Hi, Mon, 2021-05-10 at 17:36 +0200, Daniel Vetter wrote: > ... > > If DRM allows user-space to exhaust all of system memory, this seems > > to be a gap in enforcement of MEMCG limits for system memory. > > I tried to look into it when this was discussed in the past.... > > My guess is that shmem_read_mapping_page_gfp() -> > > shmem_getpage_gfp() is not choosing the correct MM to charge against > > in the use case of drivers using shmemfs for backing gem buffers. > > Yeah we know about this one since forever. The bug report is roughly > as old as the gem/ttm memory managers :-/ So another problem might be > that if we now suddenly include gpu memory in the memcg accounting, we > start breaking a bunch of workloads that worked just fine beforehand. It's not the first time tightening security requires adapting settings for running workloads... Workload GPU memory usage needs to be significant portion of application's real memory usage, to cause workload to hit limits that have been set for it earlier. Therefore I think it to definitely be something that user setting such limits actually cares about. => I think the important thing is that reason for the failures is clear from the OOM message. This works much better if GPU related memory usage is specifically stated in that message, once that memory starts to be accounted for system memory. - Eero
