On Thu, Jul 10, 2014 at 02:14:16PM +0100, Arnd Bergmann wrote:

On Thursday 10 July 2014, Catalin Marinas wrote:

quoted

quoted

@@ -235,8 +155,7 @@ extern void __iomem *ioremap_cache(phys_addr_t phys_addr, size_t size);
 #define ioremap_wc(addr, size)               __ioremap((addr), (size), __pgprot(PROT_NORMAL_NC))
 #define iounmap                              __iounmap
 
-#define ARCH_HAS_IOREMAP_WC
-#include <asm-generic/iomap.h>
+#include <asm-generic/io.h>

We don't currently have PCIe support in mainline for arm64 but what I
had in mind with the generic iomap is that functions like ioread32_rep
first check whether the address is an IO address or a memory one and
calls the insl or mmio_insl accordingly. With your generic
implementation, this check disappears.

The question is whether this functionality would still be needed.

I think it's much better not to have that check on architectures that have a memory
mapped I/O space like arm64. The main advantage is that ioread32() is just a trivial
alias for readl(). The only reason for needing generic_iomap is architectures that
do something very different for I/O ports.

ioread32() is indeed an alias for readl on arm64 but if the address it
gets as argument looks like an I/O port, it adds PCI_IOBASE (via the inl
macro). This would no longer happen with Thierry's patches.

So I guess that for this patch to work, we also need ioport_map() to
return an address from the PCI_IOBASE range. Should we merge this now as
well:

http://linux-arm.org/git?p=linux-ld.git;a=commitdiff;h=3423064d8e42a38164a8436bcdf7434cf9cd2192

-- 
Catalin