On Wed, Jul 06, 2016 at 10:54:58PM +0100, Andr? Przywara wrote:

On 06/07/16 22:06, Christoffer Dall wrote:

quoted

On Tue, Jul 05, 2016 at 12:22:55PM +0100, Andre Przywara wrote:

quoted

The ARM GICv3 ITS MSI controller requires a device ID to be able to
assign the proper interrupt vector. On real hardware, this ID is
sampled from the bus. To be able to emulate an ITS controller, extend
the KVM MSI interface to let userspace provide such a device ID. For
PCI devices, the device ID is simply the 16-bit bus-device-function
triplet, which should be easily available to the userland tool.

Also there is a new KVM capability which advertises whether the
current VM requires a device ID to be set along with the MSI data.
This flag is still reported as not available everywhere, later we will
enable it when ITS emulation is used.

Signed-off-by: Andre Przywara <andre.przywara@arm.com>
Reviewed-by: Eric Auger <redacted>
---
 Documentation/virtual/kvm/api.txt | 12 ++++++++++--
 include/uapi/linux/kvm.h          |  5 ++++-
 2 files changed, 14 insertions(+), 3 deletions(-)

diff --git a/Documentation/virtual/kvm/api.txt b/Documentation/virtual/kvm/api.txt
index 09efa9e..6551311 100644
--- a/Documentation/virtual/kvm/api.txt
+++ b/Documentation/virtual/kvm/api.txt

@@ -2175,10 +2175,18 @@ struct kvm_msi {
 	__u32 address_hi;
 	__u32 data;
 	__u32 flags;
-	__u8  pad[16];
+	__u32 devid;
+	__u8  pad[12];
 };
 
-No flags are defined so far. The corresponding field must be 0.
+flags: KVM_MSI_VALID_DEVID: devid contains a valid value
+devid: If KVM_MSI_VALID_DEVID is set, contains a unique device identifier
+       for the device that wrote the MSI message.
+       For PCI, this is usually a BFD identifier in the lower 16 bits.
+
+The per-VM KVM_CAP_MSI_DEVID capability advertises the need to provide
+the device ID. If this capability is not set, userland cannot rely on
+the kernel to allow the KVM_MSI_VALID_DEVID flag being set.

If KVM_CAP_MSI_DEVID is set, is it an error to provide a struct kvm_msi
without the KVM_MSI_VALID_DEVID flag set, or not necessarily?

In the moment we return an error when the bit is not set. But
theoretically a guest could have other (in-kernel emulated) MSI
controllers which don't require a device ID, though currently there is none.
The reason for both the KVM capability and the flag on the ioctl is for
mutual assurance that both userland and the kernel know about this ITS
specific extension.
This KVM_SIGNAL_MSI API is rather generic, so I don't want to
unnecessarily slam the door on other emulated ITS controllers.

That's a reasonable argument, that you could have multiple MSI
controllers with/without device ids.

Thanks,
-Christoffer