Re: [linux-linus test] 25478: regressions - FAIL [and 1 more messages]

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

On 16/04/14 12:47, Ian Jackson wrote:
> David Vrabel writes ("Re: [Xen-devel] [linux-linus test] 25478: regressions - FAIL [and 1 more messages]"):
>> On 16/04/14 11:12, Ian Jackson wrote:
>>> Ian Jackson writes ("Re: [Xen-devel] [linux-linus test] 25478: regressions - FAIL [and 1 more messages]"):
>>>> Are we any closer to figuring out how to get a fix for this past the
>>>> x86 maintainers ?
>>>>
>>>> Does the Linux kernel not have a "no regressions" policy ?  This is a
>>>> regression, after all...
>>>
>>> Another week has gone by and Linux tip still fails this test.
>>
>> Yes, You've not fixed it yet.
> 
> My understanding from private emails is that the problem is mostly
> political rather than technical.
> 
> You wrote:
>>> mptsas is a driver that uses dma_get_required_mask() to determine a
>>> "suitable" DMA mask -- but under Xen dma_get_required_mask() may return
>>> the wrong mask since it only gets the physical RAM range and not machine
>>> addresses.

For this specific problem with the mptsas driver, yes.  But there other
problems with swiotlb usage such as by skbs using compound pages.

We (XenServer) plan to fix all these by making use of the IOMMU. See
Malcolm's recent design doc.  This is still quite a ways off though.

> I'm not familiar with the Linux kernel's VM system.  However, if
> someone would write a patch which provides an arch override for this
> (which from the private emails seems like is what required and would
> not be too hard for someone who knew what they were doing), then I can
> try to do the political work of negotiating with the Linux community.

This fix doesn't require any particular knowledge of the VM subsystem
and there's already some infrastructure for providing an arch specific
implementation of dma_get_required_mask().

David