[vfio-users] vfio help needed (nvidia GTX 750 Ti persistent code 43)

[Blank Field]

>VFIO configures memory regions
The documentation is in the source code as usual or there is a text file
somewhere telling more about this?
And this is not related to MTRR, right?
On Sep 17, 2015 7:35 AM, "Alex Williamson" <alex.williamson at redhat.com>
wrote:

> On Wed, Sep 16, 2015 at 10:18 PM, Blank Field <ihatethisfield at gmail.com>
> wrote:
>
>> You have an AMD FM2 CPU.
>> Do use SeaBIOS and make use of VGA I/O lines, you must have a VideoBIOS
>> Extension support in your GPU's firmware. Since we are in the 21'st
>> century, you have it.
>> But the problem is that in order to make VGA work in a VM we must
>> translate MMIO, PIO and IRQs needed into the VM.
>> That is where IOMMU kicks in, and the related part of software is
>> vfio-pci's x-vga option.
>> Usually you enable x-vga, vfio tells IOMMU to translate all VGA related
>> stuff from GPU to the VM and vice versa, but on that particular
>> platform(AMD FM2) something is broken in hardware that crashes the IOMMU
>> and the CPU when trying to work with VGA.
>>
>
> Actually the IOMMU isn't involved with VGA access.  The IOMMU is only
> involved for DMA to or from the device.  These are device initiated memory
> transfers.  VGA is not a DMA model, it's a programmed I/O model, where
> every access is an explicit read or write initiated by the CPU.  In the
> case of a VM, the vCPU is the initiator and vfio configures memory regions
> for the VGA address space such that a vCPU traps into QEMU on access.  The
> access is then completed through the vfio kernel module by way of the VGA
> arbitration code.
>
> Something may still be broken in FM2 hardware, but hopefully this clears
> up how the IOMMU (doesn't) participate in VGA access.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/vfio-users/attachments/20150917/09f6bc95/attachment.htm>