Re: Summary of LPC guest MSI discussion in Santa Fe

[Alex Williamson]

On Wed, 9 Nov 2016 20:23:03 +0100
Christoffer Dall <christoffer.dall@xxxxxxxxxx> wrote:

> On Wed, Nov 09, 2016 at 01:59:07PM -0500, Don Dutile wrote:
> > On 11/09/2016 12:03 PM, Will Deacon wrote:  
> > >On Tue, Nov 08, 2016 at 09:52:33PM -0500, Don Dutile wrote:  
> > >>On 11/08/2016 06:35 PM, Alex Williamson wrote:  
> > >>>On Tue, 8 Nov 2016 21:29:22 +0100
> > >>>Christoffer Dall <christoffer.dall@xxxxxxxxxx> wrote:  
> > >>>>Is my understanding correct, that you need to tell userspace about the
> > >>>>location of the doorbell (in the IOVA space) in case (2), because even
> > >>>>though the configuration of the device is handled by the (host) kernel
> > >>>>through trapping of the BARs, we have to avoid the VFIO user programming
> > >>>>the device to create other DMA transactions to this particular address,
> > >>>>since that will obviously conflict and either not produce the desired
> > >>>>DMA transactions or result in unintended weird interrupts?  
> > >
> > >Yes, that's the crux of the issue.
> > >  
> > >>>Correct, if the MSI doorbell IOVA range overlaps RAM in the VM, then
> > >>>it's potentially a DMA target and we'll get bogus data on DMA read from
> > >>>the device, and lose data and potentially trigger spurious interrupts on
> > >>>DMA write from the device.  Thanks,
> > >>>  
> > >>That's b/c the MSI doorbells are not positioned *above* the SMMU, i.e.,
> > >>they address match before the SMMU checks are done.  if
> > >>all DMA addrs had to go through SMMU first, then the DMA access could
> > >>be ignored/rejected.  
> > >
> > >That's actually not true :( The SMMU can't generally distinguish between MSI
> > >writes and DMA writes, so it would just see a write transaction to the
> > >doorbell address, regardless of how it was generated by the endpoint.
> > >
> > >Will
> > >  
> > So, we have real systems where MSI doorbells are placed at the same IOVA
> > that could have memory for a guest  
> 
> I don't think this is a property of a hardware system.  THe problem is
> userspace not knowing where in the IOVA space the kernel is going to
> place the doorbell, so you can end up (basically by chance) that some
> IPA range of guest memory overlaps with the IOVA space for the doorbell.
> 
> 
> >, but not at the same IOVA as memory on real hw ?  
> 
> On real hardware without an IOMMU the system designer would have to
> separate the IOVA and RAM in the physical address space.  With an IOMMU,
> the SMMU driver just makes sure to allocate separate regions in the IOVA
> space.
> 
> The challenge, as I understand it, happens with the VM, because the VM
> doesn't allocate the IOVA for the MSI doorbell itself, but the host
> kernel does this, independently from the attributes (e.g. memory map) of
> the VM.
> 
> Because the IOVA is a single resource, but with two independent entities
> allocating chunks of it (the host kernel for the MSI doorbell IOVA, and
> the VFIO user for other DMA operations), you have to provide some
> coordination between those to entities to avoid conflicts.  In the case
> of KVM, the two entities are the host kernel and the VFIO user (QEMU/the
> VM), and the host kernel informs the VFIO user to never attempt to use
> the doorbell IOVA already reserved by the host kernel for DMA.
> 
> One way to do that is to ensure that the IPA space of the VFIO user
> corresponding to the doorbell IOVA is simply not valid, ie. the reserved
> regions that avoid for example QEMU to allocate RAM there.
> 
> (I suppose it's technically possible to get around this issue by letting
> QEMU place RAM wherever it wants but tell the guest to never use a
> particular subset of its RAM for DMA, because that would conflict with
> the doorbell IOVA or be seen as p2p transactions.  But I think we all
> probably agree that it's a disgusting idea.)

Well, it's not like QEMU or libvirt stumbling through sysfs to figure
out where holes could be in order to instantiate a VM with matching
holes, just in case someone might decide to hot-add a device into the
VM, at some point, and hopefully they don't migrate the VM to another
host with a different layout first, is all that much less disgusting or
foolproof. It's just that in order to dynamically remove a page as a
possible DMA target we require a paravirt channel, such as a balloon
driver that's able to pluck a specific page.  In some ways it's
actually less disgusting, but it puts some prerequisites on
enlightening the guest OS.  Thanks,

Alex