On Fri, May 15, 2020 at 10:49:22AM -0700, Shakeel Butt wrote: > On Fri, May 15, 2020 at 8:00 AM Roman Gushchin <g...@fb.com> wrote: > > On Fri, May 15, 2020 at 06:44:44AM -0700, Shakeel Butt wrote: > > > On Fri, May 15, 2020 at 6:24 AM Johannes Weiner <han...@cmpxchg.org> > > > wrote: > > > > You're right. It should only bypass the page_counter, but still set > > > > page->mem_cgroup = root_mem_cgroup, just like user pages. > > > > What about kernel threads? We consider them belonging to the root memory > > cgroup. Should their memory consumption being considered in root-level > > stats? > > > > I'm not sure we really want it, but I guess we need to document how > > kernel threads are handled. > > What will be the cons of updating root-level stats for kthreads?
Should kernel threads be doing GFP_ACCOUNT allocations without memalloc_use_memcg()? GFP_ACCOUNT implies that the memory consumption can be significant and should be attributed to userspace activity. If the kernel thread has no userspace entity to blame, it seems to imply the same thing as a !GFP_ACCOUNT allocation: shared public infrastructure, not interesting to account to any specific cgroup. I'm not sure if we have such allocations right now. But IMO we should not account anything from kthreads, or interrupts for that matter, /unless/ there is a specific active_memcg that was set by the kthread or the interrupt.