Hi Michal,

On 10/20/21 1:10 AM, Michal Koutný wrote:

Hi.

On Tue, Oct 19, 2021 at 06:41:14PM +0800, Quanyang Wang [off-list ref] wrote:

quoted

So I add 2 "Fixes tags" here to indicate that 2 commits introduce two
different issues.

AFAIU, both the changes are needed to cause the leak, a single patch
alone won't cause the issue. Is that correct? (Perhaps not as I realize,
see below.)

Yes, I back to the earlier commit 4bfc0bb2c60e and no memory leak is 
observed.

But on second thought, the problem is the missing percpu_ref_exit() in
the (root) cgroup release path and percpu counter would allocate the
percpu_count_ptr anyway, so 4bfc0bb2c60e is only making the leak more
visible. Is this correct?

No, the earlier commit 4bfc0bb2c60e introduces a imbalance and the later
commit 2b0d3d3e4fcf introduces a visible leak.

Thanks,
Quanyang

I agree the commit 2b0d3d3e4fcf ("percpu_ref: reduce memory footprint of
percpu_ref in fast path") alone did nothing wrong.

[On a related (but independent) note, there seems to be an optimization
opportunity in not dealing with cgroup_bpf at all on the non-default
hierarchies.]

Regards,
Michal