KVM call agenda for 2013-06-25

[Michael S. Tsirkin]

Please, send any topic that you are interested in covering.

Thanks, MST

[Qemu-devel] KVM call agenda for 2013-06-25

[Alexander Graf]

On 20.06.2013, at 14:47, Michael S. Tsirkin wrote:
> Please, send any topic that you are interested in covering.

VFIO with platform devices

Alex

[Qemu-devel] KVM call agenda for 2013-06-25

[Alexander Graf]

On 22.06.2013, at 03:01, Alexander Graf wrote:
> On 20.06.2013, at 14:47, Michael S. Tsirkin wrote:
> > Please, send any topic that you are interested in covering.
>
> VFIO with platform devices

A few notes ahead of the discussion:

-problem we are trying to solve is:
  1.  how the kernel can expose device tree based platform
      devices to user space
  2.  how QEMU exposes these devices in virtual machines

-general
  -this is not just a QEMU/KVM issue, we have other user space 
   drivers that should use the same framework (e.g. USDPAA)
  -acpi/dsdt based device discovery is on the horizon, but this 
   should be treated separately
  -we want the same linux drivers to be used in host and
   guest...don't want 'paravirt' drivers in guest

-3 aspects to this

1. kernel side
     -a device must be able to be unbound from the host
      and bound to vfio
     -vfio can determine the mappable register regions and
      interrupts from the associated device tree node and
      expose these to user space...details TBD
     -keep it simple: expose register regions and interrupts

2. user-space/QEMU - device tree

-there is potentially a large variety of devices, the
      simplest are represented by a simple self-contained
      node, others get complicated and could have phandles to
      other nodes (e.g. etsec has phandle to phy node).  Some
      devices may need device tree tweaks.

-think what we need is custom handling in QEMU for
      each type of device.  QEMU uses what the kernel 
      exposed via vfio to pass the device through and
     For example to pass through
      a SATA controller something like:

-device platfom-vfio-sata,dev=/soc/sata@221000,bus=platform.0

-the device specific support in QEMU knows how to
      create a device tree node for the given device type

-it is up to user space to deal with device inter-
      relationships if they exist

-SEC engine job ring example-- what we want ideally is a 
      Linux driver that can operate on plain job rings
      so we don't need the complexity of dealing with the
      main SEC node:

[cut]
        crypto@300000 {
           compatible = "fsl,sec-v4.0";
           #address-cells = <0x1>;
           #size-cells = <0x1>;
           reg = <0x300000 0x10000>;
           ranges = <0x0 0x300000 0x10000>;
           interrupts = <0x5c 0x2 0x0 0x0>;

jr@1000 {
              compatible = "fsl,sec-v4.0-job-ring";
              reg = <0x1000 0x1000>;
              interrupts = <0x58 0x2 0x0 0x0>;
           };

jr@2000 {
              compatible = "fsl,sec-v4.0-job-ring";
              reg = <0x2000 0x1000>;
              interrupts = <0x59 0x2 0x0 0x0>;
           };
        [cut]

3. user-space/QEMU - dynamic interrupt allocation

-currently devices and interrupts on our system/platform
      bus are statically allocated.  We need a way to
      dynamically get devices added

-simple example would be adding a virtual UART via -device

-it would be nice to solve this generally for both
      ARM and Power