Home | History | Annotate | Download | only in fs
      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 2009 Sun Microsystems, Inc.  All rights reserved.
     23  * Use is subject to license terms.
     24  */
     25 
     26 /*	Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T	*/
     27 /*	  All Rights Reserved  	*/
     28 
     29 /*
     30  * University Copyright- Copyright (c) 1982, 1986, 1988
     31  * The Regents of the University of California
     32  * All Rights Reserved
     33  *
     34  * University Acknowledgment- Portions of this document are derived from
     35  * software developed by the University of California, Berkeley, and its
     36  * contributors.
     37  */
     38 
     39 #include <sys/types.h>
     40 #include <sys/param.h>
     41 #include <sys/systm.h>
     42 #include <sys/cpuvar.h>
     43 #include <sys/errno.h>
     44 #include <sys/cred.h>
     45 #include <sys/user.h>
     46 #include <sys/uio.h>
     47 #include <sys/vfs.h>
     48 #include <sys/vnode.h>
     49 #include <sys/pathname.h>
     50 #include <sys/proc.h>
     51 #include <sys/vtrace.h>
     52 #include <sys/sysmacros.h>
     53 #include <sys/debug.h>
     54 #include <sys/dirent.h>
     55 #include <c2/audit.h>
     56 #include <sys/zone.h>
     57 #include <sys/dnlc.h>
     58 #include <sys/fs/snode.h>
     59 
     60 /* Controls whether paths are stored with vnodes. */
     61 int vfs_vnode_path = 1;
     62 
     63 int
     64 lookupname(
     65 	char *fnamep,
     66 	enum uio_seg seg,
     67 	enum symfollow followlink,
     68 	vnode_t **dirvpp,
     69 	vnode_t **compvpp)
     70 {
     71 	return (lookupnameatcred(fnamep, seg, followlink, dirvpp, compvpp, NULL,
     72 	    CRED()));
     73 }
     74 
     75 /*
     76  * Lookup the user file name,
     77  * Handle allocation and freeing of pathname buffer, return error.
     78  */
     79 int
     80 lookupnameatcred(
     81 	char *fnamep,			/* user pathname */
     82 	enum uio_seg seg,		/* addr space that name is in */
     83 	enum symfollow followlink,	/* follow sym links */
     84 	vnode_t **dirvpp,		/* ret for ptr to parent dir vnode */
     85 	vnode_t **compvpp,		/* ret for ptr to component vnode */
     86 	vnode_t *startvp,		/* start path search from vp */
     87 	cred_t *cr)			/* credential */
     88 {
     89 	char namebuf[TYPICALMAXPATHLEN];
     90 	struct pathname lookpn;
     91 	int error;
     92 
     93 	error = pn_get_buf(fnamep, seg, &lookpn, namebuf, sizeof (namebuf));
     94 	if (error == 0) {
     95 		if (audit_active)
     96 			audit_lookupname();
     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, enum symfollow 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 	enum symfollow 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 	enum symfollow 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     enum symfollow 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 	vnode_t *tvp;	/* addressable temp ptr */
    223 	char component[MAXNAMELEN];	/* buffer for component (incl null) */
    224 	int error;
    225 	int nlink;
    226 	int lookup_flags;
    227 	struct pathname presrvd; /* case preserved name */
    228 	struct pathname *pp = NULL;
    229 	vnode_t *startvp;
    230 	vnode_t *zonevp = curproc->p_zone->zone_rootvp;		/* zone root */
    231 	int must_be_directory = 0;
    232 	boolean_t retry_with_kcred = B_FALSE;
    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 (audit_active)
    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 	/*
    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 		if (audit_active)
    278 			audit_addcomponent(pnp);
    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, &tvp, 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, &tvp, pnp, lookup_flags,
    396 		    rootvp, zone_kcred(), NULL, NULL, pp);
    397 
    398 	cvp = tvp;
    399 	if (error) {
    400 		cvp = NULL;
    401 		/*
    402 		 * On error, return hard error if
    403 		 * (a) we're not at the end of the pathname yet, or
    404 		 * (b) the caller didn't want the parent directory, or
    405 		 * (c) we failed for some reason other than a missing entry.
    406 		 */
    407 		if (pn_pathleft(pnp) || dirvpp == NULL || error != ENOENT)
    408 			goto bad;
    409 		if (audit_active) {	/* directory access */
    410 			if (error = audit_savepath(pnp, vp, error, cr))
    411 				goto bad_noaudit;
    412 		}
    413 		pn_setlast(pnp);
    414 		/*
    415 		 * We inform the caller that the desired entry must be
    416 		 * a directory by adding a '/' to the component name.
    417 		 */
    418 		if (must_be_directory && (error = pn_addslash(pnp)) != 0)
    419 			goto bad;
    420 		*dirvpp = vp;
    421 		if (compvpp != NULL)
    422 			*compvpp = NULL;
    423 		if (rootvp != rootdir)
    424 			VN_RELE(rootvp);
    425 		if (pp)
    426 			pn_free(pp);
    427 		return (0);
    428 	}
    429 
    430 	/*
    431 	 * Traverse mount points.
    432 	 * XXX why don't we need to hold a read lock here (call vn_vfsrlock)?
    433 	 * What prevents a concurrent update to v_vfsmountedhere?
    434 	 * 	Possible answer: if mounting, we might not see the mount
    435 	 *	if it is concurrently coming into existence, but that's
    436 	 *	really not much different from the thread running a bit slower.
    437 	 *	If unmounting, we may get into traverse() when we shouldn't,
    438 	 *	but traverse() will catch this case for us.
    439 	 *	(For this to work, fetching v_vfsmountedhere had better
    440 	 *	be atomic!)
    441 	 */
    442 	if (vn_mountedvfs(cvp) != NULL) {
    443 		tvp = cvp;
    444 		if ((error = traverse(&tvp)) != 0) {
    445 			/*
    446 			 * It is required to assign cvp here, because
    447 			 * traverse() will return a held vnode which
    448 			 * may different than the vnode that was passed
    449 			 * in (even in the error case).  If traverse()
    450 			 * changes the vnode it releases the original,
    451 			 * and holds the new one.
    452 			 */
    453 			cvp = tvp;
    454 			goto bad;
    455 		}
    456 		cvp = tvp;
    457 	}
    458 
    459 	/*
    460 	 * If we hit a symbolic link and there is more path to be
    461 	 * translated or this operation does not wish to apply
    462 	 * to a link, then place the contents of the link at the
    463 	 * front of the remaining pathname.
    464 	 */
    465 	if (cvp->v_type == VLNK && ((flags & FOLLOW) || pn_pathleft(pnp))) {
    466 		struct pathname linkpath;
    467 		if (audit_active) {
    468 			if (error = audit_pathcomp(pnp, cvp, cr))
    469 				goto bad;
    470 		}
    471 
    472 		if (++nlink > MAXSYMLINKS) {
    473 			error = ELOOP;
    474 			goto bad;
    475 		}
    476 		pn_alloc(&linkpath);
    477 		if (error = pn_getsymlink(cvp, &linkpath, cr)) {
    478 			pn_free(&linkpath);
    479 			goto bad;
    480 		}
    481 
    482 		if (audit_active)
    483 			audit_symlink(pnp, &linkpath);
    484 
    485 		if (pn_pathleft(&linkpath) == 0)
    486 			(void) pn_set(&linkpath, ".");
    487 		error = pn_insert(pnp, &linkpath, strlen(component));
    488 		pn_free(&linkpath);
    489 		if (error)
    490 			goto bad;
    491 		VN_RELE(cvp);
    492 		cvp = NULL;
    493 		if (pnp->pn_pathlen == 0) {
    494 			error = ENOENT;
    495 			goto bad;
    496 		}
    497 		if (pnp->pn_path[0] == '/') {
    498 			do {
    499 				pnp->pn_path++;
    500 				pnp->pn_pathlen--;
    501 			} while (pnp->pn_path[0] == '/');
    502 			VN_RELE(vp);
    503 			vp = rootvp;
    504 			VN_HOLD(vp);
    505 		}
    506 		if (audit_active)
    507 			audit_anchorpath(pnp, vp == rootvp);
    508 		if (pn_fixslash(pnp)) {
    509 			flags |= FOLLOW;
    510 			must_be_directory = 1;
    511 		}
    512 		goto next;
    513 	}
    514 
    515 	/*
    516 	 * If rpnp is non-NULL, remember the resolved path name therein.
    517 	 * Do not include "." components.  Collapse occurrences of
    518 	 * "previous/..", so long as "previous" is not itself "..".
    519 	 * Exhausting rpnp results in error ENAMETOOLONG.
    520 	 */
    521 	if (rpnp && strcmp(component, ".") != 0) {
    522 		size_t len;
    523 
    524 		if (strcmp(component, "..") == 0 &&
    525 		    rpnp->pn_pathlen != 0 &&
    526 		    !((rpnp->pn_pathlen > 2 &&
    527 		    strncmp(rpnp->pn_path+rpnp->pn_pathlen-3, "/..", 3) == 0) ||
    528 		    (rpnp->pn_pathlen == 2 &&
    529 		    strncmp(rpnp->pn_path, "..", 2) == 0))) {
    530 			while (rpnp->pn_pathlen &&
    531 			    rpnp->pn_path[rpnp->pn_pathlen-1] != '/')
    532 				rpnp->pn_pathlen--;
    533 			if (rpnp->pn_pathlen > 1)
    534 				rpnp->pn_pathlen--;
    535 			rpnp->pn_path[rpnp->pn_pathlen] = '\0';
    536 		} else {
    537 			if (rpnp->pn_pathlen != 0 &&
    538 			    rpnp->pn_path[rpnp->pn_pathlen-1] != '/')
    539 				rpnp->pn_path[rpnp->pn_pathlen++] = '/';
    540 			if (flags & FIGNORECASE) {
    541 				/*
    542 				 * Return the case-preserved name
    543 				 * within the resolved path.
    544 				 */
    545 				error = copystr(pp->pn_buf,
    546 				    rpnp->pn_path + rpnp->pn_pathlen,
    547 				    rpnp->pn_bufsize - rpnp->pn_pathlen, &len);
    548 			} else {
    549 				error = copystr(component,
    550 				    rpnp->pn_path + rpnp->pn_pathlen,
    551 				    rpnp->pn_bufsize - rpnp->pn_pathlen, &len);
    552 			}
    553 			if (error)	/* copystr() returns ENAMETOOLONG */
    554 				goto bad;
    555 			rpnp->pn_pathlen += (len - 1);
    556 			ASSERT(rpnp->pn_bufsize > rpnp->pn_pathlen);
    557 		}
    558 	}
    559 
    560 	/*
    561 	 * If no more components, return last directory (if wanted) and
    562 	 * last component (if wanted).
    563 	 */
    564 	if (pn_pathleft(pnp) == 0) {
    565 		/*
    566 		 * If there was a trailing slash in the pathname,
    567 		 * make sure the last component is a directory.
    568 		 */
    569 		if (must_be_directory && cvp->v_type != VDIR) {
    570 			error = ENOTDIR;
    571 			goto bad;
    572 		}
    573 		if (dirvpp != NULL) {
    574 			/*
    575 			 * Check that we have the real parent and not
    576 			 * an alias of the last component.
    577 			 */
    578 			if (vn_compare(vp, cvp)) {
    579 				if (audit_active)
    580 					(void) audit_savepath(pnp, cvp,
    581 					    EINVAL, cr);
    582 				pn_setlast(pnp);
    583 				VN_RELE(vp);
    584 				VN_RELE(cvp);
    585 				if (rootvp != rootdir)
    586 					VN_RELE(rootvp);
    587 				if (pp)
    588 					pn_free(pp);
    589 				return (EINVAL);
    590 			}
    591 			if (audit_active) {
    592 				if (error = audit_pathcomp(pnp, vp, cr))
    593 					goto bad;
    594 			}
    595 			*dirvpp = vp;
    596 		} else
    597 			VN_RELE(vp);
    598 		if (audit_active)
    599 			(void) audit_savepath(pnp, cvp, 0, cr);
    600 		if (pnp->pn_path == pnp->pn_buf)
    601 			(void) pn_set(pnp, ".");
    602 		else
    603 			pn_setlast(pnp);
    604 		if (rpnp) {
    605 			if (VN_CMP(cvp, rootvp))
    606 				(void) pn_set(rpnp, "/");
    607 			else if (rpnp->pn_pathlen == 0)
    608 				(void) pn_set(rpnp, ".");
    609 		}
    610 
    611 		if (compvpp != NULL)
    612 			*compvpp = cvp;
    613 		else
    614 			VN_RELE(cvp);
    615 		if (rootvp != rootdir)
    616 			VN_RELE(rootvp);
    617 		if (pp)
    618 			pn_free(pp);
    619 		return (0);
    620 	}
    621 
    622 	if (audit_active) {
    623 		if (error = audit_pathcomp(pnp, cvp, cr))
    624 			goto bad;
    625 	}
    626 
    627 	/*
    628 	 * Skip over slashes from end of last component.
    629 	 */
    630 	while (pnp->pn_path[0] == '/') {
    631 		pnp->pn_path++;
    632 		pnp->pn_pathlen--;
    633 	}
    634 
    635 	/*
    636 	 * Searched through another level of directory:
    637 	 * release previous directory handle and save new (result
    638 	 * of lookup) as current directory.
    639 	 */
    640 	VN_RELE(vp);
    641 	vp = cvp;
    642 	cvp = NULL;
    643 	goto next;
    644 
    645 bad:
    646 	if (audit_active)	/* reached end of path */
    647 		(void) audit_savepath(pnp, cvp, error, cr);
    648 bad_noaudit:
    649 	/*
    650 	 * Error.  Release vnodes and return.
    651 	 */
    652 	if (cvp)
    653 		VN_RELE(cvp);
    654 	/*
    655 	 * If the error was ESTALE and the current directory to look in
    656 	 * was the root for this lookup, the root for a mounted file
    657 	 * system, or the starting directory for lookups, then
    658 	 * return ENOENT instead of ESTALE.  In this case, no recovery
    659 	 * is possible by the higher level.  If ESTALE was returned for
    660 	 * some intermediate directory along the path, then recovery
    661 	 * is potentially possible and retrying from the higher level
    662 	 * will either correct the situation by purging stale cache
    663 	 * entries or eventually get back to the point where no recovery
    664 	 * is possible.
    665 	 */
    666 	if (error == ESTALE &&
    667 	    (VN_CMP(vp, rootvp) || (vp->v_flag & VROOT) || vp == startvp))
    668 		error = ENOENT;
    669 	VN_RELE(vp);
    670 	if (rootvp != rootdir)
    671 		VN_RELE(rootvp);
    672 	if (pp)
    673 		pn_free(pp);
    674 	return (error);
    675 }
    676 
    677 /*
    678  * Traverse a mount point.  Routine accepts a vnode pointer as a reference
    679  * parameter and performs the indirection, releasing the original vnode.
    680  */
    681 int
    682 traverse(vnode_t **cvpp)
    683 {
    684 	int error = 0;
    685 	vnode_t *cvp;
    686 	vnode_t *tvp;
    687 	vfs_t *vfsp;
    688 
    689 	cvp = *cvpp;
    690 
    691 	/*
    692 	 * If this vnode is mounted on, then we transparently indirect
    693 	 * to the vnode which is the root of the mounted file system.
    694 	 * Before we do this we must check that an unmount is not in
    695 	 * progress on this vnode.
    696 	 */
    697 
    698 	for (;;) {
    699 		/*
    700 		 * Try to read lock the vnode.  If this fails because
    701 		 * the vnode is already write locked, then check to
    702 		 * see whether it is the current thread which locked
    703 		 * the vnode.  If it is not, then read lock the vnode
    704 		 * by waiting to acquire the lock.
    705 		 *
    706 		 * The code path in domount() is an example of support
    707 		 * which needs to look up two pathnames and locks one
    708 		 * of them in between the two lookups.
    709 		 */
    710 		error = vn_vfsrlock(cvp);
    711 		if (error) {
    712 			if (!vn_vfswlock_held(cvp))
    713 				error = vn_vfsrlock_wait(cvp);
    714 			if (error != 0) {
    715 				/*
    716 				 * lookuppn() expects a held vnode to be
    717 				 * returned because it promptly calls
    718 				 * VN_RELE after the error return
    719 				 */
    720 				*cvpp = cvp;
    721 				return (error);
    722 			}
    723 		}
    724 
    725 		/*
    726 		 * Reached the end of the mount chain?
    727 		 */
    728 		vfsp = vn_mountedvfs(cvp);
    729 		if (vfsp == NULL) {
    730 			vn_vfsunlock(cvp);
    731 			break;
    732 		}
    733 
    734 		/*
    735 		 * The read lock must be held across the call to VFS_ROOT() to
    736 		 * prevent a concurrent unmount from destroying the vfs.
    737 		 */
    738 		error = VFS_ROOT(vfsp, &tvp);
    739 		vn_vfsunlock(cvp);
    740 
    741 		if (error)
    742 			break;
    743 
    744 		VN_RELE(cvp);
    745 
    746 		cvp = tvp;
    747 	}
    748 
    749 	*cvpp = cvp;
    750 	return (error);
    751 }
    752 
    753 /*
    754  * Return the lowermost vnode if this is a mountpoint.
    755  */
    756 static vnode_t *
    757 vn_under(vnode_t *vp)
    758 {
    759 	vnode_t *uvp;
    760 	vfs_t *vfsp;
    761 
    762 	while (vp->v_flag & VROOT) {
    763 
    764 		vfsp = vp->v_vfsp;
    765 		vfs_rlock_wait(vfsp);
    766 		if ((uvp = vfsp->vfs_vnodecovered) == NULL ||
    767 		    (vfsp->vfs_flag & VFS_UNMOUNTED)) {
    768 			vfs_unlock(vfsp);
    769 			break;
    770 		}
    771 		VN_HOLD(uvp);
    772 		vfs_unlock(vfsp);
    773 		VN_RELE(vp);
    774 		vp = uvp;
    775 	}
    776 
    777 	return (vp);
    778 }
    779 
    780 static int
    781 vnode_match(vnode_t *v1, vnode_t *v2, cred_t *cr)
    782 {
    783 	vattr_t	v1attr, v2attr;
    784 
    785 	/*
    786 	 * If we have a device file, check to see if is a cloned open of the
    787 	 * same device.  For self-cloning devices, the major numbers will match.
    788 	 * For devices cloned through the 'clone' driver, the minor number of
    789 	 * the source device will be the same as the major number of the cloned
    790 	 * device.
    791 	 */
    792 	if ((v1->v_type == VCHR || v1->v_type == VBLK) &&
    793 	    v1->v_type == v2->v_type) {
    794 		if ((spec_is_selfclone(v1) || spec_is_selfclone(v2)) &&
    795 		    getmajor(v1->v_rdev) == getmajor(v2->v_rdev))
    796 			return (1);
    797 
    798 		if (spec_is_clone(v1) &&
    799 		    getmajor(v1->v_rdev) == getminor(v2->v_rdev))
    800 			return (1);
    801 
    802 		if (spec_is_clone(v2) &&
    803 		    getmajor(v2->v_rdev) == getminor(v1->v_rdev))
    804 			return (1);
    805 	}
    806 
    807 	v1attr.va_mask = v2attr.va_mask = AT_TYPE;
    808 
    809 	/*
    810 	 * This check for symbolic links handles the pseudo-symlinks in procfs.
    811 	 * These particular links have v_type of VDIR, but the attributes have a
    812 	 * type of VLNK.  We need to avoid these links because otherwise if we
    813 	 * are currently in '/proc/self/fd', then '/proc/self/cwd' will compare
    814 	 * as the same vnode.
    815 	 */
    816 	if (VOP_GETATTR(v1, &v1attr, 0, cr, NULL) != 0 ||
    817 	    VOP_GETATTR(v2, &v2attr, 0, cr, NULL) != 0 ||
    818 	    v1attr.va_type == VLNK || v2attr.va_type == VLNK)
    819 		return (0);
    820 
    821 	v1attr.va_mask = v2attr.va_mask = AT_TYPE | AT_FSID | AT_NODEID;
    822 
    823 	if (VOP_GETATTR(v1, &v1attr, ATTR_REAL, cr, NULL) != 0 ||
    824 	    VOP_GETATTR(v2, &v2attr, ATTR_REAL, cr, NULL) != 0)
    825 		return (0);
    826 
    827 	return (v1attr.va_fsid == v2attr.va_fsid &&
    828 	    v1attr.va_nodeid == v2attr.va_nodeid);
    829 }
    830 
    831 
    832 /*
    833  * Find the entry in the directory corresponding to the target vnode.
    834  */
    835 int
    836 dirfindvp(vnode_t *vrootp, vnode_t *dvp, vnode_t *tvp, cred_t *cr, char *dbuf,
    837     size_t dlen, dirent64_t **rdp)
    838 {
    839 	size_t dbuflen;
    840 	struct iovec iov;
    841 	struct uio uio;
    842 	int error;
    843 	int eof;
    844 	vnode_t *cmpvp;
    845 	struct dirent64 *dp;
    846 	pathname_t pnp;
    847 
    848 	ASSERT(dvp->v_type == VDIR);
    849 
    850 	/*
    851 	 * This is necessary because of the strange semantics of VOP_LOOKUP().
    852 	 */
    853 	bzero(&pnp, sizeof (pnp));
    854 
    855 	eof = 0;
    856 
    857 	uio.uio_iov = &iov;
    858 	uio.uio_iovcnt = 1;
    859 	uio.uio_segflg = UIO_SYSSPACE;
    860 	uio.uio_fmode = 0;
    861 	uio.uio_extflg = UIO_COPY_CACHED;
    862 	uio.uio_loffset = 0;
    863 
    864 	if ((error = VOP_ACCESS(dvp, VREAD, 0, cr, NULL)) != 0)
    865 		return (error);
    866 
    867 	while (!eof) {
    868 		uio.uio_resid = dlen;
    869 		iov.iov_base = dbuf;
    870 		iov.iov_len = dlen;
    871 
    872 		(void) VOP_RWLOCK(dvp, V_WRITELOCK_FALSE, NULL);
    873 		error = VOP_READDIR(dvp, &uio, cr, &eof, NULL, 0);
    874 		VOP_RWUNLOCK(dvp, V_WRITELOCK_FALSE, NULL);
    875 
    876 		dbuflen = dlen - uio.uio_resid;
    877 
    878 		if (error || dbuflen == 0)
    879 			break;
    880 
    881 		dp = (dirent64_t *)dbuf;
    882 		while ((intptr_t)dp < (intptr_t)dbuf + dbuflen) {
    883 			/*
    884 			 * Ignore '.' and '..' entries
    885 			 */
    886 			if (strcmp(dp->d_name, ".") == 0 ||
    887 			    strcmp(dp->d_name, "..") == 0) {
    888 				dp = (dirent64_t *)((intptr_t)dp +
    889 				    dp->d_reclen);
    890 				continue;
    891 			}
    892 
    893 			error = VOP_LOOKUP(dvp, dp->d_name, &cmpvp, &pnp, 0,
    894 			    vrootp, cr, NULL, NULL, NULL);
    895 
    896 			/*
    897 			 * We only want to bail out if there was an error other
    898 			 * than ENOENT.  Otherwise, it could be that someone
    899 			 * just removed an entry since the readdir() call, and
    900 			 * the entry we want is further on in the directory.
    901 			 */
    902 			if (error == 0) {
    903 				if (vnode_match(tvp, cmpvp, cr)) {
    904 					VN_RELE(cmpvp);
    905 					*rdp = dp;
    906 					return (0);
    907 				}
    908 
    909 				VN_RELE(cmpvp);
    910 			} else if (error != ENOENT) {
    911 				return (error);
    912 			}
    913 
    914 			dp = (dirent64_t *)((intptr_t)dp + dp->d_reclen);
    915 		}
    916 	}
    917 
    918 	/*
    919 	 * Something strange has happened, this directory does not contain the
    920 	 * specified vnode.  This should never happen in the normal case, since
    921 	 * we ensured that dvp is the parent of vp.  This is possible in some
    922 	 * rare conditions (races and the special .zfs directory).
    923 	 */
    924 	if (error == 0) {
    925 		error = VOP_LOOKUP(dvp, ".zfs", &cmpvp, &pnp, 0, vrootp, cr,
    926 		    NULL, NULL, NULL);
    927 		if (error == 0) {
    928 			if (vnode_match(tvp, cmpvp, cr)) {
    929 				(void) strcpy(dp->d_name, ".zfs");
    930 				dp->d_reclen = strlen(".zfs");
    931 				dp->d_off = 2;
    932 				dp->d_ino = 1;
    933 				*rdp = dp;
    934 			} else {
    935 				error = ENOENT;
    936 			}
    937 			VN_RELE(cmpvp);
    938 		}
    939 	}
    940 
    941 	return (error);
    942 }
    943 
    944 /*
    945  * Given a global path (from rootdir), and a vnode that is the current root,
    946  * return the portion of the path that is beneath the current root or NULL on
    947  * failure.  The path MUST be a resolved path (no '..' entries or symlinks),
    948  * otherwise this function will fail.
    949  */
    950 static char *
    951 localpath(char *path, struct vnode *vrootp, cred_t *cr)
    952 {
    953 	vnode_t *vp;
    954 	vnode_t *cvp;
    955 	char component[MAXNAMELEN];
    956 	char *ret = NULL;
    957 	pathname_t pn;
    958 
    959 	/*
    960 	 * We use vn_compare() instead of VN_CMP() in order to detect lofs
    961 	 * mounts and stacked vnodes.
    962 	 */
    963 	if (vn_compare(vrootp, rootdir))
    964 		return (path);
    965 
    966 	if (pn_get(path, UIO_SYSSPACE, &pn) != 0)
    967 		return (NULL);
    968 
    969 	vp = rootdir;
    970 	VN_HOLD(vp);
    971 
    972 	if (vn_ismntpt(vp) && traverse(&vp) != 0) {
    973 		VN_RELE(vp);
    974 		pn_free(&pn);
    975 		return (NULL);
    976 	}
    977 
    978 	while (pn_pathleft(&pn)) {
    979 		pn_skipslash(&pn);
    980 
    981 		if (pn_getcomponent(&pn, component) != 0)
    982 			break;
    983 
    984 		if (VOP_LOOKUP(vp, component, &cvp, &pn, 0, rootdir, cr,
    985 		    NULL, NULL, NULL) != 0)
    986 			break;
    987 		VN_RELE(vp);
    988 		vp = cvp;
    989 
    990 		if (vn_ismntpt(vp) && traverse(&vp) != 0)
    991 			break;
    992 
    993 		if (vn_compare(vp, vrootp)) {
    994 			ret = path + (pn.pn_path - pn.pn_buf);
    995 			break;
    996 		}
    997 	}
    998 
    999 	VN_RELE(vp);
   1000 	pn_free(&pn);
   1001 
   1002 	return (ret);
   1003 }
   1004 
   1005 /*
   1006  * Given a directory, return the full, resolved path.  This looks up "..",
   1007  * searches for the given vnode in the parent, appends the component, etc.  It
   1008  * is used to implement vnodetopath() and getcwd() when the cached path fails
   1009  * (or vfs_vnode_path is not set).
   1010  */
   1011 static int
   1012 dirtopath(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen, int flags,
   1013     cred_t *cr)
   1014 {
   1015 	pathname_t pn, rpn, emptypn;
   1016 	vnode_t *cmpvp, *pvp = NULL;
   1017 	vnode_t *startvp = vp;
   1018 	int err = 0, vprivs;
   1019 	size_t complen;
   1020 	char *dbuf;
   1021 	dirent64_t *dp;
   1022 	char		*bufloc;
   1023 	size_t		dlen = DIRENT64_RECLEN(MAXPATHLEN);
   1024 	refstr_t	*mntpt;
   1025 
   1026 	/* Operation only allowed on directories */
   1027 	ASSERT(vp->v_type == VDIR);
   1028 
   1029 	/* We must have at least enough space for "/" */
   1030 	if (buflen < 2)
   1031 		return (ENAMETOOLONG);
   1032 
   1033 	/* Start at end of string with terminating null */
   1034 	bufloc = &buf[buflen - 1];
   1035 	*bufloc = '\0';
   1036 
   1037 	pn_alloc(&pn);
   1038 	pn_alloc(&rpn);
   1039 	dbuf = kmem_alloc(dlen, KM_SLEEP);
   1040 	bzero(&emptypn, sizeof (emptypn));
   1041 
   1042 	/*
   1043 	 * Begin with an additional reference on vp.  This will be decremented
   1044 	 * during the loop.
   1045 	 */
   1046 	VN_HOLD(vp);
   1047 
   1048 	for (;;) {
   1049 		/*
   1050 		 * Return if we've reached the root.  If the buffer is empty,
   1051 		 * return '/'.  We explicitly don't use vn_compare(), since it
   1052 		 * compares the real vnodes.  A lofs mount of '/' would produce
   1053 		 * incorrect results otherwise.
   1054 		 */
   1055 		if (VN_CMP(vrootp, vp)) {
   1056 			if (*bufloc == '\0')
   1057 				*--bufloc = '/';
   1058 			break;
   1059 		}
   1060 
   1061 		/*
   1062 		 * If we've reached the VFS root, something has gone wrong.  We
   1063 		 * should have reached the root in the above check.  The only
   1064 		 * explantation is that 'vp' is not contained withing the given
   1065 		 * root, in which case we return EPERM.
   1066 		 */
   1067 		if (VN_CMP(rootdir, vp)) {
   1068 			err = EPERM;
   1069 			goto out;
   1070 		}
   1071 
   1072 		/*
   1073 		 * Shortcut: see if this vnode is a mountpoint.  If so,
   1074 		 * grab the path information from the vfs_t.
   1075 		 */
   1076 		if (vp->v_flag & VROOT) {
   1077 
   1078 			mntpt = vfs_getmntpoint(vp->v_vfsp);
   1079 			if ((err = pn_set(&pn, (char *)refstr_value(mntpt)))
   1080 			    == 0) {
   1081 				refstr_rele(mntpt);
   1082 				rpn.pn_path = rpn.pn_buf;
   1083 
   1084 				/*
   1085 				 * Ensure the mountpoint still exists.
   1086 				 */
   1087 				VN_HOLD(vrootp);
   1088 				if (vrootp != rootdir)
   1089 					VN_HOLD(vrootp);
   1090 				if (lookuppnvp(&pn, &rpn, flags, NULL,
   1091 				    &cmpvp, vrootp, vrootp, cr) == 0) {
   1092 
   1093 					if (VN_CMP(vp, cmpvp)) {
   1094 						VN_RELE(cmpvp);
   1095 
   1096 						complen = strlen(rpn.pn_path);
   1097 						bufloc -= complen;
   1098 						if (bufloc < buf) {
   1099 							err = ERANGE;
   1100 							goto out;
   1101 						}
   1102 						bcopy(rpn.pn_path, bufloc,
   1103 						    complen);
   1104 						break;
   1105 					} else {
   1106 						VN_RELE(cmpvp);
   1107 					}
   1108 				}
   1109 			} else {
   1110 				refstr_rele(mntpt);
   1111 			}
   1112 		}
   1113 
   1114 		/*
   1115 		 * Shortcut: see if this vnode has correct v_path. If so,
   1116 		 * we have the work done.
   1117 		 */
   1118 		mutex_enter(&vp->v_lock);
   1119 		if (vp->v_path != NULL) {
   1120 
   1121 			if ((err = pn_set(&pn, vp->v_path)) == 0) {
   1122 				mutex_exit(&vp->v_lock);
   1123 				rpn.pn_path = rpn.pn_buf;
   1124 
   1125 				/*
   1126 				 * Ensure the v_path pointing to correct vnode
   1127 				 */
   1128 				VN_HOLD(vrootp);
   1129 				if (vrootp != rootdir)
   1130 					VN_HOLD(vrootp);
   1131 				if (lookuppnvp(&pn, &rpn, flags, NULL,
   1132 				    &cmpvp, vrootp, vrootp, cr) == 0) {
   1133 
   1134 					if (VN_CMP(vp, cmpvp)) {
   1135 						VN_RELE(cmpvp);
   1136 
   1137 						complen = strlen(rpn.pn_path);
   1138 						bufloc -= complen;
   1139 						if (bufloc < buf) {
   1140 							err = ERANGE;
   1141 							goto out;
   1142 						}
   1143 						bcopy(rpn.pn_path, bufloc,
   1144 						    complen);
   1145 						break;
   1146 					} else {
   1147 						VN_RELE(cmpvp);
   1148 					}
   1149 				}
   1150 			} else {
   1151 				mutex_exit(&vp->v_lock);
   1152 			}
   1153 		} else {
   1154 			mutex_exit(&vp->v_lock);
   1155 		}
   1156 
   1157 		/*
   1158 		 * Shortcuts failed, search for this vnode in its parent.  If
   1159 		 * this is a mountpoint, then get the vnode underneath.
   1160 		 */
   1161 		if (vp->v_flag & VROOT)
   1162 			vp = vn_under(vp);
   1163 		if ((err = VOP_LOOKUP(vp, "..", &pvp, &emptypn, 0, vrootp, cr,
   1164 		    NULL, NULL, NULL)) != 0)
   1165 			goto out;
   1166 
   1167 		/*
   1168 		 * With extended attributes, it's possible for a directory to
   1169 		 * have a parent that is a regular file.  Check for that here.
   1170 		 */
   1171 		if (pvp->v_type != VDIR) {
   1172 			err = ENOTDIR;
   1173 			goto out;
   1174 		}
   1175 
   1176 		/*
   1177 		 * If this is true, something strange has happened.  This is
   1178 		 * only true if we are the root of a filesystem, which should
   1179 		 * have been caught by the check above.
   1180 		 */
   1181 		if (VN_CMP(pvp, vp)) {
   1182 			err = ENOENT;
   1183 			goto out;
   1184 		}
   1185 
   1186 		/*
   1187 		 * Check if we have read and search privilege so, that
   1188 		 * we can lookup the path in the directory
   1189 		 */
   1190 		vprivs = (flags & LOOKUP_CHECKREAD) ? VREAD | VEXEC : VEXEC;
   1191 		if ((err = VOP_ACCESS(pvp, vprivs, 0, cr, NULL)) != 0) {
   1192 			goto out;
   1193 		}
   1194 
   1195 		/*
   1196 		 * Try to obtain the path component from dnlc cache
   1197 		 * before searching through the directory.
   1198 		 */
   1199 		if ((cmpvp = dnlc_reverse_lookup(vp, dbuf, dlen)) != NULL) {
   1200 			/*
   1201 			 * If we got parent vnode as a result,
   1202 			 * then the answered path is correct.
   1203 			 */
   1204 			if (VN_CMP(cmpvp, pvp)) {
   1205 				VN_RELE(cmpvp);
   1206 				complen = strlen(dbuf);
   1207 				bufloc -= complen;
   1208 				if (bufloc <= buf) {
   1209 					err = ENAMETOOLONG;
   1210 					goto out;
   1211 				}
   1212 				bcopy(dbuf, bufloc, complen);
   1213 
   1214 				/* Prepend a slash to the current path */
   1215 				*--bufloc = '/';
   1216 
   1217 				/* And continue with the next component */
   1218 				VN_RELE(vp);
   1219 				vp = pvp;
   1220 				pvp = NULL;
   1221 				continue;
   1222 			} else {
   1223 				VN_RELE(cmpvp);
   1224 			}
   1225 		}
   1226 
   1227 		/*
   1228 		 * Search the parent directory for the entry corresponding to
   1229 		 * this vnode.
   1230 		 */
   1231 		if ((err = dirfindvp(vrootp, pvp, vp, cr, dbuf, dlen, &dp))
   1232 		    != 0)
   1233 			goto out;
   1234 		complen = strlen(dp->d_name);
   1235 		bufloc -= complen;
   1236 		if (bufloc <= buf) {
   1237 			err = ENAMETOOLONG;
   1238 			goto out;
   1239 		}
   1240 		bcopy(dp->d_name, bufloc, complen);
   1241 
   1242 		/* Prepend a slash to the current path.  */
   1243 		*--bufloc = '/';
   1244 
   1245 		/* And continue with the next component */
   1246 		VN_RELE(vp);
   1247 		vp = pvp;
   1248 		pvp = NULL;
   1249 	}
   1250 
   1251 	/*
   1252 	 * Place the path at the beginning of the buffer.
   1253 	 */
   1254 	if (bufloc != buf)
   1255 		ovbcopy(bufloc, buf, buflen - (bufloc - buf));
   1256 
   1257 	/*
   1258 	 * We got here because of invalid v_path in startvp.
   1259 	 * Now, we have all info to fix it.
   1260 	 * Path must not include leading slash to let vn_renamepath
   1261 	 * pre-attach chroot'd root directory path. Also, trailing '\0'
   1262 	 * is not counted to length.
   1263 	 */
   1264 	vn_renamepath(vrootp, startvp, &buf[1], buflen - (bufloc - buf) - 2);
   1265 
   1266 out:
   1267 	/*
   1268 	 * If the error was ESTALE and the current directory to look in
   1269 	 * was the root for this lookup, the root for a mounted file
   1270 	 * system, or the starting directory for lookups, then
   1271 	 * return ENOENT instead of ESTALE.  In this case, no recovery
   1272 	 * is possible by the higher level.  If ESTALE was returned for
   1273 	 * some intermediate directory along the path, then recovery
   1274 	 * is potentially possible and retrying from the higher level
   1275 	 * will either correct the situation by purging stale cache
   1276 	 * entries or eventually get back to the point where no recovery
   1277 	 * is possible.
   1278 	 */
   1279 	if (err == ESTALE &&
   1280 	    (VN_CMP(vp, vrootp) || (vp->v_flag & VROOT) || vp == startvp))
   1281 		err = ENOENT;
   1282 
   1283 	kmem_free(dbuf, dlen);
   1284 	VN_RELE(vp);
   1285 	if (pvp)
   1286 		VN_RELE(pvp);
   1287 	pn_free(&pn);
   1288 	pn_free(&rpn);
   1289 
   1290 	return (err);
   1291 }
   1292 
   1293 /*
   1294  * The additional flag, LOOKUP_CHECKREAD, is used to enforce artificial
   1295  * constraints in order to be standards compliant.  For example, if we have
   1296  * the cached path of '/foo/bar', and '/foo' has permissions 100 (execute
   1297  * only), then we can legitimately look up the path to the current working
   1298  * directory without needing read permission.  Existing standards tests,
   1299  * however, assume that we are determining the path by repeatedly looking up
   1300  * "..".  We need to keep this behavior in order to maintain backwards
   1301  * compatibility.
   1302  */
   1303 static int
   1304 vnodetopath_common(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen,
   1305     cred_t *cr, int flags)
   1306 {
   1307 	pathname_t pn, rpn;
   1308 	int ret, len;
   1309 	vnode_t *compvp, *pvp, *realvp;
   1310 	proc_t *p = curproc;
   1311 	char path[MAXNAMELEN];
   1312 	int doclose = 0;
   1313 
   1314 	/*
   1315 	 * If vrootp is NULL, get the root for curproc.  Callers with any other
   1316 	 * requirements should pass in a different vrootp.
   1317 	 */
   1318 	if (vrootp == NULL) {
   1319 		mutex_enter(&p->p_lock);
   1320 		if ((vrootp = PTOU(p)->u_rdir) == NULL)
   1321 			vrootp = rootdir;
   1322 		VN_HOLD(vrootp);
   1323 		mutex_exit(&p->p_lock);
   1324 	} else {
   1325 		VN_HOLD(vrootp);
   1326 	}
   1327 
   1328 	/*
   1329 	 * This is to get around an annoying artifact of the /proc filesystem,
   1330 	 * which is the behavior of {cwd/root}.  Trying to resolve this path
   1331 	 * will result in /proc/pid/cwd instead of whatever the real working
   1332 	 * directory is.  We can't rely on VOP_REALVP(), since that will break
   1333 	 * lofs.  The only difference between procfs and lofs is that opening
   1334 	 * the file will return the underling vnode in the case of procfs.
   1335 	 */
   1336 	if (vp->v_type == VDIR && VOP_REALVP(vp, &realvp, NULL) == 0 &&
   1337 	    realvp != vp) {
   1338 		VN_HOLD(vp);
   1339 		if (VOP_OPEN(&vp, FREAD, cr, NULL) == 0)
   1340 			doclose = 1;
   1341 		else
   1342 			VN_RELE(vp);
   1343 	}
   1344 
   1345 	pn_alloc(&pn);
   1346 
   1347 	/*
   1348 	 * Check to see if we have a cached path in the vnode.
   1349 	 */
   1350 	mutex_enter(&vp->v_lock);
   1351 	if (vp->v_path != NULL) {
   1352 		(void) pn_set(&pn, vp->v_path);
   1353 		mutex_exit(&vp->v_lock);
   1354 
   1355 		pn_alloc(&rpn);
   1356 
   1357 		/* We should only cache absolute paths */
   1358 		ASSERT(pn.pn_buf[0] == '/');
   1359 
   1360 		/*
   1361 		 * If we are in a zone or a chroot environment, then we have to
   1362 		 * take additional steps, since the path to the root might not
   1363 		 * be readable with the current credentials, even though the
   1364 		 * process can legitmately access the file.  In this case, we
   1365 		 * do the following:
   1366 		 *
   1367 		 * lookuppnvp() with all privileges to get the resolved path.
   1368 		 * call localpath() to get the local portion of the path, and
   1369 		 * continue as normal.
   1370 		 *
   1371 		 * If the the conversion to a local path fails, then we continue
   1372 		 * as normal.  This is a heuristic to make process object file
   1373 		 * paths available from within a zone.  Because lofs doesn't
   1374 		 * support page operations, the vnode stored in the seg_t is
   1375 		 * actually the underlying real vnode, not the lofs node itself.
   1376 		 * Most of the time, the lofs path is the same as the underlying
   1377 		 * vnode (for example, /usr/lib/libc.so.1).
   1378 		 */
   1379 		if (vrootp != rootdir) {
   1380 			char *local = NULL;
   1381 			VN_HOLD(rootdir);
   1382 			if (lookuppnvp(&pn, &rpn, FOLLOW,
   1383 			    NULL, &compvp, rootdir, rootdir, kcred) == 0) {
   1384 				local = localpath(rpn.pn_path, vrootp,
   1385 				    kcred);
   1386 				VN_RELE(compvp);
   1387 			}
   1388 
   1389 			/*
   1390 			 * The original pn was changed through lookuppnvp().
   1391 			 * Set it to local for next validation attempt.
   1392 			 */
   1393 			if (local) {
   1394 				(void) pn_set(&pn, local);
   1395 			} else {
   1396 				goto notcached;
   1397 			}
   1398 		}
   1399 
   1400 		/*
   1401 		 * We should have a local path at this point, so start the
   1402 		 * search from the root of the current process.
   1403 		 */
   1404 		VN_HOLD(vrootp);
   1405 		if (vrootp != rootdir)
   1406 			VN_HOLD(vrootp);
   1407 		ret = lookuppnvp(&pn, &rpn, FOLLOW | flags, NULL,
   1408 		    &compvp, vrootp, vrootp, cr);
   1409 		if (ret == 0) {
   1410 			/*
   1411 			 * Check to see if the returned vnode is the same as
   1412 			 * the one we expect.  If not, give up.
   1413 			 */
   1414 			if (!vn_compare(vp, compvp) &&
   1415 			    !vnode_match(vp, compvp, cr)) {
   1416 				VN_RELE(compvp);
   1417 				goto notcached;
   1418 			}
   1419 
   1420 			VN_RELE(compvp);
   1421 
   1422 			/*
   1423 			 * Return the result.
   1424 			 */
   1425 			if (buflen <= rpn.pn_pathlen)
   1426 				goto notcached;
   1427 
   1428 			bcopy(rpn.pn_path, buf, rpn.pn_pathlen + 1);
   1429 			pn_free(&pn);
   1430 			pn_free(&rpn);
   1431 			VN_RELE(vrootp);
   1432 			if (doclose) {
   1433 				(void) VOP_CLOSE(vp, FREAD, 1, 0, cr, NULL);
   1434 				VN_RELE(vp);
   1435 			}
   1436 			return (0);
   1437 		}
   1438 
   1439 notcached:
   1440 		pn_free(&rpn);
   1441 	} else {
   1442 		mutex_exit(&vp->v_lock);
   1443 	}
   1444 
   1445 	pn_free(&pn);
   1446 
   1447 	if (vp->v_type != VDIR) {
   1448 		/*
   1449 		 * If we don't have a directory, try to find it in the dnlc via
   1450 		 * reverse lookup.  Once this is found, we can use the regular
   1451 		 * directory search to find the full path.
   1452 		 */
   1453 		if ((pvp = dnlc_reverse_lookup(vp, path, MAXNAMELEN)) != NULL) {
   1454 			/*
   1455 			 * Check if we have read privilege so, that
   1456 			 * we can lookup the path in the directory
   1457 			 */
   1458 			ret = 0;
   1459 			if ((flags & LOOKUP_CHECKREAD)) {
   1460 				ret = VOP_ACCESS(pvp, VREAD, 0, cr, NULL);
   1461 			}
   1462 			if (ret == 0) {
   1463 				ret = dirtopath(vrootp, pvp, buf, buflen,
   1464 				    flags, cr);
   1465 			}
   1466 			if (ret == 0) {
   1467 				len = strlen(buf);
   1468 				if (len + strlen(path) + 1 >= buflen) {
   1469 					ret = ENAMETOOLONG;
   1470 				} else {
   1471 					if (buf[len - 1] != '/')
   1472 						buf[len++] = '/';
   1473 					bcopy(path, buf + len,
   1474 					    strlen(path) + 1);
   1475 				}
   1476 			}
   1477 
   1478 			VN_RELE(pvp);
   1479 		} else
   1480 			ret = ENOENT;
   1481 	} else
   1482 		ret = dirtopath(vrootp, vp, buf, buflen, flags, cr);
   1483 
   1484 	VN_RELE(vrootp);
   1485 	if (doclose) {
   1486 		(void) VOP_CLOSE(vp, FREAD, 1, 0, cr, NULL);
   1487 		VN_RELE(vp);
   1488 	}
   1489 
   1490 	return (ret);
   1491 }
   1492 
   1493 int
   1494 vnodetopath(vnode_t *vrootp, vnode_t *vp, char *buf, size_t buflen, cred_t *cr)
   1495 {
   1496 	return (vnodetopath_common(vrootp, vp, buf, buflen, cr, 0));
   1497 }
   1498 
   1499 int
   1500 dogetcwd(char *buf, size_t buflen)
   1501 {
   1502 	int ret;
   1503 	vnode_t *vp;
   1504 	vnode_t *compvp;
   1505 	refstr_t *cwd, *oldcwd;
   1506 	const char *value;
   1507 	pathname_t rpnp, pnp;
   1508 	proc_t *p = curproc;
   1509 
   1510 	/*
   1511 	 * Check to see if there is a cached version of the cwd.  If so, lookup
   1512 	 * the cached value and make sure it is the same vnode.
   1513 	 */
   1514 	mutex_enter(&p->p_lock);
   1515 	if ((cwd = PTOU(p)->u_cwd) != NULL)
   1516 		refstr_hold(cwd);
   1517 	vp = PTOU(p)->u_cdir;
   1518 	VN_HOLD(vp);
   1519 	mutex_exit(&p->p_lock);
   1520 
   1521 	/*
   1522 	 * Make sure we have permission to access the current directory.
   1523 	 */
   1524 	if ((ret = VOP_ACCESS(vp, VEXEC, 0, CRED(), NULL)) != 0) {
   1525 		if (cwd != NULL)
   1526 			refstr_rele(cwd);
   1527 		VN_RELE(vp);
   1528 		return (ret);
   1529 	}
   1530 
   1531 	if (cwd) {
   1532 		value = refstr_value(cwd);
   1533 		if ((ret = pn_get((char *)value, UIO_SYSSPACE, &pnp)) != 0) {
   1534 			refstr_rele(cwd);
   1535 			VN_RELE(vp);
   1536 			return (ret);
   1537 		}
   1538 
   1539 		pn_alloc(&rpnp);
   1540 
   1541 		if (lookuppn(&pnp, &rpnp, NO_FOLLOW, NULL, &compvp) == 0) {
   1542 
   1543 			if (VN_CMP(vp, compvp) &&
   1544 			    strcmp(value, rpnp.pn_path) == 0) {
   1545 				VN_RELE(compvp);
   1546 				VN_RELE(vp);
   1547 				pn_free(&pnp);
   1548 				pn_free(&rpnp);
   1549 				if (strlen(value) + 1 > buflen) {
   1550 					refstr_rele(cwd);
   1551 					return (ENAMETOOLONG);
   1552 				}
   1553 				bcopy(value, buf, strlen(value) + 1);
   1554 				refstr_rele(cwd);
   1555 				return (0);
   1556 			}
   1557 
   1558 			VN_RELE(compvp);
   1559 		}
   1560 
   1561 		pn_free(&rpnp);
   1562 		pn_free(&pnp);
   1563 
   1564 		refstr_rele(cwd);
   1565 	}
   1566 
   1567 	ret = vnodetopath_common(NULL, vp, buf, buflen, CRED(),
   1568 	    LOOKUP_CHECKREAD);
   1569 
   1570 	VN_RELE(vp);
   1571 
   1572 	/*
   1573 	 * Store the new cwd and replace the existing cached copy.
   1574 	 */
   1575 	if (ret == 0)
   1576 		cwd = refstr_alloc(buf);
   1577 	else
   1578 		cwd = NULL;
   1579 
   1580 	mutex_enter(&p->p_lock);
   1581 	oldcwd = PTOU(p)->u_cwd;
   1582 	PTOU(p)->u_cwd = cwd;
   1583 	mutex_exit(&p->p_lock);
   1584 
   1585 	if (oldcwd)
   1586 		refstr_rele(oldcwd);
   1587 
   1588 	return (ret);
   1589 }
   1590