From 286ace1259992bd0c5d9016715833f2e148ac596 Mon Sep 17 00:00:00 2001
From: "Andrew G. Morgan" <morgan@kernel.org>
Date: Thu, 12 Mar 2026 07:38:05 -0700
Subject: [PATCH] Address a potential TOCTOU race condition in cap_set_file().

This issue was researched and reported by Ali Raza (@locus-x64). It
has been assigned CVE-2026-4878.

The finding is that while cap_set_file() checks if a file is a regular
file before applying or removing a capability attribute, a small
window existed after that check when the filepath could be overwritten
either with new content or a symlink to some other file. To do this
would imply that the caller of cap_set_file() was directing it to a
directory over which a local attacker has write access, and performed
the operation frequently enough that an attacker had a non-negligible
chance of exploiting the race condition. The code now locks onto the
intended file, eliminating the race condition.

Signed-off-by: Andrew G. Morgan <morgan@kernel.org>
---
 libcap/cap_file.c  | 69 +++++++++++++++++++++++++++++++++++++++-------
 progs/quicktest.sh | 14 +++++++++-
 2 files changed, 72 insertions(+), 11 deletions(-)

diff --git a/libcap/cap_file.c b/libcap/cap_file.c
index 0bc07f7..f02bf9f 100644
--- a/libcap/cap_file.c
+++ b/libcap/cap_file.c
@@ -8,8 +8,13 @@
 #define _DEFAULT_SOURCE
 #endif
 
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+
 #include <sys/types.h>
 #include <byteswap.h>
+#include <fcntl.h>
 #include <sys/stat.h>
 #include <unistd.h>
 
@@ -322,26 +327,70 @@ int cap_set_file(const char *filename, cap_t cap_d)
     struct vfs_ns_cap_data rawvfscap;
     int sizeofcaps;
     struct stat buf;
+    char fdpath[64];
+    int fd, ret;
+
+    _cap_debug("setting filename capabilities");
+    fd = open(filename, O_RDONLY|O_NOFOLLOW);
+    if (fd >= 0) {
+	ret = cap_set_fd(fd, cap_d);
+	close(fd);
+	return ret;
+    }
 
-    if (lstat(filename, &buf) != 0) {
-	_cap_debug("unable to stat file [%s]", filename);
+    /*
+     * Attempting to set a file capability on a file the process can't
+     * read the content of. This is considered a non-standard use case
+     * and the following (slower) code is complicated because it is
+     * trying to avoid a TOCTOU race condition.
+     */
+
+    fd = open(filename, O_PATH|O_NOFOLLOW);
+    if (fd < 0) {
+	_cap_debug("cannot find file at path [%s]", filename);
+	return -1;
+    }
+    if (fstat(fd, &buf) != 0) {
+	_cap_debug("unable to stat file [%s] descriptor %d",
+		   filename, fd);
+	close(fd);
 	return -1;
     }
     if (S_ISLNK(buf.st_mode) || !S_ISREG(buf.st_mode)) {
-	_cap_debug("file [%s] is not a regular file", filename);
+	_cap_debug("file [%s] descriptor %d for non-regular file",
+		   filename, fd);
+	close(fd);
 	errno = EINVAL;
 	return -1;
     }
 
-    if (cap_d == NULL) {
-	_cap_debug("removing filename capabilities");
-	return removexattr(filename, XATTR_NAME_CAPS);
+    /*
+     * While the fd remains open, this named file is locked to the
+     * origin regular file. The size of the fdpath variable is
+     * sufficient to support a 160+ bit number.
+     */
+    if (snprintf(fdpath, sizeof(fdpath), "/proc/self/fd/%d", fd)
+	>= sizeof(fdpath)) {
+	_cap_debug("file descriptor too large %d", fd);
+	errno = EINVAL;
+	ret = -1;
+
+    } else if (cap_d == NULL) {
+	_cap_debug("dropping file caps on [%s] via [%s]",
+		   filename, fdpath);
+	ret = removexattr(fdpath, XATTR_NAME_CAPS);
+
     } else if (_fcaps_save(&rawvfscap, cap_d, &sizeofcaps) != 0) {
-	return -1;
-    }
+	_cap_debug("problem converting cap_d to vfscap format");
+	ret = -1;
 
-    _cap_debug("setting filename capabilities");
-    return setxattr(filename, XATTR_NAME_CAPS, &rawvfscap, sizeofcaps, 0);
+    } else {
+	_cap_debug("setting filename capabilities");
+	ret = setxattr(fdpath, XATTR_NAME_CAPS, &rawvfscap,
+		       sizeofcaps, 0);
+    }
+    close(fd);
+    return ret;
 }
 
 /*
diff --git a/progs/quicktest.sh b/progs/quicktest.sh
index e6c48e6..5dc72f9 100755
--- a/progs/quicktest.sh
+++ b/progs/quicktest.sh
@@ -148,7 +148,19 @@ pass_capsh --caps="cap_setpcap=p" --inh=cap_chown --current
 pass_capsh --strict --caps="cap_chown=p" --inh=cap_chown --current
 
 # change the way the capability is obtained (make it inheritable)
+chmod 0000 ./privileged
 ./setcap cap_setuid,cap_setgid=ei ./privileged
+if [ $? -ne 0 ]; then
+    echo "FAILED to set file capability"
+    exit 1
+fi
+chmod 0755 ./privileged
+ln -s privileged unprivileged
+./setcap -r ./unprivileged
+if [ $? -eq 0 ]; then
+    echo "FAILED by removing a capability from a symlinked file"
+    exit 1
+fi
 
 # Note, the bounding set (edited with --drop) only limits p
 # capabilities, not i's.
@@ -246,7 +258,7 @@ EOF
     pass_capsh --iab='!%cap_chown,^cap_setpcap,cap_setuid'
     fail_capsh --mode=PURE1E --iab='!%cap_chown,^cap_setuid'
 fi
-/bin/rm -f ./privileged
+/bin/rm -f ./privileged ./unprivileged
 
 echo "testing namespaced file caps"
 
-- 
2.51.0

