Re: [PATCH] mm/slub: fix a deadlock in shuffle_freelist()

[Sebastian Andrzej Siewior]

On 2019-09-16 10:01:27 [-0400], Qian Cai wrote:
> On Mon, 2019-09-16 at 11:03 +0200, Sebastian Andrzej Siewior wrote:
> > On 2019-09-13 12:27:44 [-0400], Qian Cai wrote:
> > â
> > > Chain exists of:
> > >   random_write_wait.lock --> &rq->lock --> batched_entropy_u32.lock
> > > 
> > >  Possible unsafe locking scenario:
> > > 
> > >        CPU0                    CPU1
> > >        ----                    ----
> > >   lock(batched_entropy_u32.lock);
> > >                                lock(&rq->lock);
> > >                                lock(batched_entropy_u32.lock);
> > >   lock(random_write_wait.lock);
> > 
> > would this deadlock still occur if lockdep knew that
> > batched_entropy_u32.lock on CPU0 could be acquired at the same time
> > as CPU1 acquired its batched_entropy_u32.lock?
> 
> I suppose that might fix it too if it can teach the lockdep the trick, but it
> would be better if there is a patch if you have something in mind that could be
> tested to make sure.

get_random_bytes() is heavier than get_random_int() so I would prefer to
avoid its usage to fix what looks like a false positive report from
lockdep.
But no, I don't have a patch sitting around. A lock in per-CPU memory
could lead to the scenario mentioned above if the lock could be obtained
cross-CPU it just isn't so in that case. So I don't think it is that
simple.

Sebastian