Xen project Mailing List

All,

On Wed, Mar 12, 2014 at 7:33 AM, Gordan Bobic <gordan@xxxxxxxxxx> wrote:

On 03/12/2014 09:25 AM, H. Sieger wrote:

Thanks for the detailed reply, Gordon. Very helpful!

Can someone tell me which Xen release implements reinitializing passed
through devices to fix the AMD driver bug when rebooting the domU?

It's definitely not in 4.3.x (what I'm running). I don't know if it made it into the 4.4.x release.

When I was rebuilding the Arch Linux PKGBUILD file for 4.4, I tried enabling the old ATI patch and it applied cleanly - I didn't see any mention of its inclusion so suspect it hasn't. What probably needs to happen is a patch that allows a user to specify a VGA BIOS image to use (or access and use the existing one, like the old ATI patch does) to help with the initialization; unfortunately I've not had much experience in that particular area and don't have enough personal time to get up to speed so could do it.

I believe I read somewhere that perhaps a new AMD graphics card driver
might solve the issue. Does anybody know about that?

It could - but that would require ATI to fix their driver - which given how otherwise broken their drivers are in various ways, I wouldn't hold my breath for.

I tried getting an ATI based solution very hard, but in the end I failed to achieve anything workable. I got a Quadro 2000 for testing, which "just worked". So I cut my losses, traded in my ATIs for Nvidias and have been using rather than troubleshooting the system since. From there on I went with modifying GTS450 cards into Quadro 2000s, then went on to GTX470 and 480, modifying them to Quadro 5000 and 6000 respectively, because it was easy and they could be soft modded. More recently I've been using GTX680/GTX690 (Tesla K10 / Grid K2) and GTX780Ti (Quadro K6000) cards but that requires physically modifying the card.

Or, if your decompiling and assembly-foo is strong, and you are that way inclined you could void your EULA and modify the binary driver to neuter it's decision to not initialize the card if it's device ID isn't on the whitelist. Something similar was also done to enable SLI across dissimilar cards (google for DifferentSLI).

It's quite strange that on one side AMD supports VGA passthrough (a lot
more than Nvidia does), but on the other side they haven't fixed their
Windows driver to properly reset the card (at least that is how I
understand it).

AMD have always been very good at making positive noises about supporting various open source things, but very little has ever come of it. For example, they support GCC, whereas Intel has the closed-source commercial ICC (or whatever they call it nowdays). While AMD's efforts are ideologically commendable, Intel's compiler is still, sadly way, way ahead in features and performance of the code it produces - including when said code runs on AMD processors (although Intel did get some very bad publicity back when they released a version of ICC generated code that detected the CPU ID and if it wasn't Intel it disabled vectorization which is where most of the speed-up comes from, but this doesn't happen any more, they are happy to take your money for the compiler even if you are going to run your code on an AMD CPU).

This domU reboot issue with xl has been around for some time and I was
hoping that some fix has trickled into a recent Xen release or AMD
driver release.

It has been around for years, but then again so have a lot of other issues with ATI drivers. In the end it won't be fixed in the driver, it'll be fixed by Xen adding bus reset to the list of reset methods.

Unfortunately, the choice is between two solutions, neither of which is ideal. But one of the options is a lot more workable.

Gordan

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

Regards,

David

Re: [Xen-users] VGA passthrough with Xen 4.3 and xl toolstack - performance degradation resolved?