Add container restart policy to Libpod & Podman

This initial version does not support restart count, but it works
as advertised otherwise.

Signed-off-by: Matthew Heon <matthew.heon@pm.me>

4 files changed, 127 insertions, 1 deletions

diff --git a/libpod/container.go b/libpod/container.go
index 570110ca1..4236f2456 100644
--- a/libpod/container.go
+++ b/libpod/container.go
@@ -181,7 +181,10 @@ type ContainerState struct {
 	BindMounts map[string]string `json:"bindMounts,omitempty"`
 	// StoppedByUser indicates whether the container was stopped by an
 	// explicit call to the Stop() API.
-	StoppedByUser bool
+	StoppedByUser bool `json:"stoppedByUser,omitempty"`
+	// RestartPolicyMatch indicates whether the conditions for restart
+	// policy have been met.
+	RestartPolicyMatch bool `json:"restartPolicyMatch,omitempty"`
 
 	// ExtensionStageHooks holds hooks which will be executed by libpod
 	// and not delegated to the OCI runtime.
@@ -349,6 +352,17 @@ type ContainerConfig struct {
 	LogPath string `json:"logPath"`
 	// File containing the conmon PID
 	ConmonPidFile string `json:"conmonPidFile,omitempty"`
+	// RestartPolicy indicates what action the container will take upon
+	// exiting naturally.
+	// Allowed options are "no" (take no action), "on-failure" (restart on
+	// non-zero exit code, up an a maximum of RestartRetries times),
+	// and "always" (always restart the container on any exit code).
+	// The empty string is treated as the default ("no")
+	RestartPolicy string `json:"restart_policy,omitempty"`
+	// RestartRetries indicates the number of attempts that will be made to
+	// restart the container. Used only if RestartPolicy is set to
+	// "on-failure".
+	RestartRetries uint `json:"restart_retries,omitempty"`
 	// TODO log options for log drivers
 
 	PostConfigureNetNS bool `json:"postConfigureNetNS"`
@@ -732,6 +746,17 @@ func (c *Container) LogPath() string {
 	return c.config.LogPath
 }
 
+// RestartPolicy returns the container's restart policy.
+func (c *Container) RestartPolicy() string {
+	return c.config.RestartPolicy
+}
+
+// RestartRetries returns the number of retries that will be attempted when
+// using the "on-failure" restart policy
+func (c *Container) RestartRetries() uint {
+	return c.config.RestartRetries
+}
+
 // RuntimeName returns the name of the runtime
 func (c *Container) RuntimeName() string {
 	return c.runtime.ociRuntime.name
diff --git a/libpod/container_api.go b/libpod/container_api.go
index 5bfd869b3..46c913e99 100644
--- a/libpod/container_api.go
+++ b/libpod/container_api.go
@@ -583,6 +583,7 @@ func (c *Container) Cleanup(ctx context.Context) error {
 	if !c.batched {
 		c.lock.Lock()
 		defer c.lock.Unlock()
+
 		if err := c.syncContainer(); err != nil {
 			return err
 		}
@@ -593,6 +594,24 @@ func (c *Container) Cleanup(ctx context.Context) error {
 		return errors.Wrapf(ErrCtrStateInvalid, "container %s is running or paused, refusing to clean up", c.ID())
 	}
 
+	// If we have a restart policy match when updating the state, we need to
+	// restart the container.
+	// However, perform a full validation of restart policy first.
+	if c.state.RestartPolicyMatch {
+		// if restart policy is on-error and exit code is 0, we do
+		// nothing.
+		// TODO: if restart retries is set, handle that here.
+		if c.config.RestartRetries != 0 {
+			return errors.Wrapf(ErrNotImplemented, "restart retries not yet implemented")
+		}
+		if (c.config.RestartPolicy == "on-error" && c.state.ExitCode == 0) || c.config.RestartPolicy == "always" {
+			// The container stopped. We need to restart it.
+			return c.handleRestartPolicy(ctx)
+		}
+	}
+
+	// If we aren't hitting restart policy, we perform a normal cleanup
+
 	// Check if we have active exec sessions
 	if len(c.state.ExecSessions) != 0 {
 		return errors.Wrapf(ErrCtrStateInvalid, "container %s has active exec sessions, refusing to clean up", c.ID())
diff --git a/libpod/container_internal.go b/libpod/container_internal.go
index 0b9c5a96a..f72d86c84 100644
--- a/libpod/container_internal.go
+++ b/libpod/container_internal.go
@@ -210,6 +210,43 @@ func (c *Container) handleExitFile(exitFile string, fi os.FileInfo) error {
 	return nil
 }
 
+// Handle container restart policy.
+// This is called when a container has exited, and was not explicitly stopped by
+// an API call to stop the container or pod it is in.
+func (c *Container) handleRestartPolicy(ctx context.Context) (err error) {
+	logrus.Debugf("Restarting container %s due to restart policy %s", c.ID(), c.config.RestartPolicy)
+
+	// Need to check if dependencies are alive.
+	if err = c.checkDependenciesAndHandleError(ctx); err != nil {
+		return err
+	}
+
+	defer func() {
+		if err != nil {
+			if err2 := c.cleanup(ctx); err2 != nil {
+				logrus.Errorf("error cleaning up container %s: %v", c.ID(), err2)
+			}
+		}
+	}()
+	if err := c.prepare(); err != nil {
+		return err
+	}
+
+	if c.state.State == ContainerStateStopped {
+		// Reinitialize the container if we need to
+		if err := c.reinit(ctx); err != nil {
+			return err
+		}
+	} else if c.state.State == ContainerStateConfigured ||
+		c.state.State == ContainerStateExited {
+		// Initialize the container
+		if err := c.init(ctx); err != nil {
+			return err
+		}
+	}
+	return c.start()
+}
+
 // Sync this container with on-disk state and runtime status
 // Should only be called with container lock held
 // This function should suffice to ensure a container's state is accurate and
@@ -230,6 +267,14 @@ func (c *Container) syncContainer() error {
 		}
 		// Only save back to DB if state changed
 		if c.state.State != oldState {
+			// Check for a restart policy match
+			if c.config.RestartPolicy != "" && c.config.RestartPolicy != "no" &&
+				(oldState == ContainerStateRunning || oldState == ContainerStatePaused) &&
+				(c.state.State == ContainerStateStopped || c.state.State == ContainerStateExited) &&
+				!c.state.StoppedByUser {
+				c.state.RestartPolicyMatch = true
+			}
+
 			if err := c.save(); err != nil {
 				return err
 			}
@@ -377,6 +422,7 @@ func resetState(state *ContainerState) error {
 	state.NetworkStatus = nil
 	state.BindMounts = make(map[string]string)
 	state.StoppedByUser = false
+	state.RestartPolicyMatch = false
 
 	return nil
 }
@@ -791,6 +837,7 @@ func (c *Container) init(ctx context.Context) error {
 	c.state.Exited = false
 	c.state.State = ContainerStateCreated
 	c.state.StoppedByUser = false
+	c.state.RestartPolicyMatch = false
 
 	if err := c.save(); err != nil {
 		return err
diff --git a/libpod/options.go b/libpod/options.go
index 86c04db09..e83515822 100644
--- a/libpod/options.go
+++ b/libpod/options.go
@@ -1239,6 +1239,41 @@ func WithUseImageHosts() CtrCreateOption {
 	}
 }
 
+// WithRestartPolicy sets the container's restart policy. Valid values are
+// "no", "on-failure", and "always". The empty string is allowed, and will be
+// equivalent to "no".
+func WithRestartPolicy(policy string) CtrCreateOption {
+	return func(ctr *Container) error {
+		if ctr.valid {
+			return ErrCtrFinalized
+		}
+
+		switch policy {
+		case "", "no", "on-failure", "always":
+			ctr.config.RestartPolicy = policy
+		default:
+			return errors.Wrapf(ErrInvalidArg, "%q is not a valid restart policy", policy)
+		}
+
+		return nil
+	}
+}
+
+// WithRestartRetries sets the number of retries to use when restarting a
+// container with the "on-failure" restart policy.
+// 0 is an allowed value, and indicates infinite retries.
+func WithRestartRetries(tries uint) CtrCreateOption {
+	return func(ctr *Container) error {
+		if ctr.valid {
+			return ErrCtrFinalized
+		}
+
+		ctr.config.RestartRetries = tries
+
+		return nil
+	}
+}
+
 // withIsInfra sets the container to be an infra container. This means the container will be sometimes hidden
 // and expected to be the first container in the pod.
 func withIsInfra() CtrCreateOption {