xref: /illumos-gate/usr/src/uts/common/fs/lookup.c (revision 6e062f4a)
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  */
26 
27 /*	Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T	*/
28 /*	  All Rights Reserved  	*/
29 
30 /*
31  * University Copyright- Copyright (c) 1982, 1986, 1988
32  * The Regents of the University of California
33  * All Rights Reserved
34  *
35  * University Acknowledgment- Portions of this document are derived from
36  * software developed by the University of California, Berkeley, and its
37  * contributors.
38  */
39 
40 #include <sys/types.h>
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/cpuvar.h>
44 #include <sys/errno.h>
45 #include <sys/cred.h>
46 #include <sys/user.h>
47 #include <sys/uio.h>
48 #include <sys/vfs.h>
49 #include <sys/vnode.h>
50 #include <sys/pathname.h>
51 #include <sys/proc.h>
52 #include <sys/vtrace.h>
53 #include <sys/sysmacros.h>
54 #include <sys/debug.h>
55 #include <sys/dirent.h>
56 #include <c2/audit.h>
57 #include <sys/zone.h>
58 #include <sys/dnlc.h>
59 #include <sys/fs/snode.h>
60 
61 /* Controls whether paths are stored with vnodes. */
62 int vfs_vnode_path = 1;
63 
64 int
65 lookupname(
66 	char *fnamep,
67 	enum uio_seg seg,
68 	int followlink,
69 	vnode_t **dirvpp,
70 	vnode_t **compvpp)
71 {
72 	return (lookupnameatcred(fnamep, seg, followlink, dirvpp, compvpp, NULL,
73 	    CRED()));
74 }
75 
76 /*
77  * Lookup the user file name,
78  * Handle allocation and freeing of pathname buffer, return error.
79  */
80 int
81 lookupnameatcred(
82 	char *fnamep,			/* user pathname */
83 	enum uio_seg seg,		/* addr space that name is in */
84 	int followlink,			/* follow sym links */
85 	vnode_t **dirvpp,		/* ret for ptr to parent dir vnode */
86 	vnode_t **compvpp,		/* ret for ptr to component vnode */
87 	vnode_t *startvp,		/* start path search from vp */
88 	cred_t *cr)			/* credential */
89 {
90 	char namebuf[TYPICALMAXPATHLEN];
91 	struct pathname lookpn;
92 	int error;
93 
94 	error = pn_get_buf(fnamep, seg, &lookpn, namebuf, sizeof (namebuf));
95 	if (error == 0) {
96 		error = lookuppnatcred(&lookpn, NULL, followlink,
97 		    dirvpp, compvpp, startvp, cr);
98 	}
99 	if (error == ENAMETOOLONG) {
100 		/*
101 		 * This thread used a pathname > TYPICALMAXPATHLEN bytes long.
102 		 */
103 		if (error = pn_get(fnamep, seg, &lookpn))
104 			return (error);
105 		error = lookuppnatcred(&lookpn, NULL, followlink,
106 		    dirvpp, compvpp, startvp, cr);
107 		pn_free(&lookpn);
108 	}
109 
110 	return (error);
111 }
112 
113 int
114 lookupnameat(char *fnamep, enum uio_seg seg, int followlink,
115     vnode_t **dirvpp, vnode_t **compvpp, vnode_t *startvp)
116 {
117 	return (lookupnameatcred(fnamep, seg, followlink, dirvpp, compvpp,
118 	    startvp, CRED()));
119 }
120 
121 int
122 lookuppn(
123 	struct pathname *pnp,
124 	struct pathname *rpnp,
125 	int followlink,
126 	vnode_t **dirvpp,
127 	vnode_t **compvpp)
128 {
129 	return (lookuppnatcred(pnp, rpnp, followlink, dirvpp, compvpp, NULL,
130 	    CRED()));
131 }
132 
133 /*
134  * Lookup the user file name from a given vp, using a specific credential.
135  */
136 int
137 lookuppnatcred(
138 	struct pathname *pnp,		/* pathname to lookup */
139 	struct pathname *rpnp,		/* if non-NULL, return resolved path */
140 	int followlink,			/* (don't) follow sym links */
141 	vnode_t **dirvpp,		/* ptr for parent vnode */
142 	vnode_t **compvpp,		/* ptr for entry vnode */
143 	vnode_t *startvp,		/* start search from this vp */
144 	cred_t *cr)			/* user credential */
145 {
146 	vnode_t *vp;	/* current directory vp */
147 	vnode_t *rootvp;
148 	proc_t *p = curproc;
149 
150 	if (pnp->pn_pathlen == 0)
151 		return (ENOENT);
152 
153 	mutex_enter(&p->p_lock);	/* for u_rdir and u_cdir */
154 	if ((rootvp = PTOU(p)->u_rdir) == NULL)
155 		rootvp = rootdir;
156 	else if (rootvp != rootdir)	/* no need to VN_HOLD rootdir */
157 		VN_HOLD(rootvp);
158 
159 	if (pnp->pn_path[0] == '/') {
160 		vp = rootvp;
161 	} else {
162 		vp = (startvp == NULL) ? PTOU(p)->u_cdir : startvp;
163 	}
164 	VN_HOLD(vp);
165 	mutex_exit(&p->p_lock);
166 
167 	/*
168 	 * Skip over leading slashes
169 	 */
170 	if (pnp->pn_path[0] == '/') {
171 		do {
172 			pnp->pn_path++;
173 			pnp->pn_pathlen--;
174 		} while (pnp->pn_path[0] == '/');
175 	}
176 
177 	return (lookuppnvp(pnp, rpnp, followlink, dirvpp,
178 	    compvpp, rootvp, vp, cr));
179 }
180 
181 int
182 lookuppnat(struct pathname *pnp, struct pathname *rpnp,
183     int followlink, vnode_t **dirvpp, vnode_t **compvpp,
184     vnode_t *startvp)
185 {
186 	return (lookuppnatcred(pnp, rpnp, followlink, dirvpp, compvpp, startvp,
187 	    CRED()));
188 }
189 
190 /* Private flag to do our getcwd() dirty work */
191 #define	LOOKUP_CHECKREAD	0x10
192 #define	LOOKUP_MASK		(~LOOKUP_CHECKREAD)
193 
194 /*
195  * Starting at current directory, translate pathname pnp to end.
196  * Leave pathname of final component in pnp, return the vnode
197  * for the final component in *compvpp, and return the vnode
198  * for the parent of the final component in dirvpp.
199  *
200  * This is the central routine in pathname translation and handles
201  * multiple components in pathnames, separating them at /'s.  It also
202  * implements mounted file systems and processes symbolic links.
203  *
204  * vp is the vnode where the directory search should start.
205  *
206  * Reference counts: vp must be held prior to calling this function.  rootvp
207  * should only be held if rootvp != rootdir.
208  */
209 int
210 lookuppnvp(
211 	struct pathname *pnp,		/* pathname to lookup */
212 	struct pathname *rpnp,		/* if non-NULL, return resolved path */
213 	int flags,			/* follow symlinks */
214 	vnode_t **dirvpp,		/* ptr for parent vnode */
215 	vnode_t **compvpp,		/* ptr for entry vnode */
216 	vnode_t *rootvp,		/* rootvp */
217 	vnode_t *vp,			/* directory to start search at */
218 	cred_t *cr)			/* user's credential */
219 {
220 	vnode_t *cvp;	/* current component vp */
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, &cvp, 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, &cvp, pnp, lookup_flags,
395 		    rootvp, zone_kcred(), NULL, NULL, pp);
396 
397 	if (error) {
398 		cvp = NULL;
399 		/*
400 		 * On error, return hard error if
401 		 * (a) we're not at the end of the pathname yet, or
402 		 * (b) the caller didn't want the parent directory, or
403 		 * (c) we failed for some reason other than a missing entry.
404 		 */
405 		if (pn_pathleft(pnp) || dirvpp == NULL || error != ENOENT)
406 			goto bad;
407 		if (auditing) {	/* directory access */
408 			if (error = audit_savepath(pnp, vp, vp, error, cr))
409 				goto bad_noaudit;
410 		}
411 
412 		pn_setlast(pnp);
413 		/*
414 		 * We inform the caller that the desired entry must be
415 		 * a directory by adding a '/' to the component name.
416 		 */
417 		if (must_be_directory && (error = pn_addslash(pnp)) != 0)
418 			goto bad;
419 		*dirvpp = vp;
420 		if (compvpp != NULL)
421 			*compvpp = NULL;
422 		if (rootvp != rootdir)
423 			VN_RELE(rootvp);
424 		if (pp)
425 			pn_free(pp);
426 		return (0);
427 	}
428 
429 	/*
430 	 * Traverse mount points.
431 	 * XXX why don't we need to hold a read lock here (call vn_vfsrlock)?
432 	 * What prevents a concurrent update to v_vfsmountedhere?
433 	 * 	Possible answer: if mounting, we might not see the mount
434 	 *	if it is concurrently coming into existence, but that's
435 	 *	really not much different from the thread running a bit slower.
436 	 *	If unmounting, we may get into traverse() when we shouldn't,
437 	 *	but traverse() will catch this case for us.
438 	 *	(For this to work, fetching v_vfsmountedhere had better
439 	 *	be atomic!)
440 	 */
441 	if (vn_mountedvfs(cvp) != NULL) {
442 		if ((error = traverse(&cvp)) != 0)
443 			goto bad;
444 	}
445 
446 	/*
447 	 * If we hit a symbolic link and there is more path to be
448 	 * translated or this operation does not wish to apply
449 	 * to a link, then place the contents of the link at the
450 	 * front of the remaining pathname.
451 	 */
452 	if (cvp->v_type == VLNK && ((flags & FOLLOW) || pn_pathleft(pnp))) {
453 		struct pathname linkpath;
454 
455 		if (++nlink > MAXSYMLINKS) {
456 			error = ELOOP;
457 			goto bad;
458 		}
459 		pn_alloc(&linkpath);
460 		if (error = pn_getsymlink(cvp, &linkpath, cr)) {
461 			pn_free(&linkpath);
462 			goto bad;
463 		}
464 
465 		if (auditing)
466 			audit_symlink(pnp, &linkpath);
467 
468 		if (pn_pathleft(&linkpath) == 0)
469 			(void) pn_set(&linkpath, ".");
470 		error = pn_insert(pnp, &linkpath, strlen(component));
471 		pn_free(&linkpath);
472 		if (error)
473 			goto bad;
474 		VN_RELE(cvp);
475 		cvp = NULL;
476 		if (pnp->pn_pathlen == 0) {
477 			error = ENOENT;
478 			goto bad;
479 		}
480 		if (pnp->pn_path[0] == '/') {
481 			do {
482 				pnp->pn_path++;
483 				pnp->pn_pathlen--;
484 			} while (pnp->pn_path[0] == '/');
485 			VN_RELE(vp);
486 			vp = rootvp;
487 			VN_HOLD(vp);
488 		}
489 		if (auditing)
490 			audit_anchorpath(pnp, vp == rootvp);
491 		if (pn_fixslash(pnp)) {
492 			flags |= FOLLOW;
493 			must_be_directory = 1;
494 		}
495 		goto next;
496 	}
497 
498 	/*
499 	 * If rpnp is non-NULL, remember the resolved path name therein.
500 	 * Do not include "." components.  Collapse occurrences of
501 	 * "previous/..", so long as "previous" is not itself "..".
502 	 * Exhausting rpnp results in error ENAMETOOLONG.
503 	 */
504 	if (rpnp && strcmp(component, ".") != 0) {
505 		size_t len;
506 
507 		if (strcmp(component, "..") == 0 &&
508 		    rpnp->pn_pathlen != 0 &&
509 		    !((rpnp->pn_pathlen > 2 &&
510 		    strncmp(rpnp->pn_path+rpnp->pn_pathlen-3, "/..", 3) == 0) ||
511 		    (rpnp->pn_pathlen == 2 &&
512 		    strncmp(rpnp->pn_path, "..", 2) == 0))) {
513 			while (rpnp->pn_pathlen &&
514 			    rpnp->pn_path[rpnp->pn_pathlen-1] != '/')
515 				rpnp->pn_pathlen--;
516 			if (rpnp->pn_pathlen > 1)
517 				rpnp->pn_pathlen--;
518 			rpnp->pn_path[rpnp->pn_pathlen] = '\0';
519 		} else {
520 			if (rpnp->pn_pathlen != 0 &&
521 			    rpnp->pn_path[rpnp->pn_pathlen-1] != '/')
522 				rpnp->pn_path[rpnp->pn_pathlen++] = '/';
523 			if (flags & FIGNORECASE) {
524 				/*
525 				 * Return the case-preserved name
526 				 * within the resolved path.
527 				 */
528 				error = copystr(pp->pn_buf,
529 				    rpnp->pn_path + rpnp->pn_pathlen,
530 				    rpnp->pn_bufsize - rpnp->pn_pathlen, &len);
531 			} else {
532 				error = copystr(component,
533 				    rpnp->pn_path + rpnp->pn_pathlen,
534 				    rpnp->pn_bufsize - rpnp->pn_pathlen, &len);
535 			}
536 			if (error)	/* copystr() returns ENAMETOOLONG */
537 				goto bad;
538 			rpnp->pn_pathlen += (len - 1);
539 			ASSERT(rpnp->pn_bufsize > rpnp->pn_pathlen);
540 		}
541 	}
542 
543 	/*
544 	 * If no more components, return last directory (if wanted) and
545 	 * last component (if wanted).
546 	 */
547 	if (pn_pathleft(pnp) == 0) {
548 		/*
549 		 * If there was a trailing slash in the pathname,
550 		 * make sure the last component is a directory.
551 		 */
552 		if (must_be_directory && cvp->v_type != VDIR) {
553 			error = ENOTDIR;
554 			goto bad;
555 		}
556 		if (dirvpp != NULL) {
557 			/*
558 			 * Check that we have the real parent and not
559 			 * an alias of the last component.
560 			 */
561 			if (vn_compare(vp, cvp)) {
562 				if (auditing)
563 					(void) audit_savepath(pnp, cvp, vp,
564 					    EINVAL, cr);
565 				pn_setlast(pnp);
566 				VN_RELE(vp);
567 				VN_RELE(cvp);
568 				if (rootvp != rootdir)
569 					VN_RELE(rootvp);
570 				if (pp)
571 					pn_free(pp);
572 				return (EINVAL);
573 			}
574 			*dirvpp = vp;
575 		} else
576 			VN_RELE(vp);
577 		if (auditing)
578 			(void) audit_savepath(pnp, cvp, vp, 0, cr);
579 		if (pnp->pn_path == pnp->pn_buf)
580 			(void) pn_set(pnp, ".");
581 		else
582 			pn_setlast(pnp);
583 		if (rpnp) {
584 			if (VN_CMP(cvp, rootvp))
585 				(void) pn_set(rpnp, "/");
586 			else if (rpnp->pn_pathlen == 0)
587 				(void) pn_set(rpnp, ".");
588 		}
589 
590 		if (compvpp != NULL)
591 			*compvpp = cvp;
592 		else
593 			VN_RELE(cvp);
594 		if (rootvp != rootdir)
595 			VN_RELE(rootvp);
596 		if (pp)
597 			pn_free(pp);
598 		return (0);
599 	}
600 
601 	/*
602 	 * Skip over slashes from end of last component.
603 	 */
604 	while (pnp->pn_path[0] == '/') {
605 		pnp->pn_path++;
606 		pnp->pn_pathlen--;
607 	}
608 
609 	/*
610 	 * Searched through another level of directory:
611 	 * release previous directory handle and save new (result
612 	 * of lookup) as current directory.
613 	 */
614 	VN_RELE(vp);
615 	vp = cvp;
616 	cvp = NULL;
617 	goto next;
618 
619 bad:
620 	if (auditing)	/* reached end of path */
621 		(void) audit_savepath(pnp, cvp, vp, error, cr);
622 bad_noaudit:
623 	/*
624 	 * Error.  Release vnodes and return.
625 	 */
626 	if (cvp)
627 		VN_RELE(cvp);
628 	/*
629 	 * If the error was ESTALE and the current directory to look in
630 	 * was the root for this lookup, the root for a mounted file
631 	 * system, or the starting directory for lookups, then
632 	 * return ENOENT instead of ESTALE.  In this case, no recovery
633 	 * is possible by the higher level.  If ESTALE was returned for
634 	 * some intermediate directory along the path, then recovery
635 	 * is potentially possible and retrying from the higher level
636 	 * will either correct the situation by purging stale cache
637 	 * entries or eventually get back to the point where no recovery
638 	 * is possible.
639 	 */
640 	if (error == ESTALE &&
641 	    (VN_CMP(vp, rootvp) || (vp->v_flag & VROOT) || vp == startvp))
642 		error = ENOENT;
643 	VN_RELE(vp);
644 	if (rootvp != rootdir)
645 		VN_RELE(rootvp);
646 	if (pp)
647 		pn_free(pp);
648 	return (error);
649 }
650 
651 /*
652  * Traverse a mount point.  Routine accepts a vnode pointer as a reference
653  * parameter and performs the indirection, releasing the original vnode.
654  */
655 int
656 traverse(vnode_t **cvpp)
657 {
658 	int error = 0;
659 	vnode_t *cvp;
660 	vnode_t *tvp;
661 	vfs_t *vfsp;
662 
663 	cvp = *cvpp;
664 
665 	/*
666 	 * If this vnode is mounted on, then we transparently indirect
667 	 * to the vnode which is the root of the mounted file system.
668 	 * Before we do this we must check that an unmount is not in
669 	 * progress on this vnode.
670 	 */
671 
672 	for (;;) {
673 		/*
674 		 * Try to read lock the vnode.  If this fails because
675 		 * the vnode is already write locked, then check to
676 		 * see whether it is the current thread which locked
677 		 * the vnode.  If it is not, then read lock the vnode
678 		 * by waiting to acquire the lock.
679 		 *
680 		 * The code path in domount() is an example of support
681 		 * which needs to look up two pathnames and locks one
682 		 * of them in between the two lookups.
683 		 */
684 		error = vn_vfsrlock(cvp);
685 		if (error) {
686 			if (!vn_vfswlock_held(cvp))
687 				error = vn_vfsrlock_wait(cvp);
688 			if (error != 0) {
689 				/*
690 				 * lookuppn() expects a held vnode to be
691 				 * returned because it promptly calls
692 				 * VN_RELE after the error return
693 				 */
694 				*cvpp = cvp;
695 				return (error);
696 			}
697 		}
698 
699 		/*
700 		 * Reached the end of the mount chain?
701 		 */
702 		vfsp = vn_mountedvfs(cvp);
703 		if (vfsp == NULL) {
704 			vn_vfsunlock(cvp);
705 			break;
706 		}
707 
708 		/*
709 		 * The read lock must be held across the call to VFS_ROOT() to
710 		 * prevent a concurrent unmount from destroying the vfs.
711 		 */
712 		error = VFS_ROOT(vfsp, &tvp);
713 		vn_vfsunlock(cvp);
714 
715 		if (error)
716 			break;
717 
718 		VN_RELE(cvp);
719 
720 		cvp = tvp;
721 	}
722 
723 	*cvpp = cvp;
724 	return (error);
725 }
726 
727 /*
728  * Return the lowermost vnode if this is a mountpoint.
729  */
730 static vnode_t *
731 vn_under(vnode_t *vp)
732 {
733 	vnode_t *uvp;
734 	vfs_t *vfsp;
735 
736 	while (vp->v_flag & VROOT) {
737 
738 		vfsp = vp->v_vfsp;
739 		vfs_rlock_wait(vfsp);
740 		if ((uvp = vfsp->vfs_vnodecovered) == NULL ||
741 		    (vfsp->vfs_flag & VFS_UNMOUNTED)) {
742 			vfs_unlock(vfsp);
743 			break;
744 		}
745 		VN_HOLD(uvp);
746 		vfs_unlock(vfsp);
747 		VN_RELE(vp);
748 		vp = uvp;
749 	}
750 
751 	return (vp);
752 }
753 
754 static int
755 vnode_match(vnode_t *v1, vnode_t *v2, cred_t *cr)
756 {
757 	vattr_t	v1attr, v2attr;
758 
759 	/*
760 	 * If we have a device file, check to see if is a cloned open of the
761 	 * same device.  For self-cloning devices, the major numbers will match.
762 	 * For devices cloned through the 'clone' driver, the minor number of
763 	 * the source device will be the same as the major number of the cloned
764 	 * device.
765 	 */
766 	if ((v1->v_type == VCHR || v1->v_type == VBLK) &&
767 	    v1->v_type == v2->v_type) {
768 		if ((spec_is_selfclone(v1) || spec_is_selfclone(v2)) &&
769 		    getmajor(v1->v_rdev) == getmajor(v2->v_rdev))
770 			return (1);
771 
772 		if (spec_is_clone(v1) &&
773 		    getmajor(v1->v_rdev) == getminor(v2->v_rdev))
774 			return (1);
775 
776 		if (spec_is_clone(v2) &&
777 		    getmajor(v2->v_rdev) == getminor(v1->v_rdev))
778 			return (1);
779 	}
780 
781 	v1attr.va_mask = v2attr.va_mask = AT_TYPE;
782 
783 	/*
784 	 * This check for symbolic links handles the pseudo-symlinks in procfs.
785 	 * These particular links have v_type of VDIR, but the attributes have a
786 	 * type of VLNK.  We need to avoid these links because otherwise if we
787 	 * are currently in '/proc/self/fd', then '/proc/self/cwd' will compare
788 	 * as the same vnode.
789 	 */
790 	if (VOP_GETATTR(v1, &v1attr, 0, cr, NULL) != 0 ||
791 	    VOP_GETATTR(v2, &v2attr, 0, cr, NULL) != 0 ||
792 	    v1attr.va_type == VLNK || v2attr.va_type == VLNK)
793 		return (0);
794 
795 	v1attr.va_mask = v2attr.va_mask = AT_TYPE | AT_FSID | AT_NODEID;
796 
797 	if (VOP_GETATTR(v1, &v1attr, ATTR_REAL, cr, NULL) != 0 ||
798 	    VOP_GETATTR(v2, &v2attr, ATTR_REAL, cr, NULL) != 0)
799 		return (0);
800 
801 	return (v1attr.va_fsid == v2attr.va_fsid &&
802 	    v1attr.va_nodeid == v2attr.va_nodeid);
803 }
804 
805 
806 /*
807  * Find the entry in the directory corresponding to the target vnode.
808  */
809 int
810 dirfindvp(vnode_t *vrootp, vnode_t *dvp, vnode_t *tvp, cred_t *cr, char *dbuf,
811     size_t dlen, dirent64_t **rdp)
812 {
813 	size_t dbuflen;
814 	struct iovec iov;
815 	struct uio uio;
816 	int error;
817 	int eof;
818 	vnode_t *cmpvp;
819 	struct dirent64 *dp;
820 	pathname_t pnp;
821 
822 	ASSERT(dvp->v_type == VDIR);
823 
824 	/*
825 	 * This is necessary because of the strange semantics of VOP_LOOKUP().
826 	 */
827 	bzero(&pnp, sizeof (pnp));
828 
829 	eof = 0;
830 
831 	uio.uio_iov = &iov;
832 	uio.uio_iovcnt = 1;
833 	uio.uio_segflg = UIO_SYSSPACE;
834 	uio.uio_fmode = 0;
835 	uio.uio_extflg = UIO_COPY_CACHED;
836 	uio.uio_loffset = 0;
837 
838 	if ((error = VOP_ACCESS(dvp, VREAD, 0, cr, NULL)) != 0)
839 		return (error);
840 
841 	while (!eof) {
842 		uio.uio_resid = dlen;
843 		iov.iov_base = dbuf;
844 		iov.iov_len = dlen;
845 
846 		(void) VOP_RWLOCK(dvp, V_WRITELOCK_FALSE, NULL);
847 		error = VOP_READDIR(dvp, &uio, cr, &eof, NULL, 0);
848 		VOP_RWUNLOCK(dvp, V_WRITELOCK_FALSE, NULL);
849 
850 		dbuflen = dlen - uio.uio_resid;
851 
852 		if (error || dbuflen == 0)
853 			break;
854 
855 		dp = (dirent64_t *)dbuf;
856 		while ((intptr_t)dp < (intptr_t)dbuf + dbuflen) {
857 			/*
858 			 * Ignore '.' and '..' entries
859 			 */
860 			if (strcmp(dp->d_name, ".") == 0 ||
861 			    strcmp(dp->d_name, "..") == 0) {
862 				dp = (dirent64_t *)((intptr_t)dp +
863 				    dp->d_reclen);
864 				continue;
865 			}
866 
867 			error = VOP_LOOKUP(dvp, dp->d_name, &cmpvp, &pnp, 0,
868 			    vrootp, cr, NULL, NULL, NULL);
869 
870 			/*
871 			 * We only want to bail out if there was an error other
872 			 * than ENOENT.  Otherwise, it could be that someone
873 			 * just removed an entry since the readdir() call, and
874 			 * the entry we want is further on in the directory.
875 			 */
876 			if (error == 0) {
877 				if (vnode_match(tvp, cmpvp, cr)) {
878 					VN_RELE(cmpvp);
879 					*rdp = dp;
880 					return (0);
881 				}
882 
883 				VN_RELE(cmpvp);
884 			} else if (error != ENOENT) {
885 				return (error);
886 			}
887 
888 			dp = (dirent64_t *)((intptr_t)dp + dp->d_reclen);
889 		}
890 	}
891 
892 	/*
893 	 * Something strange has happened, this directory does not contain the
894 	 * specified vnode.  This should never happen in the normal case, since
895 	 * we ensured that dvp is the parent of vp.  This is possible in some
896 	 * rare conditions (races and the special .zfs directory).
897 	 */
898 	if (error == 0) {
899 		error = VOP_LOOKUP(dvp, ".zfs", &cmpvp, &pnp, 0, vrootp, cr,
900 		    NULL, NULL, NULL);
901 		if (error == 0) {
902 			if (vnode_match(tvp, cmpvp, cr)) {
903 				(void) strcpy(dp->d_name, ".zfs");
904 				dp->d_reclen = strlen(".zfs");
905 				dp->d_off = 2;
906 				dp->d_ino = 1;
907 				*rdp = dp;
908 			} else {
909 				error = ENOENT;
910 			}
911 			VN_RELE(cmpvp);
912 		}
913 	}
914 
915 	return (error);
916 }
917 
918 /*
919  * Given a global path (from rootdir), and a vnode that is the current root,
920  * return the portion of the path that is beneath the current root or NULL on
921  * failure.  The path MUST be a resolved path (no '..' entries or symlinks),
922  * otherwise this function will fail.
923  */
924 static char *
925 localpath(char *path, struct vnode *vrootp, cred_t *cr)
926 {
927 	vnode_t *vp;
928 	vnode_t *cvp;
929 	char component[MAXNAMELEN];
930 	char *ret = NULL;
931 	pathname_t pn;
932 
933 	/*
934 	 * We use vn_compare() instead of VN_CMP() in order to detect lofs
935 	 * mounts and stacked vnodes.
936 	 */
937 	if (vn_compare(vrootp, rootdir))
938 		return (path);
939 
940 	if (pn_get(path, UIO_SYSSPACE, &pn) != 0)
941 		return (NULL);
942 
943 	vp = rootdir;
944 	VN_HOLD(vp);
945 
946 	if (vn_ismntpt(vp) && traverse(&vp) != 0) {
947 		VN_RELE(vp);
948 		pn_free(&pn);
949 		return (NULL);
950 	}
951 
952 	while (pn_pathleft(&pn)) {
953 		pn_skipslash(&pn);
954 
955 		if (pn_getcomponent(&pn, component) != 0)
956 			break;
957 
958 		if (VOP_LOOKUP(vp, component, &cvp, &pn, 0, rootdir, cr,
959 		    NULL, NULL, NULL) != 0)
960 			break;
961 		VN_RELE(vp);
962 		vp = cvp;
963 
964 		if (vn_ismntpt(vp) && traverse(&vp) != 0)
965 			break;
966 
967 		if (vn_compare(vp, vrootp)) {
968 			ret = path + (pn.pn_path - pn.pn_buf);
969 			break;
970 		}
971 	}
972 
973 	VN_RELE(vp);
974 	pn_free(&pn);
975 
976 	return (ret);
977 }
978 
979 /*
980  * Given a directory, return the full, resolved path.  This looks up "..",
981  * searches for the given vnode in the parent, appends the component, etc.  It
982  * is used to implement vnodetopath() and getcwd() when the cached path fails.
983  */
984 static int
985 dirtopath(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen, int flags,
986     cred_t *cr)
987 {
988 	pathname_t pn, rpn, emptypn;
989 	vnode_t *cmpvp, *pvp = NULL;
990 	vnode_t *startvp = vp;
991 	int err = 0, vprivs;
992 	size_t complen;
993 	char *dbuf;
994 	dirent64_t *dp;
995 	char		*bufloc;
996 	size_t		dlen = DIRENT64_RECLEN(MAXPATHLEN);
997 	refstr_t	*mntpt;
998 
999 	/* Operation only allowed on directories */
1000 	ASSERT(vp->v_type == VDIR);
1001 
1002 	/* We must have at least enough space for "/" */
1003 	if (buflen < 2)
1004 		return (ENAMETOOLONG);
1005 
1006 	/* Start at end of string with terminating null */
1007 	bufloc = &buf[buflen - 1];
1008 	*bufloc = '\0';
1009 
1010 	pn_alloc(&pn);
1011 	pn_alloc(&rpn);
1012 	dbuf = kmem_alloc(dlen, KM_SLEEP);
1013 	bzero(&emptypn, sizeof (emptypn));
1014 
1015 	/*
1016 	 * Begin with an additional reference on vp.  This will be decremented
1017 	 * during the loop.
1018 	 */
1019 	VN_HOLD(vp);
1020 
1021 	for (;;) {
1022 		/*
1023 		 * Return if we've reached the root.  If the buffer is empty,
1024 		 * return '/'.  We explicitly don't use vn_compare(), since it
1025 		 * compares the real vnodes.  A lofs mount of '/' would produce
1026 		 * incorrect results otherwise.
1027 		 */
1028 		if (VN_CMP(vrootp, vp)) {
1029 			if (*bufloc == '\0')
1030 				*--bufloc = '/';
1031 			break;
1032 		}
1033 
1034 		/*
1035 		 * If we've reached the VFS root, something has gone wrong.  We
1036 		 * should have reached the root in the above check.  The only
1037 		 * explantation is that 'vp' is not contained withing the given
1038 		 * root, in which case we return EPERM.
1039 		 */
1040 		if (VN_CMP(rootdir, vp)) {
1041 			err = EPERM;
1042 			goto out;
1043 		}
1044 
1045 		/*
1046 		 * Shortcut: see if this vnode is a mountpoint.  If so,
1047 		 * grab the path information from the vfs_t.
1048 		 */
1049 		if (vp->v_flag & VROOT) {
1050 
1051 			mntpt = vfs_getmntpoint(vp->v_vfsp);
1052 			if ((err = pn_set(&pn, (char *)refstr_value(mntpt)))
1053 			    == 0) {
1054 				refstr_rele(mntpt);
1055 				rpn.pn_path = rpn.pn_buf;
1056 
1057 				/*
1058 				 * Ensure the mountpoint still exists.
1059 				 */
1060 				VN_HOLD(vrootp);
1061 				if (vrootp != rootdir)
1062 					VN_HOLD(vrootp);
1063 				if (lookuppnvp(&pn, &rpn, flags, NULL,
1064 				    &cmpvp, vrootp, vrootp, cr) == 0) {
1065 
1066 					if (VN_CMP(vp, cmpvp)) {
1067 						VN_RELE(cmpvp);
1068 
1069 						complen = strlen(rpn.pn_path);
1070 						bufloc -= complen;
1071 						if (bufloc < buf) {
1072 							err = ERANGE;
1073 							goto out;
1074 						}
1075 						bcopy(rpn.pn_path, bufloc,
1076 						    complen);
1077 						break;
1078 					} else {
1079 						VN_RELE(cmpvp);
1080 					}
1081 				}
1082 			} else {
1083 				refstr_rele(mntpt);
1084 			}
1085 		}
1086 
1087 		/*
1088 		 * Shortcut: see if this vnode has correct v_path. If so,
1089 		 * we have the work done.
1090 		 */
1091 		mutex_enter(&vp->v_lock);
1092 		if (vp->v_path != NULL) {
1093 
1094 			if ((err = pn_set(&pn, vp->v_path)) == 0) {
1095 				mutex_exit(&vp->v_lock);
1096 				rpn.pn_path = rpn.pn_buf;
1097 
1098 				/*
1099 				 * Ensure the v_path pointing to correct vnode
1100 				 */
1101 				VN_HOLD(vrootp);
1102 				if (vrootp != rootdir)
1103 					VN_HOLD(vrootp);
1104 				if (lookuppnvp(&pn, &rpn, flags, NULL,
1105 				    &cmpvp, vrootp, vrootp, cr) == 0) {
1106 
1107 					if (VN_CMP(vp, cmpvp)) {
1108 						VN_RELE(cmpvp);
1109 
1110 						complen = strlen(rpn.pn_path);
1111 						bufloc -= complen;
1112 						if (bufloc < buf) {
1113 							err = ERANGE;
1114 							goto out;
1115 						}
1116 						bcopy(rpn.pn_path, bufloc,
1117 						    complen);
1118 						break;
1119 					} else {
1120 						VN_RELE(cmpvp);
1121 					}
1122 				}
1123 			} else {
1124 				mutex_exit(&vp->v_lock);
1125 			}
1126 		} else {
1127 			mutex_exit(&vp->v_lock);
1128 		}
1129 
1130 		/*
1131 		 * Shortcuts failed, search for this vnode in its parent.  If
1132 		 * this is a mountpoint, then get the vnode underneath.
1133 		 */
1134 		if (vp->v_flag & VROOT)
1135 			vp = vn_under(vp);
1136 		if ((err = VOP_LOOKUP(vp, "..", &pvp, &emptypn, 0, vrootp, cr,
1137 		    NULL, NULL, NULL)) != 0)
1138 			goto out;
1139 
1140 		/*
1141 		 * With extended attributes, it's possible for a directory to
1142 		 * have a parent that is a regular file.  Check for that here.
1143 		 */
1144 		if (pvp->v_type != VDIR) {
1145 			err = ENOTDIR;
1146 			goto out;
1147 		}
1148 
1149 		/*
1150 		 * If this is true, something strange has happened.  This is
1151 		 * only true if we are the root of a filesystem, which should
1152 		 * have been caught by the check above.
1153 		 */
1154 		if (VN_CMP(pvp, vp)) {
1155 			err = ENOENT;
1156 			goto out;
1157 		}
1158 
1159 		/*
1160 		 * Check if we have read and search privilege so, that
1161 		 * we can lookup the path in the directory
1162 		 */
1163 		vprivs = (flags & LOOKUP_CHECKREAD) ? VREAD | VEXEC : VEXEC;
1164 		if ((err = VOP_ACCESS(pvp, vprivs, 0, cr, NULL)) != 0) {
1165 			goto out;
1166 		}
1167 
1168 		/*
1169 		 * Try to obtain the path component from dnlc cache
1170 		 * before searching through the directory.
1171 		 */
1172 		if ((cmpvp = dnlc_reverse_lookup(vp, dbuf, dlen)) != NULL) {
1173 			/*
1174 			 * If we got parent vnode as a result,
1175 			 * then the answered path is correct.
1176 			 */
1177 			if (VN_CMP(cmpvp, pvp)) {
1178 				VN_RELE(cmpvp);
1179 				complen = strlen(dbuf);
1180 				bufloc -= complen;
1181 				if (bufloc <= buf) {
1182 					err = ENAMETOOLONG;
1183 					goto out;
1184 				}
1185 				bcopy(dbuf, bufloc, complen);
1186 
1187 				/* Prepend a slash to the current path */
1188 				*--bufloc = '/';
1189 
1190 				/* And continue with the next component */
1191 				VN_RELE(vp);
1192 				vp = pvp;
1193 				pvp = NULL;
1194 				continue;
1195 			} else {
1196 				VN_RELE(cmpvp);
1197 			}
1198 		}
1199 
1200 		/*
1201 		 * Search the parent directory for the entry corresponding to
1202 		 * this vnode.
1203 		 */
1204 		if ((err = dirfindvp(vrootp, pvp, vp, cr, dbuf, dlen, &dp))
1205 		    != 0)
1206 			goto out;
1207 		complen = strlen(dp->d_name);
1208 		bufloc -= complen;
1209 		if (bufloc <= buf) {
1210 			err = ENAMETOOLONG;
1211 			goto out;
1212 		}
1213 		bcopy(dp->d_name, bufloc, complen);
1214 
1215 		/* Prepend a slash to the current path.  */
1216 		*--bufloc = '/';
1217 
1218 		/* And continue with the next component */
1219 		VN_RELE(vp);
1220 		vp = pvp;
1221 		pvp = NULL;
1222 	}
1223 
1224 	/*
1225 	 * Place the path at the beginning of the buffer.
1226 	 */
1227 	if (bufloc != buf)
1228 		ovbcopy(bufloc, buf, buflen - (bufloc - buf));
1229 
1230 out:
1231 	/*
1232 	 * If the error was ESTALE and the current directory to look in
1233 	 * was the root for this lookup, the root for a mounted file
1234 	 * system, or the starting directory for lookups, then
1235 	 * return ENOENT instead of ESTALE.  In this case, no recovery
1236 	 * is possible by the higher level.  If ESTALE was returned for
1237 	 * some intermediate directory along the path, then recovery
1238 	 * is potentially possible and retrying from the higher level
1239 	 * will either correct the situation by purging stale cache
1240 	 * entries or eventually get back to the point where no recovery
1241 	 * is possible.
1242 	 */
1243 	if (err == ESTALE &&
1244 	    (VN_CMP(vp, vrootp) || (vp->v_flag & VROOT) || vp == startvp))
1245 		err = ENOENT;
1246 
1247 	kmem_free(dbuf, dlen);
1248 	VN_RELE(vp);
1249 	if (pvp)
1250 		VN_RELE(pvp);
1251 	pn_free(&pn);
1252 	pn_free(&rpn);
1253 
1254 	return (err);
1255 }
1256 
1257 /*
1258  * The additional flag, LOOKUP_CHECKREAD, is used to enforce artificial
1259  * constraints in order to be standards compliant.  For example, if we have
1260  * the cached path of '/foo/bar', and '/foo' has permissions 100 (execute
1261  * only), then we can legitimately look up the path to the current working
1262  * directory without needing read permission.  Existing standards tests,
1263  * however, assume that we are determining the path by repeatedly looking up
1264  * "..".  We need to keep this behavior in order to maintain backwards
1265  * compatibility.
1266  */
1267 static int
1268 vnodetopath_common(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen,
1269     cred_t *cr, int flags)
1270 {
1271 	pathname_t pn, rpn;
1272 	int ret, len;
1273 	vnode_t *compvp, *pvp, *realvp;
1274 	proc_t *p = curproc;
1275 	char path[MAXNAMELEN];
1276 	int doclose = 0;
1277 
1278 	/*
1279 	 * If vrootp is NULL, get the root for curproc.  Callers with any other
1280 	 * requirements should pass in a different vrootp.
1281 	 */
1282 	if (vrootp == NULL) {
1283 		mutex_enter(&p->p_lock);
1284 		if ((vrootp = PTOU(p)->u_rdir) == NULL)
1285 			vrootp = rootdir;
1286 		VN_HOLD(vrootp);
1287 		mutex_exit(&p->p_lock);
1288 	} else {
1289 		VN_HOLD(vrootp);
1290 	}
1291 
1292 	/*
1293 	 * This is to get around an annoying artifact of the /proc filesystem,
1294 	 * which is the behavior of {cwd/root}.  Trying to resolve this path
1295 	 * will result in /proc/pid/cwd instead of whatever the real working
1296 	 * directory is.  We can't rely on VOP_REALVP(), since that will break
1297 	 * lofs.  The only difference between procfs and lofs is that opening
1298 	 * the file will return the underling vnode in the case of procfs.
1299 	 */
1300 	if (vp->v_type == VDIR && VOP_REALVP(vp, &realvp, NULL) == 0 &&
1301 	    realvp != vp) {
1302 		VN_HOLD(vp);
1303 		if (VOP_OPEN(&vp, FREAD, cr, NULL) == 0)
1304 			doclose = 1;
1305 		else
1306 			VN_RELE(vp);
1307 	}
1308 
1309 	pn_alloc(&pn);
1310 
1311 	/*
1312 	 * Check to see if we have a cached path in the vnode.
1313 	 */
1314 	mutex_enter(&vp->v_lock);
1315 	if (vp->v_path != NULL) {
1316 		(void) pn_set(&pn, vp->v_path);
1317 		mutex_exit(&vp->v_lock);
1318 
1319 		pn_alloc(&rpn);
1320 
1321 		/* We should only cache absolute paths */
1322 		ASSERT(pn.pn_buf[0] == '/');
1323 
1324 		/*
1325 		 * If we are in a zone or a chroot environment, then we have to
1326 		 * take additional steps, since the path to the root might not
1327 		 * be readable with the current credentials, even though the
1328 		 * process can legitmately access the file.  In this case, we
1329 		 * do the following:
1330 		 *
1331 		 * lookuppnvp() with all privileges to get the resolved path.
1332 		 * call localpath() to get the local portion of the path, and
1333 		 * continue as normal.
1334 		 *
1335 		 * If the the conversion to a local path fails, then we continue
1336 		 * as normal.  This is a heuristic to make process object file
1337 		 * paths available from within a zone.  Because lofs doesn't
1338 		 * support page operations, the vnode stored in the seg_t is
1339 		 * actually the underlying real vnode, not the lofs node itself.
1340 		 * Most of the time, the lofs path is the same as the underlying
1341 		 * vnode (for example, /usr/lib/libc.so.1).
1342 		 */
1343 		if (vrootp != rootdir) {
1344 			char *local = NULL;
1345 			VN_HOLD(rootdir);
1346 			if (lookuppnvp(&pn, &rpn, FOLLOW,
1347 			    NULL, &compvp, rootdir, rootdir, kcred) == 0) {
1348 				local = localpath(rpn.pn_path, vrootp,
1349 				    kcred);
1350 				VN_RELE(compvp);
1351 			}
1352 
1353 			/*
1354 			 * The original pn was changed through lookuppnvp().
1355 			 * Set it to local for next validation attempt.
1356 			 */
1357 			if (local) {
1358 				(void) pn_set(&pn, local);
1359 			} else {
1360 				goto notcached;
1361 			}
1362 		}
1363 
1364 		/*
1365 		 * We should have a local path at this point, so start the
1366 		 * search from the root of the current process.
1367 		 */
1368 		VN_HOLD(vrootp);
1369 		if (vrootp != rootdir)
1370 			VN_HOLD(vrootp);
1371 		ret = lookuppnvp(&pn, &rpn, FOLLOW | flags, NULL,
1372 		    &compvp, vrootp, vrootp, cr);
1373 		if (ret == 0) {
1374 			/*
1375 			 * Check to see if the returned vnode is the same as
1376 			 * the one we expect.  If not, give up.
1377 			 */
1378 			if (!vn_compare(vp, compvp) &&
1379 			    !vnode_match(vp, compvp, cr)) {
1380 				VN_RELE(compvp);
1381 				goto notcached;
1382 			}
1383 
1384 			VN_RELE(compvp);
1385 
1386 			/*
1387 			 * Return the result.
1388 			 */
1389 			if (buflen <= rpn.pn_pathlen)
1390 				goto notcached;
1391 
1392 			bcopy(rpn.pn_path, buf, rpn.pn_pathlen + 1);
1393 			pn_free(&pn);
1394 			pn_free(&rpn);
1395 			VN_RELE(vrootp);
1396 			if (doclose) {
1397 				(void) VOP_CLOSE(vp, FREAD, 1, 0, cr, NULL);
1398 				VN_RELE(vp);
1399 			}
1400 			return (0);
1401 		}
1402 
1403 notcached:
1404 		pn_free(&rpn);
1405 	} else {
1406 		mutex_exit(&vp->v_lock);
1407 	}
1408 
1409 	pn_free(&pn);
1410 
1411 	if (vp->v_type != VDIR) {
1412 		/*
1413 		 * If we don't have a directory, try to find it in the dnlc via
1414 		 * reverse lookup.  Once this is found, we can use the regular
1415 		 * directory search to find the full path.
1416 		 */
1417 		if ((pvp = dnlc_reverse_lookup(vp, path, MAXNAMELEN)) != NULL) {
1418 			/*
1419 			 * Check if we have read privilege so, that
1420 			 * we can lookup the path in the directory
1421 			 */
1422 			ret = 0;
1423 			if ((flags & LOOKUP_CHECKREAD)) {
1424 				ret = VOP_ACCESS(pvp, VREAD, 0, cr, NULL);
1425 			}
1426 			if (ret == 0) {
1427 				ret = dirtopath(vrootp, pvp, buf, buflen,
1428 				    flags, cr);
1429 			}
1430 			if (ret == 0) {
1431 				len = strlen(buf);
1432 				if (len + strlen(path) + 1 >= buflen) {
1433 					ret = ENAMETOOLONG;
1434 				} else {
1435 					if (buf[len - 1] != '/')
1436 						buf[len++] = '/';
1437 					bcopy(path, buf + len,
1438 					    strlen(path) + 1);
1439 				}
1440 			}
1441 
1442 			VN_RELE(pvp);
1443 		} else
1444 			ret = ENOENT;
1445 	} else
1446 		ret = dirtopath(vrootp, vp, buf, buflen, flags, cr);
1447 
1448 	VN_RELE(vrootp);
1449 	if (doclose) {
1450 		(void) VOP_CLOSE(vp, FREAD, 1, 0, cr, NULL);
1451 		VN_RELE(vp);
1452 	}
1453 
1454 	return (ret);
1455 }
1456 
1457 int
1458 vnodetopath(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen, cred_t *cr)
1459 {
1460 	return (vnodetopath_common(vrootp, vp, buf, buflen, cr, 0));
1461 }
1462 
1463 int
1464 dogetcwd(char *buf, size_t buflen)
1465 {
1466 	int ret;
1467 	vnode_t *vp;
1468 	vnode_t *compvp;
1469 	refstr_t *cwd, *oldcwd;
1470 	const char *value;
1471 	pathname_t rpnp, pnp;
1472 	proc_t *p = curproc;
1473 
1474 	/*
1475 	 * Check to see if there is a cached version of the cwd.  If so, lookup
1476 	 * the cached value and make sure it is the same vnode.
1477 	 */
1478 	mutex_enter(&p->p_lock);
1479 	if ((cwd = PTOU(p)->u_cwd) != NULL)
1480 		refstr_hold(cwd);
1481 	vp = PTOU(p)->u_cdir;
1482 	VN_HOLD(vp);
1483 	mutex_exit(&p->p_lock);
1484 
1485 	/*
1486 	 * Make sure we have permission to access the current directory.
1487 	 */
1488 	if ((ret = VOP_ACCESS(vp, VEXEC, 0, CRED(), NULL)) != 0) {
1489 		if (cwd != NULL)
1490 			refstr_rele(cwd);
1491 		VN_RELE(vp);
1492 		return (ret);
1493 	}
1494 
1495 	if (cwd) {
1496 		value = refstr_value(cwd);
1497 		if ((ret = pn_get((char *)value, UIO_SYSSPACE, &pnp)) != 0) {
1498 			refstr_rele(cwd);
1499 			VN_RELE(vp);
1500 			return (ret);
1501 		}
1502 
1503 		pn_alloc(&rpnp);
1504 
1505 		if (lookuppn(&pnp, &rpnp, NO_FOLLOW, NULL, &compvp) == 0) {
1506 
1507 			if (VN_CMP(vp, compvp) &&
1508 			    strcmp(value, rpnp.pn_path) == 0) {
1509 				VN_RELE(compvp);
1510 				VN_RELE(vp);
1511 				pn_free(&pnp);
1512 				pn_free(&rpnp);
1513 				if (strlen(value) + 1 > buflen) {
1514 					refstr_rele(cwd);
1515 					return (ENAMETOOLONG);
1516 				}
1517 				bcopy(value, buf, strlen(value) + 1);
1518 				refstr_rele(cwd);
1519 				return (0);
1520 			}
1521 
1522 			VN_RELE(compvp);
1523 		}
1524 
1525 		pn_free(&rpnp);
1526 		pn_free(&pnp);
1527 
1528 		refstr_rele(cwd);
1529 	}
1530 
1531 	ret = vnodetopath_common(NULL, vp, buf, buflen, CRED(),
1532 	    LOOKUP_CHECKREAD);
1533 
1534 	VN_RELE(vp);
1535 
1536 	/*
1537 	 * Store the new cwd and replace the existing cached copy.
1538 	 */
1539 	if (ret == 0)
1540 		cwd = refstr_alloc(buf);
1541 	else
1542 		cwd = NULL;
1543 
1544 	mutex_enter(&p->p_lock);
1545 	oldcwd = PTOU(p)->u_cwd;
1546 	PTOU(p)->u_cwd = cwd;
1547 	mutex_exit(&p->p_lock);
1548 
1549 	if (oldcwd)
1550 		refstr_rele(oldcwd);
1551 
1552 	return (ret);
1553 }
1554