1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
4  * Copyright (c) 1994 Jan-Simon Pendry
11  * Jan-Simon Pendry.
66 #define UNIONFSHASHMASK (NUNIONFSNODECACHE - 1)
122 		MPASS(unp->un_dvp != NULL);  in unionfs_deferred_rele()
123 		vrele(unp->un_dvp);  in unionfs_deferred_rele()
127 	/* We expect this function to be single-threaded, thus no atomic */  in unionfs_deferred_rele()
138 	return (&(unp->un_hashtbl[vfs_hash_index(lookup) & UNIONFSHASHMASK]));  in unionfs_get_hashhead()
155 		if (unp->un_uppervp == lookup ||  in unionfs_get_cached_vnode_locked()
156 		    unp->un_lowervp == lookup) {  in unionfs_get_cached_vnode_locked()
159 			vp->v_iflag &= ~VI_OWEINACT;  in unionfs_get_cached_vnode_locked()
161 			    ((vp->v_iflag & VI_DOINGINACT) != 0)) {  in unionfs_get_cached_vnode_locked()
208 	if (uncp->un_uppervp != NULL) {  in unionfs_ins_cached_vnode()
209 		ASSERT_VOP_ELOCKED(uncp->un_uppervp, __func__);  in unionfs_ins_cached_vnode()
210 		KASSERT(uncp->un_uppervp->v_type == VDIR,  in unionfs_ins_cached_vnode()
212 		vp = unionfs_get_cached_vnode_locked(uncp->un_uppervp, dvp);  in unionfs_ins_cached_vnode()
213 	} else if (uncp->un_lowervp != NULL) {  in unionfs_ins_cached_vnode()
214 		ASSERT_VOP_ELOCKED(uncp->un_lowervp, __func__);  in unionfs_ins_cached_vnode()
215 		KASSERT(uncp->un_lowervp->v_type == VDIR,  in unionfs_ins_cached_vnode()
217 		vp = unionfs_get_cached_vnode_locked(uncp->un_lowervp, dvp);  in unionfs_ins_cached_vnode()
220 		hd = unionfs_get_hashhead(dvp, (uncp->un_uppervp != NULL ?  in unionfs_ins_cached_vnode()
221 						uncp->un_uppervp : uncp->un_lowervp));  in unionfs_ins_cached_vnode()
237 		("%s: null parent vnode", __func__));  in unionfs_rem_cached_vnode()
240 	if (unp->un_hash.le_prev != NULL) {  in unionfs_rem_cached_vnode()
242 		unp->un_hash.le_next = NULL;  in unionfs_rem_cached_vnode()
243 		unp->un_hash.le_prev = NULL;  in unionfs_rem_cached_vnode()
250  * Upper, lower, and parent directory vnodes are expected to be referenced by
251  * the caller.  Upper and lower vnodes, if non-NULL, are also expected to be
252  * exclusively locked by the caller.
260 	 * Lock and reset the default vnode lock; vgone() expects a locked  in unionfs_nodeget_cleanup()
263 	lockmgr(&vp->v_lock, LK_EXCLUSIVE, NULL);  in unionfs_nodeget_cleanup()
270 	vp->v_data = NULL;  in unionfs_nodeget_cleanup()
271 	vp->v_vnlock = &vp->v_lock;  in unionfs_nodeget_cleanup()
272 	vp->v_op = &dead_vnodeops;  in unionfs_nodeget_cleanup()
277 	if (unp->un_dvp != NULL)  in unionfs_nodeget_cleanup()
278 		vrele(unp->un_dvp);  in unionfs_nodeget_cleanup()
279 	if (unp->un_uppervp != NULL) {  in unionfs_nodeget_cleanup()
280 		vput(unp->un_uppervp);  in unionfs_nodeget_cleanup()
281 		if (unp->un_lowervp != NULL)  in unionfs_nodeget_cleanup()
282 			vrele(unp->un_lowervp);  in unionfs_nodeget_cleanup()
283 	} else if (unp->un_lowervp != NULL)  in unionfs_nodeget_cleanup()
284 		vput(unp->un_lowervp);  in unionfs_nodeget_cleanup()
285 	if (unp->un_hashtbl != NULL)  in unionfs_nodeget_cleanup()
286 		hashdestroy(unp->un_hashtbl, M_UNIONFSHASH, UNIONFSHASHMASK);  in unionfs_nodeget_cleanup()
287 	free(unp->un_path, M_UNIONFSPATH);  in unionfs_nodeget_cleanup()
295  * locked, uppervp or lowervp is locked too. In order to prevent dead lock,
314 	lkflags = (cnp ? cnp->cn_lkflags : 0);  in unionfs_nodeget()
315 	path = (cnp ? cnp->cn_nameptr : NULL);  in unionfs_nodeget()
321 	vt = (uppervp != NULL ? uppervp->v_type : lowervp->v_type);  in unionfs_nodeget()
324 	if (cnp && !(cnp->cn_flags & ISLASTCN))  in unionfs_nodeget()
354 		unp->un_hashtbl = hashinit(NUNIONFSNODECACHE, M_UNIONFSHASH,  in unionfs_nodeget()
360 	unp->un_vnode = vp;  in unionfs_nodeget()
361 	unp->un_uppervp = uppervp;  in unionfs_nodeget()
362 	unp->un_lowervp = lowervp;  in unionfs_nodeget()
363 	unp->un_dvp = dvp;  in unionfs_nodeget()
365 		vp->v_vnlock = uppervp->v_vnlock;  in unionfs_nodeget()
367 		vp->v_vnlock = lowervp->v_vnlock;  in unionfs_nodeget()
370 		unp->un_path = malloc(cnp->cn_namelen + 1,  in unionfs_nodeget()
372 		bcopy(cnp->cn_nameptr, unp->un_path, cnp->cn_namelen);  in unionfs_nodeget()
373 		unp->un_path[cnp->cn_namelen] = '\0';  in unionfs_nodeget()
374 		unp->un_pathlen = cnp->cn_namelen;  in unionfs_nodeget()
376 	vp->v_type = vt;  in unionfs_nodeget()
377 	vp->v_data = unp;  in unionfs_nodeget()
387 	if (ump->um_uppervp == uppervp || ump->um_lowervp == lowervp)  in unionfs_nodeget()
388 		vp->v_vflag |= VV_ROOT;  in unionfs_nodeget()
389 	KASSERT(dvp != NULL || (vp->v_vflag & VV_ROOT) != 0,  in unionfs_nodeget()
390 	    ("%s: NULL dvp for non-root vp %p", __func__, vp));  in unionfs_nodeget()
394 	 * NOTE: There is still a possibility for cross-filesystem locking here.  in unionfs_nodeget()
395 	 * If dvp has an upper FS component and is locked, while the new vnode  in unionfs_nodeget()
396 	 * created here only has a lower-layer FS component, then we will end  in unionfs_nodeget()
397 	 * up taking a lower-FS lock while holding an upper-FS lock.  in unionfs_nodeget()
400 	 * a child vnode lock is taken while holding a lock on its parent  in unionfs_nodeget()
402 	 * a few places within unionfs (which could lead to the same cross-FS  in unionfs_nodeget()
473 	 * which is locked across VFS_UNMOUNT().  This lock will then be  in unionfs_noderem()
479 	KASSERT(vp->v_vnlock->lk_recurse == 0 || (vp->v_vflag & VV_ROOT) != 0,  in unionfs_noderem()
480 	    ("%s: vnode %p locked recursively", __func__, vp));  in unionfs_noderem()
484 	lvp = unp->un_lowervp;  in unionfs_noderem()
485 	uvp = unp->un_uppervp;  in unionfs_noderem()
486 	dvp = unp->un_dvp;  in unionfs_noderem()
496 	 * Moreover, during unmount of a non-"below" unionfs mount, the lower  in unionfs_noderem()
497 	 * root vnode will already be locked as it is the covered vnode.  in unionfs_noderem()
499 	if (uvp != NULL && lvp != NULL && (vp->v_vflag & VV_ROOT) == 0) {  in unionfs_noderem()
504 	if (lockmgr(&vp->v_lock, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0)  in unionfs_noderem()
511 	unp->un_lowervp = unp->un_uppervp = NULL;  in unionfs_noderem()
512 	vp->v_vnlock = &(vp->v_lock);  in unionfs_noderem()
513 	vp->v_data = NULL;  in unionfs_noderem()
514 	vp->v_object = NULL;  in unionfs_noderem()
515 	if (unp->un_hashtbl != NULL) {  in unionfs_noderem()
519 		 * be reclaimed with a non-zero use count.  Otherwise the  in unionfs_noderem()
523 			hd = unp->un_hashtbl + count;  in unionfs_noderem()
526 				unp_t1->un_hash.le_next = NULL;  in unionfs_noderem()
527 				unp_t1->un_hash.le_prev = NULL;  in unionfs_noderem()
533 	writerefs = atomic_load_int(&vp->v_writecount);  in unionfs_noderem()
544 		VOP_ADD_WRITECOUNT(uvp, -writerefs);  in unionfs_noderem()
556 	if (unp->un_path != NULL) {  in unionfs_noderem()
557 		free(unp->un_path, M_UNIONFSPATH);  in unionfs_noderem()
558 		unp->un_path = NULL;  in unionfs_noderem()
559 		unp->un_pathlen = 0;  in unionfs_noderem()
562 	if (unp->un_hashtbl != NULL) {  in unionfs_noderem()
563 		hashdestroy(unp->un_hashtbl, M_UNIONFSHASH, UNIONFSHASHMASK);  in unionfs_noderem()
566 	LIST_FOREACH_SAFE(unsp, &(unp->un_unshead), uns_list, unsp_tmp) {  in unionfs_noderem()
591 	pid = td->td_proc->p_pid;  in unionfs_find_node_status()
595 	LIST_FOREACH(unsp, &(unp->un_unshead), uns_list) {  in unionfs_find_node_status()
596 		if (unsp->uns_pid == pid) {  in unionfs_find_node_status()
617 	pid = td->td_proc->p_pid;  in unionfs_get_node_status()
626 		unsp->uns_pid = pid;  in unionfs_get_node_status()
627 		LIST_INSERT_HEAD(&(unp->un_unshead), unsp, uns_list);  in unionfs_get_node_status()
644 	if (0 < unsp->uns_lower_opencnt || 0 < unsp->uns_upper_opencnt)  in unionfs_tryrem_node_status()
659 	uva->va_type = lva->va_type;  in unionfs_create_uppervattr_core()
660 	uva->va_atime = lva->va_atime;  in unionfs_create_uppervattr_core()
661 	uva->va_mtime = lva->va_mtime;  in unionfs_create_uppervattr_core()
662 	uva->va_ctime = lva->va_ctime;  in unionfs_create_uppervattr_core()
664 	switch (ump->um_copymode) {  in unionfs_create_uppervattr_core()
666 		uva->va_mode = lva->va_mode;  in unionfs_create_uppervattr_core()
667 		uva->va_uid = lva->va_uid;  in unionfs_create_uppervattr_core()
668 		uva->va_gid = lva->va_gid;  in unionfs_create_uppervattr_core()
671 		if (ump->um_uid == lva->va_uid) {  in unionfs_create_uppervattr_core()
672 			uva->va_mode = lva->va_mode & 077077;  in unionfs_create_uppervattr_core()
673 			uva->va_mode |= (lva->va_type == VDIR ?  in unionfs_create_uppervattr_core()
674 			    ump->um_udir : ump->um_ufile) & 0700;  in unionfs_create_uppervattr_core()
675 			uva->va_uid = lva->va_uid;  in unionfs_create_uppervattr_core()
676 			uva->va_gid = lva->va_gid;  in unionfs_create_uppervattr_core()
678 			uva->va_mode = (lva->va_type == VDIR ?  in unionfs_create_uppervattr_core()
679 			    ump->um_udir : ump->um_ufile);  in unionfs_create_uppervattr_core()
680 			uva->va_uid = ump->um_uid;  in unionfs_create_uppervattr_core()
681 			uva->va_gid = ump->um_gid;  in unionfs_create_uppervattr_core()
685 		uva->va_mode = 0777 & ~td->td_proc->p_pd->pd_cmask;  in unionfs_create_uppervattr_core()
686 		uva->va_uid = ump->um_uid;  in unionfs_create_uppervattr_core()
687 		uva->va_gid = ump->um_gid;  in unionfs_create_uppervattr_core()
713  * dvp should be locked on entry and will be locked on return.
716  * locked, referenced vnode. If *vpp == dvp then remember that only one
727 	cn->cn_namelen = pathlen;  in unionfs_relookup()
728 	cn->cn_pnbuf = path;  in unionfs_relookup()
729 	cn->cn_nameiop = nameiop;  in unionfs_relookup()
730 	cn->cn_flags = (LOCKPARENT | LOCKLEAF | ISLASTCN);  in unionfs_relookup()
731 	cn->cn_lkflags = LK_EXCLUSIVE;  in unionfs_relookup()
732 	cn->cn_cred = cnp->cn_cred;  in unionfs_relookup()
733 	cn->cn_nameptr = cn->cn_pnbuf;  in unionfs_relookup()
737 		cn->cn_flags |= (cnp->cn_flags & DOWHITEOUT);  in unionfs_relookup()
741 		cn->cn_flags |= NOCACHE;  in unionfs_relookup()
752 	KASSERT(cn->cn_pnbuf == path, ("%s: cn_pnbuf changed", __func__));  in unionfs_relookup()
760  * uvp is new locked upper vnode. unionfs vnode's lock will be exchanged to the
774 	lvp = unp->un_lowervp;  in unionfs_node_update()
777 	dvp = unp->un_dvp;  in unionfs_node_update()
779 	VNASSERT(vp->v_writecount == 0, vp,  in unionfs_node_update()
780 	    ("%s: non-zero writecount", __func__));  in unionfs_node_update()
785 	lockrec = lvp->v_vnlock->lk_recurse;  in unionfs_node_update()
789 	unp->un_uppervp = uvp;  in unionfs_node_update()
790 	vp->v_vnlock = uvp->v_vnlock;  in unionfs_node_update()
796 	 * Re-cache the unionfs vnode against the upper vnode  in unionfs_node_update()
798 	if (dvp != NULL && vp->v_type == VDIR) {  in unionfs_node_update()
800 		if (unp->un_hash.le_prev != NULL) {  in unionfs_node_update()
805 		VI_UNLOCK(unp->un_dvp);  in unionfs_node_update()
812  * This is useful, for example, during copy-up operations in which
814  * possibility of a concurrent copy-up on the same vnode triggering
827 	while (error == 0 && (unp->un_flag & flag) != 0) {  in unionfs_set_in_progress_flag()
834 			 * If we waited on a concurrent copy-up and that  in unionfs_set_in_progress_flag()
835 			 * copy-up was successful, return a non-fatal  in unionfs_set_in_progress_flag()
839 			 * be re-queried to avoid creating a duplicate unionfs  in unionfs_set_in_progress_flag()
846 			    unp->un_uppervp != NULL)  in unionfs_set_in_progress_flag()
853 		unp->un_flag |= flag;  in unionfs_set_in_progress_flag()
868 		VNASSERT((unp->un_flag & flag) != 0, vp,  in unionfs_clear_in_progress_flag()
870 		unp->un_flag &= ~flag;  in unionfs_clear_in_progress_flag()
879  * dvp and vp are unionfs vnodes representing a parent directory and
880  * child file, should be locked on entry, and will be locked on return.
905 	ump = MOUNTTOUNIONFSMOUNT(vp->v_mount);  in unionfs_mkshadowdir()
907 	if (unp->un_uppervp != NULL)  in unionfs_mkshadowdir()
910 	udvp = dunp->un_uppervp;  in unionfs_mkshadowdir()
918 	lvp = unp->un_lowervp;  in unionfs_mkshadowdir()
920 	credbk = cnp->cn_cred;  in unionfs_mkshadowdir()
924 	cred = crdup(cnp->cn_cred);  in unionfs_mkshadowdir()
929 	cnp->cn_cred = cred;  in unionfs_mkshadowdir()
934 	if ((error = VOP_GETATTR(lvp, &lva, cnp->cn_cred)))  in unionfs_mkshadowdir()
940 	    cnp->cn_nameptr, cnp->cn_namelen, CREATE))) {  in unionfs_mkshadowdir()
976 	 * Temporarily NUL-terminate the current pathname component.  in unionfs_mkshadowdir()
986 	 * NUL-termination will not have an effect on other threads.  in unionfs_mkshadowdir()
1045 	cnp->cn_cred = credbk;  in unionfs_mkshadowdir()
1063  * The unionfs vnode shares its lock with the base-layer vnode(s); if the
1077  *    unionfs vnode holds a reference on the base-layer vnode(s).  But
1082  * basevp1 and basevp2 represent two base-layer vnodes which are
1083  * expected to be locked when this function is called.  basevp2
1085  * a parent directory and a filed linked to it, respectively.
1097 	 * Take an additional reference on the base-layer vnodes to  in unionfs_forward_vop_start_pair()
1133  * basevp1 and basevp2 (if not NULL) must be locked when this function
1137  * used to re-lock unionvp1 and unionvp2, respectively, if either
1152 	 * a flag indicating that it needs to be re-locked.  in unionfs_forward_vop_finish_pair()
1153 	 * Otherwise, simply drop the base-vnode reference that  in unionfs_forward_vop_finish_pair()
1164 	 * If any of the unionfs vnodes need to be re-locked, that  in unionfs_forward_vop_finish_pair()
1165 	 * means the unionfs vnode's lock is now de-coupled from the  in unionfs_forward_vop_finish_pair()
1176 		 * be able to safely unlock it with basevp2 still locked,  in unionfs_forward_vop_finish_pair()
1184 		 * scenario would require at least re-locking the unionfs  in unionfs_forward_vop_finish_pair()
1205  * dvp and vp are unionfs vnodes representing a parent directory and
1206  * child file, should be locked on entry, and will be locked on return.
1222 	udvp = VTOUNIONFS(dvp)->un_uppervp;  in unionfs_mkwhiteout()
1265  * locked, referenced and opened vnode.
1285 	ump = MOUNTTOUNIONFSMOUNT(UNIONFSTOV(unp)->v_mount);  in unionfs_vn_create_on_upper()
1287 	lvp = unp->un_lowervp;  in unionfs_vn_create_on_upper()
1288 	cred = td->td_ucred;  in unionfs_vn_create_on_upper()
1296 	if (unp->un_path == NULL)  in unionfs_vn_create_on_upper()
1299 	nd.ni_cnd.cn_namelen = unp->un_pathlen;  in unionfs_vn_create_on_upper()
1300 	nd.ni_cnd.cn_pnbuf = unp->un_path;  in unionfs_vn_create_on_upper()
1333 	    __func__, uvp, uvp->v_writecount);  in unionfs_vn_create_on_upper()
1348  * lvp and uvp should be locked and opened on entry and will be locked and
1384 		if ((count = MAXBSIZE - uio.uio_resid) == 0)  in unionfs_copyfile_core()
1392 			iov.iov_len = count - bufoffset;  in unionfs_copyfile_core()
1400 			bufoffset += (count - bufoffset) - uio.uio_resid;  in unionfs_copyfile_core()
1417  * vp is a unionfs vnode that should be locked on entry and will be
1418  * locked on return.
1437 	lvp = unp->un_lowervp;  in unionfs_copyfile()
1440 	if ((UNIONFSTOV(unp)->v_mount->mnt_flag & MNT_RDONLY))  in unionfs_copyfile()
1442 	if (unp->un_dvp == NULL)  in unionfs_copyfile()
1444 	if (unp->un_uppervp != NULL)  in unionfs_copyfile()
1448 	VI_LOCK(unp->un_dvp);  in unionfs_copyfile()
1449 	dunp = VTOUNIONFS(unp->un_dvp);  in unionfs_copyfile()
1451 		udvp = dunp->un_uppervp;  in unionfs_copyfile()
1452 	VI_UNLOCK(unp->un_dvp);  in unionfs_copyfile()
1456 	if ((udvp->v_mount->mnt_flag & MNT_RDONLY))  in unionfs_copyfile()
1481 	 * uvp below while holding vp[=lvp] locked.  Replacing  in unionfs_copyfile()
1501 	VOP_ADD_WRITECOUNT_CHECKED(uvp, -1);  in unionfs_copyfile()
1503 	    __func__, uvp, uvp->v_writecount);  in unionfs_copyfile()
1527  * then cross-check any files we find against the upper FS to see if
1529  * non-present).
1533  * vp should be a locked unionfs node, and vp's lowervp should also be
1534  * locked.
1558 	 * parent directory lock, which it can do in a manner that avoids  in unionfs_check_rmdir()
1559 	 * deadlock.  However, if the cross-check logic below needs to call  in unionfs_check_rmdir()
1563 	 * below, such as VOP_OPEN() and VOP_READDIR(), may also re-lock the  in unionfs_check_rmdir()
1566 	 * (and its parent directory lock) and the lower vnode lock as needed,  in unionfs_check_rmdir()
1570 	 * read-only) may change while the relevant lock is dropped.  But  in unionfs_check_rmdir()
1571 	 * since re-locking may happen here and open up such a window anyway,  in unionfs_check_rmdir()
1575 	 * of out-of-band state changes) that we can avoid these costly checks  in unionfs_check_rmdir()
1618 		if (dp->d_type == DT_WHT)  in unionfs_check_rmdir()
1625 		switch (dp->d_namlen) {  in unionfs_check_rmdir()
1627 			if (dp->d_name[1] != '.') {  in unionfs_check_rmdir()
1633 			if (dp->d_name[0] != '.') {  in unionfs_check_rmdir()
1642 		cn.cn_namelen = dp->d_namlen;  in unionfs_check_rmdir()
1644 		cn.cn_nameptr = dp->d_name;  in unionfs_check_rmdir()