Re: [Xen-devel] [PATCH OSSTEST v5] Add arm64 build and test jobs

[Ian Campbell <ian.campbell@xxxxxxxxxx>]

On Tue, 2015-09-29 at 16:33 +0100, Ian Jackson wrote:
> Ian Campbell writes ("[PATCH OSSTEST v5] Add arm64 build and test jobs"):
> > Runvars for the xen-unstable flight build jobs and an illustrative
> > test:
> ...
> 
> Reading this again:
> 
> This is approximately doubling our ARM test bandwidth requirements,
> isn't it ?

Not quite, since the arm32 and arm64 h/w is distinct. These tests won't run
on any of the arm32 hardware we have and the existing armhf tests won't run
on the new hwardware (for reasons I'll hopefully explain in the course of
answering the next question below).

> > test-arm64-arm64-libvirt
> > test-arm64-arm64-libvirt-qcow2
> > test-arm64-arm64-libvirt-raw
> > test-arm64-arm64-libvirt-vhd
> > test-arm64-arm64-libvirt-xsm
> > test-arm64-arm64-xl
> > test-arm64-arm64-xl-credit2
> > test-arm64-arm64-xl-multivcpu
> > test-arm64-arm64-xl-qcow2
> > test-arm64-arm64-xl-raw
> > test-arm64-arm64-xl-rtds
> > test-arm64-arm64-xl-vhd
> > test-arm64-arm64-xl-xsm
> 
> Maybe it would be worth doing some of these stripy in arch and disk
> format ?
> 
> Also, why are there no
>   test-arm64-armhf
> jobs ?

"It's complicated".

On x86 it is expected that a 64-bit processor will support 32-bit just fine
and everything (drivers etc) will be available etc, and people even
routinely test such processors running in 32-bit mode.

For ARMv8 the standard allows for implementations which include AArch32 (32
-bit ISA mostly compatible with ARMv7) at all privilege levels.

But each new ARMv8 processor implementation involves a certain amount of
per vendor code in Linux in arch/arm64 (the AArch64 AKA 64-bit port). That
support is rarely (never?) also added to arch/arm (the 32-bit port). Hence
it would be unusual to even find a 32-bit kernel to run on a 64-bit capable
processor.

Unlike with x86 it's unclear if there will ever be sufficient legacy drag
to cause such support to become a thing. I personally doubt it will.

Now there is another possibility, which is 32-bit userspace on a 64-bit
processor, which is something which Xen on ARM could support. I just didn't
think it was worth adding testing of that just now. It adds another nested
loop to make-flight and some complexity in the test naming etc (test-arm64
-armhf-armhf-foo?). In reality I think it unlikely people will use that for
dom0 though (there's no real reason to do so AFAICT).

I suppose there is also a fourth possibility which is a fully 64-bit host
with 32-bit guests, which is probably the most plausible and useful one,
and something we ought to consider testing at some point.

Ian.

> > In addition test-arm64-arm64-xl and appropriate build jobs are added
> > to xen-unstable-smoke.
> 
> I think this is probably fine provided the arm test bandwidth doesn't
> start to be the limiting factor.
> 
> Ian.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel