cgroup2: exec: join the cgroup of the init process on EBUSY

Signed-off-by: Akihiro Suda <[email protected]>

libcontainer/container_linux.go

@@ -565,6 +565,7 @@ func (c *linuxContainer) newSetnsProcess(p *Process, cmd *exec.Cmd, messageSockP
 		config:          c.newInitConfig(p),
 		process:         p,
 		bootstrapData:   data,
+		initProcessPid:  state.InitProcessPid,
 	}, nil
 }
 

libcontainer/process_linux.go

@@ -5,19 +5,22 @@ package libcontainer
 import (
 	"encoding/json"
 	"errors"
+	"fmt"
 	"io"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"strconv"
 
 	"github.com/opencontainers/runc/libcontainer/cgroups"
+	"github.com/opencontainers/runc/libcontainer/cgroups/fs2"
 	"github.com/opencontainers/runc/libcontainer/configs"
 	"github.com/opencontainers/runc/libcontainer/intelrdt"
 	"github.com/opencontainers/runc/libcontainer/logs"
 	"github.com/opencontainers/runc/libcontainer/system"
 	"github.com/opencontainers/runc/libcontainer/utils"
 
+	"github.com/sirupsen/logrus"
 	"golang.org/x/sys/unix"
 )
 
@@ -66,6 +69,7 @@ type setnsProcess struct {
 	fds             []string
 	process         *Process
 	bootstrapData   io.Reader
+	initProcessPid  int
 }
 
 func (p *setnsProcess) startTime() (uint64, error) {
@@ -100,7 +104,25 @@ func (p *setnsProcess) start() (err error) {
 	}
 	if len(p.cgroupPaths) > 0 {
 		if err := cgroups.EnterPid(p.cgroupPaths, p.pid()); err != nil && !p.rootlessCgroups {
-			return newSystemErrorWithCausef(err, "adding pid %d to cgroups", p.pid())
+			// On cgroup v2 + nesting + domain controllers, EnterPid may fail with EBUSY.
+			// https://github.com/opencontainers/runc/issues/2356#issuecomment-621277643
+			// Try to join the cgroup of InitProcessPid.
+			if cgroups.IsCgroup2UnifiedMode() {
+				initProcCgroupFile := fmt.Sprintf("/proc/%d/cgroup", p.initProcessPid)
+				initCg, initCgErr := cgroups.ParseCgroupFile(initProcCgroupFile)
+				if initCgErr == nil {
+					if initCgPath, ok := initCg[""]; ok {
+						initCgDirpath := filepath.Join(fs2.UnifiedMountpoint, initCgPath)
+						logrus.Warnf("adding pid %d to cgroups %v failed (%v), attempting to join %q (obtained from %s)",
+							p.pid(), p.cgroupPaths, err, initCg, initCgDirpath)
+						// NOTE: initCgDirPath is not guaranteed to exist because we didn't pause the container.
+						err = cgroups.WriteCgroupProc(initCgDirpath, p.pid())
+					}
+				}
+			}
+			if err != nil {
+				return newSystemErrorWithCausef(err, "adding pid %d to cgroups", p.pid())
+			}
 		}
 	}
 	if p.intelRdtPath != "" {

tests/integration/cgroups.bats

@@ -125,3 +125,55 @@ EOF
     [ "$status" -eq 0 ]
     [[ ${lines[0]} == *"cgroups_exec"* ]]
 }
+
+@test "runc exec (cgroup v2 + init process in non-root cgroup) succeeds" {
+    requires root cgroups_v2
+
+    set_cgroups_path "$BUSYBOX_BUNDLE"
+# set_memory_limit is currently required to make sure "memory" appear in /sys/fs/cgroup/cgroup.controllers .
+# set_memory_limit will no longer be needed after merging "Partially revert "CreateCgroupPath: only enable needed controllers" (#2395)
+    set_memory_limit "$BUSYBOX_BUNDLE"
+    set_cgroup_mount_writable "$BUSYBOX_BUNDLE"
+
+    runc run -d --console-socket $CONSOLE_SOCKET test_cgroups_group
+    [ "$status" -eq 0 ]
+
+    runc exec test_cgroups_group cat /sys/fs/cgroup/cgroup.controllers
+    [ "$status" -eq 0 ]
+    [[ ${lines[0]} == *"memory"* ]]
+
+    runc exec test_cgroups_group cat /proc/self/cgroup
+    [ "$status" -eq 0 ]
+    [[ ${lines[0]} == "0::/" ]]
+
+    runc exec test_cgroups_group mkdir /sys/fs/cgroup/foo
+    [ "$status" -eq 0 ]
+
+    runc exec test_cgroups_group sh -c "echo 1 > /sys/fs/cgroup/foo/cgroup.procs"
+    [ "$status" -eq 0 ]
+
+# the init process is now in "/foo", but an exec process can still join "/"
+# because we haven't enabled any domain controller.
+    runc exec test_cgroups_group cat /proc/self/cgroup
+    [ "$status" -eq 0 ]
+    [[ ${lines[0]} == "0::/" ]]
+
+# turn on a domain controller (memory)
+    runc exec test_cgroups_group sh -euxc 'echo $$ > /sys/fs/cgroup/foo/cgroup.procs; echo +memory > /sys/fs/cgroup/cgroup.subtree_control'
+    [ "$status" -eq 0 ]
+
+# an exec process can no longer join "/" after turning on a domain controller.
+# falls back to "/foo".
+    runc exec test_cgroups_group cat /proc/self/cgroup
+    [ "$status" -eq 0 ]
+# warning log from runc (stderr)
+    [[ ${lines[0]} == *"attempting to join"* ]]
+# actual container log
+    [[ ${lines[1]} == "0::/foo" ]]
+
+# teardown: remove "/foo"
+    runc exec test_cgroups_group sh -uxc 'echo -memory > /sys/fs/cgroup/cgroup.subtree_control; for f in $(cat /sys/fs/cgroup/foo/cgroup.procs); do echo $f > /sys/fs/cgroup/cgroup.procs; done; rmdir /sys/fs/cgroup/foo'
+    runc exec test_cgroups_group test ! -d /sys/fs/cgroup/foo
+    [ "$status" -eq 0 ]
+#
+}

tests/integration/helpers.bash

@@ -197,6 +197,24 @@ function set_resources_limit() {
   sed -i 's/\("linux": {\)/\1\n   "resources": { "pids": { "limit": 100 } },/'  "$bundle/config.json"
 }
 
+# Helper function to set a memory limit
+function set_memory_limit(){
+	bundle="${1:-.}"
+	cat "$bundle/config.json" \
+        |  jq '.linux.resources += {"memory": {"limit":33554432}}' \
+		>"$bundle/config.json.tmp"
+	mv "$bundle/config.json"{.tmp,}
+}
+
+# Helper function to make /sys/fs/cgroup writable
+function set_cgroup_mount_writable(){
+	bundle="${1:-.}"
+	cat "$bundle/config.json" \
+        |  jq '.mounts |= map((select(.type == "cgroup") | .options -= ["ro"]) // .)' \
+		>"$bundle/config.json.tmp"
+	mv "$bundle/config.json"{.tmp,}
+}
+
 # Fails the current test, providing the error given.
 function fail() {
 	echo "$@" >&2