Add support for pod inside of user namespace.

Add the --userns flag to podman pod create and keep
track of the userns setting that pod was created with
so that all containers created within the pod will inherit
that userns setting.

Specifically we need to be able to launch a pod with
--userns=keep-id

Signed-off-by: Daniel J Walsh <dwalsh@redhat.com>
Signed-off-by: Urvashi Mohnani <umohnani@redhat.com>

7 files changed, 95 insertions, 44 deletions

diff --git a/pkg/api/handlers/libpod/pods.go b/pkg/api/handlers/libpod/pods.go
index 4dc8740e2..ff105bc48 100644
--- a/pkg/api/handlers/libpod/pods.go
+++ b/pkg/api/handlers/libpod/pods.go
@@ -30,6 +30,12 @@ func PodCreate(w http.ResponseWriter, r *http.Request) {
 		utils.Error(w, "failed to decode specgen", http.StatusInternalServerError, errors.Wrap(err, "failed to decode specgen"))
 		return
 	}
+	// parse userns so we get the valid default value of userns
+	psg.Userns, err = specgen.ParseUserNamespace(psg.Userns.String())
+	if err != nil {
+		utils.Error(w, "failed to parse userns", http.StatusInternalServerError, errors.Wrap(err, "failed to parse userns"))
+		return
+	}
 	pod, err := generate.MakePod(&psg, runtime)
 	if err != nil {
 		httpCode := http.StatusInternalServerError
diff --git a/pkg/domain/entities/pods.go b/pkg/domain/entities/pods.go
index 68e335f8d..c66bf96fc 100644
--- a/pkg/domain/entities/pods.go
+++ b/pkg/domain/entities/pods.go
@@ -122,6 +122,7 @@ type PodCreateOptions struct {
 	Pid                string
 	Cpus               float64
 	CpusetCpus         string
+	Userns             specgen.Namespace
 }
 
 type PodCreateReport struct {
@@ -217,6 +218,7 @@ func (p *PodCreateOptions) ToPodSpecGen(s *specgen.PodSpecGenerator) error {
 			s.CPUQuota = *cpuDat.Quota
 		}
 	}
+	s.Userns = p.Userns
 	return nil
 }
 
diff --git a/pkg/specgen/generate/kube/kube.go b/pkg/specgen/generate/kube/kube.go
index fb7eb99a2..d6d479c67 100644
--- a/pkg/specgen/generate/kube/kube.go
+++ b/pkg/specgen/generate/kube/kube.go
@@ -303,6 +303,8 @@ func ToSpecGen(ctx context.Context, opts *CtrSpecGenOptions) (*specgen.SpecGener
 	if opts.NetNSIsHost {
 		s.NetNS.NSMode = specgen.Host
 	}
+	// Always set the userns to host since k8s doesn't have support for userns yet
+	s.UserNS.NSMode = specgen.Host
 
 	// Add labels that come from kube
 	if len(s.Labels) == 0 {
diff --git a/pkg/specgen/generate/namespaces.go b/pkg/specgen/generate/namespaces.go
index f41186ae4..1d7373093 100644
--- a/pkg/specgen/generate/namespaces.go
+++ b/pkg/specgen/generate/namespaces.go
@@ -175,6 +175,11 @@ func namespaceOptions(ctx context.Context, s *specgen.SpecGenerator, rt *libpod.
 		if pod == nil || infraCtr == nil {
 			return nil, errNoInfra
 		}
+		// Inherit the user from the infra container if it is set and --user has not
+		// been set explicitly
+		if infraCtr.User() != "" && s.User == "" {
+			toReturn = append(toReturn, libpod.WithUser(infraCtr.User()))
+		}
 		toReturn = append(toReturn, libpod.WithUserNSFrom(infraCtr))
 	case specgen.FromContainer:
 		userCtr, err := rt.LookupContainer(s.UserNS.Value)
@@ -184,7 +189,10 @@ func namespaceOptions(ctx context.Context, s *specgen.SpecGenerator, rt *libpod.
 		toReturn = append(toReturn, libpod.WithUserNSFrom(userCtr))
 	}
 
-	if s.IDMappings != nil {
+	// This wipes the UserNS settings that get set from the infra container
+	// when we are inheritting from the pod. So only apply this if the container
+	// is not being created in a pod.
+	if s.IDMappings != nil && pod == nil {
 		toReturn = append(toReturn, libpod.WithIDMappings(*s.IDMappings))
 	}
 	if s.User != "" {
@@ -379,46 +387,8 @@ func specConfigureNamespaces(s *specgen.SpecGenerator, g *generate.Generator, rt
 	}
 
 	// User
-	switch s.UserNS.NSMode {
-	case specgen.Path:
-		if _, err := os.Stat(s.UserNS.Value); err != nil {
-			return errors.Wrap(err, "cannot find specified user namespace path")
-		}
-		if err := g.AddOrReplaceLinuxNamespace(string(spec.UserNamespace), s.UserNS.Value); err != nil {
-			return err
-		}
-		// runc complains if no mapping is specified, even if we join another ns.  So provide a dummy mapping
-		g.AddLinuxUIDMapping(uint32(0), uint32(0), uint32(1))
-		g.AddLinuxGIDMapping(uint32(0), uint32(0), uint32(1))
-	case specgen.Host:
-		if err := g.RemoveLinuxNamespace(string(spec.UserNamespace)); err != nil {
-			return err
-		}
-	case specgen.KeepID:
-		var (
-			err      error
-			uid, gid int
-		)
-		s.IDMappings, uid, gid, err = util.GetKeepIDMapping()
-		if err != nil {
-			return err
-		}
-		g.SetProcessUID(uint32(uid))
-		g.SetProcessGID(uint32(gid))
-		fallthrough
-	case specgen.Private:
-		if err := g.AddOrReplaceLinuxNamespace(string(spec.UserNamespace), ""); err != nil {
-			return err
-		}
-		if s.IDMappings == nil || (len(s.IDMappings.UIDMap) == 0 && len(s.IDMappings.GIDMap) == 0) {
-			return errors.Errorf("must provide at least one UID or GID mapping to configure a user namespace")
-		}
-		for _, uidmap := range s.IDMappings.UIDMap {
-			g.AddLinuxUIDMapping(uint32(uidmap.HostID), uint32(uidmap.ContainerID), uint32(uidmap.Size))
-		}
-		for _, gidmap := range s.IDMappings.GIDMap {
-			g.AddLinuxGIDMapping(uint32(gidmap.HostID), uint32(gidmap.ContainerID), uint32(gidmap.Size))
-		}
+	if _, err := specgen.SetupUserNS(s.IDMappings, s.UserNS, g); err != nil {
+		return err
 	}
 
 	// Cgroup
@@ -474,7 +444,7 @@ func specConfigureNamespaces(s *specgen.SpecGenerator, g *generate.Generator, rt
 // GetNamespaceOptions transforms a slice of kernel namespaces
 // into a slice of pod create options. Currently, not all
 // kernel namespaces are supported, and they will be returned in an error
-func GetNamespaceOptions(ns []string) ([]libpod.PodCreateOption, error) {
+func GetNamespaceOptions(ns []string, netnsIsHost bool) ([]libpod.PodCreateOption, error) {
 	var options []libpod.PodCreateOption
 	var erroredOptions []libpod.PodCreateOption
 	if ns == nil {
@@ -486,7 +456,10 @@ func GetNamespaceOptions(ns []string) ([]libpod.PodCreateOption, error) {
 		case "cgroup":
 			options = append(options, libpod.WithPodCgroups())
 		case "net":
-			options = append(options, libpod.WithPodNet())
+			// share the netns setting with other containers in the pod only when it is not set to host
+			if !netnsIsHost {
+				options = append(options, libpod.WithPodNet())
+			}
 		case "mnt":
 			return erroredOptions, errors.Errorf("Mount sharing functionality not supported on pod level")
 		case "pid":
diff --git a/pkg/specgen/generate/pod_create.go b/pkg/specgen/generate/pod_create.go
index aab29499e..426cf1b6d 100644
--- a/pkg/specgen/generate/pod_create.go
+++ b/pkg/specgen/generate/pod_create.go
@@ -27,11 +27,16 @@ func createPodOptions(p *specgen.PodSpecGenerator, rt *libpod.Runtime) ([]libpod
 	)
 	if !p.NoInfra {
 		options = append(options, libpod.WithInfraContainer())
-		nsOptions, err := GetNamespaceOptions(p.SharedNamespaces)
+		nsOptions, err := GetNamespaceOptions(p.SharedNamespaces, p.NetNS.IsHost())
 		if err != nil {
 			return nil, err
 		}
 		options = append(options, nsOptions...)
+		// Use pod user and infra userns only when --userns is not set to host
+		if !p.Userns.IsHost() {
+			options = append(options, libpod.WithPodUser())
+			options = append(options, libpod.WithPodUserns(p.Userns))
+		}
 
 		// Make our exit command
 		storageConfig := rt.StorageConfig()
@@ -154,5 +159,6 @@ func createPodOptions(p *specgen.PodSpecGenerator, rt *libpod.Runtime) ([]libpod
 	if len(p.InfraConmonPidFile) > 0 {
 		options = append(options, libpod.WithInfraConmonPidFile(p.InfraConmonPidFile))
 	}
+
 	return options, nil
 }
diff --git a/pkg/specgen/namespaces.go b/pkg/specgen/namespaces.go
index 76fa66bc7..2f4c48811 100644
--- a/pkg/specgen/namespaces.go
+++ b/pkg/specgen/namespaces.go
@@ -1,10 +1,16 @@
 package specgen
 
 import (
+	"fmt"
+	"os"
 	"strings"
 
 	"github.com/containers/podman/v3/pkg/cgroups"
 	"github.com/containers/podman/v3/pkg/rootless"
+	"github.com/containers/podman/v3/pkg/util"
+	"github.com/containers/storage"
+	spec "github.com/opencontainers/runtime-spec/specs-go"
+	"github.com/opencontainers/runtime-tools/generate"
 	"github.com/pkg/errors"
 )
 
@@ -103,6 +109,13 @@ func (n *Namespace) IsKeepID() bool {
 	return n.NSMode == KeepID
 }
 
+func (n *Namespace) String() string {
+	if n.Value != "" {
+		return fmt.Sprintf("%s:%s", n.NSMode, n.Value)
+	}
+	return string(n.NSMode)
+}
+
 func validateUserNS(n *Namespace) error {
 	if n == nil {
 		return nil
@@ -323,3 +336,48 @@ func ParseNetworkString(network string) (Namespace, []string, map[string][]strin
 	}
 	return ns, cniNets, networkOptions, nil
 }
+
+func SetupUserNS(idmappings *storage.IDMappingOptions, userns Namespace, g *generate.Generator) (string, error) {
+	// User
+	var user string
+	switch userns.NSMode {
+	case Path:
+		if _, err := os.Stat(userns.Value); err != nil {
+			return user, errors.Wrap(err, "cannot find specified user namespace path")
+		}
+		if err := g.AddOrReplaceLinuxNamespace(string(spec.UserNamespace), userns.Value); err != nil {
+			return user, err
+		}
+		// runc complains if no mapping is specified, even if we join another ns.  So provide a dummy mapping
+		g.AddLinuxUIDMapping(uint32(0), uint32(0), uint32(1))
+		g.AddLinuxGIDMapping(uint32(0), uint32(0), uint32(1))
+	case Host:
+		if err := g.RemoveLinuxNamespace(string(spec.UserNamespace)); err != nil {
+			return user, err
+		}
+	case KeepID:
+		mappings, uid, gid, err := util.GetKeepIDMapping()
+		if err != nil {
+			return user, err
+		}
+		idmappings = mappings
+		g.SetProcessUID(uint32(uid))
+		g.SetProcessGID(uint32(gid))
+		user = fmt.Sprintf("%d:%d", uid, gid)
+		fallthrough
+	case Private:
+		if err := g.AddOrReplaceLinuxNamespace(string(spec.UserNamespace), ""); err != nil {
+			return user, err
+		}
+		if idmappings == nil || (len(idmappings.UIDMap) == 0 && len(idmappings.GIDMap) == 0) {
+			return user, errors.Errorf("must provide at least one UID or GID mapping to configure a user namespace")
+		}
+		for _, uidmap := range idmappings.UIDMap {
+			g.AddLinuxUIDMapping(uint32(uidmap.HostID), uint32(uidmap.ContainerID), uint32(uidmap.Size))
+		}
+		for _, gidmap := range idmappings.GIDMap {
+			g.AddLinuxGIDMapping(uint32(gidmap.HostID), uint32(gidmap.ContainerID), uint32(gidmap.Size))
+		}
+	}
+	return user, nil
+}
diff --git a/pkg/specgen/podspecgen.go b/pkg/specgen/podspecgen.go
index 02237afe9..893ebf675 100644
--- a/pkg/specgen/podspecgen.go
+++ b/pkg/specgen/podspecgen.go
@@ -67,6 +67,10 @@ type PodBasicConfig struct {
 	// Optional (defaults to private if unset). This sets the PID namespace of the infra container
 	// This configuration will then be shared with the entire pod if PID namespace sharing is enabled via --share
 	Pid Namespace `json:"pid,omitempty:"`
+	// Userns is used to indicate which kind of Usernamespace to enter.
+	// Any containers created within the pod will inherit the pod's userns settings.
+	// Optional
+	Userns Namespace `json:"userns,omitempty"`
 }
 
 // PodNetworkConfig contains networking configuration for a pod.