Re: [RFC PATCH bpf-next 0/3] bpf: freeze a task cgroup from bpf

[Djalal Harouni <tixxdz@gmail.com>]

Hello Michal,

On 4/2/24 18:16, Michal Koutný wrote:
> Hello.
> 
> On Wed, Mar 27, 2024 at 11:53:22PM +0100, Djalal Harouni <tixxdz@gmail.com> wrote:
>> ...
>> For some cases we want to freeze the cgroup of a task based on some
>> signals, doing so from bpf is better than user space which could be
>> too late.
> 
> Notice that freezer itself is not immediate -- tasks are frozen as if a
> signal (kill(2)) was delivered to them (i.e. returning to userspace).

Thanks yes, I would expect freeze to behave like signal, and if one
wants to block immediately there is the LSM override return. The
selftest attached tries to do exactly that.

> What kind of signals (also kill?) are you talking about for
> illustration?

Could be security signals, reading sensitive files or related to any
operation management, for X reasons this user session should be freezed
or killed.

The kill is an effective defense against fork-bombs as an example.

>> Planned users of this feature are: tetragon and systemd when freezing
>> a cgroup hierarchy that could be a K8s pod, container, system service
>> or a user session.
> 
> It sounds like the signals are related to a particular process. If so
> what is it good for to freeze unrelated processes in the same cgroup?

Today some container/pod operations are performed at bpf level, having
the freeze and kill available is straightforward to perform this.


> I think those answers better clarify why this is needed.

Alright will add those in v2.

> 
> As for the generalization to any cgroup attribute (or kernfs). Can this
> be compared with sysctls -- I see there are helpers to intercept user
> writes but no helpers to affect sysctl values without an outer writer.
> What would justify different approaches between kernfs attributes and
> sysctls (direct writes vs modified writes)?

For generalizing this, haven't thought about it that much. First use
case is to try to get freeze and possibly kill support, and use a common
interface as requested.

Thank you!

> 
> Thanks,
> Michal