Re: [PATCH] slab: distinguish lock and trylock for sheaf_flush_main()

[Hao Li <hao.li@linux.dev>]

On Wed, Mar 04, 2026 at 10:05:27AM +0900, Harry Yoo wrote:
> On Mon, Mar 02, 2026 at 10:56:44AM +0100, Vlastimil Babka (SUSE) wrote:
> > On 2/26/26 15:50, Vlastimil Babka (SUSE) wrote:
> > > On 2/11/26 10:42, Vlastimil Babka wrote:
> > >> sheaf_flush_main() can be called from __pcs_replace_full_main() where
> > >> the trylock can in theory fail, and pcs_flush_all() where it's not
> > >> expected to and it would be actually a problem if it failed and left the
> > >> main sheaf not flushed.
> > > 
> > > Thinking about this more, I now think it's not a theoretical issue because
> > > on PREEMPT_RT I think pcs_flush_all() can preempt someone holding the lock
> 
> Agreed!

Exactly, not just theoretical.

> 
> > > (on PREEMPT_RT it doesn't have to be an irq handler preempting a holder),
> > > and then fail to flush the main sheaf silently.
> > >
> > > The impact is probably limited though - if this failure to flush happens in
> > > __kmem_cache_shutdown(), it means someone was destroying a cache while using
> > > it, so that was already buggy.

Exactly.

> > > slab_mem_going_offline_callback() could be
> > > where this matters although it's unlikely someone would do memory hotplug
> > > together with PREEMPT_RT.

Yes, given the semantics of slab_mem_going_offline_callback(),
sheaf_flush_main() really should ensure that all objects are flushed out.

-- 
Thanks,
Hao

> > > 
> > > But maybe still worth tagging this as Fixes: 2d517aa09bbc ("slab: add opt-in
> > > caching layer of percpu sheaves") and Cc stable and sending it as a hotfix.
> > 
> > Added to slab/for-next-fixes with adjusted changelog:
> >
> > https://git.kernel.org/pub/scm/linux/kernel/git/vbabka/slab.git/commit/?h=slab*for-next-fixes&id=48647d3f9a644d1e81af6558102d43cdb260597b
> 
> -- 
> Cheers,
> Harry / Hyeonggon