On Tue, Aug 1, 2017, at 09:18, Eric Dumazet wrote: > On Tue, 2017-08-01 at 02:17 -0400, Hannes Frederic Sowa wrote: > > > We do account rmem as well as wmem allocated memory to the apropriate > > mem_cgs. In theory this should be okay. > > Last time I checked, rmem was not memcg ready yet. > > Can you describe the details ?
As long as our packets pass __sock_queue_rcv_skb (what we do with udp, raw, packet) we call down to sk_rmem_schedule, which in __sk_mem_schedule -> sk_mem_raise_allocated will charge the mem_cg or suppress the allocation. For tcp, as I remember from the last discussion, we simply drop packets and don't handle that very sensible, albeit we should not get out of the mem_cg limits. tcp_try_rmem_schedule in front of the ofo as well as the data queue should make sure of that. Did I overlook anything? OTOH, I am not too fond of the change. I just think that it shouldn't deplete memory but rather stall connections. Thanks, Hannes
