xref: /dflybsd-src/sys/kern/vfs_cache.c (revision 5270936ced7738caef7b7302e4b02466f8e913bc)
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. Neither the name of the University nor the names of its contributors
49  *    may be used to endorse or promote products derived from this software
50  *    without specific prior written permission.
51  *
52  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
53  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
54  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
55  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
56  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
57  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
58  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
59  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
60  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
61  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
62  * SUCH DAMAGE.
63  */
64 
65 #include <sys/param.h>
66 #include <sys/systm.h>
67 #include <sys/kernel.h>
68 #include <sys/sysctl.h>
69 #include <sys/mount.h>
70 #include <sys/vnode.h>
71 #include <sys/malloc.h>
72 #include <sys/sysproto.h>
73 #include <sys/spinlock.h>
74 #include <sys/proc.h>
75 #include <sys/namei.h>
76 #include <sys/nlookup.h>
77 #include <sys/filedesc.h>
78 #include <sys/fnv_hash.h>
79 #include <sys/globaldata.h>
80 #include <sys/kern_syscall.h>
81 #include <sys/dirent.h>
82 #include <ddb/ddb.h>
83 
84 #include <sys/sysref2.h>
85 #include <sys/spinlock2.h>
86 #include <sys/mplock2.h>
87 
88 #define MAX_RECURSION_DEPTH	64
89 
90 /*
91  * Random lookups in the cache are accomplished with a hash table using
92  * a hash key of (nc_src_vp, name).  Each hash chain has its own spin lock.
93  *
94  * Negative entries may exist and correspond to resolved namecache
95  * structures where nc_vp is NULL.  In a negative entry, NCF_WHITEOUT
96  * will be set if the entry corresponds to a whited-out directory entry
97  * (verses simply not finding the entry at all).   ncneglist is locked
98  * with a global spinlock (ncspin).
99  *
100  * MPSAFE RULES:
101  *
102  * (1) A ncp must be referenced before it can be locked.
103  *
104  * (2) A ncp must be locked in order to modify it.
105  *
106  * (3) ncp locks are always ordered child -> parent.  That may seem
107  *     backwards but forward scans use the hash table and thus can hold
108  *     the parent unlocked when traversing downward.
109  *
110  *     This allows insert/rename/delete/dot-dot and other operations
111  *     to use ncp->nc_parent links.
112  *
113  *     This also prevents a locked up e.g. NFS node from creating a
114  *     chain reaction all the way back to the root vnode / namecache.
115  *
116  * (4) parent linkages require both the parent and child to be locked.
117  */
118 
119 /*
120  * Structures associated with name cacheing.
121  */
122 #define NCHHASH(hash)		(&nchashtbl[(hash) & nchash])
123 #define MINNEG			1024
124 #define MINPOS			1024
125 #define NCMOUNT_NUMCACHE	1009	/* prime number */
126 
127 MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
128 
129 LIST_HEAD(nchash_list, namecache);
130 
131 struct nchash_head {
132        struct nchash_list list;
133        struct spinlock	spin;
134 };
135 
136 struct ncmount_cache {
137 	struct spinlock	spin;
138 	struct namecache *ncp;
139 	struct mount *mp;
140 	int isneg;		/* if != 0 mp is originator and not target */
141 };
142 
143 static struct nchash_head	*nchashtbl;
144 static struct namecache_list	ncneglist;
145 static struct spinlock		ncspin;
146 static struct ncmount_cache	ncmount_cache[NCMOUNT_NUMCACHE];
147 
148 /*
149  * ncvp_debug - debug cache_fromvp().  This is used by the NFS server
150  * to create the namecache infrastructure leading to a dangling vnode.
151  *
152  * 0	Only errors are reported
153  * 1	Successes are reported
154  * 2	Successes + the whole directory scan is reported
155  * 3	Force the directory scan code run as if the parent vnode did not
156  *	have a namecache record, even if it does have one.
157  */
158 static int	ncvp_debug;
159 SYSCTL_INT(_debug, OID_AUTO, ncvp_debug, CTLFLAG_RW, &ncvp_debug, 0,
160     "Namecache debug level (0-3)");
161 
162 static u_long	nchash;			/* size of hash table */
163 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
164     "Size of namecache hash table");
165 
166 static int	ncnegflush = 10;	/* burst for negative flush */
167 SYSCTL_INT(_debug, OID_AUTO, ncnegflush, CTLFLAG_RW, &ncnegflush, 0,
168     "Batch flush negative entries");
169 
170 static int	ncposflush = 10;	/* burst for positive flush */
171 SYSCTL_INT(_debug, OID_AUTO, ncposflush, CTLFLAG_RW, &ncposflush, 0,
172     "Batch flush positive entries");
173 
174 static int	ncnegfactor = 16;	/* ratio of negative entries */
175 SYSCTL_INT(_debug, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0,
176     "Ratio of namecache negative entries");
177 
178 static int	nclockwarn;		/* warn on locked entries in ticks */
179 SYSCTL_INT(_debug, OID_AUTO, nclockwarn, CTLFLAG_RW, &nclockwarn, 0,
180     "Warn on locked namecache entries in ticks");
181 
182 static int	numdefered;		/* number of cache entries allocated */
183 SYSCTL_INT(_debug, OID_AUTO, numdefered, CTLFLAG_RD, &numdefered, 0,
184     "Number of cache entries allocated");
185 
186 static int	ncposlimit;		/* number of cache entries allocated */
187 SYSCTL_INT(_debug, OID_AUTO, ncposlimit, CTLFLAG_RW, &ncposlimit, 0,
188     "Number of cache entries allocated");
189 
190 static int	ncp_shared_lock_disable = 0;
191 SYSCTL_INT(_debug, OID_AUTO, ncp_shared_lock_disable, CTLFLAG_RW,
192 	   &ncp_shared_lock_disable, 0, "Disable shared namecache locks");
193 
194 SYSCTL_INT(_debug, OID_AUTO, vnsize, CTLFLAG_RD, 0, sizeof(struct vnode),
195     "sizeof(struct vnode)");
196 SYSCTL_INT(_debug, OID_AUTO, ncsize, CTLFLAG_RD, 0, sizeof(struct namecache),
197     "sizeof(struct namecache)");
198 
199 static int	ncmount_cache_enable = 1;
200 SYSCTL_INT(_debug, OID_AUTO, ncmount_cache_enable, CTLFLAG_RW,
201 	   &ncmount_cache_enable, 0, "mount point cache");
202 static long	ncmount_cache_hit;
203 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_hit, CTLFLAG_RW,
204 	    &ncmount_cache_hit, 0, "mpcache hits");
205 static long	ncmount_cache_miss;
206 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_miss, CTLFLAG_RW,
207 	    &ncmount_cache_miss, 0, "mpcache misses");
208 static long	ncmount_cache_overwrite;
209 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_overwrite, CTLFLAG_RW,
210 	    &ncmount_cache_overwrite, 0, "mpcache entry overwrites");
211 
212 static int cache_resolve_mp(struct mount *mp);
213 static struct vnode *cache_dvpref(struct namecache *ncp);
214 static void _cache_lock(struct namecache *ncp);
215 static void _cache_setunresolved(struct namecache *ncp);
216 static void _cache_cleanneg(int count);
217 static void _cache_cleanpos(int count);
218 static void _cache_cleandefered(void);
219 static void _cache_unlink(struct namecache *ncp);
220 
221 /*
222  * The new name cache statistics
223  */
224 SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics");
225 static int numneg;
226 SYSCTL_INT(_vfs_cache, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0,
227     "Number of negative namecache entries");
228 static int numcache;
229 SYSCTL_INT(_vfs_cache, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0,
230     "Number of namecaches entries");
231 static u_long numcalls;
232 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcalls, CTLFLAG_RD, &numcalls, 0,
233     "Number of namecache lookups");
234 static u_long numchecks;
235 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numchecks, CTLFLAG_RD, &numchecks, 0,
236     "Number of checked entries in namecache lookups");
237 
238 struct nchstats nchstats[SMP_MAXCPU];
239 /*
240  * Export VFS cache effectiveness statistics to user-land.
241  *
242  * The statistics are left for aggregation to user-land so
243  * neat things can be achieved, like observing per-CPU cache
244  * distribution.
245  */
246 static int
247 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
248 {
249 	struct globaldata *gd;
250 	int i, error;
251 
252 	error = 0;
253 	for (i = 0; i < ncpus; ++i) {
254 		gd = globaldata_find(i);
255 		if ((error = SYSCTL_OUT(req, (void *)&(*gd->gd_nchstats),
256 			sizeof(struct nchstats))))
257 			break;
258 	}
259 
260 	return (error);
261 }
262 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE|CTLFLAG_RD,
263   0, 0, sysctl_nchstats, "S,nchstats", "VFS cache effectiveness statistics");
264 
265 static struct namecache *cache_zap(struct namecache *ncp, int nonblock);
266 
267 /*
268  * Namespace locking.  The caller must already hold a reference to the
269  * namecache structure in order to lock/unlock it.  This function prevents
270  * the namespace from being created or destroyed by accessors other then
271  * the lock holder.
272  *
273  * Note that holding a locked namecache structure prevents other threads
274  * from making namespace changes (e.g. deleting or creating), prevents
275  * vnode association state changes by other threads, and prevents the
276  * namecache entry from being resolved or unresolved by other threads.
277  *
278  * An exclusive lock owner has full authority to associate/disassociate
279  * vnodes and resolve/unresolve the locked ncp.
280  *
281  * A shared lock owner only has authority to acquire the underlying vnode,
282  * if any.
283  *
284  * The primary lock field is nc_lockstatus.  nc_locktd is set after the
285  * fact (when locking) or cleared prior to unlocking.
286  *
287  * WARNING!  Holding a locked ncp will prevent a vnode from being destroyed
288  *	     or recycled, but it does NOT help you if the vnode had already
289  *	     initiated a recyclement.  If this is important, use cache_get()
290  *	     rather then cache_lock() (and deal with the differences in the
291  *	     way the refs counter is handled).  Or, alternatively, make an
292  *	     unconditional call to cache_validate() or cache_resolve()
293  *	     after cache_lock() returns.
294  */
295 static
296 void
297 _cache_lock(struct namecache *ncp)
298 {
299 	thread_t td;
300 	int didwarn;
301 	int begticks;
302 	int error;
303 	u_int count;
304 
305 	KKASSERT(ncp->nc_refs != 0);
306 	didwarn = 0;
307 	begticks = 0;
308 	td = curthread;
309 
310 	for (;;) {
311 		count = ncp->nc_lockstatus;
312 		cpu_ccfence();
313 
314 		if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 0) {
315 			if (atomic_cmpset_int(&ncp->nc_lockstatus,
316 					      count, count + 1)) {
317 				/*
318 				 * The vp associated with a locked ncp must
319 				 * be held to prevent it from being recycled.
320 				 *
321 				 * WARNING!  If VRECLAIMED is set the vnode
322 				 * could already be in the middle of a recycle.
323 				 * Callers must use cache_vref() or
324 				 * cache_vget() on the locked ncp to
325 				 * validate the vp or set the cache entry
326 				 * to unresolved.
327 				 *
328 				 * NOTE! vhold() is allowed if we hold a
329 				 *	 lock on the ncp (which we do).
330 				 */
331 				ncp->nc_locktd = td;
332 				if (ncp->nc_vp)
333 					vhold(ncp->nc_vp);
334 				break;
335 			}
336 			/* cmpset failed */
337 			continue;
338 		}
339 		if (ncp->nc_locktd == td) {
340 			KKASSERT((count & NC_SHLOCK_FLAG) == 0);
341 			if (atomic_cmpset_int(&ncp->nc_lockstatus,
342 					      count, count + 1)) {
343 				break;
344 			}
345 			/* cmpset failed */
346 			continue;
347 		}
348 		tsleep_interlock(&ncp->nc_locktd, 0);
349 		if (atomic_cmpset_int(&ncp->nc_lockstatus, count,
350 				      count | NC_EXLOCK_REQ) == 0) {
351 			/* cmpset failed */
352 			continue;
353 		}
354 		if (begticks == 0)
355 			begticks = ticks;
356 		error = tsleep(&ncp->nc_locktd, PINTERLOCKED,
357 			       "clock", nclockwarn);
358 		if (error == EWOULDBLOCK) {
359 			if (didwarn == 0) {
360 				didwarn = ticks;
361 				kprintf("[diagnostic] cache_lock: "
362 					"blocked on %p %08x",
363 					ncp, count);
364 				kprintf(" \"%*.*s\"\n",
365 					ncp->nc_nlen, ncp->nc_nlen,
366 					ncp->nc_name);
367 			}
368 		}
369 		/* loop */
370 	}
371 	if (didwarn) {
372 		kprintf("[diagnostic] cache_lock: unblocked %*.*s after "
373 			"%d secs\n",
374 			ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name,
375 			(int)(ticks + (hz / 2) - begticks) / hz);
376 	}
377 }
378 
379 /*
380  * The shared lock works similarly to the exclusive lock except
381  * nc_locktd is left NULL and we need an interlock (VHOLD) to
382  * prevent vhold() races, since the moment our cmpset_int succeeds
383  * another cpu can come in and get its own shared lock.
384  *
385  * A critical section is needed to prevent interruption during the
386  * VHOLD interlock.
387  */
388 static
389 void
390 _cache_lock_shared(struct namecache *ncp)
391 {
392 	int didwarn;
393 	int error;
394 	u_int count;
395 	u_int optreq = NC_EXLOCK_REQ;
396 
397 	KKASSERT(ncp->nc_refs != 0);
398 	didwarn = 0;
399 
400 	for (;;) {
401 		count = ncp->nc_lockstatus;
402 		cpu_ccfence();
403 
404 		if ((count & ~NC_SHLOCK_REQ) == 0) {
405 			crit_enter();
406 			if (atomic_cmpset_int(&ncp->nc_lockstatus,
407 				      count,
408 				      (count + 1) | NC_SHLOCK_FLAG |
409 						    NC_SHLOCK_VHOLD)) {
410 				/*
411 				 * The vp associated with a locked ncp must
412 				 * be held to prevent it from being recycled.
413 				 *
414 				 * WARNING!  If VRECLAIMED is set the vnode
415 				 * could already be in the middle of a recycle.
416 				 * Callers must use cache_vref() or
417 				 * cache_vget() on the locked ncp to
418 				 * validate the vp or set the cache entry
419 				 * to unresolved.
420 				 *
421 				 * NOTE! vhold() is allowed if we hold a
422 				 *	 lock on the ncp (which we do).
423 				 */
424 				if (ncp->nc_vp)
425 					vhold(ncp->nc_vp);
426 				atomic_clear_int(&ncp->nc_lockstatus,
427 						 NC_SHLOCK_VHOLD);
428 				crit_exit();
429 				break;
430 			}
431 			/* cmpset failed */
432 			crit_exit();
433 			continue;
434 		}
435 
436 		/*
437 		 * If already held shared we can just bump the count, but
438 		 * only allow this if nobody is trying to get the lock
439 		 * exclusively.  If we are blocking too long ignore excl
440 		 * requests (which can race/deadlock us).
441 		 *
442 		 * VHOLD is a bit of a hack.  Even though we successfully
443 		 * added another shared ref, the cpu that got the first
444 		 * shared ref might not yet have held the vnode.
445 		 */
446 		if ((count & (optreq|NC_SHLOCK_FLAG)) == NC_SHLOCK_FLAG) {
447 			KKASSERT((count & ~(NC_EXLOCK_REQ |
448 					    NC_SHLOCK_REQ |
449 					    NC_SHLOCK_FLAG)) > 0);
450 			if (atomic_cmpset_int(&ncp->nc_lockstatus,
451 					      count, count + 1)) {
452 				while (ncp->nc_lockstatus & NC_SHLOCK_VHOLD)
453 					cpu_pause();
454 				break;
455 			}
456 			continue;
457 		}
458 		tsleep_interlock(ncp, 0);
459 		if (atomic_cmpset_int(&ncp->nc_lockstatus, count,
460 				      count | NC_SHLOCK_REQ) == 0) {
461 			/* cmpset failed */
462 			continue;
463 		}
464 		error = tsleep(ncp, PINTERLOCKED, "clocksh", nclockwarn);
465 		if (error == EWOULDBLOCK) {
466 			optreq = 0;
467 			if (didwarn == 0) {
468 				didwarn = ticks;
469 				kprintf("[diagnostic] cache_lock_shared: "
470 					"blocked on %p %08x",
471 					ncp, count);
472 				kprintf(" \"%*.*s\"\n",
473 					ncp->nc_nlen, ncp->nc_nlen,
474 					ncp->nc_name);
475 			}
476 		}
477 		/* loop */
478 	}
479 	if (didwarn) {
480 		kprintf("[diagnostic] cache_lock_shared: "
481 			"unblocked %*.*s after %d secs\n",
482 			ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name,
483 			(int)(ticks - didwarn) / hz);
484 	}
485 }
486 
487 /*
488  * Lock ncp exclusively, return 0 on success.
489  *
490  * NOTE: nc_refs may be zero if the ncp is interlocked by circumstance,
491  *	 such as the case where one of its children is locked.
492  */
493 static
494 int
495 _cache_lock_nonblock(struct namecache *ncp)
496 {
497 	thread_t td;
498 	u_int count;
499 
500 	td = curthread;
501 
502 	for (;;) {
503 		count = ncp->nc_lockstatus;
504 
505 		if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 0) {
506 			if (atomic_cmpset_int(&ncp->nc_lockstatus,
507 					      count, count + 1)) {
508 				/*
509 				 * The vp associated with a locked ncp must
510 				 * be held to prevent it from being recycled.
511 				 *
512 				 * WARNING!  If VRECLAIMED is set the vnode
513 				 * could already be in the middle of a recycle.
514 				 * Callers must use cache_vref() or
515 				 * cache_vget() on the locked ncp to
516 				 * validate the vp or set the cache entry
517 				 * to unresolved.
518 				 *
519 				 * NOTE! vhold() is allowed if we hold a
520 				 *	 lock on the ncp (which we do).
521 				 */
522 				ncp->nc_locktd = td;
523 				if (ncp->nc_vp)
524 					vhold(ncp->nc_vp);
525 				break;
526 			}
527 			/* cmpset failed */
528 			continue;
529 		}
530 		if (ncp->nc_locktd == td) {
531 			if (atomic_cmpset_int(&ncp->nc_lockstatus,
532 					      count, count + 1)) {
533 				break;
534 			}
535 			/* cmpset failed */
536 			continue;
537 		}
538 		return(EWOULDBLOCK);
539 	}
540 	return(0);
541 }
542 
543 /*
544  * The shared lock works similarly to the exclusive lock except
545  * nc_locktd is left NULL and we need an interlock (VHOLD) to
546  * prevent vhold() races, since the moment our cmpset_int succeeds
547  * another cpu can come in and get its own shared lock.
548  *
549  * A critical section is needed to prevent interruption during the
550  * VHOLD interlock.
551  */
552 static
553 int
554 _cache_lock_shared_nonblock(struct namecache *ncp)
555 {
556 	u_int count;
557 
558 	for (;;) {
559 		count = ncp->nc_lockstatus;
560 
561 		if ((count & ~NC_SHLOCK_REQ) == 0) {
562 			crit_enter();
563 			if (atomic_cmpset_int(&ncp->nc_lockstatus,
564 				      count,
565 				      (count + 1) | NC_SHLOCK_FLAG |
566 						    NC_SHLOCK_VHOLD)) {
567 				/*
568 				 * The vp associated with a locked ncp must
569 				 * be held to prevent it from being recycled.
570 				 *
571 				 * WARNING!  If VRECLAIMED is set the vnode
572 				 * could already be in the middle of a recycle.
573 				 * Callers must use cache_vref() or
574 				 * cache_vget() on the locked ncp to
575 				 * validate the vp or set the cache entry
576 				 * to unresolved.
577 				 *
578 				 * NOTE! vhold() is allowed if we hold a
579 				 *	 lock on the ncp (which we do).
580 				 */
581 				if (ncp->nc_vp)
582 					vhold(ncp->nc_vp);
583 				atomic_clear_int(&ncp->nc_lockstatus,
584 						 NC_SHLOCK_VHOLD);
585 				crit_exit();
586 				break;
587 			}
588 			/* cmpset failed */
589 			crit_exit();
590 			continue;
591 		}
592 
593 		/*
594 		 * If already held shared we can just bump the count, but
595 		 * only allow this if nobody is trying to get the lock
596 		 * exclusively.
597 		 *
598 		 * VHOLD is a bit of a hack.  Even though we successfully
599 		 * added another shared ref, the cpu that got the first
600 		 * shared ref might not yet have held the vnode.
601 		 */
602 		if ((count & (NC_EXLOCK_REQ|NC_SHLOCK_FLAG)) ==
603 		    NC_SHLOCK_FLAG) {
604 			KKASSERT((count & ~(NC_EXLOCK_REQ |
605 					    NC_SHLOCK_REQ |
606 					    NC_SHLOCK_FLAG)) > 0);
607 			if (atomic_cmpset_int(&ncp->nc_lockstatus,
608 					      count, count + 1)) {
609 				while (ncp->nc_lockstatus & NC_SHLOCK_VHOLD)
610 					cpu_pause();
611 				break;
612 			}
613 			continue;
614 		}
615 		return(EWOULDBLOCK);
616 	}
617 	return(0);
618 }
619 
620 /*
621  * Helper function
622  *
623  * NOTE: nc_refs can be 0 (degenerate case during _cache_drop).
624  *
625  *	 nc_locktd must be NULLed out prior to nc_lockstatus getting cleared.
626  */
627 static
628 void
629 _cache_unlock(struct namecache *ncp)
630 {
631 	thread_t td __debugvar = curthread;
632 	u_int count;
633 	u_int ncount;
634 	struct vnode *dropvp;
635 
636 	KKASSERT(ncp->nc_refs >= 0);
637 	KKASSERT((ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) > 0);
638 	KKASSERT((ncp->nc_lockstatus & NC_SHLOCK_FLAG) || ncp->nc_locktd == td);
639 
640 	count = ncp->nc_lockstatus;
641 	cpu_ccfence();
642 
643 	/*
644 	 * Clear nc_locktd prior to the atomic op (excl lock only)
645 	 */
646 	if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 1)
647 		ncp->nc_locktd = NULL;
648 	dropvp = NULL;
649 
650 	for (;;) {
651 		if ((count &
652 		     ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ|NC_SHLOCK_FLAG)) == 1) {
653 			dropvp = ncp->nc_vp;
654 			if (count & NC_EXLOCK_REQ)
655 				ncount = count & NC_SHLOCK_REQ; /* cnt->0 */
656 			else
657 				ncount = 0;
658 
659 			if (atomic_cmpset_int(&ncp->nc_lockstatus,
660 					      count, ncount)) {
661 				if (count & NC_EXLOCK_REQ)
662 					wakeup(&ncp->nc_locktd);
663 				else if (count & NC_SHLOCK_REQ)
664 					wakeup(ncp);
665 				break;
666 			}
667 			dropvp = NULL;
668 		} else {
669 			KKASSERT((count & NC_SHLOCK_VHOLD) == 0);
670 			KKASSERT((count & ~(NC_EXLOCK_REQ |
671 					    NC_SHLOCK_REQ |
672 					    NC_SHLOCK_FLAG)) > 1);
673 			if (atomic_cmpset_int(&ncp->nc_lockstatus,
674 					      count, count - 1)) {
675 				break;
676 			}
677 		}
678 		count = ncp->nc_lockstatus;
679 		cpu_ccfence();
680 	}
681 
682 	/*
683 	 * Don't actually drop the vp until we successfully clean out
684 	 * the lock, otherwise we may race another shared lock.
685 	 */
686 	if (dropvp)
687 		vdrop(dropvp);
688 }
689 
690 static
691 int
692 _cache_lockstatus(struct namecache *ncp)
693 {
694 	if (ncp->nc_locktd == curthread)
695 		return(LK_EXCLUSIVE);
696 	if (ncp->nc_lockstatus & NC_SHLOCK_FLAG)
697 		return(LK_SHARED);
698 	return(-1);
699 }
700 
701 /*
702  * cache_hold() and cache_drop() prevent the premature deletion of a
703  * namecache entry but do not prevent operations (such as zapping) on
704  * that namecache entry.
705  *
706  * This routine may only be called from outside this source module if
707  * nc_refs is already at least 1.
708  *
709  * This is a rare case where callers are allowed to hold a spinlock,
710  * so we can't ourselves.
711  */
712 static __inline
713 struct namecache *
714 _cache_hold(struct namecache *ncp)
715 {
716 	atomic_add_int(&ncp->nc_refs, 1);
717 	return(ncp);
718 }
719 
720 /*
721  * Drop a cache entry, taking care to deal with races.
722  *
723  * For potential 1->0 transitions we must hold the ncp lock to safely
724  * test its flags.  An unresolved entry with no children must be zapped
725  * to avoid leaks.
726  *
727  * The call to cache_zap() itself will handle all remaining races and
728  * will decrement the ncp's refs regardless.  If we are resolved or
729  * have children nc_refs can safely be dropped to 0 without having to
730  * zap the entry.
731  *
732  * NOTE: cache_zap() will re-check nc_refs and nc_list in a MPSAFE fashion.
733  *
734  * NOTE: cache_zap() may return a non-NULL referenced parent which must
735  *	 be dropped in a loop.
736  */
737 static __inline
738 void
739 _cache_drop(struct namecache *ncp)
740 {
741 	int refs;
742 
743 	while (ncp) {
744 		KKASSERT(ncp->nc_refs > 0);
745 		refs = ncp->nc_refs;
746 
747 		if (refs == 1) {
748 			if (_cache_lock_nonblock(ncp) == 0) {
749 				ncp->nc_flag &= ~NCF_DEFEREDZAP;
750 				if ((ncp->nc_flag & NCF_UNRESOLVED) &&
751 				    TAILQ_EMPTY(&ncp->nc_list)) {
752 					ncp = cache_zap(ncp, 1);
753 					continue;
754 				}
755 				if (atomic_cmpset_int(&ncp->nc_refs, 1, 0)) {
756 					_cache_unlock(ncp);
757 					break;
758 				}
759 				_cache_unlock(ncp);
760 			}
761 		} else {
762 			if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1))
763 				break;
764 		}
765 		cpu_pause();
766 	}
767 }
768 
769 /*
770  * Link a new namecache entry to its parent and to the hash table.  Be
771  * careful to avoid races if vhold() blocks in the future.
772  *
773  * Both ncp and par must be referenced and locked.
774  *
775  * NOTE: The hash table spinlock is held during this call, we can't do
776  *	 anything fancy.
777  */
778 static void
779 _cache_link_parent(struct namecache *ncp, struct namecache *par,
780 		   struct nchash_head *nchpp)
781 {
782 	KKASSERT(ncp->nc_parent == NULL);
783 	ncp->nc_parent = par;
784 	ncp->nc_head = nchpp;
785 
786 	/*
787 	 * Set inheritance flags.  Note that the parent flags may be
788 	 * stale due to getattr potentially not having been run yet
789 	 * (it gets run during nlookup()'s).
790 	 */
791 	ncp->nc_flag &= ~(NCF_SF_PNOCACHE | NCF_UF_PCACHE);
792 	if (par->nc_flag & (NCF_SF_NOCACHE | NCF_SF_PNOCACHE))
793 		ncp->nc_flag |= NCF_SF_PNOCACHE;
794 	if (par->nc_flag & (NCF_UF_CACHE | NCF_UF_PCACHE))
795 		ncp->nc_flag |= NCF_UF_PCACHE;
796 
797 	LIST_INSERT_HEAD(&nchpp->list, ncp, nc_hash);
798 
799 	if (TAILQ_EMPTY(&par->nc_list)) {
800 		TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
801 		/*
802 		 * Any vp associated with an ncp which has children must
803 		 * be held to prevent it from being recycled.
804 		 */
805 		if (par->nc_vp)
806 			vhold(par->nc_vp);
807 	} else {
808 		TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry);
809 	}
810 }
811 
812 /*
813  * Remove the parent and hash associations from a namecache structure.
814  * If this is the last child of the parent the cache_drop(par) will
815  * attempt to recursively zap the parent.
816  *
817  * ncp must be locked.  This routine will acquire a temporary lock on
818  * the parent as wlel as the appropriate hash chain.
819  */
820 static void
821 _cache_unlink_parent(struct namecache *ncp)
822 {
823 	struct namecache *par;
824 	struct vnode *dropvp;
825 
826 	if ((par = ncp->nc_parent) != NULL) {
827 		KKASSERT(ncp->nc_parent == par);
828 		_cache_hold(par);
829 		_cache_lock(par);
830 		spin_lock(&ncp->nc_head->spin);
831 		LIST_REMOVE(ncp, nc_hash);
832 		TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
833 		dropvp = NULL;
834 		if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
835 			dropvp = par->nc_vp;
836 		spin_unlock(&ncp->nc_head->spin);
837 		ncp->nc_parent = NULL;
838 		ncp->nc_head = NULL;
839 		_cache_unlock(par);
840 		_cache_drop(par);
841 
842 		/*
843 		 * We can only safely vdrop with no spinlocks held.
844 		 */
845 		if (dropvp)
846 			vdrop(dropvp);
847 	}
848 }
849 
850 /*
851  * Allocate a new namecache structure.  Most of the code does not require
852  * zero-termination of the string but it makes vop_compat_ncreate() easier.
853  */
854 static struct namecache *
855 cache_alloc(int nlen)
856 {
857 	struct namecache *ncp;
858 
859 	ncp = kmalloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO);
860 	if (nlen)
861 		ncp->nc_name = kmalloc(nlen + 1, M_VFSCACHE, M_WAITOK);
862 	ncp->nc_nlen = nlen;
863 	ncp->nc_flag = NCF_UNRESOLVED;
864 	ncp->nc_error = ENOTCONN;	/* needs to be resolved */
865 	ncp->nc_refs = 1;
866 
867 	TAILQ_INIT(&ncp->nc_list);
868 	_cache_lock(ncp);
869 	return(ncp);
870 }
871 
872 /*
873  * Can only be called for the case where the ncp has never been
874  * associated with anything (so no spinlocks are needed).
875  */
876 static void
877 _cache_free(struct namecache *ncp)
878 {
879 	KKASSERT(ncp->nc_refs == 1 && ncp->nc_lockstatus == 1);
880 	if (ncp->nc_name)
881 		kfree(ncp->nc_name, M_VFSCACHE);
882 	kfree(ncp, M_VFSCACHE);
883 }
884 
885 /*
886  * [re]initialize a nchandle.
887  */
888 void
889 cache_zero(struct nchandle *nch)
890 {
891 	nch->ncp = NULL;
892 	nch->mount = NULL;
893 }
894 
895 /*
896  * Ref and deref a namecache structure.
897  *
898  * The caller must specify a stable ncp pointer, typically meaning the
899  * ncp is already referenced but this can also occur indirectly through
900  * e.g. holding a lock on a direct child.
901  *
902  * WARNING: Caller may hold an unrelated read spinlock, which means we can't
903  *	    use read spinlocks here.
904  *
905  * MPSAFE if nch is
906  */
907 struct nchandle *
908 cache_hold(struct nchandle *nch)
909 {
910 	_cache_hold(nch->ncp);
911 	atomic_add_int(&nch->mount->mnt_refs, 1);
912 	return(nch);
913 }
914 
915 /*
916  * Create a copy of a namecache handle for an already-referenced
917  * entry.
918  *
919  * MPSAFE if nch is
920  */
921 void
922 cache_copy(struct nchandle *nch, struct nchandle *target)
923 {
924 	*target = *nch;
925 	if (target->ncp)
926 		_cache_hold(target->ncp);
927 	atomic_add_int(&nch->mount->mnt_refs, 1);
928 }
929 
930 /*
931  * MPSAFE if nch is
932  */
933 void
934 cache_changemount(struct nchandle *nch, struct mount *mp)
935 {
936 	atomic_add_int(&nch->mount->mnt_refs, -1);
937 	nch->mount = mp;
938 	atomic_add_int(&nch->mount->mnt_refs, 1);
939 }
940 
941 void
942 cache_drop(struct nchandle *nch)
943 {
944 	atomic_add_int(&nch->mount->mnt_refs, -1);
945 	_cache_drop(nch->ncp);
946 	nch->ncp = NULL;
947 	nch->mount = NULL;
948 }
949 
950 int
951 cache_lockstatus(struct nchandle *nch)
952 {
953 	return(_cache_lockstatus(nch->ncp));
954 }
955 
956 void
957 cache_lock(struct nchandle *nch)
958 {
959 	_cache_lock(nch->ncp);
960 }
961 
962 void
963 cache_lock_maybe_shared(struct nchandle *nch, int excl)
964 {
965 	struct namecache *ncp = nch->ncp;
966 
967 	if (ncp_shared_lock_disable || excl ||
968 	    (ncp->nc_flag & NCF_UNRESOLVED)) {
969 		_cache_lock(ncp);
970 	} else {
971 		_cache_lock_shared(ncp);
972 		if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
973 			if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) {
974 				_cache_unlock(ncp);
975 				_cache_lock(ncp);
976 			}
977 		} else {
978 			_cache_unlock(ncp);
979 			_cache_lock(ncp);
980 		}
981 	}
982 }
983 
984 /*
985  * Relock nch1 given an unlocked nch1 and a locked nch2.  The caller
986  * is responsible for checking both for validity on return as they
987  * may have become invalid.
988  *
989  * We have to deal with potential deadlocks here, just ping pong
990  * the lock until we get it (we will always block somewhere when
991  * looping so this is not cpu-intensive).
992  *
993  * which = 0	nch1 not locked, nch2 is locked
994  * which = 1	nch1 is locked, nch2 is not locked
995  */
996 void
997 cache_relock(struct nchandle *nch1, struct ucred *cred1,
998 	     struct nchandle *nch2, struct ucred *cred2)
999 {
1000 	int which;
1001 
1002 	which = 0;
1003 
1004 	for (;;) {
1005 		if (which == 0) {
1006 			if (cache_lock_nonblock(nch1) == 0) {
1007 				cache_resolve(nch1, cred1);
1008 				break;
1009 			}
1010 			cache_unlock(nch2);
1011 			cache_lock(nch1);
1012 			cache_resolve(nch1, cred1);
1013 			which = 1;
1014 		} else {
1015 			if (cache_lock_nonblock(nch2) == 0) {
1016 				cache_resolve(nch2, cred2);
1017 				break;
1018 			}
1019 			cache_unlock(nch1);
1020 			cache_lock(nch2);
1021 			cache_resolve(nch2, cred2);
1022 			which = 0;
1023 		}
1024 	}
1025 }
1026 
1027 int
1028 cache_lock_nonblock(struct nchandle *nch)
1029 {
1030 	return(_cache_lock_nonblock(nch->ncp));
1031 }
1032 
1033 void
1034 cache_unlock(struct nchandle *nch)
1035 {
1036 	_cache_unlock(nch->ncp);
1037 }
1038 
1039 /*
1040  * ref-and-lock, unlock-and-deref functions.
1041  *
1042  * This function is primarily used by nlookup.  Even though cache_lock
1043  * holds the vnode, it is possible that the vnode may have already
1044  * initiated a recyclement.
1045  *
1046  * We want cache_get() to return a definitively usable vnode or a
1047  * definitively unresolved ncp.
1048  */
1049 static
1050 struct namecache *
1051 _cache_get(struct namecache *ncp)
1052 {
1053 	_cache_hold(ncp);
1054 	_cache_lock(ncp);
1055 	if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1056 		_cache_setunresolved(ncp);
1057 	return(ncp);
1058 }
1059 
1060 /*
1061  * Attempt to obtain a shared lock on the ncp.  A shared lock will only
1062  * be obtained if the ncp is resolved and the vnode (if not ENOENT) is
1063  * valid.  Otherwise an exclusive lock will be acquired instead.
1064  */
1065 static
1066 struct namecache *
1067 _cache_get_maybe_shared(struct namecache *ncp, int excl)
1068 {
1069 	if (ncp_shared_lock_disable || excl ||
1070 	    (ncp->nc_flag & NCF_UNRESOLVED)) {
1071 		return(_cache_get(ncp));
1072 	}
1073 	_cache_hold(ncp);
1074 	_cache_lock_shared(ncp);
1075 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1076 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) {
1077 			_cache_unlock(ncp);
1078 			ncp = _cache_get(ncp);
1079 			_cache_drop(ncp);
1080 		}
1081 	} else {
1082 		_cache_unlock(ncp);
1083 		ncp = _cache_get(ncp);
1084 		_cache_drop(ncp);
1085 	}
1086 	return(ncp);
1087 }
1088 
1089 /*
1090  * This is a special form of _cache_lock() which only succeeds if
1091  * it can get a pristine, non-recursive lock.  The caller must have
1092  * already ref'd the ncp.
1093  *
1094  * On success the ncp will be locked, on failure it will not.  The
1095  * ref count does not change either way.
1096  *
1097  * We want _cache_lock_special() (on success) to return a definitively
1098  * usable vnode or a definitively unresolved ncp.
1099  */
1100 static int
1101 _cache_lock_special(struct namecache *ncp)
1102 {
1103 	if (_cache_lock_nonblock(ncp) == 0) {
1104 		if ((ncp->nc_lockstatus &
1105 		     ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 1) {
1106 			if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
1107 				_cache_setunresolved(ncp);
1108 			return(0);
1109 		}
1110 		_cache_unlock(ncp);
1111 	}
1112 	return(EWOULDBLOCK);
1113 }
1114 
1115 /*
1116  * This function tries to get a shared lock but will back-off to an exclusive
1117  * lock if:
1118  *
1119  * (1) Some other thread is trying to obtain an exclusive lock
1120  *     (to prevent the exclusive requester from getting livelocked out
1121  *     by many shared locks).
1122  *
1123  * (2) The current thread already owns an exclusive lock (to avoid
1124  *     deadlocking).
1125  *
1126  * WARNING! On machines with lots of cores we really want to try hard to
1127  *	    get a shared lock or concurrent path lookups can chain-react
1128  *	    into a very high-latency exclusive lock.
1129  */
1130 static int
1131 _cache_lock_shared_special(struct namecache *ncp)
1132 {
1133 	/*
1134 	 * Only honor a successful shared lock (returning 0) if there is
1135 	 * no exclusive request pending and the vnode, if present, is not
1136 	 * in a reclaimed state.
1137 	 */
1138 	if (_cache_lock_shared_nonblock(ncp) == 0) {
1139 		if ((ncp->nc_lockstatus & NC_EXLOCK_REQ) == 0) {
1140 			if (ncp->nc_vp == NULL ||
1141 			    (ncp->nc_vp->v_flag & VRECLAIMED) == 0) {
1142 				return(0);
1143 			}
1144 		}
1145 		_cache_unlock(ncp);
1146 		return(EWOULDBLOCK);
1147 	}
1148 
1149 	/*
1150 	 * Non-blocking shared lock failed.  If we already own the exclusive
1151 	 * lock just acquire another exclusive lock (instead of deadlocking).
1152 	 * Otherwise acquire a shared lock.
1153 	 */
1154 	if (ncp->nc_locktd == curthread) {
1155 		_cache_lock(ncp);
1156 		return(0);
1157 	}
1158 	_cache_lock_shared(ncp);
1159 	return(0);
1160 }
1161 
1162 
1163 /*
1164  * NOTE: The same nchandle can be passed for both arguments.
1165  */
1166 void
1167 cache_get(struct nchandle *nch, struct nchandle *target)
1168 {
1169 	KKASSERT(nch->ncp->nc_refs > 0);
1170 	target->mount = nch->mount;
1171 	target->ncp = _cache_get(nch->ncp);
1172 	atomic_add_int(&target->mount->mnt_refs, 1);
1173 }
1174 
1175 void
1176 cache_get_maybe_shared(struct nchandle *nch, struct nchandle *target, int excl)
1177 {
1178 	KKASSERT(nch->ncp->nc_refs > 0);
1179 	target->mount = nch->mount;
1180 	target->ncp = _cache_get_maybe_shared(nch->ncp, excl);
1181 	atomic_add_int(&target->mount->mnt_refs, 1);
1182 }
1183 
1184 /*
1185  *
1186  */
1187 static __inline
1188 void
1189 _cache_put(struct namecache *ncp)
1190 {
1191 	_cache_unlock(ncp);
1192 	_cache_drop(ncp);
1193 }
1194 
1195 /*
1196  *
1197  */
1198 void
1199 cache_put(struct nchandle *nch)
1200 {
1201 	atomic_add_int(&nch->mount->mnt_refs, -1);
1202 	_cache_put(nch->ncp);
1203 	nch->ncp = NULL;
1204 	nch->mount = NULL;
1205 }
1206 
1207 /*
1208  * Resolve an unresolved ncp by associating a vnode with it.  If the
1209  * vnode is NULL, a negative cache entry is created.
1210  *
1211  * The ncp should be locked on entry and will remain locked on return.
1212  */
1213 static
1214 void
1215 _cache_setvp(struct mount *mp, struct namecache *ncp, struct vnode *vp)
1216 {
1217 	KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
1218 	KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE);
1219 
1220 	if (vp != NULL) {
1221 		/*
1222 		 * Any vp associated with an ncp which has children must
1223 		 * be held.  Any vp associated with a locked ncp must be held.
1224 		 */
1225 		if (!TAILQ_EMPTY(&ncp->nc_list))
1226 			vhold(vp);
1227 		spin_lock(&vp->v_spin);
1228 		ncp->nc_vp = vp;
1229 		TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode);
1230 		spin_unlock(&vp->v_spin);
1231 		if (ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ))
1232 			vhold(vp);
1233 
1234 		/*
1235 		 * Set auxiliary flags
1236 		 */
1237 		switch(vp->v_type) {
1238 		case VDIR:
1239 			ncp->nc_flag |= NCF_ISDIR;
1240 			break;
1241 		case VLNK:
1242 			ncp->nc_flag |= NCF_ISSYMLINK;
1243 			/* XXX cache the contents of the symlink */
1244 			break;
1245 		default:
1246 			break;
1247 		}
1248 		atomic_add_int(&numcache, 1);
1249 		ncp->nc_error = 0;
1250 		/* XXX: this is a hack to work-around the lack of a real pfs vfs
1251 		 * implementation*/
1252 		if (mp != NULL)
1253 			if (strncmp(mp->mnt_stat.f_fstypename, "null", 5) == 0)
1254 				vp->v_pfsmp = mp;
1255 	} else {
1256 		/*
1257 		 * When creating a negative cache hit we set the
1258 		 * namecache_gen.  A later resolve will clean out the
1259 		 * negative cache hit if the mount point's namecache_gen
1260 		 * has changed.  Used by devfs, could also be used by
1261 		 * other remote FSs.
1262 		 */
1263 		ncp->nc_vp = NULL;
1264 		spin_lock(&ncspin);
1265 		TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
1266 		++numneg;
1267 		spin_unlock(&ncspin);
1268 		ncp->nc_error = ENOENT;
1269 		if (mp)
1270 			VFS_NCPGEN_SET(mp, ncp);
1271 	}
1272 	ncp->nc_flag &= ~(NCF_UNRESOLVED | NCF_DEFEREDZAP);
1273 }
1274 
1275 /*
1276  *
1277  */
1278 void
1279 cache_setvp(struct nchandle *nch, struct vnode *vp)
1280 {
1281 	_cache_setvp(nch->mount, nch->ncp, vp);
1282 }
1283 
1284 /*
1285  *
1286  */
1287 void
1288 cache_settimeout(struct nchandle *nch, int nticks)
1289 {
1290 	struct namecache *ncp = nch->ncp;
1291 
1292 	if ((ncp->nc_timeout = ticks + nticks) == 0)
1293 		ncp->nc_timeout = 1;
1294 }
1295 
1296 /*
1297  * Disassociate the vnode or negative-cache association and mark a
1298  * namecache entry as unresolved again.  Note that the ncp is still
1299  * left in the hash table and still linked to its parent.
1300  *
1301  * The ncp should be locked and refd on entry and will remain locked and refd
1302  * on return.
1303  *
1304  * This routine is normally never called on a directory containing children.
1305  * However, NFS often does just that in its rename() code as a cop-out to
1306  * avoid complex namespace operations.  This disconnects a directory vnode
1307  * from its namecache and can cause the OLDAPI and NEWAPI to get out of
1308  * sync.
1309  *
1310  */
1311 static
1312 void
1313 _cache_setunresolved(struct namecache *ncp)
1314 {
1315 	struct vnode *vp;
1316 
1317 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1318 		ncp->nc_flag |= NCF_UNRESOLVED;
1319 		ncp->nc_timeout = 0;
1320 		ncp->nc_error = ENOTCONN;
1321 		if ((vp = ncp->nc_vp) != NULL) {
1322 			atomic_add_int(&numcache, -1);
1323 			spin_lock(&vp->v_spin);
1324 			ncp->nc_vp = NULL;
1325 			TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode);
1326 			spin_unlock(&vp->v_spin);
1327 
1328 			/*
1329 			 * Any vp associated with an ncp with children is
1330 			 * held by that ncp.  Any vp associated with a locked
1331 			 * ncp is held by that ncp.  These conditions must be
1332 			 * undone when the vp is cleared out from the ncp.
1333 			 */
1334 			if (!TAILQ_EMPTY(&ncp->nc_list))
1335 				vdrop(vp);
1336 			if (ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ))
1337 				vdrop(vp);
1338 		} else {
1339 			spin_lock(&ncspin);
1340 			TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
1341 			--numneg;
1342 			spin_unlock(&ncspin);
1343 		}
1344 		ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK);
1345 	}
1346 }
1347 
1348 /*
1349  * The cache_nresolve() code calls this function to automatically
1350  * set a resolved cache element to unresolved if it has timed out
1351  * or if it is a negative cache hit and the mount point namecache_gen
1352  * has changed.
1353  */
1354 static __inline int
1355 _cache_auto_unresolve_test(struct mount *mp, struct namecache *ncp)
1356 {
1357 	/*
1358 	 * Try to zap entries that have timed out.  We have
1359 	 * to be careful here because locked leafs may depend
1360 	 * on the vnode remaining intact in a parent, so only
1361 	 * do this under very specific conditions.
1362 	 */
1363 	if (ncp->nc_timeout && (int)(ncp->nc_timeout - ticks) < 0 &&
1364 	    TAILQ_EMPTY(&ncp->nc_list)) {
1365 		return 1;
1366 	}
1367 
1368 	/*
1369 	 * If a resolved negative cache hit is invalid due to
1370 	 * the mount's namecache generation being bumped, zap it.
1371 	 */
1372 	if (ncp->nc_vp == NULL && VFS_NCPGEN_TEST(mp, ncp)) {
1373 		return 1;
1374 	}
1375 
1376 	/*
1377 	 * Otherwise we are good
1378 	 */
1379 	return 0;
1380 }
1381 
1382 static __inline void
1383 _cache_auto_unresolve(struct mount *mp, struct namecache *ncp)
1384 {
1385 	/*
1386 	 * Already in an unresolved state, nothing to do.
1387 	 */
1388 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
1389 		if (_cache_auto_unresolve_test(mp, ncp))
1390 			_cache_setunresolved(ncp);
1391 	}
1392 }
1393 
1394 /*
1395  *
1396  */
1397 void
1398 cache_setunresolved(struct nchandle *nch)
1399 {
1400 	_cache_setunresolved(nch->ncp);
1401 }
1402 
1403 /*
1404  * Determine if we can clear NCF_ISMOUNTPT by scanning the mountlist
1405  * looking for matches.  This flag tells the lookup code when it must
1406  * check for a mount linkage and also prevents the directories in question
1407  * from being deleted or renamed.
1408  */
1409 static
1410 int
1411 cache_clrmountpt_callback(struct mount *mp, void *data)
1412 {
1413 	struct nchandle *nch = data;
1414 
1415 	if (mp->mnt_ncmounton.ncp == nch->ncp)
1416 		return(1);
1417 	if (mp->mnt_ncmountpt.ncp == nch->ncp)
1418 		return(1);
1419 	return(0);
1420 }
1421 
1422 /*
1423  *
1424  */
1425 void
1426 cache_clrmountpt(struct nchandle *nch)
1427 {
1428 	int count;
1429 
1430 	count = mountlist_scan(cache_clrmountpt_callback, nch,
1431 			       MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
1432 	if (count == 0)
1433 		nch->ncp->nc_flag &= ~NCF_ISMOUNTPT;
1434 }
1435 
1436 /*
1437  * Invalidate portions of the namecache topology given a starting entry.
1438  * The passed ncp is set to an unresolved state and:
1439  *
1440  * The passed ncp must be referencxed and locked.  The routine may unlock
1441  * and relock ncp several times, and will recheck the children and loop
1442  * to catch races.  When done the passed ncp will be returned with the
1443  * reference and lock intact.
1444  *
1445  * CINV_DESTROY		- Set a flag in the passed ncp entry indicating
1446  *			  that the physical underlying nodes have been
1447  *			  destroyed... as in deleted.  For example, when
1448  *			  a directory is removed.  This will cause record
1449  *			  lookups on the name to no longer be able to find
1450  *			  the record and tells the resolver to return failure
1451  *			  rather then trying to resolve through the parent.
1452  *
1453  *			  The topology itself, including ncp->nc_name,
1454  *			  remains intact.
1455  *
1456  *			  This only applies to the passed ncp, if CINV_CHILDREN
1457  *			  is specified the children are not flagged.
1458  *
1459  * CINV_CHILDREN	- Set all children (recursively) to an unresolved
1460  *			  state as well.
1461  *
1462  *			  Note that this will also have the side effect of
1463  *			  cleaning out any unreferenced nodes in the topology
1464  *			  from the leaves up as the recursion backs out.
1465  *
1466  * Note that the topology for any referenced nodes remains intact, but
1467  * the nodes will be marked as having been destroyed and will be set
1468  * to an unresolved state.
1469  *
1470  * It is possible for cache_inval() to race a cache_resolve(), meaning that
1471  * the namecache entry may not actually be invalidated on return if it was
1472  * revalidated while recursing down into its children.  This code guarentees
1473  * that the node(s) will go through an invalidation cycle, but does not
1474  * guarentee that they will remain in an invalidated state.
1475  *
1476  * Returns non-zero if a revalidation was detected during the invalidation
1477  * recursion, zero otherwise.  Note that since only the original ncp is
1478  * locked the revalidation ultimately can only indicate that the original ncp
1479  * *MIGHT* no have been reresolved.
1480  *
1481  * DEEP RECURSION HANDLING - If a recursive invalidation recurses deeply we
1482  * have to avoid blowing out the kernel stack.  We do this by saving the
1483  * deep namecache node and aborting the recursion, then re-recursing at that
1484  * node using a depth-first algorithm in order to allow multiple deep
1485  * recursions to chain through each other, then we restart the invalidation
1486  * from scratch.
1487  */
1488 
1489 struct cinvtrack {
1490 	struct namecache *resume_ncp;
1491 	int depth;
1492 };
1493 
1494 static int _cache_inval_internal(struct namecache *, int, struct cinvtrack *);
1495 
1496 static
1497 int
1498 _cache_inval(struct namecache *ncp, int flags)
1499 {
1500 	struct cinvtrack track;
1501 	struct namecache *ncp2;
1502 	int r;
1503 
1504 	track.depth = 0;
1505 	track.resume_ncp = NULL;
1506 
1507 	for (;;) {
1508 		r = _cache_inval_internal(ncp, flags, &track);
1509 		if (track.resume_ncp == NULL)
1510 			break;
1511 		kprintf("Warning: deep namecache recursion at %s\n",
1512 			ncp->nc_name);
1513 		_cache_unlock(ncp);
1514 		while ((ncp2 = track.resume_ncp) != NULL) {
1515 			track.resume_ncp = NULL;
1516 			_cache_lock(ncp2);
1517 			_cache_inval_internal(ncp2, flags & ~CINV_DESTROY,
1518 					     &track);
1519 			_cache_put(ncp2);
1520 		}
1521 		_cache_lock(ncp);
1522 	}
1523 	return(r);
1524 }
1525 
1526 int
1527 cache_inval(struct nchandle *nch, int flags)
1528 {
1529 	return(_cache_inval(nch->ncp, flags));
1530 }
1531 
1532 /*
1533  * Helper for _cache_inval().  The passed ncp is refd and locked and
1534  * remains that way on return, but may be unlocked/relocked multiple
1535  * times by the routine.
1536  */
1537 static int
1538 _cache_inval_internal(struct namecache *ncp, int flags, struct cinvtrack *track)
1539 {
1540 	struct namecache *nextkid;
1541 	int rcnt = 0;
1542 
1543 	KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE);
1544 
1545 	_cache_setunresolved(ncp);
1546 	if (flags & CINV_DESTROY) {
1547 		ncp->nc_flag |= NCF_DESTROYED;
1548 		++ncp->nc_generation;
1549 	}
1550 	while ((flags & CINV_CHILDREN) &&
1551 	       (nextkid = TAILQ_FIRST(&ncp->nc_list)) != NULL
1552 	) {
1553 		struct namecache *kid;
1554 		int restart;
1555 
1556 		restart = 0;
1557 		_cache_hold(nextkid);
1558 		if (++track->depth > MAX_RECURSION_DEPTH) {
1559 			track->resume_ncp = ncp;
1560 			_cache_hold(ncp);
1561 			++rcnt;
1562 		}
1563 		while ((kid = nextkid) != NULL) {
1564 			/*
1565 			 * Parent (ncp) must be locked for the iteration.
1566 			 */
1567 			nextkid = NULL;
1568 			if (kid->nc_parent != ncp) {
1569 				_cache_drop(kid);
1570 				kprintf("cache_inval_internal restartA %s\n", ncp->nc_name);
1571 				restart = 1;
1572 				break;
1573 			}
1574 			if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL)
1575 				_cache_hold(nextkid);
1576 
1577 			/*
1578 			 * Parent unlocked for this section to avoid
1579 			 * deadlocks.
1580 			 */
1581 			_cache_unlock(ncp);
1582 			if (track->resume_ncp) {
1583 				_cache_drop(kid);
1584 				_cache_lock(ncp);
1585 				break;
1586 			}
1587 			if ((kid->nc_flag & NCF_UNRESOLVED) == 0 ||
1588 			    TAILQ_FIRST(&kid->nc_list)
1589 			) {
1590 				_cache_lock(kid);
1591 				if (kid->nc_parent != ncp) {
1592 					kprintf("cache_inval_internal restartB %s\n", ncp->nc_name);
1593 					restart = 1;
1594 					_cache_unlock(kid);
1595 					_cache_drop(kid);
1596 					_cache_lock(ncp);
1597 					break;
1598 				}
1599 
1600 				rcnt += _cache_inval_internal(kid, flags & ~CINV_DESTROY, track);
1601 				_cache_unlock(kid);
1602 			}
1603 			_cache_drop(kid);
1604 			_cache_lock(ncp);
1605 		}
1606 		if (nextkid)
1607 			_cache_drop(nextkid);
1608 		--track->depth;
1609 		if (restart == 0)
1610 			break;
1611 	}
1612 
1613 	/*
1614 	 * Someone could have gotten in there while ncp was unlocked,
1615 	 * retry if so.
1616 	 */
1617 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
1618 		++rcnt;
1619 	return (rcnt);
1620 }
1621 
1622 /*
1623  * Invalidate a vnode's namecache associations.  To avoid races against
1624  * the resolver we do not invalidate a node which we previously invalidated
1625  * but which was then re-resolved while we were in the invalidation loop.
1626  *
1627  * Returns non-zero if any namecache entries remain after the invalidation
1628  * loop completed.
1629  *
1630  * NOTE: Unlike the namecache topology which guarentees that ncp's will not
1631  *	 be ripped out of the topology while held, the vnode's v_namecache
1632  *	 list has no such restriction.  NCP's can be ripped out of the list
1633  *	 at virtually any time if not locked, even if held.
1634  *
1635  *	 In addition, the v_namecache list itself must be locked via
1636  *	 the vnode's spinlock.
1637  */
1638 int
1639 cache_inval_vp(struct vnode *vp, int flags)
1640 {
1641 	struct namecache *ncp;
1642 	struct namecache *next;
1643 
1644 restart:
1645 	spin_lock(&vp->v_spin);
1646 	ncp = TAILQ_FIRST(&vp->v_namecache);
1647 	if (ncp)
1648 		_cache_hold(ncp);
1649 	while (ncp) {
1650 		/* loop entered with ncp held and vp spin-locked */
1651 		if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
1652 			_cache_hold(next);
1653 		spin_unlock(&vp->v_spin);
1654 		_cache_lock(ncp);
1655 		if (ncp->nc_vp != vp) {
1656 			kprintf("Warning: cache_inval_vp: race-A detected on "
1657 				"%s\n", ncp->nc_name);
1658 			_cache_put(ncp);
1659 			if (next)
1660 				_cache_drop(next);
1661 			goto restart;
1662 		}
1663 		_cache_inval(ncp, flags);
1664 		_cache_put(ncp);		/* also releases reference */
1665 		ncp = next;
1666 		spin_lock(&vp->v_spin);
1667 		if (ncp && ncp->nc_vp != vp) {
1668 			spin_unlock(&vp->v_spin);
1669 			kprintf("Warning: cache_inval_vp: race-B detected on "
1670 				"%s\n", ncp->nc_name);
1671 			_cache_drop(ncp);
1672 			goto restart;
1673 		}
1674 	}
1675 	spin_unlock(&vp->v_spin);
1676 	return(TAILQ_FIRST(&vp->v_namecache) != NULL);
1677 }
1678 
1679 /*
1680  * This routine is used instead of the normal cache_inval_vp() when we
1681  * are trying to recycle otherwise good vnodes.
1682  *
1683  * Return 0 on success, non-zero if not all namecache records could be
1684  * disassociated from the vnode (for various reasons).
1685  */
1686 int
1687 cache_inval_vp_nonblock(struct vnode *vp)
1688 {
1689 	struct namecache *ncp;
1690 	struct namecache *next;
1691 
1692 	spin_lock(&vp->v_spin);
1693 	ncp = TAILQ_FIRST(&vp->v_namecache);
1694 	if (ncp)
1695 		_cache_hold(ncp);
1696 	while (ncp) {
1697 		/* loop entered with ncp held */
1698 		if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL)
1699 			_cache_hold(next);
1700 		spin_unlock(&vp->v_spin);
1701 		if (_cache_lock_nonblock(ncp)) {
1702 			_cache_drop(ncp);
1703 			if (next)
1704 				_cache_drop(next);
1705 			goto done;
1706 		}
1707 		if (ncp->nc_vp != vp) {
1708 			kprintf("Warning: cache_inval_vp: race-A detected on "
1709 				"%s\n", ncp->nc_name);
1710 			_cache_put(ncp);
1711 			if (next)
1712 				_cache_drop(next);
1713 			goto done;
1714 		}
1715 		_cache_inval(ncp, 0);
1716 		_cache_put(ncp);		/* also releases reference */
1717 		ncp = next;
1718 		spin_lock(&vp->v_spin);
1719 		if (ncp && ncp->nc_vp != vp) {
1720 			spin_unlock(&vp->v_spin);
1721 			kprintf("Warning: cache_inval_vp: race-B detected on "
1722 				"%s\n", ncp->nc_name);
1723 			_cache_drop(ncp);
1724 			goto done;
1725 		}
1726 	}
1727 	spin_unlock(&vp->v_spin);
1728 done:
1729 	return(TAILQ_FIRST(&vp->v_namecache) != NULL);
1730 }
1731 
1732 /*
1733  * The source ncp has been renamed to the target ncp.  Both fncp and tncp
1734  * must be locked.  The target ncp is destroyed (as a normal rename-over
1735  * would destroy the target file or directory).
1736  *
1737  * Because there may be references to the source ncp we cannot copy its
1738  * contents to the target.  Instead the source ncp is relinked as the target
1739  * and the target ncp is removed from the namecache topology.
1740  */
1741 void
1742 cache_rename(struct nchandle *fnch, struct nchandle *tnch)
1743 {
1744 	struct namecache *fncp = fnch->ncp;
1745 	struct namecache *tncp = tnch->ncp;
1746 	struct namecache *tncp_par;
1747 	struct nchash_head *nchpp;
1748 	u_int32_t hash;
1749 	char *oname;
1750 	char *nname;
1751 
1752 	++fncp->nc_generation;
1753 	++tncp->nc_generation;
1754 	if (tncp->nc_nlen) {
1755 		nname = kmalloc(tncp->nc_nlen + 1, M_VFSCACHE, M_WAITOK);
1756 		bcopy(tncp->nc_name, nname, tncp->nc_nlen);
1757 		nname[tncp->nc_nlen] = 0;
1758 	} else {
1759 		nname = NULL;
1760 	}
1761 
1762 	/*
1763 	 * Rename fncp (unlink)
1764 	 */
1765 	_cache_unlink_parent(fncp);
1766 	oname = fncp->nc_name;
1767 	fncp->nc_name = nname;
1768 	fncp->nc_nlen = tncp->nc_nlen;
1769 	if (oname)
1770 		kfree(oname, M_VFSCACHE);
1771 
1772 	tncp_par = tncp->nc_parent;
1773 	_cache_hold(tncp_par);
1774 	_cache_lock(tncp_par);
1775 
1776 	/*
1777 	 * Rename fncp (relink)
1778 	 */
1779 	hash = fnv_32_buf(fncp->nc_name, fncp->nc_nlen, FNV1_32_INIT);
1780 	hash = fnv_32_buf(&tncp_par, sizeof(tncp_par), hash);
1781 	nchpp = NCHHASH(hash);
1782 
1783 	spin_lock(&nchpp->spin);
1784 	_cache_link_parent(fncp, tncp_par, nchpp);
1785 	spin_unlock(&nchpp->spin);
1786 
1787 	_cache_put(tncp_par);
1788 
1789 	/*
1790 	 * Get rid of the overwritten tncp (unlink)
1791 	 */
1792 	_cache_unlink(tncp);
1793 }
1794 
1795 /*
1796  * Perform actions consistent with unlinking a file.  The passed-in ncp
1797  * must be locked.
1798  *
1799  * The ncp is marked DESTROYED so it no longer shows up in searches,
1800  * and will be physically deleted when the vnode goes away.
1801  *
1802  * If the related vnode has no refs then we cycle it through vget()/vput()
1803  * to (possibly if we don't have a ref race) trigger a deactivation,
1804  * allowing the VFS to trivially detect and recycle the deleted vnode
1805  * via VOP_INACTIVE().
1806  *
1807  * NOTE: _cache_rename() will automatically call _cache_unlink() on the
1808  *	 target ncp.
1809  */
1810 void
1811 cache_unlink(struct nchandle *nch)
1812 {
1813 	_cache_unlink(nch->ncp);
1814 }
1815 
1816 static void
1817 _cache_unlink(struct namecache *ncp)
1818 {
1819 	struct vnode *vp;
1820 
1821 	/*
1822 	 * Causes lookups to fail and allows another ncp with the same
1823 	 * name to be created under ncp->nc_parent.
1824 	 */
1825 	ncp->nc_flag |= NCF_DESTROYED;
1826 	++ncp->nc_generation;
1827 
1828 	/*
1829 	 * Attempt to trigger a deactivation.  Set VREF_FINALIZE to
1830 	 * force action on the 1->0 transition.
1831 	 */
1832 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
1833 	    (vp = ncp->nc_vp) != NULL) {
1834 		atomic_set_int(&vp->v_refcnt, VREF_FINALIZE);
1835 		if (VREFCNT(vp) <= 0) {
1836 			if (vget(vp, LK_SHARED) == 0)
1837 				vput(vp);
1838 		}
1839 	}
1840 }
1841 
1842 /*
1843  * Return non-zero if the nch might be associated with an open and/or mmap()'d
1844  * file.  The easy solution is to just return non-zero if the vnode has refs.
1845  * Used to interlock hammer2 reclaims (VREF_FINALIZE should already be set to
1846  * force the reclaim).
1847  */
1848 int
1849 cache_isopen(struct nchandle *nch)
1850 {
1851 	struct vnode *vp;
1852 	struct namecache *ncp = nch->ncp;
1853 
1854 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
1855 	    (vp = ncp->nc_vp) != NULL &&
1856 	    VREFCNT(vp)) {
1857 		return 1;
1858 	}
1859 	return 0;
1860 }
1861 
1862 
1863 /*
1864  * vget the vnode associated with the namecache entry.  Resolve the namecache
1865  * entry if necessary.  The passed ncp must be referenced and locked.  If
1866  * the ncp is resolved it might be locked shared.
1867  *
1868  * lk_type may be LK_SHARED, LK_EXCLUSIVE.  A ref'd, possibly locked
1869  * (depending on the passed lk_type) will be returned in *vpp with an error
1870  * of 0, or NULL will be returned in *vpp with a non-0 error code.  The
1871  * most typical error is ENOENT, meaning that the ncp represents a negative
1872  * cache hit and there is no vnode to retrieve, but other errors can occur
1873  * too.
1874  *
1875  * The vget() can race a reclaim.  If this occurs we re-resolve the
1876  * namecache entry.
1877  *
1878  * There are numerous places in the kernel where vget() is called on a
1879  * vnode while one or more of its namecache entries is locked.  Releasing
1880  * a vnode never deadlocks against locked namecache entries (the vnode
1881  * will not get recycled while referenced ncp's exist).  This means we
1882  * can safely acquire the vnode.  In fact, we MUST NOT release the ncp
1883  * lock when acquiring the vp lock or we might cause a deadlock.
1884  *
1885  * NOTE: The passed-in ncp must be locked exclusively if it is initially
1886  *	 unresolved.  If a reclaim race occurs the passed-in ncp will be
1887  *	 relocked exclusively before being re-resolved.
1888  */
1889 int
1890 cache_vget(struct nchandle *nch, struct ucred *cred,
1891 	   int lk_type, struct vnode **vpp)
1892 {
1893 	struct namecache *ncp;
1894 	struct vnode *vp;
1895 	int error;
1896 
1897 	ncp = nch->ncp;
1898 again:
1899 	vp = NULL;
1900 	if (ncp->nc_flag & NCF_UNRESOLVED)
1901 		error = cache_resolve(nch, cred);
1902 	else
1903 		error = 0;
1904 
1905 	if (error == 0 && (vp = ncp->nc_vp) != NULL) {
1906 		error = vget(vp, lk_type);
1907 		if (error) {
1908 			/*
1909 			 * VRECLAIM race
1910 			 *
1911 			 * The ncp may have been locked shared, we must relock
1912 			 * it exclusively before we can set it to unresolved.
1913 			 */
1914 			if (error == ENOENT) {
1915 				kprintf("Warning: vnode reclaim race detected "
1916 					"in cache_vget on %p (%s)\n",
1917 					vp, ncp->nc_name);
1918 				_cache_unlock(ncp);
1919 				_cache_lock(ncp);
1920 				_cache_setunresolved(ncp);
1921 				goto again;
1922 			}
1923 
1924 			/*
1925 			 * Not a reclaim race, some other error.
1926 			 */
1927 			KKASSERT(ncp->nc_vp == vp);
1928 			vp = NULL;
1929 		} else {
1930 			KKASSERT(ncp->nc_vp == vp);
1931 			KKASSERT((vp->v_flag & VRECLAIMED) == 0);
1932 		}
1933 	}
1934 	if (error == 0 && vp == NULL)
1935 		error = ENOENT;
1936 	*vpp = vp;
1937 	return(error);
1938 }
1939 
1940 /*
1941  * Similar to cache_vget() but only acquires a ref on the vnode.
1942  *
1943  * NOTE: The passed-in ncp must be locked exclusively if it is initially
1944  *	 unresolved.  If a reclaim race occurs the passed-in ncp will be
1945  *	 relocked exclusively before being re-resolved.
1946  */
1947 int
1948 cache_vref(struct nchandle *nch, struct ucred *cred, struct vnode **vpp)
1949 {
1950 	struct namecache *ncp;
1951 	struct vnode *vp;
1952 	int error;
1953 
1954 	ncp = nch->ncp;
1955 again:
1956 	vp = NULL;
1957 	if (ncp->nc_flag & NCF_UNRESOLVED)
1958 		error = cache_resolve(nch, cred);
1959 	else
1960 		error = 0;
1961 
1962 	if (error == 0 && (vp = ncp->nc_vp) != NULL) {
1963 		error = vget(vp, LK_SHARED);
1964 		if (error) {
1965 			/*
1966 			 * VRECLAIM race
1967 			 */
1968 			if (error == ENOENT) {
1969 				kprintf("Warning: vnode reclaim race detected "
1970 					"in cache_vget on %p (%s)\n",
1971 					vp, ncp->nc_name);
1972 				_cache_unlock(ncp);
1973 				_cache_lock(ncp);
1974 				_cache_setunresolved(ncp);
1975 				goto again;
1976 			}
1977 
1978 			/*
1979 			 * Not a reclaim race, some other error.
1980 			 */
1981 			KKASSERT(ncp->nc_vp == vp);
1982 			vp = NULL;
1983 		} else {
1984 			KKASSERT(ncp->nc_vp == vp);
1985 			KKASSERT((vp->v_flag & VRECLAIMED) == 0);
1986 			/* caller does not want a lock */
1987 			vn_unlock(vp);
1988 		}
1989 	}
1990 	if (error == 0 && vp == NULL)
1991 		error = ENOENT;
1992 	*vpp = vp;
1993 	return(error);
1994 }
1995 
1996 /*
1997  * Return a referenced vnode representing the parent directory of
1998  * ncp.
1999  *
2000  * Because the caller has locked the ncp it should not be possible for
2001  * the parent ncp to go away.  However, the parent can unresolve its
2002  * dvp at any time so we must be able to acquire a lock on the parent
2003  * to safely access nc_vp.
2004  *
2005  * We have to leave par unlocked when vget()ing dvp to avoid a deadlock,
2006  * so use vhold()/vdrop() while holding the lock to prevent dvp from
2007  * getting destroyed.
2008  *
2009  * NOTE: vhold() is allowed when dvp has 0 refs if we hold a
2010  *	 lock on the ncp in question..
2011  */
2012 static struct vnode *
2013 cache_dvpref(struct namecache *ncp)
2014 {
2015 	struct namecache *par;
2016 	struct vnode *dvp;
2017 
2018 	dvp = NULL;
2019 	if ((par = ncp->nc_parent) != NULL) {
2020 		_cache_hold(par);
2021 		_cache_lock(par);
2022 		if ((par->nc_flag & NCF_UNRESOLVED) == 0) {
2023 			if ((dvp = par->nc_vp) != NULL)
2024 				vhold(dvp);
2025 		}
2026 		_cache_unlock(par);
2027 		if (dvp) {
2028 			if (vget(dvp, LK_SHARED) == 0) {
2029 				vn_unlock(dvp);
2030 				vdrop(dvp);
2031 				/* return refd, unlocked dvp */
2032 			} else {
2033 				vdrop(dvp);
2034 				dvp = NULL;
2035 			}
2036 		}
2037 		_cache_drop(par);
2038 	}
2039 	return(dvp);
2040 }
2041 
2042 /*
2043  * Convert a directory vnode to a namecache record without any other
2044  * knowledge of the topology.  This ONLY works with directory vnodes and
2045  * is ONLY used by the NFS server.  dvp must be refd but unlocked, and the
2046  * returned ncp (if not NULL) will be held and unlocked.
2047  *
2048  * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned.
2049  * If 'makeit' is 1 we attempt to track-down and create the namecache topology
2050  * for dvp.  This will fail only if the directory has been deleted out from
2051  * under the caller.
2052  *
2053  * Callers must always check for a NULL return no matter the value of 'makeit'.
2054  *
2055  * To avoid underflowing the kernel stack each recursive call increments
2056  * the makeit variable.
2057  */
2058 
2059 static int cache_inefficient_scan(struct nchandle *nch, struct ucred *cred,
2060 				  struct vnode *dvp, char *fakename);
2061 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
2062 				  struct vnode **saved_dvp);
2063 
2064 int
2065 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit,
2066 	      struct nchandle *nch)
2067 {
2068 	struct vnode *saved_dvp;
2069 	struct vnode *pvp;
2070 	char *fakename;
2071 	int error;
2072 
2073 	nch->ncp = NULL;
2074 	nch->mount = dvp->v_mount;
2075 	saved_dvp = NULL;
2076 	fakename = NULL;
2077 
2078 	/*
2079 	 * Handle the makeit == 0 degenerate case
2080 	 */
2081 	if (makeit == 0) {
2082 		spin_lock_shared(&dvp->v_spin);
2083 		nch->ncp = TAILQ_FIRST(&dvp->v_namecache);
2084 		if (nch->ncp)
2085 			cache_hold(nch);
2086 		spin_unlock_shared(&dvp->v_spin);
2087 	}
2088 
2089 	/*
2090 	 * Loop until resolution, inside code will break out on error.
2091 	 */
2092 	while (makeit) {
2093 		/*
2094 		 * Break out if we successfully acquire a working ncp.
2095 		 */
2096 		spin_lock_shared(&dvp->v_spin);
2097 		nch->ncp = TAILQ_FIRST(&dvp->v_namecache);
2098 		if (nch->ncp) {
2099 			cache_hold(nch);
2100 			spin_unlock_shared(&dvp->v_spin);
2101 			break;
2102 		}
2103 		spin_unlock_shared(&dvp->v_spin);
2104 
2105 		/*
2106 		 * If dvp is the root of its filesystem it should already
2107 		 * have a namecache pointer associated with it as a side
2108 		 * effect of the mount, but it may have been disassociated.
2109 		 */
2110 		if (dvp->v_flag & VROOT) {
2111 			nch->ncp = _cache_get(nch->mount->mnt_ncmountpt.ncp);
2112 			error = cache_resolve_mp(nch->mount);
2113 			_cache_put(nch->ncp);
2114 			if (ncvp_debug) {
2115 				kprintf("cache_fromdvp: resolve root of mount %p error %d",
2116 					dvp->v_mount, error);
2117 			}
2118 			if (error) {
2119 				if (ncvp_debug)
2120 					kprintf(" failed\n");
2121 				nch->ncp = NULL;
2122 				break;
2123 			}
2124 			if (ncvp_debug)
2125 				kprintf(" succeeded\n");
2126 			continue;
2127 		}
2128 
2129 		/*
2130 		 * If we are recursed too deeply resort to an O(n^2)
2131 		 * algorithm to resolve the namecache topology.  The
2132 		 * resolved pvp is left referenced in saved_dvp to
2133 		 * prevent the tree from being destroyed while we loop.
2134 		 */
2135 		if (makeit > 20) {
2136 			error = cache_fromdvp_try(dvp, cred, &saved_dvp);
2137 			if (error) {
2138 				kprintf("lookupdotdot(longpath) failed %d "
2139 				       "dvp %p\n", error, dvp);
2140 				nch->ncp = NULL;
2141 				break;
2142 			}
2143 			continue;
2144 		}
2145 
2146 		/*
2147 		 * Get the parent directory and resolve its ncp.
2148 		 */
2149 		if (fakename) {
2150 			kfree(fakename, M_TEMP);
2151 			fakename = NULL;
2152 		}
2153 		error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred,
2154 					  &fakename);
2155 		if (error) {
2156 			kprintf("lookupdotdot failed %d dvp %p\n", error, dvp);
2157 			break;
2158 		}
2159 		vn_unlock(pvp);
2160 
2161 		/*
2162 		 * Reuse makeit as a recursion depth counter.  On success
2163 		 * nch will be fully referenced.
2164 		 */
2165 		cache_fromdvp(pvp, cred, makeit + 1, nch);
2166 		vrele(pvp);
2167 		if (nch->ncp == NULL)
2168 			break;
2169 
2170 		/*
2171 		 * Do an inefficient scan of pvp (embodied by ncp) to look
2172 		 * for dvp.  This will create a namecache record for dvp on
2173 		 * success.  We loop up to recheck on success.
2174 		 *
2175 		 * ncp and dvp are both held but not locked.
2176 		 */
2177 		error = cache_inefficient_scan(nch, cred, dvp, fakename);
2178 		if (error) {
2179 			kprintf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n",
2180 				pvp, nch->ncp->nc_name, dvp);
2181 			cache_drop(nch);
2182 			/* nch was NULLed out, reload mount */
2183 			nch->mount = dvp->v_mount;
2184 			break;
2185 		}
2186 		if (ncvp_debug) {
2187 			kprintf("cache_fromdvp: scan %p (%s) succeeded\n",
2188 				pvp, nch->ncp->nc_name);
2189 		}
2190 		cache_drop(nch);
2191 		/* nch was NULLed out, reload mount */
2192 		nch->mount = dvp->v_mount;
2193 	}
2194 
2195 	/*
2196 	 * If nch->ncp is non-NULL it will have been held already.
2197 	 */
2198 	if (fakename)
2199 		kfree(fakename, M_TEMP);
2200 	if (saved_dvp)
2201 		vrele(saved_dvp);
2202 	if (nch->ncp)
2203 		return (0);
2204 	return (EINVAL);
2205 }
2206 
2207 /*
2208  * Go up the chain of parent directories until we find something
2209  * we can resolve into the namecache.  This is very inefficient.
2210  */
2211 static
2212 int
2213 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred,
2214 		  struct vnode **saved_dvp)
2215 {
2216 	struct nchandle nch;
2217 	struct vnode *pvp;
2218 	int error;
2219 	static time_t last_fromdvp_report;
2220 	char *fakename;
2221 
2222 	/*
2223 	 * Loop getting the parent directory vnode until we get something we
2224 	 * can resolve in the namecache.
2225 	 */
2226 	vref(dvp);
2227 	nch.mount = dvp->v_mount;
2228 	nch.ncp = NULL;
2229 	fakename = NULL;
2230 
2231 	for (;;) {
2232 		if (fakename) {
2233 			kfree(fakename, M_TEMP);
2234 			fakename = NULL;
2235 		}
2236 		error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred,
2237 					  &fakename);
2238 		if (error) {
2239 			vrele(dvp);
2240 			break;
2241 		}
2242 		vn_unlock(pvp);
2243 		spin_lock_shared(&pvp->v_spin);
2244 		if ((nch.ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) {
2245 			_cache_hold(nch.ncp);
2246 			spin_unlock_shared(&pvp->v_spin);
2247 			vrele(pvp);
2248 			break;
2249 		}
2250 		spin_unlock_shared(&pvp->v_spin);
2251 		if (pvp->v_flag & VROOT) {
2252 			nch.ncp = _cache_get(pvp->v_mount->mnt_ncmountpt.ncp);
2253 			error = cache_resolve_mp(nch.mount);
2254 			_cache_unlock(nch.ncp);
2255 			vrele(pvp);
2256 			if (error) {
2257 				_cache_drop(nch.ncp);
2258 				nch.ncp = NULL;
2259 				vrele(dvp);
2260 			}
2261 			break;
2262 		}
2263 		vrele(dvp);
2264 		dvp = pvp;
2265 	}
2266 	if (error == 0) {
2267 		if (last_fromdvp_report != time_uptime) {
2268 			last_fromdvp_report = time_uptime;
2269 			kprintf("Warning: extremely inefficient path "
2270 				"resolution on %s\n",
2271 				nch.ncp->nc_name);
2272 		}
2273 		error = cache_inefficient_scan(&nch, cred, dvp, fakename);
2274 
2275 		/*
2276 		 * Hopefully dvp now has a namecache record associated with
2277 		 * it.  Leave it referenced to prevent the kernel from
2278 		 * recycling the vnode.  Otherwise extremely long directory
2279 		 * paths could result in endless recycling.
2280 		 */
2281 		if (*saved_dvp)
2282 		    vrele(*saved_dvp);
2283 		*saved_dvp = dvp;
2284 		_cache_drop(nch.ncp);
2285 	}
2286 	if (fakename)
2287 		kfree(fakename, M_TEMP);
2288 	return (error);
2289 }
2290 
2291 /*
2292  * Do an inefficient scan of the directory represented by ncp looking for
2293  * the directory vnode dvp.  ncp must be held but not locked on entry and
2294  * will be held on return.  dvp must be refd but not locked on entry and
2295  * will remain refd on return.
2296  *
2297  * Why do this at all?  Well, due to its stateless nature the NFS server
2298  * converts file handles directly to vnodes without necessarily going through
2299  * the namecache ops that would otherwise create the namecache topology
2300  * leading to the vnode.  We could either (1) Change the namecache algorithms
2301  * to allow disconnect namecache records that are re-merged opportunistically,
2302  * or (2) Make the NFS server backtrack and scan to recover a connected
2303  * namecache topology in order to then be able to issue new API lookups.
2304  *
2305  * It turns out that (1) is a huge mess.  It takes a nice clean set of
2306  * namecache algorithms and introduces a lot of complication in every subsystem
2307  * that calls into the namecache to deal with the re-merge case, especially
2308  * since we are using the namecache to placehold negative lookups and the
2309  * vnode might not be immediately assigned. (2) is certainly far less
2310  * efficient then (1), but since we are only talking about directories here
2311  * (which are likely to remain cached), the case does not actually run all
2312  * that often and has the supreme advantage of not polluting the namecache
2313  * algorithms.
2314  *
2315  * If a fakename is supplied just construct a namecache entry using the
2316  * fake name.
2317  */
2318 static int
2319 cache_inefficient_scan(struct nchandle *nch, struct ucred *cred,
2320 		       struct vnode *dvp, char *fakename)
2321 {
2322 	struct nlcomponent nlc;
2323 	struct nchandle rncp;
2324 	struct dirent *den;
2325 	struct vnode *pvp;
2326 	struct vattr vat;
2327 	struct iovec iov;
2328 	struct uio uio;
2329 	int blksize;
2330 	int eofflag;
2331 	int bytes;
2332 	char *rbuf;
2333 	int error;
2334 
2335 	vat.va_blocksize = 0;
2336 	if ((error = VOP_GETATTR(dvp, &vat)) != 0)
2337 		return (error);
2338 	cache_lock(nch);
2339 	error = cache_vref(nch, cred, &pvp);
2340 	cache_unlock(nch);
2341 	if (error)
2342 		return (error);
2343 	if (ncvp_debug) {
2344 		kprintf("inefficient_scan of (%p,%s): directory iosize %ld "
2345 			"vattr fileid = %lld\n",
2346 			nch->ncp, nch->ncp->nc_name,
2347 			vat.va_blocksize,
2348 			(long long)vat.va_fileid);
2349 	}
2350 
2351 	/*
2352 	 * Use the supplied fakename if not NULL.  Fake names are typically
2353 	 * not in the actual filesystem hierarchy.  This is used by HAMMER
2354 	 * to glue @@timestamp recursions together.
2355 	 */
2356 	if (fakename) {
2357 		nlc.nlc_nameptr = fakename;
2358 		nlc.nlc_namelen = strlen(fakename);
2359 		rncp = cache_nlookup(nch, &nlc);
2360 		goto done;
2361 	}
2362 
2363 	if ((blksize = vat.va_blocksize) == 0)
2364 		blksize = DEV_BSIZE;
2365 	rbuf = kmalloc(blksize, M_TEMP, M_WAITOK);
2366 	rncp.ncp = NULL;
2367 
2368 	eofflag = 0;
2369 	uio.uio_offset = 0;
2370 again:
2371 	iov.iov_base = rbuf;
2372 	iov.iov_len = blksize;
2373 	uio.uio_iov = &iov;
2374 	uio.uio_iovcnt = 1;
2375 	uio.uio_resid = blksize;
2376 	uio.uio_segflg = UIO_SYSSPACE;
2377 	uio.uio_rw = UIO_READ;
2378 	uio.uio_td = curthread;
2379 
2380 	if (ncvp_debug >= 2)
2381 		kprintf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset);
2382 	error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL);
2383 	if (error == 0) {
2384 		den = (struct dirent *)rbuf;
2385 		bytes = blksize - uio.uio_resid;
2386 
2387 		while (bytes > 0) {
2388 			if (ncvp_debug >= 2) {
2389 				kprintf("cache_inefficient_scan: %*.*s\n",
2390 					den->d_namlen, den->d_namlen,
2391 					den->d_name);
2392 			}
2393 			if (den->d_type != DT_WHT &&
2394 			    den->d_ino == vat.va_fileid) {
2395 				if (ncvp_debug) {
2396 					kprintf("cache_inefficient_scan: "
2397 					       "MATCHED inode %lld path %s/%*.*s\n",
2398 					       (long long)vat.va_fileid,
2399 					       nch->ncp->nc_name,
2400 					       den->d_namlen, den->d_namlen,
2401 					       den->d_name);
2402 				}
2403 				nlc.nlc_nameptr = den->d_name;
2404 				nlc.nlc_namelen = den->d_namlen;
2405 				rncp = cache_nlookup(nch, &nlc);
2406 				KKASSERT(rncp.ncp != NULL);
2407 				break;
2408 			}
2409 			bytes -= _DIRENT_DIRSIZ(den);
2410 			den = _DIRENT_NEXT(den);
2411 		}
2412 		if (rncp.ncp == NULL && eofflag == 0 && uio.uio_resid != blksize)
2413 			goto again;
2414 	}
2415 	kfree(rbuf, M_TEMP);
2416 done:
2417 	vrele(pvp);
2418 	if (rncp.ncp) {
2419 		if (rncp.ncp->nc_flag & NCF_UNRESOLVED) {
2420 			_cache_setvp(rncp.mount, rncp.ncp, dvp);
2421 			if (ncvp_debug >= 2) {
2422 				kprintf("cache_inefficient_scan: setvp %s/%s = %p\n",
2423 					nch->ncp->nc_name, rncp.ncp->nc_name, dvp);
2424 			}
2425 		} else {
2426 			if (ncvp_debug >= 2) {
2427 				kprintf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n",
2428 					nch->ncp->nc_name, rncp.ncp->nc_name, dvp,
2429 					rncp.ncp->nc_vp);
2430 			}
2431 		}
2432 		if (rncp.ncp->nc_vp == NULL)
2433 			error = rncp.ncp->nc_error;
2434 		/*
2435 		 * Release rncp after a successful nlookup.  rncp was fully
2436 		 * referenced.
2437 		 */
2438 		cache_put(&rncp);
2439 	} else {
2440 		kprintf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n",
2441 			dvp, nch->ncp->nc_name);
2442 		error = ENOENT;
2443 	}
2444 	return (error);
2445 }
2446 
2447 /*
2448  * Zap a namecache entry.  The ncp is unconditionally set to an unresolved
2449  * state, which disassociates it from its vnode or ncneglist.
2450  *
2451  * Then, if there are no additional references to the ncp and no children,
2452  * the ncp is removed from the topology and destroyed.
2453  *
2454  * References and/or children may exist if the ncp is in the middle of the
2455  * topology, preventing the ncp from being destroyed.
2456  *
2457  * This function must be called with the ncp held and locked and will unlock
2458  * and drop it during zapping.
2459  *
2460  * If nonblock is non-zero and the parent ncp cannot be locked we give up.
2461  * This case can occur in the cache_drop() path.
2462  *
2463  * This function may returned a held (but NOT locked) parent node which the
2464  * caller must drop.  We do this so _cache_drop() can loop, to avoid
2465  * blowing out the kernel stack.
2466  *
2467  * WARNING!  For MPSAFE operation this routine must acquire up to three
2468  *	     spin locks to be able to safely test nc_refs.  Lock order is
2469  *	     very important.
2470  *
2471  *	     hash spinlock if on hash list
2472  *	     parent spinlock if child of parent
2473  *	     (the ncp is unresolved so there is no vnode association)
2474  */
2475 static struct namecache *
2476 cache_zap(struct namecache *ncp, int nonblock)
2477 {
2478 	struct namecache *par;
2479 	struct vnode *dropvp;
2480 	struct nchash_head *nchpp;
2481 	int refs;
2482 
2483 	/*
2484 	 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED.
2485 	 */
2486 	_cache_setunresolved(ncp);
2487 
2488 	/*
2489 	 * Try to scrap the entry and possibly tail-recurse on its parent.
2490 	 * We only scrap unref'd (other then our ref) unresolved entries,
2491 	 * we do not scrap 'live' entries.
2492 	 *
2493 	 * Note that once the spinlocks are acquired if nc_refs == 1 no
2494 	 * other references are possible.  If it isn't, however, we have
2495 	 * to decrement but also be sure to avoid a 1->0 transition.
2496 	 */
2497 	KKASSERT(ncp->nc_flag & NCF_UNRESOLVED);
2498 	KKASSERT(ncp->nc_refs > 0);
2499 
2500 	/*
2501 	 * Acquire locks.  Note that the parent can't go away while we hold
2502 	 * a child locked.
2503 	 */
2504 	nchpp = NULL;
2505 	if ((par = ncp->nc_parent) != NULL) {
2506 		if (nonblock) {
2507 			for (;;) {
2508 				if (_cache_lock_nonblock(par) == 0)
2509 					break;
2510 				refs = ncp->nc_refs;
2511 				ncp->nc_flag |= NCF_DEFEREDZAP;
2512 				++numdefered;	/* MP race ok */
2513 				if (atomic_cmpset_int(&ncp->nc_refs,
2514 						      refs, refs - 1)) {
2515 					_cache_unlock(ncp);
2516 					return(NULL);
2517 				}
2518 				cpu_pause();
2519 			}
2520 			_cache_hold(par);
2521 		} else {
2522 			_cache_hold(par);
2523 			_cache_lock(par);
2524 		}
2525 		nchpp = ncp->nc_head;
2526 		spin_lock(&nchpp->spin);
2527 	}
2528 
2529 	/*
2530 	 * At this point if we find refs == 1 it should not be possible for
2531 	 * anyone else to have access to the ncp.  We are holding the only
2532 	 * possible access point left (nchpp) spin-locked.
2533 	 *
2534 	 * If someone other then us has a ref or we have children
2535 	 * we cannot zap the entry.  The 1->0 transition and any
2536 	 * further list operation is protected by the spinlocks
2537 	 * we have acquired but other transitions are not.
2538 	 */
2539 	for (;;) {
2540 		refs = ncp->nc_refs;
2541 		cpu_ccfence();
2542 		if (refs == 1 && TAILQ_EMPTY(&ncp->nc_list))
2543 			break;
2544 		if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1)) {
2545 			if (par) {
2546 				spin_unlock(&nchpp->spin);
2547 				_cache_put(par);
2548 			}
2549 			_cache_unlock(ncp);
2550 			return(NULL);
2551 		}
2552 		cpu_pause();
2553 	}
2554 
2555 	/*
2556 	 * We are the only ref and with the spinlocks held no further
2557 	 * refs can be acquired by others.
2558 	 *
2559 	 * Remove us from the hash list and parent list.  We have to
2560 	 * drop a ref on the parent's vp if the parent's list becomes
2561 	 * empty.
2562 	 */
2563 	dropvp = NULL;
2564 	if (par) {
2565 		KKASSERT(nchpp == ncp->nc_head);
2566 		LIST_REMOVE(ncp, nc_hash);
2567 		TAILQ_REMOVE(&par->nc_list, ncp, nc_entry);
2568 		if (par->nc_vp && TAILQ_EMPTY(&par->nc_list))
2569 			dropvp = par->nc_vp;
2570 		ncp->nc_head = NULL;
2571 		ncp->nc_parent = NULL;
2572 		spin_unlock(&nchpp->spin);
2573 		_cache_unlock(par);
2574 	} else {
2575 		KKASSERT(ncp->nc_head == NULL);
2576 	}
2577 
2578 	/*
2579 	 * ncp should not have picked up any refs.  Physically
2580 	 * destroy the ncp.
2581 	 */
2582 	if (ncp->nc_refs != 1) {
2583 		int save_refs = ncp->nc_refs;
2584 		cpu_ccfence();
2585 		panic("cache_zap: %p bad refs %d (%d)\n",
2586 			ncp, save_refs, atomic_fetchadd_int(&ncp->nc_refs, 0));
2587 	}
2588 	KKASSERT(ncp->nc_refs == 1);
2589 	/* _cache_unlock(ncp) not required */
2590 	ncp->nc_refs = -1;	/* safety */
2591 	if (ncp->nc_name)
2592 		kfree(ncp->nc_name, M_VFSCACHE);
2593 	kfree(ncp, M_VFSCACHE);
2594 
2595 	/*
2596 	 * Delayed drop (we had to release our spinlocks)
2597 	 *
2598 	 * The refed parent (if not  NULL) must be dropped.  The
2599 	 * caller is responsible for looping.
2600 	 */
2601 	if (dropvp)
2602 		vdrop(dropvp);
2603 	return(par);
2604 }
2605 
2606 /*
2607  * Clean up dangling negative cache and defered-drop entries in the
2608  * namecache.
2609  *
2610  * This routine is called in the critical path and also called from
2611  * vnlru().  When called from vnlru we use a lower limit to try to
2612  * deal with the negative cache before the critical path has to start
2613  * dealing with it.
2614  */
2615 typedef enum { CHI_LOW, CHI_HIGH } cache_hs_t;
2616 
2617 static cache_hs_t neg_cache_hysteresis_state[2] = { CHI_LOW, CHI_LOW };
2618 static cache_hs_t pos_cache_hysteresis_state[2] = { CHI_LOW, CHI_LOW };
2619 
2620 void
2621 cache_hysteresis(int critpath)
2622 {
2623 	int poslimit;
2624 	int neglimit = desiredvnodes / ncnegfactor;
2625 	int xnumcache = numcache;
2626 
2627 	if (critpath == 0)
2628 		neglimit = neglimit * 8 / 10;
2629 
2630 	/*
2631 	 * Don't cache too many negative hits.  We use hysteresis to reduce
2632 	 * the impact on the critical path.
2633 	 */
2634 	switch(neg_cache_hysteresis_state[critpath]) {
2635 	case CHI_LOW:
2636 		if (numneg > MINNEG && numneg > neglimit) {
2637 			if (critpath)
2638 				_cache_cleanneg(ncnegflush);
2639 			else
2640 				_cache_cleanneg(ncnegflush +
2641 						numneg - neglimit);
2642 			neg_cache_hysteresis_state[critpath] = CHI_HIGH;
2643 		}
2644 		break;
2645 	case CHI_HIGH:
2646 		if (numneg > MINNEG * 9 / 10 &&
2647 		    numneg * 9 / 10 > neglimit
2648 		) {
2649 			if (critpath)
2650 				_cache_cleanneg(ncnegflush);
2651 			else
2652 				_cache_cleanneg(ncnegflush +
2653 						numneg * 9 / 10 - neglimit);
2654 		} else {
2655 			neg_cache_hysteresis_state[critpath] = CHI_LOW;
2656 		}
2657 		break;
2658 	}
2659 
2660 	/*
2661 	 * Don't cache too many positive hits.  We use hysteresis to reduce
2662 	 * the impact on the critical path.
2663 	 *
2664 	 * Excessive positive hits can accumulate due to large numbers of
2665 	 * hardlinks (the vnode cache will not prevent hl ncps from growing
2666 	 * into infinity).
2667 	 */
2668 	if ((poslimit = ncposlimit) == 0)
2669 		poslimit = desiredvnodes * 2;
2670 	if (critpath == 0)
2671 		poslimit = poslimit * 8 / 10;
2672 
2673 	switch(pos_cache_hysteresis_state[critpath]) {
2674 	case CHI_LOW:
2675 		if (xnumcache > poslimit && xnumcache > MINPOS) {
2676 			if (critpath)
2677 				_cache_cleanpos(ncposflush);
2678 			else
2679 				_cache_cleanpos(ncposflush +
2680 						xnumcache - poslimit);
2681 			pos_cache_hysteresis_state[critpath] = CHI_HIGH;
2682 		}
2683 		break;
2684 	case CHI_HIGH:
2685 		if (xnumcache > poslimit * 5 / 6 && xnumcache > MINPOS) {
2686 			if (critpath)
2687 				_cache_cleanpos(ncposflush);
2688 			else
2689 				_cache_cleanpos(ncposflush +
2690 						xnumcache - poslimit * 5 / 6);
2691 		} else {
2692 			pos_cache_hysteresis_state[critpath] = CHI_LOW;
2693 		}
2694 		break;
2695 	}
2696 
2697 	/*
2698 	 * Clean out dangling defered-zap ncps which could not
2699 	 * be cleanly dropped if too many build up.  Note
2700 	 * that numdefered is not an exact number as such ncps
2701 	 * can be reused and the counter is not handled in a MP
2702 	 * safe manner by design.
2703 	 */
2704 	if (numdefered > neglimit) {
2705 		_cache_cleandefered();
2706 	}
2707 }
2708 
2709 /*
2710  * NEW NAMECACHE LOOKUP API
2711  *
2712  * Lookup an entry in the namecache.  The passed par_nch must be referenced
2713  * and unlocked.  A referenced and locked nchandle with a non-NULL nch.ncp
2714  * is ALWAYS returned, eve if the supplied component is illegal.
2715  *
2716  * The resulting namecache entry should be returned to the system with
2717  * cache_put() or cache_unlock() + cache_drop().
2718  *
2719  * namecache locks are recursive but care must be taken to avoid lock order
2720  * reversals (hence why the passed par_nch must be unlocked).  Locking
2721  * rules are to order for parent traversals, not for child traversals.
2722  *
2723  * Nobody else will be able to manipulate the associated namespace (e.g.
2724  * create, delete, rename, rename-target) until the caller unlocks the
2725  * entry.
2726  *
2727  * The returned entry will be in one of three states:  positive hit (non-null
2728  * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set).
2729  * Unresolved entries must be resolved through the filesystem to associate the
2730  * vnode and/or determine whether a positive or negative hit has occured.
2731  *
2732  * It is not necessary to lock a directory in order to lock namespace under
2733  * that directory.  In fact, it is explicitly not allowed to do that.  A
2734  * directory is typically only locked when being created, renamed, or
2735  * destroyed.
2736  *
2737  * The directory (par) may be unresolved, in which case any returned child
2738  * will likely also be marked unresolved.  Likely but not guarenteed.  Since
2739  * the filesystem lookup requires a resolved directory vnode the caller is
2740  * responsible for resolving the namecache chain top-down.  This API
2741  * specifically allows whole chains to be created in an unresolved state.
2742  */
2743 struct nchandle
2744 cache_nlookup(struct nchandle *par_nch, struct nlcomponent *nlc)
2745 {
2746 	struct nchandle nch;
2747 	struct namecache *ncp;
2748 	struct namecache *new_ncp;
2749 	struct nchash_head *nchpp;
2750 	struct mount *mp;
2751 	u_int32_t hash;
2752 	globaldata_t gd;
2753 	int par_locked;
2754 
2755 	numcalls++;
2756 	gd = mycpu;
2757 	mp = par_nch->mount;
2758 	par_locked = 0;
2759 
2760 	/*
2761 	 * This is a good time to call it, no ncp's are locked by
2762 	 * the caller or us.
2763 	 */
2764 	cache_hysteresis(1);
2765 
2766 	/*
2767 	 * Try to locate an existing entry
2768 	 */
2769 	hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
2770 	hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
2771 	new_ncp = NULL;
2772 	nchpp = NCHHASH(hash);
2773 restart:
2774 	if (new_ncp)
2775 		spin_lock(&nchpp->spin);
2776 	else
2777 		spin_lock_shared(&nchpp->spin);
2778 
2779 	LIST_FOREACH(ncp, &nchpp->list, nc_hash) {
2780 		numchecks++;
2781 
2782 		/*
2783 		 * Break out if we find a matching entry.  Note that
2784 		 * UNRESOLVED entries may match, but DESTROYED entries
2785 		 * do not.
2786 		 */
2787 		if (ncp->nc_parent == par_nch->ncp &&
2788 		    ncp->nc_nlen == nlc->nlc_namelen &&
2789 		    bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
2790 		    (ncp->nc_flag & NCF_DESTROYED) == 0
2791 		) {
2792 			_cache_hold(ncp);
2793 			if (new_ncp)
2794 				spin_unlock(&nchpp->spin);
2795 			else
2796 				spin_unlock_shared(&nchpp->spin);
2797 			if (par_locked) {
2798 				_cache_unlock(par_nch->ncp);
2799 				par_locked = 0;
2800 			}
2801 			if (_cache_lock_special(ncp) == 0) {
2802 				/*
2803 				 * Successfully locked but we must re-test
2804 				 * conditions that might have changed since
2805 				 * we did not have the lock before.
2806 				 */
2807 				if (ncp->nc_parent != par_nch->ncp ||
2808 				    ncp->nc_nlen != nlc->nlc_namelen ||
2809 				    bcmp(ncp->nc_name, nlc->nlc_nameptr,
2810 					 ncp->nc_nlen) ||
2811 				    (ncp->nc_flag & NCF_DESTROYED)) {
2812 					_cache_put(ncp);
2813 					goto restart;
2814 				}
2815 				_cache_auto_unresolve(mp, ncp);
2816 				if (new_ncp)
2817 					_cache_free(new_ncp);
2818 				goto found;
2819 			}
2820 			_cache_get(ncp);	/* cycle the lock to block */
2821 			_cache_put(ncp);
2822 			_cache_drop(ncp);
2823 			goto restart;
2824 		}
2825 	}
2826 
2827 	/*
2828 	 * We failed to locate an entry, create a new entry and add it to
2829 	 * the cache.  The parent ncp must also be locked so we
2830 	 * can link into it.
2831 	 *
2832 	 * We have to relookup after possibly blocking in kmalloc or
2833 	 * when locking par_nch.
2834 	 *
2835 	 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special
2836 	 *	 mount case, in which case nc_name will be NULL.
2837 	 */
2838 	if (new_ncp == NULL) {
2839 		spin_unlock_shared(&nchpp->spin);
2840 		new_ncp = cache_alloc(nlc->nlc_namelen);
2841 		if (nlc->nlc_namelen) {
2842 			bcopy(nlc->nlc_nameptr, new_ncp->nc_name,
2843 			      nlc->nlc_namelen);
2844 			new_ncp->nc_name[nlc->nlc_namelen] = 0;
2845 		}
2846 		goto restart;
2847 	}
2848 
2849 	/*
2850 	 * NOTE! The spinlock is held exclusively here because new_ncp
2851 	 *	 is non-NULL.
2852 	 */
2853 	if (par_locked == 0) {
2854 		spin_unlock(&nchpp->spin);
2855 		_cache_lock(par_nch->ncp);
2856 		par_locked = 1;
2857 		goto restart;
2858 	}
2859 
2860 	/*
2861 	 * WARNING!  We still hold the spinlock.  We have to set the hash
2862 	 *	     table entry atomically.
2863 	 */
2864 	ncp = new_ncp;
2865 	_cache_link_parent(ncp, par_nch->ncp, nchpp);
2866 	spin_unlock(&nchpp->spin);
2867 	_cache_unlock(par_nch->ncp);
2868 	/* par_locked = 0 - not used */
2869 found:
2870 	/*
2871 	 * stats and namecache size management
2872 	 */
2873 	if (ncp->nc_flag & NCF_UNRESOLVED)
2874 		++gd->gd_nchstats->ncs_miss;
2875 	else if (ncp->nc_vp)
2876 		++gd->gd_nchstats->ncs_goodhits;
2877 	else
2878 		++gd->gd_nchstats->ncs_neghits;
2879 	nch.mount = mp;
2880 	nch.ncp = ncp;
2881 	atomic_add_int(&nch.mount->mnt_refs, 1);
2882 	return(nch);
2883 }
2884 
2885 /*
2886  * Attempt to lookup a namecache entry and return with a shared namecache
2887  * lock.
2888  */
2889 int
2890 cache_nlookup_maybe_shared(struct nchandle *par_nch, struct nlcomponent *nlc,
2891 			   int excl, struct nchandle *res_nch)
2892 {
2893 	struct namecache *ncp;
2894 	struct nchash_head *nchpp;
2895 	struct mount *mp;
2896 	u_int32_t hash;
2897 	globaldata_t gd;
2898 
2899 	/*
2900 	 * If exclusive requested or shared namecache locks are disabled,
2901 	 * return failure.
2902 	 */
2903 	if (ncp_shared_lock_disable || excl)
2904 		return(EWOULDBLOCK);
2905 
2906 	numcalls++;
2907 	gd = mycpu;
2908 	mp = par_nch->mount;
2909 
2910 	/*
2911 	 * This is a good time to call it, no ncp's are locked by
2912 	 * the caller or us.
2913 	 */
2914 	cache_hysteresis(1);
2915 
2916 	/*
2917 	 * Try to locate an existing entry
2918 	 */
2919 	hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
2920 	hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
2921 	nchpp = NCHHASH(hash);
2922 
2923 	spin_lock_shared(&nchpp->spin);
2924 
2925 	LIST_FOREACH(ncp, &nchpp->list, nc_hash) {
2926 		numchecks++;
2927 
2928 		/*
2929 		 * Break out if we find a matching entry.  Note that
2930 		 * UNRESOLVED entries may match, but DESTROYED entries
2931 		 * do not.
2932 		 */
2933 		if (ncp->nc_parent == par_nch->ncp &&
2934 		    ncp->nc_nlen == nlc->nlc_namelen &&
2935 		    bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
2936 		    (ncp->nc_flag & NCF_DESTROYED) == 0
2937 		) {
2938 			_cache_hold(ncp);
2939 			spin_unlock_shared(&nchpp->spin);
2940 			if (_cache_lock_shared_special(ncp) == 0) {
2941 				if (ncp->nc_parent == par_nch->ncp &&
2942 				    ncp->nc_nlen == nlc->nlc_namelen &&
2943 				    bcmp(ncp->nc_name, nlc->nlc_nameptr,
2944 					 ncp->nc_nlen) == 0 &&
2945 				    (ncp->nc_flag & NCF_DESTROYED) == 0 &&
2946 				    (ncp->nc_flag & NCF_UNRESOLVED) == 0 &&
2947 				    _cache_auto_unresolve_test(mp, ncp) == 0) {
2948 					goto found;
2949 				}
2950 				_cache_unlock(ncp);
2951 			}
2952 			_cache_drop(ncp);
2953 			spin_lock_shared(&nchpp->spin);
2954 			break;
2955 		}
2956 	}
2957 
2958 	/*
2959 	 * Failure
2960 	 */
2961 	spin_unlock_shared(&nchpp->spin);
2962 	return(EWOULDBLOCK);
2963 
2964 	/*
2965 	 * Success
2966 	 *
2967 	 * Note that nc_error might be non-zero (e.g ENOENT).
2968 	 */
2969 found:
2970 	res_nch->mount = mp;
2971 	res_nch->ncp = ncp;
2972 	++gd->gd_nchstats->ncs_goodhits;
2973 	atomic_add_int(&res_nch->mount->mnt_refs, 1);
2974 
2975 	KKASSERT(ncp->nc_error != EWOULDBLOCK);
2976 	return(ncp->nc_error);
2977 }
2978 
2979 /*
2980  * This is a non-blocking verison of cache_nlookup() used by
2981  * nfs_readdirplusrpc_uio().  It can fail for any reason and
2982  * will return nch.ncp == NULL in that case.
2983  */
2984 struct nchandle
2985 cache_nlookup_nonblock(struct nchandle *par_nch, struct nlcomponent *nlc)
2986 {
2987 	struct nchandle nch;
2988 	struct namecache *ncp;
2989 	struct namecache *new_ncp;
2990 	struct nchash_head *nchpp;
2991 	struct mount *mp;
2992 	u_int32_t hash;
2993 	globaldata_t gd;
2994 	int par_locked;
2995 
2996 	numcalls++;
2997 	gd = mycpu;
2998 	mp = par_nch->mount;
2999 	par_locked = 0;
3000 
3001 	/*
3002 	 * Try to locate an existing entry
3003 	 */
3004 	hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT);
3005 	hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash);
3006 	new_ncp = NULL;
3007 	nchpp = NCHHASH(hash);
3008 restart:
3009 	spin_lock(&nchpp->spin);
3010 	LIST_FOREACH(ncp, &nchpp->list, nc_hash) {
3011 		numchecks++;
3012 
3013 		/*
3014 		 * Break out if we find a matching entry.  Note that
3015 		 * UNRESOLVED entries may match, but DESTROYED entries
3016 		 * do not.
3017 		 */
3018 		if (ncp->nc_parent == par_nch->ncp &&
3019 		    ncp->nc_nlen == nlc->nlc_namelen &&
3020 		    bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 &&
3021 		    (ncp->nc_flag & NCF_DESTROYED) == 0
3022 		) {
3023 			_cache_hold(ncp);
3024 			spin_unlock(&nchpp->spin);
3025 			if (par_locked) {
3026 				_cache_unlock(par_nch->ncp);
3027 				par_locked = 0;
3028 			}
3029 			if (_cache_lock_special(ncp) == 0) {
3030 				if (ncp->nc_parent != par_nch->ncp ||
3031 				    ncp->nc_nlen != nlc->nlc_namelen ||
3032 				    bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) ||
3033 				    (ncp->nc_flag & NCF_DESTROYED)) {
3034 					kprintf("cache_lookup_nonblock: "
3035 						"ncp-race %p %*.*s\n",
3036 						ncp,
3037 						nlc->nlc_namelen,
3038 						nlc->nlc_namelen,
3039 						nlc->nlc_nameptr);
3040 					_cache_unlock(ncp);
3041 					_cache_drop(ncp);
3042 					goto failed;
3043 				}
3044 				_cache_auto_unresolve(mp, ncp);
3045 				if (new_ncp) {
3046 					_cache_free(new_ncp);
3047 					new_ncp = NULL;
3048 				}
3049 				goto found;
3050 			}
3051 			_cache_drop(ncp);
3052 			goto failed;
3053 		}
3054 	}
3055 
3056 	/*
3057 	 * We failed to locate an entry, create a new entry and add it to
3058 	 * the cache.  The parent ncp must also be locked so we
3059 	 * can link into it.
3060 	 *
3061 	 * We have to relookup after possibly blocking in kmalloc or
3062 	 * when locking par_nch.
3063 	 *
3064 	 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special
3065 	 *	 mount case, in which case nc_name will be NULL.
3066 	 */
3067 	if (new_ncp == NULL) {
3068 		spin_unlock(&nchpp->spin);
3069 		new_ncp = cache_alloc(nlc->nlc_namelen);
3070 		if (nlc->nlc_namelen) {
3071 			bcopy(nlc->nlc_nameptr, new_ncp->nc_name,
3072 			      nlc->nlc_namelen);
3073 			new_ncp->nc_name[nlc->nlc_namelen] = 0;
3074 		}
3075 		goto restart;
3076 	}
3077 	if (par_locked == 0) {
3078 		spin_unlock(&nchpp->spin);
3079 		if (_cache_lock_nonblock(par_nch->ncp) == 0) {
3080 			par_locked = 1;
3081 			goto restart;
3082 		}
3083 		goto failed;
3084 	}
3085 
3086 	/*
3087 	 * WARNING!  We still hold the spinlock.  We have to set the hash
3088 	 *	     table entry atomically.
3089 	 */
3090 	ncp = new_ncp;
3091 	_cache_link_parent(ncp, par_nch->ncp, nchpp);
3092 	spin_unlock(&nchpp->spin);
3093 	_cache_unlock(par_nch->ncp);
3094 	/* par_locked = 0 - not used */
3095 found:
3096 	/*
3097 	 * stats and namecache size management
3098 	 */
3099 	if (ncp->nc_flag & NCF_UNRESOLVED)
3100 		++gd->gd_nchstats->ncs_miss;
3101 	else if (ncp->nc_vp)
3102 		++gd->gd_nchstats->ncs_goodhits;
3103 	else
3104 		++gd->gd_nchstats->ncs_neghits;
3105 	nch.mount = mp;
3106 	nch.ncp = ncp;
3107 	atomic_add_int(&nch.mount->mnt_refs, 1);
3108 	return(nch);
3109 failed:
3110 	if (new_ncp) {
3111 		_cache_free(new_ncp);
3112 		new_ncp = NULL;
3113 	}
3114 	nch.mount = NULL;
3115 	nch.ncp = NULL;
3116 	return(nch);
3117 }
3118 
3119 /*
3120  * The namecache entry is marked as being used as a mount point.
3121  * Locate the mount if it is visible to the caller.  The DragonFly
3122  * mount system allows arbitrary loops in the topology and disentangles
3123  * those loops by matching against (mp, ncp) rather than just (ncp).
3124  * This means any given ncp can dive any number of mounts, depending
3125  * on the relative mount (e.g. nullfs) the caller is at in the topology.
3126  *
3127  * We use a very simple frontend cache to reduce SMP conflicts,
3128  * which we have to do because the mountlist scan needs an exclusive
3129  * lock around its ripout info list.  Not to mention that there might
3130  * be a lot of mounts.
3131  */
3132 struct findmount_info {
3133 	struct mount *result;
3134 	struct mount *nch_mount;
3135 	struct namecache *nch_ncp;
3136 };
3137 
3138 static
3139 struct ncmount_cache *
3140 ncmount_cache_lookup(struct mount *mp, struct namecache *ncp)
3141 {
3142 	int hash;
3143 
3144 	hash = ((int)(intptr_t)mp / sizeof(*mp)) ^
3145 	       ((int)(intptr_t)ncp / sizeof(*ncp));
3146 	hash = (hash & 0x7FFFFFFF) % NCMOUNT_NUMCACHE;
3147 	return (&ncmount_cache[hash]);
3148 }
3149 
3150 static
3151 int
3152 cache_findmount_callback(struct mount *mp, void *data)
3153 {
3154 	struct findmount_info *info = data;
3155 
3156 	/*
3157 	 * Check the mount's mounted-on point against the passed nch.
3158 	 */
3159 	if (mp->mnt_ncmounton.mount == info->nch_mount &&
3160 	    mp->mnt_ncmounton.ncp == info->nch_ncp
3161 	) {
3162 	    info->result = mp;
3163 	    atomic_add_int(&mp->mnt_refs, 1);
3164 	    return(-1);
3165 	}
3166 	return(0);
3167 }
3168 
3169 struct mount *
3170 cache_findmount(struct nchandle *nch)
3171 {
3172 	struct findmount_info info;
3173 	struct ncmount_cache *ncc;
3174 	struct mount *mp;
3175 
3176 	/*
3177 	 * Fast
3178 	 */
3179 	if (ncmount_cache_enable == 0) {
3180 		ncc = NULL;
3181 		goto skip;
3182 	}
3183 	ncc = ncmount_cache_lookup(nch->mount, nch->ncp);
3184 	if (ncc->ncp == nch->ncp) {
3185 		spin_lock_shared(&ncc->spin);
3186 		if (ncc->isneg == 0 &&
3187 		    ncc->ncp == nch->ncp && (mp = ncc->mp) != NULL) {
3188 			if (mp->mnt_ncmounton.mount == nch->mount &&
3189 			    mp->mnt_ncmounton.ncp == nch->ncp) {
3190 				/*
3191 				 * Cache hit (positive)
3192 				 */
3193 				atomic_add_int(&mp->mnt_refs, 1);
3194 				spin_unlock_shared(&ncc->spin);
3195 				++ncmount_cache_hit;
3196 				return(mp);
3197 			}
3198 			/* else cache miss */
3199 		}
3200 		if (ncc->isneg &&
3201 		    ncc->ncp == nch->ncp && ncc->mp == nch->mount) {
3202 			/*
3203 			 * Cache hit (negative)
3204 			 */
3205 			spin_unlock_shared(&ncc->spin);
3206 			++ncmount_cache_hit;
3207 			return(NULL);
3208 		}
3209 		spin_unlock_shared(&ncc->spin);
3210 	}
3211 skip:
3212 
3213 	/*
3214 	 * Slow
3215 	 */
3216 	info.result = NULL;
3217 	info.nch_mount = nch->mount;
3218 	info.nch_ncp = nch->ncp;
3219 	mountlist_scan(cache_findmount_callback, &info,
3220 			       MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
3221 
3222 	/*
3223 	 * Cache the result.
3224 	 *
3225 	 * Negative lookups: We cache the originating {ncp,mp}. (mp) is
3226 	 *		     only used for pointer comparisons and is not
3227 	 *		     referenced (otherwise there would be dangling
3228 	 *		     refs).
3229 	 *
3230 	 * Positive lookups: We cache the originating {ncp} and the target
3231 	 *		     (mp).  (mp) is referenced.
3232 	 *
3233 	 * Indeterminant:    If the match is undergoing an unmount we do
3234 	 *		     not cache it to avoid racing cache_unmounting(),
3235 	 *		     but still return the match.
3236 	 */
3237 	if (ncc) {
3238 		spin_lock(&ncc->spin);
3239 		if (info.result == NULL) {
3240 			if (ncc->isneg == 0 && ncc->mp)
3241 				atomic_add_int(&ncc->mp->mnt_refs, -1);
3242 			ncc->ncp = nch->ncp;
3243 			ncc->mp = nch->mount;
3244 			ncc->isneg = 1;
3245 			spin_unlock(&ncc->spin);
3246 			++ncmount_cache_overwrite;
3247 		} else if ((info.result->mnt_kern_flag & MNTK_UNMOUNT) == 0) {
3248 			if (ncc->isneg == 0 && ncc->mp)
3249 				atomic_add_int(&ncc->mp->mnt_refs, -1);
3250 			atomic_add_int(&info.result->mnt_refs, 1);
3251 			ncc->ncp = nch->ncp;
3252 			ncc->mp = info.result;
3253 			ncc->isneg = 0;
3254 			spin_unlock(&ncc->spin);
3255 			++ncmount_cache_overwrite;
3256 		} else {
3257 			spin_unlock(&ncc->spin);
3258 		}
3259 		++ncmount_cache_miss;
3260 	}
3261 	return(info.result);
3262 }
3263 
3264 void
3265 cache_dropmount(struct mount *mp)
3266 {
3267 	atomic_add_int(&mp->mnt_refs, -1);
3268 }
3269 
3270 void
3271 cache_ismounting(struct mount *mp)
3272 {
3273 	struct nchandle *nch = &mp->mnt_ncmounton;
3274 	struct ncmount_cache *ncc;
3275 
3276 	ncc = ncmount_cache_lookup(nch->mount, nch->ncp);
3277 	if (ncc->isneg &&
3278 	    ncc->ncp == nch->ncp && ncc->mp == nch->mount) {
3279 		spin_lock(&ncc->spin);
3280 		if (ncc->isneg &&
3281 		    ncc->ncp == nch->ncp && ncc->mp == nch->mount) {
3282 			ncc->ncp = NULL;
3283 			ncc->mp = NULL;
3284 		}
3285 		spin_unlock(&ncc->spin);
3286 	}
3287 }
3288 
3289 void
3290 cache_unmounting(struct mount *mp)
3291 {
3292 	struct nchandle *nch = &mp->mnt_ncmounton;
3293 	struct ncmount_cache *ncc;
3294 
3295 	ncc = ncmount_cache_lookup(nch->mount, nch->ncp);
3296 	if (ncc->isneg == 0 &&
3297 	    ncc->ncp == nch->ncp && ncc->mp == mp) {
3298 		spin_lock(&ncc->spin);
3299 		if (ncc->isneg == 0 &&
3300 		    ncc->ncp == nch->ncp && ncc->mp == mp) {
3301 			atomic_add_int(&mp->mnt_refs, -1);
3302 			ncc->ncp = NULL;
3303 			ncc->mp = NULL;
3304 		}
3305 		spin_unlock(&ncc->spin);
3306 	}
3307 }
3308 
3309 /*
3310  * Resolve an unresolved namecache entry, generally by looking it up.
3311  * The passed ncp must be locked and refd.
3312  *
3313  * Theoretically since a vnode cannot be recycled while held, and since
3314  * the nc_parent chain holds its vnode as long as children exist, the
3315  * direct parent of the cache entry we are trying to resolve should
3316  * have a valid vnode.  If not then generate an error that we can
3317  * determine is related to a resolver bug.
3318  *
3319  * However, if a vnode was in the middle of a recyclement when the NCP
3320  * got locked, ncp->nc_vp might point to a vnode that is about to become
3321  * invalid.  cache_resolve() handles this case by unresolving the entry
3322  * and then re-resolving it.
3323  *
3324  * Note that successful resolution does not necessarily return an error
3325  * code of 0.  If the ncp resolves to a negative cache hit then ENOENT
3326  * will be returned.
3327  */
3328 int
3329 cache_resolve(struct nchandle *nch, struct ucred *cred)
3330 {
3331 	struct namecache *par_tmp;
3332 	struct namecache *par;
3333 	struct namecache *ncp;
3334 	struct nchandle nctmp;
3335 	struct mount *mp;
3336 	struct vnode *dvp;
3337 	int error;
3338 
3339 	ncp = nch->ncp;
3340 	mp = nch->mount;
3341 	KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE);
3342 restart:
3343 	/*
3344 	 * If the ncp is already resolved we have nothing to do.  However,
3345 	 * we do want to guarentee that a usable vnode is returned when
3346 	 * a vnode is present, so make sure it hasn't been reclaimed.
3347 	 */
3348 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
3349 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
3350 			_cache_setunresolved(ncp);
3351 		if ((ncp->nc_flag & NCF_UNRESOLVED) == 0)
3352 			return (ncp->nc_error);
3353 	}
3354 
3355 	/*
3356 	 * If the ncp was destroyed it will never resolve again.  This
3357 	 * can basically only happen when someone is chdir'd into an
3358 	 * empty directory which is then rmdir'd.  We want to catch this
3359 	 * here and not dive the VFS because the VFS might actually
3360 	 * have a way to re-resolve the disconnected ncp, which will
3361 	 * result in inconsistencies in the cdir/nch for proc->p_fd.
3362 	 */
3363 	if (ncp->nc_flag & NCF_DESTROYED) {
3364 		kprintf("Warning: cache_resolve: ncp '%s' was unlinked\n",
3365 			ncp->nc_name);
3366 		return(EINVAL);
3367 	}
3368 
3369 	/*
3370 	 * Mount points need special handling because the parent does not
3371 	 * belong to the same filesystem as the ncp.
3372 	 */
3373 	if (ncp == mp->mnt_ncmountpt.ncp)
3374 		return (cache_resolve_mp(mp));
3375 
3376 	/*
3377 	 * We expect an unbroken chain of ncps to at least the mount point,
3378 	 * and even all the way to root (but this code doesn't have to go
3379 	 * past the mount point).
3380 	 */
3381 	if (ncp->nc_parent == NULL) {
3382 		kprintf("EXDEV case 1 %p %*.*s\n", ncp,
3383 			ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
3384 		ncp->nc_error = EXDEV;
3385 		return(ncp->nc_error);
3386 	}
3387 
3388 	/*
3389 	 * The vp's of the parent directories in the chain are held via vhold()
3390 	 * due to the existance of the child, and should not disappear.
3391 	 * However, there are cases where they can disappear:
3392 	 *
3393 	 *	- due to filesystem I/O errors.
3394 	 *	- due to NFS being stupid about tracking the namespace and
3395 	 *	  destroys the namespace for entire directories quite often.
3396 	 *	- due to forced unmounts.
3397 	 *	- due to an rmdir (parent will be marked DESTROYED)
3398 	 *
3399 	 * When this occurs we have to track the chain backwards and resolve
3400 	 * it, looping until the resolver catches up to the current node.  We
3401 	 * could recurse here but we might run ourselves out of kernel stack
3402 	 * so we do it in a more painful manner.  This situation really should
3403 	 * not occur all that often, or if it does not have to go back too
3404 	 * many nodes to resolve the ncp.
3405 	 */
3406 	while ((dvp = cache_dvpref(ncp)) == NULL) {
3407 		/*
3408 		 * This case can occur if a process is CD'd into a
3409 		 * directory which is then rmdir'd.  If the parent is marked
3410 		 * destroyed there is no point trying to resolve it.
3411 		 */
3412 		if (ncp->nc_parent->nc_flag & NCF_DESTROYED)
3413 			return(ENOENT);
3414 		par = ncp->nc_parent;
3415 		_cache_hold(par);
3416 		_cache_lock(par);
3417 		while ((par_tmp = par->nc_parent) != NULL &&
3418 		       par_tmp->nc_vp == NULL) {
3419 			_cache_hold(par_tmp);
3420 			_cache_lock(par_tmp);
3421 			_cache_put(par);
3422 			par = par_tmp;
3423 		}
3424 		if (par->nc_parent == NULL) {
3425 			kprintf("EXDEV case 2 %*.*s\n",
3426 				par->nc_nlen, par->nc_nlen, par->nc_name);
3427 			_cache_put(par);
3428 			return (EXDEV);
3429 		}
3430 		/*
3431 		 * The parent is not set in stone, ref and lock it to prevent
3432 		 * it from disappearing.  Also note that due to renames it
3433 		 * is possible for our ncp to move and for par to no longer
3434 		 * be one of its parents.  We resolve it anyway, the loop
3435 		 * will handle any moves.
3436 		 */
3437 		_cache_get(par);	/* additional hold/lock */
3438 		_cache_put(par);	/* from earlier hold/lock */
3439 		if (par == nch->mount->mnt_ncmountpt.ncp) {
3440 			cache_resolve_mp(nch->mount);
3441 		} else if ((dvp = cache_dvpref(par)) == NULL) {
3442 			kprintf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name);
3443 			_cache_put(par);
3444 			continue;
3445 		} else {
3446 			if (par->nc_flag & NCF_UNRESOLVED) {
3447 				nctmp.mount = mp;
3448 				nctmp.ncp = par;
3449 				par->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred);
3450 			}
3451 			vrele(dvp);
3452 		}
3453 		if ((error = par->nc_error) != 0) {
3454 			if (par->nc_error != EAGAIN) {
3455 				kprintf("EXDEV case 3 %*.*s error %d\n",
3456 				    par->nc_nlen, par->nc_nlen, par->nc_name,
3457 				    par->nc_error);
3458 				_cache_put(par);
3459 				return(error);
3460 			}
3461 			kprintf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n",
3462 				par, par->nc_nlen, par->nc_nlen, par->nc_name);
3463 		}
3464 		_cache_put(par);
3465 		/* loop */
3466 	}
3467 
3468 	/*
3469 	 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected
3470 	 * ncp's and reattach them.  If this occurs the original ncp is marked
3471 	 * EAGAIN to force a relookup.
3472 	 *
3473 	 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed
3474 	 * ncp must already be resolved.
3475 	 */
3476 	if (dvp) {
3477 		nctmp.mount = mp;
3478 		nctmp.ncp = ncp;
3479 		ncp->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred);
3480 		vrele(dvp);
3481 	} else {
3482 		ncp->nc_error = EPERM;
3483 	}
3484 	if (ncp->nc_error == EAGAIN) {
3485 		kprintf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n",
3486 			ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name);
3487 		goto restart;
3488 	}
3489 	return(ncp->nc_error);
3490 }
3491 
3492 /*
3493  * Resolve the ncp associated with a mount point.  Such ncp's almost always
3494  * remain resolved and this routine is rarely called.  NFS MPs tends to force
3495  * re-resolution more often due to its mac-truck-smash-the-namecache
3496  * method of tracking namespace changes.
3497  *
3498  * The semantics for this call is that the passed ncp must be locked on
3499  * entry and will be locked on return.  However, if we actually have to
3500  * resolve the mount point we temporarily unlock the entry in order to
3501  * avoid race-to-root deadlocks due to e.g. dead NFS mounts.  Because of
3502  * the unlock we have to recheck the flags after we relock.
3503  */
3504 static int
3505 cache_resolve_mp(struct mount *mp)
3506 {
3507 	struct namecache *ncp = mp->mnt_ncmountpt.ncp;
3508 	struct vnode *vp;
3509 	int error;
3510 
3511 	KKASSERT(mp != NULL);
3512 
3513 	/*
3514 	 * If the ncp is already resolved we have nothing to do.  However,
3515 	 * we do want to guarentee that a usable vnode is returned when
3516 	 * a vnode is present, so make sure it hasn't been reclaimed.
3517 	 */
3518 	if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) {
3519 		if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED))
3520 			_cache_setunresolved(ncp);
3521 	}
3522 
3523 	if (ncp->nc_flag & NCF_UNRESOLVED) {
3524 		_cache_unlock(ncp);
3525 		while (vfs_busy(mp, 0))
3526 			;
3527 		error = VFS_ROOT(mp, &vp);
3528 		_cache_lock(ncp);
3529 
3530 		/*
3531 		 * recheck the ncp state after relocking.
3532 		 */
3533 		if (ncp->nc_flag & NCF_UNRESOLVED) {
3534 			ncp->nc_error = error;
3535 			if (error == 0) {
3536 				_cache_setvp(mp, ncp, vp);
3537 				vput(vp);
3538 			} else {
3539 				kprintf("[diagnostic] cache_resolve_mp: failed"
3540 					" to resolve mount %p err=%d ncp=%p\n",
3541 					mp, error, ncp);
3542 				_cache_setvp(mp, ncp, NULL);
3543 			}
3544 		} else if (error == 0) {
3545 			vput(vp);
3546 		}
3547 		vfs_unbusy(mp);
3548 	}
3549 	return(ncp->nc_error);
3550 }
3551 
3552 /*
3553  * Clean out negative cache entries when too many have accumulated.
3554  */
3555 static void
3556 _cache_cleanneg(int count)
3557 {
3558 	struct namecache *ncp;
3559 
3560 	/*
3561 	 * Attempt to clean out the specified number of negative cache
3562 	 * entries.
3563 	 */
3564 	while (count) {
3565 		spin_lock(&ncspin);
3566 		ncp = TAILQ_FIRST(&ncneglist);
3567 		if (ncp == NULL) {
3568 			spin_unlock(&ncspin);
3569 			break;
3570 		}
3571 		TAILQ_REMOVE(&ncneglist, ncp, nc_vnode);
3572 		TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode);
3573 		_cache_hold(ncp);
3574 		spin_unlock(&ncspin);
3575 
3576 		/*
3577 		 * This can race, so we must re-check that the ncp
3578 		 * is on the ncneglist after successfully locking it.
3579 		 */
3580 		if (_cache_lock_special(ncp) == 0) {
3581 			if (ncp->nc_vp == NULL &&
3582 			    (ncp->nc_flag & NCF_UNRESOLVED) == 0) {
3583 				ncp = cache_zap(ncp, 1);
3584 				if (ncp)
3585 					_cache_drop(ncp);
3586 			} else {
3587 				kprintf("cache_cleanneg: race avoided\n");
3588 				_cache_unlock(ncp);
3589 			}
3590 		} else {
3591 			_cache_drop(ncp);
3592 		}
3593 		--count;
3594 	}
3595 }
3596 
3597 /*
3598  * Clean out positive cache entries when too many have accumulated.
3599  */
3600 static void
3601 _cache_cleanpos(int count)
3602 {
3603 	static volatile int rover;
3604 	struct nchash_head *nchpp;
3605 	struct namecache *ncp;
3606 	int rover_copy;
3607 
3608 	/*
3609 	 * Attempt to clean out the specified number of negative cache
3610 	 * entries.
3611 	 */
3612 	while (count) {
3613 		rover_copy = ++rover;	/* MPSAFEENOUGH */
3614 		cpu_ccfence();
3615 		nchpp = NCHHASH(rover_copy);
3616 
3617 		spin_lock_shared(&nchpp->spin);
3618 		ncp = LIST_FIRST(&nchpp->list);
3619 		while (ncp && (ncp->nc_flag & NCF_DESTROYED))
3620 			ncp = LIST_NEXT(ncp, nc_hash);
3621 		if (ncp)
3622 			_cache_hold(ncp);
3623 		spin_unlock_shared(&nchpp->spin);
3624 
3625 		if (ncp) {
3626 			if (_cache_lock_special(ncp) == 0) {
3627 				ncp = cache_zap(ncp, 1);
3628 				if (ncp)
3629 					_cache_drop(ncp);
3630 			} else {
3631 				_cache_drop(ncp);
3632 			}
3633 		}
3634 		--count;
3635 	}
3636 }
3637 
3638 /*
3639  * This is a kitchen sink function to clean out ncps which we
3640  * tried to zap from cache_drop() but failed because we were
3641  * unable to acquire the parent lock.
3642  *
3643  * Such entries can also be removed via cache_inval_vp(), such
3644  * as when unmounting.
3645  */
3646 static void
3647 _cache_cleandefered(void)
3648 {
3649 	struct nchash_head *nchpp;
3650 	struct namecache *ncp;
3651 	struct namecache dummy;
3652 	int i;
3653 
3654 	numdefered = 0;
3655 	bzero(&dummy, sizeof(dummy));
3656 	dummy.nc_flag = NCF_DESTROYED;
3657 	dummy.nc_refs = 1;
3658 
3659 	for (i = 0; i <= nchash; ++i) {
3660 		nchpp = &nchashtbl[i];
3661 
3662 		spin_lock(&nchpp->spin);
3663 		LIST_INSERT_HEAD(&nchpp->list, &dummy, nc_hash);
3664 		ncp = &dummy;
3665 		while ((ncp = LIST_NEXT(ncp, nc_hash)) != NULL) {
3666 			if ((ncp->nc_flag & NCF_DEFEREDZAP) == 0)
3667 				continue;
3668 			LIST_REMOVE(&dummy, nc_hash);
3669 			LIST_INSERT_AFTER(ncp, &dummy, nc_hash);
3670 			_cache_hold(ncp);
3671 			spin_unlock(&nchpp->spin);
3672 			if (_cache_lock_nonblock(ncp) == 0) {
3673 				ncp->nc_flag &= ~NCF_DEFEREDZAP;
3674 				_cache_unlock(ncp);
3675 			}
3676 			_cache_drop(ncp);
3677 			spin_lock(&nchpp->spin);
3678 			ncp = &dummy;
3679 		}
3680 		LIST_REMOVE(&dummy, nc_hash);
3681 		spin_unlock(&nchpp->spin);
3682 	}
3683 }
3684 
3685 /*
3686  * Name cache initialization, from vfsinit() when we are booting
3687  */
3688 void
3689 nchinit(void)
3690 {
3691 	int i;
3692 	globaldata_t gd;
3693 
3694 	/* initialise per-cpu namecache effectiveness statistics. */
3695 	for (i = 0; i < ncpus; ++i) {
3696 		gd = globaldata_find(i);
3697 		gd->gd_nchstats = &nchstats[i];
3698 	}
3699 	TAILQ_INIT(&ncneglist);
3700 	spin_init(&ncspin, "nchinit");
3701 	nchashtbl = hashinit_ext(desiredvnodes / 2,
3702 				 sizeof(struct nchash_head),
3703 				 M_VFSCACHE, &nchash);
3704 	for (i = 0; i <= (int)nchash; ++i) {
3705 		LIST_INIT(&nchashtbl[i].list);
3706 		spin_init(&nchashtbl[i].spin, "nchinit_hash");
3707 	}
3708 	for (i = 0; i < NCMOUNT_NUMCACHE; ++i)
3709 		spin_init(&ncmount_cache[i].spin, "nchinit_cache");
3710 	nclockwarn = 5 * hz;
3711 }
3712 
3713 /*
3714  * Called from start_init() to bootstrap the root filesystem.  Returns
3715  * a referenced, unlocked namecache record.
3716  */
3717 void
3718 cache_allocroot(struct nchandle *nch, struct mount *mp, struct vnode *vp)
3719 {
3720 	nch->ncp = cache_alloc(0);
3721 	nch->mount = mp;
3722 	atomic_add_int(&mp->mnt_refs, 1);
3723 	if (vp)
3724 		_cache_setvp(nch->mount, nch->ncp, vp);
3725 }
3726 
3727 /*
3728  * vfs_cache_setroot()
3729  *
3730  *	Create an association between the root of our namecache and
3731  *	the root vnode.  This routine may be called several times during
3732  *	booting.
3733  *
3734  *	If the caller intends to save the returned namecache pointer somewhere
3735  *	it must cache_hold() it.
3736  */
3737 void
3738 vfs_cache_setroot(struct vnode *nvp, struct nchandle *nch)
3739 {
3740 	struct vnode *ovp;
3741 	struct nchandle onch;
3742 
3743 	ovp = rootvnode;
3744 	onch = rootnch;
3745 	rootvnode = nvp;
3746 	if (nch)
3747 		rootnch = *nch;
3748 	else
3749 		cache_zero(&rootnch);
3750 	if (ovp)
3751 		vrele(ovp);
3752 	if (onch.ncp)
3753 		cache_drop(&onch);
3754 }
3755 
3756 /*
3757  * XXX OLD API COMPAT FUNCTION.  This really messes up the new namecache
3758  * topology and is being removed as quickly as possible.  The new VOP_N*()
3759  * API calls are required to make specific adjustments using the supplied
3760  * ncp pointers rather then just bogusly purging random vnodes.
3761  *
3762  * Invalidate all namecache entries to a particular vnode as well as
3763  * any direct children of that vnode in the namecache.  This is a
3764  * 'catch all' purge used by filesystems that do not know any better.
3765  *
3766  * Note that the linkage between the vnode and its namecache entries will
3767  * be removed, but the namecache entries themselves might stay put due to
3768  * active references from elsewhere in the system or due to the existance of
3769  * the children.   The namecache topology is left intact even if we do not
3770  * know what the vnode association is.  Such entries will be marked
3771  * NCF_UNRESOLVED.
3772  */
3773 void
3774 cache_purge(struct vnode *vp)
3775 {
3776 	cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN);
3777 }
3778 
3779 static int disablecwd;
3780 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0,
3781     "Disable getcwd");
3782 
3783 static u_long numcwdcalls;
3784 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdcalls, CTLFLAG_RD, &numcwdcalls, 0,
3785     "Number of current directory resolution calls");
3786 static u_long numcwdfailnf;
3787 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailnf, CTLFLAG_RD, &numcwdfailnf, 0,
3788     "Number of current directory failures due to lack of file");
3789 static u_long numcwdfailsz;
3790 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailsz, CTLFLAG_RD, &numcwdfailsz, 0,
3791     "Number of current directory failures due to large result");
3792 static u_long numcwdfound;
3793 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfound, CTLFLAG_RD, &numcwdfound, 0,
3794     "Number of current directory resolution successes");
3795 
3796 /*
3797  * MPALMOSTSAFE
3798  */
3799 int
3800 sys___getcwd(struct __getcwd_args *uap)
3801 {
3802 	u_int buflen;
3803 	int error;
3804 	char *buf;
3805 	char *bp;
3806 
3807 	if (disablecwd)
3808 		return (ENODEV);
3809 
3810 	buflen = uap->buflen;
3811 	if (buflen == 0)
3812 		return (EINVAL);
3813 	if (buflen > MAXPATHLEN)
3814 		buflen = MAXPATHLEN;
3815 
3816 	buf = kmalloc(buflen, M_TEMP, M_WAITOK);
3817 	bp = kern_getcwd(buf, buflen, &error);
3818 	if (error == 0)
3819 		error = copyout(bp, uap->buf, strlen(bp) + 1);
3820 	kfree(buf, M_TEMP);
3821 	return (error);
3822 }
3823 
3824 char *
3825 kern_getcwd(char *buf, size_t buflen, int *error)
3826 {
3827 	struct proc *p = curproc;
3828 	char *bp;
3829 	int i, slash_prefixed;
3830 	struct filedesc *fdp;
3831 	struct nchandle nch;
3832 	struct namecache *ncp;
3833 
3834 	numcwdcalls++;
3835 	bp = buf;
3836 	bp += buflen - 1;
3837 	*bp = '\0';
3838 	fdp = p->p_fd;
3839 	slash_prefixed = 0;
3840 
3841 	nch = fdp->fd_ncdir;
3842 	ncp = nch.ncp;
3843 	if (ncp)
3844 		_cache_hold(ncp);
3845 
3846 	while (ncp && (ncp != fdp->fd_nrdir.ncp ||
3847 	       nch.mount != fdp->fd_nrdir.mount)
3848 	) {
3849 		/*
3850 		 * While traversing upwards if we encounter the root
3851 		 * of the current mount we have to skip to the mount point
3852 		 * in the underlying filesystem.
3853 		 */
3854 		if (ncp == nch.mount->mnt_ncmountpt.ncp) {
3855 			nch = nch.mount->mnt_ncmounton;
3856 			_cache_drop(ncp);
3857 			ncp = nch.ncp;
3858 			if (ncp)
3859 				_cache_hold(ncp);
3860 			continue;
3861 		}
3862 
3863 		/*
3864 		 * Prepend the path segment
3865 		 */
3866 		for (i = ncp->nc_nlen - 1; i >= 0; i--) {
3867 			if (bp == buf) {
3868 				numcwdfailsz++;
3869 				*error = ERANGE;
3870 				bp = NULL;
3871 				goto done;
3872 			}
3873 			*--bp = ncp->nc_name[i];
3874 		}
3875 		if (bp == buf) {
3876 			numcwdfailsz++;
3877 			*error = ERANGE;
3878 			bp = NULL;
3879 			goto done;
3880 		}
3881 		*--bp = '/';
3882 		slash_prefixed = 1;
3883 
3884 		/*
3885 		 * Go up a directory.  This isn't a mount point so we don't
3886 		 * have to check again.
3887 		 */
3888 		while ((nch.ncp = ncp->nc_parent) != NULL) {
3889 			if (ncp_shared_lock_disable)
3890 				_cache_lock(ncp);
3891 			else
3892 				_cache_lock_shared(ncp);
3893 			if (nch.ncp != ncp->nc_parent) {
3894 				_cache_unlock(ncp);
3895 				continue;
3896 			}
3897 			_cache_hold(nch.ncp);
3898 			_cache_unlock(ncp);
3899 			break;
3900 		}
3901 		_cache_drop(ncp);
3902 		ncp = nch.ncp;
3903 	}
3904 	if (ncp == NULL) {
3905 		numcwdfailnf++;
3906 		*error = ENOENT;
3907 		bp = NULL;
3908 		goto done;
3909 	}
3910 	if (!slash_prefixed) {
3911 		if (bp == buf) {
3912 			numcwdfailsz++;
3913 			*error = ERANGE;
3914 			bp = NULL;
3915 			goto done;
3916 		}
3917 		*--bp = '/';
3918 	}
3919 	numcwdfound++;
3920 	*error = 0;
3921 done:
3922 	if (ncp)
3923 		_cache_drop(ncp);
3924 	return (bp);
3925 }
3926 
3927 /*
3928  * Thus begins the fullpath magic.
3929  *
3930  * The passed nchp is referenced but not locked.
3931  */
3932 static int disablefullpath;
3933 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW,
3934     &disablefullpath, 0,
3935     "Disable fullpath lookups");
3936 
3937 static u_int numfullpathcalls;
3938 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathcalls, CTLFLAG_RD,
3939     &numfullpathcalls, 0,
3940     "Number of full path resolutions in progress");
3941 static u_int numfullpathfailnf;
3942 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfailnf, CTLFLAG_RD,
3943     &numfullpathfailnf, 0,
3944     "Number of full path resolution failures due to lack of file");
3945 static u_int numfullpathfailsz;
3946 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfailsz, CTLFLAG_RD,
3947     &numfullpathfailsz, 0,
3948     "Number of full path resolution failures due to insufficient memory");
3949 static u_int numfullpathfound;
3950 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfound, CTLFLAG_RD,
3951     &numfullpathfound, 0,
3952     "Number of full path resolution successes");
3953 
3954 int
3955 cache_fullpath(struct proc *p, struct nchandle *nchp, struct nchandle *nchbase,
3956 	       char **retbuf, char **freebuf, int guess)
3957 {
3958 	struct nchandle fd_nrdir;
3959 	struct nchandle nch;
3960 	struct namecache *ncp;
3961 	struct mount *mp, *new_mp;
3962 	char *bp, *buf;
3963 	int slash_prefixed;
3964 	int error = 0;
3965 	int i;
3966 
3967 	atomic_add_int(&numfullpathcalls, -1);
3968 
3969 	*retbuf = NULL;
3970 	*freebuf = NULL;
3971 
3972 	buf = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK);
3973 	bp = buf + MAXPATHLEN - 1;
3974 	*bp = '\0';
3975 	if (nchbase)
3976 		fd_nrdir = *nchbase;
3977 	else if (p != NULL)
3978 		fd_nrdir = p->p_fd->fd_nrdir;
3979 	else
3980 		fd_nrdir = rootnch;
3981 	slash_prefixed = 0;
3982 	nch = *nchp;
3983 	ncp = nch.ncp;
3984 	if (ncp)
3985 		_cache_hold(ncp);
3986 	mp = nch.mount;
3987 
3988 	while (ncp && (ncp != fd_nrdir.ncp || mp != fd_nrdir.mount)) {
3989 		new_mp = NULL;
3990 
3991 		/*
3992 		 * If we are asked to guess the upwards path, we do so whenever
3993 		 * we encounter an ncp marked as a mountpoint. We try to find
3994 		 * the actual mountpoint by finding the mountpoint with this
3995 		 * ncp.
3996 		 */
3997 		if (guess && (ncp->nc_flag & NCF_ISMOUNTPT)) {
3998 			new_mp = mount_get_by_nc(ncp);
3999 		}
4000 		/*
4001 		 * While traversing upwards if we encounter the root
4002 		 * of the current mount we have to skip to the mount point.
4003 		 */
4004 		if (ncp == mp->mnt_ncmountpt.ncp) {
4005 			new_mp = mp;
4006 		}
4007 		if (new_mp) {
4008 			nch = new_mp->mnt_ncmounton;
4009 			_cache_drop(ncp);
4010 			ncp = nch.ncp;
4011 			if (ncp)
4012 				_cache_hold(ncp);
4013 			mp = nch.mount;
4014 			continue;
4015 		}
4016 
4017 		/*
4018 		 * Prepend the path segment
4019 		 */
4020 		for (i = ncp->nc_nlen - 1; i >= 0; i--) {
4021 			if (bp == buf) {
4022 				numfullpathfailsz++;
4023 				kfree(buf, M_TEMP);
4024 				error = ENOMEM;
4025 				goto done;
4026 			}
4027 			*--bp = ncp->nc_name[i];
4028 		}
4029 		if (bp == buf) {
4030 			numfullpathfailsz++;
4031 			kfree(buf, M_TEMP);
4032 			error = ENOMEM;
4033 			goto done;
4034 		}
4035 		*--bp = '/';
4036 		slash_prefixed = 1;
4037 
4038 		/*
4039 		 * Go up a directory.  This isn't a mount point so we don't
4040 		 * have to check again.
4041 		 *
4042 		 * We can only safely access nc_parent with ncp held locked.
4043 		 */
4044 		while ((nch.ncp = ncp->nc_parent) != NULL) {
4045 			_cache_lock(ncp);
4046 			if (nch.ncp != ncp->nc_parent) {
4047 				_cache_unlock(ncp);
4048 				continue;
4049 			}
4050 			_cache_hold(nch.ncp);
4051 			_cache_unlock(ncp);
4052 			break;
4053 		}
4054 		_cache_drop(ncp);
4055 		ncp = nch.ncp;
4056 	}
4057 	if (ncp == NULL) {
4058 		numfullpathfailnf++;
4059 		kfree(buf, M_TEMP);
4060 		error = ENOENT;
4061 		goto done;
4062 	}
4063 
4064 	if (!slash_prefixed) {
4065 		if (bp == buf) {
4066 			numfullpathfailsz++;
4067 			kfree(buf, M_TEMP);
4068 			error = ENOMEM;
4069 			goto done;
4070 		}
4071 		*--bp = '/';
4072 	}
4073 	numfullpathfound++;
4074 	*retbuf = bp;
4075 	*freebuf = buf;
4076 	error = 0;
4077 done:
4078 	if (ncp)
4079 		_cache_drop(ncp);
4080 	return(error);
4081 }
4082 
4083 int
4084 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf,
4085 	    char **freebuf, int guess)
4086 {
4087 	struct namecache *ncp;
4088 	struct nchandle nch;
4089 	int error;
4090 
4091 	*freebuf = NULL;
4092 	atomic_add_int(&numfullpathcalls, 1);
4093 	if (disablefullpath)
4094 		return (ENODEV);
4095 
4096 	if (p == NULL)
4097 		return (EINVAL);
4098 
4099 	/* vn is NULL, client wants us to use p->p_textvp */
4100 	if (vn == NULL) {
4101 		if ((vn = p->p_textvp) == NULL)
4102 			return (EINVAL);
4103 	}
4104 	spin_lock_shared(&vn->v_spin);
4105 	TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) {
4106 		if (ncp->nc_nlen)
4107 			break;
4108 	}
4109 	if (ncp == NULL) {
4110 		spin_unlock_shared(&vn->v_spin);
4111 		return (EINVAL);
4112 	}
4113 	_cache_hold(ncp);
4114 	spin_unlock_shared(&vn->v_spin);
4115 
4116 	atomic_add_int(&numfullpathcalls, -1);
4117 	nch.ncp = ncp;
4118 	nch.mount = vn->v_mount;
4119 	error = cache_fullpath(p, &nch, NULL, retbuf, freebuf, guess);
4120 	_cache_drop(ncp);
4121 	return (error);
4122 }
4123