Re: RCU stall in af_unix.c, should use spin_lock_irqsave?

[Thomas Petazzoni <thomas.petazzoni@free-electrons.com>]

Dear Eric Dumazet,

On Tue, 21 Oct 2014 03:04:34 -0700, Eric Dumazet wrote:

> > So, the question is: is this patch the correct solution (but then other
> > usage of spin_lock in af_unix.c might also need fixing) ? Or is the
> > network driver at fault?
> > 
> > Thanks for your input,
> > 
> > Thomas
> 
> Locks in af_unix do not need to mask irqs. Ever.
> 
> skb_queue_tail() uses an irqsave variant because its a generic function,
> and _some_ skb list might be manipulated from hard irq handlers in pre
> NAPI drivers. But af_unix does not have an interrupt handler that could
> potentially try to lock sk_receive_queue.lock

Ok. So it's actually safe to mix spin_lock() and spin_lock_irqsave() on
the same lock, if you know that this lock will never ever be taken in
an interrupt context?

> mvpp2 is seriously brain damaged : on_each_cpu() cannot be used from
> a bottom half handler.

That's what I thought. Back to the drawing board then, to fix mvpp2.

Do you think there is a place where we can write down those
assumptions? It isn't easy to spot whether on_each_cpu() is safe to use
in a bottom half or not.

Anyway, thanks a lot for your quick feedback!

Best regards,

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com