xref: /titanic_44/usr/src/uts/common/fs/lookup.c (revision dd49f125507979bb2ab505a8daf2a46d1be27051)
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  * Copyright (c) 1988, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /*	Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T	*/
26 /*	  All Rights Reserved  	*/
27 
28 /*
29  * University Copyright- Copyright (c) 1982, 1986, 1988
30  * The Regents of the University of California
31  * All Rights Reserved
32  *
33  * University Acknowledgment- Portions of this document are derived from
34  * software developed by the University of California, Berkeley, and its
35  * contributors.
36  */
37 
38 #include <sys/types.h>
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/cpuvar.h>
42 #include <sys/errno.h>
43 #include <sys/cred.h>
44 #include <sys/user.h>
45 #include <sys/uio.h>
46 #include <sys/vfs.h>
47 #include <sys/vnode.h>
48 #include <sys/pathname.h>
49 #include <sys/proc.h>
50 #include <sys/vtrace.h>
51 #include <sys/sysmacros.h>
52 #include <sys/debug.h>
53 #include <sys/dirent.h>
54 #include <c2/audit.h>
55 #include <sys/zone.h>
56 #include <sys/dnlc.h>
57 #include <sys/fs/snode.h>
58 
59 /* Controls whether paths are stored with vnodes. */
60 int vfs_vnode_path = 1;
61 
62 int
63 lookupname(
64 	char *fnamep,
65 	enum uio_seg seg,
66 	enum symfollow followlink,
67 	vnode_t **dirvpp,
68 	vnode_t **compvpp)
69 {
70 	return (lookupnameatcred(fnamep, seg, followlink, dirvpp, compvpp, NULL,
71 	    CRED()));
72 }
73 
74 /*
75  * Lookup the user file name,
76  * Handle allocation and freeing of pathname buffer, return error.
77  */
78 int
79 lookupnameatcred(
80 	char *fnamep,			/* user pathname */
81 	enum uio_seg seg,		/* addr space that name is in */
82 	enum symfollow followlink,	/* follow sym links */
83 	vnode_t **dirvpp,		/* ret for ptr to parent dir vnode */
84 	vnode_t **compvpp,		/* ret for ptr to component vnode */
85 	vnode_t *startvp,		/* start path search from vp */
86 	cred_t *cr)			/* credential */
87 {
88 	char namebuf[TYPICALMAXPATHLEN];
89 	struct pathname lookpn;
90 	int error;
91 
92 	error = pn_get_buf(fnamep, seg, &lookpn, namebuf, sizeof (namebuf));
93 	if (error == 0) {
94 		error = lookuppnatcred(&lookpn, NULL, followlink,
95 		    dirvpp, compvpp, startvp, cr);
96 	}
97 	if (error == ENAMETOOLONG) {
98 		/*
99 		 * This thread used a pathname > TYPICALMAXPATHLEN bytes long.
100 		 */
101 		if (error = pn_get(fnamep, seg, &lookpn))
102 			return (error);
103 		error = lookuppnatcred(&lookpn, NULL, followlink,
104 		    dirvpp, compvpp, startvp, cr);
105 		pn_free(&lookpn);
106 	}
107 
108 	return (error);
109 }
110 
111 int
112 lookupnameat(char *fnamep, enum uio_seg seg, enum symfollow followlink,
113     vnode_t **dirvpp, vnode_t **compvpp, vnode_t *startvp)
114 {
115 	return (lookupnameatcred(fnamep, seg, followlink, dirvpp, compvpp,
116 	    startvp, CRED()));
117 }
118 
119 int
120 lookuppn(
121 	struct pathname *pnp,
122 	struct pathname *rpnp,
123 	enum symfollow followlink,
124 	vnode_t **dirvpp,
125 	vnode_t **compvpp)
126 {
127 	return (lookuppnatcred(pnp, rpnp, followlink, dirvpp, compvpp, NULL,
128 	    CRED()));
129 }
130 
131 /*
132  * Lookup the user file name from a given vp, using a specific credential.
133  */
134 int
135 lookuppnatcred(
136 	struct pathname *pnp,		/* pathname to lookup */
137 	struct pathname *rpnp,		/* if non-NULL, return resolved path */
138 	enum symfollow followlink,	/* (don't) follow sym links */
139 	vnode_t **dirvpp,		/* ptr for parent vnode */
140 	vnode_t **compvpp,		/* ptr for entry vnode */
141 	vnode_t *startvp,		/* start search from this vp */
142 	cred_t *cr)			/* user credential */
143 {
144 	vnode_t *vp;	/* current directory vp */
145 	vnode_t *rootvp;
146 	proc_t *p = curproc;
147 
148 	if (pnp->pn_pathlen == 0)
149 		return (ENOENT);
150 
151 	mutex_enter(&p->p_lock);	/* for u_rdir and u_cdir */
152 	if ((rootvp = PTOU(p)->u_rdir) == NULL)
153 		rootvp = rootdir;
154 	else if (rootvp != rootdir)	/* no need to VN_HOLD rootdir */
155 		VN_HOLD(rootvp);
156 
157 	if (pnp->pn_path[0] == '/') {
158 		vp = rootvp;
159 	} else {
160 		vp = (startvp == NULL) ? PTOU(p)->u_cdir : startvp;
161 	}
162 	VN_HOLD(vp);
163 	mutex_exit(&p->p_lock);
164 
165 	/*
166 	 * Skip over leading slashes
167 	 */
168 	if (pnp->pn_path[0] == '/') {
169 		do {
170 			pnp->pn_path++;
171 			pnp->pn_pathlen--;
172 		} while (pnp->pn_path[0] == '/');
173 	}
174 
175 	return (lookuppnvp(pnp, rpnp, followlink, dirvpp,
176 	    compvpp, rootvp, vp, cr));
177 }
178 
179 int
180 lookuppnat(struct pathname *pnp, struct pathname *rpnp,
181     enum symfollow followlink, vnode_t **dirvpp, vnode_t **compvpp,
182     vnode_t *startvp)
183 {
184 	return (lookuppnatcred(pnp, rpnp, followlink, dirvpp, compvpp, startvp,
185 	    CRED()));
186 }
187 
188 /* Private flag to do our getcwd() dirty work */
189 #define	LOOKUP_CHECKREAD	0x10
190 #define	LOOKUP_MASK		(~LOOKUP_CHECKREAD)
191 
192 /*
193  * Starting at current directory, translate pathname pnp to end.
194  * Leave pathname of final component in pnp, return the vnode
195  * for the final component in *compvpp, and return the vnode
196  * for the parent of the final component in dirvpp.
197  *
198  * This is the central routine in pathname translation and handles
199  * multiple components in pathnames, separating them at /'s.  It also
200  * implements mounted file systems and processes symbolic links.
201  *
202  * vp is the vnode where the directory search should start.
203  *
204  * Reference counts: vp must be held prior to calling this function.  rootvp
205  * should only be held if rootvp != rootdir.
206  */
207 int
208 lookuppnvp(
209 	struct pathname *pnp,		/* pathname to lookup */
210 	struct pathname *rpnp,		/* if non-NULL, return resolved path */
211 	int flags,			/* follow symlinks */
212 	vnode_t **dirvpp,		/* ptr for parent vnode */
213 	vnode_t **compvpp,		/* ptr for entry vnode */
214 	vnode_t *rootvp,		/* rootvp */
215 	vnode_t *vp,			/* directory to start search at */
216 	cred_t *cr)			/* user's credential */
217 {
218 	vnode_t *cvp;	/* current component vp */
219 	vnode_t *tvp;	/* addressable temp ptr */
220 	char component[MAXNAMELEN];	/* buffer for component (incl null) */
221 	int error;
222 	int nlink;
223 	int lookup_flags;
224 	struct pathname presrvd; /* case preserved name */
225 	struct pathname *pp = NULL;
226 	vnode_t *startvp;
227 	vnode_t *zonevp = curproc->p_zone->zone_rootvp;		/* zone root */
228 	int must_be_directory = 0;
229 	boolean_t retry_with_kcred;
230 	uint32_t auditing = AU_AUDITING();
231 
232 	CPU_STATS_ADDQ(CPU, sys, namei, 1);
233 	nlink = 0;
234 	cvp = NULL;
235 	if (rpnp)
236 		rpnp->pn_pathlen = 0;
237 
238 	lookup_flags = dirvpp ? LOOKUP_DIR : 0;
239 	if (flags & FIGNORECASE) {
240 		lookup_flags |= FIGNORECASE;
241 		pn_alloc(&presrvd);
242 		pp = &presrvd;
243 	}
244 
245 	if (auditing)
246 		audit_anchorpath(pnp, vp == rootvp);
247 
248 	/*
249 	 * Eliminate any trailing slashes in the pathname.
250 	 * If there are any, we must follow all symlinks.
251 	 * Also, we must guarantee that the last component is a directory.
252 	 */
253 	if (pn_fixslash(pnp)) {
254 		flags |= FOLLOW;
255 		must_be_directory = 1;
256 	}
257 
258 	startvp = vp;
259 next:
260 	retry_with_kcred = B_FALSE;
261 
262 	/*
263 	 * Make sure we have a directory.
264 	 */
265 	if (vp->v_type != VDIR) {
266 		error = ENOTDIR;
267 		goto bad;
268 	}
269 
270 	if (rpnp && VN_CMP(vp, rootvp))
271 		(void) pn_set(rpnp, "/");
272 
273 	/*
274 	 * Process the next component of the pathname.
275 	 */
276 	if (error = pn_getcomponent(pnp, component)) {
277 		goto bad;
278 	}
279 
280 	/*
281 	 * Handle "..": two special cases.
282 	 * 1. If we're at the root directory (e.g. after chroot or
283 	 *    zone_enter) then change ".." to "." so we can't get
284 	 *    out of this subtree.
285 	 * 2. If this vnode is the root of a mounted file system,
286 	 *    then replace it with the vnode that was mounted on
287 	 *    so that we take the ".." in the other file system.
288 	 */
289 	if (component[0] == '.' && component[1] == '.' && component[2] == 0) {
290 checkforroot:
291 		if (VN_CMP(vp, rootvp) || VN_CMP(vp, zonevp)) {
292 			component[1] = '\0';
293 		} else if (vp->v_flag & VROOT) {
294 			vfs_t *vfsp;
295 			cvp = vp;
296 
297 			/*
298 			 * While we deal with the vfs pointer from the vnode
299 			 * the filesystem could have been forcefully unmounted
300 			 * and the vnode's v_vfsp could have been invalidated
301 			 * by VFS_UNMOUNT. Hence, we cache v_vfsp and use it
302 			 * with vfs_rlock_wait/vfs_unlock.
303 			 * It is safe to use the v_vfsp even it is freed by
304 			 * VFS_UNMOUNT because vfs_rlock_wait/vfs_unlock
305 			 * do not dereference v_vfsp. It is just used as a
306 			 * magic cookie.
307 			 * One more corner case here is the memory getting
308 			 * reused for another vfs structure. In this case
309 			 * lookuppnvp's vfs_rlock_wait will succeed, domount's
310 			 * vfs_lock will fail and domount will bail out with an
311 			 * error (EBUSY).
312 			 */
313 			vfsp = cvp->v_vfsp;
314 
315 			/*
316 			 * This lock is used to synchronize
317 			 * mounts/unmounts and lookups.
318 			 * Threads doing mounts/unmounts hold the
319 			 * writers version vfs_lock_wait().
320 			 */
321 
322 			vfs_rlock_wait(vfsp);
323 
324 			/*
325 			 * If this vnode is on a file system that
326 			 * has been forcibly unmounted,
327 			 * we can't proceed. Cancel this operation
328 			 * and return EIO.
329 			 *
330 			 * vfs_vnodecovered is NULL if unmounted.
331 			 * Currently, nfs uses VFS_UNMOUNTED to
332 			 * check if it's a forced-umount. Keep the
333 			 * same checking here as well even though it
334 			 * may not be needed.
335 			 */
336 			if (((vp = cvp->v_vfsp->vfs_vnodecovered) == NULL) ||
337 			    (cvp->v_vfsp->vfs_flag & VFS_UNMOUNTED)) {
338 				vfs_unlock(vfsp);
339 				VN_RELE(cvp);
340 				if (pp)
341 					pn_free(pp);
342 				return (EIO);
343 			}
344 			VN_HOLD(vp);
345 			vfs_unlock(vfsp);
346 			VN_RELE(cvp);
347 			cvp = NULL;
348 			/*
349 			 * Crossing mount points. For eg: We are doing
350 			 * a lookup of ".." for file systems root vnode
351 			 * mounted here, and VOP_LOOKUP() (with covered vnode)
352 			 * will be on underlying file systems mount point
353 			 * vnode. Set retry_with_kcred flag as we might end
354 			 * up doing VOP_LOOKUP() with kcred if required.
355 			 */
356 			retry_with_kcred = B_TRUE;
357 			goto checkforroot;
358 		}
359 	}
360 
361 	/*
362 	 * LOOKUP_CHECKREAD is a private flag used by vnodetopath() to indicate
363 	 * that we need to have read permission on every directory in the entire
364 	 * path.  This is used to ensure that a forward-lookup of a cached value
365 	 * has the same effect as a reverse-lookup when the cached value cannot
366 	 * be found.
367 	 */
368 	if ((flags & LOOKUP_CHECKREAD) &&
369 	    (error = VOP_ACCESS(vp, VREAD, 0, cr, NULL)) != 0)
370 		goto bad;
371 
372 	/*
373 	 * Perform a lookup in the current directory.
374 	 */
375 	error = VOP_LOOKUP(vp, component, &tvp, pnp, lookup_flags,
376 	    rootvp, cr, NULL, NULL, pp);
377 
378 	/*
379 	 * Retry with kcred - If crossing mount points & error is EACCES.
380 	 *
381 	 * If we are crossing mount points here and doing ".." lookup,
382 	 * VOP_LOOKUP() might fail if the underlying file systems
383 	 * mount point has no execute permission. In cases like these,
384 	 * we retry VOP_LOOKUP() by giving as much privilage as possible
385 	 * by passing kcred credentials.
386 	 *
387 	 * In case of hierarchical file systems, passing kcred still may
388 	 * or may not work.
389 	 * For eg: UFS FS --> Mount NFS FS --> Again mount UFS on some
390 	 *			directory inside NFS FS.
391 	 */
392 	if ((error == EACCES) && retry_with_kcred)
393 		error = VOP_LOOKUP(vp, component, &tvp, pnp, lookup_flags,
394 		    rootvp, zone_kcred(), NULL, NULL, pp);
395 
396 	cvp = tvp;
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 		tvp = cvp;
443 		if ((error = traverse(&tvp)) != 0) {
444 			/*
445 			 * It is required to assign cvp here, because
446 			 * traverse() will return a held vnode which
447 			 * may different than the vnode that was passed
448 			 * in (even in the error case).  If traverse()
449 			 * changes the vnode it releases the original,
450 			 * and holds the new one.
451 			 */
452 			cvp = tvp;
453 			goto bad;
454 		}
455 		cvp = tvp;
456 	}
457 
458 	/*
459 	 * If we hit a symbolic link and there is more path to be
460 	 * translated or this operation does not wish to apply
461 	 * to a link, then place the contents of the link at the
462 	 * front of the remaining pathname.
463 	 */
464 	if (cvp->v_type == VLNK && ((flags & FOLLOW) || pn_pathleft(pnp))) {
465 		struct pathname linkpath;
466 
467 		if (++nlink > MAXSYMLINKS) {
468 			error = ELOOP;
469 			goto bad;
470 		}
471 		pn_alloc(&linkpath);
472 		if (error = pn_getsymlink(cvp, &linkpath, cr)) {
473 			pn_free(&linkpath);
474 			goto bad;
475 		}
476 
477 		if (auditing)
478 			audit_symlink(pnp, &linkpath);
479 
480 		if (pn_pathleft(&linkpath) == 0)
481 			(void) pn_set(&linkpath, ".");
482 		error = pn_insert(pnp, &linkpath, strlen(component));
483 		pn_free(&linkpath);
484 		if (error)
485 			goto bad;
486 		VN_RELE(cvp);
487 		cvp = NULL;
488 		if (pnp->pn_pathlen == 0) {
489 			error = ENOENT;
490 			goto bad;
491 		}
492 		if (pnp->pn_path[0] == '/') {
493 			do {
494 				pnp->pn_path++;
495 				pnp->pn_pathlen--;
496 			} while (pnp->pn_path[0] == '/');
497 			VN_RELE(vp);
498 			vp = rootvp;
499 			VN_HOLD(vp);
500 		}
501 		if (auditing)
502 			audit_anchorpath(pnp, vp == rootvp);
503 		if (pn_fixslash(pnp)) {
504 			flags |= FOLLOW;
505 			must_be_directory = 1;
506 		}
507 		goto next;
508 	}
509 
510 	/*
511 	 * If rpnp is non-NULL, remember the resolved path name therein.
512 	 * Do not include "." components.  Collapse occurrences of
513 	 * "previous/..", so long as "previous" is not itself "..".
514 	 * Exhausting rpnp results in error ENAMETOOLONG.
515 	 */
516 	if (rpnp && strcmp(component, ".") != 0) {
517 		size_t len;
518 
519 		if (strcmp(component, "..") == 0 &&
520 		    rpnp->pn_pathlen != 0 &&
521 		    !((rpnp->pn_pathlen > 2 &&
522 		    strncmp(rpnp->pn_path+rpnp->pn_pathlen-3, "/..", 3) == 0) ||
523 		    (rpnp->pn_pathlen == 2 &&
524 		    strncmp(rpnp->pn_path, "..", 2) == 0))) {
525 			while (rpnp->pn_pathlen &&
526 			    rpnp->pn_path[rpnp->pn_pathlen-1] != '/')
527 				rpnp->pn_pathlen--;
528 			if (rpnp->pn_pathlen > 1)
529 				rpnp->pn_pathlen--;
530 			rpnp->pn_path[rpnp->pn_pathlen] = '\0';
531 		} else {
532 			if (rpnp->pn_pathlen != 0 &&
533 			    rpnp->pn_path[rpnp->pn_pathlen-1] != '/')
534 				rpnp->pn_path[rpnp->pn_pathlen++] = '/';
535 			if (flags & FIGNORECASE) {
536 				/*
537 				 * Return the case-preserved name
538 				 * within the resolved path.
539 				 */
540 				error = copystr(pp->pn_buf,
541 				    rpnp->pn_path + rpnp->pn_pathlen,
542 				    rpnp->pn_bufsize - rpnp->pn_pathlen, &len);
543 			} else {
544 				error = copystr(component,
545 				    rpnp->pn_path + rpnp->pn_pathlen,
546 				    rpnp->pn_bufsize - rpnp->pn_pathlen, &len);
547 			}
548 			if (error)	/* copystr() returns ENAMETOOLONG */
549 				goto bad;
550 			rpnp->pn_pathlen += (len - 1);
551 			ASSERT(rpnp->pn_bufsize > rpnp->pn_pathlen);
552 		}
553 	}
554 
555 	/*
556 	 * If no more components, return last directory (if wanted) and
557 	 * last component (if wanted).
558 	 */
559 	if (pn_pathleft(pnp) == 0) {
560 		/*
561 		 * If there was a trailing slash in the pathname,
562 		 * make sure the last component is a directory.
563 		 */
564 		if (must_be_directory && cvp->v_type != VDIR) {
565 			error = ENOTDIR;
566 			goto bad;
567 		}
568 		if (dirvpp != NULL) {
569 			/*
570 			 * Check that we have the real parent and not
571 			 * an alias of the last component.
572 			 */
573 			if (vn_compare(vp, cvp)) {
574 				if (auditing)
575 					(void) audit_savepath(pnp, cvp, vp,
576 					    EINVAL, cr);
577 				pn_setlast(pnp);
578 				VN_RELE(vp);
579 				VN_RELE(cvp);
580 				if (rootvp != rootdir)
581 					VN_RELE(rootvp);
582 				if (pp)
583 					pn_free(pp);
584 				return (EINVAL);
585 			}
586 			*dirvpp = vp;
587 		} else
588 			VN_RELE(vp);
589 		if (auditing)
590 			(void) audit_savepath(pnp, cvp, vp, 0, cr);
591 		if (pnp->pn_path == pnp->pn_buf)
592 			(void) pn_set(pnp, ".");
593 		else
594 			pn_setlast(pnp);
595 		if (rpnp) {
596 			if (VN_CMP(cvp, rootvp))
597 				(void) pn_set(rpnp, "/");
598 			else if (rpnp->pn_pathlen == 0)
599 				(void) pn_set(rpnp, ".");
600 		}
601 
602 		if (compvpp != NULL)
603 			*compvpp = cvp;
604 		else
605 			VN_RELE(cvp);
606 		if (rootvp != rootdir)
607 			VN_RELE(rootvp);
608 		if (pp)
609 			pn_free(pp);
610 		return (0);
611 	}
612 
613 	/*
614 	 * Skip over slashes from end of last component.
615 	 */
616 	while (pnp->pn_path[0] == '/') {
617 		pnp->pn_path++;
618 		pnp->pn_pathlen--;
619 	}
620 
621 	/*
622 	 * Searched through another level of directory:
623 	 * release previous directory handle and save new (result
624 	 * of lookup) as current directory.
625 	 */
626 	VN_RELE(vp);
627 	vp = cvp;
628 	cvp = NULL;
629 	goto next;
630 
631 bad:
632 	if (auditing)	/* reached end of path */
633 		(void) audit_savepath(pnp, cvp, vp, error, cr);
634 bad_noaudit:
635 	/*
636 	 * Error.  Release vnodes and return.
637 	 */
638 	if (cvp)
639 		VN_RELE(cvp);
640 	/*
641 	 * If the error was ESTALE and the current directory to look in
642 	 * was the root for this lookup, the root for a mounted file
643 	 * system, or the starting directory for lookups, then
644 	 * return ENOENT instead of ESTALE.  In this case, no recovery
645 	 * is possible by the higher level.  If ESTALE was returned for
646 	 * some intermediate directory along the path, then recovery
647 	 * is potentially possible and retrying from the higher level
648 	 * will either correct the situation by purging stale cache
649 	 * entries or eventually get back to the point where no recovery
650 	 * is possible.
651 	 */
652 	if (error == ESTALE &&
653 	    (VN_CMP(vp, rootvp) || (vp->v_flag & VROOT) || vp == startvp))
654 		error = ENOENT;
655 	VN_RELE(vp);
656 	if (rootvp != rootdir)
657 		VN_RELE(rootvp);
658 	if (pp)
659 		pn_free(pp);
660 	return (error);
661 }
662 
663 /*
664  * Traverse a mount point.  Routine accepts a vnode pointer as a reference
665  * parameter and performs the indirection, releasing the original vnode.
666  */
667 int
668 traverse(vnode_t **cvpp)
669 {
670 	int error = 0;
671 	vnode_t *cvp;
672 	vnode_t *tvp;
673 	vfs_t *vfsp;
674 
675 	cvp = *cvpp;
676 
677 	/*
678 	 * If this vnode is mounted on, then we transparently indirect
679 	 * to the vnode which is the root of the mounted file system.
680 	 * Before we do this we must check that an unmount is not in
681 	 * progress on this vnode.
682 	 */
683 
684 	for (;;) {
685 		/*
686 		 * Try to read lock the vnode.  If this fails because
687 		 * the vnode is already write locked, then check to
688 		 * see whether it is the current thread which locked
689 		 * the vnode.  If it is not, then read lock the vnode
690 		 * by waiting to acquire the lock.
691 		 *
692 		 * The code path in domount() is an example of support
693 		 * which needs to look up two pathnames and locks one
694 		 * of them in between the two lookups.
695 		 */
696 		error = vn_vfsrlock(cvp);
697 		if (error) {
698 			if (!vn_vfswlock_held(cvp))
699 				error = vn_vfsrlock_wait(cvp);
700 			if (error != 0) {
701 				/*
702 				 * lookuppn() expects a held vnode to be
703 				 * returned because it promptly calls
704 				 * VN_RELE after the error return
705 				 */
706 				*cvpp = cvp;
707 				return (error);
708 			}
709 		}
710 
711 		/*
712 		 * Reached the end of the mount chain?
713 		 */
714 		vfsp = vn_mountedvfs(cvp);
715 		if (vfsp == NULL) {
716 			vn_vfsunlock(cvp);
717 			break;
718 		}
719 
720 		/*
721 		 * The read lock must be held across the call to VFS_ROOT() to
722 		 * prevent a concurrent unmount from destroying the vfs.
723 		 */
724 		error = VFS_ROOT(vfsp, &tvp);
725 		vn_vfsunlock(cvp);
726 
727 		if (error)
728 			break;
729 
730 		VN_RELE(cvp);
731 
732 		cvp = tvp;
733 	}
734 
735 	*cvpp = cvp;
736 	return (error);
737 }
738 
739 /*
740  * Return the lowermost vnode if this is a mountpoint.
741  */
742 static vnode_t *
743 vn_under(vnode_t *vp)
744 {
745 	vnode_t *uvp;
746 	vfs_t *vfsp;
747 
748 	while (vp->v_flag & VROOT) {
749 
750 		vfsp = vp->v_vfsp;
751 		vfs_rlock_wait(vfsp);
752 		if ((uvp = vfsp->vfs_vnodecovered) == NULL ||
753 		    (vfsp->vfs_flag & VFS_UNMOUNTED)) {
754 			vfs_unlock(vfsp);
755 			break;
756 		}
757 		VN_HOLD(uvp);
758 		vfs_unlock(vfsp);
759 		VN_RELE(vp);
760 		vp = uvp;
761 	}
762 
763 	return (vp);
764 }
765 
766 static int
767 vnode_match(vnode_t *v1, vnode_t *v2, cred_t *cr)
768 {
769 	vattr_t	v1attr, v2attr;
770 
771 	/*
772 	 * If we have a device file, check to see if is a cloned open of the
773 	 * same device.  For self-cloning devices, the major numbers will match.
774 	 * For devices cloned through the 'clone' driver, the minor number of
775 	 * the source device will be the same as the major number of the cloned
776 	 * device.
777 	 */
778 	if ((v1->v_type == VCHR || v1->v_type == VBLK) &&
779 	    v1->v_type == v2->v_type) {
780 		if ((spec_is_selfclone(v1) || spec_is_selfclone(v2)) &&
781 		    getmajor(v1->v_rdev) == getmajor(v2->v_rdev))
782 			return (1);
783 
784 		if (spec_is_clone(v1) &&
785 		    getmajor(v1->v_rdev) == getminor(v2->v_rdev))
786 			return (1);
787 
788 		if (spec_is_clone(v2) &&
789 		    getmajor(v2->v_rdev) == getminor(v1->v_rdev))
790 			return (1);
791 	}
792 
793 	v1attr.va_mask = v2attr.va_mask = AT_TYPE;
794 
795 	/*
796 	 * This check for symbolic links handles the pseudo-symlinks in procfs.
797 	 * These particular links have v_type of VDIR, but the attributes have a
798 	 * type of VLNK.  We need to avoid these links because otherwise if we
799 	 * are currently in '/proc/self/fd', then '/proc/self/cwd' will compare
800 	 * as the same vnode.
801 	 */
802 	if (VOP_GETATTR(v1, &v1attr, 0, cr, NULL) != 0 ||
803 	    VOP_GETATTR(v2, &v2attr, 0, cr, NULL) != 0 ||
804 	    v1attr.va_type == VLNK || v2attr.va_type == VLNK)
805 		return (0);
806 
807 	v1attr.va_mask = v2attr.va_mask = AT_TYPE | AT_FSID | AT_NODEID;
808 
809 	if (VOP_GETATTR(v1, &v1attr, ATTR_REAL, cr, NULL) != 0 ||
810 	    VOP_GETATTR(v2, &v2attr, ATTR_REAL, cr, NULL) != 0)
811 		return (0);
812 
813 	return (v1attr.va_fsid == v2attr.va_fsid &&
814 	    v1attr.va_nodeid == v2attr.va_nodeid);
815 }
816 
817 
818 /*
819  * Find the entry in the directory corresponding to the target vnode.
820  */
821 int
822 dirfindvp(vnode_t *vrootp, vnode_t *dvp, vnode_t *tvp, cred_t *cr, char *dbuf,
823     size_t dlen, dirent64_t **rdp)
824 {
825 	size_t dbuflen;
826 	struct iovec iov;
827 	struct uio uio;
828 	int error;
829 	int eof;
830 	vnode_t *cmpvp;
831 	struct dirent64 *dp;
832 	pathname_t pnp;
833 
834 	ASSERT(dvp->v_type == VDIR);
835 
836 	/*
837 	 * This is necessary because of the strange semantics of VOP_LOOKUP().
838 	 */
839 	bzero(&pnp, sizeof (pnp));
840 
841 	eof = 0;
842 
843 	uio.uio_iov = &iov;
844 	uio.uio_iovcnt = 1;
845 	uio.uio_segflg = UIO_SYSSPACE;
846 	uio.uio_fmode = 0;
847 	uio.uio_extflg = UIO_COPY_CACHED;
848 	uio.uio_loffset = 0;
849 
850 	if ((error = VOP_ACCESS(dvp, VREAD, 0, cr, NULL)) != 0)
851 		return (error);
852 
853 	while (!eof) {
854 		uio.uio_resid = dlen;
855 		iov.iov_base = dbuf;
856 		iov.iov_len = dlen;
857 
858 		(void) VOP_RWLOCK(dvp, V_WRITELOCK_FALSE, NULL);
859 		error = VOP_READDIR(dvp, &uio, cr, &eof, NULL, 0);
860 		VOP_RWUNLOCK(dvp, V_WRITELOCK_FALSE, NULL);
861 
862 		dbuflen = dlen - uio.uio_resid;
863 
864 		if (error || dbuflen == 0)
865 			break;
866 
867 		dp = (dirent64_t *)dbuf;
868 		while ((intptr_t)dp < (intptr_t)dbuf + dbuflen) {
869 			/*
870 			 * Ignore '.' and '..' entries
871 			 */
872 			if (strcmp(dp->d_name, ".") == 0 ||
873 			    strcmp(dp->d_name, "..") == 0) {
874 				dp = (dirent64_t *)((intptr_t)dp +
875 				    dp->d_reclen);
876 				continue;
877 			}
878 
879 			error = VOP_LOOKUP(dvp, dp->d_name, &cmpvp, &pnp, 0,
880 			    vrootp, cr, NULL, NULL, NULL);
881 
882 			/*
883 			 * We only want to bail out if there was an error other
884 			 * than ENOENT.  Otherwise, it could be that someone
885 			 * just removed an entry since the readdir() call, and
886 			 * the entry we want is further on in the directory.
887 			 */
888 			if (error == 0) {
889 				if (vnode_match(tvp, cmpvp, cr)) {
890 					VN_RELE(cmpvp);
891 					*rdp = dp;
892 					return (0);
893 				}
894 
895 				VN_RELE(cmpvp);
896 			} else if (error != ENOENT) {
897 				return (error);
898 			}
899 
900 			dp = (dirent64_t *)((intptr_t)dp + dp->d_reclen);
901 		}
902 	}
903 
904 	/*
905 	 * Something strange has happened, this directory does not contain the
906 	 * specified vnode.  This should never happen in the normal case, since
907 	 * we ensured that dvp is the parent of vp.  This is possible in some
908 	 * rare conditions (races and the special .zfs directory).
909 	 */
910 	if (error == 0) {
911 		error = VOP_LOOKUP(dvp, ".zfs", &cmpvp, &pnp, 0, vrootp, cr,
912 		    NULL, NULL, NULL);
913 		if (error == 0) {
914 			if (vnode_match(tvp, cmpvp, cr)) {
915 				(void) strcpy(dp->d_name, ".zfs");
916 				dp->d_reclen = strlen(".zfs");
917 				dp->d_off = 2;
918 				dp->d_ino = 1;
919 				*rdp = dp;
920 			} else {
921 				error = ENOENT;
922 			}
923 			VN_RELE(cmpvp);
924 		}
925 	}
926 
927 	return (error);
928 }
929 
930 /*
931  * Given a global path (from rootdir), and a vnode that is the current root,
932  * return the portion of the path that is beneath the current root or NULL on
933  * failure.  The path MUST be a resolved path (no '..' entries or symlinks),
934  * otherwise this function will fail.
935  */
936 static char *
937 localpath(char *path, struct vnode *vrootp, cred_t *cr)
938 {
939 	vnode_t *vp;
940 	vnode_t *cvp;
941 	char component[MAXNAMELEN];
942 	char *ret = NULL;
943 	pathname_t pn;
944 
945 	/*
946 	 * We use vn_compare() instead of VN_CMP() in order to detect lofs
947 	 * mounts and stacked vnodes.
948 	 */
949 	if (vn_compare(vrootp, rootdir))
950 		return (path);
951 
952 	if (pn_get(path, UIO_SYSSPACE, &pn) != 0)
953 		return (NULL);
954 
955 	vp = rootdir;
956 	VN_HOLD(vp);
957 
958 	if (vn_ismntpt(vp) && traverse(&vp) != 0) {
959 		VN_RELE(vp);
960 		pn_free(&pn);
961 		return (NULL);
962 	}
963 
964 	while (pn_pathleft(&pn)) {
965 		pn_skipslash(&pn);
966 
967 		if (pn_getcomponent(&pn, component) != 0)
968 			break;
969 
970 		if (VOP_LOOKUP(vp, component, &cvp, &pn, 0, rootdir, cr,
971 		    NULL, NULL, NULL) != 0)
972 			break;
973 		VN_RELE(vp);
974 		vp = cvp;
975 
976 		if (vn_ismntpt(vp) && traverse(&vp) != 0)
977 			break;
978 
979 		if (vn_compare(vp, vrootp)) {
980 			ret = path + (pn.pn_path - pn.pn_buf);
981 			break;
982 		}
983 	}
984 
985 	VN_RELE(vp);
986 	pn_free(&pn);
987 
988 	return (ret);
989 }
990 
991 /*
992  * Given a directory, return the full, resolved path.  This looks up "..",
993  * searches for the given vnode in the parent, appends the component, etc.  It
994  * is used to implement vnodetopath() and getcwd() when the cached path fails.
995  */
996 static int
997 dirtopath(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen, int flags,
998     cred_t *cr)
999 {
1000 	pathname_t pn, rpn, emptypn;
1001 	vnode_t *cmpvp, *pvp = NULL;
1002 	vnode_t *startvp = vp;
1003 	int err = 0, vprivs;
1004 	size_t complen;
1005 	char *dbuf;
1006 	dirent64_t *dp;
1007 	char		*bufloc;
1008 	size_t		dlen = DIRENT64_RECLEN(MAXPATHLEN);
1009 	refstr_t	*mntpt;
1010 
1011 	/* Operation only allowed on directories */
1012 	ASSERT(vp->v_type == VDIR);
1013 
1014 	/* We must have at least enough space for "/" */
1015 	if (buflen < 2)
1016 		return (ENAMETOOLONG);
1017 
1018 	/* Start at end of string with terminating null */
1019 	bufloc = &buf[buflen - 1];
1020 	*bufloc = '\0';
1021 
1022 	pn_alloc(&pn);
1023 	pn_alloc(&rpn);
1024 	dbuf = kmem_alloc(dlen, KM_SLEEP);
1025 	bzero(&emptypn, sizeof (emptypn));
1026 
1027 	/*
1028 	 * Begin with an additional reference on vp.  This will be decremented
1029 	 * during the loop.
1030 	 */
1031 	VN_HOLD(vp);
1032 
1033 	for (;;) {
1034 		/*
1035 		 * Return if we've reached the root.  If the buffer is empty,
1036 		 * return '/'.  We explicitly don't use vn_compare(), since it
1037 		 * compares the real vnodes.  A lofs mount of '/' would produce
1038 		 * incorrect results otherwise.
1039 		 */
1040 		if (VN_CMP(vrootp, vp)) {
1041 			if (*bufloc == '\0')
1042 				*--bufloc = '/';
1043 			break;
1044 		}
1045 
1046 		/*
1047 		 * If we've reached the VFS root, something has gone wrong.  We
1048 		 * should have reached the root in the above check.  The only
1049 		 * explantation is that 'vp' is not contained withing the given
1050 		 * root, in which case we return EPERM.
1051 		 */
1052 		if (VN_CMP(rootdir, vp)) {
1053 			err = EPERM;
1054 			goto out;
1055 		}
1056 
1057 		/*
1058 		 * Shortcut: see if this vnode is a mountpoint.  If so,
1059 		 * grab the path information from the vfs_t.
1060 		 */
1061 		if (vp->v_flag & VROOT) {
1062 
1063 			mntpt = vfs_getmntpoint(vp->v_vfsp);
1064 			if ((err = pn_set(&pn, (char *)refstr_value(mntpt)))
1065 			    == 0) {
1066 				refstr_rele(mntpt);
1067 				rpn.pn_path = rpn.pn_buf;
1068 
1069 				/*
1070 				 * Ensure the mountpoint still exists.
1071 				 */
1072 				VN_HOLD(vrootp);
1073 				if (vrootp != rootdir)
1074 					VN_HOLD(vrootp);
1075 				if (lookuppnvp(&pn, &rpn, flags, NULL,
1076 				    &cmpvp, vrootp, vrootp, cr) == 0) {
1077 
1078 					if (VN_CMP(vp, cmpvp)) {
1079 						VN_RELE(cmpvp);
1080 
1081 						complen = strlen(rpn.pn_path);
1082 						bufloc -= complen;
1083 						if (bufloc < buf) {
1084 							err = ERANGE;
1085 							goto out;
1086 						}
1087 						bcopy(rpn.pn_path, bufloc,
1088 						    complen);
1089 						break;
1090 					} else {
1091 						VN_RELE(cmpvp);
1092 					}
1093 				}
1094 			} else {
1095 				refstr_rele(mntpt);
1096 			}
1097 		}
1098 
1099 		/*
1100 		 * Shortcut: see if this vnode has correct v_path. If so,
1101 		 * we have the work done.
1102 		 */
1103 		mutex_enter(&vp->v_lock);
1104 		if (vp->v_path != NULL) {
1105 
1106 			if ((err = pn_set(&pn, vp->v_path)) == 0) {
1107 				mutex_exit(&vp->v_lock);
1108 				rpn.pn_path = rpn.pn_buf;
1109 
1110 				/*
1111 				 * Ensure the v_path pointing to correct vnode
1112 				 */
1113 				VN_HOLD(vrootp);
1114 				if (vrootp != rootdir)
1115 					VN_HOLD(vrootp);
1116 				if (lookuppnvp(&pn, &rpn, flags, NULL,
1117 				    &cmpvp, vrootp, vrootp, cr) == 0) {
1118 
1119 					if (VN_CMP(vp, cmpvp)) {
1120 						VN_RELE(cmpvp);
1121 
1122 						complen = strlen(rpn.pn_path);
1123 						bufloc -= complen;
1124 						if (bufloc < buf) {
1125 							err = ERANGE;
1126 							goto out;
1127 						}
1128 						bcopy(rpn.pn_path, bufloc,
1129 						    complen);
1130 						break;
1131 					} else {
1132 						VN_RELE(cmpvp);
1133 					}
1134 				}
1135 			} else {
1136 				mutex_exit(&vp->v_lock);
1137 			}
1138 		} else {
1139 			mutex_exit(&vp->v_lock);
1140 		}
1141 
1142 		/*
1143 		 * Shortcuts failed, search for this vnode in its parent.  If
1144 		 * this is a mountpoint, then get the vnode underneath.
1145 		 */
1146 		if (vp->v_flag & VROOT)
1147 			vp = vn_under(vp);
1148 		if ((err = VOP_LOOKUP(vp, "..", &pvp, &emptypn, 0, vrootp, cr,
1149 		    NULL, NULL, NULL)) != 0)
1150 			goto out;
1151 
1152 		/*
1153 		 * With extended attributes, it's possible for a directory to
1154 		 * have a parent that is a regular file.  Check for that here.
1155 		 */
1156 		if (pvp->v_type != VDIR) {
1157 			err = ENOTDIR;
1158 			goto out;
1159 		}
1160 
1161 		/*
1162 		 * If this is true, something strange has happened.  This is
1163 		 * only true if we are the root of a filesystem, which should
1164 		 * have been caught by the check above.
1165 		 */
1166 		if (VN_CMP(pvp, vp)) {
1167 			err = ENOENT;
1168 			goto out;
1169 		}
1170 
1171 		/*
1172 		 * Check if we have read and search privilege so, that
1173 		 * we can lookup the path in the directory
1174 		 */
1175 		vprivs = (flags & LOOKUP_CHECKREAD) ? VREAD | VEXEC : VEXEC;
1176 		if ((err = VOP_ACCESS(pvp, vprivs, 0, cr, NULL)) != 0) {
1177 			goto out;
1178 		}
1179 
1180 		/*
1181 		 * Try to obtain the path component from dnlc cache
1182 		 * before searching through the directory.
1183 		 */
1184 		if ((cmpvp = dnlc_reverse_lookup(vp, dbuf, dlen)) != NULL) {
1185 			/*
1186 			 * If we got parent vnode as a result,
1187 			 * then the answered path is correct.
1188 			 */
1189 			if (VN_CMP(cmpvp, pvp)) {
1190 				VN_RELE(cmpvp);
1191 				complen = strlen(dbuf);
1192 				bufloc -= complen;
1193 				if (bufloc <= buf) {
1194 					err = ENAMETOOLONG;
1195 					goto out;
1196 				}
1197 				bcopy(dbuf, bufloc, complen);
1198 
1199 				/* Prepend a slash to the current path */
1200 				*--bufloc = '/';
1201 
1202 				/* And continue with the next component */
1203 				VN_RELE(vp);
1204 				vp = pvp;
1205 				pvp = NULL;
1206 				continue;
1207 			} else {
1208 				VN_RELE(cmpvp);
1209 			}
1210 		}
1211 
1212 		/*
1213 		 * Search the parent directory for the entry corresponding to
1214 		 * this vnode.
1215 		 */
1216 		if ((err = dirfindvp(vrootp, pvp, vp, cr, dbuf, dlen, &dp))
1217 		    != 0)
1218 			goto out;
1219 		complen = strlen(dp->d_name);
1220 		bufloc -= complen;
1221 		if (bufloc <= buf) {
1222 			err = ENAMETOOLONG;
1223 			goto out;
1224 		}
1225 		bcopy(dp->d_name, bufloc, complen);
1226 
1227 		/* Prepend a slash to the current path.  */
1228 		*--bufloc = '/';
1229 
1230 		/* And continue with the next component */
1231 		VN_RELE(vp);
1232 		vp = pvp;
1233 		pvp = NULL;
1234 	}
1235 
1236 	/*
1237 	 * Place the path at the beginning of the buffer.
1238 	 */
1239 	if (bufloc != buf)
1240 		ovbcopy(bufloc, buf, buflen - (bufloc - buf));
1241 
1242 out:
1243 	/*
1244 	 * If the error was ESTALE and the current directory to look in
1245 	 * was the root for this lookup, the root for a mounted file
1246 	 * system, or the starting directory for lookups, then
1247 	 * return ENOENT instead of ESTALE.  In this case, no recovery
1248 	 * is possible by the higher level.  If ESTALE was returned for
1249 	 * some intermediate directory along the path, then recovery
1250 	 * is potentially possible and retrying from the higher level
1251 	 * will either correct the situation by purging stale cache
1252 	 * entries or eventually get back to the point where no recovery
1253 	 * is possible.
1254 	 */
1255 	if (err == ESTALE &&
1256 	    (VN_CMP(vp, vrootp) || (vp->v_flag & VROOT) || vp == startvp))
1257 		err = ENOENT;
1258 
1259 	kmem_free(dbuf, dlen);
1260 	VN_RELE(vp);
1261 	if (pvp)
1262 		VN_RELE(pvp);
1263 	pn_free(&pn);
1264 	pn_free(&rpn);
1265 
1266 	return (err);
1267 }
1268 
1269 /*
1270  * The additional flag, LOOKUP_CHECKREAD, is used to enforce artificial
1271  * constraints in order to be standards compliant.  For example, if we have
1272  * the cached path of '/foo/bar', and '/foo' has permissions 100 (execute
1273  * only), then we can legitimately look up the path to the current working
1274  * directory without needing read permission.  Existing standards tests,
1275  * however, assume that we are determining the path by repeatedly looking up
1276  * "..".  We need to keep this behavior in order to maintain backwards
1277  * compatibility.
1278  */
1279 static int
1280 vnodetopath_common(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen,
1281     cred_t *cr, int flags)
1282 {
1283 	pathname_t pn, rpn;
1284 	int ret, len;
1285 	vnode_t *compvp, *pvp, *realvp;
1286 	proc_t *p = curproc;
1287 	char path[MAXNAMELEN];
1288 	int doclose = 0;
1289 
1290 	/*
1291 	 * If vrootp is NULL, get the root for curproc.  Callers with any other
1292 	 * requirements should pass in a different vrootp.
1293 	 */
1294 	if (vrootp == NULL) {
1295 		mutex_enter(&p->p_lock);
1296 		if ((vrootp = PTOU(p)->u_rdir) == NULL)
1297 			vrootp = rootdir;
1298 		VN_HOLD(vrootp);
1299 		mutex_exit(&p->p_lock);
1300 	} else {
1301 		VN_HOLD(vrootp);
1302 	}
1303 
1304 	/*
1305 	 * This is to get around an annoying artifact of the /proc filesystem,
1306 	 * which is the behavior of {cwd/root}.  Trying to resolve this path
1307 	 * will result in /proc/pid/cwd instead of whatever the real working
1308 	 * directory is.  We can't rely on VOP_REALVP(), since that will break
1309 	 * lofs.  The only difference between procfs and lofs is that opening
1310 	 * the file will return the underling vnode in the case of procfs.
1311 	 */
1312 	if (vp->v_type == VDIR && VOP_REALVP(vp, &realvp, NULL) == 0 &&
1313 	    realvp != vp) {
1314 		VN_HOLD(vp);
1315 		if (VOP_OPEN(&vp, FREAD, cr, NULL) == 0)
1316 			doclose = 1;
1317 		else
1318 			VN_RELE(vp);
1319 	}
1320 
1321 	pn_alloc(&pn);
1322 
1323 	/*
1324 	 * Check to see if we have a cached path in the vnode.
1325 	 */
1326 	mutex_enter(&vp->v_lock);
1327 	if (vp->v_path != NULL) {
1328 		(void) pn_set(&pn, vp->v_path);
1329 		mutex_exit(&vp->v_lock);
1330 
1331 		pn_alloc(&rpn);
1332 
1333 		/* We should only cache absolute paths */
1334 		ASSERT(pn.pn_buf[0] == '/');
1335 
1336 		/*
1337 		 * If we are in a zone or a chroot environment, then we have to
1338 		 * take additional steps, since the path to the root might not
1339 		 * be readable with the current credentials, even though the
1340 		 * process can legitmately access the file.  In this case, we
1341 		 * do the following:
1342 		 *
1343 		 * lookuppnvp() with all privileges to get the resolved path.
1344 		 * call localpath() to get the local portion of the path, and
1345 		 * continue as normal.
1346 		 *
1347 		 * If the the conversion to a local path fails, then we continue
1348 		 * as normal.  This is a heuristic to make process object file
1349 		 * paths available from within a zone.  Because lofs doesn't
1350 		 * support page operations, the vnode stored in the seg_t is
1351 		 * actually the underlying real vnode, not the lofs node itself.
1352 		 * Most of the time, the lofs path is the same as the underlying
1353 		 * vnode (for example, /usr/lib/libc.so.1).
1354 		 */
1355 		if (vrootp != rootdir) {
1356 			char *local = NULL;
1357 			VN_HOLD(rootdir);
1358 			if (lookuppnvp(&pn, &rpn, FOLLOW,
1359 			    NULL, &compvp, rootdir, rootdir, kcred) == 0) {
1360 				local = localpath(rpn.pn_path, vrootp,
1361 				    kcred);
1362 				VN_RELE(compvp);
1363 			}
1364 
1365 			/*
1366 			 * The original pn was changed through lookuppnvp().
1367 			 * Set it to local for next validation attempt.
1368 			 */
1369 			if (local) {
1370 				(void) pn_set(&pn, local);
1371 			} else {
1372 				goto notcached;
1373 			}
1374 		}
1375 
1376 		/*
1377 		 * We should have a local path at this point, so start the
1378 		 * search from the root of the current process.
1379 		 */
1380 		VN_HOLD(vrootp);
1381 		if (vrootp != rootdir)
1382 			VN_HOLD(vrootp);
1383 		ret = lookuppnvp(&pn, &rpn, FOLLOW | flags, NULL,
1384 		    &compvp, vrootp, vrootp, cr);
1385 		if (ret == 0) {
1386 			/*
1387 			 * Check to see if the returned vnode is the same as
1388 			 * the one we expect.  If not, give up.
1389 			 */
1390 			if (!vn_compare(vp, compvp) &&
1391 			    !vnode_match(vp, compvp, cr)) {
1392 				VN_RELE(compvp);
1393 				goto notcached;
1394 			}
1395 
1396 			VN_RELE(compvp);
1397 
1398 			/*
1399 			 * Return the result.
1400 			 */
1401 			if (buflen <= rpn.pn_pathlen)
1402 				goto notcached;
1403 
1404 			bcopy(rpn.pn_path, buf, rpn.pn_pathlen + 1);
1405 			pn_free(&pn);
1406 			pn_free(&rpn);
1407 			VN_RELE(vrootp);
1408 			if (doclose) {
1409 				(void) VOP_CLOSE(vp, FREAD, 1, 0, cr, NULL);
1410 				VN_RELE(vp);
1411 			}
1412 			return (0);
1413 		}
1414 
1415 notcached:
1416 		pn_free(&rpn);
1417 	} else {
1418 		mutex_exit(&vp->v_lock);
1419 	}
1420 
1421 	pn_free(&pn);
1422 
1423 	if (vp->v_type != VDIR) {
1424 		/*
1425 		 * If we don't have a directory, try to find it in the dnlc via
1426 		 * reverse lookup.  Once this is found, we can use the regular
1427 		 * directory search to find the full path.
1428 		 */
1429 		if ((pvp = dnlc_reverse_lookup(vp, path, MAXNAMELEN)) != NULL) {
1430 			/*
1431 			 * Check if we have read privilege so, that
1432 			 * we can lookup the path in the directory
1433 			 */
1434 			ret = 0;
1435 			if ((flags & LOOKUP_CHECKREAD)) {
1436 				ret = VOP_ACCESS(pvp, VREAD, 0, cr, NULL);
1437 			}
1438 			if (ret == 0) {
1439 				ret = dirtopath(vrootp, pvp, buf, buflen,
1440 				    flags, cr);
1441 			}
1442 			if (ret == 0) {
1443 				len = strlen(buf);
1444 				if (len + strlen(path) + 1 >= buflen) {
1445 					ret = ENAMETOOLONG;
1446 				} else {
1447 					if (buf[len - 1] != '/')
1448 						buf[len++] = '/';
1449 					bcopy(path, buf + len,
1450 					    strlen(path) + 1);
1451 				}
1452 			}
1453 
1454 			VN_RELE(pvp);
1455 		} else
1456 			ret = ENOENT;
1457 	} else
1458 		ret = dirtopath(vrootp, vp, buf, buflen, flags, cr);
1459 
1460 	VN_RELE(vrootp);
1461 	if (doclose) {
1462 		(void) VOP_CLOSE(vp, FREAD, 1, 0, cr, NULL);
1463 		VN_RELE(vp);
1464 	}
1465 
1466 	return (ret);
1467 }
1468 
1469 int
1470 vnodetopath(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen, cred_t *cr)
1471 {
1472 	return (vnodetopath_common(vrootp, vp, buf, buflen, cr, 0));
1473 }
1474 
1475 int
1476 dogetcwd(char *buf, size_t buflen)
1477 {
1478 	int ret;
1479 	vnode_t *vp;
1480 	vnode_t *compvp;
1481 	refstr_t *cwd, *oldcwd;
1482 	const char *value;
1483 	pathname_t rpnp, pnp;
1484 	proc_t *p = curproc;
1485 
1486 	/*
1487 	 * Check to see if there is a cached version of the cwd.  If so, lookup
1488 	 * the cached value and make sure it is the same vnode.
1489 	 */
1490 	mutex_enter(&p->p_lock);
1491 	if ((cwd = PTOU(p)->u_cwd) != NULL)
1492 		refstr_hold(cwd);
1493 	vp = PTOU(p)->u_cdir;
1494 	VN_HOLD(vp);
1495 	mutex_exit(&p->p_lock);
1496 
1497 	/*
1498 	 * Make sure we have permission to access the current directory.
1499 	 */
1500 	if ((ret = VOP_ACCESS(vp, VEXEC, 0, CRED(), NULL)) != 0) {
1501 		if (cwd != NULL)
1502 			refstr_rele(cwd);
1503 		VN_RELE(vp);
1504 		return (ret);
1505 	}
1506 
1507 	if (cwd) {
1508 		value = refstr_value(cwd);
1509 		if ((ret = pn_get((char *)value, UIO_SYSSPACE, &pnp)) != 0) {
1510 			refstr_rele(cwd);
1511 			VN_RELE(vp);
1512 			return (ret);
1513 		}
1514 
1515 		pn_alloc(&rpnp);
1516 
1517 		if (lookuppn(&pnp, &rpnp, NO_FOLLOW, NULL, &compvp) == 0) {
1518 
1519 			if (VN_CMP(vp, compvp) &&
1520 			    strcmp(value, rpnp.pn_path) == 0) {
1521 				VN_RELE(compvp);
1522 				VN_RELE(vp);
1523 				pn_free(&pnp);
1524 				pn_free(&rpnp);
1525 				if (strlen(value) + 1 > buflen) {
1526 					refstr_rele(cwd);
1527 					return (ENAMETOOLONG);
1528 				}
1529 				bcopy(value, buf, strlen(value) + 1);
1530 				refstr_rele(cwd);
1531 				return (0);
1532 			}
1533 
1534 			VN_RELE(compvp);
1535 		}
1536 
1537 		pn_free(&rpnp);
1538 		pn_free(&pnp);
1539 
1540 		refstr_rele(cwd);
1541 	}
1542 
1543 	ret = vnodetopath_common(NULL, vp, buf, buflen, CRED(),
1544 	    LOOKUP_CHECKREAD);
1545 
1546 	VN_RELE(vp);
1547 
1548 	/*
1549 	 * Store the new cwd and replace the existing cached copy.
1550 	 */
1551 	if (ret == 0)
1552 		cwd = refstr_alloc(buf);
1553 	else
1554 		cwd = NULL;
1555 
1556 	mutex_enter(&p->p_lock);
1557 	oldcwd = PTOU(p)->u_cwd;
1558 	PTOU(p)->u_cwd = cwd;
1559 	mutex_exit(&p->p_lock);
1560 
1561 	if (oldcwd)
1562 		refstr_rele(oldcwd);
1563 
1564 	return (ret);
1565 }
1566