[PATCH v2] arm64: Introduce IRQ stack

[Catalin Marinas]

On Fri, Sep 18, 2015 at 04:03:02PM +0100, Catalin Marinas wrote:
> On Fri, Sep 18, 2015 at 09:57:56PM +0900, Jungseok Lee wrote:
> > On Sep 18, 2015, at 1:21 AM, Catalin Marinas wrote:
> > > So, without any better suggestion for current_thread_info(), I'm giving
> > > up the idea of using SPSel == 0 in the kernel. I'll look at your patch
> > > in more detail. BTW, I don't think we need the any count for the irq
> > > stack as we don't re-enter the same IRQ stack.
> > 
> > Another interrupt could come in since IRQ is enabled when handling softirq
> > according to the following information which are self-evident.
> > 
> > (Am I missing something?)
> 
> No. I had the wrong impression that we switch to the softirqd stack for
> softirqs but you are right, if we run them on the same stack the IRQs
> are enabled and they can be re-entered before we return from this
> exception handler.
> 
> I've seen other architectures implementing a dedicated softirq stack but
> for now we should just re-use the current IRQ stack.
> 
> > In my first approach using SPSel = 0, current_thread_info function was inefficient
> > in order to handle this case correctly.
> 
> I agree. And we don't have any other scratch register left as we use
> tpidr_el1 for per-cpu areas.
> 
> > BTW, in this context, it is only meaningful to decide whether a current interrupt
> > is re-enterrant or not. Its actual value is not important, but I could not figure
> > out a better implementation than this one yet. Any suggestions are welcome!
> 
> James' idea of checking the lower SP bits instead of a count may work.

Another thought (it seems that x86 does something similar): we know the
IRQ stack is not re-entered until interrupts are enabled in
__do_softirq. If we enable __ARCH_HAS_DO_SOFTIRQ, we can implement an
arm64-specific do_softirq_own_stack() which increments a counter before
calling __do_softirq. The difference from your patch is that
irq_stack_entry only reads such counter, doesn't need to write it.

Yet another idea is to reserve some space in the lower address part of
the stack with a "stack type" information. It still requires another
read, so I think the x86 approach is probably better.

-- 
Catalin