Re: [Xen-devel] VGA passthrough and AMD drivers

[Sander Eikelenboom <linux@xxxxxxxxxxxxxx>]

Monday, December 10, 2012, 5:49:16 PM, you wrote:

>>>>>>> Hi all,
>>>>>>> I have made some tests to find a good driver for FirePro V8800 on 
>>>>>>> windows 7 64bit HVM.
>>>>>>> I have been focused on ?advanced features?: quad buffer and active 
>>>>>>> stereoscopy, synchronization ?
>>>>>>> The results, for all FirePro drivers (of this year); I can?t get 
>>>>>>> the quad buffer/active stereoscopy feature.
>>>>>>> But they work on a native installation.
>>>>>> Can you describe the setup a little more?
>>>>> I?ve got 2 HP Z800 workstation with FirePro V8800, one per computer.
>>>>>
>>>>> It?s a setup used in CAVE system, I try (and its works, minus some
>>>>> issues) to virtualize ?virtual reality contexts? that needs full 
>>>>> graphics card features.
>>>>>
>>>>> Intel Xeon E5640 CPU with Intel 5520 chipset
>>>>>
>>>>> cores_per_socket : 4
>>>>>
>>>>> threads_per_core : 2
>>>>>
>>>>> cpu_mhz : 2660
>>>>>
>>>>> total_memory : 4079
>>>>>
>>>>>> How many graphic cards per guest?
>>>>> One card per guest.
>>>>>
>>>>>> How many guests? On how many hosts?
>>>>> One guest per computer.
>>>>>
>>>> And of course, I just thought of some other questions:
>>>> What version of Xen are you using?
>>>> What kernel are you using in Dom0?
>>> release                : 2.6.32-5-xen-amd64
>>> version                : #1 SMP Sun May 6 08:57:29 UTC 2012
>>> machine                : x86_64
>>> nr_cpus                : 8
>>> nr_nodes               : 1
>>> cores_per_socket       : 4
>>> threads_per_core       : 2
>>> cpu_mhz                : 2660
>>> hw_caps                : 
>>> bfebfbff:2c100800:00000000:00003f40:029ee3ff:00000000:00000001:00000000
>>> virt_caps              : hvm hvm_directio
>>> total_memory           : 4079
>>> free_cpus              : 0
>>> xen_major              : 4
>>> xen_minor              : 2
>>> xen_extra              : -unstable
>>> xen_caps               : xen-3.0-x86_64 xen-3.0-x86_32p hvm-3.0-x86_32 
>>> hvm-3.0-x86_32p hvm-3.0-x86_64
>>> xen_scheduler          : credit
>>> xen_pagesize           : 4096
>>> platform_params        : virt_start=0xffff800000000000
>>> xen_changeset          : Sun Jul 22 16:37:25 2012 +0100 25622:3c426da4788e
>>> xen_commandline        : placeholder
>>> cc_compiler            : gcc version 4.4.5 (Debian 4.4.5-8)
>>> xend_config_format     : 4
>>>
>>> I will change to a newer version and use  xl toolstack when VGA passthrough 
>>> will be supported.
>>>
>>>> And just to be clear, there is only Dom0 and one Windows 7 HVM guest on 
>>>> each machine?
>>> Yes
>>>
>>>>>>> The only driver that allows this feature is a Radeon HD driver 
>>>>>>> (Catalyst 12.10 WHQL).
>>>>>>> But this driver becomes unstable when an application using active 
>>>>>>> stereo and synchronization is closed:
>>>>>>> -The synchronization between two computers is lost.
>>>>>>> -The CCC can crash when the synchronization is made again.
>>>>>>> Someone have any clues about this?
>>>>>> I don't know exactly how this works on AMD/ATI graphics cards, but 
>>>>>> I have worked with synchronisation on other graphics cards about 7 
>>>>>> years ago, so I have some idea of how you solve the various 
>>>>>> problems.
>>>>>> What I don't quite understand is why it would be different between 
>>>>>> a virtual environment and the bare-metal ("native") install. My 
>>>>>> immediate guess is that there is a timing difference, for one of 
>>>>>> two reasons:
>>>>>> 1. IOMMU is adding extra delays to the graphics card reading system 
>>>>>> memory.
>>>>>> 2. Interrupt delays due to hypervisor.
>>>>>> 3. Dom0 or other guest domains "stealing" CPU from the guest.
>>>>>> I don't think those are easy to work around (as they all have to 
>>>>>> "happen" in a virtual system), but I also don't REALLY understand 
>>>>>> why this should cause problems in the first place, as there isn't 
>>>>>> any guarantee as to the timings of either memory reads, interrupt 
>>>>>> latency/responsiveness or CPU availability in Windows, so the same 
>>>>>> problem would appear in native systems as well, given "the right"
>>>>>> circumstances.
>>>>>> What exactly is the crash in CCC?
>>>>>> (CCC stands for "Catalyst Control Center" - which I think is a 
>>>>>> Windows "service" to handle certain requests from the driver that 
>>>>>> can't be done in kernel mode [or shouldn't be done in the driver in 
>>>>>> general]).
>>>>> After the application is closed, I launch the Catalyst Control 
>>>>> Center, the synchronization state seems to be good. But there is no 
>>>>> synchronization.
>>>>>
>>>>> If I try to apply any modifications of synchronization (synchro 
>>>>> server or client), CCC freeze and I need to kill it manually.
>>>>>
>>>>> I can set the synchronization back after this.
>>>>>
>>>> This clearly sounds like a software issue in the CCC itself. I could be 
>>>> wrong, but that's what I think right now. It would be rather difficult to 
>>>> figure out what is going wrong without at least a repro environment.
>>> I've made a bunch of tests this morning:
>>> -CCC crash when I've got two displays: I set one to be the synchronization 
>>> server and the other a client at the same time. When I set the server, 
>>> apply this configuration and set the client after, it didn't crash.
>>> -If my application (Virtools) crash, synchronization is reset.
>>> -Eyes are sometimes inverted with the same trigger edge.
>>I saw that problem with the product I was working on once or twice. 
>>Makes it look really "confusing". This was a settings problem in my case 
>>(because I wrote my own "controls", I could set almost every aspect of 
>>everything that could possibly be changed, with a very basic command line 
>>>application that interacted pretty straight down to the driver - with the 
>>usual caveat of "make sure you know what you are doing" - the normal GUI 
>>Control panel setup was much more "you can only set things that make sense 
>>>for you to set"). That is probably not really what your problem is... But 
>>could be a configuration of driver or application issue, of course.
>>
>>>
>>> I've got all this behaviors with both HVM and native installation under 7 
>>> 64bits.  So I think it's clearly a software issue.
>>>
>>> Next step:  7 32bits.
>>So, this is not a Xen issue... Report it to the ATI/AMD folks!
>>
> Yes, but it doesn't explain why I can't get active stereoscopy with FirePro 
> drivers on HVM.

>>>> Whilst I'm all for using Xen for everything, there are sometimes 
>>>> situations when "not using Xen" may actually be the right choice. Can you 
>>>> explain why running your guests in Xen is of benefit? [If you'd like to 
>>>> answer "none of >>>your business", that's fine, but it may help to 
>>>> understand what the "business case" is for this].
>>> The objective is to mutualize graphical cluster for immersive systems. 
>>> Virtual Reality applications are sensitive in their configurations; it's a 
>>> pain to manage multiple users and it's nearly impossible to have different 
>>> configurations for these users. Usually immersive systems are stuck in one 
>>> configuration (OS, drivers, applications ...), and only few people are 
>>> allowed to change settings.
>>> The idea is to use Xen and VGA passthrough, for create personals 
>>> environments that allow every user to make their own configurations without 
>>> impacts on others.
>>>
>>> Be able to have VR configurations in virtual machines and to able to run it 
>>> with 3D features, is a serious benefit for Virtual Reality users.
>>
>>Thanks for your explanation. Makes some sense, however, I feel that it also 
>>makes things more complex - if the system is so sensitive, it may get "upset" 
>>simply by having the differences in system behaviour that you >automatically 
>>get from running on a virtual machine vs. "bare metal". Don't let that stop 
>>you, I'm just saying there may be issues caused by Xen (or other 
>>Virtualisation products) are not quite as transparent as they really should 
>>>be.
>>

> It's not the hardware configurations that is so sensitive but more the 
> software configurations and drivers versions.  
> I've already made some demonstrations of Xen capabilities in our use case, 
> there is no negative feedback. I think that HVM behavior is perfect for our 
> uses, except these driver issues. 

> I found one minor bug (for us), if the first HVM executed (id=1) has the VGA 
> card, the computer reboot without logs. 
> My workaround is to launch an HVM without VGA first, stop it properly, and 
> launch my usual HVM with VGA passthrough. 
> I think, it's a bug due to my installation (Xen 4.2.0-unstable).

> I just got a new test computer, a Dell Precision T7500 with a V9800 FirePro, 
> maybe I will have the time to test something tomorrow!

Hi,

My experiences with CCC and vga passthrough aren't great either (bluescreen 
most of the time).
It works now with the 12.11 catalyst beta drivers and not installing CCC, but 
just installing the driver and opencl packages from the c:\AMD\packages dir 
after stopping the install after the total package is unpacked.
Don't know if the stereoscopy also comes in a seperate package.

It runs opencl fine now, with near native performance (with luxmark benchmark) 
:-)
(with xen-unstable, qemu-upstream, linux 3.7 dom0, win7 guest, 12.11 catalyst 
drivers, ati radeon HD 6570 at the moment)

--
Sander

> Aurelien 
>>--
>>Mats
>>




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