Merge pull request #37701 from dperny/add-swarmkit-sysctl-support

Add support for sysctl options in services

api/server/router/swarm/cluster_routes.go

@@ -182,8 +182,17 @@ func (sr *swarmRouter) createService(ctx context.Context, w http.ResponseWriter,
 	encodedAuth := r.Header.Get("X-Registry-Auth")
 	cliVersion := r.Header.Get("version")
 	queryRegistry := false
-	if cliVersion != "" && versions.LessThan(cliVersion, "1.30") {
+	if cliVersion != "" {
-		queryRegistry = true
+		if versions.LessThan(cliVersion, "1.30") {
+			queryRegistry = true
+		}
+		if versions.LessThan(cliVersion, "1.39") {
+			if service.TaskTemplate.ContainerSpec != nil {
+				// Sysctls for docker swarm services weren't supported before
+				// API version 1.39
+				service.TaskTemplate.ContainerSpec.Sysctls = nil
+			}
+		}
 	}
 
 	resp, err := sr.backend.CreateService(service, encodedAuth, queryRegistry)
@@ -216,8 +225,17 @@ func (sr *swarmRouter) updateService(ctx context.Context, w http.ResponseWriter,
 	flags.Rollback = r.URL.Query().Get("rollback")
 	cliVersion := r.Header.Get("version")
 	queryRegistry := false
-	if cliVersion != "" && versions.LessThan(cliVersion, "1.30") {
+	if cliVersion != "" {
-		queryRegistry = true
+		if versions.LessThan(cliVersion, "1.30") {
+			queryRegistry = true
+		}
+		if versions.LessThan(cliVersion, "1.39") {
+			if service.TaskTemplate.ContainerSpec != nil {
+				// Sysctls for docker swarm services weren't supported before
+				// API version 1.39
+				service.TaskTemplate.ContainerSpec.Sysctls = nil
+			}
+		}
 	}
 
 	resp, err := sr.backend.UpdateService(vars["id"], version, service, flags, queryRegistry)

api/swagger.yaml

@@ -2750,6 +2750,18 @@ definitions:
             description: "Run an init inside the container that forwards signals and reaps processes. This field is omitted if empty, and the default (as configured on the daemon) is used."
             type: "boolean"
             x-nullable: true
+          Sysctls:
+            description: |
+              Set kernel namedspaced parameters (sysctls) in the container.
+              The Sysctls option on services accepts the same sysctls as the
+              are supported on containers. Note that while the same sysctls are
+              supported, no guarantees or checks are made about their
+              suitability for a clustered environment, and it's up to the user
+              to determine whether a given sysctl will work properly in a
+              Service.
+            type: "object"
+            additionalProperties:
+              type: "string"
       NetworkAttachmentSpec:
         description: |
           Read-only spec type for non-swarm containers attached to swarm overlay

api/types/swarm/container.go

@@ -71,4 +71,5 @@ type ContainerSpec struct {
 	Secrets   []*SecretReference  `json:",omitempty"`
 	Configs   []*ConfigReference  `json:",omitempty"`
 	Isolation container.Isolation `json:",omitempty"`
+	Sysctls   map[string]string   `json:",omitempty"`
 }

daemon/cluster/convert/container.go

@@ -36,6 +36,7 @@ func containerSpecFromGRPC(c *swarmapi.ContainerSpec) *types.ContainerSpec {
 		Configs:    configReferencesFromGRPC(c.Configs),
 		Isolation:  IsolationFromGRPC(c.Isolation),
 		Init:       initFromGRPC(c.Init),
+		Sysctls:    c.Sysctls,
 	}
 
 	if c.DNSConfig != nil {
@@ -251,6 +252,7 @@ func containerToGRPC(c *types.ContainerSpec) (*swarmapi.ContainerSpec, error) {
 		Configs:    configReferencesToGRPC(c.Configs),
 		Isolation:  isolationToGRPC(c.Isolation),
 		Init:       initToGRPC(c.Init),
+		Sysctls:    c.Sysctls,
 	}
 
 	if c.DNSConfig != nil {

daemon/cluster/executor/container/container.go

@@ -364,6 +364,7 @@ func (c *containerConfig) hostConfig() *enginecontainer.HostConfig {
 		ReadonlyRootfs: c.spec().ReadOnly,
 		Isolation:      c.isolation(),
 		Init:           c.init(),
+		Sysctls:        c.spec().Sysctls,
 	}
 
 	if c.spec().DNSConfig != nil {

docs/api/version-history.md

@@ -30,6 +30,12 @@ keywords: "API, Docker, rcli, REST, documentation"
   on the node.label. The format of the label filter is `node.label=<key>`/`node.label=<key>=<value>`
   to return those with the specified labels, or `node.label!=<key>`/`node.label!=<key>=<value>`
   to return those without the specified labels.
+* `GET /services` now returns `Sysctls` as part of the `ContainerSpec`.
+* `GET /services/{id}` now returns `Sysctls` as part of the `ContainerSpec`.
+* `POST /services/create` now accepts `Sysctls` as part of the `ContainerSpec`.
+* `POST /services/{id}/update` now accepts `Sysctls` as part of the `ContainerSpec`.
+* `GET /tasks` now  returns `Sysctls` as part of the `ContainerSpec`.
+* `GET /tasks/{id}` now  returns `Sysctls` as part of the `ContainerSpec`.
 
 ## V1.38 API changes
 

integration/internal/swarm/service.go

@@ -159,6 +159,14 @@ func ServiceWithEndpoint(endpoint *swarmtypes.EndpointSpec) ServiceSpecOpt {
 	}
 }
 
+// ServiceWithSysctls sets the Sysctls option of the service's ContainerSpec.
+func ServiceWithSysctls(sysctls map[string]string) ServiceSpecOpt {
+	return func(spec *swarmtypes.ServiceSpec) {
+		ensureContainerSpec(spec)
+		spec.TaskTemplate.ContainerSpec.Sysctls = sysctls
+	}
+}
+
 // GetRunningTasks gets the list of running tasks for a service
 func GetRunningTasks(t *testing.T, d *daemon.Daemon, serviceID string) []swarmtypes.Task {
 	t.Helper()

integration/service/create_test.go

@@ -10,6 +10,7 @@ import (
 	"github.com/docker/docker/api/types"
 	"github.com/docker/docker/api/types/filters"
 	swarmtypes "github.com/docker/docker/api/types/swarm"
+	"github.com/docker/docker/api/types/versions"
 	"github.com/docker/docker/client"
 	"github.com/docker/docker/integration/internal/network"
 	"github.com/docker/docker/integration/internal/swarm"
@@ -17,6 +18,7 @@ import (
 	"gotest.tools/assert"
 	is "gotest.tools/assert/cmp"
 	"gotest.tools/poll"
+	"gotest.tools/skip"
 )
 
 func TestServiceCreateInit(t *testing.T) {
@@ -309,6 +311,101 @@ func TestCreateServiceConfigFileMode(t *testing.T) {
 	assert.NilError(t, err)
 }
 
+// TestServiceCreateSysctls tests that a service created with sysctl options in
+// the ContainerSpec correctly applies those options.
+//
+// To test this, we're going to create a service with the sysctl option
+//
+//   {"net.ipv4.ip_nonlocal_bind": "0"}
+//
+// We'll get the service's tasks to get the container ID, and then we'll
+// inspect the container. If the output of the container inspect contains the
+// sysctl option with the correct value, we can assume that the sysctl has been
+// plumbed correctly.
+//
+// Next, we'll remove that service and create a new service with that option
+// set to 1. This means that no matter what the default is, we can be confident
+// that the sysctl option is applying as intended.
+//
+// Additionally, we'll do service and task inspects to verify that the inspect
+// output includes the desired sysctl option.
+//
+// We're using net.ipv4.ip_nonlocal_bind because it's something that I'm fairly
+// confident won't be modified by the container runtime, and won't blow
+// anything up in the test environment
+func TestCreateServiceSysctls(t *testing.T) {
+	skip.If(
+		t, versions.LessThan(testEnv.DaemonAPIVersion(), "1.39"),
+		"setting service sysctls is unsupported before api v1.39",
+	)
+
+	defer setupTest(t)()
+	d := swarm.NewSwarm(t, testEnv)
+	defer d.Stop(t)
+	client := d.NewClientT(t)
+	defer client.Close()
+
+	ctx := context.Background()
+
+	// run thie block twice, so that no matter what the default value of
+	// net.ipv4.ip_nonlocal_bind is, we can verify that setting the sysctl
+	// options works
+	for _, expected := range []string{"0", "1"} {
+
+		// store the map we're going to be using everywhere.
+		expectedSysctls := map[string]string{"net.ipv4.ip_nonlocal_bind": expected}
+
+		// Create the service with the sysctl options
+		var instances uint64 = 1
+		serviceID := swarm.CreateService(t, d,
+			swarm.ServiceWithSysctls(expectedSysctls),
+		)
+
+		// wait for the service to converge to 1 running task as expected
+		poll.WaitOn(t, serviceRunningTasksCount(client, serviceID, instances))
+
+		// we're going to check 3 things:
+		//
+		//   1. Does the container, when inspected, have the sysctl option set?
+		//   2. Does the task have the sysctl in the spec?
+		//   3. Does the service have the sysctl in the spec?
+		//
+		// if all 3 of these things are true, we know that the sysctl has been
+		// plumbed correctly through the engine.
+		//
+		// We don't actually have to get inside the container and check its
+		// logs or anything. If we see the sysctl set on the container inspect,
+		// we know that the sysctl is plumbed correctly. everything below that
+		// level has been tested elsewhere. (thanks @thaJeztah, because an
+		// earlier version of this test had to get container logs and was much
+		// more complex)
+
+		// get all of the tasks of the service, so we can get the container
+		filter := filters.NewArgs()
+		filter.Add("service", serviceID)
+		tasks, err := client.TaskList(ctx, types.TaskListOptions{
+			Filters: filter,
+		})
+		assert.NilError(t, err)
+		assert.Check(t, is.Equal(len(tasks), 1))
+
+		// verify that the container has the sysctl option set
+		ctnr, err := client.ContainerInspect(ctx, tasks[0].Status.ContainerStatus.ContainerID)
+		assert.NilError(t, err)
+		assert.DeepEqual(t, ctnr.HostConfig.Sysctls, expectedSysctls)
+
+		// verify that the task has the sysctl option set in the task object
+		assert.DeepEqual(t, tasks[0].Spec.ContainerSpec.Sysctls, expectedSysctls)
+
+		// verify that the service also has the sysctl set in the spec.
+		service, _, err := client.ServiceInspectWithRaw(ctx, serviceID, types.ServiceInspectOptions{})
+		assert.NilError(t, err)
+		assert.DeepEqual(t,
+			service.Spec.TaskTemplate.ContainerSpec.Sysctls, expectedSysctls,
+		)
+	}
+}
+
 func serviceRunningTasksCount(client client.ServiceAPIClient, serviceID string, instances uint64) func(log poll.LogT) poll.Result {
 	return func(log poll.LogT) poll.Result {
 		filter := filters.NewArgs()