Let OCI runtime create netns when userns is used

As explained in the comments, this patch lets the OCI runtime create the
netns when userns are in use. This is needed because the netns needs to
be owned by the userns (otherwise can't modify the IP, etc.).

Before this patch, we are creating the netns and then starting the pod
sandbox asking to join this netns. This can't never work with userns, as
the userns needs to be created first for the netns ownership to be
correct.

One option would be to also create the userns in containerd, then create
the netns. But this is painful (needs tricks with the go runtime,
special care to write the mapping, etc.).

So, we just let the OCI runtime create the userns and netns, that
creates them with the proper ownership.

As requested by Mike Brown, the current code when userns is not used is
left unchanged. We can unify the cases (with and without userns) in a
future release.

Signed-off-by: Rodrigo Campos <rodrigoca@microsoft.com>

pkg/cri/server/sandbox_run.go

@@ -23,6 +23,7 @@ import (
 	"fmt"
 	"math"
 	"path/filepath"
+	goruntime "runtime"
 	"strings"
 	"time"
 
@@ -244,8 +245,27 @@ func (c *criService) RunPodSandbox(ctx context.Context, r *runtime.RunPodSandbox
 		return nil, fmt.Errorf("failed to get sandbox container info: %w", err)
 	}
 
+	userNsEnabled := false
+	if goruntime.GOOS != "windows" {
+		usernsOpts := config.GetLinux().GetSecurityContext().GetNamespaceOptions().GetUsernsOptions()
+		if usernsOpts != nil && usernsOpts.GetMode() == runtime.NamespaceMode_POD {
+			userNsEnabled = true
+		}
+	}
+
 	// Setup the network namespace if host networking wasn't requested.
-	if !hostNetwork(config) {
+	if !hostNetwork(config) && !userNsEnabled {
+		// XXX: We do c&p of this code later for the podNetwork && userNsEnabled case too.
+		// We can't move this to a function, as the defer calls need to be executed if other
+		// errors are returned in this function. So, we would need more refactors to move
+		// this code to a function and the idea was to not change the current code for
+		// !userNsEnabled case, therefore doing it would defeat the purpose.
+		//
+		// The difference between the cases is the use of netns.NewNetNS() vs
+		// netns.NewNetNSFromPID() and we verify the task is still running in the other case.
+		//
+		// To simplify this, in the future, we should just remove this case (podNetwork &&
+		// !userNsEnabled) and just keep the other case (podNetwork && userNsEnabled).
 		netStart := time.Now()
 
 		// If it is not in host network namespace then create a namespace and set the sandbox
@@ -353,6 +373,88 @@ func (c *criService) RunPodSandbox(ctx context.Context, r *runtime.RunPodSandbox
 		return nil, fmt.Errorf("failed to wait for sandbox container task: %w", err)
 	}
 
+	if !hostNetwork(config) && userNsEnabled {
+		// If userns is enabled, then the netns was created by the OCI runtime
+		// when creating "task". The OCI runtime needs to create the netns
+		// because, if userns is in use, the netns needs to be owned by the
+		// userns. So, let the OCI runtime just handle this for us.
+		// If the netns is not owned by the userns several problems will happen.
+		// For instance, the container will lack permission (even if
+		// capabilities are present) to modify the netns or, even worse, the OCI
+		// runtime will fail to mount sysfs:
+		// 	https://github.com/torvalds/linux/commit/7dc5dbc879bd0779924b5132a48b731a0bc04a1e#diff-4839664cd0c8eab716e064323c7cd71fR1164
+		netStart := time.Now()
+
+		// If it is not in host network namespace then create a namespace and set the sandbox
+		// handle. NetNSPath in sandbox metadata and NetNS is non empty only for non host network
+		// namespaces. If the pod is in host network namespace then both are empty and should not
+		// be used.
+		var netnsMountDir = "/var/run/netns"
+		if c.config.NetNSMountsUnderStateDir {
+			netnsMountDir = filepath.Join(c.config.StateDir, "netns")
+		}
+		sandbox.NetNS, err = netns.NewNetNSFromPID(netnsMountDir, task.Pid())
+		if err != nil {
+			return nil, fmt.Errorf("failed to create network namespace for sandbox %q: %w", id, err)
+		}
+
+		// Verify task is still in created state.
+		if st, err := task.Status(ctx); err != nil || st.Status != containerd.Created {
+			return nil, fmt.Errorf("failed to create pod sandbox %q: err is %v - status is %q and is expected %q", id, err, st.Status, containerd.Created)
+		}
+		sandbox.NetNSPath = sandbox.NetNS.GetPath()
+
+		defer func() {
+			// Remove the network namespace only if all the resource cleanup is done.
+			if retErr != nil && cleanupErr == nil {
+				if cleanupErr = sandbox.NetNS.Remove(); cleanupErr != nil {
+					log.G(ctx).WithError(cleanupErr).Errorf("Failed to remove network namespace %s for sandbox %q", sandbox.NetNSPath, id)
+					return
+				}
+				sandbox.NetNSPath = ""
+			}
+		}()
+
+		// Update network namespace in the container's spec
+		c.updateNetNamespacePath(spec, sandbox.NetNSPath)
+
+		if err := container.Update(ctx,
+			// Update spec of the container
+			containerd.UpdateContainerOpts(containerd.WithSpec(spec)),
+			// Update sandbox metadata to include NetNS info
+			containerd.UpdateContainerOpts(containerd.WithContainerExtension(sandboxMetadataExtension, &sandbox.Metadata))); err != nil {
+			return nil, fmt.Errorf("failed to update the network namespace for the sandbox container %q: %w", id, err)
+		}
+
+		// Define this defer to teardownPodNetwork prior to the setupPodNetwork function call.
+		// This is because in setupPodNetwork the resource is allocated even if it returns error, unlike other resource creation functions.
+		defer func() {
+			// Teardown the network only if all the resource cleanup is done.
+			if retErr != nil && cleanupErr == nil {
+				deferCtx, deferCancel := ctrdutil.DeferContext()
+				defer deferCancel()
+				// Teardown network if an error is returned.
+				if cleanupErr = c.teardownPodNetwork(deferCtx, sandbox); cleanupErr != nil {
+					log.G(ctx).WithError(cleanupErr).Errorf("Failed to destroy network for sandbox %q", id)
+				}
+			}
+		}()
+
+		// Setup network for sandbox.
+		// Certain VM based solutions like clear containers (Issue containerd/cri-containerd#524)
+		// rely on the assumption that CRI shim will not be querying the network namespace to check the
+		// network states such as IP.
+		// In future runtime implementation should avoid relying on CRI shim implementation details.
+		// In this case however caching the IP will add a subtle performance enhancement by avoiding
+		// calls to network namespace of the pod to query the IP of the veth interface on every
+		// SandboxStatus request.
+		if err := c.setupPodNetwork(ctx, &sandbox); err != nil {
+			return nil, fmt.Errorf("failed to setup network for sandbox %q: %w", id, err)
+		}
+
+		sandboxCreateNetworkTimer.UpdateSince(netStart)
+	}
+
 	if c.nri.isEnabled() {
 		err = c.nri.runPodSandbox(ctx, &sandbox)
 		if err != nil {