On Fri, 3 Jan 2014, Stefano Stabellini wrote:

On Mon, 30 Dec 2013, Mark Salter wrote:

quoted

On Mon, 2013-12-30 at 14:55 +0800, Chen Baozi wrote:

quoted

xen_remap used to be defined as ioremap_cached on arm64. In commit
c04e8e2fe, a new ioremap_cache was implemented, and ioremap_cached
was deleted, while xen_remap stays the same. This would lead to
the failure when building with CONFIG_HVC_XEN. Redefined xen_remap
on arm64 as ioremap_cache on arm64 to fix it.

I missed that include of arm header by arm64 when looking for users
of arm64's ioremap_cached() when working on commit c04e8e2fe. Anyway,
grepping the kernel tree, I see:

  ioremap_cached()
    defined by: arm, metag, unicore32
    used by: arch/arm/include/asm/xen/page.h
             drivers/mtd/maps/pxa2xx-flash.c

  ioremap_cache()
    defined by: arm64, sh, xtensa, ia64, x86
    used by: drivers/video/vesafb.c
             drivers/char/toshiba.c
             drivers/acpi/apei
             drivers/lguest/lguest_device.c
             drivers/sfi/sfi_core.c
             include/linux/acpi_io.h

I think it would be better to just avoid the confusion and the ifdef in
asm/xen/page.h by globally changing ioremap_cached to ioremap_cache.

While I welcome the suggestion, this is a critical fix for a regression
that I think should go in as soon as possible, maybe 3.13-rc7, while I
don't think that a global s/ioremap_cached/ioremap_cache would be
acceptable at this stage.

BTW Rob Herring sent a patch in November to solve the problem by
renaming ioremap_cached to ioremap_cache under arch/arm:

http://marc.info/?l=linux-arm-kernel&m=138394601804783

I wrongly assumed that was going to go in.
I think it is too late for that now. We'll have to go this this patch
for 3.13.