Re: [PATCH 02/20] PCI: fix pci_remap_iospace() remap attribute

[Bjorn Helgaas]

[+cc Luis]

On Mon, Feb 27, 2017 at 03:14:13PM +0000, Lorenzo Pieralisi wrote:
> According to the PCI local bus specifications (Revision 3.0, 3.2.5),
> I/O Address space transactions are non-posted. On architectures where
> I/O space is implemented through a chunk of memory mapped space mapped
> to PCI address space (ie IA64/ARM/ARM64) the memory mapping for the
> region backing I/O Address Space transactions determines the I/O
> transactions attributes (before the transactions actually reaches the
> PCI bus where it is handled according to the PCI specifications).
> 
> Current pci_remap_iospace() interface, that is used to map the PCI I/O
> Address Space into virtual address space, use pgprot_device() as memory
> attribute for the virtual address mapping, that in some architectures
> (ie ARM64) provides non-cacheable but write bufferable mappings (ie
> posted writes), which clash with the non-posted write behaviour for I/O
> Address Space mandated by the PCI specifications.
> 
> Update the prot ioremap_page_range() parameter in pci_remap_iospace()
> to pgprot_noncached to ensure that the virtual mapping backing
> I/O Address Space guarantee non-posted write transactions issued
> when addressing I/O Address Space through the MMIO mapping.
> 
> Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi@xxxxxxx>
> Cc: Arnd Bergmann <arnd@xxxxxxxx>
> Cc: Will Deacon <will.deacon@xxxxxxx>
> Cc: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
> Cc: Russell King <linux@xxxxxxxxxxxxxxx>
> Cc: Catalin Marinas <catalin.marinas@xxxxxxx>
> ---
>  drivers/pci/pci.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> index bd98674..bfb3c6e 100644
> --- a/drivers/pci/pci.c
> +++ b/drivers/pci/pci.c
> @@ -3375,7 +3375,7 @@ int pci_remap_iospace(const struct resource *res, phys_addr_t phys_addr)
>  		return -EINVAL;
>  
>  	return ioremap_page_range(vaddr, vaddr + resource_size(res), phys_addr,
> -				  pgprot_device(PAGE_KERNEL));
> +				  pgprot_noncached(PAGE_KERNEL));

pgprot_device() is equivalent to pgprot_noncached() on all arches
except ARM64, and I trust you're doing the right thing on ARM64, so
I'm fine with this from a PCI perspective.

I do find this puzzling because I naively expected pgprot_noncached()
to match up with ioremap_nocache(), and apparently it doesn't.

For example, ARM64 ioremap_nocache() uses PROT_DEVICE_nGnRE, which
doesn't match the MT_DEVICE_nGnRnE in pgprot_noncached().

The point of these patches is to use non-posted mappings.  Apparently
you can do that with pgprot_noncached() here, but ioremap_nocache()
isn't enough for the config space mappings?

I suppose that's a consequence of the pgprot_noncached() vs
ioremap_nocache() mismatch, but this is all extremely confusing.

>  #else
>  	/* this architecture does not have memory mapped I/O space,
>  	   so this function should never be called */
> -- 
> 2.10.0
> 
> 
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel