Petr Lautrbach <plautrba@xxxxxxxxxx> writes:
> Stephen Smalley <stephen.smalley.work@xxxxxxxxx> writes:
>
>> On Wed, Jan 13, 2021 at 4:52 PM Petr Lautrbach <plautrba@xxxxxxxxxx> wrote:
>>>
>>> Hi,
>>>
>>> we have few tests which uses `umount /sys/fs/selinux` trick to check how
>>> userspace works in SELinux "disabled" environment. But it's not possible
>>> with the current master:
>>>
>>> # umount /sys/fs/selinux
>>> umount: /sys/fs/selinux: target is busy.
>>>
>>> # lsof /sys/fs/selinux
>>> COMMAND PID USER FD TYPE DEVICE SIZE/OFF NODE NAME
>>> systemd 1 root mem REG 0,21 0 19 /sys/fs/selinux/status
>>> systemd 1 root 55r REG 0,21 0 19 /sys/fs/selinux/status
>>> systemd-u 875 root mem REG 0,21 0 19 /sys/fs/selinux/status
>>> systemd-u 875 root 6r REG 0,21 0 19 /sys/fs/selinux/status
>>> dbus-brok 1116 dbus mem REG 0,21 0 19 /sys/fs/selinux/status
>>> dbus-brok 1116 dbus 5r REG 0,21 0 19 /sys/fs/selinux/status
>>> systemd-l 1134 root mem REG 0,21 0 19 /sys/fs/selinux/status
>>> systemd-l 1134 root 4r REG 0,21 0 19 /sys/fs/selinux/status
>>> systemd 1643 root mem REG 0,21 0 19 /sys/fs/selinux/status
>>> systemd 1643 root 28r REG 0,21 0 19 /sys/fs/selinux/status
>>> (sd-pam) 1645 root mem REG 0,21 0 19 /sys/fs/selinux/status
>>> ...
>>> sshd 218874 root mem REG 0,21 0 19 /sys/fs/selinux/status
>>> sshd 218874 root 3r REG 0,21 0 19 /sys/fs/selinux/status
>>> sshd 218880 plautrba mem REG 0,21 0 19 /sys/fs/selinux/status
>>> sshd 218880 plautrba 3r REG 0,21 0 19 /sys/fs/selinux/status
>>>
>>> It seems to be caused by commit 05bdc03130d7 ("libselinux: use kernel
>>> status page by default") which replaced avc_netlink_open() in
>>> avc_init_internal() with selinux_status_open()
>>>
>>> In case of sshd process, /sys/fs/selinux/status seems to be mapped by
>>> selinux_check_access() which is called from pam_selinux and it's left
>>> open as there's no selinux_status_close() in selinux_check_access().
>>> The similar situations probably happen in systemd and dbus.
>>>
>>> So is it expected? Is it a bug? Do we need to change other components so
>>> that they would call selinux_status_close() when they use check access?
>>
>> What if we just close the fd after mmap and not keep it open? I don't
>> see any use of selinux_status_fd beyond assignment and closing.
>> Tearing down the mapping and re-creating it on every access check
>> would defeat the purpose.
>
> Thanks for this hint! I've checked mmap(2) and it's there: After the
> mmap() call has returned, the file descriptor, fd, can be closed
> immediately without invalidating the mapping.
>
> I'll try it tomorrow.
https://patchwork.kernel.org/project/selinux/patch/20210114133910.282686-1-plautrba@xxxxxxxxxx/
With this patch fd's are closed but the mapped memory is still there:
# umount /sys/fs/selinux
umount: /sys/fs/selinux: target is busy.
# lsof /sys/fs/selinux
COMMAND PID USER FD TYPE DEVICE SIZE/OFF NODE NAME
systemd 1 root mem REG 0,21 0 19 /sys/fs/selinux/status
systemd-u 363 root mem REG 0,21 0 19 /sys/fs/selinux/status
systemd-r 393 systemd-resolve mem REG 0,21 0 19 /sys/fs/selinux/status
dbus-brok 432 dbus mem REG 0,21 0 19 /sys/fs/selinux/status
...
For now we have a workaround for our tests - `umount -l /sys/fs/selinux` works.
But we should document the side effect of selinux status page change in
release notes if there's no better solution.
Petr
