gpu: Add ColdPlug of VFIO devices with devManager

If we have a VFIO device and cold-plug is enabled
we mark each device as ColdPlug=true and let the VFIO
module do the attaching.

Signed-off-by: Zvonko Kaiser <zkaiser@nvidia.com>

src/runtime/virtcontainers/hypervisor.go

@@ -509,9 +509,9 @@ type HypervisorConfig struct {
 	// The PCIe Root Port device is used to hot-plug the PCIe device
 	PCIeRootPort uint32
 
-	// RawDevics are used to get PCIe device info early before the sandbox
+	// VFIODevics are used to get PCIe device info early before the sandbox
 	// is started to make better PCIe topology decisions
-	RawDevices []config.DeviceInfo
+	VFIODevices []config.DeviceInfo
 
 	// ColdPlugVFIO is used to indicate if devices need to be coldplugged on the
 	// root port, switch or no port

src/runtime/virtcontainers/kata_agent.go

@@ -1177,8 +1177,7 @@ func (k *kataAgent) appendDevices(deviceList []*grpc.Device, c *Container) []*gr
 		case config.VhostUserBlk:
 			kataDevice = k.appendVhostUserBlkDevice(dev, device, c)
 		case config.DeviceVFIO:
-			k.Logger().Infof("### ColdPlugging container is not adding any VFIO devices")
-			//kataDevice = k.appendVfioDevice(dev, device, c)
+			kataDevice = k.appendVfioDevice(dev, device, c)
 		}
 
 		if kataDevice == nil {

src/runtime/virtcontainers/qemu.go

@@ -83,6 +83,7 @@ type QemuState struct {
 	VirtiofsDaemonPid    int
 	PCIeRootPort         int
 	HotplugVFIOOnRootBus bool
+	ColdPlugVFIO         hv.PCIePort
 }
 
 // qemu is an Hypervisor interface implementation for the Linux qemu hypervisor.
@@ -282,6 +283,7 @@ func (q *qemu) setup(ctx context.Context, id string, hypervisorConfig *Hyperviso
 		q.Logger().Debug("Creating UUID")
 		q.state.UUID = uuid.Generate().String()
 
+		q.state.ColdPlugVFIO = q.config.ColdPlugVFIO
 		q.state.HotplugVFIOOnRootBus = q.config.HotplugVFIOOnRootBus
 		q.state.PCIeRootPort = int(q.config.PCIeRootPort)
 
@@ -708,9 +710,25 @@ func (q *qemu) CreateVM(ctx context.Context, id string, network Network, hypervi
 		qemuConfig.Devices = q.arch.appendPCIeRootPortDevice(qemuConfig.Devices, hypervisorConfig.PCIeRootPort, memSize32bit, memSize64bit)
 	}
 
+	q.virtiofsDaemon, err = q.createVirtiofsDaemon(hypervisorConfig.SharedPath)
+
+	// If we have a VFIO device we need to update the firmware configuration
+	// if executed in a trusted execution environment.
+	if hypervisorConfig.ConfidentialGuest {
+		// At the sandbox level we alreaady checked that we have a
+		// VFIO device, pass-through of a PCIe device needs allocated
+		// mmemory in the firmware otherwise BARs cannot be mapped
+		if len(hypervisorConfig.VFIODevices) > 0 {
+			fwCfg := govmmQemu.FwCfg{
+				Name: "opt/ovmf/X-PciMmio64Mb",
+				Str:  "262144",
+			}
+			qemuConfig.FwCfg = append(qemuConfig.FwCfg, fwCfg)
+		}
+	}
+
 	q.qemuConfig = qemuConfig
 
-	q.virtiofsDaemon, err = q.createVirtiofsDaemon(hypervisorConfig.SharedPath)
 	return err
 }
 

src/runtime/virtcontainers/sandbox.go

@@ -32,6 +32,7 @@ import (
 	"github.com/kata-containers/kata-containers/src/runtime/pkg/device/config"
 	"github.com/kata-containers/kata-containers/src/runtime/pkg/device/drivers"
 	deviceManager "github.com/kata-containers/kata-containers/src/runtime/pkg/device/manager"
+	hv "github.com/kata-containers/kata-containers/src/runtime/pkg/hypervisors"
 	"github.com/kata-containers/kata-containers/src/runtime/pkg/katautils/katatrace"
 	resCtrl "github.com/kata-containers/kata-containers/src/runtime/pkg/resourcecontrol"
 	exp "github.com/kata-containers/kata-containers/src/runtime/virtcontainers/experimental"
@@ -619,12 +620,30 @@ func newSandbox(ctx context.Context, sandboxConfig SandboxConfig, factory Factor
 	if err := validateHypervisorConfig(&sandboxConfig.HypervisorConfig); err != nil {
 		return nil, err
 	}
-	// Aggregate all the container devices and update the HV config
-	var devices []config.DeviceInfo
-	for _, ct := range sandboxConfig.Containers {
-		devices = append(devices, ct.DeviceInfos...)
+
+	// If we have a confidential guest we need to cold-plug the PCIe VFIO devices
+	// until we have TDISP/IDE PCIe support.
+	coldPlugVFIO := (sandboxConfig.HypervisorConfig.ColdPlugVFIO != hv.NoPort)
+	var devs []config.DeviceInfo
+	for cnt, containers := range sandboxConfig.Containers {
+		for dev, device := range containers.DeviceInfos {
+			if coldPlugVFIO && deviceManager.IsVFIO(device.ContainerPath) {
+				device.ColdPlug = true
+				devs = append(devs, device)
+				// We need to remove the devices marked for cold-plug
+				// otherwise at the container level the kata-agent
+				// will try to hot-plug them.
+				infos := sandboxConfig.Containers[cnt].DeviceInfos
+				infos = append(infos[:dev], infos[dev+1:]...)
+				sandboxConfig.Containers[cnt].DeviceInfos = infos
+			}
+		}
 	}
-	sandboxConfig.HypervisorConfig.RawDevices = devices
+	// If we have a confidential guest, we need to add a specific
+	// firmware configuration to the hypervisor. We cannot do it here at
+	// the sandbox level we need to do that at the hypervisor level, capturing
+	// the devices here and processing in CreateVM().
+	sandboxConfig.HypervisorConfig.VFIODevices = devs
 
 	// store doesn't require hypervisor to be stored immediately
 	if err = s.hypervisor.CreateVM(ctx, s.id, s.network, &sandboxConfig.HypervisorConfig); err != nil {
@@ -635,6 +654,17 @@ func newSandbox(ctx context.Context, sandboxConfig SandboxConfig, factory Factor
 		return nil, err
 	}
 
+	if !coldPlugVFIO {
+		return s, nil
+	}
+
+	for _, dev := range devs {
+		_, err := s.AddDevice(ctx, dev)
+		if err != nil {
+			s.Logger().WithError(err).Debug("Cannot cold-plug add device")
+			return nil, err
+		}
+	}
 	return s, nil
 }