Re: [PATCH] powerpc/64s: relocation, register save fixes for system reset interrupt

[Nicholas Piggin <npiggin@gmail.com>]

On Thu, 3 Nov 2016 01:56:46 -0400
"Shreyas B. Prabhu" <shreyasbp@gmail.com> wrote:

> On Thu, Nov 3, 2016 at 1:21 AM, Nicholas Piggin <npiggin@gmail.com> wrote:
> > On Wed, 2 Nov 2016 14:15:48 +0530
> > Gautham R Shenoy <ego@linux.vnet.ibm.com> wrote:
> >  
> >> Hi Nick,
> >>
> >> On Wed, Nov 02, 2016 at 07:36:24PM +1100, Nicholas Piggin wrote:  
> >> >
> >> > Okay, I'll work with that. What's the best way to make a P8 do
> >> > winkle sleeps?  
> >>
> >> From the userspace, offlining the CPUs of the core will put them to
> >> winkle.  
> >
> > Thanks for this. Hum, that r13 manipulation throughout the idle
> > and exception code is a bit interesting. I'll do the minimal patch
> > for 4.9, but what's the reason not to just use the winkle state
> > in the PACA rather than storing it into HSPRG0 bit, can you (or
> > Shreyas) explain?
> >  
> Hi Nick,
> 
> Before deep winkle, checking SRR1's wakeup bits (Bits 46:47) was enough to
> figure out which idle state we are waking up from. But in P8, SRR1's wakeup
> bits aren't enough since bits 46:47 are 0b11 for both fast sleep and
> deep winkle.
> So to distinguish bw fastsleep and deep winkle, we use the current HSPRG0/PORE
> trick. We program the PORE engine (which is used for state restore when waking
> up from deep winkle) to restore HSPRG0 with the last bit set (we do this in
> pnv_save_sprs_for_winkle()). R13 bit manipulation in pnv_restore_hyp_resource
> is related to this.

Right, I didn't realize how that exactly worked until I had to go read
the code just now. It's a neat little trick. I'm wondering can we use PACA_THREAD_IDLE_STATE==PNV_THREAD_WINKLE for this instead? It would just
make the early PACA usage in the exception handlers able to use more common
code.

Thanks,
Nick