Re: [xen-4.4-testing test] 33394: regressions - FAIL

[David Vrabel <david.vrabel@citrix.com>]

On 14/01/15 09:52, Ian Campbell wrote:
> On Wed, 2015-01-14 at 09:12 +0000, Jan Beulich wrote:
>>>>> On 13.01.15 at 21:30, <Ian.Jackson@eu.citrix.com> wrote:
>>> flight 33394 xen-4.4-testing real [real]
>>> http://www.chiark.greenend.org.uk/~xensrcts/logs/33394/ 
>>>
>>> Regressions :-(
>>>
>>> Tests which did not succeed and are blocking,
>>> including tests which could not be run:
>>>  test-amd64-i386-xl-qemut-winxpsp3-vcpus1  5 xen-boot      fail REGR. vs. 33292
>>
>> In all recent xen-boot failures I looked at the boot process seems
>> stalled starting ntpd - is that perhaps infrastructure related?
> 
> That's not impossible, I'd suspect DNS related. Hopefully all this sort
> of thing will be much better (or at least under our control) in the new
> colo.
> 
>>>  test-amd64-i386-pair   17 guest-migrate/src_host/dst_host fail REGR. vs. 33292
>>
>> Both involved hosts appear to still be affected by the swiotlb issue
>> - wasn't that supposed to long be addressed?
> 
> The patches to increase the swiotlb size in osstest didn't help, I
> think.
> 
> David V posted some patches upstream to fix the problem at the root, but
> I don't remember seeing any responses from the upstream maintainers
> (they touched some common and/or x86 code IIRC)

There are two main users of the swiotlb.

1. The mptsas driver incorrectly setting the DMA mask based on max_pfn.

The series I posted addresses this. But after only 1 (rude) response
from the relevant maintainers in 5 different versions of this series I
have no wish to pursue this further.

Someone else will need to pick it up if they care.

You can work around this by giving more than 4 GiB of PFN space to dom0
(e.g, dom0_mem=2G,max:4G would work).

2. > 4k buffers (from block I/O or network packets using compound pages.

This is only a problem with debug builds of Xen where pages are
deliberately allocated in reverse order.  In non-debug builds this is
much less of a problem (most of dom0's memory ends up being machine
contiguous).

Our (XenServer's) long term plan is to introduce the PV IOMMU support to
make bus addresses contiguous.  Malcolm will be able to advise on what
work remains if you wish to pick this up.

David