On Mon, Oct 02, 2023 at 05:08:34PM +0200, Michal Hocko wrote:
> On Mon 02-10-23 10:50:26, Johannes Weiner wrote:
> > On Mon, Oct 02, 2023 at 03:43:19PM +0200, Michal Hocko wrote:
> > > On Wed 27-09-23 17:57:22, Nhat Pham wrote:
> [...]
> > > - memcg limit reclaim doesn't assist hugetlb pages allocation when
> > > hugetlb overcommit is configured (i.e. pages are not consumed from the
> > > pool) which means that the page allocation might disrupt workloads
> > > from other memcgs.
> > > - failure to charge a hugetlb page results in SIGBUS rather
> > > than memcg oom killer. That could be the case even if the
> > > hugetlb pool still has pages available and there is
> > > reclaimable memory in the memcg.
> >
> > Are these actually true? AFAICS, regardless of whether the page comes
> > from the pool or the buddy allocator, the memcg code will go through
> > the regular charge path, attempt reclaim, and OOM if that fails.
>
> OK, I should have been more explicit. Let me expand. Charges are
> accounted only _after_ the actual allocation is done. So the actual
> allocation is not constrained by the memcg context. It might reclaim
> from the memcg at that time but the disruption could have already
> happened. Not really any different from regular memory allocation
> attempt but much more visible with GB pages and one could reasonably
> expect that memcg should stop such a GB allocation if the local reclaim
> would be hopeless to free up enough from its own consumption.
>
> Makes more sense?
Yes, that makes sense.
This should be fairly easy to address by having hugetlb do the split
transaction that charge_memcg() does in one go, similar to what we do
for the hugetlb controller as well. IOW,
alloc_hugetlb_folio()
{
if (mem_cgroup_hugetlb_try_charge())
return ERR_PTR(-ENOMEM);
folio = dequeue();
if (!folio) {
folio = alloc_buddy();
if (!folio)
goto uncharge;
}
mem_cgroup_hugetlb_commit_charge();
}
