On Sat, 2013-02-23 at 12:05 -0800, Jonathan Nieder wrote:
found 697029 linux/3.7.3-1~experimental.1
tags 697029 - wontfix
affects 697029 + release-notes
quit
Hi,
Chris Wilson wrote:
The performance issue on 3.7 is not due to the missed irq, but a
combination of using UXA and VT-d. In order to workaround an erratum
on Ironlake, every time we touch the GPU's page tables, we have to
idle the GPU before doing so. This causes extremely noticeable display
lag.
For reference, the workaround seems to be implemented by:
commit 6fbcfb3e467adb414e235eeefaeaf51ad12f2461
Author: David Woodhouse <dwmw2@infradead.org>
Date: Sun Sep 25 19:11:14 2011 -0700
intel-iommu: Workaround IOTLB hang on Ironlake GPU
commit 5c0422878fcdc279ae9a8e8b66972a15b5efb67f
Author: Ben Widawsky <ben@bwidawsk.net>
Date: Mon Oct 17 15:51:55 2011 -0700
drm/i915: ILK + VT-d workaround
However the latter is applied only to Ironlake M (mobile). So either
there are two different bugs or there is some confusion about which
chips have the bug.
Pierre AUSSAGUEL wrote:
Appending intel_iommu=off seems to fix the problem (I tested a few
days befor posting).
Daniel Vetter wrote:
Since we can't fix the hw, closing this as wontfix. Thanks for
reporting this issue anyway.
That makes this a distro issue, I suppose.
Ben and X team, any ideas? Would it makes sense to disable
intel_iommu by default on this hardware and require intel_iommu=on to
reenable it?
Use of an IOMMU is in part a performance vs security trade-off. I tend
to think that security settings should have consistent defaults, as
otherwise users may assume that a security feature is enabled when it is
not.
Aside from that, the Intel IOMMU can be enabled separately per device
(except behind PCI bridges). Since IGPs aren't real PCI(e) devices and
Intel has not always validated their interaction with the IOMMU, they
often don't work with it. There is already a kernel parameter to
disable it for the IGP ('intel_iommu=igfx_off') and a quirk to do so
automatically for the G4x and GM45. Maybe the thing to do is to log a
message about this parameter when enabling the workaround for Ironlake.
Should GNOME somehow detect that it should use classic
mode by default when the iommu is enabled?
I think this would be a very poor heuristic.
If we can't come up with a workaround, this should be mentioned in the
release notes to prevent a regression on upgrade. Please feel free to
remind me in that case so I can come up with some wording (though I
also wouldn't mind if someone else does).
"Graphics rendering may be very slow on Intel 'Ironlake M' GPUs (PCI ID
8086:0046) when an IOMMU (VT-d) is enabled. The IOMMU functionality can
be disabled for the GPU by adding the kernel parameter
'intel_iommu=igfx_off'."