Re: possible deadlock in sk_clone_lock

[Michal Hocko]

On Wed 03-03-21 09:59:45, Paul E. McKenney wrote:
> On Wed, Mar 03, 2021 at 09:03:27AM +0100, Michal Hocko wrote:
[...]
> > Paul what is the current plan with in_atomic to be usable for !PREEMPT
> > configurations?
> 
> Ah, thank you for the reminder!  I have rebased that series on top of
> v5.12-rc1 on -rcu branch tglx-pc.2021.03.03a.
> 
> The current state is that Linus is not convinced that this change is
> worthwhile given that only RCU and printk() want it.  (BPF would use
> it if it were available, but is willing to live without it, at least in
> the short term.)
> 
> But maybe Linus will be convinced given your additional use case.
> Here is hoping!

Yes, hugetlb freeing path would benefit from this. You can reference
this lockdep report (http://lkml.kernel.org/r/000000000000f1c03b05bc43aadc@xxxxxxxxxx)
with an additional argument that making hugetlb_lock irq safe is a
larger undertaking and we will need something reasonably backportable
for older kernels as well.
-- 
Michal Hocko
SUSE Labs