xref: /illumos-gate/usr/src/uts/common/fs/lookup.c (revision 6f1fa39e3cf1b335f342bbca41590e9d76ab29b7)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
24  * Copyright (c) 1988, 2010, Oracle and/or its affiliates. All rights reserved.
25  * Copyright 2016 Joyent, Inc.
26  */
27 
28 /*	Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T	*/
29 /*	  All Rights Reserved  	*/
30 
31 /*
32  * University Copyright- Copyright (c) 1982, 1986, 1988
33  * The Regents of the University of California
34  * All Rights Reserved
35  *
36  * University Acknowledgment- Portions of this document are derived from
37  * software developed by the University of California, Berkeley, and its
38  * contributors.
39  */
40 
41 #include <sys/types.h>
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/cpuvar.h>
45 #include <sys/errno.h>
46 #include <sys/cred.h>
47 #include <sys/user.h>
48 #include <sys/uio.h>
49 #include <sys/vfs.h>
50 #include <sys/vnode.h>
51 #include <sys/pathname.h>
52 #include <sys/proc.h>
53 #include <sys/vtrace.h>
54 #include <sys/sysmacros.h>
55 #include <sys/debug.h>
56 #include <sys/dirent.h>
57 #include <c2/audit.h>
58 #include <sys/zone.h>
59 #include <sys/dnlc.h>
60 #include <sys/fs/snode.h>
61 
62 /* Controls whether paths are stored with vnodes. */
63 int vfs_vnode_path = 1;
64 
65 int
66 lookupname(
67 	char *fnamep,
68 	enum uio_seg seg,
69 	int followlink,
70 	vnode_t **dirvpp,
71 	vnode_t **compvpp)
72 {
73 	return (lookupnameatcred(fnamep, seg, followlink, dirvpp, compvpp, NULL,
74 	    CRED()));
75 }
76 
77 /*
78  * Lookup the user file name,
79  * Handle allocation and freeing of pathname buffer, return error.
80  */
81 int
82 lookupnameatcred(
83 	char *fnamep,			/* user pathname */
84 	enum uio_seg seg,		/* addr space that name is in */
85 	int followlink,			/* follow sym links */
86 	vnode_t **dirvpp,		/* ret for ptr to parent dir vnode */
87 	vnode_t **compvpp,		/* ret for ptr to component vnode */
88 	vnode_t *startvp,		/* start path search from vp */
89 	cred_t *cr)			/* credential */
90 {
91 	char namebuf[TYPICALMAXPATHLEN];
92 	struct pathname lookpn;
93 	int error;
94 
95 	error = pn_get_buf(fnamep, seg, &lookpn, namebuf, sizeof (namebuf));
96 	if (error == 0) {
97 		error = lookuppnatcred(&lookpn, NULL, followlink,
98 		    dirvpp, compvpp, startvp, cr);
99 	}
100 	if (error == ENAMETOOLONG) {
101 		/*
102 		 * This thread used a pathname > TYPICALMAXPATHLEN bytes long.
103 		 */
104 		if (error = pn_get(fnamep, seg, &lookpn))
105 			return (error);
106 		error = lookuppnatcred(&lookpn, NULL, followlink,
107 		    dirvpp, compvpp, startvp, cr);
108 		pn_free(&lookpn);
109 	}
110 
111 	return (error);
112 }
113 
114 int
115 lookupnameat(char *fnamep, enum uio_seg seg, int followlink,
116     vnode_t **dirvpp, vnode_t **compvpp, vnode_t *startvp)
117 {
118 	return (lookupnameatcred(fnamep, seg, followlink, dirvpp, compvpp,
119 	    startvp, CRED()));
120 }
121 
122 int
123 lookuppn(
124 	struct pathname *pnp,
125 	struct pathname *rpnp,
126 	int followlink,
127 	vnode_t **dirvpp,
128 	vnode_t **compvpp)
129 {
130 	return (lookuppnatcred(pnp, rpnp, followlink, dirvpp, compvpp, NULL,
131 	    CRED()));
132 }
133 
134 /*
135  * Lookup the user file name from a given vp, using a specific credential.
136  */
137 int
138 lookuppnatcred(
139 	struct pathname *pnp,		/* pathname to lookup */
140 	struct pathname *rpnp,		/* if non-NULL, return resolved path */
141 	int followlink,			/* (don't) follow sym links */
142 	vnode_t **dirvpp,		/* ptr for parent vnode */
143 	vnode_t **compvpp,		/* ptr for entry vnode */
144 	vnode_t *startvp,		/* start search from this vp */
145 	cred_t *cr)			/* user credential */
146 {
147 	vnode_t *vp;	/* current directory vp */
148 	vnode_t *rootvp;
149 	proc_t *p = curproc;
150 
151 	if (pnp->pn_pathlen == 0)
152 		return (ENOENT);
153 
154 	mutex_enter(&p->p_lock);	/* for u_rdir and u_cdir */
155 	if ((rootvp = PTOU(p)->u_rdir) == NULL)
156 		rootvp = rootdir;
157 	else if (rootvp != rootdir)	/* no need to VN_HOLD rootdir */
158 		VN_HOLD(rootvp);
159 
160 	if (pnp->pn_path[0] == '/') {
161 		vp = rootvp;
162 	} else {
163 		vp = (startvp == NULL) ? PTOU(p)->u_cdir : startvp;
164 	}
165 	VN_HOLD(vp);
166 	mutex_exit(&p->p_lock);
167 
168 	/*
169 	 * Skip over leading slashes
170 	 */
171 	if (pnp->pn_path[0] == '/') {
172 		do {
173 			pnp->pn_path++;
174 			pnp->pn_pathlen--;
175 		} while (pnp->pn_path[0] == '/');
176 	}
177 
178 	return (lookuppnvp(pnp, rpnp, followlink, dirvpp,
179 	    compvpp, rootvp, vp, cr));
180 }
181 
182 int
183 lookuppnat(struct pathname *pnp, struct pathname *rpnp,
184     int followlink, vnode_t **dirvpp, vnode_t **compvpp,
185     vnode_t *startvp)
186 {
187 	return (lookuppnatcred(pnp, rpnp, followlink, dirvpp, compvpp, startvp,
188 	    CRED()));
189 }
190 
191 /* Private flag to do our getcwd() dirty work */
192 #define	LOOKUP_CHECKREAD	0x10
193 #define	LOOKUP_MASK		(~LOOKUP_CHECKREAD)
194 
195 /*
196  * Starting at current directory, translate pathname pnp to end.
197  * Leave pathname of final component in pnp, return the vnode
198  * for the final component in *compvpp, and return the vnode
199  * for the parent of the final component in dirvpp.
200  *
201  * This is the central routine in pathname translation and handles
202  * multiple components in pathnames, separating them at /'s.  It also
203  * implements mounted file systems and processes symbolic links.
204  *
205  * vp is the vnode where the directory search should start.
206  *
207  * Reference counts: vp must be held prior to calling this function.  rootvp
208  * should only be held if rootvp != rootdir.
209  */
210 int
211 lookuppnvp(
212 	struct pathname *pnp,		/* pathname to lookup */
213 	struct pathname *rpnp,		/* if non-NULL, return resolved path */
214 	int flags,			/* follow symlinks */
215 	vnode_t **dirvpp,		/* ptr for parent vnode */
216 	vnode_t **compvpp,		/* ptr for entry vnode */
217 	vnode_t *rootvp,		/* rootvp */
218 	vnode_t *vp,			/* directory to start search at */
219 	cred_t *cr)			/* user's credential */
220 {
221 	vnode_t *cvp;	/* current component vp */
222 	char component[MAXNAMELEN];	/* buffer for component (incl null) */
223 	int error;
224 	int nlink;
225 	int lookup_flags;
226 	struct pathname presrvd; /* case preserved name */
227 	struct pathname *pp = NULL;
228 	vnode_t *startvp;
229 	vnode_t *zonevp = curproc->p_zone->zone_rootvp;		/* zone root */
230 	int must_be_directory = 0;
231 	boolean_t retry_with_kcred;
232 	uint32_t auditing = AU_AUDITING();
233 
234 	CPU_STATS_ADDQ(CPU, sys, namei, 1);
235 	nlink = 0;
236 	cvp = NULL;
237 	if (rpnp)
238 		rpnp->pn_pathlen = 0;
239 
240 	lookup_flags = dirvpp ? LOOKUP_DIR : 0;
241 	if (flags & FIGNORECASE) {
242 		lookup_flags |= FIGNORECASE;
243 		pn_alloc(&presrvd);
244 		pp = &presrvd;
245 	}
246 
247 	if (auditing)
248 		audit_anchorpath(pnp, vp == rootvp);
249 
250 	/*
251 	 * Eliminate any trailing slashes in the pathname.
252 	 * If there are any, we must follow all symlinks.
253 	 * Also, we must guarantee that the last component is a directory.
254 	 */
255 	if (pn_fixslash(pnp)) {
256 		flags |= FOLLOW;
257 		must_be_directory = 1;
258 	}
259 
260 	startvp = vp;
261 next:
262 	retry_with_kcred = B_FALSE;
263 
264 	/*
265 	 * Make sure we have a directory.
266 	 */
267 	if (vp->v_type != VDIR) {
268 		error = ENOTDIR;
269 		goto bad;
270 	}
271 
272 	if (rpnp && VN_CMP(vp, rootvp))
273 		(void) pn_set(rpnp, "/");
274 
275 	/*
276 	 * Process the next component of the pathname.
277 	 */
278 	if (error = pn_getcomponent(pnp, component)) {
279 		goto bad;
280 	}
281 
282 	/*
283 	 * Handle "..": two special cases.
284 	 * 1. If we're at the root directory (e.g. after chroot or
285 	 *    zone_enter) then change ".." to "." so we can't get
286 	 *    out of this subtree.
287 	 * 2. If this vnode is the root of a mounted file system,
288 	 *    then replace it with the vnode that was mounted on
289 	 *    so that we take the ".." in the other file system.
290 	 */
291 	if (component[0] == '.' && component[1] == '.' && component[2] == 0) {
292 checkforroot:
293 		if (VN_CMP(vp, rootvp) || VN_CMP(vp, zonevp)) {
294 			component[1] = '\0';
295 		} else if (vp->v_flag & VROOT) {
296 			vfs_t *vfsp;
297 			cvp = vp;
298 
299 			/*
300 			 * While we deal with the vfs pointer from the vnode
301 			 * the filesystem could have been forcefully unmounted
302 			 * and the vnode's v_vfsp could have been invalidated
303 			 * by VFS_UNMOUNT. Hence, we cache v_vfsp and use it
304 			 * with vfs_rlock_wait/vfs_unlock.
305 			 * It is safe to use the v_vfsp even it is freed by
306 			 * VFS_UNMOUNT because vfs_rlock_wait/vfs_unlock
307 			 * do not dereference v_vfsp. It is just used as a
308 			 * magic cookie.
309 			 * One more corner case here is the memory getting
310 			 * reused for another vfs structure. In this case
311 			 * lookuppnvp's vfs_rlock_wait will succeed, domount's
312 			 * vfs_lock will fail and domount will bail out with an
313 			 * error (EBUSY).
314 			 */
315 			vfsp = cvp->v_vfsp;
316 
317 			/*
318 			 * This lock is used to synchronize
319 			 * mounts/unmounts and lookups.
320 			 * Threads doing mounts/unmounts hold the
321 			 * writers version vfs_lock_wait().
322 			 */
323 
324 			vfs_rlock_wait(vfsp);
325 
326 			/*
327 			 * If this vnode is on a file system that
328 			 * has been forcibly unmounted,
329 			 * we can't proceed. Cancel this operation
330 			 * and return EIO.
331 			 *
332 			 * vfs_vnodecovered is NULL if unmounted.
333 			 * Currently, nfs uses VFS_UNMOUNTED to
334 			 * check if it's a forced-umount. Keep the
335 			 * same checking here as well even though it
336 			 * may not be needed.
337 			 */
338 			if (((vp = cvp->v_vfsp->vfs_vnodecovered) == NULL) ||
339 			    (cvp->v_vfsp->vfs_flag & VFS_UNMOUNTED)) {
340 				vfs_unlock(vfsp);
341 				VN_RELE(cvp);
342 				if (pp)
343 					pn_free(pp);
344 				return (EIO);
345 			}
346 			VN_HOLD(vp);
347 			vfs_unlock(vfsp);
348 			VN_RELE(cvp);
349 			cvp = NULL;
350 			/*
351 			 * Crossing mount points. For eg: We are doing
352 			 * a lookup of ".." for file systems root vnode
353 			 * mounted here, and VOP_LOOKUP() (with covered vnode)
354 			 * will be on underlying file systems mount point
355 			 * vnode. Set retry_with_kcred flag as we might end
356 			 * up doing VOP_LOOKUP() with kcred if required.
357 			 */
358 			retry_with_kcred = B_TRUE;
359 			goto checkforroot;
360 		}
361 	}
362 
363 	/*
364 	 * LOOKUP_CHECKREAD is a private flag used by vnodetopath() to indicate
365 	 * that we need to have read permission on every directory in the entire
366 	 * path.  This is used to ensure that a forward-lookup of a cached value
367 	 * has the same effect as a reverse-lookup when the cached value cannot
368 	 * be found.
369 	 */
370 	if ((flags & LOOKUP_CHECKREAD) &&
371 	    (error = VOP_ACCESS(vp, VREAD, 0, cr, NULL)) != 0)
372 		goto bad;
373 
374 	/*
375 	 * Perform a lookup in the current directory.
376 	 */
377 	error = VOP_LOOKUP(vp, component, &cvp, pnp, lookup_flags,
378 	    rootvp, cr, NULL, NULL, pp);
379 
380 	/*
381 	 * Retry with kcred - If crossing mount points & error is EACCES.
382 	 *
383 	 * If we are crossing mount points here and doing ".." lookup,
384 	 * VOP_LOOKUP() might fail if the underlying file systems
385 	 * mount point has no execute permission. In cases like these,
386 	 * we retry VOP_LOOKUP() by giving as much privilage as possible
387 	 * by passing kcred credentials.
388 	 *
389 	 * In case of hierarchical file systems, passing kcred still may
390 	 * or may not work.
391 	 * For eg: UFS FS --> Mount NFS FS --> Again mount UFS on some
392 	 *			directory inside NFS FS.
393 	 */
394 	if ((error == EACCES) && retry_with_kcred)
395 		error = VOP_LOOKUP(vp, component, &cvp, pnp, lookup_flags,
396 		    rootvp, zone_kcred(), NULL, NULL, pp);
397 
398 	if (error) {
399 		cvp = NULL;
400 		/*
401 		 * On error, return hard error if
402 		 * (a) we're not at the end of the pathname yet, or
403 		 * (b) the caller didn't want the parent directory, or
404 		 * (c) we failed for some reason other than a missing entry.
405 		 */
406 		if (pn_pathleft(pnp) || dirvpp == NULL || error != ENOENT)
407 			goto bad;
408 		if (auditing) {	/* directory access */
409 			if (error = audit_savepath(pnp, vp, vp, error, cr))
410 				goto bad_noaudit;
411 		}
412 
413 		pn_setlast(pnp);
414 		/*
415 		 * We inform the caller that the desired entry must be
416 		 * a directory by adding a '/' to the component name.
417 		 */
418 		if (must_be_directory && (error = pn_addslash(pnp)) != 0)
419 			goto bad;
420 		*dirvpp = vp;
421 		if (compvpp != NULL)
422 			*compvpp = NULL;
423 		if (rootvp != rootdir)
424 			VN_RELE(rootvp);
425 		if (pp)
426 			pn_free(pp);
427 		return (0);
428 	}
429 
430 	/*
431 	 * Traverse mount points.
432 	 * XXX why don't we need to hold a read lock here (call vn_vfsrlock)?
433 	 * What prevents a concurrent update to v_vfsmountedhere?
434 	 * 	Possible answer: if mounting, we might not see the mount
435 	 *	if it is concurrently coming into existence, but that's
436 	 *	really not much different from the thread running a bit slower.
437 	 *	If unmounting, we may get into traverse() when we shouldn't,
438 	 *	but traverse() will catch this case for us.
439 	 *	(For this to work, fetching v_vfsmountedhere had better
440 	 *	be atomic!)
441 	 */
442 	if (vn_mountedvfs(cvp) != NULL) {
443 		if ((error = traverse(&cvp)) != 0)
444 			goto bad;
445 	}
446 
447 	/*
448 	 * If we hit a symbolic link and there is more path to be
449 	 * translated or this operation does not wish to apply
450 	 * to a link, then place the contents of the link at the
451 	 * front of the remaining pathname.
452 	 */
453 	if (cvp->v_type == VLNK && ((flags & FOLLOW) || pn_pathleft(pnp))) {
454 		struct pathname linkpath;
455 
456 		if (++nlink > MAXSYMLINKS) {
457 			error = ELOOP;
458 			goto bad;
459 		}
460 		pn_alloc(&linkpath);
461 		if (error = pn_getsymlink(cvp, &linkpath, cr)) {
462 			pn_free(&linkpath);
463 			goto bad;
464 		}
465 
466 		if (auditing)
467 			audit_symlink(pnp, &linkpath);
468 
469 		if (pn_pathleft(&linkpath) == 0)
470 			(void) pn_set(&linkpath, ".");
471 		error = pn_insert(pnp, &linkpath, strlen(component));
472 		pn_free(&linkpath);
473 		if (error)
474 			goto bad;
475 		VN_RELE(cvp);
476 		cvp = NULL;
477 		if (pnp->pn_pathlen == 0) {
478 			error = ENOENT;
479 			goto bad;
480 		}
481 		if (pnp->pn_path[0] == '/') {
482 			do {
483 				pnp->pn_path++;
484 				pnp->pn_pathlen--;
485 			} while (pnp->pn_path[0] == '/');
486 			VN_RELE(vp);
487 			vp = rootvp;
488 			VN_HOLD(vp);
489 		}
490 		if (auditing)
491 			audit_anchorpath(pnp, vp == rootvp);
492 		if (pn_fixslash(pnp)) {
493 			flags |= FOLLOW;
494 			must_be_directory = 1;
495 		}
496 		goto next;
497 	}
498 
499 	/*
500 	 * If rpnp is non-NULL, remember the resolved path name therein.
501 	 * Do not include "." components.  Collapse occurrences of
502 	 * "previous/..", so long as "previous" is not itself "..".
503 	 * Exhausting rpnp results in error ENAMETOOLONG.
504 	 */
505 	if (rpnp && strcmp(component, ".") != 0) {
506 		size_t len;
507 
508 		if (strcmp(component, "..") == 0 &&
509 		    rpnp->pn_pathlen != 0 &&
510 		    !((rpnp->pn_pathlen > 2 &&
511 		    strncmp(rpnp->pn_path+rpnp->pn_pathlen-3, "/..", 3) == 0) ||
512 		    (rpnp->pn_pathlen == 2 &&
513 		    strncmp(rpnp->pn_path, "..", 2) == 0))) {
514 			while (rpnp->pn_pathlen &&
515 			    rpnp->pn_path[rpnp->pn_pathlen-1] != '/')
516 				rpnp->pn_pathlen--;
517 			if (rpnp->pn_pathlen > 1)
518 				rpnp->pn_pathlen--;
519 			rpnp->pn_path[rpnp->pn_pathlen] = '\0';
520 		} else {
521 			if (rpnp->pn_pathlen != 0 &&
522 			    rpnp->pn_path[rpnp->pn_pathlen-1] != '/')
523 				rpnp->pn_path[rpnp->pn_pathlen++] = '/';
524 			if (flags & FIGNORECASE) {
525 				/*
526 				 * Return the case-preserved name
527 				 * within the resolved path.
528 				 */
529 				error = copystr(pp->pn_buf,
530 				    rpnp->pn_path + rpnp->pn_pathlen,
531 				    rpnp->pn_bufsize - rpnp->pn_pathlen, &len);
532 			} else {
533 				error = copystr(component,
534 				    rpnp->pn_path + rpnp->pn_pathlen,
535 				    rpnp->pn_bufsize - rpnp->pn_pathlen, &len);
536 			}
537 			if (error)	/* copystr() returns ENAMETOOLONG */
538 				goto bad;
539 			rpnp->pn_pathlen += (len - 1);
540 			ASSERT(rpnp->pn_bufsize > rpnp->pn_pathlen);
541 		}
542 	}
543 
544 	/*
545 	 * If no more components, return last directory (if wanted) and
546 	 * last component (if wanted).
547 	 */
548 	if (pn_pathleft(pnp) == 0) {
549 		/*
550 		 * If there was a trailing slash in the pathname,
551 		 * make sure the last component is a directory.
552 		 */
553 		if (must_be_directory && cvp->v_type != VDIR) {
554 			error = ENOTDIR;
555 			goto bad;
556 		}
557 		if (dirvpp != NULL) {
558 			/*
559 			 * Check that we have the real parent and not
560 			 * an alias of the last component.
561 			 */
562 			if (vn_compare(vp, cvp)) {
563 				if (auditing)
564 					(void) audit_savepath(pnp, cvp, vp,
565 					    EINVAL, cr);
566 				pn_setlast(pnp);
567 				VN_RELE(vp);
568 				VN_RELE(cvp);
569 				if (rootvp != rootdir)
570 					VN_RELE(rootvp);
571 				if (pp)
572 					pn_free(pp);
573 				return (EINVAL);
574 			}
575 			*dirvpp = vp;
576 		} else
577 			VN_RELE(vp);
578 		if (auditing)
579 			(void) audit_savepath(pnp, cvp, vp, 0, cr);
580 		if (pnp->pn_path == pnp->pn_buf)
581 			(void) pn_set(pnp, ".");
582 		else
583 			pn_setlast(pnp);
584 		if (rpnp) {
585 			if (VN_CMP(cvp, rootvp))
586 				(void) pn_set(rpnp, "/");
587 			else if (rpnp->pn_pathlen == 0)
588 				(void) pn_set(rpnp, ".");
589 		}
590 
591 		if (compvpp != NULL)
592 			*compvpp = cvp;
593 		else
594 			VN_RELE(cvp);
595 		if (rootvp != rootdir)
596 			VN_RELE(rootvp);
597 		if (pp)
598 			pn_free(pp);
599 		return (0);
600 	}
601 
602 	/*
603 	 * Skip over slashes from end of last component.
604 	 */
605 	while (pnp->pn_path[0] == '/') {
606 		pnp->pn_path++;
607 		pnp->pn_pathlen--;
608 	}
609 
610 	/*
611 	 * Searched through another level of directory:
612 	 * release previous directory handle and save new (result
613 	 * of lookup) as current directory.
614 	 */
615 	VN_RELE(vp);
616 	vp = cvp;
617 	cvp = NULL;
618 	goto next;
619 
620 bad:
621 	if (auditing)	/* reached end of path */
622 		(void) audit_savepath(pnp, cvp, vp, error, cr);
623 bad_noaudit:
624 	/*
625 	 * Error.  Release vnodes and return.
626 	 */
627 	if (cvp)
628 		VN_RELE(cvp);
629 	/*
630 	 * If the error was ESTALE and the current directory to look in
631 	 * was the root for this lookup, the root for a mounted file
632 	 * system, or the starting directory for lookups, then
633 	 * return ENOENT instead of ESTALE.  In this case, no recovery
634 	 * is possible by the higher level.  If ESTALE was returned for
635 	 * some intermediate directory along the path, then recovery
636 	 * is potentially possible and retrying from the higher level
637 	 * will either correct the situation by purging stale cache
638 	 * entries or eventually get back to the point where no recovery
639 	 * is possible.
640 	 */
641 	if (error == ESTALE &&
642 	    (VN_CMP(vp, rootvp) || (vp->v_flag & VROOT) || vp == startvp))
643 		error = ENOENT;
644 	VN_RELE(vp);
645 	if (rootvp != rootdir)
646 		VN_RELE(rootvp);
647 	if (pp)
648 		pn_free(pp);
649 	return (error);
650 }
651 
652 /*
653  * Traverse a mount point.  Routine accepts a vnode pointer as a reference
654  * parameter and performs the indirection, releasing the original vnode.
655  */
656 int
657 traverse(vnode_t **cvpp)
658 {
659 	int error = 0;
660 	vnode_t *cvp;
661 	vnode_t *tvp;
662 	vfs_t *vfsp;
663 
664 	cvp = *cvpp;
665 
666 	/*
667 	 * If this vnode is mounted on, then we transparently indirect
668 	 * to the vnode which is the root of the mounted file system.
669 	 * Before we do this we must check that an unmount is not in
670 	 * progress on this vnode.
671 	 */
672 
673 	for (;;) {
674 		/*
675 		 * Try to read lock the vnode.  If this fails because
676 		 * the vnode is already write locked, then check to
677 		 * see whether it is the current thread which locked
678 		 * the vnode.  If it is not, then read lock the vnode
679 		 * by waiting to acquire the lock.
680 		 *
681 		 * The code path in domount() is an example of support
682 		 * which needs to look up two pathnames and locks one
683 		 * of them in between the two lookups.
684 		 */
685 		error = vn_vfsrlock(cvp);
686 		if (error) {
687 			if (!vn_vfswlock_held(cvp))
688 				error = vn_vfsrlock_wait(cvp);
689 			if (error != 0) {
690 				/*
691 				 * lookuppn() expects a held vnode to be
692 				 * returned because it promptly calls
693 				 * VN_RELE after the error return
694 				 */
695 				*cvpp = cvp;
696 				return (error);
697 			}
698 		}
699 
700 		/*
701 		 * Reached the end of the mount chain?
702 		 */
703 		vfsp = vn_mountedvfs(cvp);
704 		if (vfsp == NULL) {
705 			vn_vfsunlock(cvp);
706 			break;
707 		}
708 
709 		/*
710 		 * The read lock must be held across the call to VFS_ROOT() to
711 		 * prevent a concurrent unmount from destroying the vfs.
712 		 */
713 		error = VFS_ROOT(vfsp, &tvp);
714 		vn_vfsunlock(cvp);
715 
716 		if (error)
717 			break;
718 
719 		VN_RELE(cvp);
720 
721 		cvp = tvp;
722 	}
723 
724 	*cvpp = cvp;
725 	return (error);
726 }
727 
728 /*
729  * Return the lowermost vnode if this is a mountpoint.
730  */
731 static vnode_t *
732 vn_under(vnode_t *vp)
733 {
734 	vnode_t *uvp;
735 	vfs_t *vfsp;
736 
737 	while (vp->v_flag & VROOT) {
738 
739 		vfsp = vp->v_vfsp;
740 		vfs_rlock_wait(vfsp);
741 		if ((uvp = vfsp->vfs_vnodecovered) == NULL ||
742 		    (vfsp->vfs_flag & VFS_UNMOUNTED)) {
743 			vfs_unlock(vfsp);
744 			break;
745 		}
746 		VN_HOLD(uvp);
747 		vfs_unlock(vfsp);
748 		VN_RELE(vp);
749 		vp = uvp;
750 	}
751 
752 	return (vp);
753 }
754 
755 static int
756 vnode_match(vnode_t *v1, vnode_t *v2, cred_t *cr)
757 {
758 	vattr_t	v1attr, v2attr;
759 
760 	/*
761 	 * If we have a device file, check to see if is a cloned open of the
762 	 * same device.  For self-cloning devices, the major numbers will match.
763 	 * For devices cloned through the 'clone' driver, the minor number of
764 	 * the source device will be the same as the major number of the cloned
765 	 * device.
766 	 */
767 	if ((v1->v_type == VCHR || v1->v_type == VBLK) &&
768 	    v1->v_type == v2->v_type) {
769 		if ((spec_is_selfclone(v1) || spec_is_selfclone(v2)) &&
770 		    getmajor(v1->v_rdev) == getmajor(v2->v_rdev))
771 			return (1);
772 
773 		if (spec_is_clone(v1) &&
774 		    getmajor(v1->v_rdev) == getminor(v2->v_rdev))
775 			return (1);
776 
777 		if (spec_is_clone(v2) &&
778 		    getmajor(v2->v_rdev) == getminor(v1->v_rdev))
779 			return (1);
780 	}
781 
782 	v1attr.va_mask = v2attr.va_mask = AT_TYPE;
783 
784 	/*
785 	 * This check for symbolic links handles the pseudo-symlinks in procfs.
786 	 * These particular links have v_type of VDIR, but the attributes have a
787 	 * type of VLNK.  We need to avoid these links because otherwise if we
788 	 * are currently in '/proc/self/fd', then '/proc/self/cwd' will compare
789 	 * as the same vnode.
790 	 */
791 	if (VOP_GETATTR(v1, &v1attr, 0, cr, NULL) != 0 ||
792 	    VOP_GETATTR(v2, &v2attr, 0, cr, NULL) != 0 ||
793 	    v1attr.va_type == VLNK || v2attr.va_type == VLNK)
794 		return (0);
795 
796 	v1attr.va_mask = v2attr.va_mask = AT_TYPE | AT_FSID | AT_NODEID;
797 
798 	if (VOP_GETATTR(v1, &v1attr, ATTR_REAL, cr, NULL) != 0 ||
799 	    VOP_GETATTR(v2, &v2attr, ATTR_REAL, cr, NULL) != 0)
800 		return (0);
801 
802 	return (v1attr.va_fsid == v2attr.va_fsid &&
803 	    v1attr.va_nodeid == v2attr.va_nodeid);
804 }
805 
806 
807 /*
808  * Find the entry in the directory corresponding to the target vnode.
809  */
810 int
811 dirfindvp(vnode_t *vrootp, vnode_t *dvp, vnode_t *tvp, cred_t *cr, char *dbuf,
812     size_t dlen, dirent64_t **rdp)
813 {
814 	size_t dbuflen;
815 	struct iovec iov;
816 	struct uio uio;
817 	int error;
818 	int eof;
819 	vnode_t *cmpvp;
820 	struct dirent64 *dp;
821 	pathname_t pnp;
822 
823 	ASSERT(dvp->v_type == VDIR);
824 
825 	/*
826 	 * This is necessary because of the strange semantics of VOP_LOOKUP().
827 	 */
828 	bzero(&pnp, sizeof (pnp));
829 
830 	uio.uio_iov = &iov;
831 	uio.uio_iovcnt = 1;
832 	uio.uio_segflg = UIO_SYSSPACE;
833 	uio.uio_fmode = 0;
834 	uio.uio_extflg = UIO_COPY_CACHED;
835 	uio.uio_loffset = 0;
836 
837 	if ((error = VOP_ACCESS(dvp, VREAD, 0, cr, NULL)) != 0)
838 		return (error);
839 
840 	dp = NULL;
841 	eof = 0;
842 
843 	while (!eof) {
844 		uio.uio_resid = dlen;
845 		iov.iov_base = dbuf;
846 		iov.iov_len = dlen;
847 
848 		(void) VOP_RWLOCK(dvp, V_WRITELOCK_FALSE, NULL);
849 		error = VOP_READDIR(dvp, &uio, cr, &eof, NULL, 0);
850 		VOP_RWUNLOCK(dvp, V_WRITELOCK_FALSE, NULL);
851 
852 		dbuflen = dlen - uio.uio_resid;
853 
854 		if (error || dbuflen == 0)
855 			break;
856 
857 		dp = (dirent64_t *)dbuf;
858 		while ((intptr_t)dp < (intptr_t)dbuf + dbuflen) {
859 			/*
860 			 * Ignore '.' and '..' entries
861 			 */
862 			if (strcmp(dp->d_name, ".") == 0 ||
863 			    strcmp(dp->d_name, "..") == 0) {
864 				dp = (dirent64_t *)((intptr_t)dp +
865 				    dp->d_reclen);
866 				continue;
867 			}
868 
869 			error = VOP_LOOKUP(dvp, dp->d_name, &cmpvp, &pnp, 0,
870 			    vrootp, cr, NULL, NULL, NULL);
871 
872 			/*
873 			 * We only want to bail out if there was an error other
874 			 * than ENOENT.  Otherwise, it could be that someone
875 			 * just removed an entry since the readdir() call, and
876 			 * the entry we want is further on in the directory.
877 			 */
878 			if (error == 0) {
879 				if (vnode_match(tvp, cmpvp, cr)) {
880 					VN_RELE(cmpvp);
881 					*rdp = dp;
882 					return (0);
883 				}
884 
885 				VN_RELE(cmpvp);
886 			} else if (error != ENOENT) {
887 				return (error);
888 			}
889 
890 			dp = (dirent64_t *)((intptr_t)dp + dp->d_reclen);
891 		}
892 	}
893 
894 	/*
895 	 * Something strange has happened, this directory does not contain the
896 	 * specified vnode.  This should never happen in the normal case, since
897 	 * we ensured that dvp is the parent of vp.  This is possible in some
898 	 * rare conditions (races and the special .zfs directory).
899 	 */
900 	if (error == 0) {
901 		error = VOP_LOOKUP(dvp, ".zfs", &cmpvp, &pnp, 0, vrootp, cr,
902 		    NULL, NULL, NULL);
903 		if (error == 0) {
904 			if (vnode_match(tvp, cmpvp, cr)) {
905 				(void) strcpy(dp->d_name, ".zfs");
906 				dp->d_reclen = strlen(".zfs");
907 				dp->d_off = 2;
908 				dp->d_ino = 1;
909 				*rdp = dp;
910 			} else {
911 				error = ENOENT;
912 			}
913 			VN_RELE(cmpvp);
914 		}
915 	}
916 
917 	return (error);
918 }
919 
920 /*
921  * Given a global path (from rootdir), and a vnode that is the current root,
922  * return the portion of the path that is beneath the current root or NULL on
923  * failure.  The path MUST be a resolved path (no '..' entries or symlinks),
924  * otherwise this function will fail.
925  */
926 static char *
927 localpath(char *path, struct vnode *vrootp, cred_t *cr)
928 {
929 	vnode_t *vp;
930 	vnode_t *cvp;
931 	char component[MAXNAMELEN];
932 	char *ret = NULL;
933 	pathname_t pn;
934 
935 	/*
936 	 * We use vn_compare() instead of VN_CMP() in order to detect lofs
937 	 * mounts and stacked vnodes.
938 	 */
939 	if (vn_compare(vrootp, rootdir))
940 		return (path);
941 
942 	if (pn_get(path, UIO_SYSSPACE, &pn) != 0)
943 		return (NULL);
944 
945 	vp = rootdir;
946 	VN_HOLD(vp);
947 
948 	if (vn_ismntpt(vp) && traverse(&vp) != 0) {
949 		VN_RELE(vp);
950 		pn_free(&pn);
951 		return (NULL);
952 	}
953 
954 	while (pn_pathleft(&pn)) {
955 		pn_skipslash(&pn);
956 
957 		if (pn_getcomponent(&pn, component) != 0)
958 			break;
959 
960 		if (VOP_LOOKUP(vp, component, &cvp, &pn, 0, rootdir, cr,
961 		    NULL, NULL, NULL) != 0)
962 			break;
963 		VN_RELE(vp);
964 		vp = cvp;
965 
966 		if (vn_ismntpt(vp) && traverse(&vp) != 0)
967 			break;
968 
969 		if (vn_compare(vp, vrootp)) {
970 			ret = path + (pn.pn_path - pn.pn_buf);
971 			break;
972 		}
973 	}
974 
975 	VN_RELE(vp);
976 	pn_free(&pn);
977 
978 	return (ret);
979 }
980 
981 /*
982  * Clean a stale v_path from a vnode.  This is only performed if the v_path has
983  * not been altered since it was found to be stale
984  */
985 static void
986 vnode_clear_vpath(vnode_t *vp, char *vpath_old)
987 {
988 	mutex_enter(&vp->v_lock);
989 	if (vp->v_path != vn_vpath_empty && vp->v_path == vpath_old) {
990 		vp->v_path = vn_vpath_empty;
991 		mutex_exit(&vp->v_lock);
992 		kmem_free(vpath_old, strlen(vpath_old) + 1);
993 	} else {
994 		mutex_exit(&vp->v_lock);
995 	}
996 }
997 
998 /*
999  * Validate that a pathname refers to a given vnode.
1000  */
1001 static int
1002 vnode_valid_pn(vnode_t *vp, vnode_t *vrootp, pathname_t *pn, pathname_t *rpn,
1003     int flags, cred_t *cr)
1004 {
1005 	vnode_t *compvp;
1006 	/*
1007 	 * If we are in a zone or a chroot environment, then we have to
1008 	 * take additional steps, since the path to the root might not
1009 	 * be readable with the current credentials, even though the
1010 	 * process can legitmately access the file.  In this case, we
1011 	 * do the following:
1012 	 *
1013 	 * lookuppnvp() with all privileges to get the resolved path.
1014 	 * call localpath() to get the local portion of the path, and
1015 	 * continue as normal.
1016 	 *
1017 	 * If the the conversion to a local path fails, then we continue
1018 	 * as normal.  This is a heuristic to make process object file
1019 	 * paths available from within a zone.  Because lofs doesn't
1020 	 * support page operations, the vnode stored in the seg_t is
1021 	 * actually the underlying real vnode, not the lofs node itself.
1022 	 * Most of the time, the lofs path is the same as the underlying
1023 	 * vnode (for example, /usr/lib/libc.so.1).
1024 	 */
1025 	if (vrootp != rootdir) {
1026 		char *local = NULL;
1027 
1028 		VN_HOLD(rootdir);
1029 		if (lookuppnvp(pn, rpn, FOLLOW, NULL, &compvp, rootdir,
1030 		    rootdir, kcred) == 0) {
1031 			local = localpath(rpn->pn_path, vrootp, kcred);
1032 			VN_RELE(compvp);
1033 		}
1034 
1035 		/*
1036 		 * The original pn was changed through lookuppnvp().
1037 		 * Set it to local for next validation attempt.
1038 		 */
1039 		if (local) {
1040 			(void) pn_set(pn, local);
1041 		} else {
1042 			return (1);
1043 		}
1044 	}
1045 
1046 	/*
1047 	 * We should have a local path at this point, so start the search from
1048 	 * the root of the current process.
1049 	 */
1050 	VN_HOLD(vrootp);
1051 	if (vrootp != rootdir)
1052 		VN_HOLD(vrootp);
1053 	if (lookuppnvp(pn, rpn, FOLLOW | flags, NULL, &compvp, vrootp, vrootp,
1054 	    cr) == 0) {
1055 		/*
1056 		 * Check to see if the returned vnode is the same as the one we
1057 		 * expect.
1058 		 */
1059 		if (vn_compare(vp, compvp) ||
1060 		    vnode_match(vp, compvp, cr)) {
1061 			VN_RELE(compvp);
1062 			return (0);
1063 		} else {
1064 			VN_RELE(compvp);
1065 		}
1066 	}
1067 
1068 	return (1);
1069 }
1070 
1071 /*
1072  * Struct for tracking vnodes with invalidated v_path entries during a
1073  * dirtopath reverse lookup.  By keeping adequate state, those vnodes can be
1074  * revisted to populate v_path.
1075  */
1076 struct dirpath_walk {
1077 	struct dirpath_walk	*dw_next;
1078 	vnode_t			*dw_vnode;
1079 	vnode_t			*dw_pvnode;
1080 	size_t			dw_len;
1081 	char			*dw_name;
1082 };
1083 
1084 /*
1085  * Given a directory, return the full, resolved path.  This looks up "..",
1086  * searches for the given vnode in the parent, appends the component, etc.  It
1087  * is used to implement vnodetopath() and getcwd() when the cached path fails.
1088  */
1089 static int
1090 dirtopath(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen, int flags,
1091     cred_t *cr)
1092 {
1093 	pathname_t	pn, rpn, emptypn;
1094 	vnode_t		*pvp = NULL, *startvp = vp;
1095 	int		err = 0;
1096 	size_t		complen;
1097 	dirent64_t	*dp;
1098 	char		*bufloc, *dbuf;
1099 	const size_t	dlen = DIRENT64_RECLEN(MAXPATHLEN);
1100 	struct dirpath_walk *dw_chain = NULL, *dw_entry;
1101 
1102 	/* Operation only allowed on directories */
1103 	ASSERT(vp->v_type == VDIR);
1104 
1105 	/* We must have at least enough space for "/" */
1106 	if (buflen < 2)
1107 		return (ENAMETOOLONG);
1108 
1109 	/* Start at end of string with terminating null */
1110 	bufloc = &buf[buflen - 1];
1111 	*bufloc = '\0';
1112 
1113 	pn_alloc(&pn);
1114 	pn_alloc(&rpn);
1115 	dbuf = kmem_alloc(dlen, KM_SLEEP);
1116 	bzero(&emptypn, sizeof (emptypn));
1117 
1118 	/*
1119 	 * Begin with an additional reference on vp.  This will be decremented
1120 	 * during the loop.
1121 	 */
1122 	VN_HOLD(vp);
1123 
1124 	for (;;) {
1125 		int vprivs;
1126 		hrtime_t cached_stamp;
1127 
1128 		/*
1129 		 * Return if we've reached the root.  If the buffer is empty,
1130 		 * return '/'.  We explicitly don't use vn_compare(), since it
1131 		 * compares the real vnodes.  A lofs mount of '/' would produce
1132 		 * incorrect results otherwise.
1133 		 */
1134 		if (VN_CMP(vrootp, vp)) {
1135 			if (*bufloc == '\0')
1136 				*--bufloc = '/';
1137 			break;
1138 		}
1139 
1140 		/*
1141 		 * If we've reached the VFS root, something has gone wrong.  We
1142 		 * should have reached the root in the above check.  The only
1143 		 * explantation is that 'vp' is not contained withing the given
1144 		 * root, in which case we return EPERM.
1145 		 */
1146 		if (VN_CMP(rootdir, vp)) {
1147 			err = EPERM;
1148 			goto out;
1149 		}
1150 
1151 		/*
1152 		 * Shortcut: see if this vnode has correct v_path. If so,
1153 		 * we have the work done.
1154 		 */
1155 		mutex_enter(&vp->v_lock);
1156 		if (vp->v_path != vn_vpath_empty &&
1157 		    pn_set(&pn, vp->v_path) == 0) {
1158 			cached_stamp = vp->v_path_stamp;
1159 			mutex_exit(&vp->v_lock);
1160 			rpn.pn_path = rpn.pn_buf;
1161 
1162 			/* Ensure the v_path pointing to correct vnode */
1163 			if (vnode_valid_pn(vp, vrootp, &pn, &rpn, flags,
1164 			    cr) == 0) {
1165 				complen = strlen(rpn.pn_path);
1166 				bufloc -= complen;
1167 				if (bufloc < buf) {
1168 					err = ERANGE;
1169 					goto out;
1170 				}
1171 				bcopy(rpn.pn_path, bufloc, complen);
1172 				break;
1173 			} else {
1174 				/*
1175 				 * Immediately nuke cached v_path entries known
1176 				 * to be invalid.
1177 				 */
1178 				vn_clearpath(vp, cached_stamp);
1179 			}
1180 		} else {
1181 			mutex_exit(&vp->v_lock);
1182 		}
1183 
1184 		/*
1185 		 * Shortcuts failed, search for this vnode in its parent.  If
1186 		 * this is a mountpoint, then get the vnode underneath.
1187 		 */
1188 		if (vp->v_flag & VROOT)
1189 			vp = vn_under(vp);
1190 		if ((err = VOP_LOOKUP(vp, "..", &pvp, &emptypn, 0, vrootp, cr,
1191 		    NULL, NULL, NULL)) != 0)
1192 			goto out;
1193 
1194 		/*
1195 		 * With extended attributes, it's possible for a directory to
1196 		 * have a parent that is a regular file.  Check for that here.
1197 		 */
1198 		if (pvp->v_type != VDIR) {
1199 			err = ENOTDIR;
1200 			goto out;
1201 		}
1202 
1203 		/*
1204 		 * If this is true, something strange has happened.  This is
1205 		 * only true if we are the root of a filesystem, which should
1206 		 * have been caught by the check above.
1207 		 */
1208 		if (VN_CMP(pvp, vp)) {
1209 			err = ENOENT;
1210 			goto out;
1211 		}
1212 
1213 		/*
1214 		 * Check if we have read and search privilege so, that
1215 		 * we can lookup the path in the directory
1216 		 */
1217 		vprivs = (flags & LOOKUP_CHECKREAD) ? VREAD | VEXEC : VEXEC;
1218 		if ((err = VOP_ACCESS(pvp, vprivs, 0, cr, NULL)) != 0) {
1219 			goto out;
1220 		}
1221 
1222 		/*
1223 		 * Search the parent directory for the entry corresponding to
1224 		 * this vnode.
1225 		 */
1226 		if ((err = dirfindvp(vrootp, pvp, vp, cr, dbuf, dlen, &dp))
1227 		    != 0)
1228 			goto out;
1229 		complen = strlen(dp->d_name);
1230 		bufloc -= complen;
1231 		if (bufloc <= buf) {
1232 			err = ENAMETOOLONG;
1233 			goto out;
1234 		}
1235 		bcopy(dp->d_name, bufloc, complen);
1236 
1237 		/* Prepend a slash to the current path.  */
1238 		*--bufloc = '/';
1239 
1240 		/*
1241 		 * Record the name and directory for later reconstruction and
1242 		 * link it up with the others.
1243 		 */
1244 		dw_entry = kmem_alloc(sizeof (*dw_entry), KM_SLEEP);
1245 		dw_entry->dw_name = kmem_alloc(complen + 1, KM_SLEEP);
1246 		VN_HOLD(dw_entry->dw_vnode = vp);
1247 		VN_HOLD(dw_entry->dw_pvnode = pvp);
1248 		bcopy(dp->d_name, dw_entry->dw_name, complen + 1);
1249 		dw_entry->dw_len = complen;
1250 		dw_entry->dw_next = dw_chain;
1251 		dw_chain = dw_entry;
1252 
1253 		/* And continue with the next component */
1254 		VN_RELE(vp);
1255 		vp = pvp;
1256 		pvp = NULL;
1257 	}
1258 
1259 	/*
1260 	 * Place the path at the beginning of the buffer.
1261 	 */
1262 	if (bufloc != buf)
1263 		ovbcopy(bufloc, buf, buflen - (bufloc - buf));
1264 
1265 out:
1266 	/*
1267 	 * Walk over encountered directory entries which were afflicted with a
1268 	 * stale or absent v_path.  If the dirtopath was successful, we should
1269 	 * possess the necessary information to populate all of them with a
1270 	 * valid v_path.
1271 	 *
1272 	 * While processing this list, it is safe to call vn_setpath despite
1273 	 * the fact that racing vnode actions may have altered v_path entries
1274 	 * while the above loopwas still executing.  Any updated entries will
1275 	 * have a newer v_path_stamp value which prevents an invalid overwrite.
1276 	 *
1277 	 * If an error was encountered during the search, freeing the chain is
1278 	 * still required.
1279 	 */
1280 	dw_entry = dw_chain;
1281 	while (dw_entry != NULL) {
1282 		struct dirpath_walk *next = dw_entry->dw_next;
1283 
1284 		if (err == 0) {
1285 			vn_setpath(NULL, dw_entry->dw_pvnode,
1286 			    dw_entry->dw_vnode, dw_entry->dw_name,
1287 			    dw_entry->dw_len);
1288 		}
1289 
1290 		VN_RELE(dw_entry->dw_vnode);
1291 		VN_RELE(dw_entry->dw_pvnode);
1292 		kmem_free(dw_entry->dw_name, dw_entry->dw_len + 1);
1293 		kmem_free(dw_entry, sizeof (*dw_entry));
1294 		dw_entry = next;
1295 	}
1296 
1297 	/*
1298 	 * If the error was ESTALE and the current directory to look in
1299 	 * was the root for this lookup, the root for a mounted file
1300 	 * system, or the starting directory for lookups, then
1301 	 * return ENOENT instead of ESTALE.  In this case, no recovery
1302 	 * is possible by the higher level.  If ESTALE was returned for
1303 	 * some intermediate directory along the path, then recovery
1304 	 * is potentially possible and retrying from the higher level
1305 	 * will either correct the situation by purging stale cache
1306 	 * entries or eventually get back to the point where no recovery
1307 	 * is possible.
1308 	 */
1309 	if (err == ESTALE &&
1310 	    (VN_CMP(vp, vrootp) || (vp->v_flag & VROOT) || vp == startvp))
1311 		err = ENOENT;
1312 
1313 	kmem_free(dbuf, dlen);
1314 	VN_RELE(vp);
1315 	if (pvp)
1316 		VN_RELE(pvp);
1317 	pn_free(&pn);
1318 	pn_free(&rpn);
1319 
1320 	return (err);
1321 }
1322 
1323 /*
1324  * The additional flag, LOOKUP_CHECKREAD, is used to enforce artificial
1325  * constraints in order to be standards compliant.  For example, if we have
1326  * the cached path of '/foo/bar', and '/foo' has permissions 100 (execute
1327  * only), then we can legitimately look up the path to the current working
1328  * directory without needing read permission.  Existing standards tests,
1329  * however, assume that we are determining the path by repeatedly looking up
1330  * "..".  We need to keep this behavior in order to maintain backwards
1331  * compatibility.
1332  */
1333 static int
1334 vnodetopath_common(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen,
1335     cred_t *cr, int flags)
1336 {
1337 	pathname_t pn;
1338 	int ret = 0;
1339 	vnode_t *realvp;
1340 	boolean_t doclose = B_FALSE;
1341 
1342 	/*
1343 	 * If vrootp is NULL, get the root for curproc.  Callers with any other
1344 	 * requirements should pass in a different vrootp.
1345 	 */
1346 	if (vrootp == NULL) {
1347 		proc_t *p = curproc;
1348 
1349 		mutex_enter(&p->p_lock);
1350 		if ((vrootp = PTOU(p)->u_rdir) == NULL)
1351 			vrootp = rootdir;
1352 		VN_HOLD(vrootp);
1353 		mutex_exit(&p->p_lock);
1354 	} else {
1355 		VN_HOLD(vrootp);
1356 	}
1357 
1358 	/*
1359 	 * This is to get around an annoying artifact of the /proc filesystem,
1360 	 * which is the behavior of {cwd/root}.  Trying to resolve this path
1361 	 * will result in /proc/pid/cwd instead of whatever the real working
1362 	 * directory is.  We can't rely on VOP_REALVP(), since that will break
1363 	 * lofs.  The only difference between procfs and lofs is that opening
1364 	 * the file will return the underling vnode in the case of procfs.
1365 	 */
1366 	if (vp->v_type == VDIR && VOP_REALVP(vp, &realvp, NULL) == 0 &&
1367 	    realvp != vp) {
1368 		VN_HOLD(vp);
1369 		if (VOP_OPEN(&vp, FREAD, cr, NULL) == 0)
1370 			doclose = B_TRUE;
1371 		else
1372 			VN_RELE(vp);
1373 	}
1374 
1375 	/*
1376 	 * Check to see if we have a valid cached path in the vnode.
1377 	 */
1378 	pn_alloc(&pn);
1379 	mutex_enter(&vp->v_lock);
1380 	if (vp->v_path != vn_vpath_empty) {
1381 		hrtime_t cached_stamp;
1382 		pathname_t rpn;
1383 
1384 		cached_stamp = vp->v_path_stamp;
1385 		(void) pn_set(&pn, vp->v_path);
1386 		mutex_exit(&vp->v_lock);
1387 
1388 		/* We should only cache absolute paths */
1389 		ASSERT(pn.pn_buf[0] == '/');
1390 
1391 		pn_alloc(&rpn);
1392 		if (vnode_valid_pn(vp, vrootp, &pn, &rpn, flags, cr) == 0) {
1393 			/* Return the result, if we're able. */
1394 			if (buflen > rpn.pn_pathlen) {
1395 				bcopy(rpn.pn_path, buf, rpn.pn_pathlen + 1);
1396 			} else {
1397 				ret = ENAMETOOLONG;
1398 			}
1399 			pn_free(&pn);
1400 			pn_free(&rpn);
1401 			goto out;
1402 		}
1403 		pn_free(&rpn);
1404 		vn_clearpath(vp, cached_stamp);
1405 	} else {
1406 		mutex_exit(&vp->v_lock);
1407 	}
1408 	pn_free(&pn);
1409 
1410 	if (vp->v_type != VDIR) {
1411 		/*
1412 		 * The reverse lookup tricks used by dirtopath aren't possible
1413 		 * for non-directory entries.  The best which can be done is
1414 		 * clearing any stale v_path so later lookups can potentially
1415 		 * repopulate it with a valid path.
1416 		 */
1417 		ret = ENOENT;
1418 	} else {
1419 		ret = dirtopath(vrootp, vp, buf, buflen, flags, cr);
1420 	}
1421 
1422 out:
1423 	VN_RELE(vrootp);
1424 	if (doclose) {
1425 		(void) VOP_CLOSE(vp, FREAD, 1, 0, cr, NULL);
1426 		VN_RELE(vp);
1427 	}
1428 
1429 	return (ret);
1430 }
1431 
1432 int
1433 vnodetopath(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen, cred_t *cr)
1434 {
1435 	return (vnodetopath_common(vrootp, vp, buf, buflen, cr, 0));
1436 }
1437 
1438 int
1439 dogetcwd(char *buf, size_t buflen)
1440 {
1441 	int ret;
1442 	vnode_t *vp;
1443 	vnode_t *compvp;
1444 	refstr_t *cwd, *oldcwd;
1445 	const char *value;
1446 	pathname_t rpnp, pnp;
1447 	proc_t *p = curproc;
1448 
1449 	/*
1450 	 * Check to see if there is a cached version of the cwd.  If so, lookup
1451 	 * the cached value and make sure it is the same vnode.
1452 	 */
1453 	mutex_enter(&p->p_lock);
1454 	if ((cwd = PTOU(p)->u_cwd) != NULL)
1455 		refstr_hold(cwd);
1456 	vp = PTOU(p)->u_cdir;
1457 	VN_HOLD(vp);
1458 	mutex_exit(&p->p_lock);
1459 
1460 	/*
1461 	 * Make sure we have permission to access the current directory.
1462 	 */
1463 	if ((ret = VOP_ACCESS(vp, VEXEC, 0, CRED(), NULL)) != 0) {
1464 		if (cwd != NULL)
1465 			refstr_rele(cwd);
1466 		VN_RELE(vp);
1467 		return (ret);
1468 	}
1469 
1470 	if (cwd) {
1471 		value = refstr_value(cwd);
1472 		if ((ret = pn_get((char *)value, UIO_SYSSPACE, &pnp)) != 0) {
1473 			refstr_rele(cwd);
1474 			VN_RELE(vp);
1475 			return (ret);
1476 		}
1477 
1478 		pn_alloc(&rpnp);
1479 
1480 		if (lookuppn(&pnp, &rpnp, NO_FOLLOW, NULL, &compvp) == 0) {
1481 
1482 			if (VN_CMP(vp, compvp) &&
1483 			    strcmp(value, rpnp.pn_path) == 0) {
1484 				VN_RELE(compvp);
1485 				VN_RELE(vp);
1486 				pn_free(&pnp);
1487 				pn_free(&rpnp);
1488 				if (strlen(value) + 1 > buflen) {
1489 					refstr_rele(cwd);
1490 					return (ENAMETOOLONG);
1491 				}
1492 				bcopy(value, buf, strlen(value) + 1);
1493 				refstr_rele(cwd);
1494 				return (0);
1495 			}
1496 
1497 			VN_RELE(compvp);
1498 		}
1499 
1500 		pn_free(&rpnp);
1501 		pn_free(&pnp);
1502 
1503 		refstr_rele(cwd);
1504 	}
1505 
1506 	ret = vnodetopath_common(NULL, vp, buf, buflen, CRED(),
1507 	    LOOKUP_CHECKREAD);
1508 
1509 	VN_RELE(vp);
1510 
1511 	/*
1512 	 * Store the new cwd and replace the existing cached copy.
1513 	 */
1514 	if (ret == 0)
1515 		cwd = refstr_alloc(buf);
1516 	else
1517 		cwd = NULL;
1518 
1519 	mutex_enter(&p->p_lock);
1520 	oldcwd = PTOU(p)->u_cwd;
1521 	PTOU(p)->u_cwd = cwd;
1522 	mutex_exit(&p->p_lock);
1523 
1524 	if (oldcwd)
1525 		refstr_rele(oldcwd);
1526 
1527 	return (ret);
1528 }
1529