vfs: provide F_ISUNIONSTACK as a kludge for libc

Prior to introduction of this op libc's readdir would call fstatfs(2), in
effect unnecessarily copying kilobytes of data just to check fs name and a
mount flag.

Reviewed by:	kib (previous version)
Differential Revision:	https://reviews.freebsd.org/D23162

lib/libc/gen/opendir.c

@@ -273,6 +273,24 @@ _filldir(DIR *dirp, bool use_current_pos)
 	return (true);
 }
 
+static bool
+is_unionstack(int fd)
+{
+	struct statfs sfb;
+	int unionstack;
+
+	unionstack = _fcntl(fd, F_ISUNIONSTACK);
+	if (unionstack != -1)
+		return (unionstack);
+
+	/*
+	 * Temporary compat for kernels which don't provide F_ISUNIONSTACK.
+	 */
+	if (_fstatfs(fd, &sfb) < 0)
+		return (true);
+	return (strcmp(sfb.f_fstypename, "unionfs") == 0 ||
+	    (sfb.f_flags & MNT_UNION));
+}
 
 /*
  * Common routine for opendir(3), __opendir2(3) and fdopendir(3).
@@ -312,12 +330,7 @@ __opendir_common(int fd, int flags, bool use_current_pos)
 	 */
 	unionstack = false;
 	if (flags & DTF_NODUP) {
-		struct statfs sfb;
-
-		if (_fstatfs(fd, &sfb) == 0) {
-			unionstack = strcmp(sfb.f_fstypename, "unionfs") == 0 ||
-			    (sfb.f_flags & MNT_UNION);
-		}
+		unionstack = is_unionstack(fd);
 	}
 
 	if (unionstack) {

lib/libc/sys/fcntl.2

@@ -28,7 +28,7 @@
 .\"     @(#)fcntl.2	8.2 (Berkeley) 1/12/94
 .\" $FreeBSD$
 .\"
-.Dd September 4, 2019
+.Dd January 17, 2020
 .Dt FCNTL 2
 .Os
 .Sh NAME
@@ -185,6 +185,11 @@ Add seals to the file as described below, if the underlying filesystem supports
 seals.
 .It Dv F_GET_SEALS
 Get seals associated with the file, if the underlying filesystem supports seals.
+.It Dv F_ISUNIONSTACK
+Check if the vnode is part of a union stack (either the "union" flag from
+.Xr mount 2
+or unionfs).
+This is a hack not intended to be used outside of libc.
 .El
 .Pp
 The flags for the

sys/fs/unionfs/union_vfsops.c

@@ -296,7 +296,7 @@ unionfs_domount(struct mount *mp)
 	if ((ump->um_lowervp->v_mount->mnt_flag & MNT_LOCAL) &&
 	    (ump->um_uppervp->v_mount->mnt_flag & MNT_LOCAL))
 		mp->mnt_flag |= MNT_LOCAL;
-	mp->mnt_kern_flag |= MNTK_NOMSYNC;
+	mp->mnt_kern_flag |= MNTK_NOMSYNC | MNTK_UNIONFS;
 	MNT_IUNLOCK(mp);
 
 	/*

sys/kern/kern_descrip.c

@@ -489,6 +489,7 @@ kern_fcntl(struct thread *td, int fd, int cmd, intptr_t arg)
 	struct filedescent *fde;
 	struct proc *p;
 	struct vnode *vp;
+	struct mount *mp;
 	int error, flg, seals, tmp;
 	uint64_t bsize;
 	off_t foffset;
@@ -816,6 +817,49 @@ kern_fcntl(struct thread *td, int fd, int cmd, intptr_t arg)
 		fdrop(fp, td);
 		break;
 
+	case F_ISUNIONSTACK:
+		/*
+		 * Check if the vnode is part of a union stack (either the
+		 * "union" flag from mount(2) or unionfs).
+		 *
+		 * Prior to introduction of this op libc's readdir would call
+		 * fstatfs(2), in effect unnecessarily copying kilobytes of
+		 * data just to check fs name and a mount flag.
+		 *
+		 * Fixing the code to handle everything in the kernel instead
+		 * is a non-trivial endeavor and has low priority, thus this
+		 * horrible kludge facilitates the current behavior in a much
+		 * cheaper manner until someone(tm) sorts this out.
+		 */
+		error = fget_unlocked(fdp, fd, &cap_no_rights, &fp, NULL);
+		if (error != 0)
+			break;
+		if (fp->f_type != DTYPE_VNODE) {
+			fdrop(fp, td);
+			error = EBADF;
+			break;
+		}
+		vp = fp->f_vnode;
+		/*
+		 * Since we don't prevent dooming the vnode even non-null mp
+		 * found can become immediately stale. This is tolerable since
+		 * mount points are type-stable (providing safe memory access)
+		 * and any vfs op on this vnode going forward will return an
+		 * error (meaning return value in this case is meaningless).
+		 */
+		mp = (struct mount *)atomic_load_ptr(&vp->v_mount);
+		if (__predict_false(mp == NULL)) {
+			fdrop(fp, td);
+			error = EBADF;
+			break;
+		}
+		td->td_retval[0] = 0;
+		if (mp->mnt_kern_flag & MNTK_UNIONFS ||
+		    mp->mnt_flag & MNT_UNION)
+			td->td_retval[0] = 1;
+		fdrop(fp, td);
+		break;
+
 	default:
 		error = EINVAL;
 		break;

sys/sys/fcntl.h

@@ -250,6 +250,7 @@ typedef	__pid_t		pid_t;
 #define	F_DUP2FD_CLOEXEC 18		/* Like F_DUP2FD, but FD_CLOEXEC is set */
 #define	F_ADD_SEALS	19
 #define	F_GET_SEALS	20
+#define	F_ISUNIONSTACK	21		/* Kludge for libc, don't use it. */
 
 /* Seals (F_ADD_SEALS, F_GET_SEALS). */
 #define	F_SEAL_SEAL	0x0001		/* Prevent adding sealings */

sys/sys/mount.h

@@ -414,6 +414,7 @@ void          __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp);
 #define	MNTK_USES_BCACHE	0x00004000 /* FS uses the buffer cache. */
 #define	MNTK_TEXT_REFS		0x00008000 /* Keep use ref for text */
 #define	MNTK_VMSETSIZE_BUG	0x00010000
+#define	MNTK_UNIONFS	0x00020000	/* A hack for F_ISUNIONSTACK */
 #define MNTK_NOASYNC	0x00800000	/* disable async */
 #define MNTK_UNMOUNT	0x01000000	/* unmount in progress */
 #define	MNTK_MWAIT	0x02000000	/* waiting for unmount to finish */