Re: [PATCH V7 4/5] arm64: apei: implement arch_apei_get_mem_attributes()

[Will Deacon <will.deacon@arm.com>]

On Mon, Jul 27, 2015 at 10:38:11AM +0100, Matt Fleming wrote:
> On Fri, 2015-07-24 at 17:26 +0100, Will Deacon wrote:
> > On Fri, Jul 24, 2015 at 05:21:49PM +0100, Catalin Marinas wrote:
> > > On Fri, Jul 24, 2015 at 03:57:08PM +0100, Will Deacon wrote:
> > > > On Tue, Jul 21, 2015 at 10:59:19PM +0100, Jonathan (Zhixiong) Zhang wrote:
> > > > > diff --git a/arch/arm64/include/asm/acpi.h b/arch/arm64/include/asm/acpi.h
> > > > > +static inline pgprot_t arch_apei_get_mem_attribute(phys_addr_t addr)
> > > > > +{
> > > > > +	pgprot_t prot;
> > > > > +
> > > > > +	prot = efi_mem_attributes(addr);
> > > > > +	if (prot & EFI_MEMORY_UC)
> > > > > +		return PROT_DEVICE_nGnRnE;
> > > > > +	if (prot & EFI_MEMORY_WC)
> > > > > +		return PROT_NORMAL_NC;
> > > > 
> > > > Can we not use pgprot_noncached and pgprot_writecombine for these two?
> > > 
> > > Actually, why do we even use pgprot_t for prot here? EFI_MEMORY_* don't
> > > have anything to do with the arch-specific pgprot_t.
> > 
> > Good point; the pgprot_t confused me, so my suggestion is much use after
> > ll. We're better off with a u64 to avoid further confusion.
> 
> Isn't the whole point of arch_apei_get_mem_attribute() to turn an
> arch-independent memory attribute (EFI_MEMORY_*) into an arch-specific
> value to pass to ioremap_page_range()?

That bit's fine. The weird bit is:

  pgprot_t prot;

  prot = efi_mem_attributes(addr);

Since that's putting the arch-independent format into the pg_prot.

> I don't see how you can do that any other way than by using pgprot_t.
> 
> Really, the problem here is that ioremap_page_caller() has no notion of
> "map this range in a firmware-compatible manner". If we could do, for
> example,
> 
> 	ioremap_page_range(vaddr, vend, paddr, PAGE_FW_COMPAT);
> 
> that would allow the innards of the arch-ioremap to figure out exactly
> how to map this range so that the firmware could access it coherently.
> 
> I suggested this previously but it didn't gain any traction.

Yeah, or just ioremap_efi.

</me runs away>

Will