On Fri, Feb 5, 2021 at 6:21 PM Michal Hocko <mhocko@xxxxxxxx> wrote: > > On Fri 05-02-21 17:55:10, Muchun Song wrote: > > On Fri, Feb 5, 2021 at 4:24 PM Michal Hocko <mhocko@xxxxxxxx> wrote: > > > > > > On Fri 05-02-21 14:23:10, Muchun Song wrote: > > > > We call memcg_oom_recover() in the uncharge_batch() to wakeup OOM task > > > > when page uncharged, but for the slab pages, we do not do this when page > > > > uncharged. > > > > > > How does the patch deal with this? > > > > When we uncharge a slab page via __memcg_kmem_uncharge, > > actually, this path forgets to do this for us compared to > > uncharge_batch(). Right? > > Yes this was more more or less clear (still would have been nicer to be > explicit). But you still haven't replied to my question I believe. I > assume you rely on refill_stock doing draining but how does this address > the problem? Is it sufficient to do wakeups in the batched way? Sorry, the subject title may not be suitable. IIUC, memcg_oom_recover aims to wake up the OOM task when we uncharge the page. I see uncharge_batch always do this. I am confused why __memcg_kmem_uncharge does not. Both paths do the same thing (uncharge pages). So actually, this patch want to keep the two paths consistent. Thanks. > > > > > When we drain per cpu stock, we also should do this. > > > > > > Can we have anything the per-cpu stock while entering the OOM path. IIRC > > > we do drain all cpus before entering oom path. > > > > You are right. I did not notice this. Thank you. > > > > > > > > > The memcg_oom_recover() is small, so make it inline. > > > > > > Does this lead to any code generation improvements? I would expect > > > compiler to be clever enough to inline static functions if that pays > > > off. If yes make this a patch on its own. > > > > I have disassembled the code, I see memcg_oom_recover is not > > inline. Maybe because memcg_oom_recover has a lot of callers. > > Just guess. > > > > (gdb) disassemble uncharge_batch > > [...] > > 0xffffffff81341c73 <+227>: callq 0xffffffff8133c420 <page_counter_uncharge> > > 0xffffffff81341c78 <+232>: jmpq 0xffffffff81341bc0 <uncharge_batch+48> > > 0xffffffff81341c7d <+237>: callq 0xffffffff8133e2c0 <memcg_oom_recover> > > So does it really help to do the inlining? I just think memcg_oom_recover is very small, inline maybe a good choice. Maybe I am wrong. > -- > Michal Hocko > SUSE Labs
