Re: [Qemu-devel] [PATCH] hw/misc: slavepci_passthru driver

[Alex Williamson <alex.williamson@redhat.com>]

On Mon, 2016-01-18 at 10:16 -0500, Marc-André Lureau wrote:
> Hi
> 
> ----- Original Message -----
> > Hi there,
> > 
> > I'd like to submit this new pci driver ( hw/misc )for inclusion,
> > if you think it could be useful to other as well as ourself.
> > 
> > The driver "worked for our needs" BUT we haven't done extensive
> > testing and this is our first attempt to submit a patch so I kindly
> > ask for extra-forgiveness .
> > 
> > The "slavepci_passthru" driver is useful in the scenario described
> > below to implement a simplified passthru when the host CPU does not
> > support IOMMU and one is interested only in pci target-mode (slave
> > devices).
> 
> Let's CC Alex, who worked on the most recent framework for something related to that (VFIO).
> 
> > 
> > Embedded system cpu (e.g. Atom, AMD G-Series) often lack the VT-d
> > extensions (IOMMU) needed to be able to pass-thru pci peripherals to
> > the guest machine (i.e. the pci pass-thru feature cannot be used).
> > 
> > If one is only interested in using the pci board as a pci-target
> > (slave device), this driver mmap(s) the host-pci-bars into the guest
> > within a virtual pci-device.

What exactly do you mean by pci-target/slave device?  Does this mean
that the device is not DMA capable, ie. cannot enable BusMaster?

> > This is useful in our case for debugging via qemu gsbserver facility
> > (i.e. '-s' option in qemu) a system running barebone-executable .
> > 
> > Currently the driver assumes the custom pci card has four 32-bit bars
> > to be mapped (in current patch this is mandatory)
> > 
> > HowTo:
> > To use the new driver one shall:
> > - define two environment variables for assigning proper VID and DID to
> >   associate to the guest pci card
> > - give the host pci bar address to map in the guest.
> > 
> > Example Usage:
> > 
> > Let us suppose that we have in the host a slave pci device with the
> > following 4 bars (i.e. output of lspci -v -s YOUR-CARD | grep Memory)
> >   Memory at db800000 (32-bit, non-prefetchable) [size=4K]
> >   Memory at db900000 (32-bit, non-prefetchable) [size=8K]
> >   Memory at dba00000 (32-bit, non-prefetchable) [size=4K]
> >   Memory at dbb00000 (32-bit, non-prefetchable) [size=4K]
> > 
> > We can map these bars in a guest-pci with VID=0xe33e DID=0x000a using
> > 
> > SLAVEPASSTHRU_VID="0xe33e" SLAVEPASSTHRU_DID="0xa" qemu-system-x86_64 \
> >   YOUR-SET-OF-FLAGS \
> >   -device
> >   slavepassthru,size1=4096,baseaddr1=0xdb900000,size2=8192,baseaddr2=0xdba00000,size3=4096,baseaddr3=0xdbd00000,size4=4096,baseaddr4=0xdbe00000
> > 
> > Please note that if your device has less than four bars you can give
> > the same size and baseaddress to the unused bars.

Those are some pretty serious usage restrictions and using /dev/mem is
really not practical.  The resource files in pci-sysfs would even be a
better option.  I didn't see how IO and MMIO BARs get enabled on the
physical device or whether you support any kind of interrupt scheme.  I
had never really intended QEMU use of this, but you might want to
consider vfio no-iommu mode:

http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/drivers/vfio/vfio.c?id=03a76b60f8ba27974e2d252bc555d2c103420e15

Using this taints the kernel, but maybe that's nothing you mind if
you're already letting QEMU access /dev/mem.  The QEMU vfio-pci driver
would need to be modified to use the new device and of course it
wouldn't have IOMMU translation capabilities.  That means that the
BusMaster bit should protected and MSI/X capabilities should be hidden
from the VM.  It seems more flexible and featureful than what you have
here.  Thanks,

Alex