On Wed 30-04-25 14:53:50, Roman Gushchin wrote: > On Wed, Apr 30, 2025 at 09:27:39AM +0200, Michal Hocko wrote: > > On Tue 29-04-25 21:31:35, Roman Gushchin wrote: > > > On Tue, Apr 29, 2025 at 01:46:07PM +0200, Michal Hocko wrote: > > > > On Mon 28-04-25 03:36:15, Roman Gushchin wrote: > > > > > Introduce bpf_out_of_memory() bpf kfunc, which allows to declare > > > > > an out of memory events and trigger the corresponding kernel OOM > > > > > handling mechanism. > > > > > > > > > > It takes a trusted memcg pointer (or NULL for system-wide OOMs) > > > > > as an argument, as well as the page order. > > > > > > > > > > Only one OOM can be declared and handled in the system at once, > > > > > so if the function is called in parallel to another OOM handling, > > > > > it bails out with -EBUSY. > > > > > > > > This makes sense for the global OOM handler because concurrent handlers > > > > are cooperative. But is this really correct for memcg ooms which could > > > > happen for different hierarchies? Currently we do block on oom_lock in > > > > that case to make sure one oom doesn't starve others. Do we want the > > > > same behavior for custom OOM handlers? > > > > > > It's a good point and I had similar thoughts when I was working on it. > > > But I think it's orthogonal to the customization of the oom handling. > > > Even for the existing oom killer it makes no sense to serialize memcg ooms > > > in independent memcg subtrees. But I'm worried about the dmesg reporting, > > > it can become really messy for 2+ concurrent OOMs. > > > > > > Also, some memory can be shared, so one OOM can eliminate a need for another > > > OOM, even if they look independent. > > > > > > So my conclusion here is to leave things as they are until we'll get signs > > > of real world problems with the (lack of) concurrency between ooms. > > > > How do we learn about that happening though? I do not think we have any > > counters to watch to suspect that some oom handlers cannot run. > > The bpf program which declares an OOM can handle this: e.g. retry, wait > and retry, etc. We can also try to mimick the existing behavior and wait > on oom_lock (potentially splitting it into multiple locks to support > concurrent ooms in various memcgs). Do you think it's preferable? Yes, I would just provide different callbacks for global and memcg ooms and do the blockin for the latter. It will be consistent with the in kernel implementation (therefore less surprising behavior). -- Michal Hocko SUSE Labs
