Re: Permissions in the udevadm_t domain

On 7/09/19 15:14, Nicolas Iooss wrote:
> [Adding [email protected], which is better suited for
> questions related to refpolicy]
>
> On Sat, Sep 7, 2019 at 12:19 PM Laurent Bigonville <[email protected]> wrote:
> [...]
>> But when
>> looking at the current policy code I'm seeing that udevadm is allowed to
>> delete files/directory/.. in /var/run and I'm wondering why. I've never
>> seen this happening during my (limited) test, an idea?
> Which policy? On my test system based on a lightly-patched refpolicy, I have:
>
> # sesearch -A -s udevadm_t -t var_run_t
> allow udevadm_t var_run_t:dir { getattr open search };
> allow udevadm_t var_run_t:lnk_file { getattr read };

It's the refpolicy, but I meant udev_var_run_t (/var/run/udev) rather 
than var_run_t

# sesearch -A -s udevadm_t -t udev_var_run_t
allow udevadm_t udev_var_run_t:dir { getattr ioctl lock open read 
remove_name rmdir search write };
allow udevadm_t udev_var_run_t:file { getattr unlink };
allow udevadm_t udev_var_run_t:lnk_file { getattr unlink };
allow udevadm_t udev_var_run_t:sock_file { append getattr open write };

>
>> For the later, it seems that the kernel the mode to 400 on some files in
>> /sys (ie. --w------- 1 root root 4096 sep  5 17:06
>> /sys/module/snd_hda_codec_generic/uevent) looking at the code it seems
>> that udev is ready to handle EACCES already, so I was wondering, should
>> we just allow dac_read_search or don't audit dac_read_search (and
>> dac_override as well then)?
> For the record, even with dac_read_search, opening this file fails
> with EACCES, because the kernel did not implement show() on this sysfs
> file (cf. https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/kernel/module.c?h=v5.2#n1204
> and kernfs_fop_open() in fs/kernfs/file.c). More precisely, openat()
> returns EACCES because kernfs_fop_open() returns this error code when
> trying to open a write-only file for reading. Unfortunately this check
> happens after the capability checks, which is why you got the denials
> for dac_override and dac_read_search.
>
> I do not have an opinion about allowing dac_read_search vs.
> dontaudit-ing dac_read_search and dac_override.
Grift was saying that "the cap_dac_read_search could maybe be 
dontaudited, but then cap_dac_override would have to be dontaudited as 
well. cap_dac_read_search would also be triggered when you run `sudo 
udevadm ...` where pwd or/and oldpwd is ~"

So I guess I will just allow it