Re: [linux-linus test] 30356: regressions - FAIL

[Ian Campbell <Ian.Campbell@citrix.com>]

On Wed, 2014-09-24 at 00:35 +0100, Julien Grall wrote:
> 
> On 23/09/2014 19:40, Ian Campbell wrote:
> > On Tue, 2014-09-23 at 18:31 +0100, Julien Grall wrote:
> >> I guess we will have to select LPAE when XEN is enabled, right? If
> >> it's the case that would mean the user won't be able to compile a
> >> Linux guest with short page table and Xen.
> >>
> >> Any though?
> >
> > Things must work without in guest LPAE too, so something somewhere else
> > will need fixing.
> >
> > Apart from restricting the user in an unwanted way requiring LPAE will
> > mean that practically no distro installer will work in a Xen guest.
> 
> Xen does an identity mapping for the host physical address into DOM0
> for the grant mapping. DOM0 will use a scratch page (see commit 340720b
> "xen/arm: reimplement xen_dma_unmap_page & friends") and map and this 
> physical address.
> 
> That means on platform with an address space higher than 32 bits, which 
> is the case on Midway, we have to handle 64 bits physical address in DOM0.
> 
> With the current implementation in Linux we can only use LPAE when a 
> guest is started. The distro installer will still be able to work with
> short page table.
> 
> The drawback is we are requiring LPAE from DOM0 and a different kernel
> in the guest if the user doesn't want to use LPAE.
> 
> As the code is already pushed in Linux 3.17, I don't find a simpler 
> solution to fix Linux boot without requiring LPAE.

We will have to try harder then, requiring LPAE simply isn't acceptable
IMHO.

Ian.