Re: [PATCH] VT-d: add iommu=igfx_off option to workaround graphics issues

From: Andrew Cooper <andrew.cooper3@citrix.com>
To: Jan Beulich <JBeulich@suse.com>
Cc: Ting-Wei Lan <lantw44@gmail.com>,
	Yang Zhang <yang.z.zhang@intel.com>,
	Kevin Tian <kevin.tian@intel.com>,
	xen-devel@lists.xen.org
Subject: Re: [PATCH] VT-d: add iommu=igfx_off option to workaround graphics issues
Date: Mon, 20 Jul 2015 13:34:03 +0100	[thread overview]
Message-ID: <55ACEABB.300@citrix.com> (raw)
In-Reply-To: <55AD04890200007800093114@mail.emea.novell.com>

On 20/07/15 13:24, Jan Beulich wrote:
>>>> On 20.07.15 at 14:12, <andrew.cooper3@citrix.com> wrote:
>> On 17/07/15 20:05, Ting-Wei Lan wrote:
>>> When using Linux >= 3.19 (commit 47591df) as dom0 on some Intel Ironlake
>>> devices, It is possible to encounter graphics issues that make screen
>>> unreadable or crash the system. It was reported in freedesktop bugzilla:
>>>
>>> https://bugs.freedesktop.org/show_bug.cgi?id=90037 
>>>
>>> As we still cannot find a proper fix for this problem, this patch adds
>>> iommu=igfx_off option that is similar to Linux intel_iommu=igfx_off for
>>> users to manually workaround the problem.
>>>
>>> Signed-off-by: Ting-Wei Lan <lantw44@gmail.com>
>> Having looked into this issue, the i915 driver has several workarounds
>> in it for systems when the IOMMU is in use.  In some cases there are
>> plain errata, while in other cases there are specific hardware features
>> which don't function if the IOMMU is enabled.
>>
>> In all cases this is gated on Linux's idea of whether the IOMMU is
>> enabled.  When used under Xen, Linux has no clue that the IOMMU exists,
>> or that Xen has turned it on.
> Perhaps it should just assume an IOMMU is in use when running under
> Xen. Having inspected all those code places quite some time ago, I
> came to the conclusion that making this assumption is better than
> the current one of there not being an enabled IOMMU (and I adjusted
> our kernels accordingly).

In at least one case, an errata workaround involves issuing extra IOMMU
commands.  We cannot safely let even dom0 perform this.

>
>> Longterm, we should probably see about not clobbering the DMAR table and
>> making dom0 aware that there is an IOMMU.  (I still don't actually know
>> why we have this asymmetry between Intel and AMD).
> The presence of a DMAR table is not a 1:1 representation of whether
> Xen has the IOMMU enabled.

It would (in theory) allow dom0 to get RO mappings of the IOMMU control
pages and observe that the IOMMU is enabled.  Whether this is useful in
practice is a different matter.

~Andrew