Re: do_softirq() vs __do_softirq() in irq_exit() and stack overflow

[Frederic Weisbecker <fweisbec@gmail.com>]

On Thu, Sep 05, 2013 at 07:39:56AM +1000, Benjamin Herrenschmidt wrote:
> Hi Folks !
> 
> It appears that the current version of irq_exit() calls __do_softirq()
> directly rather than do_softirq().
> 
> That means we are going to call the softirq's in the current interrupt
> frame rather than on the separate softirq stack.
> 
> The current frame is also still the normal kernel stack, because
> do_IRQ() itself only switches to the interrupt stack for processing
> the handlers (it's back to the original stack by the time it calls
> irq_exit).
> 
> That means that we end up stacking the normal stack, the actually HW
> interrupt stack frame (which can be pretty big on ppc) + do_IRQ's own,
> then the softirq (networks stack can create HUGE stack frames) and ...
> we are in softirq, so HW irqs are enable, we can thus can another irq
> stack frame piled up on top of that (or a perf stack).
> 
> We are observing actual overflows, here's an example blowing up our 16k
> stack on ppc64, you notice that it's all on the normal kernel stack:

I see, __do_softirq() is sometimes called to avoid irqsafe and softirq_pending
check they are not necessary but OTOH this bypass the arch overriden handler.

I'm going to try something and post soon.

Thanks.