On Tue, 2013-11-12 at 07:55 +1000, Dave Airlie wrote: > On Tue, Nov 12, 2013 at 7:43 AM, Alex Williamson > <alex.william...@redhat.com> wrote: > > When MSI is enabled on Nvidia GeForce cards the driver seems to > > acknowledge the interrupt by writing a 0xff byte to the MSI capability > > ID register using the PCI config space mirror at offset 0x88000 from > > BAR0. Without this, the device will only fire a single interrupt. > > VFIO handles the PCI capability ID/next registers as virtual w/o write > > support, so any write through config space is currently dropped. Add > > a check for this and allow the write through the BAR window. The > > registers are read-only anyway. > > This is only half the truth, I'm afraid if I'm right its much worse than that. > > At least on some GPUs the MSI ack is done via PCI config space itself, > and on some its done via the mirror, and yes it matters on some cards > which way it works.
I was hoping that wouldn't be the case since it seems fairly universal that PCI config space access should be considered slow and avoided for things like this. But, I suppose with MMConfig it's no worse than device MMIO space. On my GTX660 I did actually test both paths. The existing quirk converts all accesses into the config space mirror into proper config space accesses, so I changed the kernel to not drop the write. Then I figured I'd rather solve this in QEMU using the mirror directly. So in my case it didn't matter which path. I'll keep this in mind though, I may want to enable kernel support for this first and advertise it via a flag on the config space region, then I can avoid doing a double write in QEMU and really breaking the state machine. Thanks for the feedback, Alex