Re: [Qemu-devel] [RFC patch 0/6] vfio based pci pass-through for qemu/KVM on s390

[Frank Blaschka]

On Fri, Sep 26, 2014 at 01:59:40PM -0600, Alex Williamson wrote:
> On Fri, 2014-09-26 at 08:45 +0200, Frank Blaschka wrote:
> > On Wed, Sep 24, 2014 at 10:05:57AM -0600, Alex Williamson wrote:
> > > On Wed, 2014-09-24 at 10:47 +0200, Frank Blaschka wrote:
> > > > On Mon, Sep 22, 2014 at 02:47:31PM -0600, Alex Williamson wrote:
> > > > > On Fri, 2014-09-19 at 13:54 +0200, address@hidden wrote:
> > > > > > This set of patches implements a vfio based solution for pci
> > > > > > pass-through on the s390 platform. The kernel stuff is pretty
> > > > > > much straight forward, but qemu needs more work.
> > > > > > 
> > > > > > Most interesting patch is:
> > > > > >   vfio: make vfio run on s390 platform
> > > > > > 
> > > > > > I hope Alex & Alex can give me some guidance how to do the changes
> > > > > > in an appropriate way. After creating a separate iommmu address 
> > > > > > space
> > > > > > for each attached PCI device I can successfully run the vfio type1
> > > > > > iommu. So If we could extend type1 not registering all guest memory
> > > > > > (see patch) I think we do not need a special vfio iommu for s390
> > > > > > for the moment.
> > > > > > 
> > > > > > The patches implement the base pass-through support. s390 specific
> > > > > > virtualization functions are currently not included. This would
> > > > > > be a second step after the base support is done.
> > > > > > 
> > > > > > kernel patches apply to linux-kvm-next
> > > > > > 
> > > > > > KVM: s390: Enable PCI instructions
> > > > > > iommu: add iommu for s390 platform
> > > > > > vfio: make vfio build on s390
> > > > > > 
> > > > > > qemu patches apply to qemu-master
> > > > > > 
> > > > > > s390: Add PCI bus support
> > > > > > s390: implement pci instruction
> > > > > > vfio: make vfio run on s390 platform
> > > > > > 
> > > > > > Thx for feedback and review comments
> > > > > 
> > > > > Sending patches as attachments makes it difficult to comment inline.
> > > > >
> > > > Sorry, don't understand this. I sent every patch as separate email so
> > > > you can comment directly on the patch. What do you prefer?
> > > 
> > > The patches in each email are showing up as attachments in my mail
> > > client.  Is it just me?
> > > 
> > > > > 2/6
> > > > >  - careful of the namespace as you're changing functions from static 
> > > > > and
> > > > > exporting them
> > > > >  - doesn't seem like functions need to be exported, just non-static to
> > > > > call from s390-iommu.c
> > > > > 
> > > > Ok, will change this.
> > > > 
> > > > > 6/6
> > > > >  - We shouldn't need to globally disable mmap, each VFIO region 
> > > > > reports
> > > > > whether it supports mmap and vfio-pci on s390 should indicate mmap is
> > > > > not supported on the platform.
> > > > Yes, this is even better to let the kernel announce a BAR can not be
> > > > mmap'ed. Checking the kernel code I realized the BARs are valid for
> > > > mmap'ing but the s390 platform does simply not allow this. So I feal we
> > > > have to introduce a platform switch in kernel. How about this ...
> > > > 
> > > > --- a/drivers/vfio/pci/vfio_pci.c
> > > > +++ b/drivers/vfio/pci/vfio_pci.c
> > > > @@ -377,9 +377,11 @@ static long vfio_pci_ioctl(void *device_
> > > > 
> > > >                         info.flags = VFIO_REGION_INFO_FLAG_READ |
> > > >                                      VFIO_REGION_INFO_FLAG_WRITE;
> > > > +#ifndef CONFIG_S390
> > > >                         if (pci_resource_flags(pdev, info.index) &
> > > >                             IORESOURCE_MEM && info.size >= PAGE_SIZE)
> > > >                                 info.flags |= 
> > > > VFIO_REGION_INFO_FLAG_MMAP;
> > > > +#endif
> > > >                         break;
> > > >                 case VFIO_PCI_ROM_REGION_INDEX:
> > > >                 {
> > > 
> > > Maybe pull it out into a function.  Also, is there some capability or
> > > feature we can test rather than just the architecture?  I'd prefer it to
> > > be excluded because of a platform feature that prevents it rather than
> > > the overall architecture itself.
> > >
> > 
> > Ok, understand this. There is no capability of feature so I will go with
> > the function.
> >  
> > > > >  - INTx should be done the same way, the interrupt index for INTx 
> > > > > should
> > > > > report 0 count.  The current code likely doesn't handle this, but it
> > > > > should be easy to fix.
> > > > The current code is fine. Problem is the card reports an interrupt index
> > > > (PCI_INTERRUPT_PIN) but again the platform does not support INTx at all.
> > > > So we need a platform switch as well. 
> > > 
> > > Yep, let's try to do something consistent with the MMAP testing.
> > >
> > 
> > Do you mean let the kernel announce this also?
> 
> Yes, the kernel reports a count of 0 in vfio_irq_info when the interrupt
> type is not supported.  We do this for MSI/X already, but it's assumed
> that INTx is always present since it's part of what most platforms would
> consider the minimal feature set.
> 
> > > > >  - s390_msix_notify() vs msix_notify() should be abstracted somewhere
> > > > 
> > > > Platform does not have have an apic so there is nothing we could emulate
> > > > in qemu to make the existing msix_notify() work.
> > > > 
> > > > > else.  How would an emulated PCI device with MSI-X support work?
> > > > >  - same for add_msi_route
> > > > Same here, we have to setup an adapter route due to the fact MSIX
> > > > notifications are delivered as adapter/thin IRQs on the platform.
> > > > 
> > > > Any suggestion or idea how a better abstraction could look like?
> > > > 
> > > > With all the platform constraints I was not able to find a suitable
> > > > emulated device. Remember s390:
> > > > - does not support IO BARs
> > > > - does not support INTx only MSIX
> > > 
> > > What about MSI (non-X)?
> > 
> > In theory MSI should work also but I have not seen in reality.
> > 
> > > 
> > > > - in reality currently there is only a PCI network card available
> > > 
> > > On the physical hardware?
> > > 
> > 
> > yes
> > 
> > > > - platform does not support fancy I/O like usb or audio :-)
> > > >   So we don't even have kernel (host and guest) support for this
> > > >   kind of devices.
> > > 
> > > Does that mean you couldn't?  What about virtio-net-pci with MSI-X
> > > interrupts or emulated xhci with MSI-X interrupts, couldn't those be
> > > supported if s390 MSI-X were properly integrated into the QEMU MSI-X
> > > API?  vfio-pci isn't the right level to be switching between the
> > > standard API and the s390 API.
> > > 
> > 
> > Yes, I also think vfio might not be the best place to switch API. Will try
> > to move s390 specifics to MSI-X level. 
> > 
> > > > >  - We can probably come up with a better way to determine which 
> > > > > address
> > > > > space to connect to the memory listener.
> > > > Any suggestion or idea for that?
> > > 
> > > I imagine you can tell by the address space of the device whether it
> > > lives behind an emulated IOMMU or not and therefore pick the closest
> > > address space for the notifier, the IOMMU or the system.  Thanks,
> > >
> > 
> > I do not undertand this in detail, can you elaborate a little bit more on 
> > this?
> > Or maybe provide a code snip?
> 
> Well, I'm mostly making things up, but my assumption is that the device
> appears behind an IOMMU in the guest and by walking through address
> spaces from the device, we should be able to figure that out and avoid
> using a platform #ifdef.  IOW, it's not s390 that makes us need to use a
> different address space, it's the guest topology of having an emulated
> IOMMU for the device, and that's what we should be keying on rather than
> the arch.  Thanks,
>

Do you think this would be sufficient?

@@ -3689,8 +3701,13 @@ static int vfio_connect_container(VFIOGr
         container->iommu_data.type1.listener = vfio_memory_listener;
         container->iommu_data.release = vfio_listener_release;

-        memory_listener_register(&container->iommu_data.type1.listener,
-                                 &address_space_memory);
+        if (memory_region_is_iommu(as->root)) {
+            memory_listener_register(&container->iommu_data.type1.listener,
+                                     container->space->as);
+        } else {
+            memory_listener_register(&container->iommu_data.type1.listener,
+                                     &address_space_memory);
+        }

         if (container->iommu_data.type1.error) {
             ret = container->iommu_data.type1.error;

If not what else has to be checked? What are the indications to add the memory
listener to container address space or to address_space_memory?
Thx for your help.
 
> Alex
> 
>