch;
2434 	struct namecache *ncp;
2435 	struct namecache *new_ncp;
2436 	struct nchash_head *nchpp;
2437 	struct mount *mp;
2438 	u_int32_t hash;
2439 	globaldata_t gd;
2440 	int par_locked;
2441 
2442 	numcalls++;
2443 	gd = mycpu;
2444 	mp = par_nch->mount;
2445 	par_locked = 0;
2446 
2447 	/*
2448 	 * Try to locate an existing entry
2449 	 */
2450 	hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
2451 	hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
2452 	new_ncp = NULL;
2453 	nchpp = NCHHASH(hash);
2454 restart:
2455 	spin_lock(&nchpp->spin);
2456 	LIST_FOREACH(ncp, &nchpp->list, nc_hash) {
2457 		numchecks++;
2458 
2459 		/*
2460 		 * Break out if we find a matching entry.  Note that
2461 		 * UNRESOLVED entries may match, but DESTROYED entries
2462 		 * do not.
2463 		 */
2464 		if (ncp->nc_parent == par_nch->ncp &&
2465 		    ncp->nc_nlen == nlc->nlc_namelen &&
2466 		    bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
2467 		    (ncp->nc_flag & NCF_DESTROYED) == 0
2468 		) {
2469 			_cache_hold(ncp);
2470 			spin_unlock(&nchpp->spin);
2471 			if (par_locked) {
2472 				_cache_unlock(par_nch->ncp);
2473 				par_locked = 0;
2474 			}
2475 			if (_cache_lock_special(ncp) == 0) {
2476 				_cache_auto_unresolve(mp, ncp);
2477 				if (new_ncp) {
2478 					_cache_free(new_ncp);
2479 					new_ncp = NULL;
2480 				}
2481 				goto found;
2482 			}
2483 			_cache_drop(ncp);
2484 			goto failed;
2485 		}
2486 	}
2487 
2488 	/*
2489 	 * We failed to locate an entry, create a new entry and add it to
2490 	 * the cache.  The parent ncp must also be locked so we
2491 	 * can link into it.
2492 	 *
2493 	 * We have to relookup after possibly blocking in kmalloc or
2494 	 * when locking par_nch.
2495 	 *
2496 	 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special
2497 	 *	 mount case, in which case nc_name will be NULL.
2498 	 */
2499 	if (new_ncp == NULL) {
2500 		spin_unlock(&nchpp->spin);
2501 		new_ncp = cache_alloc(nlc->nlc_namelen);
2502 		if (nlc->nlc_namelen) {
2503 			bcopy(nlc->nlc_nameptr, new_ncp->nc_name,
2504 			      nlc->nlc_namelen);
2505 			new_ncp->nc_name[nlc->nlc_namelen] = 0;
2506 		}
2507 		goto restart;
2508 	}
2509 	if (par_locked == 0) {
2510 		spin_unlock(&nchpp->spin);
2511 		if (_cache_lock_nonblock(par_nch->ncp) == 0) {
2512 			par_locked = 1;
2513 			goto restart;
2514 		}
2515 		goto failed;
2516 	}
2517 
2518 	/*
2519 	 * WARNING!  We still hold the spinlock.  We have to set the hash
2520 	 *	     table entry atomically.
2521 	 */
2522 	ncp = new_ncp;
2523 	_cache_link_parent(ncp, par_nch->ncp, nchpp);
2524 	spin_unlock(&nchpp->spin);
2525 	_cache_unlock(par_nch->ncp);
2526 	/* par_locked = 0 - not used */
2527 found:
2528 	/*
2529 	 * stats and namecache size management
2530 	 */
2531 	if (ncp->nc_flag & NCF_UNRESOLVED)
2532 		++gd->gd_nchstats->ncs_miss;
2533 	else if (ncp->nc_vp)
2534 		++gd->gd_nchstats->ncs_goodhits;
2535 	else
2536 		++gd->gd_nchstats->ncs_neghits;
2537 	nch.mount = mp;
2538 	nch.ncp = ncp;
2539 	atomic_add_int(&nch.mount->mnt_refs, 1);
2540 	return(nch);
2541 failed:
2542 	if (new_ncp) {
2543 		_cache_free(new_ncp);
2544 		new_ncp = NULL;
2545 	}
2546 	nch.mount = NULL;
2547 	nch.ncp = NULL;
2548 	return(nch);
2549 }
2550 
2551 /*
2552  * The namecache entry is marked as being used as a mount point.
2553  * Locate the mount if it is visible to the caller.  The DragonFly
2554  * mount system allows arbitrary loops in the topology and disentangles
2555  * those loops by matching against (mp, ncp) rather than just (ncp).
2556  * This means any given ncp can dive any number of mounts, depending
2557  * on the relative mount (e.g. nullfs) the caller is at in the topology.
2558  *
2559  * We use a very simple frontend cache to reduce SMP conflicts,
2560  * which we have to do because the mountlist scan needs an exclusive
2561  * lock around its ripout info list.  Not to mention that there might
2562  * be a lot of mounts.
2563  */
2564 struct findmount_info {
2565 	struct mount *result;
2566 	struct mount *nch_mount;
2567 	struct namecache *nch_ncp;
2568 };
2569 
2570 static
2571 struct ncmount_cache *
2572 ncmount_cache_lookup(struct mount *mp, struct namecache *ncp)
2573 {
2574 	int hash;
2575 
2576 	hash = ((int)(intptr_t)mp / sizeof(*mp)) ^
2577 	       ((int)(intptr_t)ncp / sizeof(*ncp));
2578 	hash = (hash & 0x7FFFFFFF) % NCMOUNT_NUMCACHE;
2579 	return (&ncmount_cache[hash]);
2580 }
2581 
2582 static
2583 int
2584 cache_findmount_callback(struct mount *mp, void *data)
2585 {
2586 	struct findmount_info *info = data;
2587 
2588 	/*
2589 	 * Check the mount's mounted-on point against the passed nch.
2590 	 */
2591 	if (mp->mnt_ncmounton.mount == info->nch_mount &&
2592 	    mp->mnt_ncmounton.ncp == info->nch_ncp
2593 	) {
2594 	    info->result = mp;
2595 	    atomic_add_int(&mp->mnt_refs, 1);
2596 	    return(-1);
2597 	}
2598 	return(0);
2599 }
2600 
2601 struct mount *
2602 cache_findmount(struct nchandle *nch)
2603 {
2604 	struct findmount_info info;
2605 	struct ncmount_cache *ncc;
2606 	struct mount *mp;
2607 
2608 	/*
2609 	 * Fast
2610 	 */
2611 	if (ncmount_cache_enable == 0) {
2612 		ncc = NULL;
2613 		goto skip;
2614 	}
2615 	ncc = ncmount_cache_lookup(nch->mount, nch->ncp);
2616 	if (ncc->ncp == nch->ncp) {
2617 		spin_lock_shared(&ncc->spin);
2618 		if (ncc->isneg == 0 &&
2619 		    ncc->ncp == nch->ncp && (mp = ncc->mp) != NULL) {
2620 			if (mp->mnt_ncmounton.mount == nch->mount &&
2621 			    mp->mnt_ncmounton.ncp == nch->ncp) {
2622 				/*
2623 				 * Cache hit (positive)
2624 				 */
2625 				atomic_add_int(&mp->mnt_refs, 1);
2626 				spin_unlock_shared(&ncc->spin);
2627 				++ncmount_cache_hit;
2628 				return(mp);
2629 			}
2630 			/* else cache miss */
2631 		}
2632 		if (ncc->isneg &&
2633 		    ncc->ncp == nch->ncp && ncc->mp == nch->mount) {
2634 			/*
2635 			 * Cache hit (negative)
2636 			 */
2637 			spin_unlock_shared(&ncc->spin);
2638 			++ncmount_cache_hit;
2639 			return(NULL);
2640 		}
2641 		spin_unlock_shared(&ncc->spin);
2642 	}
2643 skip:
2644 
2645 	/*
2646 	 * Slow
2647 	 */
2648 	info.result = NULL;
2649 	info.nch_mount = nch->mount;
2650 	info.nch_ncp = nch->ncp;
2651 	mountlist_scan(cache_findmount_callback, &info,
2652 			       MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
2653 
2654 	/*
2655 	 * Cache the result.
2656 	 *
2657 	 * Negative lookups: We cache the originating {ncp,mp}. (mp) is
2658 	 *		     only used for pointer comparisons and is not
2659 	 *		     referenced (otherwise there would be dangling
2660 	 *		     refs).
2661 	 *
2662 	 * Positive lookups: We cache the originating {ncp} and the target
2663 	 *		     (mp).  (mp) is referenced.
2664 	 *
2665 	 * Indeterminant:    If the match is undergoing an unmount we do
2666 	 *		     not cache it to avoid racing cache_unmounting(),
2667 	 *		     but still return the match.
2668 	 */
2669 	if (ncc) {
2670 		spin_lock(&ncc->spin);
2671 		if (info.result == NULL) {
2672 			if (ncc->isneg == 0 && ncc->mp)
2673 				atomic_add_int(&ncc->mp->mnt_refs, -1);
2674 			ncc->ncp = nch->ncp;
2675 			ncc->mp = nch->mount;
2676 			ncc->isneg = 1;
2677 			spin_unlock(&ncc->spin);
2678 			++ncmount_cache_overwrite;
2679 		} else if ((info.result->mnt_kern_flag & MNTK_UNMOUNT) == 0) {
2680 			if (ncc->isneg == 0 && ncc->mp)
2681 				atomic_add_int(&ncc->mp->mnt_refs, -1);
2682 			atomic_add_int(&info.result->mnt_refs, 1);
2683 			ncc->ncp = nch->ncp;
2684 			ncc->mp = info.result;
2685 			ncc->isneg = 0;
2686 			spin_unlock(&ncc->spin);
2687 			++ncmount_cache_overwrite;
2688 		} else {
2689 			spin_unlock(&ncc->spin);
2690 		}
2691 		++ncmount_cache_miss;
2692 	}
2693 	return(info.result);
2694 }
2695 
2696 void
2697 cache_dropmount(struct mount *mp)
2698 {
2699 	atomic_add_int(&mp->mnt_refs, -1);
2700 }
2701 
2702 void
2703 cache_ismounting(struct mount *mp)
2704 {
2705 	struct nchandle *nch = &mp->mnt_ncmounton;
2706 	struct ncmount_cache *ncc;
2707 
2708 	ncc = ncmount_cache_lookup(nch->mount, nch->ncp);
2709 	if (ncc->isneg &&
2710 	    ncc->ncp == nch->ncp && ncc->mp == nch->mount) {
2711 		spin_lock(&ncc->spin);
2712 		if (ncc->isneg &&
2713 		    ncc->ncp == nch->ncp && ncc->mp == nch->mount) {
2714 			ncc->ncp = NULL;
2715 			ncc->mp = NULL;
2716 		}
2717 		spin_unlock(&ncc->spin);
2718 	}
2719 }
2720 
2721 void
2722 cache_unmounting(struct mount *mp)
2723 {
2724 	struct nchandle *nch = &mp->mnt_ncmounton;
2725 	struct ncmount_cache *ncc;
2726 
2727 	ncc = ncmount_cache_lookup(nch->mount, nch->ncp);
2728 	if (ncc->isneg == 0 &&
2729 	    ncc->ncp == nch->ncp && ncc->mp == mp) {
2730 		spin_lock(&ncc->spin);
2731 		if (ncc->isneg == 0 &&
2732 		    ncc->ncp == nch->ncp && ncc->mp == mp) {
2733 			atomic_add_int(&mp->mnt_refs, -1);
2734 			ncc->ncp = NULL;
2735 			ncc->mp = NULL;
2736 		}
2737 		spin_unlock(&ncc->spin);
2738 	}
2739 }
2740 
2741 /*
2742  * Resolve an unresolved namecache entry, generally by looking it up.
2743  * The passed ncp must be locked and refd.
2744  *
2745  * Theoretically since a vnode cannot be recycled while held, and since
2746  * the nc_parent chain holds its vnode as long as children exist, the
2747  * direct parent of the cache entry we are trying to resolve should
2748  * have a valid vnode.  If not then generate an error that we can
2749  * determine is related to a resolver bug.
2750  *
2751  * However, if a vnode was in the middle of a recyclement when the NCP
2752  * got locked, ncp->nc_vp might point to a vnode that is about to become
2753  * invalid.  cache_resolve() handles this case by unresolving the entry
2754  * and then re-resolving it.
2755  *
2756  * Note that successful resolution does not necessarily return an error
2757  * code of 0.  If the ncp resolves to a negative cache hit then ENOENT
2758  * will be returned.
2759  *
2760  * MPSAFE
2761  */
2762 int
2763 cache_resolve(struct nchandle *nch, struct ucred *cred)
2764 {
2765 	struct namecache *par_tmp;
2766 	struct namecache *par;
2767 	struct namecache *ncp;
2768 	struct nchandle nctmp;
2769 	struct mount *mp;
2770 	struct vnode *dvp;
2771 	int error;
2772 
2773 	ncp = nch->ncp;
2774 	mp = nch->mount;
2775 restart:
2776 	/*
2777 	 * If the ncp is already resolved we have nothing to do.  However,
2778 	 * we do want to guarentee that a usable vnode is returned when
2779 	 * a vnode is present, so make sure it hasn't been reclaimed.
2780 	 */
2781 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
2782 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
2783 			_cache_setunresolved(ncp);
2784 		if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
2785 			return (ncp->nc_error);
2786 	}
2787 
2788 	/*
2789 	 * If the ncp was destroyed it will never resolve again.  This
2790 	 * can basically only happen when someone is chdir'd into an
2791 	 * empty directory which is then rmdir'd.  We want to catch this
2792 	 * here and not dive the VFS because the VFS might actually
2793 	 * have a way to re-resolve the disconnected ncp, which will
2794 	 * result in inconsistencies in the cdir/nch for proc->p_fd.
2795 	 */
2796 	if (ncp->nc_flag & NCF_DESTROYED) {
2797 		kprintf("Warning: cache_resolve: ncp '%s' was unlinked\n",
2798 			ncp->nc_name);
2799 		return(EINVAL);
2800 	}
2801 
2802 	/*
2803 	 * Mount points need special handling because the parent does not
2804 	 * belong to the same filesystem as the ncp.
2805 	 */
2806 	if (ncp == mp->mnt_ncmountpt.ncp)
2807 		return (cache_resolve_mp(mp));
2808 
2809 	/*
2810 	 * We expect an unbroken chain of ncps to at least the mount point,
2811 	 * and even all the way to root (but this code doesn't have to go
2812 	 * past the mount point).
2813 	 */
2814 	if (ncp->nc_parent == NULL) {
2815 		kprintf("EXDEV case 1 %p %*.*s\n", ncp,
2816 			ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
2817 		ncp->nc_error = EXDEV;
2818 		return(ncp->nc_error);
2819 	}
2820 
2821 	/*
2822 	 * The vp's of the parent directories in the chain are held via vhold()
2823 	 * due to the existance of the child, and should not disappear.
2824 	 * However, there are cases where they can disappear:
2825 	 *
2826 	 *	- due to filesystem I/O errors.
2827 	 *	- due to NFS being stupid about tracking the namespace and
2828 	 *	  destroys the namespace for entire directories quite often.
2829 	 *	- due to forced unmounts.
2830 	 *	- due to an rmdir (parent will be marked DESTROYED)
2831 	 *
2832 	 * When this occurs we have to track the chain backwards and resolve
2833 	 * it, looping until the resolver catches up to the current node.  We
2834 	 * could recurse here but we might run ourselves out of kernel stack
2835 	 * so we do it in a more painful manner.  This situation really should
2836 	 * not occur all that often, or if it does not have to go back too
2837 	 * many nodes to resolve the ncp.
2838 	 */
2839 	while ((dvp = cache_dvpref(ncp)) == NULL) {
2840 		/*
2841 		 * This case can occur if a process is CD'd into a
2842 		 * directory which is then rmdir'd.  If the parent is marked
2843 		 * destroyed there is no point trying to resolve it.
2844 		 */
2845 		if (ncp->nc_parent->nc_flag & NCF_DESTROYED)
2846 			return(ENOENT);
2847 		par = ncp->nc_parent;
2848 		_cache_hold(par);
2849 		_cache_lock(par);
2850 		while ((par_tmp = par->nc_parent) != NULL &&
2851 		       par_tmp->nc_vp == NULL) {
2852 			_cache_hold(par_tmp);
2853 			_cache_lock(par_tmp);
2854 			_cache_put(par);
2855 			par = par_tmp;
2856 		}
2857 		if (par->nc_parent == NULL) {
2858 			kprintf("EXDEV case 2 %*.*s\n",
2859 				par->nc_nlen, par->nc_nlen, par->nc_name);
2860 			_cache_put(par);
2861 			return (EXDEV);
2862 		}
2863 		kprintf("[diagnostic] cache_resolve: had to recurse on %*.*s\n",
2864 			par->nc_nlen, par->nc_nlen, par->nc_name);
2865 		/*
2866 		 * The parent is not set in stone, ref and lock it to prevent
2867 		 * it from disappearing.  Also note that due to renames it
2868 		 * is possible for our ncp to move and for par to no longer
2869 		 * be one of its parents.  We resolve it anyway, the loop
2870 		 * will handle any moves.
2871 		 */
2872 		_cache_get(par);	/* additional hold/lock */
2873 		_cache_put(par);	/* from earlier hold/lock */
2874 		if (par == nch->mount->mnt_ncmountpt.ncp) {
2875 			cache_resolve_mp(nch->mount);
2876 		} else if ((dvp = cache_dvpref(par)) == NULL) {
2877 			kprintf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name);
2878 			_cache_put(par);
2879 			continue;
2880 		} else {
2881 			if (par->nc_flag & NCF_UNRESOLVED) {
2882 				nctmp.mount = mp;
2883 				nctmp.ncp = par;
2884 				par->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred);
2885 			}
2886 			vrele(dvp);
2887 		}
2888 		if ((error = par->nc_error) != 0) {
2889 			if (par->nc_error != EAGAIN) {
2890 				kprintf("EXDEV case 3 %*.*s error %d\n",
2891 				    par->nc_nlen, par->nc_nlen, par->nc_name,
2892 				    par->nc_error);
2893 				_cache_put(par);
2894 				return(error);
2895 			}
2896 			kprintf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n",
2897 				par, par->nc_nlen, par->nc_nlen, par->nc_name);
2898 		}
2899 		_cache_put(par);
2900 		/* loop */
2901 	}
2902 
2903 	/*
2904 	 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected
2905 	 * ncp's and reattach them.  If this occurs the original ncp is marked
2906 	 * EAGAIN to force a relookup.
2907 	 *
2908 	 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed
2909 	 * ncp must already be resolved.
2910 	 */
2911 	if (dvp) {
2912 		nctmp.mount = mp;
2913 		nctmp.ncp = ncp;
2914 		ncp->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred);
2915 		vrele(dvp);
2916 	} else {
2917 		ncp->nc_error = EPERM;
2918 	}
2919 	if (ncp->nc_error == EAGAIN) {
2920 		kprintf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n",
2921 			ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
2922 		goto restart;
2923 	}
2924 	return(ncp->nc_error);
2925 }
2926 
2927 /*
2928  * Resolve the ncp associated with a mount point.  Such ncp's almost always
2929  * remain resolved and this routine is rarely called.  NFS MPs tends to force
2930  * re-resolution more often due to its mac-truck-smash-the-namecache
2931  * method of tracking namespace changes.
2932  *
2933  * The semantics for this call is that the passed ncp must be locked on
2934  * entry and will be locked on return.  However, if we actually have to
2935  * resolve the mount point we temporarily unlock the entry in order to
2936  * avoid race-to-root deadlocks due to e.g. dead NFS mounts.  Because of
2937  * the unlock we have to recheck the flags after we relock.
2938  */
2939 static int
2940 cache_resolve_mp(struct mount *mp)
2941 {
2942 	struct namecache *ncp = mp->mnt_ncmountpt.ncp;
2943 	struct vnode *vp;
2944 	int error;
2945 
2946 	KKASSERT(mp != NULL);
2947 
2948 	/*
2949 	 * If the ncp is already resolved we have nothing to do.  However,
2950 	 * we do want to guarentee that a usable vnode is returned when
2951 	 * a vnode is present, so make sure it hasn't been reclaimed.
2952 	 */
2953 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
2954 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
2955 			_cache_setunresolved(ncp);
2956 	}
2957 
2958 	if (ncp->nc_flag & NCF_UNRESOLVED) {
2959 		_cache_unlock(ncp);
2960 		while (vfs_busy(mp, 0))
2961 			;
2962 		error = VFS_ROOT(mp, &vp);
2963 		_cache_lock(ncp);
2964 
2965 		/*
2966 		 * recheck the ncp state after relocking.
2967 		 */
2968 		if (ncp->nc_flag & NCF_UNRESOLVED) {
2969 			ncp->nc_error = error;
2970 			if (error == 0) {
2971 				_cache_setvp(mp, ncp, vp);
2972 				vput(vp);
2973 			} else {
2974 				kprintf("[diagnostic] cache_resolve_mp: failed"
2975 					" to resolve mount %p err=%d ncp=%p\n",
2976 					mp, error, ncp);
2977 				_cache_setvp(mp, ncp, NULL);
2978 			}
2979 		} else if (error == 0) {
2980 			vput(vp);
2981 		}
2982 		vfs_unbusy(mp);
2983 	}
2984 	return(ncp->nc_error);
2985 }
2986 
2987 /*
2988  * Clean out negative cache entries when too many have accumulated.
2989  *
2990  * MPSAFE
2991  */
2992 static void
2993 _cache_cleanneg(int count)
2994 {
2995 	struct namecache *ncp;
2996 
2997 	/*
2998 	 * Attempt to clean out the specified number of negative cache
2999 	 * entries.
3000 	 */
3001 	while (count) {
3002 		spin_lock(&ncspin);
3003 		ncp = TAILQ_FIRST(&ncneglist);
3004 		if (ncp == NULL) {
3005 			spin_unlock(&ncspin);
3006 			break;
3007 		}
3008 		TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
3009 		TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
3010 		_cache_hold(ncp);
3011 		spin_unlock(&ncspin);
3012 
3013 		/*
3014 		 * This can race, so we must re-check that the ncp
3015 		 * is on the ncneglist after successfully locking it.
3016 		 */
3017 		if (_cache_lock_special(ncp) == 0) {
3018 			if (ncp->nc_vp == NULL &&
3019 			    (ncp->nc_flag & NCF_UNRESOLVED) == 0) {
3020 				ncp = cache_zap(ncp, 1);
3021 				if (ncp)
3022 					_cache_drop(ncp);
3023 			} else {
3024 				kprintf("cache_cleanneg: race avoided\n");
3025 				_cache_unlock(ncp);
3026 			}
3027 		} else {
3028 			_cache_drop(ncp);
3029 		}
3030 		--count;
3031 	}
3032 }
3033 
3034 /*
3035  * Clean out positive cache entries when too many have accumulated.
3036  *
3037  * MPSAFE
3038  */
3039 static void
3040 _cache_cleanpos(int count)
3041 {
3042 	static volatile int rover;
3043 	struct nchash_head *nchpp;
3044 	struct namecache *ncp;
3045 	int rover_copy;
3046 
3047 	/*
3048 	 * Attempt to clean out the specified number of negative cache
3049 	 * entries.
3050 	 */
3051 	while (count) {
3052 		rover_copy = ++rover;	/* MPSAFEENOUGH */
3053 		cpu_ccfence();
3054 		nchpp = NCHHASH(rover_copy);
3055 
3056 		spin_lock(&nchpp->spin);
3057 		ncp = LIST_FIRST(&nchpp->list);
3058 		if (ncp)
3059 			_cache_hold(ncp);
3060 		spin_unlock(&nchpp->spin);
3061 
3062 		if (ncp) {
3063 			if (_cache_lock_special(ncp) == 0) {
3064 				ncp = cache_zap(ncp, 1);
3065 				if (ncp)
3066 					_cache_drop(ncp);
3067 			} else {
3068 				_cache_drop(ncp);
3069 			}
3070 		}
3071 		--count;
3072 	}
3073 }
3074 
3075 /*
3076  * This is a kitchen sink function to clean out ncps which we
3077  * tried to zap from cache_drop() but failed because we were
3078  * unable to acquire the parent lock.
3079  *
3080  * Such entries can also be removed via cache_inval_vp(), such
3081  * as when unmounting.
3082  *
3083  * MPSAFE
3084  */
3085 static void
3086 _cache_cleandefered(void)
3087 {
3088 	struct nchash_head *nchpp;
3089 	struct namecache *ncp;
3090 	struct namecache dummy;
3091 	int i;
3092 
3093 	numdefered = 0;
3094 	bzero(&dummy, sizeof(dummy));
3095 	dummy.nc_flag = NCF_DESTROYED;
3096 
3097 	for (i = 0; i <= nchash; ++i) {
3098 		nchpp = &nchashtbl[i];
3099 
3100 		spin_lock(&nchpp->spin);
3101 		LIST_INSERT_HEAD(&nchpp->list, &dummy, nc_hash);
3102 		ncp = &dummy;
3103 		while ((ncp = LIST_NEXT(ncp, nc_hash)) != NULL) {
3104 			if ((ncp->nc_flag & NCF_DEFEREDZAP) == 0)
3105 				continue;
3106 			LIST_REMOVE(&dummy, nc_hash);
3107 			LIST_INSERT_AFTER(ncp, &dummy, nc_hash);
3108 			_cache_hold(ncp);
3109 			spin_unlock(&nchpp->spin);
3110 			if (_cache_lock_nonblock(ncp) == 0) {
3111 				ncp->nc_flag &= ~NCF_DEFEREDZAP;
3112 				_cache_unlock(ncp);
3113 			}
3114 			_cache_drop(ncp);
3115 			spin_lock(&nchpp->spin);
3116 			ncp = &dummy;
3117 		}
3118 		LIST_REMOVE(&dummy, nc_hash);
3119 		spin_unlock(&nchpp->spin);
3120 	}
3121 }
3122 
3123 /*
3124  * Name cache initialization, from vfsinit() when we are booting
3125  */
3126 void
3127 nchinit(void)
3128 {
3129 	int i;
3130 	globaldata_t gd;
3131 
3132 	/* initialise per-cpu namecache effectiveness statistics. */
3133 	for (i = 0; i < ncpus; ++i) {
3134 		gd = globaldata_find(i);
3135 		gd->gd_nchstats = &nchstats[i];
3136 	}
3137 	TAILQ_INIT(&ncneglist);
3138 	spin_init(&ncspin);
3139 	nchashtbl = hashinit_ext(desiredvnodes / 2,
3140 				 sizeof(struct nchash_head),
3141 				 M_VFSCACHE, &nchash);
3142 	for (i = 0; i <= (int)nchash; ++i) {
3143 		LIST_INIT(&nchashtbl[i].list);
3144 		spin_init(&nchashtbl[i].spin);
3145 	}
3146 	for (i = 0; i < NCMOUNT_NUMCACHE; ++i)
3147 		spin_init(&ncmount_cache[i].spin);
3148 	nclockwarn = 5 * hz;
3149 }
3150 
3151 /*
3152  * Called from start_init() to bootstrap the root filesystem.  Returns
3153  * a referenced, unlocked namecache record.
3154  */
3155 void
3156 cache_allocroot(struct nchandle *nch, struct mount *mp, struct vnode *vp)
3157 {
3158 	nch->ncp = cache_alloc(0);
3159 	nch->mount = mp;
3160 	atomic_add_int(&mp->mnt_refs, 1);
3161 	if (vp)
3162 		_cache_setvp(nch->mount, nch->ncp, vp);
3163 }
3164 
3165 /*
3166  * vfs_cache_setroot()
3167  *
3168  *	Create an association between the root of our namecache and
3169  *	the root vnode.  This routine may be called several times during
3170  *	booting.
3171  *
3172  *	If the caller intends to save the returned namecache pointer somewhere
3173  *	it must cache_hold() it.
3174  */
3175 void
3176 vfs_cache_setroot(struct vnode *nvp, struct nchandle *nch)
3177 {
3178 	struct vnode *ovp;
3179 	struct nchandle onch;
3180 
3181 	ovp = rootvnode;
3182 	onch = rootnch;
3183 	rootvnode = nvp;
3184 	if (nch)
3185 		rootnch = *nch;
3186 	else
3187 		cache_zero(&rootnch);
3188 	if (ovp)
3189 		vrele(ovp);
3190 	if (onch.ncp)
3191 		cache_drop(&onch);
3192 }
3193 
3194 /*
3195  * XXX OLD API COMPAT FUNCTION.  This really messes up the new namecache
3196  * topology and is being removed as quickly as possible.  The new VOP_N*()
3197  * API calls are required to make specific adjustments using the supplied
3198  * ncp pointers rather then just bogusly purging random vnodes.
3199  *
3200  * Invalidate all namecache entries to a particular vnode as well as
3201  * any direct children of that vnode in the namecache.  This is a
3202  * 'catch all' purge used by filesystems that do not know any better.
3203  *
3204  * Note that the linkage between the vnode and its namecache entries will
3205  * be removed, but the namecache entries themselves might stay put due to
3206  * active references from elsewhere in the system or due to the existance of
3207  * the children.   The namecache topology is left intact even if we do not
3208  * know what the vnode association is.  Such entries will be marked
3209  * NCF_UNRESOLVED.
3210  */
3211 void
3212 cache_purge(struct vnode *vp)
3213 {
3214 	cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN);
3215 }
3216 
3217 /*
3218  * Flush all entries referencing a particular filesystem.
3219  *
3220  * Since we need to check it anyway, we will flush all the invalid
3221  * entries at the same time.
3222  */
3223 #if 0
3224 
3225 void
3226 cache_purgevfs(struct mount *mp)
3227 {
3228 	struct nchash_head *nchpp;
3229 	struct namecache *ncp, *nnp;
3230 
3231 	/*
3232 	 * Scan hash tables for applicable entries.
3233 	 */
3234 	for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) {
3235 		spin_lock_wr(&nchpp->spin); XXX
3236 		ncp = LIST_FIRST(&nchpp->list);
3237 		if (ncp)
3238 			_cache_hold(ncp);
3239 		while (ncp) {
3240 			nnp = LIST_NEXT(ncp, nc_hash);
3241 			if (nnp)
3242 				_cache_hold(nnp);
3243 			if (ncp->nc_mount == mp) {
3244 				_cache_lock(ncp);
3245 				ncp = cache_zap(ncp, 0);
3246 				if (ncp)
3247 					_cache_drop(ncp);
3248 			} else {
3249 				_cache_drop(ncp);
3250 			}
3251 			ncp = nnp;
3252 		}
3253 		spin_unlock_wr(&nchpp->spin); XXX
3254 	}
3255 }
3256 
3257 #endif
3258 
3259 static int disablecwd;
3260 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
3261     "Disable getcwd");
3262 
3263 static u_long numcwdcalls;
3264 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdcalls, CTLFLAG_RD, &numcwdcalls, 0,
3265     "Number of current directory resolution calls");
3266 static u_long numcwdfailnf;
3267 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailnf, CTLFLAG_RD, &numcwdfailnf, 0,
3268     "Number of current directory failures due to lack of file");
3269 static u_long numcwdfailsz;
3270 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailsz, CTLFLAG_RD, &numcwdfailsz, 0,
3271     "Number of current directory failures due to large result");
3272 static u_long numcwdfound;
3273 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfound, CTLFLAG_RD, &numcwdfound, 0,
3274     "Number of current directory resolution successes");
3275 
3276 /*
3277  * MPALMOSTSAFE
3278  */
3279 int
3280 sys___getcwd(struct __getcwd_args *uap)
3281 {
3282 	u_int buflen;
3283 	int error;
3284 	char *buf;
3285 	char *bp;
3286 
3287 	if (disablecwd)
3288 		return (ENODEV);
3289 
3290 	buflen = uap->buflen;
3291 	if (buflen == 0)
3292 		return (EINVAL);
3293 	if (buflen > MAXPATHLEN)
3294 		buflen = MAXPATHLEN;
3295 
3296 	buf = kmalloc(buflen, M_TEMP, M_WAITOK);
3297 	bp = kern_getcwd(buf, buflen, &error);
3298 	if (error == 0)
3299 		error = copyout(bp, uap->buf, strlen(bp) + 1);
3300 	kfree(buf, M_TEMP);
3301 	return (error);
3302 }
3303 
3304 char *
3305 kern_getcwd(char *buf, size_t buflen, int *error)
3306 {
3307 	struct proc *p = curproc;
3308 	char *bp;
3309 	int i, slash_prefixed;
3310 	struct filedesc *fdp;
3311 	struct nchandle nch;
3312 	struct namecache *ncp;
3313 
3314 	numcwdcalls++;
3315 	bp = buf;
3316 	bp += buflen - 1;
3317 	*bp = '\0';
3318 	fdp = p->p_fd;
3319 	slash_prefixed = 0;
3320 
3321 	nch = fdp->fd_ncdir;
3322 	ncp = nch.ncp;
3323 	if (ncp)
3324 		_cache_hold(ncp);
3325 
3326 	while (ncp && (ncp != fdp->fd_nrdir.ncp ||
3327 	       nch.mount != fdp->fd_nrdir.mount)
3328 	) {
3329 		/*
3330 		 * While traversing upwards if we encounter the root
3331 		 * of the current mount we have to skip to the mount point
3332 		 * in the underlying filesystem.
3333 		 */
3334 		if (ncp == nch.mount->mnt_ncmountpt.ncp) {
3335 			nch = nch.mount->mnt_ncmounton;
3336 			_cache_drop(ncp);
3337 			ncp = nch.ncp;
3338 			if (ncp)
3339 				_cache_hold(ncp);
3340 			continue;
3341 		}
3342 
3343 		/*
3344 		 * Prepend the path segment
3345 		 */
3346 		for (i = ncp->nc_nlen - 1; i >= 0; i--) {
3347 			if (bp == buf) {
3348 				numcwdfailsz++;
3349 				*error = ERANGE;
3350 				bp = NULL;
3351 				goto done;
3352 			}
3353 			*--bp = ncp->nc_name[i];
3354 		}
3355 		if (bp == buf) {
3356 			numcwdfailsz++;
3357 			*error = ERANGE;
3358 			bp = NULL;
3359 			goto done;
3360 		}
3361 		*--bp = '/';
3362 		slash_prefixed = 1;
3363 
3364 		/*
3365 		 * Go up a directory.  This isn't a mount point so we don't
3366 		 * have to check again.
3367 		 */
3368 		while ((nch.ncp = ncp->nc_parent) != NULL) {
3369 			_cache_lock(ncp);
3370 			if (nch.ncp != ncp->nc_parent) {
3371 				_cache_unlock(ncp);
3372 				continue;
3373 			}
3374 			_cache_hold(nch.ncp);
3375 			_cache_unlock(ncp);
3376 			break;
3377 		}
3378 		_cache_drop(ncp);
3379 		ncp = nch.ncp;
3380 	}
3381 	if (ncp == NULL) {
3382 		numcwdfailnf++;
3383 		*error = ENOENT;
3384 		bp = NULL;
3385 		goto done;
3386 	}
3387 	if (!slash_prefixed) {
3388 		if (bp == buf) {
3389 			numcwdfailsz++;
3390 			*error = ERANGE;
3391 			bp = NULL;
3392 			goto done;
3393 		}
3394 		*--bp = '/';
3395 	}
3396 	numcwdfound++;
3397 	*error = 0;
3398 done:
3399 	if (ncp)
3400 		_cache_drop(ncp);
3401 	return (bp);
3402 }
3403 
3404 /*
3405  * Thus begins the fullpath magic.
3406  *
3407  * The passed nchp is referenced but not locked.
3408  */
3409 static int disablefullpath;
3410 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW,
3411     &disablefullpath, 0,
3412     "Disable fullpath lookups");
3413 
3414 static u_int numfullpathcalls;
3415 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathcalls, CTLFLAG_RD,
3416     &numfullpathcalls, 0,
3417     "Number of full path resolutions in progress");
3418 static u_int numfullpathfailnf;
3419 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfailnf, CTLFLAG_RD,
3420     &numfullpathfailnf, 0,
3421     "Number of full path resolution failures due to lack of file");
3422 static u_int numfullpathfailsz;
3423 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfailsz, CTLFLAG_RD,
3424     &numfullpathfailsz, 0,
3425     "Number of full path resolution failures due to insufficient memory");
3426 static u_int numfullpathfound;
3427 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfound, CTLFLAG_RD,
3428     &numfullpathfound, 0,
3429     "Number of full path resolution successes");
3430 
3431 int
3432 cache_fullpath(struct proc *p, struct nchandle *nchp, struct nchandle *nchbase,
3433 	       char **retbuf, char **freebuf, int guess)
3434 {
3435 	struct nchandle fd_nrdir;
3436 	struct nchandle nch;
3437 	struct namecache *ncp;
3438 	struct mount *mp, *new_mp;
3439 	char *bp, *buf;
3440 	int slash_prefixed;
3441 	int error = 0;
3442 	int i;
3443 
3444 	atomic_add_int(&numfullpathcalls, -1);
3445 
3446 	*retbuf = NULL;
3447 	*freebuf = NULL;
3448 
3449 	buf = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK);
3450 	bp = buf + MAXPATHLEN - 1;
3451 	*bp = '\0';
3452 	if (nchbase)
3453 		fd_nrdir = *nchbase;
3454 	else if (p != NULL)
3455 		fd_nrdir = p->p_fd->fd_nrdir;
3456 	else
3457 		fd_nrdir = rootnch;
3458 	slash_prefixed = 0;
3459 	nch = *nchp;
3460 	ncp = nch.ncp;
3461 	if (ncp)
3462 		_cache_hold(ncp);
3463 	mp = nch.mount;
3464 
3465 	while (ncp && (ncp != fd_nrdir.ncp || mp != fd_nrdir.mount)) {
3466 		new_mp = NULL;
3467 
3468 		/*
3469 		 * If we are asked to guess the upwards path, we do so whenever
3470 		 * we encounter an ncp marked as a mountpoint. We try to find
3471 		 * the actual mountpoint by finding the mountpoint with this
3472 		 * ncp.
3473 		 */
3474 		if (guess && (ncp->nc_flag & NCF_ISMOUNTPT)) {
3475 			new_mp = mount_get_by_nc(ncp);
3476 		}
3477 		/*
3478 		 * While traversing upwards if we encounter the root
3479 		 * of the current mount we have to skip to the mount point.
3480 		 */
3481 		if (ncp == mp->mnt_ncmountpt.ncp) {
3482 			new_mp = mp;
3483 		}
3484 		if (new_mp) {
3485 			nch = new_mp->mnt_ncmounton;
3486 			_cache_drop(ncp);
3487 			ncp = nch.ncp;
3488 			if (ncp)
3489 				_cache_hold(ncp);
3490 			mp = nch.mount;
3491 			continue;
3492 		}
3493 
3494 		/*
3495 		 * Prepend the path segment
3496 		 */
3497 		for (i = ncp->nc_nlen - 1; i >= 0; i--) {
3498 			if (bp == buf) {
3499 				numfullpathfailsz++;
3500 				kfree(buf, M_TEMP);
3501 				error = ENOMEM;
3502 				goto done;
3503 			}
3504 			*--bp = ncp->nc_name[i];
3505 		}
3506 		if (bp == buf) {
3507 			numfullpathfailsz++;
3508 			kfree(buf, M_TEMP);
3509 			error = ENOMEM;
3510 			goto done;
3511 		}
3512 		*--bp = '/';
3513 		slash_prefixed = 1;
3514 
3515 		/*
3516 		 * Go up a directory.  This isn't a mount point so we don't
3517 		 * have to check again.
3518 		 *
3519 		 * We can only safely access nc_parent with ncp held locked.
3520 		 */
3521 		while ((nch.ncp = ncp->nc_parent) != NULL) {
3522 			_cache_lock(ncp);
3523 			if (nch.ncp != ncp->nc_parent) {
3524 				_cache_unlock(ncp);
3525 				continue;
3526 			}
3527 			_cache_hold(nch.ncp);
3528 			_cache_unlock(ncp);
3529 			break;
3530 		}
3531 		_cache_drop(ncp);
3532 		ncp = nch.ncp;
3533 	}
3534 	if (ncp == NULL) {
3535 		numfullpathfailnf++;
3536 		kfree(buf, M_TEMP);
3537 		error = ENOENT;
3538 		goto done;
3539 	}
3540 
3541 	if (!slash_prefixed) {
3542 		if (bp == buf) {
3543 			numfullpathfailsz++;
3544 			kfree(buf, M_TEMP);
3545 			error = ENOMEM;
3546 			goto done;
3547 		}
3548 		*--bp = '/';
3549 	}
3550 	numfullpathfound++;
3551 	*retbuf = bp;
3552 	*freebuf = buf;
3553 	error = 0;
3554 done:
3555 	if (ncp)
3556 		_cache_drop(ncp);
3557 	return(error);
3558 }
3559 
3560 int
3561 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf,
3562     int guess)
3563 {
3564 	struct namecache *ncp;
3565 	struct nchandle nch;
3566 	int error;
3567 
3568 	*freebuf = NULL;
3569 	atomic_add_int(&numfullpathcalls, 1);
3570 	if (disablefullpath)
3571 		return (ENODEV);
3572 
3573 	if (p == NULL)
3574 		return (EINVAL);
3575 
3576 	/* vn is NULL, client wants us to use p->p_textvp */
3577 	if (vn == NULL) {
3578 		if ((vn = p->p_textvp) == NULL)
3579 			return (EINVAL);
3580 	}
3581 	spin_lock(&vn->v_spin);
3582 	TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) {
3583 		if (ncp->nc_nlen)
3584 			break;
3585 	}
3586 	if (ncp == NULL) {
3587 		spin_unlock(&vn->v_spin);
3588 		return (EINVAL);
3589 	}
3590 	_cache_hold(ncp);
3591 	spin_unlock(&vn->v_spin);
3592 
3593 	atomic_add_int(&numfullpathcalls, -1);
3594 	nch.ncp = ncp;
3595 	nch.mount = vn->v_mount;
3596 	error = cache_fullpath(p, &nch, NULL, retbuf, freebuf, guess);
3597 	_cache_drop(ncp);
3598 	return (error);
3599 }
3600