Re: [Qemu-devel] VFIO-VGA Issue

[Paolo Bonzini]

Il 10/04/2013 17:20, Alex Williamson ha scritto:
> On Wed, 2013-04-10 at 12:01 +0300, Gleb Natapov wrote:
>> On Tue, Apr 09, 2013 at 04:17:48PM -0600, Alex Williamson wrote:
>>> On Tue, 2013-04-09 at 11:18 -0600, Alex Williamson wrote:
>>>> On Tue, 2013-04-09 at 08:22 -0400, address@hidden wrote:
>>>>> Hi!
>>>>>
>>>>> I've tried passing through my ASUS Radeon HD7750, but qemu just hogs 100%
>>>>> of one core and stops. Since the display didn't light up, I couldn't see
>>>>> how far Win 7 boot went. For a Linux guest I tried redirecting kernel
>>>>> messages to serial console, but there was nothing (qemu stops even before
>>>>> kernel start?).
>>>>>
>>>>> Also, this GPU is kinda troublesome: it doesn't work with the proprietary
>>>>> driver if VT-d is enabled in BIOS (boot process stops with "dmar: DRHD:
>>>>> handling fault status reg 3"). The GPU works with the FOSS driver, though.
>>>>> Could this somehow be related to the passthrough issue?
>>>>>
>>>>> Anyway, VT-d in pci-assign mode works for network adapters on this host.
>>>>> It runs kernel 3.9.0-rc5 (with vfio-vga enabled), qemu git from yesterday,
>>>>> and seabios 1.7.2. The host has no driver installed for the GPU. Before
>>>>> assiging it to qemu I'd run "echo '1002 683f' >
>>>>> /sys/bus/pci/drivers/vfio-pci/new_id" and the same for the associated HDMI
>>>>> device. Command line:
>>>>> qemu-kvm -machine q35,accel=kvm \
>>>>>    -cpu host -smp 4 -m 2G \
>>>>>    -vga none -display none \
>>>>>    -kernel /kvm/kernel-3.7.8 -append 'root=/dev/vda console=ttyS0' \
>>>>>    -drive
>>>>> index=0,media=disk,cache=none,if=virtio,aio=native,file=/dev/vg3/gpu \
>>>>>    -device vfio-pci,host=01:00.0,x-vga=on \
>>>>>    -device vfio-pci,host=01:00.1 \
>>>>>    -serial telnet:127.0.0.1:52990,server
>>>>
>>>> You should always at least get monitor sync if it's going to work, but
>>>> you may have to reboot the host between each attempt.  I've been doing
>>>> most of the VGA development on an AMD box and swapping over to an Intel
>>>> system reminds me why.  Try removing the ",accel=kvm -cpu host -smp 4"
>>>> pieces of your command an try again.  If that works, it means there are
>>>> still broken bits of real mode handling in KVM.  My HD7850 hangs on an
>>>> Intel host the same way you describe, but posts if I disable KVM
>>>> support.  Thanks,
>>>
>>> Gleb,
>>>
>>> If I run trace-cmd while I'm in this hung state with KVM enabled, the
>>> log is filled with:
>>>
>>>  qemu-system-x86-9569  [000] 10027.806836: kvm_emulate_insn:     
>>> c0000:b2d3: ff 7d 00
>>>
>>> I hope this means more to you than it does to me ;)  Based on the
>>> segment register we're executing in the VGA BIOS for the assigned VGA
>>> device.  Let me know what more I can provide to help debug.  Thanks,
>>>
>> This is wrong encoding. Such instruction does not exists. Can you see
>> what happens before this emulation starts to happen.
> 
> Yeah, it seems like b2d3 is a bad offset, if I align the code at d2 or
> d0 gdb is able to disassemble it.  I'm not having much luck with
> breakpoints or stepping, but I'll keep fiddling with it to try to figure
> out how we're getting here.  Thanks,

How does -d in_asm look for eip values around 000cb2XX?

Paolo