On Fri, Sep 06, 2013 at 03:09:21PM +0100, Catalin Marinas wrote:

On Fri, Sep 06, 2013 at 09:58:59AM +0100, Ian Campbell wrote:

quoted

On Thu, 2013-09-05 at 17:59 +0100, Stefano Stabellini wrote:

quoted

On Thu, 5 Sep 2013, Catalin Marinas wrote:

quoted

On Thu, Aug 29, 2013 at 07:32:26PM +0100, Stefano Stabellini wrote:

quoted

Move Xen initialization earlier, before any DMA requests can be made.

Signed-off-by: Stefano Stabellini <redacted>

I guess you should cc the corresponding maintainers here.

Thanks for the reminder, I'll do that.

quoted

quoted

 arch/arm/include/asm/xen/hypervisor.h |    8 ++++++++
 arch/arm/kernel/setup.c               |    2 ++
 arch/arm/xen/enlighten.c              |   21 ++++++++++++++-------
 arch/arm64/kernel/setup.c             |    2 ++
 4 files changed, 26 insertions(+), 7 deletions(-)

[...]

quoted

--- a/arch/arm64/kernel/setup.c
+++ b/arch/arm64/kernel/setup.c

@@ -53,6 +53,7 @@
 #include <asm/traps.h>
 #include <asm/memblock.h>
 #include <asm/psci.h>
+#include <asm/xen/hypervisor.h>
 
 unsigned int processor_id;
 EXPORT_SYMBOL(processor_id);

@@ -267,6 +268,7 @@ void __init setup_arch(char **cmdline_p)
 	unflatten_device_tree();
 
 	psci_init();
+	xen_early_init();

So Xen guests don't have any hope for single Image? Basically you set
dma_ops unconditionally in xen_early_init(), even if the kernel is not
intended to run under Xen.

That should not happen: if we are not running on Xen xen_early_init
returns early, before calling xen_mm_init.

x96 has a call to init_hypervisor_platform() at approximately this
location, which detects and calls the init function for any of Xen, KVM,
hyperv and vmware.

I would rather have a core_initcall(xen_early_init()) if possible,
rather than hard-coded calls in setup_arch(). This early stuff is
DT-driven, so in theory you don't need a specific xen call. The only
thing is that you end up with swiotlb_init() and 64MB wasted if the Xen
guest does not plan to use them.

There is a swiotlb_free mechanism in case the allocation was not
neccessary.

quoted

I guess only Xen and KVM are currently relevant on Linux ARM(64), so
perhaps adding similar infrastructure on ARM would be overkill at this
point. I don't know if KVM needs such an early C-land hook, I suppose
it needs it even earlier so it can set up the hyp mode trampoline from
head.S?

head.S installs a Hyp stub if it starts in that mode and then forget
about. Later when KVM is initialised it installs its own code by doing
an HVC call.

-- 
Catalin