Re: [PATCH RT] rtmutex: Flush block plug on __down_read()

[Scott Wood <swood@redhat.com>]

On Mon, 2019-01-07 at 17:50 +0100, Sebastian Andrzej Siewior wrote:
> On 2019-01-04 15:33:21 [-0500], Scott Wood wrote:
> > __down_read() bypasses the rtmutex frontend to call
> > rt_mutex_slowlock_locked() directly, and thus it needs to call
> > blk_schedule_flush_flug() itself.
> 
> we don't do this in the spin_lock() case because !RT doesn't do it.

And because spin_lock() is called inside the flush path.

>  We
> do it for rtmutex because !RT does it for mutex.
> Now I can't remember why this was skipped for a rw_sem since it is
> performed for !RT as part of the schedule() invocation.

Without this we were seeing XFS hangs on our internal kernel.  I wasn't able
to reproduce it on a newer kernel, but it's very timing-dependant so I
wouldn't read too much into that.

> If I don't come up with a plausible explanation then I will apply this
> plus a hunk for the __down_write_common() case which should also be
> required (right?).

I don't think it's needed, as it doesn't call into the rtmutex code via a
backdoor.  When blocking on sem->rtmutex, rt_mutex_fastlock() will call the
flush.  When blocking with a direct call to schedule(), tsk_is_pi_blocked()
will not be true, and thus schedule() will do the flush via
sched_submit_work().

-Scott