fix: return last known pid when scaled down

When the pod is scaled down we need to return the last known pid, else
runc will be called, which will error as the checkpointed container
cannot be found. This fixes the shim being restarted during a containerd
restart.

runc/task/service_zeropod.go

@@ -13,6 +13,7 @@ import (
 
 	"github.com/containerd/cgroups"
 	taskAPI "github.com/containerd/containerd/api/runtime/task/v2"
+	"github.com/containerd/containerd/api/types/task"
 	"github.com/containerd/containerd/errdefs"
 	"github.com/containerd/containerd/pkg/cri/annotations"
 	"github.com/containerd/containerd/pkg/oom"
@@ -198,6 +199,18 @@ func (w *wrapper) Exec(ctx context.Context, r *taskAPI.ExecProcessRequest) (*emp
 	return w.service.Exec(ctx, r)
 }
 
+func (w *wrapper) Pids(ctx context.Context, r *taskAPI.PidsRequest) (*taskAPI.PidsResponse, error) {
+	zeropodContainer, ok := w.getZeropodContainer(r.ID)
+	if ok && zeropodContainer.ScaledDown() {
+		log.G(ctx).Debugf("pids of scaled down container, returning last known pid: %v", zeropodContainer.Pid())
+		return &taskAPI.PidsResponse{
+			Processes: []*task.ProcessInfo{{Pid: uint32(zeropodContainer.Pid())}},
+		}, nil
+	}
+
+	return w.service.Pids(ctx, r)
+}
+
 func (w *wrapper) Delete(ctx context.Context, r *taskAPI.DeleteRequest) (*taskAPI.DeleteResponse, error) {
 	zeropodContainer, ok := w.getZeropodContainer(r.ID)
 	if !ok {