Re: [Qemu-devel] vfio in the guest: no available reset mechanism

[Alex Williamson]

On Fri, 2014-08-01 at 18:39 +0200, Jan Kiszka wrote:
> On 2014-08-01 18:35, Alex Williamson wrote:
> > On Fri, 2014-08-01 at 09:25 -0600, Alex Williamson wrote:
> >> On Fri, 2014-08-01 at 09:35 +0800, Le Tan wrote:
> >>> Hi Alex,
> >>>
> >>> 2014-07-30 22:46 GMT+08:00 Alex Williamson <address@hidden>:
> >>>> On Wed, 2014-07-30 at 22:16 +0800, Le Tan wrote:
> >>>>> Hi Michael,
> >>>>>
> >>>>> 2014-07-30 21:16 GMT+08:00 Michael S. Tsirkin <address@hidden>:
> >>>>>> On Wed, Jul 30, 2014 at 08:24:04PM +0800, Le Tan wrote:
> >>>>>>> Hi,
> >>>>>>> I am testing vfio in L1 with my VT-d emulation project. I assigned one
> >>>>>>> of the two AHCI controllers in L1 to L2 via vfio. After I ran the QEMU
> >>>>>>> in L1, it complains that:
> >>>>>>> qemu-system-x86_64: vfio: Cannot reset device 0000:00:03.0, no
> >>>>>>> available reset mechanism.
> >>>>>>> qemu-system-x86_64: vfio: Cannot reset device 0000:00:03.0, no
> >>>>>>> available reset mechanism.
> >>>>>>>
> >>>>>>> Then L2 paused when the SeaBIOS executed in ahci_controller_setup(). I
> >>>>>>> look into this and found that:
> >>>>>>> val = ahci_ctrl_readl(ctrl, HOST_CTL);
> >>>>>>> ahci_ctrl_writel(ctrl, HOST_CTL, val | HOST_CTL_AHCI_EN);
> >>>>>>> When the BIOS tried to read the HOST_CTL, it returns 0x80000002, which
> >>>>>>> bit 2 (Interrupt Enable) is 1. The AHCI manual says that this bit
> >>>>>>> should be cleared by default. So maybe L1 didn't reset the device
> >>>>>>> before assigning it to L2?
> >>>>>>> Then the BIOS tried to write back to HOST_CTL and it was stuck here. 
> >>>>>>> :(
> >>>>>>>
> >>>>>>> So can anyone give me some advice? About the state of PCI device or
> >>>>>>> bus-level reset?
> >>>>>>>
> >>>>>>> Here is the detail of the environment and the way I did the vfio.
> >>>>>>> 1. lspci in L1 said:
> >>>>>>> 00:03.0 SATA controller [0106]: Intel Corporation 82801IR/IO/IH
> >>>>>>> (ICH9R/DO/DH) 6 port SATA Controller [AHCI mode] [8086:2922] (rev 02)
> >>>>>>> 00:1f.0 ISA bridge [0601]: Intel Corporation 82801IB (ICH9) LPC
> >>>>>>> Interface Controller [8086:2918] (rev 02)
> >>>>>>> 00:1f.2 SATA controller [0106]: Intel Corporation 82801IR/IO/IH
> >>>>>>> (ICH9R/DO/DH) 6 port SATA Controller [AHCI mode] [8086:2922] (rev 02)
> >>>>>>> 00:1f.3 SMBus [0c05]: Intel Corporation 82801I (ICH9 Family) SMBus
> >>>>>>> Controller [8086:2930] (rev 02)
> >>>>>>> 2. Unbind 00:03.0 and do vfio:
> >>>>>>> modprobe -r vfio_iommu_type1
> >>>>>>> modprobe vfio_iommu_type1 allow_unsafe_interrupts=1
> >>>>>>> modprobe vfio-pci
> >>>>>>> echo 0000:00:03.0 > /sys/bus/pci/devices/0000\:00\:03.0/driver/unbind
> >>>>>>> echo "8086 2922" > /sys/bus/pci/drivers/vfio-pci/new_id
> >>>>>>> 3. run L2 with "-device vfio-pci,host=00:03.0"
> >>>>>>>
> >>>>>>> Any help is appreciated! Thanks very much!
> >>>>>>>
> >>>>>>> Regards,
> >>>>>>> Le
> >>>>>>
> >>>>>> Clearly, bus level reset can't work for the root bus :)
> >>>>>
> >>>>> Thanks very much!
> >>>>> I test the vfio with a second-bus ahci controller and it didn't
> >>>>> complain about the lack of reset mechanism. :) And the return val of
> >>>>> HOST_CTL is normal now (the same as emulated ahci controller).
> >>>>> However, it still paused when the BIOS tried to write to the HOST_CTL.
> >>>>> Do you have any idea?
> >>>>> And we should just test vfio and legacy pci-assignment with second bus
> >>>>> devices, not considering the root-bus devices?
> >>>>
> >>>> AHCI seems like a poor choice of device for this work, they typically
> >>>> don't support any kind of reset and they can be troublesome even for the
> >>>> L1 assignment.  You really want something with FLR support so that both
> >>>> the host and L1 guest can potentially reset the device.  That said, you
> >>>> may still run into bugs with a L1 guest directed FLR.  Thanks,
> >>>>
> >>>> Alex
> >>>>
> >>>
> >>> So what device do you think is suitable for the pci-assign test? e1000?
> >>> I just tested it with sound card ac97 and USB controller. But I don't
> >>> know how to attach them to a pcie-to-pci bridge, so maybe they weren't
> >>> reset before being assigned to L2. But it seems that they can work.
> >>> 1. With the sound card, I assigned it to L2 via both vfio and legacy
> >>> pci-assign and I could hear the music played in L2 from host's
> >>> speakers. Of course, the vfio also complained about the lack of reset
> >>> mechanism.
> >>
> >> First off, maybe I'm a little confused, are you trying to assigned
> >> emulated devices for the L1 guest to the L2 guest or are these assigned
> >> devices to the L1 guest that you want to re-assign to L2?  AFAIK, we
> >> don't have any emulated devices that support reset, but it wouldn't be
> >> that hard to add a PM capability to one that advertises a soft reset on
> >> D3hot->D0 transition and calls the QEMU driver reset function when that
> >> occurs.  This would provide the most flexibility.
> > 
> > Or even better might be to add an Advanced Features capability to each
> > device so it can do an explicit FLR.  We don't tend to trust PM reset as
> > being reliably on all devices.
> 
> But FLR is PCIe-only, isn't it?

Nope, according to the ECN it's intended for conventional PCI devices on
the PCIe root complex where a complete conversion to PCIe cannot be
justified.  Maybe it's not 100% to the letter of the capability to add
it to a random non-RC device, but it seems like a good solution here.

>  Also, it may let some of our device
> models deviate from their real versions (I suppose, e.g., none of the
> e1000 devices we currently emulate exposed FLR).

Of course, but what are the chances that the driver will care?

> In any case, I suggested Le to focus on his GSoC task again (VT-d
> emulation) and postpone this topic to later point (likely outside GSoC).
> Only few weeks left...

Ok.  Thanks,

Alex