On Wed, Jul 6, 2022 at 3:54 PM Roman Gushchin <roman.gushchin@xxxxxxxxx> wrote: > > On Wed, Jul 06, 2022 at 03:11:46PM -0700, Alexei Starovoitov wrote: > > On Wed, Jul 6, 2022 at 12:09 PM Roman Gushchin <roman.gushchin@xxxxxxxxx> wrote: > > > > > > On Wed, Jul 06, 2022 at 09:47:32AM -0700, Alexei Starovoitov wrote: > > > > On Wed, Jul 6, 2022 at 8:59 AM Yafang Shao <laoar.shao@xxxxxxxxx> wrote: > > > > > > > > > > GFP_ATOMIC doesn't cooperate well with memcg pressure so far, especially > > > > > if we allocate too much GFP_ATOMIC memory. For example, when we set the > > > > > memcg limit to limit a non-preallocated bpf memory, the GFP_ATOMIC can > > > > > easily break the memcg limit by force charge. So it is very dangerous to > > > > > use GFP_ATOMIC in non-preallocated case. One way to make it safe is to > > > > > remove __GFP_HIGH from GFP_ATOMIC, IOW, use (__GFP_ATOMIC | > > > > > __GFP_KSWAPD_RECLAIM) instead, then it will be limited if we allocate > > > > > too much memory. > > > > > > > > > > We introduced BPF_F_NO_PREALLOC is because full map pre-allocation is > > > > > too memory expensive for some cases. That means removing __GFP_HIGH > > > > > doesn't break the rule of BPF_F_NO_PREALLOC, but has the same goal with > > > > > it-avoiding issues caused by too much memory. So let's remove it. > > > > > > > > > > The force charge of GFP_ATOMIC was introduced in > > > > > commit 869712fd3de5 ("mm: memcontrol: fix network errors from failing > > > > > __GFP_ATOMIC charges") by checking __GFP_ATOMIC, then got improved in > > > > > commit 1461e8c2b6af ("memcg: unify force charging conditions") by > > > > > checking __GFP_HIGH (that is no problem because both __GFP_HIGH and > > > > > __GFP_ATOMIC are set in GFP_AOMIC). So, if we want to fix it in memcg, > > > > > we have to carefully verify all the callsites. Now that we can fix it in > > > > > BPF, we'd better not modify the memcg code. > > > > > > > > > > This fix can also apply to other run-time allocations, for example, the > > > > > allocation in lpm trie, local storage and devmap. So let fix it > > > > > consistently over the bpf code > > > > > > > > > > __GFP_KSWAPD_RECLAIM doesn't cooperate well with memcg pressure neither > > > > > currently. But the memcg code can be improved to make > > > > > __GFP_KSWAPD_RECLAIM work well under memcg pressure if desired. > > > > > > > > Could you elaborate ? > > > > > > > > > It also fixes a typo in the comment. > > > > > > > > > > Signed-off-by: Yafang Shao <laoar.shao@xxxxxxxxx> > > > > > Reviewed-by: Roman Gushchin <roman.gushchin@xxxxxxxxx> > > > > > > > > Roman, do you agree with this change ? > > > > > > Yes, removing __GFP_HIGH makes sense to me. I can imagine we might want > > > it for *some* bpf allocations, but applying it unconditionally looks wrong. > > > > Yeah. It's a difficult trade-off to make without having the data > > to decide whether removing __GFP_HIGH can cause issues or not, > > Yeah, the change looks reasonable, but it's hard to say without giving > it a good testing in (something close to) a production environment. > > > but do you agree that __GFP_HIGH doesn't cooperate well with memcg ? > > If so it's a bug on memcg side, right? > > No. Historically we allowed high-prio allocations to exceed the memcg limit > because otherwise there were too many stability and performance issues. > It's not a memcg bug, it's a way to avoid exposing ENOMEM handling bugs all over > the kernel code. Without memory cgroups GFP_ATOMIC allocations rarely fail > and a lot of code paths in the kernel are not really ready for it (at least > it was the case several years ago, maybe things are better now). > > But it was usually thought in the context of small(ish) allocations which do not > change the global memory usage picture. Subsequent "normal" allocations are > triggering reclaim/OOM, so from a user's POV the limit works as expected. > > But with the ownership model and size of bpf maps it's a different story: > if a bpf map belongs to an abandoned cgroup, it might consume a lot of memory > and there will be no "normal" allocations. So cgroup memory limit will be never > applied. It's a valid issue, I agree with Yafang here. Understood. > > but we should probably > > apply this band-aid on bpf side to fix the bleeding. > > Later we can add a knob to allow __GFP_HIGH usage on demand from > > bpf prog. > > Yes, it sounds like a good idea. I have to think what's the best approach > here, it's not obvious for me. Ok. Applied this patch for now.