From a88af321f7c591404b0316320c46591fa8250992 Mon Sep 17 00:00:00 2001
From: Archana Shinde <archana.m.shinde@intel.com>
Date: Wed, 3 Jun 2020 08:27:36 -0700
Subject: [PATCH] device: Do not allow container access to the nvdimm rootfs

With this change, a container is not longer given access to
the underlying nvdimm root partition.
This is done by explicitly adding the nvdimm root partition
to the device cgroup of the container.

Fixes #791

Signed-off-by: Archana Shinde <archana.m.shinde@intel.com>
---
 device.go      | 26 ++++++++++++++++++++++++++
 device_test.go | 23 +++++++++++++++++++++++
 grpc.go        |  3 +++
 3 files changed, 52 insertions(+)

diff --git a/device.go b/device.go
index 3b0307f975..ec1907e0b2 100644
--- a/device.go
+++ b/device.go
@@ -34,6 +34,7 @@ const (
 	driverNvdimmType    = "nvdimm"
 	driverEphemeralType = "ephemeral"
 	driverLocalType     = "local"
+	vmRootfs            = "/"
 )
 
 const (
@@ -472,3 +473,28 @@ func addDevice(ctx context.Context, device *pb.Device, spec *pb.Spec, s *sandbox
 
 	return devHandler(ctx, *device, spec, s)
 }
+
+// updateDeviceCgroupForGuestRootfs updates the device cgroup for container
+// to not allow access to the nvdim root partition. This prevents the container
+// from being able to access the VM rootfs.
+func updateDeviceCgroupForGuestRootfs(spec *pb.Spec) {
+	var devStat unix.Stat_t
+
+	err := unix.Stat(vmRootfs, &devStat)
+	if err != nil {
+		return
+	}
+
+	devMajor := int64(unix.Major(devStat.Dev))
+	devMinor := int64(unix.Minor(devStat.Dev))
+
+	nvdimmCg := pb.LinuxDeviceCgroup{
+		Allow:  false,
+		Major:  devMajor,
+		Minor:  devMinor,
+		Type:   "b",
+		Access: "rwm",
+	}
+
+	spec.Linux.Resources.Devices = append(spec.Linux.Resources.Devices, nvdimmCg)
+}
diff --git a/device_test.go b/device_test.go
index 8c8f201930..72a9dac1d9 100644
--- a/device_test.go
+++ b/device_test.go
@@ -18,6 +18,7 @@ import (
 
 	pb "github.com/kata-containers/agent/protocols/grpc"
 	"github.com/stretchr/testify/assert"
+	"golang.org/x/sys/unix"
 )
 
 var (
@@ -728,3 +729,25 @@ func TestGetDeviceName(t *testing.T) {
 	assert.Nil(err)
 	assert.Equal(name, path.Join(devRootPath, devName))
 }
+
+func TestUpdateDeviceCgroupForGuestRootfs(t *testing.T) {
+	skipUnlessRoot(t)
+	assert := assert.New(t)
+
+	spec := &pb.Spec{}
+
+	spec.Linux = &pb.Linux{}
+	spec.Linux.Resources = &pb.LinuxResources{}
+
+	updateDeviceCgroupForGuestRootfs(spec)
+	assert.Equal(1, len(spec.Linux.Resources.Devices))
+
+	var devStat unix.Stat_t
+	err := unix.Stat(vmRootfs, &devStat)
+	if err != nil {
+		return
+	}
+
+	assert.Equal(spec.Linux.Resources.Devices[0].Major, int64(unix.Major(devStat.Dev)))
+	assert.Equal(spec.Linux.Resources.Devices[0].Minor, int64(unix.Minor(devStat.Dev)))
+}
diff --git a/grpc.go b/grpc.go
index 8a0dffddc6..829881f12f 100644
--- a/grpc.go
+++ b/grpc.go
@@ -654,6 +654,9 @@ func (a *agentGRPC) CreateContainer(ctx context.Context, req *pb.CreateContainer
 		}
 	}()
 
+	// Add the nvdimm root partition to the device cgroup to prevent access
+	updateDeviceCgroupForGuestRootfs(req.OCI)
+
 	// Convert the spec to an actual OCI specification structure.
 	ociSpec, err := pb.GRPCtoOCI(req.OCI)
 	if err != nil {