On 2021-11-05 18:18:16 [+0100], To Luis Goncalves wrote:
> On 2021-11-04 15:48:46 [-0300], Luis Goncalves wrote:
> > Yes, sure!
> >
> > Just as a reminder, I was able to reproduce that backtrace in v5.14-rt
> > up to v5.14-rc6-rt11. But starting from v5.14-rc7-rt12 I no longer
> > reproduced the problem.
>
> So I'm about to this but patch but this mail confused me. Can you bisect
> which change is responsible for that? Based on the arguing, the same
> callchain should exist in v5.15, too. And a diff
> v5.14-rc6-rt11..v5.14-rc7-rt12 didn't show any change in that area.
Just to be clear:
- raw_seq_start() does read_lock(&h->lock) which is fine on its own.
- raw_seq_show() does local_bh_disable() via sock_i_ino() which nests
into the former lock. So the problem as lockdep says is:
Thread1 Thread2
raw_seq_start()
-> read_lock(&h->lock)
raw_hash_sk()
-> write_lock_bh(&h->lock)
-> local_bh_disable()
-> write_lock(&h->lock) // blocks on T1
raw_seq_show()
-> sock_i_ino()
-> local_bh_disable() // blocks on T2
and is valid on the same CPU.
Sebastian
