On Tue, 13 Aug 2024 20:37:24 -0300 Jason Gunthorpe <j...@ziepe.ca> wrote:
> On Tue, Aug 13, 2024 at 03:03:20PM -0600, Alex Williamson wrote: > > > How does the guest know to write a remappable vector format? How does > > the guest know the host interrupt architecture? For example why would > > an aarch64 guest program an MSI vector of 0xfee... if the host is x86? > > All excellent questions. > > Emulating real interrupt controllers in the VM is probably impossible > in every scenario. But certainly x86 emulating x86 and ARM emulating > ARM would be usefully achievable. > > hyperv did a neat thing where their remapping driver seems to make VMM > traps and looks kind of like the VMM gives it the platform specific > addr/data pair. > > It is a big ugly problem for sure, and we definately have painted > ourselves into a corner where the OS has no idea if IMS techniques > work properly or it is broken. :( :( > > But I think there may not be a terribly impossible path where at least > the guest could be offered a, say, virtio-irq in addition to the > existing platform controllers that would process IMS for it. > > > The idea of guest owning the physical MSI address space sounds great, > > but is it practical? > > In many cases yes, it is, but more importantly it is the only sane way > to support these IMS like techniques broadly since IMS is by > definition not generally trappable. > > > Is it something that would be accomplished while > > this device is still relevant? > > I don't know, I fear not. But it keeps coming up. Too many things > don't work right with the trapping approach, including this. > > > The Windows driver is just programming the MSI capability to use 16 > > vectors. We configure those vectors on the host at the time the > > capability is written. Whereas the Linux driver is only using a single > > vector and therefore writing the same MSI address and data at the > > locations noted in the trace, the Windows driver is writing different > > data values at different locations to make use of those vectors. This > > note is simply describing that we can't directly write the physical > > data value into the device, we need to determine which vector offset > > the guest is using and provide the same offset from the host data > > register value. > > I see, it seems to be assuming also that these extra interrupt sources > are generating the same MSI message as the main MSI, not something > else. That is more a SW quirk of Windows, I expect. I don't think > Linux would do that.. > > This is probably the only way to approach this, trap and emulate the > places in the device that program additional interrupt sources and do > a full MSI-like flow to set them up in the kernel. Your last sentence here seems to agree with this approach, but everything else suggests disapproval, so I don't know where you're going here. I have no specs for this device, nor any involvement from the device vendor, so the idea of creating a vfio-pci variant driver to setup an irq_domain and augment a device specific SET_IRQs ioctls not only sounds tremendously more complicated (host and VMM), it's simply not possible with the knowledge we have at hand. Making this device work in a VM is dead in the water if that's the bar to achieve. I observe that the device configures MSI vectors and then writes that same vector address/data elsewhere into the device. Whether the device can trigger those vectors based only on the MSI capability programming and a secondary source piggybacks on those vectors or if this is just a hack by Qualcomm to use an MSI capability to acquire some vectors which are exclusively used by the secondary hardware, I have no idea. Who can even say if this is just a cost saving measure that a PCI config space is slapped into a platform device and there's simply no hw/fw support to push the vector data into the hardware and the driver bridges the gap. The solution here is arguably fragile, we're relying on having a sufficiently unique MSI address that we can recognize writes with that value in order to both replace it with the host value and mark the location of the data register. If someone with some hardware insight wants to come along and provide a reference for static locations of these writes, I'd certainly welcome it. My sample size is one, which is why this is posted as an RFT. I do not believe that introducing a vfio device feature that disables virtualization of the MSI address/data _only_ at the vfio interface (not to a QEMU VM) provides some implicit support of this device behavior. These values are already available to a privileged user in the host and the same is available for an MSI-X use case by directly reading the MSI-X vector table. The only point of the vfio device feature is that we need a vehicle to expose the MSI phsyical address/data values through he vfio channel, without additional host privileges. The virtualized values are essentially unused by QEMU, so why not give QEMU a way to turn off the virtualization to expose the host values. Thanks, Alex