On Sat, Oct 28, 2017 at 05:19:36PM +0300, Amir Goldstein wrote: > <suggest rephrase> > Due to the current design of kmemcg, the memcg of the process who does the > allocation gets the accounting, so event allocations get accounted for > the memcg of > the event producer process, even though the misbehaving process is the listener. > The event allocations won't be freed if the producer exits, only if > the listener exists. > Nevertheless, it is still better to account event allocations to memcg > of producer > process and not to root memcg, because heuristically producer is many > time in the > same memcg as the listener. For example, this is the case with listeners inside > containers that listen on events for files or mounts that are private > to the container. > <\suggest rephrase> Well, if we're nitpicking ... Due to the current design of kmemcg, the event allocation is accounted to the memcg of the process producing the event, even though the misbehaving process is the listener. The event allocations won't be freed if the producer exits, only if the listener exits. Nevertheless, it is still better to account event allocations to the producer's memcg than the root memcg, because the producer is frequently in the same memcg as the listener. For example, this is the case with listeners inside containers that listen to events for files or mounts that are private to the container.
