xref: /titanic_51/usr/src/uts/common/fs/lookup.c (revision a1bf3f785ae05c419b339c3a2061f2b18c024f61)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
24  * Copyright (c) 1988, 2010, Oracle and/or its affiliates. All rights reserved.
25  * Copyright (c) 2015, Joyent, Inc.
26  */
27 
28 /*	Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T	*/
29 /*	  All Rights Reserved  	*/
30 
31 /*
32  * University Copyright- Copyright (c) 1982, 1986, 1988
33  * The Regents of the University of California
34  * All Rights Reserved
35  *
36  * University Acknowledgment- Portions of this document are derived from
37  * software developed by the University of California, Berkeley, and its
38  * contributors.
39  */
40 
41 #include <sys/types.h>
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/cpuvar.h>
45 #include <sys/errno.h>
46 #include <sys/cred.h>
47 #include <sys/user.h>
48 #include <sys/uio.h>
49 #include <sys/vfs.h>
50 #include <sys/vnode.h>
51 #include <sys/pathname.h>
52 #include <sys/proc.h>
53 #include <sys/vtrace.h>
54 #include <sys/sysmacros.h>
55 #include <sys/debug.h>
56 #include <sys/dirent.h>
57 #include <c2/audit.h>
58 #include <sys/zone.h>
59 #include <sys/dnlc.h>
60 #include <sys/fs/snode.h>
61 
62 /* Controls whether paths are stored with vnodes. */
63 int vfs_vnode_path = 1;
64 
65 int
66 lookupname(
67 	char *fnamep,
68 	enum uio_seg seg,
69 	int followlink,
70 	vnode_t **dirvpp,
71 	vnode_t **compvpp)
72 {
73 	return (lookupnameatcred(fnamep, seg, followlink, dirvpp, compvpp, NULL,
74 	    CRED()));
75 }
76 
77 /*
78  * Lookup the user file name,
79  * Handle allocation and freeing of pathname buffer, return error.
80  */
81 int
82 lookupnameatcred(
83 	char *fnamep,			/* user pathname */
84 	enum uio_seg seg,		/* addr space that name is in */
85 	int followlink,			/* follow sym links */
86 	vnode_t **dirvpp,		/* ret for ptr to parent dir vnode */
87 	vnode_t **compvpp,		/* ret for ptr to component vnode */
88 	vnode_t *startvp,		/* start path search from vp */
89 	cred_t *cr)			/* credential */
90 {
91 	char namebuf[TYPICALMAXPATHLEN];
92 	struct pathname lookpn;
93 	int error;
94 
95 	error = pn_get_buf(fnamep, seg, &lookpn, namebuf, sizeof (namebuf));
96 	if (error == 0) {
97 		error = lookuppnatcred(&lookpn, NULL, followlink,
98 		    dirvpp, compvpp, startvp, cr);
99 	}
100 	if (error == ENAMETOOLONG) {
101 		/*
102 		 * This thread used a pathname > TYPICALMAXPATHLEN bytes long.
103 		 */
104 		if (error = pn_get(fnamep, seg, &lookpn))
105 			return (error);
106 		error = lookuppnatcred(&lookpn, NULL, followlink,
107 		    dirvpp, compvpp, startvp, cr);
108 		pn_free(&lookpn);
109 	}
110 
111 	return (error);
112 }
113 
114 int
115 lookupnameat(char *fnamep, enum uio_seg seg, int followlink,
116     vnode_t **dirvpp, vnode_t **compvpp, vnode_t *startvp)
117 {
118 	return (lookupnameatcred(fnamep, seg, followlink, dirvpp, compvpp,
119 	    startvp, CRED()));
120 }
121 
122 int
123 lookuppn(
124 	struct pathname *pnp,
125 	struct pathname *rpnp,
126 	int followlink,
127 	vnode_t **dirvpp,
128 	vnode_t **compvpp)
129 {
130 	return (lookuppnatcred(pnp, rpnp, followlink, dirvpp, compvpp, NULL,
131 	    CRED()));
132 }
133 
134 /*
135  * Lookup the user file name from a given vp, using a specific credential.
136  */
137 int
138 lookuppnatcred(
139 	struct pathname *pnp,		/* pathname to lookup */
140 	struct pathname *rpnp,		/* if non-NULL, return resolved path */
141 	int followlink,			/* (don't) follow sym links */
142 	vnode_t **dirvpp,		/* ptr for parent vnode */
143 	vnode_t **compvpp,		/* ptr for entry vnode */
144 	vnode_t *startvp,		/* start search from this vp */
145 	cred_t *cr)			/* user credential */
146 {
147 	vnode_t *vp;	/* current directory vp */
148 	vnode_t *rootvp;
149 	proc_t *p = curproc;
150 
151 	if (pnp->pn_pathlen == 0)
152 		return (ENOENT);
153 
154 	mutex_enter(&p->p_lock);	/* for u_rdir and u_cdir */
155 	if ((rootvp = PTOU(p)->u_rdir) == NULL)
156 		rootvp = rootdir;
157 	else if (rootvp != rootdir)	/* no need to VN_HOLD rootdir */
158 		VN_HOLD(rootvp);
159 
160 	if (pnp->pn_path[0] == '/') {
161 		vp = rootvp;
162 	} else {
163 		vp = (startvp == NULL) ? PTOU(p)->u_cdir : startvp;
164 	}
165 	VN_HOLD(vp);
166 	mutex_exit(&p->p_lock);
167 
168 	/*
169 	 * Skip over leading slashes
170 	 */
171 	if (pnp->pn_path[0] == '/') {
172 		do {
173 			pnp->pn_path++;
174 			pnp->pn_pathlen--;
175 		} while (pnp->pn_path[0] == '/');
176 	}
177 
178 	return (lookuppnvp(pnp, rpnp, followlink, dirvpp,
179 	    compvpp, rootvp, vp, cr));
180 }
181 
182 int
183 lookuppnat(struct pathname *pnp, struct pathname *rpnp,
184     int followlink, vnode_t **dirvpp, vnode_t **compvpp,
185     vnode_t *startvp)
186 {
187 	return (lookuppnatcred(pnp, rpnp, followlink, dirvpp, compvpp, startvp,
188 	    CRED()));
189 }
190 
191 /* Private flag to do our getcwd() dirty work */
192 #define	LOOKUP_CHECKREAD	0x10
193 #define	LOOKUP_MASK		(~LOOKUP_CHECKREAD)
194 
195 /*
196  * Starting at current directory, translate pathname pnp to end.
197  * Leave pathname of final component in pnp, return the vnode
198  * for the final component in *compvpp, and return the vnode
199  * for the parent of the final component in dirvpp.
200  *
201  * This is the central routine in pathname translation and handles
202  * multiple components in pathnames, separating them at /'s.  It also
203  * implements mounted file systems and processes symbolic links.
204  *
205  * vp is the vnode where the directory search should start.
206  *
207  * Reference counts: vp must be held prior to calling this function.  rootvp
208  * should only be held if rootvp != rootdir.
209  */
210 int
211 lookuppnvp(
212 	struct pathname *pnp,		/* pathname to lookup */
213 	struct pathname *rpnp,		/* if non-NULL, return resolved path */
214 	int flags,			/* follow symlinks */
215 	vnode_t **dirvpp,		/* ptr for parent vnode */
216 	vnode_t **compvpp,		/* ptr for entry vnode */
217 	vnode_t *rootvp,		/* rootvp */
218 	vnode_t *vp,			/* directory to start search at */
219 	cred_t *cr)			/* user's credential */
220 {
221 	vnode_t *cvp;	/* current component vp */
222 	char component[MAXNAMELEN];	/* buffer for component (incl null) */
223 	int error;
224 	int nlink;
225 	int lookup_flags;
226 	struct pathname presrvd; /* case preserved name */
227 	struct pathname *pp = NULL;
228 	vnode_t *startvp;
229 	vnode_t *zonevp = curproc->p_zone->zone_rootvp;		/* zone root */
230 	int must_be_directory = 0;
231 	boolean_t retry_with_kcred;
232 	uint32_t auditing = AU_AUDITING();
233 
234 	CPU_STATS_ADDQ(CPU, sys, namei, 1);
235 	nlink = 0;
236 	cvp = NULL;
237 	if (rpnp)
238 		rpnp->pn_pathlen = 0;
239 
240 	lookup_flags = dirvpp ? LOOKUP_DIR : 0;
241 	if (flags & FIGNORECASE) {
242 		lookup_flags |= FIGNORECASE;
243 		pn_alloc(&presrvd);
244 		pp = &presrvd;
245 	}
246 
247 	if (auditing)
248 		audit_anchorpath(pnp, vp == rootvp);
249 
250 	/*
251 	 * Eliminate any trailing slashes in the pathname.
252 	 * If there are any, we must follow all symlinks.
253 	 * Also, we must guarantee that the last component is a directory.
254 	 */
255 	if (pn_fixslash(pnp)) {
256 		flags |= FOLLOW;
257 		must_be_directory = 1;
258 	}
259 
260 	startvp = vp;
261 next:
262 	retry_with_kcred = B_FALSE;
263 
264 	/*
265 	 * Make sure we have a directory.
266 	 */
267 	if (vp->v_type != VDIR) {
268 		error = ENOTDIR;
269 		goto bad;
270 	}
271 
272 	if (rpnp && VN_CMP(vp, rootvp))
273 		(void) pn_set(rpnp, "/");
274 
275 	/*
276 	 * Process the next component of the pathname.
277 	 */
278 	if (error = pn_getcomponent(pnp, component)) {
279 		goto bad;
280 	}
281 
282 	/*
283 	 * Handle "..": two special cases.
284 	 * 1. If we're at the root directory (e.g. after chroot or
285 	 *    zone_enter) then change ".." to "." so we can't get
286 	 *    out of this subtree.
287 	 * 2. If this vnode is the root of a mounted file system,
288 	 *    then replace it with the vnode that was mounted on
289 	 *    so that we take the ".." in the other file system.
290 	 */
291 	if (component[0] == '.' && component[1] == '.' && component[2] == 0) {
292 checkforroot:
293 		if (VN_CMP(vp, rootvp) || VN_CMP(vp, zonevp)) {
294 			component[1] = '\0';
295 		} else if (vp->v_flag & VROOT) {
296 			vfs_t *vfsp;
297 			cvp = vp;
298 
299 			/*
300 			 * While we deal with the vfs pointer from the vnode
301 			 * the filesystem could have been forcefully unmounted
302 			 * and the vnode's v_vfsp could have been invalidated
303 			 * by VFS_UNMOUNT. Hence, we cache v_vfsp and use it
304 			 * with vfs_rlock_wait/vfs_unlock.
305 			 * It is safe to use the v_vfsp even it is freed by
306 			 * VFS_UNMOUNT because vfs_rlock_wait/vfs_unlock
307 			 * do not dereference v_vfsp. It is just used as a
308 			 * magic cookie.
309 			 * One more corner case here is the memory getting
310 			 * reused for another vfs structure. In this case
311 			 * lookuppnvp's vfs_rlock_wait will succeed, domount's
312 			 * vfs_lock will fail and domount will bail out with an
313 			 * error (EBUSY).
314 			 */
315 			vfsp = cvp->v_vfsp;
316 
317 			/*
318 			 * This lock is used to synchronize
319 			 * mounts/unmounts and lookups.
320 			 * Threads doing mounts/unmounts hold the
321 			 * writers version vfs_lock_wait().
322 			 */
323 
324 			vfs_rlock_wait(vfsp);
325 
326 			/*
327 			 * If this vnode is on a file system that
328 			 * has been forcibly unmounted,
329 			 * we can't proceed. Cancel this operation
330 			 * and return EIO.
331 			 *
332 			 * vfs_vnodecovered is NULL if unmounted.
333 			 * Currently, nfs uses VFS_UNMOUNTED to
334 			 * check if it's a forced-umount. Keep the
335 			 * same checking here as well even though it
336 			 * may not be needed.
337 			 */
338 			if (((vp = cvp->v_vfsp->vfs_vnodecovered) == NULL) ||
339 			    (cvp->v_vfsp->vfs_flag & VFS_UNMOUNTED)) {
340 				vfs_unlock(vfsp);
341 				VN_RELE(cvp);
342 				if (pp)
343 					pn_free(pp);
344 				return (EIO);
345 			}
346 			VN_HOLD(vp);
347 			vfs_unlock(vfsp);
348 			VN_RELE(cvp);
349 			cvp = NULL;
350 			/*
351 			 * Crossing mount points. For eg: We are doing
352 			 * a lookup of ".." for file systems root vnode
353 			 * mounted here, and VOP_LOOKUP() (with covered vnode)
354 			 * will be on underlying file systems mount point
355 			 * vnode. Set retry_with_kcred flag as we might end
356 			 * up doing VOP_LOOKUP() with kcred if required.
357 			 */
358 			retry_with_kcred = B_TRUE;
359 			goto checkforroot;
360 		}
361 	}
362 
363 	/*
364 	 * LOOKUP_CHECKREAD is a private flag used by vnodetopath() to indicate
365 	 * that we need to have read permission on every directory in the entire
366 	 * path.  This is used to ensure that a forward-lookup of a cached value
367 	 * has the same effect as a reverse-lookup when the cached value cannot
368 	 * be found.
369 	 */
370 	if ((flags & LOOKUP_CHECKREAD) &&
371 	    (error = VOP_ACCESS(vp, VREAD, 0, cr, NULL)) != 0)
372 		goto bad;
373 
374 	/*
375 	 * Perform a lookup in the current directory.
376 	 */
377 	error = VOP_LOOKUP(vp, component, &cvp, pnp, lookup_flags,
378 	    rootvp, cr, NULL, NULL, pp);
379 
380 	/*
381 	 * Retry with kcred - If crossing mount points & error is EACCES.
382 	 *
383 	 * If we are crossing mount points here and doing ".." lookup,
384 	 * VOP_LOOKUP() might fail if the underlying file systems
385 	 * mount point has no execute permission. In cases like these,
386 	 * we retry VOP_LOOKUP() by giving as much privilage as possible
387 	 * by passing kcred credentials.
388 	 *
389 	 * In case of hierarchical file systems, passing kcred still may
390 	 * or may not work.
391 	 * For eg: UFS FS --> Mount NFS FS --> Again mount UFS on some
392 	 *			directory inside NFS FS.
393 	 */
394 	if ((error == EACCES) && retry_with_kcred)
395 		error = VOP_LOOKUP(vp, component, &cvp, pnp, lookup_flags,
396 		    rootvp, zone_kcred(), NULL, NULL, pp);
397 
398 	if (error) {
399 		cvp = NULL;
400 		/*
401 		 * On error, return hard error if
402 		 * (a) we're not at the end of the pathname yet, or
403 		 * (b) the caller didn't want the parent directory, or
404 		 * (c) we failed for some reason other than a missing entry.
405 		 */
406 		if (pn_pathleft(pnp) || dirvpp == NULL || error != ENOENT)
407 			goto bad;
408 		if (auditing) {	/* directory access */
409 			if (error = audit_savepath(pnp, vp, vp, error, cr))
410 				goto bad_noaudit;
411 		}
412 
413 		pn_setlast(pnp);
414 		/*
415 		 * We inform the caller that the desired entry must be
416 		 * a directory by adding a '/' to the component name.
417 		 */
418 		if (must_be_directory && (error = pn_addslash(pnp)) != 0)
419 			goto bad;
420 		*dirvpp = vp;
421 		if (compvpp != NULL)
422 			*compvpp = NULL;
423 		if (rootvp != rootdir)
424 			VN_RELE(rootvp);
425 		if (pp)
426 			pn_free(pp);
427 		return (0);
428 	}
429 
430 	/*
431 	 * Traverse mount points.
432 	 * XXX why don't we need to hold a read lock here (call vn_vfsrlock)?
433 	 * What prevents a concurrent update to v_vfsmountedhere?
434 	 * 	Possible answer: if mounting, we might not see the mount
435 	 *	if it is concurrently coming into existence, but that's
436 	 *	really not much different from the thread running a bit slower.
437 	 *	If unmounting, we may get into traverse() when we shouldn't,
438 	 *	but traverse() will catch this case for us.
439 	 *	(For this to work, fetching v_vfsmountedhere had better
440 	 *	be atomic!)
441 	 */
442 	if (vn_mountedvfs(cvp) != NULL) {
443 		if ((error = traverse(&cvp)) != 0)
444 			goto bad;
445 	}
446 
447 	/*
448 	 * If we hit a symbolic link and there is more path to be
449 	 * translated or this operation does not wish to apply
450 	 * to a link, then place the contents of the link at the
451 	 * front of the remaining pathname.
452 	 */
453 	if (cvp->v_type == VLNK && ((flags & FOLLOW) || pn_pathleft(pnp))) {
454 		struct pathname linkpath;
455 
456 		if (cvp->v_vfsp->vfs_flag & VFS_NOFOLLOW) {
457 			/*
458 			 * mounts with the nofollow option set never resolve
459 			 * symlinks. They are treated as dangling symlink,
460 			 * always returning ENOENT
461 			 */
462 			error = ENOENT;
463 			goto bad;
464 		}
465 
466 		if (++nlink > MAXSYMLINKS) {
467 			error = ELOOP;
468 			goto bad;
469 		}
470 		pn_alloc(&linkpath);
471 		if (error = pn_getsymlink(cvp, &linkpath, cr)) {
472 			pn_free(&linkpath);
473 			goto bad;
474 		}
475 
476 		if (auditing)
477 			audit_symlink(pnp, &linkpath);
478 
479 		if (pn_pathleft(&linkpath) == 0)
480 			(void) pn_set(&linkpath, ".");
481 		error = pn_insert(pnp, &linkpath, strlen(component));
482 		pn_free(&linkpath);
483 		if (error)
484 			goto bad;
485 		VN_RELE(cvp);
486 		cvp = NULL;
487 		if (pnp->pn_pathlen == 0) {
488 			error = ENOENT;
489 			goto bad;
490 		}
491 		if (pnp->pn_path[0] == '/') {
492 			do {
493 				pnp->pn_path++;
494 				pnp->pn_pathlen--;
495 			} while (pnp->pn_path[0] == '/');
496 			VN_RELE(vp);
497 			vp = rootvp;
498 			VN_HOLD(vp);
499 		}
500 		if (auditing)
501 			audit_anchorpath(pnp, vp == rootvp);
502 		if (pn_fixslash(pnp)) {
503 			flags |= FOLLOW;
504 			must_be_directory = 1;
505 		}
506 		goto next;
507 	}
508 
509 	/*
510 	 * If rpnp is non-NULL, remember the resolved path name therein.
511 	 * Do not include "." components.  Collapse occurrences of
512 	 * "previous/..", so long as "previous" is not itself "..".
513 	 * Exhausting rpnp results in error ENAMETOOLONG.
514 	 */
515 	if (rpnp && strcmp(component, ".") != 0) {
516 		size_t len;
517 
518 		if (strcmp(component, "..") == 0 &&
519 		    rpnp->pn_pathlen != 0 &&
520 		    !((rpnp->pn_pathlen > 2 &&
521 		    strncmp(rpnp->pn_path+rpnp->pn_pathlen-3, "/..", 3) == 0) ||
522 		    (rpnp->pn_pathlen == 2 &&
523 		    strncmp(rpnp->pn_path, "..", 2) == 0))) {
524 			while (rpnp->pn_pathlen &&
525 			    rpnp->pn_path[rpnp->pn_pathlen-1] != '/')
526 				rpnp->pn_pathlen--;
527 			if (rpnp->pn_pathlen > 1)
528 				rpnp->pn_pathlen--;
529 			rpnp->pn_path[rpnp->pn_pathlen] = '\0';
530 		} else {
531 			if (rpnp->pn_pathlen != 0 &&
532 			    rpnp->pn_path[rpnp->pn_pathlen-1] != '/')
533 				rpnp->pn_path[rpnp->pn_pathlen++] = '/';
534 			if (flags & FIGNORECASE) {
535 				/*
536 				 * Return the case-preserved name
537 				 * within the resolved path.
538 				 */
539 				error = copystr(pp->pn_buf,
540 				    rpnp->pn_path + rpnp->pn_pathlen,
541 				    rpnp->pn_bufsize - rpnp->pn_pathlen, &len);
542 			} else {
543 				error = copystr(component,
544 				    rpnp->pn_path + rpnp->pn_pathlen,
545 				    rpnp->pn_bufsize - rpnp->pn_pathlen, &len);
546 			}
547 			if (error)	/* copystr() returns ENAMETOOLONG */
548 				goto bad;
549 			rpnp->pn_pathlen += (len - 1);
550 			ASSERT(rpnp->pn_bufsize > rpnp->pn_pathlen);
551 		}
552 	}
553 
554 	/*
555 	 * If no more components, return last directory (if wanted) and
556 	 * last component (if wanted).
557 	 */
558 	if (pn_pathleft(pnp) == 0) {
559 		/*
560 		 * If there was a trailing slash in the pathname,
561 		 * make sure the last component is a directory.
562 		 */
563 		if (must_be_directory && cvp->v_type != VDIR) {
564 			error = ENOTDIR;
565 			goto bad;
566 		}
567 		if (dirvpp != NULL) {
568 			/*
569 			 * Check that we have the real parent and not
570 			 * an alias of the last component.
571 			 */
572 			if (vn_compare(vp, cvp)) {
573 				if (auditing)
574 					(void) audit_savepath(pnp, cvp, vp,
575 					    EINVAL, cr);
576 				pn_setlast(pnp);
577 				VN_RELE(vp);
578 				VN_RELE(cvp);
579 				if (rootvp != rootdir)
580 					VN_RELE(rootvp);
581 				if (pp)
582 					pn_free(pp);
583 				return (EINVAL);
584 			}
585 			*dirvpp = vp;
586 		} else
587 			VN_RELE(vp);
588 		if (auditing)
589 			(void) audit_savepath(pnp, cvp, vp, 0, cr);
590 		if (pnp->pn_path == pnp->pn_buf)
591 			(void) pn_set(pnp, ".");
592 		else
593 			pn_setlast(pnp);
594 		if (rpnp) {
595 			if (VN_CMP(cvp, rootvp))
596 				(void) pn_set(rpnp, "/");
597 			else if (rpnp->pn_pathlen == 0)
598 				(void) pn_set(rpnp, ".");
599 		}
600 
601 		if (compvpp != NULL)
602 			*compvpp = cvp;
603 		else
604 			VN_RELE(cvp);
605 		if (rootvp != rootdir)
606 			VN_RELE(rootvp);
607 		if (pp)
608 			pn_free(pp);
609 		return (0);
610 	}
611 
612 	/*
613 	 * Skip over slashes from end of last component.
614 	 */
615 	while (pnp->pn_path[0] == '/') {
616 		pnp->pn_path++;
617 		pnp->pn_pathlen--;
618 	}
619 
620 	/*
621 	 * Searched through another level of directory:
622 	 * release previous directory handle and save new (result
623 	 * of lookup) as current directory.
624 	 */
625 	VN_RELE(vp);
626 	vp = cvp;
627 	cvp = NULL;
628 	goto next;
629 
630 bad:
631 	if (auditing)	/* reached end of path */
632 		(void) audit_savepath(pnp, cvp, vp, error, cr);
633 bad_noaudit:
634 	/*
635 	 * Error.  Release vnodes and return.
636 	 */
637 	if (cvp)
638 		VN_RELE(cvp);
639 	/*
640 	 * If the error was ESTALE and the current directory to look in
641 	 * was the root for this lookup, the root for a mounted file
642 	 * system, or the starting directory for lookups, then
643 	 * return ENOENT instead of ESTALE.  In this case, no recovery
644 	 * is possible by the higher level.  If ESTALE was returned for
645 	 * some intermediate directory along the path, then recovery
646 	 * is potentially possible and retrying from the higher level
647 	 * will either correct the situation by purging stale cache
648 	 * entries or eventually get back to the point where no recovery
649 	 * is possible.
650 	 */
651 	if (error == ESTALE &&
652 	    (VN_CMP(vp, rootvp) || (vp->v_flag & VROOT) || vp == startvp))
653 		error = ENOENT;
654 	VN_RELE(vp);
655 	if (rootvp != rootdir)
656 		VN_RELE(rootvp);
657 	if (pp)
658 		pn_free(pp);
659 	return (error);
660 }
661 
662 /*
663  * Traverse a mount point.  Routine accepts a vnode pointer as a reference
664  * parameter and performs the indirection, releasing the original vnode.
665  */
666 int
667 traverse(vnode_t **cvpp)
668 {
669 	int error = 0;
670 	vnode_t *cvp;
671 	vnode_t *tvp;
672 	vfs_t *vfsp;
673 
674 	cvp = *cvpp;
675 
676 	/*
677 	 * If this vnode is mounted on, then we transparently indirect
678 	 * to the vnode which is the root of the mounted file system.
679 	 * Before we do this we must check that an unmount is not in
680 	 * progress on this vnode.
681 	 */
682 
683 	for (;;) {
684 		/*
685 		 * Try to read lock the vnode.  If this fails because
686 		 * the vnode is already write locked, then check to
687 		 * see whether it is the current thread which locked
688 		 * the vnode.  If it is not, then read lock the vnode
689 		 * by waiting to acquire the lock.
690 		 *
691 		 * The code path in domount() is an example of support
692 		 * which needs to look up two pathnames and locks one
693 		 * of them in between the two lookups.
694 		 */
695 		error = vn_vfsrlock(cvp);
696 		if (error) {
697 			if (!vn_vfswlock_held(cvp))
698 				error = vn_vfsrlock_wait(cvp);
699 			if (error != 0) {
700 				/*
701 				 * lookuppn() expects a held vnode to be
702 				 * returned because it promptly calls
703 				 * VN_RELE after the error return
704 				 */
705 				*cvpp = cvp;
706 				return (error);
707 			}
708 		}
709 
710 		/*
711 		 * Reached the end of the mount chain?
712 		 */
713 		vfsp = vn_mountedvfs(cvp);
714 		if (vfsp == NULL) {
715 			vn_vfsunlock(cvp);
716 			break;
717 		}
718 
719 		/*
720 		 * The read lock must be held across the call to VFS_ROOT() to
721 		 * prevent a concurrent unmount from destroying the vfs.
722 		 */
723 		error = VFS_ROOT(vfsp, &tvp);
724 		vn_vfsunlock(cvp);
725 
726 		if (error)
727 			break;
728 
729 		VN_RELE(cvp);
730 
731 		cvp = tvp;
732 	}
733 
734 	*cvpp = cvp;
735 	return (error);
736 }
737 
738 /*
739  * Return the lowermost vnode if this is a mountpoint.
740  */
741 static vnode_t *
742 vn_under(vnode_t *vp)
743 {
744 	vnode_t *uvp;
745 	vfs_t *vfsp;
746 
747 	while (vp->v_flag & VROOT) {
748 
749 		vfsp = vp->v_vfsp;
750 		vfs_rlock_wait(vfsp);
751 		if ((uvp = vfsp->vfs_vnodecovered) == NULL ||
752 		    (vfsp->vfs_flag & VFS_UNMOUNTED)) {
753 			vfs_unlock(vfsp);
754 			break;
755 		}
756 		VN_HOLD(uvp);
757 		vfs_unlock(vfsp);
758 		VN_RELE(vp);
759 		vp = uvp;
760 	}
761 
762 	return (vp);
763 }
764 
765 static int
766 vnode_match(vnode_t *v1, vnode_t *v2, cred_t *cr)
767 {
768 	vattr_t	v1attr, v2attr;
769 
770 	/*
771 	 * If we have a device file, check to see if is a cloned open of the
772 	 * same device.  For self-cloning devices, the major numbers will match.
773 	 * For devices cloned through the 'clone' driver, the minor number of
774 	 * the source device will be the same as the major number of the cloned
775 	 * device.
776 	 */
777 	if ((v1->v_type == VCHR || v1->v_type == VBLK) &&
778 	    v1->v_type == v2->v_type) {
779 		if ((spec_is_selfclone(v1) || spec_is_selfclone(v2)) &&
780 		    getmajor(v1->v_rdev) == getmajor(v2->v_rdev))
781 			return (1);
782 
783 		if (spec_is_clone(v1) &&
784 		    getmajor(v1->v_rdev) == getminor(v2->v_rdev))
785 			return (1);
786 
787 		if (spec_is_clone(v2) &&
788 		    getmajor(v2->v_rdev) == getminor(v1->v_rdev))
789 			return (1);
790 	}
791 
792 	v1attr.va_mask = v2attr.va_mask = AT_TYPE;
793 
794 	/*
795 	 * This check for symbolic links handles the pseudo-symlinks in procfs.
796 	 * These particular links have v_type of VDIR, but the attributes have a
797 	 * type of VLNK.  We need to avoid these links because otherwise if we
798 	 * are currently in '/proc/self/fd', then '/proc/self/cwd' will compare
799 	 * as the same vnode.
800 	 */
801 	if (VOP_GETATTR(v1, &v1attr, 0, cr, NULL) != 0 ||
802 	    VOP_GETATTR(v2, &v2attr, 0, cr, NULL) != 0 ||
803 	    v1attr.va_type == VLNK || v2attr.va_type == VLNK)
804 		return (0);
805 
806 	v1attr.va_mask = v2attr.va_mask = AT_TYPE | AT_FSID | AT_NODEID;
807 
808 	if (VOP_GETATTR(v1, &v1attr, ATTR_REAL, cr, NULL) != 0 ||
809 	    VOP_GETATTR(v2, &v2attr, ATTR_REAL, cr, NULL) != 0)
810 		return (0);
811 
812 	return (v1attr.va_fsid == v2attr.va_fsid &&
813 	    v1attr.va_nodeid == v2attr.va_nodeid);
814 }
815 
816 
817 /*
818  * Find the entry in the directory corresponding to the target vnode.
819  */
820 int
821 dirfindvp(vnode_t *vrootp, vnode_t *dvp, vnode_t *tvp, cred_t *cr, char *dbuf,
822     size_t dlen, dirent64_t **rdp)
823 {
824 	size_t dbuflen;
825 	struct iovec iov;
826 	struct uio uio;
827 	int error;
828 	int eof;
829 	vnode_t *cmpvp;
830 	struct dirent64 *dp;
831 	pathname_t pnp;
832 
833 	ASSERT(dvp->v_type == VDIR);
834 
835 	/*
836 	 * This is necessary because of the strange semantics of VOP_LOOKUP().
837 	 */
838 	bzero(&pnp, sizeof (pnp));
839 
840 	eof = 0;
841 
842 	uio.uio_iov = &iov;
843 	uio.uio_iovcnt = 1;
844 	uio.uio_segflg = UIO_SYSSPACE;
845 	uio.uio_fmode = 0;
846 	uio.uio_extflg = UIO_COPY_CACHED;
847 	uio.uio_loffset = 0;
848 
849 	if ((error = VOP_ACCESS(dvp, VREAD, 0, cr, NULL)) != 0)
850 		return (error);
851 
852 	while (!eof) {
853 		uio.uio_resid = dlen;
854 		iov.iov_base = dbuf;
855 		iov.iov_len = dlen;
856 
857 		(void) VOP_RWLOCK(dvp, V_WRITELOCK_FALSE, NULL);
858 		error = VOP_READDIR(dvp, &uio, cr, &eof, NULL, 0);
859 		VOP_RWUNLOCK(dvp, V_WRITELOCK_FALSE, NULL);
860 
861 		dbuflen = dlen - uio.uio_resid;
862 
863 		if (error || dbuflen == 0)
864 			break;
865 
866 		dp = (dirent64_t *)dbuf;
867 		while ((intptr_t)dp < (intptr_t)dbuf + dbuflen) {
868 			/*
869 			 * Ignore '.' and '..' entries
870 			 */
871 			if (strcmp(dp->d_name, ".") == 0 ||
872 			    strcmp(dp->d_name, "..") == 0) {
873 				dp = (dirent64_t *)((intptr_t)dp +
874 				    dp->d_reclen);
875 				continue;
876 			}
877 
878 			error = VOP_LOOKUP(dvp, dp->d_name, &cmpvp, &pnp, 0,
879 			    vrootp, cr, NULL, NULL, NULL);
880 
881 			/*
882 			 * We only want to bail out if there was an error other
883 			 * than ENOENT.  Otherwise, it could be that someone
884 			 * just removed an entry since the readdir() call, and
885 			 * the entry we want is further on in the directory.
886 			 */
887 			if (error == 0) {
888 				if (vnode_match(tvp, cmpvp, cr)) {
889 					VN_RELE(cmpvp);
890 					*rdp = dp;
891 					return (0);
892 				}
893 
894 				VN_RELE(cmpvp);
895 			} else if (error != ENOENT) {
896 				return (error);
897 			}
898 
899 			dp = (dirent64_t *)((intptr_t)dp + dp->d_reclen);
900 		}
901 	}
902 
903 	/*
904 	 * Something strange has happened, this directory does not contain the
905 	 * specified vnode.  This should never happen in the normal case, since
906 	 * we ensured that dvp is the parent of vp.  This is possible in some
907 	 * rare conditions (races and the special .zfs directory).
908 	 */
909 	if (error == 0) {
910 		error = VOP_LOOKUP(dvp, ".zfs", &cmpvp, &pnp, 0, vrootp, cr,
911 		    NULL, NULL, NULL);
912 		if (error == 0) {
913 			if (vnode_match(tvp, cmpvp, cr)) {
914 				(void) strcpy(dp->d_name, ".zfs");
915 				dp->d_reclen = strlen(".zfs");
916 				dp->d_off = 2;
917 				dp->d_ino = 1;
918 				*rdp = dp;
919 			} else {
920 				error = ENOENT;
921 			}
922 			VN_RELE(cmpvp);
923 		}
924 	}
925 
926 	return (error);
927 }
928 
929 /*
930  * Given a global path (from rootdir), and a vnode that is the current root,
931  * return the portion of the path that is beneath the current root or NULL on
932  * failure.  The path MUST be a resolved path (no '..' entries or symlinks),
933  * otherwise this function will fail.
934  */
935 static char *
936 localpath(char *path, struct vnode *vrootp, cred_t *cr)
937 {
938 	vnode_t *vp;
939 	vnode_t *cvp;
940 	char component[MAXNAMELEN];
941 	char *ret = NULL;
942 	pathname_t pn;
943 
944 	/*
945 	 * We use vn_compare() instead of VN_CMP() in order to detect lofs
946 	 * mounts and stacked vnodes.
947 	 */
948 	if (vn_compare(vrootp, rootdir))
949 		return (path);
950 
951 	if (pn_get(path, UIO_SYSSPACE, &pn) != 0)
952 		return (NULL);
953 
954 	vp = rootdir;
955 	VN_HOLD(vp);
956 
957 	if (vn_ismntpt(vp) && traverse(&vp) != 0) {
958 		VN_RELE(vp);
959 		pn_free(&pn);
960 		return (NULL);
961 	}
962 
963 	while (pn_pathleft(&pn)) {
964 		pn_skipslash(&pn);
965 
966 		if (pn_getcomponent(&pn, component) != 0)
967 			break;
968 
969 		if (VOP_LOOKUP(vp, component, &cvp, &pn, 0, rootdir, cr,
970 		    NULL, NULL, NULL) != 0)
971 			break;
972 		VN_RELE(vp);
973 		vp = cvp;
974 
975 		if (vn_ismntpt(vp) && traverse(&vp) != 0)
976 			break;
977 
978 		if (vn_compare(vp, vrootp)) {
979 			ret = path + (pn.pn_path - pn.pn_buf);
980 			break;
981 		}
982 	}
983 
984 	VN_RELE(vp);
985 	pn_free(&pn);
986 
987 	return (ret);
988 }
989 
990 /*
991  * Given a directory, return the full, resolved path.  This looks up "..",
992  * searches for the given vnode in the parent, appends the component, etc.  It
993  * is used to implement vnodetopath() and getcwd() when the cached path fails.
994  */
995 static int
996 dirtopath(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen, int flags,
997     cred_t *cr)
998 {
999 	pathname_t pn, rpn, emptypn;
1000 	vnode_t *cmpvp, *pvp = NULL;
1001 	vnode_t *startvp = vp;
1002 	int err = 0, vprivs;
1003 	size_t complen;
1004 	char *dbuf;
1005 	dirent64_t *dp;
1006 	char		*bufloc;
1007 	size_t		dlen = DIRENT64_RECLEN(MAXPATHLEN);
1008 	refstr_t	*mntpt;
1009 
1010 	/* Operation only allowed on directories */
1011 	ASSERT(vp->v_type == VDIR);
1012 
1013 	/* We must have at least enough space for "/" */
1014 	if (buflen < 2)
1015 		return (ENAMETOOLONG);
1016 
1017 	/* Start at end of string with terminating null */
1018 	bufloc = &buf[buflen - 1];
1019 	*bufloc = '\0';
1020 
1021 	pn_alloc(&pn);
1022 	pn_alloc(&rpn);
1023 	dbuf = kmem_alloc(dlen, KM_SLEEP);
1024 	bzero(&emptypn, sizeof (emptypn));
1025 
1026 	/*
1027 	 * Begin with an additional reference on vp.  This will be decremented
1028 	 * during the loop.
1029 	 */
1030 	VN_HOLD(vp);
1031 
1032 	for (;;) {
1033 		/*
1034 		 * Return if we've reached the root.  If the buffer is empty,
1035 		 * return '/'.  We explicitly don't use vn_compare(), since it
1036 		 * compares the real vnodes.  A lofs mount of '/' would produce
1037 		 * incorrect results otherwise.
1038 		 */
1039 		if (VN_CMP(vrootp, vp)) {
1040 			if (*bufloc == '\0')
1041 				*--bufloc = '/';
1042 			break;
1043 		}
1044 
1045 		/*
1046 		 * If we've reached the VFS root, something has gone wrong.  We
1047 		 * should have reached the root in the above check.  The only
1048 		 * explantation is that 'vp' is not contained withing the given
1049 		 * root, in which case we return EPERM.
1050 		 */
1051 		if (VN_CMP(rootdir, vp)) {
1052 			err = EPERM;
1053 			goto out;
1054 		}
1055 
1056 		/*
1057 		 * Shortcut: see if this vnode is a mountpoint.  If so,
1058 		 * grab the path information from the vfs_t.
1059 		 */
1060 		if (vp->v_flag & VROOT) {
1061 
1062 			mntpt = vfs_getmntpoint(vp->v_vfsp);
1063 			if ((err = pn_set(&pn, (char *)refstr_value(mntpt)))
1064 			    == 0) {
1065 				refstr_rele(mntpt);
1066 				rpn.pn_path = rpn.pn_buf;
1067 
1068 				/*
1069 				 * Ensure the mountpoint still exists.
1070 				 */
1071 				VN_HOLD(vrootp);
1072 				if (vrootp != rootdir)
1073 					VN_HOLD(vrootp);
1074 				if (lookuppnvp(&pn, &rpn, flags, NULL,
1075 				    &cmpvp, vrootp, vrootp, cr) == 0) {
1076 
1077 					if (VN_CMP(vp, cmpvp)) {
1078 						VN_RELE(cmpvp);
1079 
1080 						complen = strlen(rpn.pn_path);
1081 						bufloc -= complen;
1082 						if (bufloc < buf) {
1083 							err = ERANGE;
1084 							goto out;
1085 						}
1086 						bcopy(rpn.pn_path, bufloc,
1087 						    complen);
1088 						break;
1089 					} else {
1090 						VN_RELE(cmpvp);
1091 					}
1092 				}
1093 			} else {
1094 				refstr_rele(mntpt);
1095 			}
1096 		}
1097 
1098 		/*
1099 		 * Shortcut: see if this vnode has correct v_path. If so,
1100 		 * we have the work done.
1101 		 */
1102 		mutex_enter(&vp->v_lock);
1103 		if (vp->v_path != NULL) {
1104 
1105 			if ((err = pn_set(&pn, vp->v_path)) == 0) {
1106 				mutex_exit(&vp->v_lock);
1107 				rpn.pn_path = rpn.pn_buf;
1108 
1109 				/*
1110 				 * Ensure the v_path pointing to correct vnode
1111 				 */
1112 				VN_HOLD(vrootp);
1113 				if (vrootp != rootdir)
1114 					VN_HOLD(vrootp);
1115 				if (lookuppnvp(&pn, &rpn, flags, NULL,
1116 				    &cmpvp, vrootp, vrootp, cr) == 0) {
1117 
1118 					if (VN_CMP(vp, cmpvp)) {
1119 						VN_RELE(cmpvp);
1120 
1121 						complen = strlen(rpn.pn_path);
1122 						bufloc -= complen;
1123 						if (bufloc < buf) {
1124 							err = ERANGE;
1125 							goto out;
1126 						}
1127 						bcopy(rpn.pn_path, bufloc,
1128 						    complen);
1129 						break;
1130 					} else {
1131 						VN_RELE(cmpvp);
1132 					}
1133 				}
1134 			} else {
1135 				mutex_exit(&vp->v_lock);
1136 			}
1137 		} else {
1138 			mutex_exit(&vp->v_lock);
1139 		}
1140 
1141 		/*
1142 		 * Shortcuts failed, search for this vnode in its parent.  If
1143 		 * this is a mountpoint, then get the vnode underneath.
1144 		 */
1145 		if (vp->v_flag & VROOT)
1146 			vp = vn_under(vp);
1147 		if ((err = VOP_LOOKUP(vp, "..", &pvp, &emptypn, 0, vrootp, cr,
1148 		    NULL, NULL, NULL)) != 0)
1149 			goto out;
1150 
1151 		/*
1152 		 * With extended attributes, it's possible for a directory to
1153 		 * have a parent that is a regular file.  Check for that here.
1154 		 */
1155 		if (pvp->v_type != VDIR) {
1156 			err = ENOTDIR;
1157 			goto out;
1158 		}
1159 
1160 		/*
1161 		 * If this is true, something strange has happened.  This is
1162 		 * only true if we are the root of a filesystem, which should
1163 		 * have been caught by the check above.
1164 		 */
1165 		if (VN_CMP(pvp, vp)) {
1166 			err = ENOENT;
1167 			goto out;
1168 		}
1169 
1170 		/*
1171 		 * Check if we have read and search privilege so, that
1172 		 * we can lookup the path in the directory
1173 		 */
1174 		vprivs = (flags & LOOKUP_CHECKREAD) ? VREAD | VEXEC : VEXEC;
1175 		if ((err = VOP_ACCESS(pvp, vprivs, 0, cr, NULL)) != 0) {
1176 			goto out;
1177 		}
1178 
1179 		/*
1180 		 * Search the parent directory for the entry corresponding to
1181 		 * this vnode.
1182 		 */
1183 		if ((err = dirfindvp(vrootp, pvp, vp, cr, dbuf, dlen, &dp))
1184 		    != 0)
1185 			goto out;
1186 		complen = strlen(dp->d_name);
1187 		bufloc -= complen;
1188 		if (bufloc <= buf) {
1189 			err = ENAMETOOLONG;
1190 			goto out;
1191 		}
1192 		bcopy(dp->d_name, bufloc, complen);
1193 
1194 		/* Prepend a slash to the current path.  */
1195 		*--bufloc = '/';
1196 
1197 		/* And continue with the next component */
1198 		VN_RELE(vp);
1199 		vp = pvp;
1200 		pvp = NULL;
1201 	}
1202 
1203 	/*
1204 	 * Place the path at the beginning of the buffer.
1205 	 */
1206 	if (bufloc != buf)
1207 		ovbcopy(bufloc, buf, buflen - (bufloc - buf));
1208 
1209 out:
1210 	/*
1211 	 * If the error was ESTALE and the current directory to look in
1212 	 * was the root for this lookup, the root for a mounted file
1213 	 * system, or the starting directory for lookups, then
1214 	 * return ENOENT instead of ESTALE.  In this case, no recovery
1215 	 * is possible by the higher level.  If ESTALE was returned for
1216 	 * some intermediate directory along the path, then recovery
1217 	 * is potentially possible and retrying from the higher level
1218 	 * will either correct the situation by purging stale cache
1219 	 * entries or eventually get back to the point where no recovery
1220 	 * is possible.
1221 	 */
1222 	if (err == ESTALE &&
1223 	    (VN_CMP(vp, vrootp) || (vp->v_flag & VROOT) || vp == startvp))
1224 		err = ENOENT;
1225 
1226 	kmem_free(dbuf, dlen);
1227 	VN_RELE(vp);
1228 	if (pvp)
1229 		VN_RELE(pvp);
1230 	pn_free(&pn);
1231 	pn_free(&rpn);
1232 
1233 	return (err);
1234 }
1235 
1236 /*
1237  * The additional flag, LOOKUP_CHECKREAD, is used to enforce artificial
1238  * constraints in order to be standards compliant.  For example, if we have
1239  * the cached path of '/foo/bar', and '/foo' has permissions 100 (execute
1240  * only), then we can legitimately look up the path to the current working
1241  * directory without needing read permission.  Existing standards tests,
1242  * however, assume that we are determining the path by repeatedly looking up
1243  * "..".  We need to keep this behavior in order to maintain backwards
1244  * compatibility.
1245  */
1246 static int
1247 vnodetopath_common(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen,
1248     cred_t *cr, int flags)
1249 {
1250 	pathname_t pn, rpn;
1251 	int ret;
1252 	vnode_t *compvp, *realvp;
1253 	proc_t *p = curproc;
1254 	int doclose = 0;
1255 
1256 	/*
1257 	 * If vrootp is NULL, get the root for curproc.  Callers with any other
1258 	 * requirements should pass in a different vrootp.
1259 	 */
1260 	if (vrootp == NULL) {
1261 		mutex_enter(&p->p_lock);
1262 		if ((vrootp = PTOU(p)->u_rdir) == NULL)
1263 			vrootp = rootdir;
1264 		VN_HOLD(vrootp);
1265 		mutex_exit(&p->p_lock);
1266 	} else {
1267 		VN_HOLD(vrootp);
1268 	}
1269 
1270 	/*
1271 	 * This is to get around an annoying artifact of the /proc filesystem,
1272 	 * which is the behavior of {cwd/root}.  Trying to resolve this path
1273 	 * will result in /proc/pid/cwd instead of whatever the real working
1274 	 * directory is.  We can't rely on VOP_REALVP(), since that will break
1275 	 * lofs.  The only difference between procfs and lofs is that opening
1276 	 * the file will return the underling vnode in the case of procfs.
1277 	 */
1278 	if (vp->v_type == VDIR && VOP_REALVP(vp, &realvp, NULL) == 0 &&
1279 	    realvp != vp) {
1280 		VN_HOLD(vp);
1281 		if (VOP_OPEN(&vp, FREAD, cr, NULL) == 0)
1282 			doclose = 1;
1283 		else
1284 			VN_RELE(vp);
1285 	}
1286 
1287 	pn_alloc(&pn);
1288 
1289 	/*
1290 	 * Check to see if we have a cached path in the vnode.
1291 	 */
1292 	mutex_enter(&vp->v_lock);
1293 	if (vp->v_path != NULL) {
1294 		(void) pn_set(&pn, vp->v_path);
1295 		mutex_exit(&vp->v_lock);
1296 
1297 		pn_alloc(&rpn);
1298 
1299 		/* We should only cache absolute paths */
1300 		ASSERT(pn.pn_buf[0] == '/');
1301 
1302 		/*
1303 		 * If we are in a zone or a chroot environment, then we have to
1304 		 * take additional steps, since the path to the root might not
1305 		 * be readable with the current credentials, even though the
1306 		 * process can legitmately access the file.  In this case, we
1307 		 * do the following:
1308 		 *
1309 		 * lookuppnvp() with all privileges to get the resolved path.
1310 		 * call localpath() to get the local portion of the path, and
1311 		 * continue as normal.
1312 		 *
1313 		 * If the the conversion to a local path fails, then we continue
1314 		 * as normal.  This is a heuristic to make process object file
1315 		 * paths available from within a zone.  Because lofs doesn't
1316 		 * support page operations, the vnode stored in the seg_t is
1317 		 * actually the underlying real vnode, not the lofs node itself.
1318 		 * Most of the time, the lofs path is the same as the underlying
1319 		 * vnode (for example, /usr/lib/libc.so.1).
1320 		 */
1321 		if (vrootp != rootdir) {
1322 			char *local = NULL;
1323 			VN_HOLD(rootdir);
1324 			if (lookuppnvp(&pn, &rpn, FOLLOW,
1325 			    NULL, &compvp, rootdir, rootdir, kcred) == 0) {
1326 				local = localpath(rpn.pn_path, vrootp,
1327 				    kcred);
1328 				VN_RELE(compvp);
1329 			}
1330 
1331 			/*
1332 			 * The original pn was changed through lookuppnvp().
1333 			 * Set it to local for next validation attempt.
1334 			 */
1335 			if (local) {
1336 				(void) pn_set(&pn, local);
1337 			} else {
1338 				goto notcached;
1339 			}
1340 		}
1341 
1342 		/*
1343 		 * We should have a local path at this point, so start the
1344 		 * search from the root of the current process.
1345 		 */
1346 		VN_HOLD(vrootp);
1347 		if (vrootp != rootdir)
1348 			VN_HOLD(vrootp);
1349 		ret = lookuppnvp(&pn, &rpn, FOLLOW | flags, NULL,
1350 		    &compvp, vrootp, vrootp, cr);
1351 		if (ret == 0) {
1352 			/*
1353 			 * Check to see if the returned vnode is the same as
1354 			 * the one we expect.  If not, give up.
1355 			 */
1356 			if (!vn_compare(vp, compvp) &&
1357 			    !vnode_match(vp, compvp, cr)) {
1358 				VN_RELE(compvp);
1359 				goto notcached;
1360 			}
1361 
1362 			VN_RELE(compvp);
1363 
1364 			/*
1365 			 * Return the result.
1366 			 */
1367 			if (buflen <= rpn.pn_pathlen)
1368 				goto notcached;
1369 
1370 			bcopy(rpn.pn_path, buf, rpn.pn_pathlen + 1);
1371 			pn_free(&pn);
1372 			pn_free(&rpn);
1373 			VN_RELE(vrootp);
1374 			if (doclose) {
1375 				(void) VOP_CLOSE(vp, FREAD, 1, 0, cr, NULL);
1376 				VN_RELE(vp);
1377 			}
1378 			return (0);
1379 		}
1380 
1381 notcached:
1382 		pn_free(&rpn);
1383 	} else {
1384 		mutex_exit(&vp->v_lock);
1385 	}
1386 
1387 	pn_free(&pn);
1388 
1389 	if (vp->v_type != VDIR) {
1390 		ret = ENOENT;
1391 	} else {
1392 		ret = dirtopath(vrootp, vp, buf, buflen, flags, cr);
1393 	}
1394 
1395 	VN_RELE(vrootp);
1396 	if (doclose) {
1397 		(void) VOP_CLOSE(vp, FREAD, 1, 0, cr, NULL);
1398 		VN_RELE(vp);
1399 	}
1400 
1401 	return (ret);
1402 }
1403 
1404 int
1405 vnodetopath(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen, cred_t *cr)
1406 {
1407 	return (vnodetopath_common(vrootp, vp, buf, buflen, cr, 0));
1408 }
1409 
1410 int
1411 dogetcwd(char *buf, size_t buflen)
1412 {
1413 	int ret;
1414 	vnode_t *vp;
1415 	vnode_t *compvp;
1416 	refstr_t *cwd, *oldcwd;
1417 	const char *value;
1418 	pathname_t rpnp, pnp;
1419 	proc_t *p = curproc;
1420 
1421 	/*
1422 	 * Check to see if there is a cached version of the cwd.  If so, lookup
1423 	 * the cached value and make sure it is the same vnode.
1424 	 */
1425 	mutex_enter(&p->p_lock);
1426 	if ((cwd = PTOU(p)->u_cwd) != NULL)
1427 		refstr_hold(cwd);
1428 	vp = PTOU(p)->u_cdir;
1429 	VN_HOLD(vp);
1430 	mutex_exit(&p->p_lock);
1431 
1432 	/*
1433 	 * Make sure we have permission to access the current directory.
1434 	 */
1435 	if ((ret = VOP_ACCESS(vp, VEXEC, 0, CRED(), NULL)) != 0) {
1436 		if (cwd != NULL)
1437 			refstr_rele(cwd);
1438 		VN_RELE(vp);
1439 		return (ret);
1440 	}
1441 
1442 	if (cwd) {
1443 		value = refstr_value(cwd);
1444 		if ((ret = pn_get((char *)value, UIO_SYSSPACE, &pnp)) != 0) {
1445 			refstr_rele(cwd);
1446 			VN_RELE(vp);
1447 			return (ret);
1448 		}
1449 
1450 		pn_alloc(&rpnp);
1451 
1452 		if (lookuppn(&pnp, &rpnp, NO_FOLLOW, NULL, &compvp) == 0) {
1453 
1454 			if (VN_CMP(vp, compvp) &&
1455 			    strcmp(value, rpnp.pn_path) == 0) {
1456 				VN_RELE(compvp);
1457 				VN_RELE(vp);
1458 				pn_free(&pnp);
1459 				pn_free(&rpnp);
1460 				if (strlen(value) + 1 > buflen) {
1461 					refstr_rele(cwd);
1462 					return (ENAMETOOLONG);
1463 				}
1464 				bcopy(value, buf, strlen(value) + 1);
1465 				refstr_rele(cwd);
1466 				return (0);
1467 			}
1468 
1469 			VN_RELE(compvp);
1470 		}
1471 
1472 		pn_free(&rpnp);
1473 		pn_free(&pnp);
1474 
1475 		refstr_rele(cwd);
1476 	}
1477 
1478 	ret = vnodetopath_common(NULL, vp, buf, buflen, CRED(),
1479 	    LOOKUP_CHECKREAD);
1480 
1481 	VN_RELE(vp);
1482 
1483 	/*
1484 	 * Store the new cwd and replace the existing cached copy.
1485 	 */
1486 	if (ret == 0)
1487 		cwd = refstr_alloc(buf);
1488 	else
1489 		cwd = NULL;
1490 
1491 	mutex_enter(&p->p_lock);
1492 	oldcwd = PTOU(p)->u_cwd;
1493 	PTOU(p)->u_cwd = cwd;
1494 	mutex_exit(&p->p_lock);
1495 
1496 	if (oldcwd)
1497 		refstr_rele(oldcwd);
1498 
1499 	return (ret);
1500 }
1501