Hi Christoffer,

On 06/07/15 10:30, Christoffer Dall wrote:

On Mon, Jul 06, 2015 at 09:30:20AM +0100, Andre Przywara wrote:

quoted

Hi Pavel,

On 06/07/15 07:42, Pavel Fedin wrote:

quoted

 Hello!

quoted

I like this approach, but it runs into problems:
As you read above the current documentation says that the flags field
must be zero and the current KVM_SET_GSI_ROUTING handler bails out if it
isn't. So userland would need to know whether it's safe to set that
field.

 This problem does not exist because:
a) Older platforms do not need this flag, so they expect to get zero.
b) ARM64 + GICv3 platform cannot work without this flag.

 This is perfectly OK combination IMO. Userland just knows, whether it needs to supply device ID or
not. For example, my modified qemu now has kvm_msi_flags global variable which defaults to 0. ITS
code, then, if activated, changes it to KVM_MSI_VALID_DEVID, and qemu starts supplying device IDs to
the related calls.

Well, I had this solution before in kvmtool: If ARM && ITS then set the
flag. But I wasn't really happy with this, as the IRQ routing, setup and
injection code is rather architecture agnostic (implementing the generic
KVM interface), so spraying in some architecture hacks sounded not very
elegant.
Also as the flag describes a rather generic feature (provide an unique
device ID), I'd rather avoid to make this an ARM hack.

That being said this is not a show stopper for me, so if the others are
happy with this, I will go down your road.

There must be some way for userspace to discover if it's valid to set
the flag or not; either through a well-defined error-code probing
mechanism for KVM_SET_GSI_ROUTING or through advertising a capability.

Right, makes sense, I was wondering about this requirement earlier, but
couldn't find really good prior art in the code (KVM_GET_VCPU_EVENTS
seems to be bad example).
So I think we get along with a new VM specific capability like
KVM_CAP_MSIS_REQUIRE_DEVID. This isn't strictly a "capability" (as it's
more a requirement), but I guess it fits here anyway. It has to be
per-VM, as a GICv2M guest does not need it, but an ITS guest does.
We can use this very flag for both the KVM_SIGNAL_MSI and the
KVM_SET_GSI_ROUTING ioctl, so interface churn is kept minimal.

Does that make sense?

Actually I have implemented this already last week, I will send it out
along with a v2 of the ITS emulation later this week.

Cheers,
Andre.