Inconsistent capability requirements for prctl_set_mm_exe_file()

[Cyrill Gorcunov]

On Tue, Oct 27, 2020 at 08:22:11PM +0300, Kirill Tkhai wrote:
> 1)Before my commit there also were different checks
> 
>         !capable(CAP_SYS_RESOURCE))
> and
> 	uid_eq(cred->uid, make_kuid(ns, 0)) && gid_eq(cred->gid, make_kgid(ns, 0))
> 
> so it is not the initial reason. The commit even decreased the checks difference
> and it made both the checks are about capability().
> 
> 2)As I understand new PR_SET_MM_MAP interface differs in the way, that it allows to batch
> a setup of prctl_mm_map parameters. So, instead of 14 prlctl calls with different arguments:
> PR_SET_MM_START_CODE, PR_SET_MM_END_CODE, PR_SET_MM_START_DATA, .., PR_SET_MM_ENV_END,
> we set then all at once and the performance is better.
> 
> The second advantage is that the new interface is more comfortable in case of we set all
> of 14 parameters. Old interface requires special order of calls: sometimes you have to
> set PR_SET_MM_START_CODE first and then PR_SET_MM_END_CODE second, some times it is vice
> versa. Otherwise __prctl_check_order() in validate_prctl_map() will fail.

Since I've been the person who introduced the former PR_SET_MM_X interface I already
explained that the PR_SET_MM_X is simply shitty and better be vanished and forgotten.
Which we simply can't do :/ In turn PR_SET_MM_MAP is the only right way to verify
all fields in a one pass not only because of speed but otherwise the validation
of parameters is not even associative and in result (as you mentioned) _order_ of
calls does matter for start_code/end_code.

> 3)For me it looks like any combinations of parameters acceptable to be set by both interfaces
> are the same (in case of we don't see on permissions checks). In case of we can assign a set of
> parameters {A} using old interface, we can assign it from new interface and vice versa.
> Isn't this so?! If so, we should use the same permissions check.

Yup, if only I'm not missing something obvious we could drop cap(sys-resource) here
because internally after this check we jump into traditional verification procedure
where the map is filled from runtime data.

I'm quite sceptic that LSM hook Jann mentioned gonna be better. It might be an
addition but not the replacement. Moreover if we add it here someone from CRIU
camp have to verify that it won't break current CRIU tests.