xref: /dflybsd-src/sys/kern/vfs_nlookup.c (revision b227f3f50d5dc0f5fdecd8f9df23e96e8521baaf)
1 /*
2  * Copyright (c) 2004 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 /*
35  * nlookup() is the 'new' namei interface.  Rather then return directory and
36  * leaf vnodes (in various lock states) the new interface instead deals in
37  * namecache records.  Namecache records may represent both a positive or
38  * a negative hit.  The namespace is locked via the namecache record instead
39  * of via the vnode, and only the leaf namecache record (representing the
40  * filename) needs to be locked.
41  *
42  * This greatly improves filesystem parallelism and is a huge simplification
43  * of the API verses the old vnode locking / namei scheme.
44  *
45  * Filesystems must actively control the caching aspects of the namecache,
46  * and since namecache pointers are used as handles they are non-optional
47  * even for filesystems which do not generally wish to cache things.  It is
48  * intended that a separate cache coherency API will be constructed to handle
49  * these issues.
50  */
51 
52 #include "opt_ktrace.h"
53 
54 #include <sys/param.h>
55 #include <sys/systm.h>
56 #include <sys/kernel.h>
57 #include <sys/vnode.h>
58 #include <sys/mount.h>
59 #include <sys/filedesc.h>
60 #include <sys/proc.h>
61 #include <sys/namei.h>
62 #include <sys/nlookup.h>
63 #include <sys/malloc.h>
64 #include <sys/stat.h>
65 #include <sys/objcache.h>
66 #include <sys/file.h>
67 #include <sys/kcollect.h>
68 
69 #ifdef KTRACE
70 #include <sys/ktrace.h>
71 #endif
72 
73 static int naccess(struct nchandle *nch, int vmode, struct ucred *cred,
74 		int *stickyp);
75 
76 /*
77  * Initialize a nlookup() structure, early error return for copyin faults
78  * or a degenerate empty string (which is not allowed).
79  *
80  * The first process proc0's credentials are used if the calling thread
81  * is not associated with a process context.
82  *
83  * MPSAFE
84  */
85 int
86 nlookup_init(struct nlookupdata *nd,
87 	     const char *path, enum uio_seg seg, int flags)
88 {
89     size_t pathlen;
90     struct proc *p;
91     thread_t td;
92     int error;
93 
94     td = curthread;
95     p = td->td_proc;
96 
97     /*
98      * note: the pathlen set by copy*str() includes the terminating \0.
99      */
100     bzero(nd, sizeof(struct nlookupdata));
101     nd->nl_path = objcache_get(namei_oc, M_WAITOK);
102     nd->nl_flags |= NLC_HASBUF;
103     if (seg == UIO_SYSSPACE)
104 	error = copystr(path, nd->nl_path, MAXPATHLEN, &pathlen);
105     else
106 	error = copyinstr(path, nd->nl_path, MAXPATHLEN, &pathlen);
107 
108     /*
109      * Don't allow empty pathnames.
110      * POSIX.1 requirement: "" is not a vaild file name.
111      */
112     if (error == 0 && pathlen <= 1)
113 	error = ENOENT;
114 
115     if (error == 0) {
116 	if (p && p->p_fd) {
117 	    cache_copy_ncdir(p, &nd->nl_nch);
118 	    cache_copy(&p->p_fd->fd_nrdir, &nd->nl_rootnch);
119 	    if (p->p_fd->fd_njdir.ncp)
120 		cache_copy(&p->p_fd->fd_njdir, &nd->nl_jailnch);
121 	    nd->nl_cred = td->td_ucred;
122 	    nd->nl_flags |= NLC_BORROWCRED | NLC_NCDIR;
123 	} else {
124 	    cache_copy(&rootnch, &nd->nl_nch);
125 	    cache_copy(&nd->nl_nch, &nd->nl_rootnch);
126 	    cache_copy(&nd->nl_nch, &nd->nl_jailnch);
127 	    nd->nl_cred = proc0.p_ucred;
128 	    nd->nl_flags |= NLC_BORROWCRED;
129 	}
130 	nd->nl_td = td;
131 	nd->nl_flags |= flags;
132     } else {
133 	nlookup_done(nd);
134     }
135     return(error);
136 }
137 
138 
139 /*
140  * nlookup_init() for "at" family of syscalls.
141  *
142  * Works similarly to nlookup_init() but if path is relative and fd is not
143  * AT_FDCWD, path is interpreted relative to the directory pointed to by fd.
144  * In this case, the file entry pointed to by fd is ref'ed and returned in
145  * *fpp.
146  *
147  * If the call succeeds, nlookup_done_at() must be called to clean-up the nd
148  * and release the ref to the file entry.
149  */
150 int
151 nlookup_init_at(struct nlookupdata *nd, struct file **fpp, int fd,
152 		const char *path, enum uio_seg seg, int flags)
153 {
154 	struct thread *td = curthread;
155 	struct file* fp;
156 	struct vnode *vp;
157 	int error;
158 
159 	*fpp = NULL;
160 
161 	if  ((error = nlookup_init(nd, path, seg, flags)) != 0) {
162 		return (error);
163 	}
164 
165 	if (nd->nl_path[0] != '/' && fd != AT_FDCWD) {
166 		if ((error = holdvnode(td, fd, &fp)) != 0)
167 			goto done;
168 		vp = (struct vnode*)fp->f_data;
169 		if (vp->v_type != VDIR || fp->f_nchandle.ncp == NULL) {
170 			fdrop(fp);
171 			fp = NULL;
172 			error = ENOTDIR;
173 			goto done;
174 		}
175 		if (nd->nl_flags & NLC_NCDIR) {
176 			cache_drop_ncdir(&nd->nl_nch);
177 			nd->nl_flags &= ~NLC_NCDIR;
178 		} else {
179 			cache_drop(&nd->nl_nch);
180 		}
181 		cache_copy(&fp->f_nchandle, &nd->nl_nch);
182 		*fpp = fp;
183 	}
184 
185 
186 done:
187 	if (error)
188 		nlookup_done(nd);
189 	return (error);
190 
191 }
192 
193 /*
194  * This works similarly to nlookup_init() but does not assume a process
195  * context.  rootnch is always chosen for the root directory and the cred
196  * and starting directory are supplied in arguments.
197  */
198 int
199 nlookup_init_raw(struct nlookupdata *nd,
200 	     const char *path, enum uio_seg seg, int flags,
201 	     struct ucred *cred, struct nchandle *ncstart)
202 {
203     size_t pathlen;
204     thread_t td;
205     int error;
206 
207     td = curthread;
208 
209     bzero(nd, sizeof(struct nlookupdata));
210     nd->nl_path = objcache_get(namei_oc, M_WAITOK);
211     nd->nl_flags |= NLC_HASBUF;
212     if (seg == UIO_SYSSPACE)
213 	error = copystr(path, nd->nl_path, MAXPATHLEN, &pathlen);
214     else
215 	error = copyinstr(path, nd->nl_path, MAXPATHLEN, &pathlen);
216 
217     /*
218      * Don't allow empty pathnames.
219      * POSIX.1 requirement: "" is not a vaild file name.
220      */
221     if (error == 0 && pathlen <= 1)
222 	error = ENOENT;
223 
224     if (error == 0) {
225 	cache_copy(ncstart, &nd->nl_nch);
226 	cache_copy(&rootnch, &nd->nl_rootnch);
227 	cache_copy(&rootnch, &nd->nl_jailnch);
228 	nd->nl_cred = crhold(cred);
229 	nd->nl_td = td;
230 	nd->nl_flags |= flags;
231     } else {
232 	nlookup_done(nd);
233     }
234     return(error);
235 }
236 
237 /*
238  * This works similarly to nlookup_init_raw() but does not rely
239  * on rootnch being initialized yet.
240  */
241 int
242 nlookup_init_root(struct nlookupdata *nd,
243 	     const char *path, enum uio_seg seg, int flags,
244 	     struct ucred *cred, struct nchandle *ncstart,
245 	     struct nchandle *ncroot)
246 {
247     size_t pathlen;
248     thread_t td;
249     int error;
250 
251     td = curthread;
252 
253     bzero(nd, sizeof(struct nlookupdata));
254     nd->nl_path = objcache_get(namei_oc, M_WAITOK);
255     nd->nl_flags |= NLC_HASBUF;
256     if (seg == UIO_SYSSPACE)
257 	error = copystr(path, nd->nl_path, MAXPATHLEN, &pathlen);
258     else
259 	error = copyinstr(path, nd->nl_path, MAXPATHLEN, &pathlen);
260 
261     /*
262      * Don't allow empty pathnames.
263      * POSIX.1 requirement: "" is not a vaild file name.
264      */
265     if (error == 0 && pathlen <= 1)
266 	error = ENOENT;
267 
268     if (error == 0) {
269 	cache_copy(ncstart, &nd->nl_nch);
270 	cache_copy(ncroot, &nd->nl_rootnch);
271 	cache_copy(ncroot, &nd->nl_jailnch);
272 	nd->nl_cred = crhold(cred);
273 	nd->nl_td = td;
274 	nd->nl_flags |= flags;
275     } else {
276 	nlookup_done(nd);
277     }
278     return(error);
279 }
280 
281 #if 0
282 /*
283  * Set a different credential; this credential will be used by future
284  * operations performed on nd.nl_open_vp and nlookupdata structure.
285  */
286 void
287 nlookup_set_cred(struct nlookupdata *nd, struct ucred *cred)
288 {
289 	KKASSERT(nd->nl_cred != NULL);
290 
291 	if (nd->nl_cred != cred) {
292 		cred = crhold(cred);
293 		if ((nd->nl_flags & NLC_BORROWCRED) == 0)
294 			crfree(nd->nl_cred);
295 		nd->nl_flags &= ~NLC_BORROWCRED;
296 		nd->nl_cred = cred;
297 	}
298 }
299 #endif
300 
301 /*
302  * Cleanup a nlookupdata structure after we are through with it.  This may
303  * be called on any nlookupdata structure initialized with nlookup_init().
304  * Calling nlookup_done() is mandatory in all cases except where nlookup_init()
305  * returns an error, even if as a consumer you believe you have taken all
306  * dynamic elements out of the nlookupdata structure.
307  */
308 void
309 nlookup_done(struct nlookupdata *nd)
310 {
311     if (nd->nl_nch.ncp) {
312 	if (nd->nl_flags & NLC_NCPISLOCKED) {
313 	    nd->nl_flags &= ~NLC_NCPISLOCKED;
314 	    cache_unlock(&nd->nl_nch);
315 	}
316 	if (nd->nl_flags & NLC_NCDIR) {
317 		cache_drop_ncdir(&nd->nl_nch);
318 		nd->nl_flags &= ~NLC_NCDIR;
319 	} else {
320 		cache_drop(&nd->nl_nch);	/* NULL's out the nch */
321 	}
322     }
323     if (nd->nl_rootnch.ncp)
324 	cache_drop_and_cache(&nd->nl_rootnch);
325     if (nd->nl_jailnch.ncp)
326 	cache_drop_and_cache(&nd->nl_jailnch);
327     if ((nd->nl_flags & NLC_HASBUF) && nd->nl_path) {
328 	objcache_put(namei_oc, nd->nl_path);
329 	nd->nl_path = NULL;
330     }
331     if (nd->nl_cred) {
332 	if ((nd->nl_flags & NLC_BORROWCRED) == 0)
333 	    crfree(nd->nl_cred);
334 	nd->nl_cred = NULL;
335 	nd->nl_flags &= ~NLC_BORROWCRED;
336     }
337     if (nd->nl_open_vp) {
338 	if (nd->nl_flags & NLC_LOCKVP) {
339 		vn_unlock(nd->nl_open_vp);
340 		nd->nl_flags &= ~NLC_LOCKVP;
341 	}
342 	vn_close(nd->nl_open_vp, nd->nl_vp_fmode, NULL);
343 	nd->nl_open_vp = NULL;
344     }
345     if (nd->nl_dvp) {
346 	vrele(nd->nl_dvp);
347 	nd->nl_dvp = NULL;
348     }
349     nd->nl_flags = 0;	/* clear remaining flags (just clear everything) */
350 }
351 
352 /*
353  * Works similarly to nlookup_done() when nd initialized with
354  * nlookup_init_at().
355  */
356 void
357 nlookup_done_at(struct nlookupdata *nd, struct file *fp)
358 {
359 	nlookup_done(nd);
360 	if (fp != NULL)
361 		fdrop(fp);
362 }
363 
364 void
365 nlookup_zero(struct nlookupdata *nd)
366 {
367 	bzero(nd, sizeof(struct nlookupdata));
368 }
369 
370 /*
371  * Simple all-in-one nlookup.  Returns a locked namecache structure or NULL
372  * if an error occured.
373  *
374  * Note that the returned ncp is not checked for permissions, though VEXEC
375  * is checked on the directory path leading up to the result.  The caller
376  * must call naccess() to check the permissions of the returned leaf.
377  */
378 struct nchandle
379 nlookup_simple(const char *str, enum uio_seg seg,
380 	       int niflags, int *error)
381 {
382     struct nlookupdata nd;
383     struct nchandle nch;
384 
385     *error = nlookup_init(&nd, str, seg, niflags);
386     if (*error == 0) {
387 	    if ((*error = nlookup(&nd)) == 0) {
388 		    nch = nd.nl_nch;	/* keep hold ref from structure */
389 		    cache_zero(&nd.nl_nch); /* and NULL out */
390 	    } else {
391 		    cache_zero(&nch);
392 	    }
393 	    nlookup_done(&nd);
394     } else {
395 	    cache_zero(&nch);
396     }
397     return(nch);
398 }
399 
400 /*
401  * Returns non-zero if the path element is the last element
402  */
403 static
404 int
405 islastelement(const char *ptr)
406 {
407 	while (*ptr == '/')
408 		++ptr;
409 	return (*ptr == 0);
410 }
411 
412 /*
413  * Returns non-zero if we need to lock the namecache element
414  * exclusively.  Unless otherwise requested by NLC_SHAREDLOCK,
415  * the last element of the namecache lookup will be locked
416  * exclusively.
417  *
418  * NOTE: Even if we return on-zero, an unresolved namecache record
419  *	 will always be locked exclusively.
420  */
421 static __inline
422 int
423 wantsexcllock(struct nlookupdata *nd, const char *ptr)
424 {
425 	if ((nd->nl_flags & NLC_SHAREDLOCK) == 0)
426 		return(islastelement(ptr));
427 	return(0);
428 }
429 
430 
431 /*
432  * Do a generic nlookup.  Note that the passed nd is not nlookup_done()'d
433  * on return, even if an error occurs.  If no error occurs or NLC_CREATE
434  * is flagged and ENOENT is returned, then the returned nl_nch is always
435  * referenced and locked exclusively.
436  *
437  * WARNING: For any general error other than ENOENT w/NLC_CREATE, the
438  *	    the resulting nl_nch may or may not be locked and if locked
439  *	    might be locked either shared or exclusive.
440  *
441  * Intermediate directory elements, including the current directory, require
442  * execute (search) permission.  nlookup does not examine the access
443  * permissions on the returned element.
444  *
445  * If NLC_CREATE is set the last directory must allow node creation,
446  * and an error code of 0 will be returned for a non-existant
447  * target (not ENOENT).
448  *
449  * If NLC_RENAME_DST is set the last directory mut allow node deletion,
450  * plus the sticky check is made, and an error code of 0 will be returned
451  * for a non-existant target (not ENOENT).
452  *
453  * If NLC_DELETE is set the last directory mut allow node deletion,
454  * plus the sticky check is made.
455  *
456  * If NLC_REFDVP is set nd->nl_dvp will be set to the directory vnode
457  * of the returned entry.  The vnode will be referenced, but not locked,
458  * and will be released by nlookup_done() along with everything else.
459  *
460  * NOTE: As an optimization we attempt to obtain a shared namecache lock
461  *	 on any intermediate elements.  On success, the returned element
462  *	 is ALWAYS locked exclusively.
463  */
464 int
465 nlookup(struct nlookupdata *nd)
466 {
467     globaldata_t gd = mycpu;
468     struct nlcomponent nlc;
469     struct nchandle nch;
470     struct nchandle par;
471     struct nchandle nctmp;
472     struct mount *mp;
473     struct vnode *hvp;		/* hold to prevent recyclement */
474     int wasdotordotdot;
475     char *ptr;
476     char *nptr;
477     int error;
478     int len;
479     int dflags;
480     int hit = 1;
481     int saveflag = nd->nl_flags & ~NLC_NCDIR;
482     boolean_t doretry = FALSE;
483     boolean_t inretry = FALSE;
484 
485 nlookup_start:
486 #ifdef KTRACE
487     if (KTRPOINT(nd->nl_td, KTR_NAMEI))
488 	ktrnamei(nd->nl_td->td_lwp, nd->nl_path);
489 #endif
490     bzero(&nlc, sizeof(nlc));
491 
492     /*
493      * Setup for the loop.  The current working namecache element is
494      * always at least referenced.  We lock it as required, but always
495      * return a locked, resolved namecache entry.
496      */
497     nd->nl_loopcnt = 0;
498     if (nd->nl_dvp) {
499 	vrele(nd->nl_dvp);
500 	nd->nl_dvp = NULL;
501     }
502     ptr = nd->nl_path;
503 
504     /*
505      * Loop on the path components.  At the top of the loop nd->nl_nch
506      * is ref'd and unlocked and represents our current position.
507      */
508     for (;;) {
509 	/*
510 	 * Make sure nl_nch is locked so we can access the vnode, resolution
511 	 * state, etc.
512 	 */
513 	if ((nd->nl_flags & NLC_NCPISLOCKED) == 0) {
514 		nd->nl_flags |= NLC_NCPISLOCKED;
515 		cache_lock_maybe_shared(&nd->nl_nch, wantsexcllock(nd, ptr));
516 	}
517 
518 	/*
519 	 * Check if the root directory should replace the current
520 	 * directory.  This is done at the start of a translation
521 	 * or after a symbolic link has been found.  In other cases
522 	 * ptr will never be pointing at a '/'.
523 	 */
524 	if (*ptr == '/') {
525 	    do {
526 		++ptr;
527 	    } while (*ptr == '/');
528 	    cache_unlock(&nd->nl_nch);
529 	    cache_get_maybe_shared(&nd->nl_rootnch, &nch,
530 				   wantsexcllock(nd, ptr));
531 	    if (nd->nl_flags & NLC_NCDIR) {
532 		    cache_drop_ncdir(&nd->nl_nch);
533 		    nd->nl_flags &= ~NLC_NCDIR;
534 	    } else {
535 		    cache_drop(&nd->nl_nch);
536 	    }
537 	    nd->nl_nch = nch;		/* remains locked */
538 
539 	    /*
540 	     * Fast-track termination.  There is no parent directory of
541 	     * the root in the same mount from the point of view of
542 	     * the caller so return EACCES if NLC_REFDVP is specified,
543 	     * and EEXIST if NLC_CREATE is also specified.
544 	     * e.g. 'rmdir /' or 'mkdir /' are not allowed.
545 	     */
546 	    if (*ptr == 0) {
547 		if (nd->nl_flags & NLC_REFDVP)
548 			error = (nd->nl_flags & NLC_CREATE) ? EEXIST : EACCES;
549 		else
550 			error = 0;
551 		break;
552 	    }
553 	    continue;
554 	}
555 
556 	/*
557 	 * Pre-calculate next path component so we can check whether the
558 	 * current component directory is the last directory in the path
559 	 * or not.
560 	 */
561 	for (nptr = ptr; *nptr && *nptr != '/'; ++nptr)
562 		;
563 
564 	/*
565 	 * Check directory search permissions (nd->nl_nch is locked & refd).
566 	 * This will load dflags to obtain directory-special permissions to
567 	 * be checked along with the last component.
568 	 *
569 	 * We only need to pass-in &dflags for the second-to-last component.
570 	 * Optimize by passing-in NULL for any prior components, which may
571 	 * allow the code to bypass the naccess() call.
572 	 */
573 	dflags = 0;
574 	if (*nptr == '/')
575 	    error = naccess(&nd->nl_nch, NLC_EXEC, nd->nl_cred, NULL);
576 	else
577 	    error = naccess(&nd->nl_nch, NLC_EXEC, nd->nl_cred, &dflags);
578 	if (error)
579 	    break;
580 
581 	/*
582 	 * Extract the next (or last) path component.  Path components are
583 	 * limited to 255 characters.
584 	 */
585 	nlc.nlc_nameptr = ptr;
586 	nlc.nlc_namelen = nptr - ptr;
587 	ptr = nptr;
588 	if (nlc.nlc_namelen >= 256) {
589 	    error = ENAMETOOLONG;
590 	    break;
591 	}
592 
593 	/*
594 	 * Lookup the path component in the cache, creating an unresolved
595 	 * entry if necessary.  We have to handle "." and ".." as special
596 	 * cases.
597 	 *
598 	 * When handling ".." we have to detect a traversal back through a
599 	 * mount point.   If we are at the root, ".." just returns the root.
600 	 *
601 	 * When handling "." or ".." we also have to recalculate dflags
602 	 * since our dflags will be for some sub-directory instead of the
603 	 * parent dir.
604 	 *
605 	 * This subsection returns a locked, refd 'nch' unless it errors out,
606 	 * and an unlocked but still ref'd nd->nl_nch.
607 	 *
608 	 * The namecache topology is not allowed to be disconnected, so
609 	 * encountering a NULL parent will generate EINVAL.  This typically
610 	 * occurs when a directory is removed out from under a process.
611 	 *
612 	 * WARNING! The unlocking of nd->nl_nch is sensitive code.
613 	 */
614 	KKASSERT(nd->nl_flags & NLC_NCPISLOCKED);
615 
616 	if (nlc.nlc_namelen == 1 && nlc.nlc_nameptr[0] == '.') {
617 	    cache_unlock(&nd->nl_nch);
618 	    nd->nl_flags &= ~NLC_NCPISLOCKED;
619 	    cache_get_maybe_shared(&nd->nl_nch, &nch, wantsexcllock(nd, ptr));
620 	    wasdotordotdot = 1;
621 	} else if (nlc.nlc_namelen == 2 &&
622 		   nlc.nlc_nameptr[0] == '.' && nlc.nlc_nameptr[1] == '.') {
623 	    if (nd->nl_nch.mount == nd->nl_rootnch.mount &&
624 		nd->nl_nch.ncp == nd->nl_rootnch.ncp
625 	    ) {
626 		/*
627 		 * ".." at the root returns the root
628 		 */
629 		cache_unlock(&nd->nl_nch);
630 		nd->nl_flags &= ~NLC_NCPISLOCKED;
631 		cache_get_maybe_shared(&nd->nl_nch, &nch,
632 				       wantsexcllock(nd, ptr));
633 	    } else {
634 		/*
635 		 * Locate the parent ncp.  If we are at the root of a
636 		 * filesystem mount we have to skip to the mounted-on
637 		 * point in the underlying filesystem.
638 		 *
639 		 * Expect the parent to always be good since the
640 		 * mountpoint doesn't go away.  XXX hack.  cache_get()
641 		 * requires the ncp to already have a ref as a safety.
642 		 *
643 		 * However, a process which has been broken out of a chroot
644 		 * will wind up with a NULL parent if it tries to '..' above
645 		 * the real root, deal with the case.  Note that this does
646 		 * not protect us from a jail breakout, it just stops a panic
647 		 * if the jail-broken process tries to '..' past the real
648 		 * root.
649 		 */
650 		nctmp = nd->nl_nch;
651 		while (nctmp.ncp == nctmp.mount->mnt_ncmountpt.ncp) {
652 			nctmp = nctmp.mount->mnt_ncmounton;
653 			if (nctmp.ncp == NULL)
654 				break;
655 		}
656 		if (nctmp.ncp == NULL) {
657 			if (curthread->td_proc) {
658 				kprintf("vfs_nlookup: '..' traverse broke "
659 					"jail: pid %d (%s)\n",
660 					curthread->td_proc->p_pid,
661 					curthread->td_comm);
662 			}
663 			nctmp = nd->nl_rootnch;
664 		} else {
665 			nctmp.ncp = nctmp.ncp->nc_parent;
666 		}
667 		cache_hold(&nctmp);
668 		cache_unlock(&nd->nl_nch);
669 		nd->nl_flags &= ~NLC_NCPISLOCKED;
670 		cache_get_maybe_shared(&nctmp, &nch, wantsexcllock(nd, ptr));
671 		cache_drop(&nctmp);		/* NOTE: zero's nctmp */
672 	    }
673 	    wasdotordotdot = 2;
674 	} else {
675 	    /*
676 	     * Must unlock nl_nch when traversing down the path.  However,
677 	     * the child ncp has not yet been found/created and the parent's
678 	     * child list might be empty.  Thus releasing the lock can
679 	     * allow a race whereby the parent ncp's vnode is recycled.
680 	     * This case can occur especially when maxvnodes is set very low.
681 	     *
682 	     * We need the parent's ncp to remain resolved for all normal
683 	     * filesystem activities, so we vhold() the vp during the lookup
684 	     * to prevent recyclement due to vnlru / maxvnodes.
685 	     *
686 	     * If we race an unlink or rename the ncp might be marked
687 	     * DESTROYED after resolution, requiring a retry.
688 	     */
689 	    if ((hvp = nd->nl_nch.ncp->nc_vp) != NULL)
690 		vhold(hvp);
691 	    cache_unlock(&nd->nl_nch);
692 	    nd->nl_flags &= ~NLC_NCPISLOCKED;
693 	    error = cache_nlookup_maybe_shared(&nd->nl_nch, &nlc,
694 					       wantsexcllock(nd, ptr), &nch);
695 	    if (error == EWOULDBLOCK) {
696 		    nch = cache_nlookup(&nd->nl_nch, &nlc);
697 		    if (nch.ncp->nc_flag & NCF_UNRESOLVED)
698 			hit = 0;
699 		    for (;;) {
700 			error = cache_resolve(&nch, nd->nl_cred);
701 			if (error != EAGAIN &&
702 			    (nch.ncp->nc_flag & NCF_DESTROYED) == 0) {
703 				if (error == ESTALE) {
704 				    if (!inretry)
705 					error = ENOENT;
706 				    doretry = TRUE;
707 				}
708 				break;
709 			}
710 			kprintf("[diagnostic] nlookup: relookup %*.*s\n",
711 				nch.ncp->nc_nlen, nch.ncp->nc_nlen,
712 				nch.ncp->nc_name);
713 			cache_put(&nch);
714 			nch = cache_nlookup(&nd->nl_nch, &nlc);
715 		    }
716 	    }
717 	    if (hvp)
718 		vdrop(hvp);
719 	    wasdotordotdot = 0;
720 	}
721 
722 	/*
723 	 * If the last component was "." or ".." our dflags no longer
724 	 * represents the parent directory and we have to explicitly
725 	 * look it up.
726 	 *
727 	 * Expect the parent to be good since nch is locked.
728 	 */
729 	if (wasdotordotdot && error == 0) {
730 	    dflags = 0;
731 	    if ((par.ncp = nch.ncp->nc_parent) != NULL) {
732 		par.mount = nch.mount;
733 		cache_hold(&par);
734 		cache_lock_maybe_shared(&par, wantsexcllock(nd, ptr));
735 		error = naccess(&par, 0, nd->nl_cred, &dflags);
736 		cache_put(&par);
737 	    }
738 	}
739 
740 	/*
741 	 * [end of subsection]
742 	 *
743 	 * nch is locked and referenced.
744 	 * nd->nl_nch is unlocked and referenced.
745 	 *
746 	 * nl_nch must be unlocked or we could chain lock to the root
747 	 * if a resolve gets stuck (e.g. in NFS).
748 	 */
749 	KKASSERT((nd->nl_flags & NLC_NCPISLOCKED) == 0);
750 
751 	/*
752 	 * Resolve the namespace if necessary.  The ncp returned by
753 	 * cache_nlookup() is referenced and locked.
754 	 *
755 	 * XXX neither '.' nor '..' should return EAGAIN since they were
756 	 * previously resolved and thus cannot be newly created ncp's.
757 	 */
758 	if (nch.ncp->nc_flag & NCF_UNRESOLVED) {
759 	    hit = 0;
760 	    error = cache_resolve(&nch, nd->nl_cred);
761 	    if (error == ESTALE) {
762 		if (!inretry)
763 		    error = ENOENT;
764 		doretry = TRUE;
765 	    }
766 	    KKASSERT(error != EAGAIN);
767 	} else {
768 	    error = nch.ncp->nc_error;
769 	}
770 
771 	/*
772 	 * Early completion.  ENOENT is not an error if this is the last
773 	 * component and NLC_CREATE or NLC_RENAME (rename target) was
774 	 * requested.  Note that ncp->nc_error is left as ENOENT in that
775 	 * case, which we check later on.
776 	 *
777 	 * Also handle invalid '.' or '..' components terminating a path
778 	 * for a create/rename/delete.  The standard requires this and pax
779 	 * pretty stupidly depends on it.
780 	 */
781 	if (islastelement(ptr)) {
782 	    if (error == ENOENT &&
783 		(nd->nl_flags & (NLC_CREATE | NLC_RENAME_DST))
784 	    ) {
785 		if (nd->nl_flags & NLC_NFS_RDONLY) {
786 			error = EROFS;
787 		} else {
788 			error = naccess(&nch, nd->nl_flags | dflags,
789 					nd->nl_cred, NULL);
790 		}
791 	    }
792 	    if (error == 0 && wasdotordotdot &&
793 		(nd->nl_flags & (NLC_CREATE | NLC_DELETE |
794 				 NLC_RENAME_SRC | NLC_RENAME_DST))) {
795 		/*
796 		 * POSIX junk
797 		 */
798 		if (nd->nl_flags & NLC_CREATE)
799 			error = EEXIST;
800 		else if (nd->nl_flags & NLC_DELETE)
801 			error = (wasdotordotdot == 1) ? EINVAL : ENOTEMPTY;
802 		else
803 			error = EINVAL;
804 	    }
805 	}
806 
807 	/*
808 	 * Early completion on error.
809 	 */
810 	if (error) {
811 	    cache_put(&nch);
812 	    break;
813 	}
814 
815 	/*
816 	 * If the element is a symlink and it is either not the last
817 	 * element or it is the last element and we are allowed to
818 	 * follow symlinks, resolve the symlink.
819 	 */
820 	if ((nch.ncp->nc_flag & NCF_ISSYMLINK) &&
821 	    (*ptr || (nd->nl_flags & NLC_FOLLOW))
822 	) {
823 	    if (nd->nl_loopcnt++ >= MAXSYMLINKS) {
824 		error = ELOOP;
825 		cache_put(&nch);
826 		break;
827 	    }
828 	    error = nreadsymlink(nd, &nch, &nlc);
829 	    cache_put(&nch);
830 	    if (error)
831 		break;
832 
833 	    /*
834 	     * Concatenate trailing path elements onto the returned symlink.
835 	     * Note that if the path component (ptr) is not exhausted, it
836 	     * will being with a '/', so we do not have to add another one.
837 	     *
838 	     * The symlink may not be empty.
839 	     */
840 	    len = strlen(ptr);
841 	    if (nlc.nlc_namelen == 0 || nlc.nlc_namelen + len >= MAXPATHLEN) {
842 		error = nlc.nlc_namelen ? ENAMETOOLONG : ENOENT;
843 		objcache_put(namei_oc, nlc.nlc_nameptr);
844 		break;
845 	    }
846 	    bcopy(ptr, nlc.nlc_nameptr + nlc.nlc_namelen, len + 1);
847 	    if (nd->nl_flags & NLC_HASBUF)
848 		objcache_put(namei_oc, nd->nl_path);
849 	    nd->nl_path = nlc.nlc_nameptr;
850 	    nd->nl_flags |= NLC_HASBUF;
851 	    ptr = nd->nl_path;
852 
853 	    /*
854 	     * Go back up to the top to resolve any initial '/'s in the
855 	     * symlink.
856 	     */
857 	    continue;
858 	}
859 
860 	/*
861 	 * If the element is a directory and we are crossing a mount point,
862 	 * Locate the mount.
863 	 */
864 	while ((nch.ncp->nc_flag & NCF_ISMOUNTPT) &&
865 	    (nd->nl_flags & NLC_NOCROSSMOUNT) == 0 &&
866 	    (mp = cache_findmount(&nch)) != NULL
867 	) {
868 	    struct vnode *tdp;
869 	    int vfs_do_busy = 0;
870 
871 	    /*
872 	     * VFS must be busied before the namecache entry is locked,
873 	     * but we don't want to waste time calling vfs_busy() if the
874 	     * mount point is already resolved.
875 	     */
876 again:
877 	    cache_put(&nch);
878 	    if (vfs_do_busy) {
879 		while (vfs_busy(mp, 0)) {
880 		    if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
881 			kprintf("nlookup: warning umount race avoided\n");
882 			cache_dropmount(mp);
883 			error = EBUSY;
884 			vfs_do_busy = 0;
885 			goto double_break;
886 		    }
887 		}
888 	    }
889 	    cache_get_maybe_shared(&mp->mnt_ncmountpt, &nch,
890 				   wantsexcllock(nd, ptr));
891 
892 	    if (nch.ncp->nc_flag & NCF_UNRESOLVED) {
893 		if (vfs_do_busy == 0) {
894 		    vfs_do_busy = 1;
895 		    goto again;
896 		}
897 		error = VFS_ROOT(mp, &tdp);
898 		vfs_unbusy(mp);
899 		vfs_do_busy = 0;
900 		if (error) {
901 		    cache_dropmount(mp);
902 		    break;
903 		}
904 		cache_setvp(&nch, tdp);
905 		vput(tdp);
906 	    }
907 	    if (vfs_do_busy)
908 		vfs_unbusy(mp);
909 	    cache_dropmount(mp);
910 	}
911 
912 	if (error) {
913 	    cache_put(&nch);
914 double_break:
915 	    break;
916 	}
917 
918 	/*
919 	 * Skip any slashes to get to the next element.  If there
920 	 * are any slashes at all the current element must be a
921 	 * directory or, in the create case, intended to become a directory.
922 	 * If it isn't we break without incrementing ptr and fall through
923 	 * to the failure case below.
924 	 */
925 	while (*ptr == '/') {
926 	    if ((nch.ncp->nc_flag & NCF_ISDIR) == 0 &&
927 		!(nd->nl_flags & NLC_WILLBEDIR)
928 	    ) {
929 		break;
930 	    }
931 	    ++ptr;
932 	}
933 
934 	/*
935 	 * Continuation case: additional elements and the current
936 	 * element is a directory.
937 	 */
938 	if (*ptr && (nch.ncp->nc_flag & NCF_ISDIR)) {
939 	    if (nd->nl_flags & NLC_NCDIR) {
940 		    cache_drop_ncdir(&nd->nl_nch);
941 		    nd->nl_flags &= ~NLC_NCDIR;
942 	    } else {
943 		    cache_drop(&nd->nl_nch);
944 	    }
945 	    cache_unlock(&nch);
946 	    KKASSERT((nd->nl_flags & NLC_NCPISLOCKED) == 0);
947 	    nd->nl_nch = nch;
948 	    continue;
949 	}
950 
951 	/*
952 	 * Failure case: additional elements and the current element
953 	 * is not a directory
954 	 */
955 	if (*ptr) {
956 	    cache_put(&nch);
957 	    error = ENOTDIR;
958 	    break;
959 	}
960 
961 	/*
962 	 * Successful lookup of last element.
963 	 *
964 	 * Check permissions if the target exists.  If the target does not
965 	 * exist directory permissions were already tested in the early
966 	 * completion code above.
967 	 *
968 	 * nd->nl_flags will be adjusted on return with NLC_APPENDONLY
969 	 * if the file is marked append-only, and NLC_STICKY if the directory
970 	 * containing the file is sticky.
971 	 */
972 	if (nch.ncp->nc_vp && (nd->nl_flags & NLC_ALLCHKS)) {
973 	    error = naccess(&nch, nd->nl_flags | dflags,
974 			    nd->nl_cred, NULL);
975 	    if (error) {
976 		cache_put(&nch);
977 		break;
978 	    }
979 	}
980 
981 	/*
982 	 * Termination: no more elements.
983 	 *
984 	 * If NLC_REFDVP is set acquire a referenced parent dvp.
985 	 */
986 	if (nd->nl_flags & NLC_REFDVP) {
987 		cache_lock(&nd->nl_nch);
988 		error = cache_vref(&nd->nl_nch, nd->nl_cred, &nd->nl_dvp);
989 		cache_unlock(&nd->nl_nch);
990 		if (error) {
991 			kprintf("NLC_REFDVP: Cannot ref dvp of %p\n", nch.ncp);
992 			cache_put(&nch);
993 			break;
994 		}
995 	}
996 	if (nd->nl_flags & NLC_NCDIR) {
997 		cache_drop_ncdir(&nd->nl_nch);
998 		nd->nl_flags &= ~NLC_NCDIR;
999 	} else {
1000 		cache_drop(&nd->nl_nch);
1001 	}
1002 	nd->nl_nch = nch;
1003 	nd->nl_flags |= NLC_NCPISLOCKED;
1004 	error = 0;
1005 	break;
1006     }
1007 
1008     if (hit)
1009 	++gd->gd_nchstats->ncs_longhits;
1010     else
1011 	++gd->gd_nchstats->ncs_longmiss;
1012 
1013     if (nd->nl_flags & NLC_NCPISLOCKED)
1014 	KKASSERT(cache_lockstatus(&nd->nl_nch) > 0);
1015 
1016     /*
1017      * Retry the whole thing if doretry flag is set, but only once.
1018      * autofs(5) may mount another filesystem under its root directory
1019      * while resolving a path.
1020      */
1021     if (doretry && !inretry) {
1022 	inretry = TRUE;
1023 	nd->nl_flags &= NLC_NCDIR;
1024 	nd->nl_flags |= saveflag;
1025 	goto nlookup_start;
1026     }
1027 
1028     /*
1029      * NOTE: If NLC_CREATE was set the ncp may represent a negative hit
1030      * (ncp->nc_error will be ENOENT), but we will still return an error
1031      * code of 0.
1032      */
1033     return(error);
1034 }
1035 
1036 /*
1037  * Resolve a mount point's glue ncp.  This ncp connects creates the illusion
1038  * of continuity in the namecache tree by connecting the ncp related to the
1039  * vnode under the mount to the ncp related to the mount's root vnode.
1040  *
1041  * If no error occured a locked, ref'd ncp is stored in *ncpp.
1042  */
1043 int
1044 nlookup_mp(struct mount *mp, struct nchandle *nch)
1045 {
1046     struct vnode *vp;
1047     int error;
1048 
1049     error = 0;
1050     cache_get(&mp->mnt_ncmountpt, nch);
1051     if (nch->ncp->nc_flag & NCF_UNRESOLVED) {
1052 	while (vfs_busy(mp, 0))
1053 	    ;
1054 	error = VFS_ROOT(mp, &vp);
1055 	vfs_unbusy(mp);
1056 	if (error) {
1057 	    cache_put(nch);
1058 	} else {
1059 	    cache_setvp(nch, vp);
1060 	    vput(vp);
1061 	}
1062     }
1063     return(error);
1064 }
1065 
1066 /*
1067  * Read the contents of a symlink, allocate a path buffer out of the
1068  * namei_oc and initialize the supplied nlcomponent with the result.
1069  *
1070  * If an error occurs no buffer will be allocated or returned in the nlc.
1071  */
1072 int
1073 nreadsymlink(struct nlookupdata *nd, struct nchandle *nch,
1074 		struct nlcomponent *nlc)
1075 {
1076     struct vnode *vp;
1077     struct iovec aiov;
1078     struct uio auio;
1079     int linklen;
1080     int error;
1081     char *cp;
1082 
1083     nlc->nlc_nameptr = NULL;
1084     nlc->nlc_namelen = 0;
1085     if (nch->ncp->nc_vp == NULL)
1086 	return(ENOENT);
1087     if ((error = cache_vget(nch, nd->nl_cred, LK_SHARED, &vp)) != 0)
1088 	return(error);
1089     cp = objcache_get(namei_oc, M_WAITOK);
1090     aiov.iov_base = cp;
1091     aiov.iov_len = MAXPATHLEN;
1092     auio.uio_iov = &aiov;
1093     auio.uio_iovcnt = 1;
1094     auio.uio_offset = 0;
1095     auio.uio_rw = UIO_READ;
1096     auio.uio_segflg = UIO_SYSSPACE;
1097     auio.uio_td = nd->nl_td;
1098     auio.uio_resid = MAXPATHLEN - 1;
1099     error = VOP_READLINK(vp, &auio, nd->nl_cred);
1100     if (error)
1101 	goto fail;
1102     linklen = MAXPATHLEN - 1 - auio.uio_resid;
1103     if (varsym_enable) {
1104 	linklen = varsymreplace(cp, linklen, MAXPATHLEN - 1);
1105 	if (linklen < 0) {
1106 	    error = ENAMETOOLONG;
1107 	    goto fail;
1108 	}
1109     }
1110     cp[linklen] = 0;
1111     nlc->nlc_nameptr = cp;
1112     nlc->nlc_namelen = linklen;
1113     vput(vp);
1114     return(0);
1115 fail:
1116     objcache_put(namei_oc, cp);
1117     vput(vp);
1118     return(error);
1119 }
1120 
1121 /*
1122  * Check access [XXX cache vattr!] [XXX quota]
1123  *
1124  * Generally check the NLC_* access bits.   All specified bits must pass
1125  * for this function to return 0.
1126  *
1127  * The file does not have to exist when checking NLC_CREATE or NLC_RENAME_DST
1128  * access, otherwise it must exist.  No error is returned in this case.
1129  *
1130  * The file must not exist if NLC_EXCL is specified.
1131  *
1132  * Directory permissions in general are tested for NLC_CREATE if the file
1133  * does not exist, NLC_DELETE if the file does exist, and NLC_RENAME_DST
1134  * whether the file exists or not.
1135  *
1136  * The directory sticky bit is tested for NLC_DELETE and NLC_RENAME_DST,
1137  * the latter is only tested if the target exists.
1138  *
1139  * The passed ncp must be referenced and locked.  If it is already resolved
1140  * it may be locked shared but otherwise should be locked exclusively.
1141  */
1142 
1143 #define S_WXOK_MASK	(S_IRUSR|S_IXUSR|S_IRGRP|S_IXGRP|S_IROTH|S_IXOTH)
1144 
1145 static int
1146 naccess(struct nchandle *nch, int nflags, struct ucred *cred, int *nflagsp)
1147 {
1148     struct vnode *vp;
1149     struct vattr va;
1150     struct namecache *ncp;
1151     int error;
1152     int cflags;
1153 
1154     KKASSERT(cache_lockstatus(nch) > 0);
1155 
1156     ncp = nch->ncp;
1157     if (ncp->nc_flag & NCF_UNRESOLVED) {
1158 	cache_resolve(nch, cred);
1159 	ncp = nch->ncp;
1160     }
1161     error = ncp->nc_error;
1162 
1163     /*
1164      * Directory permissions checks.  Silently ignore ENOENT if these
1165      * tests pass.  It isn't an error.
1166      *
1167      * We can safely resolve ncp->nc_parent because ncp is currently
1168      * locked.
1169      */
1170     if (nflags & (NLC_CREATE | NLC_DELETE | NLC_RENAME_SRC | NLC_RENAME_DST)) {
1171 	if (((nflags & NLC_CREATE) && ncp->nc_vp == NULL) ||
1172 	    ((nflags & NLC_DELETE) && ncp->nc_vp != NULL) ||
1173 	    ((nflags & NLC_RENAME_SRC) && ncp->nc_vp != NULL) ||
1174 	    (nflags & NLC_RENAME_DST)
1175 	) {
1176 	    struct nchandle par;
1177 
1178 	    if ((par.ncp = ncp->nc_parent) == NULL) {
1179 		if (error != EAGAIN)
1180 			error = EINVAL;
1181 	    } else if (error == 0 || error == ENOENT) {
1182 		par.mount = nch->mount;
1183 		cache_hold(&par);
1184 		cache_lock_maybe_shared(&par, 0);
1185 		error = naccess(&par, NLC_WRITE, cred, NULL);
1186 		cache_put(&par);
1187 	    }
1188 	}
1189     }
1190 
1191     /*
1192      * NLC_EXCL check.  Target file must not exist.
1193      */
1194     if (error == 0 && (nflags & NLC_EXCL) && ncp->nc_vp != NULL)
1195 	error = EEXIST;
1196 
1197     /*
1198      * Try to short-cut the vnode operation for intermediate directory
1199      * components.  This is a major SMP win because it avoids having
1200      * to execute a lot of code for intermediate directory components,
1201      * including shared refs and locks on intermediate directory vnodes.
1202      *
1203      * We can only do this if the caller does not need nflagsp.
1204      */
1205     if (error == 0 && nflagsp == NULL &&
1206 	nflags == NLC_EXEC && (ncp->nc_flag & NCF_WXOK)) {
1207 	return 0;
1208     }
1209 
1210     /*
1211      * Get the vnode attributes so we can do the rest of our checks.
1212      *
1213      * NOTE: We only call naccess_va() if the target exists.
1214      */
1215     if (error == 0) {
1216 	error = cache_vget(nch, cred, LK_SHARED, &vp);
1217 	if (error == ENOENT) {
1218 	    /*
1219 	     * Silently zero-out ENOENT if creating or renaming
1220 	     * (rename target).  It isn't an error.
1221 	     */
1222 	    if (nflags & (NLC_CREATE | NLC_RENAME_DST))
1223 		error = 0;
1224 	} else if (error == 0) {
1225 	    /*
1226 	     * Get the vnode attributes and check for illegal O_TRUNC
1227 	     * requests and read-only mounts.
1228 	     *
1229 	     * NOTE: You can still open devices on read-only mounts for
1230 	     * 	     writing.
1231 	     *
1232 	     * NOTE: creates/deletes/renames are handled by the NLC_WRITE
1233 	     *	     check on the parent directory above.
1234 	     *
1235 	     * XXX cache the va in the namecache or in the vnode
1236 	     */
1237 	    error = VOP_GETATTR(vp, &va);
1238 	    if (error == 0 && (nflags & NLC_TRUNCATE)) {
1239 		switch(va.va_type) {
1240 		case VREG:
1241 		case VDATABASE:
1242 		case VCHR:
1243 		case VBLK:
1244 		case VFIFO:
1245 		    break;
1246 		case VDIR:
1247 		    error = EISDIR;
1248 		    break;
1249 		default:
1250 		    error = EINVAL;
1251 		    break;
1252 		}
1253 	    }
1254 	    if (error == 0 && (nflags & NLC_WRITE) && vp->v_mount &&
1255 		(vp->v_mount->mnt_flag & MNT_RDONLY)
1256 	    ) {
1257 		switch(va.va_type) {
1258 		case VDIR:
1259 		case VLNK:
1260 		case VREG:
1261 		case VDATABASE:
1262 		    error = EROFS;
1263 		    break;
1264 		default:
1265 		    break;
1266 		}
1267 	    }
1268 	    vput(vp);
1269 
1270 	    /*
1271 	     * Check permissions based on file attributes.  The passed
1272 	     * flags (*nflagsp) are modified with feedback based on
1273 	     * special attributes and requirements.
1274 	     */
1275 	    if (error == 0) {
1276 		/*
1277 		 * Adjust the returned (*nflagsp) if non-NULL.
1278 		 */
1279 		if (nflagsp) {
1280 		    if ((va.va_mode & VSVTX) && va.va_uid != cred->cr_uid)
1281 			*nflagsp |= NLC_STICKY;
1282 		    if (va.va_flags & APPEND)
1283 			*nflagsp |= NLC_APPENDONLY;
1284 		    if (va.va_flags & IMMUTABLE)
1285 			*nflagsp |= NLC_IMMUTABLE;
1286 		}
1287 
1288 		/*
1289 		 * NCF_WXOK can be set for world-searchable directories.
1290 		 *
1291 		 * XXX When we implement capabilities this code would also
1292 		 * need a cap check, or only set the flag if there are no
1293 		 * capabilities.
1294 		 */
1295 		cflags = 0;
1296 		if (va.va_type == VDIR &&
1297 		    (va.va_mode & S_WXOK_MASK) == S_WXOK_MASK) {
1298 			cflags |= NCF_WXOK;
1299 		}
1300 
1301 		/*
1302 		 * Track swapcache management flags in the namecache.
1303 		 *
1304 		 * Calculate the flags based on the current vattr info
1305 		 * and recalculate the inherited flags from the parent
1306 		 * (the original cache linkage may have occurred without
1307 		 * getattrs and thus have stale flags).
1308 		 */
1309 		if (va.va_flags & SF_NOCACHE)
1310 			cflags |= NCF_SF_NOCACHE;
1311 		if (va.va_flags & UF_CACHE)
1312 			cflags |= NCF_UF_CACHE;
1313 		if (ncp->nc_parent) {
1314 			if (ncp->nc_parent->nc_flag &
1315 			    (NCF_SF_NOCACHE | NCF_SF_PNOCACHE)) {
1316 				cflags |= NCF_SF_PNOCACHE;
1317 			}
1318 			if (ncp->nc_parent->nc_flag &
1319 			    (NCF_UF_CACHE | NCF_UF_PCACHE)) {
1320 				cflags |= NCF_UF_PCACHE;
1321 			}
1322 		}
1323 
1324 		/*
1325 		 * We're not supposed to update nc_flag when holding a shared
1326 		 * lock, but we allow the case for certain flags.  Note that
1327 		 * holding an exclusive lock allows updating nc_flag without
1328 		 * atomics.  nc_flag is not allowe to be updated at all unless
1329 		 * a shared or exclusive lock is held.
1330 		 */
1331 		atomic_clear_short(&ncp->nc_flag,
1332 				   (NCF_SF_NOCACHE | NCF_UF_CACHE |
1333 				   NCF_SF_PNOCACHE | NCF_UF_PCACHE |
1334 				   NCF_WXOK) & ~cflags);
1335 		atomic_set_short(&ncp->nc_flag, cflags);
1336 
1337 		/*
1338 		 * Process general access.
1339 		 */
1340 		error = naccess_va(&va, nflags, cred);
1341 	    }
1342 	}
1343     }
1344     return(error);
1345 }
1346 
1347 /*
1348  * Check the requested access against the given vattr using cred.
1349  */
1350 int
1351 naccess_va(struct vattr *va, int nflags, struct ucred *cred)
1352 {
1353     int i;
1354     int vmode;
1355 
1356     /*
1357      * Test the immutable bit.  Creations, deletions, renames (source
1358      * or destination) are not allowed.  chown/chmod/other is also not
1359      * allowed but is handled by SETATTR.  Hardlinks to the immutable
1360      * file are allowed.
1361      *
1362      * If the directory is set to immutable then creations, deletions,
1363      * renames (source or dest) and hardlinks to files within the directory
1364      * are not allowed, and regular files opened through the directory may
1365      * not be written to or truncated (unless a special device).
1366      *
1367      * NOTE!  New hardlinks to immutable files work but new hardlinks to
1368      * files, immutable or not, sitting inside an immutable directory are
1369      * not allowed.  As always if the file is hardlinked via some other
1370      * path additional hardlinks may be possible even if the file is marked
1371      * immutable.  The sysop needs to create a closure by checking the hard
1372      * link count.  Once closure is achieved you are good, and security
1373      * scripts should check link counts anyway.
1374      *
1375      * Writes and truncations are only allowed on special devices.
1376      */
1377     if ((va->va_flags & IMMUTABLE) || (nflags & NLC_IMMUTABLE)) {
1378 	if ((nflags & NLC_IMMUTABLE) && (nflags & NLC_HLINK))
1379 	    return (EPERM);
1380 	if (nflags & (NLC_CREATE | NLC_DELETE |
1381 		      NLC_RENAME_SRC | NLC_RENAME_DST)) {
1382 	    return (EPERM);
1383 	}
1384 	if (nflags & (NLC_WRITE | NLC_TRUNCATE)) {
1385 	    switch(va->va_type) {
1386 	    case VDIR:
1387 		return (EISDIR);
1388 	    case VLNK:
1389 	    case VREG:
1390 	    case VDATABASE:
1391 		return (EPERM);
1392 	    default:
1393 		break;
1394 	    }
1395 	}
1396     }
1397 
1398     /*
1399      * Test the no-unlink and append-only bits for opens, rename targets,
1400      * and deletions.  These bits are not tested for creations or
1401      * rename sources.
1402      *
1403      * Unlike FreeBSD we allow a file with APPEND set to be renamed.
1404      * If you do not wish this you must also set NOUNLINK.
1405      *
1406      * If the governing directory is marked APPEND-only it implies
1407      * NOUNLINK for all entries in the directory.
1408      */
1409     if (((va->va_flags & NOUNLINK) || (nflags & NLC_APPENDONLY)) &&
1410 	(nflags & (NLC_DELETE | NLC_RENAME_SRC | NLC_RENAME_DST))
1411     ) {
1412 	return (EPERM);
1413     }
1414 
1415     /*
1416      * A file marked append-only may not be deleted but can be renamed.
1417      */
1418     if ((va->va_flags & APPEND) &&
1419 	(nflags & (NLC_DELETE | NLC_RENAME_DST))
1420     ) {
1421 	return (EPERM);
1422     }
1423 
1424     /*
1425      * A file marked append-only which is opened for writing must also
1426      * be opened O_APPEND.
1427      */
1428     if ((va->va_flags & APPEND) && (nflags & (NLC_OPEN | NLC_TRUNCATE))) {
1429 	if (nflags & NLC_TRUNCATE)
1430 	    return (EPERM);
1431 	if ((nflags & (NLC_OPEN | NLC_WRITE)) == (NLC_OPEN | NLC_WRITE)) {
1432 	    if ((nflags & NLC_APPEND) == 0)
1433 		return (EPERM);
1434 	}
1435     }
1436 
1437     /*
1438      * root gets universal access
1439      */
1440     if (cred->cr_uid == 0)
1441 	return(0);
1442 
1443     /*
1444      * Check owner perms.
1445      *
1446      * If NLC_OWN is set the owner of the file is allowed no matter when
1447      * the owner-mode bits say (utimes).
1448      */
1449     vmode = 0;
1450     if (nflags & NLC_READ)
1451 	vmode |= S_IRUSR;
1452     if (nflags & NLC_WRITE)
1453 	vmode |= S_IWUSR;
1454     if (nflags & NLC_EXEC)
1455 	vmode |= S_IXUSR;
1456 
1457     if (cred->cr_uid == va->va_uid) {
1458 	if ((nflags & NLC_OWN) == 0) {
1459 	    if ((vmode & va->va_mode) != vmode)
1460 		return(EACCES);
1461 	}
1462 	return(0);
1463     }
1464 
1465     /*
1466      * If NLC_STICKY is set only the owner may delete or rename a file.
1467      * This bit is typically set on /tmp.
1468      *
1469      * Note that the NLC_READ/WRITE/EXEC bits are not typically set in
1470      * the specific delete or rename case.  For deletions and renames we
1471      * usually just care about directory permissions, not file permissions.
1472      */
1473     if ((nflags & NLC_STICKY) &&
1474 	(nflags & (NLC_RENAME_SRC | NLC_RENAME_DST | NLC_DELETE))) {
1475 	return(EACCES);
1476     }
1477 
1478     /*
1479      * Check group perms
1480      */
1481     vmode >>= 3;
1482     for (i = 0; i < cred->cr_ngroups; ++i) {
1483 	if (va->va_gid == cred->cr_groups[i]) {
1484 	    if ((vmode & va->va_mode) != vmode)
1485 		return(EACCES);
1486 	    return(0);
1487 	}
1488     }
1489 
1490     /*
1491      * Check world perms
1492      */
1493     vmode >>= 3;
1494     if ((vmode & va->va_mode) != vmode)
1495 	return(EACCES);
1496     return(0);
1497 }
1498 
1499 /*
1500  * Long-term (10-second interval) statistics collection
1501  */
1502 static
1503 uint64_t
1504 collect_nlookup_callback(int n)
1505 {
1506 	static uint64_t last_total;
1507 	uint64_t save;
1508 	uint64_t total;
1509 
1510 	total = 0;
1511 	for (n = 0; n < ncpus; ++n) {
1512 		globaldata_t gd = globaldata_find(n);
1513 		struct nchstats *sp;
1514 
1515 		if ((sp = gd->gd_nchstats) != NULL)
1516 			total += sp->ncs_longhits + sp->ncs_longmiss;
1517 	}
1518 	save = total;
1519 	total = total - last_total;
1520 	last_total = save;
1521 
1522 	return total;
1523 }
1524 
1525 static
1526 void
1527 nlookup_collect_init(void *dummy __unused)
1528 {
1529 	kcollect_register(KCOLLECT_NLOOKUP, "nlookup", collect_nlookup_callback,
1530 			  KCOLLECT_SCALE(KCOLLECT_NLOOKUP_FORMAT, 0));
1531 }
1532 SYSINIT(collect_nlookup, SI_SUB_PROP, SI_ORDER_ANY, nlookup_collect_init, 0);
1533