Hello, Shakeel!

On Thu, Feb 13, 2020 at 11:12:33PM -0800, Shakeel Butt wrote:

We are testing network memory accounting in our setup and noticed
inconsistent network memory usage and often unrelated memcgs network
usage correlates with testing workload. On further inspection, it seems
like mem_cgroup_sk_alloc() is broken in irq context specially for
cgroup v1.

A great catch!

mem_cgroup_sk_alloc() can be called in irq context and kind
of assumes that it can only happen from sk_clone_lock() and the source
sock object has already associated memcg. However in cgroup v1, where
network memory accounting is opt-in, the source sock can be not
associated with any memcg and the new cloned sock can get associated
with unrelated interrupted memcg.

Cgroup v2 can also suffer if the source sock object was created by
process in the root memcg or if sk_alloc() is called in irq context.

Do you mind sharing a call trace?

Also, shouldn't cgroup_sk_alloc() be changed in a similar way?

Thanks!

Roman