Re: [PATCH] Introduce CONFIG_VIRT_TO_BUS

[Arnd Bergmann <arnd@arndb.de>]

On Wednesday 23 May 2007, Al Viro wrote:
> 
> On Wed, May 23, 2007 at 05:01:55PM +1000, Stephen Rothwell wrote:
> > Make some offending drivers depend on it and set
> > CONFIG_ARCH_NO_VIRT_TO_BUS for ppc64 so that we don't build those drivers.
> 
> Seeing how few targets _do_ have that, I'd invert the that.  I.e. have
> HAS_VIRT_TO_BUS in arch/*/Kconfig for the few architectures that have
> that stuff.

Currently, that would mark x86_64 as HAS_VIRT_TO_BUS, while 64 bit powerpc
would be NO_VIRT_TO_BUS. I think they really should be the same, probably
like the other 64 bit targets, which means:

* If CONFIG_IOMMU is disabled, provide a __deprecated inline function that
maps virt_to_bus to virt_to_phys, and allow building drivers that use it.

* If we build with CONFIG_IOMMU enabled, but don't support disabling the
iommu, have no declaration for virt_to_bus at all, and disable the users
in Kconfig

* If the usage of the iommu is determined at run time, warn about users
of virt_to_bus at run time if it is enabled.

It could look like this

/* mm/memory.c or somewhere */
#if defined(CONFIG_HAS_VIRT_TO_BUS) && defined(CONFIG_IOMMU)
unsigned long virt_to_bus(void *virt)
{
	WARN_ON(iommu_enabled);
	return __virt_to_bus(virt);
}
#endif

/* include/linux/io.h */

#ifndef __virt_to_bus
#define __virt_to_bus virt_to_phys /* some architectures need 
					their own __virt_to_bus */
#endif

#ifdef CONFIG_HAS_VIRT_TO_BUS
#ifdef CONFIG_IOMMU
extern __deprecated unsigned long virt_to_bus(void *);
#else
static inline __deprecated unsigned long virt_to_bus(void *virt)
{
	return __virt_to_bus(virt);
}
#endif
#endif

	Arnd <><