Add /proc/[pid]/cgroups file

Updates #1195

PiperOrigin-RevId: 290298266

pkg/sentry/fsimpl/proc/filesystem.go

@@ -47,7 +47,12 @@ func (ft *procFSType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFile
 	procfs := &kernfs.Filesystem{}
 	procfs.VFSFilesystem().Init(vfsObj, procfs)
 
-	_, dentry := newTasksInode(procfs, k, pidns)
+	var data *InternalData
+	if opts.InternalData != nil {
+		data = opts.InternalData.(*InternalData)
+	}
+
+	_, dentry := newTasksInode(procfs, k, pidns, data.Cgroups)
 	return procfs.VFSFilesystem(), dentry.VFSDentry(), nil
 }
 
@@ -78,3 +83,9 @@ var _ dynamicInode = (*staticFile)(nil)
 func newStaticFile(data string) *staticFile {
 	return &staticFile{StaticData: vfs.StaticData{Data: data}}
 }
+
+// InternalData contains internal data passed in to the procfs mount via
+// vfs.GetFilesystemOptions.InternalData.
+type InternalData struct {
+	Cgroups map[string]string
+}

pkg/sentry/fsimpl/proc/subtasks.go

@@ -35,18 +35,20 @@ type subtasksInode struct {
 	kernfs.InodeAttrs
 	kernfs.OrderedChildren
 
-	task   *kernel.Task
-	pidns  *kernel.PIDNamespace
-	inoGen InoGenerator
+	task              *kernel.Task
+	pidns             *kernel.PIDNamespace
+	inoGen            InoGenerator
+	cgroupControllers map[string]string
 }
 
 var _ kernfs.Inode = (*subtasksInode)(nil)
 
-func newSubtasks(task *kernel.Task, pidns *kernel.PIDNamespace, inoGen InoGenerator) *kernfs.Dentry {
+func newSubtasks(task *kernel.Task, pidns *kernel.PIDNamespace, inoGen InoGenerator, cgroupControllers map[string]string) *kernfs.Dentry {
 	subInode := &subtasksInode{
-		task:   task,
-		pidns:  pidns,
-		inoGen: inoGen,
+		task:              task,
+		pidns:             pidns,
+		inoGen:            inoGen,
+		cgroupControllers: cgroupControllers,
 	}
 	// Note: credentials are overridden by taskOwnedInode.
 	subInode.InodeAttrs.Init(task.Credentials(), inoGen.NextIno(), linux.ModeDirectory|0555)
@@ -79,7 +81,7 @@ func (i *subtasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, e
 		return nil, syserror.ENOENT
 	}
 
-	subTaskDentry := newTaskInode(i.inoGen, subTask, i.pidns, false)
+	subTaskDentry := newTaskInode(i.inoGen, subTask, i.pidns, false, i.cgroupControllers)
 	return subTaskDentry.VFSDentry(), nil
 }
 

pkg/sentry/fsimpl/proc/task.go

@@ -15,6 +15,7 @@
 package proc
 
 import (
+	"bytes"
 	"fmt"
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
@@ -42,7 +43,7 @@ type taskInode struct {
 
 var _ kernfs.Inode = (*taskInode)(nil)
 
-func newTaskInode(inoGen InoGenerator, task *kernel.Task, pidns *kernel.PIDNamespace, isThreadGroup bool) *kernfs.Dentry {
+func newTaskInode(inoGen InoGenerator, task *kernel.Task, pidns *kernel.PIDNamespace, isThreadGroup bool, cgroupControllers map[string]string) *kernfs.Dentry {
 	contents := map[string]*kernfs.Dentry{
 		"auxv":    newTaskOwnedFile(task, inoGen.NextIno(), 0444, &auxvData{task: task}),
 		"cmdline": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &cmdlineData{task: task, arg: cmdlineDataArg}),
@@ -68,11 +69,11 @@ func newTaskInode(inoGen InoGenerator, task *kernel.Task, pidns *kernel.PIDNames
 		"uid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: false}),
 	}
 	if isThreadGroup {
-		contents["task"] = newSubtasks(task, pidns, inoGen)
+		contents["task"] = newSubtasks(task, pidns, inoGen, cgroupControllers)
+	}
+	if len(cgroupControllers) > 0 {
+		contents["cgroup"] = newTaskOwnedFile(task, inoGen.NextIno(), 0444, newCgroupData(cgroupControllers))
 	}
-	//if len(p.cgroupControllers) > 0 {
-	//	contents["cgroup"] = newCGroupInode(t, msrc, p.cgroupControllers)
-	//}
 
 	taskInode := &taskInode{task: task}
 	// Note: credentials are overridden by taskOwnedInode.
@@ -227,3 +228,22 @@ func newNamespaceSymlink(task *kernel.Task, ino uint64, ns string) *kernfs.Dentr
 	d.Init(taskInode)
 	return d
 }
+
+// newCgroupData creates inode that shows cgroup information.
+// From man 7 cgroups: "For each cgroup hierarchy of which the process is a
+// member, there is one entry containing three colon-separated fields:
+//   hierarchy-ID:controller-list:cgroup-path"
+func newCgroupData(controllers map[string]string) dynamicInode {
+	buf := bytes.Buffer{}
+
+	// The hierarchy ids must be positive integers (for cgroup v1), but the
+	// exact number does not matter, so long as they are unique. We can
+	// just use a counter, but since linux sorts this file in descending
+	// order, we must count down to preserve this behavior.
+	i := len(controllers)
+	for name, dir := range controllers {
+		fmt.Fprintf(&buf, "%d:%s:%s\n", i, name, dir)
+		i--
+	}
+	return newStaticFile(buf.String())
+}

pkg/sentry/fsimpl/proc/tasks.go

@@ -54,11 +54,16 @@ type tasksInode struct {
 	// Linux. So handle them outside of OrderedChildren.
 	selfSymlink       *vfs.Dentry
 	threadSelfSymlink *vfs.Dentry
+
+	// cgroupControllers is a map of controller name to directory in the
+	// cgroup hierarchy. These controllers are immutable and will be listed
+	// in /proc/pid/cgroup if not nil.
+	cgroupControllers map[string]string
 }
 
 var _ kernfs.Inode = (*tasksInode)(nil)
 
-func newTasksInode(inoGen InoGenerator, k *kernel.Kernel, pidns *kernel.PIDNamespace) (*tasksInode, *kernfs.Dentry) {
+func newTasksInode(inoGen InoGenerator, k *kernel.Kernel, pidns *kernel.PIDNamespace, cgroupControllers map[string]string) (*tasksInode, *kernfs.Dentry) {
 	root := auth.NewRootCredentials(pidns.UserNamespace())
 	contents := map[string]*kernfs.Dentry{
 		"cpuinfo": newDentry(root, inoGen.NextIno(), 0444, newStaticFile(cpuInfoData(k))),
@@ -78,6 +83,7 @@ func newTasksInode(inoGen InoGenerator, k *kernel.Kernel, pidns *kernel.PIDNames
 		inoGen:            inoGen,
 		selfSymlink:       newSelfSymlink(root, inoGen.NextIno(), 0444, pidns).VFSDentry(),
 		threadSelfSymlink: newThreadSelfSymlink(root, inoGen.NextIno(), 0444, pidns).VFSDentry(),
+		cgroupControllers: cgroupControllers,
 	}
 	inode.InodeAttrs.Init(root, inoGen.NextIno(), linux.ModeDirectory|0555)
 
@@ -111,7 +117,7 @@ func (i *tasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, erro
 		return nil, syserror.ENOENT
 	}
 
-	taskDentry := newTaskInode(i.inoGen, task, i.pidns, true)
+	taskDentry := newTaskInode(i.inoGen, task, i.pidns, true, i.cgroupControllers)
 	return taskDentry.VFSDentry(), nil
 }
 

pkg/sentry/fsimpl/proc/tasks_test.go

@@ -87,6 +87,7 @@ func checkTasksStaticFiles(gots []vfs.Dirent) ([]vfs.Dirent, error) {
 func checkTaskStaticFiles(gots []vfs.Dirent) ([]vfs.Dirent, error) {
 	wants := map[string]vfs.Dirent{
 		"auxv":    {Type: linux.DT_REG},
+		"cgroup":  {Type: linux.DT_REG},
 		"cmdline": {Type: linux.DT_REG},
 		"comm":    {Type: linux.DT_REG},
 		"environ": {Type: linux.DT_REG},
@@ -145,7 +146,15 @@ func setup() (context.Context, *vfs.VirtualFilesystem, vfs.VirtualDentry, error)
 	vfsObj.MustRegisterFilesystemType("procfs", &procFSType{}, &vfs.RegisterFilesystemTypeOptions{
 		AllowUserMount: true,
 	})
-	mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "procfs", &vfs.GetFilesystemOptions{})
+	fsOpts := vfs.GetFilesystemOptions{
+		InternalData: &InternalData{
+			Cgroups: map[string]string{
+				"cpuset": "/foo/cpuset",
+				"memory": "/foo/memory",
+			},
+		},
+	}
+	mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "procfs", &fsOpts)
 	if err != nil {
 		return nil, nil, vfs.VirtualDentry{}, fmt.Errorf("NewMountNamespace(): %v", err)
 	}