xref: /illumos-gate/usr/src/uts/common/fs/lookup.c (revision 09a48d4ca0ddda4ad26cc885769745870d989baf)
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 (c) 1988, 2010, Oracle and/or its affiliates. All rights reserved.
24  */
25 
26 /*	Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T	*/
27 /*	  All Rights Reserved  	*/
28 
29 /*
30  * University Copyright- Copyright (c) 1982, 1986, 1988
31  * The Regents of the University of California
32  * All Rights Reserved
33  *
34  * University Acknowledgment- Portions of this document are derived from
35  * software developed by the University of California, Berkeley, and its
36  * contributors.
37  */
38 
39 #include <sys/types.h>
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/cpuvar.h>
43 #include <sys/errno.h>
44 #include <sys/cred.h>
45 #include <sys/user.h>
46 #include <sys/uio.h>
47 #include <sys/vfs.h>
48 #include <sys/vnode.h>
49 #include <sys/pathname.h>
50 #include <sys/proc.h>
51 #include <sys/vtrace.h>
52 #include <sys/sysmacros.h>
53 #include <sys/debug.h>
54 #include <sys/dirent.h>
55 #include <c2/audit.h>
56 #include <sys/zone.h>
57 #include <sys/dnlc.h>
58 #include <sys/fs/snode.h>
59 
60 /* Controls whether paths are stored with vnodes. */
61 int vfs_vnode_path = 1;
62 
63 int
64 lookupname(
65 	char *fnamep,
66 	enum uio_seg seg,
67 	int followlink,
68 	vnode_t **dirvpp,
69 	vnode_t **compvpp)
70 {
71 	return (lookupnameatcred(fnamep, seg, followlink, dirvpp, compvpp, NULL,
72 	    CRED()));
73 }
74 
75 /*
76  * Lookup the user file name,
77  * Handle allocation and freeing of pathname buffer, return error.
78  */
79 int
80 lookupnameatcred(
81 	char *fnamep,			/* user pathname */
82 	enum uio_seg seg,		/* addr space that name is in */
83 	int followlink,			/* follow sym links */
84 	vnode_t **dirvpp,		/* ret for ptr to parent dir vnode */
85 	vnode_t **compvpp,		/* ret for ptr to component vnode */
86 	vnode_t *startvp,		/* start path search from vp */
87 	cred_t *cr)			/* credential */
88 {
89 	char namebuf[TYPICALMAXPATHLEN];
90 	struct pathname lookpn;
91 	int error;
92 
93 	error = pn_get_buf(fnamep, seg, &lookpn, namebuf, sizeof (namebuf));
94 	if (error == 0) {
95 		error = lookuppnatcred(&lookpn, NULL, followlink,
96 		    dirvpp, compvpp, startvp, cr);
97 	}
98 	if (error == ENAMETOOLONG) {
99 		/*
100 		 * This thread used a pathname > TYPICALMAXPATHLEN bytes long.
101 		 */
102 		if (error = pn_get(fnamep, seg, &lookpn))
103 			return (error);
104 		error = lookuppnatcred(&lookpn, NULL, followlink,
105 		    dirvpp, compvpp, startvp, cr);
106 		pn_free(&lookpn);
107 	}
108 
109 	return (error);
110 }
111 
112 int
113 lookupnameat(char *fnamep, enum uio_seg seg, int followlink,
114     vnode_t **dirvpp, vnode_t **compvpp, vnode_t *startvp)
115 {
116 	return (lookupnameatcred(fnamep, seg, followlink, dirvpp, compvpp,
117 	    startvp, CRED()));
118 }
119 
120 int
121 lookuppn(
122 	struct pathname *pnp,
123 	struct pathname *rpnp,
124 	int followlink,
125 	vnode_t **dirvpp,
126 	vnode_t **compvpp)
127 {
128 	return (lookuppnatcred(pnp, rpnp, followlink, dirvpp, compvpp, NULL,
129 	    CRED()));
130 }
131 
132 /*
133  * Lookup the user file name from a given vp, using a specific credential.
134  */
135 int
136 lookuppnatcred(
137 	struct pathname *pnp,		/* pathname to lookup */
138 	struct pathname *rpnp,		/* if non-NULL, return resolved path */
139 	int followlink,			/* (don't) follow sym links */
140 	vnode_t **dirvpp,		/* ptr for parent vnode */
141 	vnode_t **compvpp,		/* ptr for entry vnode */
142 	vnode_t *startvp,		/* start search from this vp */
143 	cred_t *cr)			/* user credential */
144 {
145 	vnode_t *vp;	/* current directory vp */
146 	vnode_t *rootvp;
147 	proc_t *p = curproc;
148 
149 	if (pnp->pn_pathlen == 0)
150 		return (ENOENT);
151 
152 	mutex_enter(&p->p_lock);	/* for u_rdir and u_cdir */
153 	if ((rootvp = PTOU(p)->u_rdir) == NULL)
154 		rootvp = rootdir;
155 	else if (rootvp != rootdir)	/* no need to VN_HOLD rootdir */
156 		VN_HOLD(rootvp);
157 
158 	if (pnp->pn_path[0] == '/') {
159 		vp = rootvp;
160 	} else {
161 		vp = (startvp == NULL) ? PTOU(p)->u_cdir : startvp;
162 	}
163 	VN_HOLD(vp);
164 	mutex_exit(&p->p_lock);
165 
166 	/*
167 	 * Skip over leading slashes
168 	 */
169 	if (pnp->pn_path[0] == '/') {
170 		do {
171 			pnp->pn_path++;
172 			pnp->pn_pathlen--;
173 		} while (pnp->pn_path[0] == '/');
174 	}
175 
176 	return (lookuppnvp(pnp, rpnp, followlink, dirvpp,
177 	    compvpp, rootvp, vp, cr));
178 }
179 
180 int
181 lookuppnat(struct pathname *pnp, struct pathname *rpnp,
182     int followlink, vnode_t **dirvpp, vnode_t **compvpp,
183     vnode_t *startvp)
184 {
185 	return (lookuppnatcred(pnp, rpnp, followlink, dirvpp, compvpp, startvp,
186 	    CRED()));
187 }
188 
189 /* Private flag to do our getcwd() dirty work */
190 #define	LOOKUP_CHECKREAD	0x10
191 #define	LOOKUP_MASK		(~LOOKUP_CHECKREAD)
192 
193 /*
194  * Starting at current directory, translate pathname pnp to end.
195  * Leave pathname of final component in pnp, return the vnode
196  * for the final component in *compvpp, and return the vnode
197  * for the parent of the final component in dirvpp.
198  *
199  * This is the central routine in pathname translation and handles
200  * multiple components in pathnames, separating them at /'s.  It also
201  * implements mounted file systems and processes symbolic links.
202  *
203  * vp is the vnode where the directory search should start.
204  *
205  * Reference counts: vp must be held prior to calling this function.  rootvp
206  * should only be held if rootvp != rootdir.
207  */
208 int
209 lookuppnvp(
210 	struct pathname *pnp,		/* pathname to lookup */
211 	struct pathname *rpnp,		/* if non-NULL, return resolved path */
212 	int flags,			/* follow symlinks */
213 	vnode_t **dirvpp,		/* ptr for parent vnode */
214 	vnode_t **compvpp,		/* ptr for entry vnode */
215 	vnode_t *rootvp,		/* rootvp */
216 	vnode_t *vp,			/* directory to start search at */
217 	cred_t *cr)			/* user's credential */
218 {
219 	vnode_t *cvp;	/* current component vp */
220 	vnode_t *tvp;	/* addressable temp ptr */
221 	char component[MAXNAMELEN];	/* buffer for component (incl null) */
222 	int error;
223 	int nlink;
224 	int lookup_flags;
225 	struct pathname presrvd; /* case preserved name */
226 	struct pathname *pp = NULL;
227 	vnode_t *startvp;
228 	vnode_t *zonevp = curproc->p_zone->zone_rootvp;		/* zone root */
229 	int must_be_directory = 0;
230 	boolean_t retry_with_kcred;
231 	uint32_t auditing = AU_AUDITING();
232 
233 	CPU_STATS_ADDQ(CPU, sys, namei, 1);
234 	nlink = 0;
235 	cvp = NULL;
236 	if (rpnp)
237 		rpnp->pn_pathlen = 0;
238 
239 	lookup_flags = dirvpp ? LOOKUP_DIR : 0;
240 	if (flags & FIGNORECASE) {
241 		lookup_flags |= FIGNORECASE;
242 		pn_alloc(&presrvd);
243 		pp = &presrvd;
244 	}
245 
246 	if (auditing)
247 		audit_anchorpath(pnp, vp == rootvp);
248 
249 	/*
250 	 * Eliminate any trailing slashes in the pathname.
251 	 * If there are any, we must follow all symlinks.
252 	 * Also, we must guarantee that the last component is a directory.
253 	 */
254 	if (pn_fixslash(pnp)) {
255 		flags |= FOLLOW;
256 		must_be_directory = 1;
257 	}
258 
259 	startvp = vp;
260 next:
261 	retry_with_kcred = B_FALSE;
262 
263 	/*
264 	 * Make sure we have a directory.
265 	 */
266 	if (vp->v_type != VDIR) {
267 		error = ENOTDIR;
268 		goto bad;
269 	}
270 
271 	if (rpnp && VN_CMP(vp, rootvp))
272 		(void) pn_set(rpnp, "/");
273 
274 	/*
275 	 * Process the next component of the pathname.
276 	 */
277 	if (error = pn_getcomponent(pnp, component)) {
278 		goto bad;
279 	}
280 
281 	/*
282 	 * Handle "..": two special cases.
283 	 * 1. If we're at the root directory (e.g. after chroot or
284 	 *    zone_enter) then change ".." to "." so we can't get
285 	 *    out of this subtree.
286 	 * 2. If this vnode is the root of a mounted file system,
287 	 *    then replace it with the vnode that was mounted on
288 	 *    so that we take the ".." in the other file system.
289 	 */
290 	if (component[0] == '.' && component[1] == '.' && component[2] == 0) {
291 checkforroot:
292 		if (VN_CMP(vp, rootvp) || VN_CMP(vp, zonevp)) {
293 			component[1] = '\0';
294 		} else if (vp->v_flag & VROOT) {
295 			vfs_t *vfsp;
296 			cvp = vp;
297 
298 			/*
299 			 * While we deal with the vfs pointer from the vnode
300 			 * the filesystem could have been forcefully unmounted
301 			 * and the vnode's v_vfsp could have been invalidated
302 			 * by VFS_UNMOUNT. Hence, we cache v_vfsp and use it
303 			 * with vfs_rlock_wait/vfs_unlock.
304 			 * It is safe to use the v_vfsp even it is freed by
305 			 * VFS_UNMOUNT because vfs_rlock_wait/vfs_unlock
306 			 * do not dereference v_vfsp. It is just used as a
307 			 * magic cookie.
308 			 * One more corner case here is the memory getting
309 			 * reused for another vfs structure. In this case
310 			 * lookuppnvp's vfs_rlock_wait will succeed, domount's
311 			 * vfs_lock will fail and domount will bail out with an
312 			 * error (EBUSY).
313 			 */
314 			vfsp = cvp->v_vfsp;
315 
316 			/*
317 			 * This lock is used to synchronize
318 			 * mounts/unmounts and lookups.
319 			 * Threads doing mounts/unmounts hold the
320 			 * writers version vfs_lock_wait().
321 			 */
322 
323 			vfs_rlock_wait(vfsp);
324 
325 			/*
326 			 * If this vnode is on a file system that
327 			 * has been forcibly unmounted,
328 			 * we can't proceed. Cancel this operation
329 			 * and return EIO.
330 			 *
331 			 * vfs_vnodecovered is NULL if unmounted.
332 			 * Currently, nfs uses VFS_UNMOUNTED to
333 			 * check if it's a forced-umount. Keep the
334 			 * same checking here as well even though it
335 			 * may not be needed.
336 			 */
337 			if (((vp = cvp->v_vfsp->vfs_vnodecovered) == NULL) ||
338 			    (cvp->v_vfsp->vfs_flag & VFS_UNMOUNTED)) {
339 				vfs_unlock(vfsp);
340 				VN_RELE(cvp);
341 				if (pp)
342 					pn_free(pp);
343 				return (EIO);
344 			}
345 			VN_HOLD(vp);
346 			vfs_unlock(vfsp);
347 			VN_RELE(cvp);
348 			cvp = NULL;
349 			/*
350 			 * Crossing mount points. For eg: We are doing
351 			 * a lookup of ".." for file systems root vnode
352 			 * mounted here, and VOP_LOOKUP() (with covered vnode)
353 			 * will be on underlying file systems mount point
354 			 * vnode. Set retry_with_kcred flag as we might end
355 			 * up doing VOP_LOOKUP() with kcred if required.
356 			 */
357 			retry_with_kcred = B_TRUE;
358 			goto checkforroot;
359 		}
360 	}
361 
362 	/*
363 	 * LOOKUP_CHECKREAD is a private flag used by vnodetopath() to indicate
364 	 * that we need to have read permission on every directory in the entire
365 	 * path.  This is used to ensure that a forward-lookup of a cached value
366 	 * has the same effect as a reverse-lookup when the cached value cannot
367 	 * be found.
368 	 */
369 	if ((flags & LOOKUP_CHECKREAD) &&
370 	    (error = VOP_ACCESS(vp, VREAD, 0, cr, NULL)) != 0)
371 		goto bad;
372 
373 	/*
374 	 * Perform a lookup in the current directory.
375 	 */
376 	error = VOP_LOOKUP(vp, component, &tvp, pnp, lookup_flags,
377 	    rootvp, cr, NULL, NULL, pp);
378 
379 	/*
380 	 * Retry with kcred - If crossing mount points & error is EACCES.
381 	 *
382 	 * If we are crossing mount points here and doing ".." lookup,
383 	 * VOP_LOOKUP() might fail if the underlying file systems
384 	 * mount point has no execute permission. In cases like these,
385 	 * we retry VOP_LOOKUP() by giving as much privilage as possible
386 	 * by passing kcred credentials.
387 	 *
388 	 * In case of hierarchical file systems, passing kcred still may
389 	 * or may not work.
390 	 * For eg: UFS FS --> Mount NFS FS --> Again mount UFS on some
391 	 *			directory inside NFS FS.
392 	 */
393 	if ((error == EACCES) && retry_with_kcred)
394 		error = VOP_LOOKUP(vp, component, &tvp, pnp, lookup_flags,
395 		    rootvp, zone_kcred(), NULL, NULL, pp);
396 
397 	cvp = tvp;
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 		tvp = cvp;
444 		if ((error = traverse(&tvp)) != 0) {
445 			/*
446 			 * It is required to assign cvp here, because
447 			 * traverse() will return a held vnode which
448 			 * may different than the vnode that was passed
449 			 * in (even in the error case).  If traverse()
450 			 * changes the vnode it releases the original,
451 			 * and holds the new one.
452 			 */
453 			cvp = tvp;
454 			goto bad;
455 		}
456 		cvp = tvp;
457 	}
458 
459 	/*
460 	 * If we hit a symbolic link and there is more path to be
461 	 * translated or this operation does not wish to apply
462 	 * to a link, then place the contents of the link at the
463 	 * front of the remaining pathname.
464 	 */
465 	if (cvp->v_type == VLNK && ((flags & FOLLOW) || pn_pathleft(pnp))) {
466 		struct pathname linkpath;
467 
468 		if (++nlink > MAXSYMLINKS) {
469 			error = ELOOP;
470 			goto bad;
471 		}
472 		pn_alloc(&linkpath);
473 		if (error = pn_getsymlink(cvp, &linkpath, cr)) {
474 			pn_free(&linkpath);
475 			goto bad;
476 		}
477 
478 		if (auditing)
479 			audit_symlink(pnp, &linkpath);
480 
481 		if (pn_pathleft(&linkpath) == 0)
482 			(void) pn_set(&linkpath, ".");
483 		error = pn_insert(pnp, &linkpath, strlen(component));
484 		pn_free(&linkpath);
485 		if (error)
486 			goto bad;
487 		VN_RELE(cvp);
488 		cvp = NULL;
489 		if (pnp->pn_pathlen == 0) {
490 			error = ENOENT;
491 			goto bad;
492 		}
493 		if (pnp->pn_path[0] == '/') {
494 			do {
495 				pnp->pn_path++;
496 				pnp->pn_pathlen--;
497 			} while (pnp->pn_path[0] == '/');
498 			VN_RELE(vp);
499 			vp = rootvp;
500 			VN_HOLD(vp);
501 		}
502 		if (auditing)
503 			audit_anchorpath(pnp, vp == rootvp);
504 		if (pn_fixslash(pnp)) {
505 			flags |= FOLLOW;
506 			must_be_directory = 1;
507 		}
508 		goto next;
509 	}
510 
511 	/*
512 	 * If rpnp is non-NULL, remember the resolved path name therein.
513 	 * Do not include "." components.  Collapse occurrences of
514 	 * "previous/..", so long as "previous" is not itself "..".
515 	 * Exhausting rpnp results in error ENAMETOOLONG.
516 	 */
517 	if (rpnp && strcmp(component, ".") != 0) {
518 		size_t len;
519 
520 		if (strcmp(component, "..") == 0 &&
521 		    rpnp->pn_pathlen != 0 &&
522 		    !((rpnp->pn_pathlen > 2 &&
523 		    strncmp(rpnp->pn_path+rpnp->pn_pathlen-3, "/..", 3) == 0) ||
524 		    (rpnp->pn_pathlen == 2 &&
525 		    strncmp(rpnp->pn_path, "..", 2) == 0))) {
526 			while (rpnp->pn_pathlen &&
527 			    rpnp->pn_path[rpnp->pn_pathlen-1] != '/')
528 				rpnp->pn_pathlen--;
529 			if (rpnp->pn_pathlen > 1)
530 				rpnp->pn_pathlen--;
531 			rpnp->pn_path[rpnp->pn_pathlen] = '\0';
532 		} else {
533 			if (rpnp->pn_pathlen != 0 &&
534 			    rpnp->pn_path[rpnp->pn_pathlen-1] != '/')
535 				rpnp->pn_path[rpnp->pn_pathlen++] = '/';
536 			if (flags & FIGNORECASE) {
537 				/*
538 				 * Return the case-preserved name
539 				 * within the resolved path.
540 				 */
541 				error = copystr(pp->pn_buf,
542 				    rpnp->pn_path + rpnp->pn_pathlen,
543 				    rpnp->pn_bufsize - rpnp->pn_pathlen, &len);
544 			} else {
545 				error = copystr(component,
546 				    rpnp->pn_path + rpnp->pn_pathlen,
547 				    rpnp->pn_bufsize - rpnp->pn_pathlen, &len);
548 			}
549 			if (error)	/* copystr() returns ENAMETOOLONG */
550 				goto bad;
551 			rpnp->pn_pathlen += (len - 1);
552 			ASSERT(rpnp->pn_bufsize > rpnp->pn_pathlen);
553 		}
554 	}
555 
556 	/*
557 	 * If no more components, return last directory (if wanted) and
558 	 * last component (if wanted).
559 	 */
560 	if (pn_pathleft(pnp) == 0) {
561 		/*
562 		 * If there was a trailing slash in the pathname,
563 		 * make sure the last component is a directory.
564 		 */
565 		if (must_be_directory && cvp->v_type != VDIR) {
566 			error = ENOTDIR;
567 			goto bad;
568 		}
569 		if (dirvpp != NULL) {
570 			/*
571 			 * Check that we have the real parent and not
572 			 * an alias of the last component.
573 			 */
574 			if (vn_compare(vp, cvp)) {
575 				if (auditing)
576 					(void) audit_savepath(pnp, cvp, vp,
577 					    EINVAL, cr);
578 				pn_setlast(pnp);
579 				VN_RELE(vp);
580 				VN_RELE(cvp);
581 				if (rootvp != rootdir)
582 					VN_RELE(rootvp);
583 				if (pp)
584 					pn_free(pp);
585 				return (EINVAL);
586 			}
587 			*dirvpp = vp;
588 		} else
589 			VN_RELE(vp);
590 		if (auditing)
591 			(void) audit_savepath(pnp, cvp, vp, 0, cr);
592 		if (pnp->pn_path == pnp->pn_buf)
593 			(void) pn_set(pnp, ".");
594 		else
595 			pn_setlast(pnp);
596 		if (rpnp) {
597 			if (VN_CMP(cvp, rootvp))
598 				(void) pn_set(rpnp, "/");
599 			else if (rpnp->pn_pathlen == 0)
600 				(void) pn_set(rpnp, ".");
601 		}
602 
603 		if (compvpp != NULL)
604 			*compvpp = cvp;
605 		else
606 			VN_RELE(cvp);
607 		if (rootvp != rootdir)
608 			VN_RELE(rootvp);
609 		if (pp)
610 			pn_free(pp);
611 		return (0);
612 	}
613 
614 	/*
615 	 * Skip over slashes from end of last component.
616 	 */
617 	while (pnp->pn_path[0] == '/') {
618 		pnp->pn_path++;
619 		pnp->pn_pathlen--;
620 	}
621 
622 	/*
623 	 * Searched through another level of directory:
624 	 * release previous directory handle and save new (result
625 	 * of lookup) as current directory.
626 	 */
627 	VN_RELE(vp);
628 	vp = cvp;
629 	cvp = NULL;
630 	goto next;
631 
632 bad:
633 	if (auditing)	/* reached end of path */
634 		(void) audit_savepath(pnp, cvp, vp, error, cr);
635 bad_noaudit:
636 	/*
637 	 * Error.  Release vnodes and return.
638 	 */
639 	if (cvp)
640 		VN_RELE(cvp);
641 	/*
642 	 * If the error was ESTALE and the current directory to look in
643 	 * was the root for this lookup, the root for a mounted file
644 	 * system, or the starting directory for lookups, then
645 	 * return ENOENT instead of ESTALE.  In this case, no recovery
646 	 * is possible by the higher level.  If ESTALE was returned for
647 	 * some intermediate directory along the path, then recovery
648 	 * is potentially possible and retrying from the higher level
649 	 * will either correct the situation by purging stale cache
650 	 * entries or eventually get back to the point where no recovery
651 	 * is possible.
652 	 */
653 	if (error == ESTALE &&
654 	    (VN_CMP(vp, rootvp) || (vp->v_flag & VROOT) || vp == startvp))
655 		error = ENOENT;
656 	VN_RELE(vp);
657 	if (rootvp != rootdir)
658 		VN_RELE(rootvp);
659 	if (pp)
660 		pn_free(pp);
661 	return (error);
662 }
663 
664 /*
665  * Traverse a mount point.  Routine accepts a vnode pointer as a reference
666  * parameter and performs the indirection, releasing the original vnode.
667  */
668 int
669 traverse(vnode_t **cvpp)
670 {
671 	int error = 0;
672 	vnode_t *cvp;
673 	vnode_t *tvp;
674 	vfs_t *vfsp;
675 
676 	cvp = *cvpp;
677 
678 	/*
679 	 * If this vnode is mounted on, then we transparently indirect
680 	 * to the vnode which is the root of the mounted file system.
681 	 * Before we do this we must check that an unmount is not in
682 	 * progress on this vnode.
683 	 */
684 
685 	for (;;) {
686 		/*
687 		 * Try to read lock the vnode.  If this fails because
688 		 * the vnode is already write locked, then check to
689 		 * see whether it is the current thread which locked
690 		 * the vnode.  If it is not, then read lock the vnode
691 		 * by waiting to acquire the lock.
692 		 *
693 		 * The code path in domount() is an example of support
694 		 * which needs to look up two pathnames and locks one
695 		 * of them in between the two lookups.
696 		 */
697 		error = vn_vfsrlock(cvp);
698 		if (error) {
699 			if (!vn_vfswlock_held(cvp))
700 				error = vn_vfsrlock_wait(cvp);
701 			if (error != 0) {
702 				/*
703 				 * lookuppn() expects a held vnode to be
704 				 * returned because it promptly calls
705 				 * VN_RELE after the error return
706 				 */
707 				*cvpp = cvp;
708 				return (error);
709 			}
710 		}
711 
712 		/*
713 		 * Reached the end of the mount chain?
714 		 */
715 		vfsp = vn_mountedvfs(cvp);
716 		if (vfsp == NULL) {
717 			vn_vfsunlock(cvp);
718 			break;
719 		}
720 
721 		/*
722 		 * The read lock must be held across the call to VFS_ROOT() to
723 		 * prevent a concurrent unmount from destroying the vfs.
724 		 */
725 		error = VFS_ROOT(vfsp, &tvp);
726 		vn_vfsunlock(cvp);
727 
728 		if (error)
729 			break;
730 
731 		VN_RELE(cvp);
732 
733 		cvp = tvp;
734 	}
735 
736 	*cvpp = cvp;
737 	return (error);
738 }
739 
740 /*
741  * Return the lowermost vnode if this is a mountpoint.
742  */
743 static vnode_t *
744 vn_under(vnode_t *vp)
745 {
746 	vnode_t *uvp;
747 	vfs_t *vfsp;
748 
749 	while (vp->v_flag & VROOT) {
750 
751 		vfsp = vp->v_vfsp;
752 		vfs_rlock_wait(vfsp);
753 		if ((uvp = vfsp->vfs_vnodecovered) == NULL ||
754 		    (vfsp->vfs_flag & VFS_UNMOUNTED)) {
755 			vfs_unlock(vfsp);
756 			break;
757 		}
758 		VN_HOLD(uvp);
759 		vfs_unlock(vfsp);
760 		VN_RELE(vp);
761 		vp = uvp;
762 	}
763 
764 	return (vp);
765 }
766 
767 static int
768 vnode_match(vnode_t *v1, vnode_t *v2, cred_t *cr)
769 {
770 	vattr_t	v1attr, v2attr;
771 
772 	/*
773 	 * If we have a device file, check to see if is a cloned open of the
774 	 * same device.  For self-cloning devices, the major numbers will match.
775 	 * For devices cloned through the 'clone' driver, the minor number of
776 	 * the source device will be the same as the major number of the cloned
777 	 * device.
778 	 */
779 	if ((v1->v_type == VCHR || v1->v_type == VBLK) &&
780 	    v1->v_type == v2->v_type) {
781 		if ((spec_is_selfclone(v1) || spec_is_selfclone(v2)) &&
782 		    getmajor(v1->v_rdev) == getmajor(v2->v_rdev))
783 			return (1);
784 
785 		if (spec_is_clone(v1) &&
786 		    getmajor(v1->v_rdev) == getminor(v2->v_rdev))
787 			return (1);
788 
789 		if (spec_is_clone(v2) &&
790 		    getmajor(v2->v_rdev) == getminor(v1->v_rdev))
791 			return (1);
792 	}
793 
794 	v1attr.va_mask = v2attr.va_mask = AT_TYPE;
795 
796 	/*
797 	 * This check for symbolic links handles the pseudo-symlinks in procfs.
798 	 * These particular links have v_type of VDIR, but the attributes have a
799 	 * type of VLNK.  We need to avoid these links because otherwise if we
800 	 * are currently in '/proc/self/fd', then '/proc/self/cwd' will compare
801 	 * as the same vnode.
802 	 */
803 	if (VOP_GETATTR(v1, &v1attr, 0, cr, NULL) != 0 ||
804 	    VOP_GETATTR(v2, &v2attr, 0, cr, NULL) != 0 ||
805 	    v1attr.va_type == VLNK || v2attr.va_type == VLNK)
806 		return (0);
807 
808 	v1attr.va_mask = v2attr.va_mask = AT_TYPE | AT_FSID | AT_NODEID;
809 
810 	if (VOP_GETATTR(v1, &v1attr, ATTR_REAL, cr, NULL) != 0 ||
811 	    VOP_GETATTR(v2, &v2attr, ATTR_REAL, cr, NULL) != 0)
812 		return (0);
813 
814 	return (v1attr.va_fsid == v2attr.va_fsid &&
815 	    v1attr.va_nodeid == v2attr.va_nodeid);
816 }
817 
818 
819 /*
820  * Find the entry in the directory corresponding to the target vnode.
821  */
822 int
823 dirfindvp(vnode_t *vrootp, vnode_t *dvp, vnode_t *tvp, cred_t *cr, char *dbuf,
824     size_t dlen, dirent64_t **rdp)
825 {
826 	size_t dbuflen;
827 	struct iovec iov;
828 	struct uio uio;
829 	int error;
830 	int eof;
831 	vnode_t *cmpvp;
832 	struct dirent64 *dp;
833 	pathname_t pnp;
834 
835 	ASSERT(dvp->v_type == VDIR);
836 
837 	/*
838 	 * This is necessary because of the strange semantics of VOP_LOOKUP().
839 	 */
840 	bzero(&pnp, sizeof (pnp));
841 
842 	eof = 0;
843 
844 	uio.uio_iov = &iov;
845 	uio.uio_iovcnt = 1;
846 	uio.uio_segflg = UIO_SYSSPACE;
847 	uio.uio_fmode = 0;
848 	uio.uio_extflg = UIO_COPY_CACHED;
849 	uio.uio_loffset = 0;
850 
851 	if ((error = VOP_ACCESS(dvp, VREAD, 0, cr, NULL)) != 0)
852 		return (error);
853 
854 	while (!eof) {
855 		uio.uio_resid = dlen;
856 		iov.iov_base = dbuf;
857 		iov.iov_len = dlen;
858 
859 		(void) VOP_RWLOCK(dvp, V_WRITELOCK_FALSE, NULL);
860 		error = VOP_READDIR(dvp, &uio, cr, &eof, NULL, 0);
861 		VOP_RWUNLOCK(dvp, V_WRITELOCK_FALSE, NULL);
862 
863 		dbuflen = dlen - uio.uio_resid;
864 
865 		if (error || dbuflen == 0)
866 			break;
867 
868 		dp = (dirent64_t *)dbuf;
869 		while ((intptr_t)dp < (intptr_t)dbuf + dbuflen) {
870 			/*
871 			 * Ignore '.' and '..' entries
872 			 */
873 			if (strcmp(dp->d_name, ".") == 0 ||
874 			    strcmp(dp->d_name, "..") == 0) {
875 				dp = (dirent64_t *)((intptr_t)dp +
876 				    dp->d_reclen);
877 				continue;
878 			}
879 
880 			error = VOP_LOOKUP(dvp, dp->d_name, &cmpvp, &pnp, 0,
881 			    vrootp, cr, NULL, NULL, NULL);
882 
883 			/*
884 			 * We only want to bail out if there was an error other
885 			 * than ENOENT.  Otherwise, it could be that someone
886 			 * just removed an entry since the readdir() call, and
887 			 * the entry we want is further on in the directory.
888 			 */
889 			if (error == 0) {
890 				if (vnode_match(tvp, cmpvp, cr)) {
891 					VN_RELE(cmpvp);
892 					*rdp = dp;
893 					return (0);
894 				}
895 
896 				VN_RELE(cmpvp);
897 			} else if (error != ENOENT) {
898 				return (error);
899 			}
900 
901 			dp = (dirent64_t *)((intptr_t)dp + dp->d_reclen);
902 		}
903 	}
904 
905 	/*
906 	 * Something strange has happened, this directory does not contain the
907 	 * specified vnode.  This should never happen in the normal case, since
908 	 * we ensured that dvp is the parent of vp.  This is possible in some
909 	 * rare conditions (races and the special .zfs directory).
910 	 */
911 	if (error == 0) {
912 		error = VOP_LOOKUP(dvp, ".zfs", &cmpvp, &pnp, 0, vrootp, cr,
913 		    NULL, NULL, NULL);
914 		if (error == 0) {
915 			if (vnode_match(tvp, cmpvp, cr)) {
916 				(void) strcpy(dp->d_name, ".zfs");
917 				dp->d_reclen = strlen(".zfs");
918 				dp->d_off = 2;
919 				dp->d_ino = 1;
920 				*rdp = dp;
921 			} else {
922 				error = ENOENT;
923 			}
924 			VN_RELE(cmpvp);
925 		}
926 	}
927 
928 	return (error);
929 }
930 
931 /*
932  * Given a global path (from rootdir), and a vnode that is the current root,
933  * return the portion of the path that is beneath the current root or NULL on
934  * failure.  The path MUST be a resolved path (no '..' entries or symlinks),
935  * otherwise this function will fail.
936  */
937 static char *
938 localpath(char *path, struct vnode *vrootp, cred_t *cr)
939 {
940 	vnode_t *vp;
941 	vnode_t *cvp;
942 	char component[MAXNAMELEN];
943 	char *ret = NULL;
944 	pathname_t pn;
945 
946 	/*
947 	 * We use vn_compare() instead of VN_CMP() in order to detect lofs
948 	 * mounts and stacked vnodes.
949 	 */
950 	if (vn_compare(vrootp, rootdir))
951 		return (path);
952 
953 	if (pn_get(path, UIO_SYSSPACE, &pn) != 0)
954 		return (NULL);
955 
956 	vp = rootdir;
957 	VN_HOLD(vp);
958 
959 	if (vn_ismntpt(vp) && traverse(&vp) != 0) {
960 		VN_RELE(vp);
961 		pn_free(&pn);
962 		return (NULL);
963 	}
964 
965 	while (pn_pathleft(&pn)) {
966 		pn_skipslash(&pn);
967 
968 		if (pn_getcomponent(&pn, component) != 0)
969 			break;
970 
971 		if (VOP_LOOKUP(vp, component, &cvp, &pn, 0, rootdir, cr,
972 		    NULL, NULL, NULL) != 0)
973 			break;
974 		VN_RELE(vp);
975 		vp = cvp;
976 
977 		if (vn_ismntpt(vp) && traverse(&vp) != 0)
978 			break;
979 
980 		if (vn_compare(vp, vrootp)) {
981 			ret = path + (pn.pn_path - pn.pn_buf);
982 			break;
983 		}
984 	}
985 
986 	VN_RELE(vp);
987 	pn_free(&pn);
988 
989 	return (ret);
990 }
991 
992 /*
993  * Given a directory, return the full, resolved path.  This looks up "..",
994  * searches for the given vnode in the parent, appends the component, etc.  It
995  * is used to implement vnodetopath() and getcwd() when the cached path fails.
996  */
997 static int
998 dirtopath(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen, int flags,
999     cred_t *cr)
1000 {
1001 	pathname_t pn, rpn, emptypn;
1002 	vnode_t *cmpvp, *pvp = NULL;
1003 	vnode_t *startvp = vp;
1004 	int err = 0, vprivs;
1005 	size_t complen;
1006 	char *dbuf;
1007 	dirent64_t *dp;
1008 	char		*bufloc;
1009 	size_t		dlen = DIRENT64_RECLEN(MAXPATHLEN);
1010 	refstr_t	*mntpt;
1011 
1012 	/* Operation only allowed on directories */
1013 	ASSERT(vp->v_type == VDIR);
1014 
1015 	/* We must have at least enough space for "/" */
1016 	if (buflen < 2)
1017 		return (ENAMETOOLONG);
1018 
1019 	/* Start at end of string with terminating null */
1020 	bufloc = &buf[buflen - 1];
1021 	*bufloc = '\0';
1022 
1023 	pn_alloc(&pn);
1024 	pn_alloc(&rpn);
1025 	dbuf = kmem_alloc(dlen, KM_SLEEP);
1026 	bzero(&emptypn, sizeof (emptypn));
1027 
1028 	/*
1029 	 * Begin with an additional reference on vp.  This will be decremented
1030 	 * during the loop.
1031 	 */
1032 	VN_HOLD(vp);
1033 
1034 	for (;;) {
1035 		/*
1036 		 * Return if we've reached the root.  If the buffer is empty,
1037 		 * return '/'.  We explicitly don't use vn_compare(), since it
1038 		 * compares the real vnodes.  A lofs mount of '/' would produce
1039 		 * incorrect results otherwise.
1040 		 */
1041 		if (VN_CMP(vrootp, vp)) {
1042 			if (*bufloc == '\0')
1043 				*--bufloc = '/';
1044 			break;
1045 		}
1046 
1047 		/*
1048 		 * If we've reached the VFS root, something has gone wrong.  We
1049 		 * should have reached the root in the above check.  The only
1050 		 * explantation is that 'vp' is not contained withing the given
1051 		 * root, in which case we return EPERM.
1052 		 */
1053 		if (VN_CMP(rootdir, vp)) {
1054 			err = EPERM;
1055 			goto out;
1056 		}
1057 
1058 		/*
1059 		 * Shortcut: see if this vnode is a mountpoint.  If so,
1060 		 * grab the path information from the vfs_t.
1061 		 */
1062 		if (vp->v_flag & VROOT) {
1063 
1064 			mntpt = vfs_getmntpoint(vp->v_vfsp);
1065 			if ((err = pn_set(&pn, (char *)refstr_value(mntpt)))
1066 			    == 0) {
1067 				refstr_rele(mntpt);
1068 				rpn.pn_path = rpn.pn_buf;
1069 
1070 				/*
1071 				 * Ensure the mountpoint still exists.
1072 				 */
1073 				VN_HOLD(vrootp);
1074 				if (vrootp != rootdir)
1075 					VN_HOLD(vrootp);
1076 				if (lookuppnvp(&pn, &rpn, flags, NULL,
1077 				    &cmpvp, vrootp, vrootp, cr) == 0) {
1078 
1079 					if (VN_CMP(vp, cmpvp)) {
1080 						VN_RELE(cmpvp);
1081 
1082 						complen = strlen(rpn.pn_path);
1083 						bufloc -= complen;
1084 						if (bufloc < buf) {
1085 							err = ERANGE;
1086 							goto out;
1087 						}
1088 						bcopy(rpn.pn_path, bufloc,
1089 						    complen);
1090 						break;
1091 					} else {
1092 						VN_RELE(cmpvp);
1093 					}
1094 				}
1095 			} else {
1096 				refstr_rele(mntpt);
1097 			}
1098 		}
1099 
1100 		/*
1101 		 * Shortcut: see if this vnode has correct v_path. If so,
1102 		 * we have the work done.
1103 		 */
1104 		mutex_enter(&vp->v_lock);
1105 		if (vp->v_path != NULL) {
1106 
1107 			if ((err = pn_set(&pn, vp->v_path)) == 0) {
1108 				mutex_exit(&vp->v_lock);
1109 				rpn.pn_path = rpn.pn_buf;
1110 
1111 				/*
1112 				 * Ensure the v_path pointing to correct vnode
1113 				 */
1114 				VN_HOLD(vrootp);
1115 				if (vrootp != rootdir)
1116 					VN_HOLD(vrootp);
1117 				if (lookuppnvp(&pn, &rpn, flags, NULL,
1118 				    &cmpvp, vrootp, vrootp, cr) == 0) {
1119 
1120 					if (VN_CMP(vp, cmpvp)) {
1121 						VN_RELE(cmpvp);
1122 
1123 						complen = strlen(rpn.pn_path);
1124 						bufloc -= complen;
1125 						if (bufloc < buf) {
1126 							err = ERANGE;
1127 							goto out;
1128 						}
1129 						bcopy(rpn.pn_path, bufloc,
1130 						    complen);
1131 						break;
1132 					} else {
1133 						VN_RELE(cmpvp);
1134 					}
1135 				}
1136 			} else {
1137 				mutex_exit(&vp->v_lock);
1138 			}
1139 		} else {
1140 			mutex_exit(&vp->v_lock);
1141 		}
1142 
1143 		/*
1144 		 * Shortcuts failed, search for this vnode in its parent.  If
1145 		 * this is a mountpoint, then get the vnode underneath.
1146 		 */
1147 		if (vp->v_flag & VROOT)
1148 			vp = vn_under(vp);
1149 		if ((err = VOP_LOOKUP(vp, "..", &pvp, &emptypn, 0, vrootp, cr,
1150 		    NULL, NULL, NULL)) != 0)
1151 			goto out;
1152 
1153 		/*
1154 		 * With extended attributes, it's possible for a directory to
1155 		 * have a parent that is a regular file.  Check for that here.
1156 		 */
1157 		if (pvp->v_type != VDIR) {
1158 			err = ENOTDIR;
1159 			goto out;
1160 		}
1161 
1162 		/*
1163 		 * If this is true, something strange has happened.  This is
1164 		 * only true if we are the root of a filesystem, which should
1165 		 * have been caught by the check above.
1166 		 */
1167 		if (VN_CMP(pvp, vp)) {
1168 			err = ENOENT;
1169 			goto out;
1170 		}
1171 
1172 		/*
1173 		 * Check if we have read and search privilege so, that
1174 		 * we can lookup the path in the directory
1175 		 */
1176 		vprivs = (flags & LOOKUP_CHECKREAD) ? VREAD | VEXEC : VEXEC;
1177 		if ((err = VOP_ACCESS(pvp, vprivs, 0, cr, NULL)) != 0) {
1178 			goto out;
1179 		}
1180 
1181 		/*
1182 		 * Try to obtain the path component from dnlc cache
1183 		 * before searching through the directory.
1184 		 */
1185 		if ((cmpvp = dnlc_reverse_lookup(vp, dbuf, dlen)) != NULL) {
1186 			/*
1187 			 * If we got parent vnode as a result,
1188 			 * then the answered path is correct.
1189 			 */
1190 			if (VN_CMP(cmpvp, pvp)) {
1191 				VN_RELE(cmpvp);
1192 				complen = strlen(dbuf);
1193 				bufloc -= complen;
1194 				if (bufloc <= buf) {
1195 					err = ENAMETOOLONG;
1196 					goto out;
1197 				}
1198 				bcopy(dbuf, bufloc, complen);
1199 
1200 				/* Prepend a slash to the current path */
1201 				*--bufloc = '/';
1202 
1203 				/* And continue with the next component */
1204 				VN_RELE(vp);
1205 				vp = pvp;
1206 				pvp = NULL;
1207 				continue;
1208 			} else {
1209 				VN_RELE(cmpvp);
1210 			}
1211 		}
1212 
1213 		/*
1214 		 * Search the parent directory for the entry corresponding to
1215 		 * this vnode.
1216 		 */
1217 		if ((err = dirfindvp(vrootp, pvp, vp, cr, dbuf, dlen, &dp))
1218 		    != 0)
1219 			goto out;
1220 		complen = strlen(dp->d_name);
1221 		bufloc -= complen;
1222 		if (bufloc <= buf) {
1223 			err = ENAMETOOLONG;
1224 			goto out;
1225 		}
1226 		bcopy(dp->d_name, bufloc, complen);
1227 
1228 		/* Prepend a slash to the current path.  */
1229 		*--bufloc = '/';
1230 
1231 		/* And continue with the next component */
1232 		VN_RELE(vp);
1233 		vp = pvp;
1234 		pvp = NULL;
1235 	}
1236 
1237 	/*
1238 	 * Place the path at the beginning of the buffer.
1239 	 */
1240 	if (bufloc != buf)
1241 		ovbcopy(bufloc, buf, buflen - (bufloc - buf));
1242 
1243 out:
1244 	/*
1245 	 * If the error was ESTALE and the current directory to look in
1246 	 * was the root for this lookup, the root for a mounted file
1247 	 * system, or the starting directory for lookups, then
1248 	 * return ENOENT instead of ESTALE.  In this case, no recovery
1249 	 * is possible by the higher level.  If ESTALE was returned for
1250 	 * some intermediate directory along the path, then recovery
1251 	 * is potentially possible and retrying from the higher level
1252 	 * will either correct the situation by purging stale cache
1253 	 * entries or eventually get back to the point where no recovery
1254 	 * is possible.
1255 	 */
1256 	if (err == ESTALE &&
1257 	    (VN_CMP(vp, vrootp) || (vp->v_flag & VROOT) || vp == startvp))
1258 		err = ENOENT;
1259 
1260 	kmem_free(dbuf, dlen);
1261 	VN_RELE(vp);
1262 	if (pvp)
1263 		VN_RELE(pvp);
1264 	pn_free(&pn);
1265 	pn_free(&rpn);
1266 
1267 	return (err);
1268 }
1269 
1270 /*
1271  * The additional flag, LOOKUP_CHECKREAD, is used to enforce artificial
1272  * constraints in order to be standards compliant.  For example, if we have
1273  * the cached path of '/foo/bar', and '/foo' has permissions 100 (execute
1274  * only), then we can legitimately look up the path to the current working
1275  * directory without needing read permission.  Existing standards tests,
1276  * however, assume that we are determining the path by repeatedly looking up
1277  * "..".  We need to keep this behavior in order to maintain backwards
1278  * compatibility.
1279  */
1280 static int
1281 vnodetopath_common(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen,
1282     cred_t *cr, int flags)
1283 {
1284 	pathname_t pn, rpn;
1285 	int ret, len;
1286 	vnode_t *compvp, *pvp, *realvp;
1287 	proc_t *p = curproc;
1288 	char path[MAXNAMELEN];
1289 	int doclose = 0;
1290 
1291 	/*
1292 	 * If vrootp is NULL, get the root for curproc.  Callers with any other
1293 	 * requirements should pass in a different vrootp.
1294 	 */
1295 	if (vrootp == NULL) {
1296 		mutex_enter(&p->p_lock);
1297 		if ((vrootp = PTOU(p)->u_rdir) == NULL)
1298 			vrootp = rootdir;
1299 		VN_HOLD(vrootp);
1300 		mutex_exit(&p->p_lock);
1301 	} else {
1302 		VN_HOLD(vrootp);
1303 	}
1304 
1305 	/*
1306 	 * This is to get around an annoying artifact of the /proc filesystem,
1307 	 * which is the behavior of {cwd/root}.  Trying to resolve this path
1308 	 * will result in /proc/pid/cwd instead of whatever the real working
1309 	 * directory is.  We can't rely on VOP_REALVP(), since that will break
1310 	 * lofs.  The only difference between procfs and lofs is that opening
1311 	 * the file will return the underling vnode in the case of procfs.
1312 	 */
1313 	if (vp->v_type == VDIR && VOP_REALVP(vp, &realvp, NULL) == 0 &&
1314 	    realvp != vp) {
1315 		VN_HOLD(vp);
1316 		if (VOP_OPEN(&vp, FREAD, cr, NULL) == 0)
1317 			doclose = 1;
1318 		else
1319 			VN_RELE(vp);
1320 	}
1321 
1322 	pn_alloc(&pn);
1323 
1324 	/*
1325 	 * Check to see if we have a cached path in the vnode.
1326 	 */
1327 	mutex_enter(&vp->v_lock);
1328 	if (vp->v_path != NULL) {
1329 		(void) pn_set(&pn, vp->v_path);
1330 		mutex_exit(&vp->v_lock);
1331 
1332 		pn_alloc(&rpn);
1333 
1334 		/* We should only cache absolute paths */
1335 		ASSERT(pn.pn_buf[0] == '/');
1336 
1337 		/*
1338 		 * If we are in a zone or a chroot environment, then we have to
1339 		 * take additional steps, since the path to the root might not
1340 		 * be readable with the current credentials, even though the
1341 		 * process can legitmately access the file.  In this case, we
1342 		 * do the following:
1343 		 *
1344 		 * lookuppnvp() with all privileges to get the resolved path.
1345 		 * call localpath() to get the local portion of the path, and
1346 		 * continue as normal.
1347 		 *
1348 		 * If the the conversion to a local path fails, then we continue
1349 		 * as normal.  This is a heuristic to make process object file
1350 		 * paths available from within a zone.  Because lofs doesn't
1351 		 * support page operations, the vnode stored in the seg_t is
1352 		 * actually the underlying real vnode, not the lofs node itself.
1353 		 * Most of the time, the lofs path is the same as the underlying
1354 		 * vnode (for example, /usr/lib/libc.so.1).
1355 		 */
1356 		if (vrootp != rootdir) {
1357 			char *local = NULL;
1358 			VN_HOLD(rootdir);
1359 			if (lookuppnvp(&pn, &rpn, FOLLOW,
1360 			    NULL, &compvp, rootdir, rootdir, kcred) == 0) {
1361 				local = localpath(rpn.pn_path, vrootp,
1362 				    kcred);
1363 				VN_RELE(compvp);
1364 			}
1365 
1366 			/*
1367 			 * The original pn was changed through lookuppnvp().
1368 			 * Set it to local for next validation attempt.
1369 			 */
1370 			if (local) {
1371 				(void) pn_set(&pn, local);
1372 			} else {
1373 				goto notcached;
1374 			}
1375 		}
1376 
1377 		/*
1378 		 * We should have a local path at this point, so start the
1379 		 * search from the root of the current process.
1380 		 */
1381 		VN_HOLD(vrootp);
1382 		if (vrootp != rootdir)
1383 			VN_HOLD(vrootp);
1384 		ret = lookuppnvp(&pn, &rpn, FOLLOW | flags, NULL,
1385 		    &compvp, vrootp, vrootp, cr);
1386 		if (ret == 0) {
1387 			/*
1388 			 * Check to see if the returned vnode is the same as
1389 			 * the one we expect.  If not, give up.
1390 			 */
1391 			if (!vn_compare(vp, compvp) &&
1392 			    !vnode_match(vp, compvp, cr)) {
1393 				VN_RELE(compvp);
1394 				goto notcached;
1395 			}
1396 
1397 			VN_RELE(compvp);
1398 
1399 			/*
1400 			 * Return the result.
1401 			 */
1402 			if (buflen <= rpn.pn_pathlen)
1403 				goto notcached;
1404 
1405 			bcopy(rpn.pn_path, buf, rpn.pn_pathlen + 1);
1406 			pn_free(&pn);
1407 			pn_free(&rpn);
1408 			VN_RELE(vrootp);
1409 			if (doclose) {
1410 				(void) VOP_CLOSE(vp, FREAD, 1, 0, cr, NULL);
1411 				VN_RELE(vp);
1412 			}
1413 			return (0);
1414 		}
1415 
1416 notcached:
1417 		pn_free(&rpn);
1418 	} else {
1419 		mutex_exit(&vp->v_lock);
1420 	}
1421 
1422 	pn_free(&pn);
1423 
1424 	if (vp->v_type != VDIR) {
1425 		/*
1426 		 * If we don't have a directory, try to find it in the dnlc via
1427 		 * reverse lookup.  Once this is found, we can use the regular
1428 		 * directory search to find the full path.
1429 		 */
1430 		if ((pvp = dnlc_reverse_lookup(vp, path, MAXNAMELEN)) != NULL) {
1431 			/*
1432 			 * Check if we have read privilege so, that
1433 			 * we can lookup the path in the directory
1434 			 */
1435 			ret = 0;
1436 			if ((flags & LOOKUP_CHECKREAD)) {
1437 				ret = VOP_ACCESS(pvp, VREAD, 0, cr, NULL);
1438 			}
1439 			if (ret == 0) {
1440 				ret = dirtopath(vrootp, pvp, buf, buflen,
1441 				    flags, cr);
1442 			}
1443 			if (ret == 0) {
1444 				len = strlen(buf);
1445 				if (len + strlen(path) + 1 >= buflen) {
1446 					ret = ENAMETOOLONG;
1447 				} else {
1448 					if (buf[len - 1] != '/')
1449 						buf[len++] = '/';
1450 					bcopy(path, buf + len,
1451 					    strlen(path) + 1);
1452 				}
1453 			}
1454 
1455 			VN_RELE(pvp);
1456 		} else
1457 			ret = ENOENT;
1458 	} else
1459 		ret = dirtopath(vrootp, vp, buf, buflen, flags, cr);
1460 
1461 	VN_RELE(vrootp);
1462 	if (doclose) {
1463 		(void) VOP_CLOSE(vp, FREAD, 1, 0, cr, NULL);
1464 		VN_RELE(vp);
1465 	}
1466 
1467 	return (ret);
1468 }
1469 
1470 int
1471 vnodetopath(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen, cred_t *cr)
1472 {
1473 	return (vnodetopath_common(vrootp, vp, buf, buflen, cr, 0));
1474 }
1475 
1476 int
1477 dogetcwd(char *buf, size_t buflen)
1478 {
1479 	int ret;
1480 	vnode_t *vp;
1481 	vnode_t *compvp;
1482 	refstr_t *cwd, *oldcwd;
1483 	const char *value;
1484 	pathname_t rpnp, pnp;
1485 	proc_t *p = curproc;
1486 
1487 	/*
1488 	 * Check to see if there is a cached version of the cwd.  If so, lookup
1489 	 * the cached value and make sure it is the same vnode.
1490 	 */
1491 	mutex_enter(&p->p_lock);
1492 	if ((cwd = PTOU(p)->u_cwd) != NULL)
1493 		refstr_hold(cwd);
1494 	vp = PTOU(p)->u_cdir;
1495 	VN_HOLD(vp);
1496 	mutex_exit(&p->p_lock);
1497 
1498 	/*
1499 	 * Make sure we have permission to access the current directory.
1500 	 */
1501 	if ((ret = VOP_ACCESS(vp, VEXEC, 0, CRED(), NULL)) != 0) {
1502 		if (cwd != NULL)
1503 			refstr_rele(cwd);
1504 		VN_RELE(vp);
1505 		return (ret);
1506 	}
1507 
1508 	if (cwd) {
1509 		value = refstr_value(cwd);
1510 		if ((ret = pn_get((char *)value, UIO_SYSSPACE, &pnp)) != 0) {
1511 			refstr_rele(cwd);
1512 			VN_RELE(vp);
1513 			return (ret);
1514 		}
1515 
1516 		pn_alloc(&rpnp);
1517 
1518 		if (lookuppn(&pnp, &rpnp, NO_FOLLOW, NULL, &compvp) == 0) {
1519 
1520 			if (VN_CMP(vp, compvp) &&
1521 			    strcmp(value, rpnp.pn_path) == 0) {
1522 				VN_RELE(compvp);
1523 				VN_RELE(vp);
1524 				pn_free(&pnp);
1525 				pn_free(&rpnp);
1526 				if (strlen(value) + 1 > buflen) {
1527 					refstr_rele(cwd);
1528 					return (ENAMETOOLONG);
1529 				}
1530 				bcopy(value, buf, strlen(value) + 1);
1531 				refstr_rele(cwd);
1532 				return (0);
1533 			}
1534 
1535 			VN_RELE(compvp);
1536 		}
1537 
1538 		pn_free(&rpnp);
1539 		pn_free(&pnp);
1540 
1541 		refstr_rele(cwd);
1542 	}
1543 
1544 	ret = vnodetopath_common(NULL, vp, buf, buflen, CRED(),
1545 	    LOOKUP_CHECKREAD);
1546 
1547 	VN_RELE(vp);
1548 
1549 	/*
1550 	 * Store the new cwd and replace the existing cached copy.
1551 	 */
1552 	if (ret == 0)
1553 		cwd = refstr_alloc(buf);
1554 	else
1555 		cwd = NULL;
1556 
1557 	mutex_enter(&p->p_lock);
1558 	oldcwd = PTOU(p)->u_cwd;
1559 	PTOU(p)->u_cwd = cwd;
1560 	mutex_exit(&p->p_lock);
1561 
1562 	if (oldcwd)
1563 		refstr_rele(oldcwd);
1564 
1565 	return (ret);
1566 }
1567