Re: [Qemu-devel] [Qemu-ppc] [PATCH 0/2] PPC: kvm: Fix incorrect remapping of in-kernel MPIC

[Alexander Graf <agraf@suse.de>]

On 05.09.14 16:31, mihai.caraman@freescale.com wrote:
>> -----Original Message-----
>> From: qemu-devel-bounces+mihai.caraman=freescale.com@nongnu.org
>> [mailto:qemu-devel-bounces+mihai.caraman=freescale.com@nongnu.org] On
>> Behalf Of Alexander Graf
>> Sent: Friday, September 05, 2014 12:08 PM
>> To: Purcareata Bogdan-B43198; qemu-ppc@nongnu.org
>> Cc: qemu-devel@nongnu.org
>> Subject: Re: [Qemu-devel] [Qemu-ppc] [PATCH 0/2] PPC: kvm: Fix incorrect
>> remapping of in-kernel MPIC
>>
>>
>>
>> On 03.09.14 20:36, Bogdan Purcareata wrote:
>>> On target-ppc, the kvm-openpic memory region is part of the E500-CCSR
>> memory
>>> region. On the kernel side, the MPIC is mapped at the same offset as
>> the
>>> kvm-openpic within the address space.
>>>
>>> When adding the PCI BAR0 memory region, an alias is created to point to
>> the
>>> E500-CCSR memory region. This results in firing the
>> kvm_openpic_region_add once
>>> more, since kvm-openpic is part of the latter. Only this time, the
>> offset is
>>> wrong - it's part of the PCI memory region. This leads to the in-kernel
>> MPIC to
>>> be remapped at a wrong address, and thus all traps to the kvm-openpic
>>> address to be emulated in userspace.
>>>
>>> The fix consists in an additional filter in
>> kvm_openpic_region_{add,del} to
>>> consider only addresses matching the start of the kvm-openpic memory
>> region.
>>
>> If this is true, wouldn't vfio and host be broken too?
> 
> You should have put the same question for 87d8354d "PPC: openpic_kvm: Filter
> memory events properly". I think vhost and vfio (except for peer to peer PCI)
> use region_add memory listener because they need to access the _RAM_ memory
> for DMA, so they skip BAR notifications (at least in FSL SDK version of qemu).
> Openpic on the other hand uses region_add as a trigger for KVM_SET_DEVICE_ATTR
> ioctl (the device base address) so it takes into account non-RAM memory regions.
> 
> Vhost uses another memory listener, eventfd_add that follows a slightly
> different path then region_add, as a trigger to call KVM_IOEVENTFD ioctl.
> Though vhost seems to work properly we can further trace the ioctl to double
> check.
> 
> Peer to peer PCI might reveal the issue on vfio but this feature is not
> supported by the current FSL PAMU driver. If you think of another platform
> which supports peer to peer PCI and registers a memory region alias like
> this patch do 3eddc1be "Adding BAR0 for e500 PCI controller", then it worth
> validating it.
> 
> I see that vfio_listener_skipped_section() changed upstream so vfio may not
> skip BAR notifications anymore. What qemu version are you using on FSL boards
> like T424QDS, are you using top of the tree? If it works we would like to try
> it to validate vfio and vhost.

Yes I'm running pure upstream code. However, I haven't tried to use VFIO
on e500 at all yet.


Alex