Re: [patch] mm, oom: prevent soft lockup on memcg oom for UP systems

[Michal Hocko]

On Thu 12-03-20 21:16:27, Michal Hocko wrote:
> On Thu 12-03-20 11:20:33, David Rientjes wrote:
> > On Thu, 12 Mar 2020, Michal Hocko wrote:
> > 
> > > > I think the changelog clearly states that we need to guarantee that a 
> > > > reclaimer will yield the processor back to allow a victim to exit.  This 
> > > > is where we make the guarantee.  If it helps for the specific reason it 
> > > > triggered in my testing, we could add:
> > > > 
> > > > "For example, mem_cgroup_protected() can prohibit reclaim and thus any 
> > > > yielding in page reclaim would not address the issue."
> > > 
> > > I would suggest something like the following:
> > > "
> > > The reclaim path (including the OOM) relies on explicit scheduling
> > > points to hand over execution to tasks which could help with the reclaim
> > > process.
> > 
> > Are there other examples where yielding in the reclaim path would "help 
> > with the reclaim process" other than oom victims?  This sentence seems 
> > vague.
> 
> In the context of UP and !PREEMPT this also includes IO flushers,
> filesystems rely on workers and there are things I am very likely not
> aware of. If you think this is vaague then feel free to reformulate.
> All I really do care about is what the next paragraph is explaining.

Btw. do you plan to send a patch with an updated changelog?

> > > Currently it is mostly shrink_page_list which yields CPU for
> > > each reclaimed page. This might be insuficient though in some
> > > configurations. E.g. when a memcg OOM path is triggered in a hierarchy
> > > which doesn't have any reclaimable memory because of memory reclaim
> > > protection (MEMCG_PROT_MIN) then there is possible to trigger a soft
> > > lockup during an out of memory situation on non preemptible kernels
> > > <PUT YOUR SOFT LOCKUP SPLAT HERE>
> > > 
> > > Fix this by adding a cond_resched up in the reclaim path and make sure
> > > there is a yield point regardless of reclaimability of the target
> > > hierarchy.
> > > "
> > > 
> 
> -- 
> Michal Hocko
> SUSE Labs

-- 
Michal Hocko
SUSE Labs