On Mon, 2023-06-05 at 11:44 +0800, Abel Wu wrote:

On 6/4/23 6:36 PM, Willem de Bruijn wrote:

quoted

On Fri, Jun 2, 2023 at 10:42 PM Shakeel Butt [off-list ref] wrote:

quoted

On Fri, Jun 02, 2023 at 04:11:34PM +0800, Abel Wu wrote:

quoted

The sk_under_memory_pressure() is called to check whether there is
memory pressure related to this socket. But now it ignores the net-
memcg's pressure if the proto of the socket doesn't care about the
global pressure, which may put burden on its memcg compaction or
reclaim path (also remember that socket memory is un-reclaimable).

So always check the memcg's vm status to alleviate memstalls when
it's in pressure.

This is interesting. UDP is the only protocol which supports memory
accounting (i.e. udp_memory_allocated) but it does not define
memory_pressure. In addition, it does have sysctl_udp_mem. So
effectively UDP supports a hard limit and ignores memcg pressure at the
moment. This patch will change its behavior to consider memcg pressure
as well. I don't have any objection but let's get opinion of UDP
maintainer.

Thanks for the head-up, I did not notice the side effect on UDP.

quoted

So this commit only affects the only other protocol-independent
caller, __sk_mem_reduce_allocated, to possibly call
sk_leave_memory_pressure if now under the global limit.

What is the expected behavioral change in practice of this commit?

Be more conservative on sockmem alloc if under memcg pressure, to
avoid worse memstall/latency.

I guess the above is for TCP sockets only, right? Or at least not for
UDP sockets?

If so, I think we should avoid change of behaviour for UDP - e.g.
keeping the initial 'if (!sk->sk_prot->memory_pressure)' in
sk_under_memory_pressure(), with some comments about the rationale for
future memory. That should preserve the whole patchset effect for other
protocols, right?

If instead you are also interested into UDP sockets under pressure, how
that is going to work? UDP sockets can reclaim memory only at send and
close time. A memcg under pressure could starve some sockets forever if
the the ones keeping the memory busy are left untouched.

Cheers,

Paolo