On Wed, Nov 15, 2017 at 1:31 AM, Jan Kara <jack@xxxxxxx> wrote: > On Wed 15-11-17 01:32:16, Yang Shi wrote: >> >> >> On 11/14/17 1:39 AM, Michal Hocko wrote: >> >On Tue 14-11-17 03:10:22, Yang Shi wrote: >> >> >> >> >> >>On 11/9/17 5:54 AM, Michal Hocko wrote: >> >>>[Sorry for the late reply] >> >>> >> >>>On Tue 31-10-17 11:12:38, Jan Kara wrote: >> >>>>On Tue 31-10-17 00:39:58, Yang Shi wrote: >> >>>[...] >> >>>>>I do agree it is not fair and not neat to account to producer rather than >> >>>>>misbehaving consumer, but current memcg design looks not support such use >> >>>>>case. And, the other question is do we know who is the listener if it >> >>>>>doesn't read the events? >> >>>> >> >>>>So you never know who will read from the notification file descriptor but >> >>>>you can simply account that to the process that created the notification >> >>>>group and that is IMO the right process to account to. >> >>> >> >>>Yes, if the creator is de-facto owner which defines the lifetime of >> >>>those objects then this should be a target of the charge. >> >>> >> >>>>I agree that current SLAB memcg accounting does not allow to account to a >> >>>>different memcg than the one of the running process. However I *think* it >> >>>>should be possible to add such interface. Michal? >> >>> >> >>>We do have memcg_kmem_charge_memcg but that would require some plumbing >> >>>to hook it into the specific allocation path. I suspect it uses kmalloc, >> >>>right? >> >> >> >>Yes. >> >> >> >>I took a look at the implementation and the callsites of >> >>memcg_kmem_charge_memcg(). It looks it is called by: >> >> >> >>* charge kmem to memcg, but it is charged to the allocator's memcg >> >>* allocate new slab page, charge to memcg_params.memcg >> >> >> >>I think this is the plumbing you mentioned, right? >> > >> >Maybe I have misunderstood, but you are using slab allocator. So you >> >would need to force it to use a different charging context than current. >> >> Yes. >> >> >I haven't checked deeply but this doesn't look trivial to me. >> >> I agree. This is also what I explained to Jan and Amir in earlier >> discussion. > > And I also agree. But the fact that it is not trivial does not mean that it > should not be done... > I am currently working on directed or remote memcg charging for a different usecase and I think that would be helpful here as well. I have two questions though: 1) Is fsnotify_group the right structure to hold the reference to target mem_cgroup for charging? 2) Remote charging can trigger an OOM in the target memcg. In this usecase, I think, there should be security concerns if the events producer can trigger OOM in the memcg of the monitor. We can either change these allocations to use __GFP_NORETRY or some new gfp flag to not trigger oom-killer. So, is this valid concern or am I over-thinking? thanks, Shakeel
