On Thu, Sep 8, 2022 at 1:01 AM Tejun Heo [off-list ref] wrote:

On Wed, Sep 07, 2022 at 09:27:09AM -0700, Alexei Starovoitov wrote:

quoted

On Wed, Sep 7, 2022 at 9:18 AM Tejun Heo [off-list ref] wrote:

quoted

Hello,

On Wed, Sep 07, 2022 at 09:13:09AM -0700, Alexei Starovoitov wrote:

quoted

Hmm. We discussed this option already. We definitely don't want
to introduce an uapi knob that will allow anyone to skip memcg
accounting today and in the future.

cgroup.memory boot parameter is how memcg provides last-resort workarounds
for this sort of problems / regressions while they're being addressed. It's
not a dynamically changeable or programmable thing. Just a boot time
opt-out. That said, if you don't want it, you don't want it.

ahh. boot param.
Are you suggesting a global off switch ? Like nosocket and nokmem.
That would be a different story.
Need to think more about it. It could be ok.

Yeah, nobpf or sth like that. An equivalent cgroup.memory parameter.

It may be a useful feature for some cases, but it can't help container users.
The memcg works well to limit the non-pinned bpf-map, that's the
reason why we, a container user, switch to memcg-based bpf charging.
Our goal is to make it also work for pinned bpf-map.

That said,  your proposal may be a useful feature,  but it should be
another different patchset.

-- 
Regards
Yafang