[PATCH 5/6] ARM: re-implement physical address space switching

[Mark Rutland]

On Wed, May 06, 2015 at 12:33:13PM +0100, Russell King - ARM Linux wrote:
> On Wed, May 06, 2015 at 11:37:37AM +0100, Mark Rutland wrote:
> > Hi Russell,
> > 
> > > > It turns out that I was incorrect in my assertion, and the reordering I
> > > > suggested above can't happen. The ARMv7 ARM states:
> > > > 
> > > > 	Any direct write to a system control register is guaranteed not
> > > > 	to affect any instruction that appears, in program
> > > > 	order, before the instruction that performed the direct write
> > > > 
> > > > Which means that the STMFD cannot be affected by the later cp15 write to
> > > > the SCTLR, and so the DSB does not need to be moved before the MCR.
> > > > 
> > > > I apologise for adding to the confusion there.
> > > 
> > > So does this mean this patch gets an ack now?
> > 
> > I assumed there was going to be a respin for the CR_W change?
> > 
> > There's also the dodginess w.r.t. the page table walkers that I can't
> > see is solvable short of disabling the MMU prior to the flush, though I
> > understand you've NAKed that approach.
> 
> Why? 

I was on about the pre-assembly portion:

	cr = get_cr();
	set_cr(cr & ~(CR_I | CR_C | CR_W));
	flush_cache_all();

With the MMU on at this point the page table walkers can race with the
set/way maintenance. It also relies on the compiler not making stack
accesses between the SCTLR write and the completion of flush_cache_all,
which is likely but not guranteed.

So this won't necessarily flush out the data it seems to be intended to.

> Are you saying that after:
> 
> +       mrc     p15, 0, r8, c1, c0, 0           @ read control reg
> +       bic     ip, r8, #CR_M                   @ disable caches and MMU
> +       mcr     p15, 0, ip, c1, c0, 0
> +       dsb
> +       isb
> 
> the page table walkers are still actively walking the page table?
> 
> That to me sounds like a hardware bug.  The point of this is to shut
> down the MMU, _then_ update the page tables, and _then_ to re-enable
> the MMU to explicitly avoid problems with the page table walkers.

I agree that after this point it would be a bug for the page table
walkers to make cacheable accesses.

Thanks,
Mark.