diff options
Diffstat (limited to 'pkg/sentry/kernel/kernel.go')
-rw-r--r-- | pkg/sentry/kernel/kernel.go | 378 |
1 files changed, 320 insertions, 58 deletions
diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index 3ee760ba2..5efeb3767 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -43,11 +43,17 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/eventchannel" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" - "gvisor.dev/gvisor/pkg/sentry/fs/timerfd" + oldtimerfd "gvisor.dev/gvisor/pkg/sentry/fs/timerfd" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/pipefs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/sockfs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/timerfd" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" "gvisor.dev/gvisor/pkg/sentry/hostcpu" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -71,6 +77,10 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" ) +// VFS2Enabled is set to true when VFS2 is enabled. Added as a global for allow +// easy access everywhere. To be removed once VFS2 becomes the default. +var VFS2Enabled = false + // Kernel represents an emulated Linux kernel. It must be initialized by calling // Init() or LoadFrom(). // @@ -105,7 +115,7 @@ type Kernel struct { timekeeper *Timekeeper tasks *TaskSet rootUserNamespace *auth.UserNamespace - networkStack inet.Stack `state:"nosave"` + rootNetworkNamespace *inet.Namespace applicationCores uint useHostCores bool extraAuxv []arch.AuxEntry @@ -238,6 +248,33 @@ type Kernel struct { // SpecialOpts contains special kernel options. SpecialOpts + + // VFS keeps the filesystem state used across the kernel. + vfs vfs.VirtualFilesystem + + // hostMount is the Mount used for file descriptors that were imported + // from the host. + hostMount *vfs.Mount + + // pipeMount is the Mount used for pipes created by the pipe() and pipe2() + // syscalls (as opposed to named pipes created by mknod()). + pipeMount *vfs.Mount + + // shmMount is the Mount used for anonymous files created by the + // memfd_create() syscalls. It is analagous to Linux's shm_mnt. + shmMount *vfs.Mount + + // socketMount is the Mount used for sockets created by the socket() and + // socketpair() syscalls. There are several cases where a socket dentry will + // not be contained in socketMount: + // 1. Socket files created by mknod() + // 2. Socket fds imported from the host (Kernel.hostMount is used for these) + // 3. Socket files created by binding Unix sockets to a file path + socketMount *vfs.Mount + + // If set to true, report address space activation waits as if the task is in + // external wait so that the watchdog doesn't report the task stuck. + SleepForAddressSpaceActivation bool } // InitKernelArgs holds arguments to Init. @@ -251,8 +288,9 @@ type InitKernelArgs struct { // RootUserNamespace is the root user namespace. RootUserNamespace *auth.UserNamespace - // NetworkStack is the TCP/IP network stack. NetworkStack may be nil. - NetworkStack inet.Stack + // RootNetworkNamespace is the root network namespace. If nil, no networking + // will be available. + RootNetworkNamespace *inet.Namespace // ApplicationCores is the number of logical CPUs visible to sandboxed // applications. The set of logical CPU IDs is [0, ApplicationCores); thus @@ -297,6 +335,9 @@ func (k *Kernel) Init(args InitKernelArgs) error { if args.Timekeeper == nil { return fmt.Errorf("Timekeeper is nil") } + if args.Timekeeper.clocks == nil { + return fmt.Errorf("Must call Timekeeper.SetClocks() before Kernel.Init()") + } if args.RootUserNamespace == nil { return fmt.Errorf("RootUserNamespace is nil") } @@ -311,7 +352,10 @@ func (k *Kernel) Init(args InitKernelArgs) error { k.rootUTSNamespace = args.RootUTSNamespace k.rootIPCNamespace = args.RootIPCNamespace k.rootAbstractSocketNamespace = args.RootAbstractSocketNamespace - k.networkStack = args.NetworkStack + k.rootNetworkNamespace = args.RootNetworkNamespace + if k.rootNetworkNamespace == nil { + k.rootNetworkNamespace = inet.NewRootNamespace(nil, nil) + } k.applicationCores = args.ApplicationCores if args.UseHostCores { k.useHostCores = true @@ -331,6 +375,47 @@ func (k *Kernel) Init(args InitKernelArgs) error { k.monotonicClock = &timekeeperClock{tk: args.Timekeeper, c: sentrytime.Monotonic} k.futexes = futex.NewManager() k.netlinkPorts = port.New() + + if VFS2Enabled { + if err := k.vfs.Init(); err != nil { + return fmt.Errorf("failed to initialize VFS: %v", err) + } + + pipeFilesystem, err := pipefs.NewFilesystem(&k.vfs) + if err != nil { + return fmt.Errorf("failed to create pipefs filesystem: %v", err) + } + defer pipeFilesystem.DecRef() + pipeMount, err := k.vfs.NewDisconnectedMount(pipeFilesystem, nil, &vfs.MountOptions{}) + if err != nil { + return fmt.Errorf("failed to create pipefs mount: %v", err) + } + k.pipeMount = pipeMount + + tmpfsFilesystem, tmpfsRoot, err := tmpfs.NewFilesystem(k.SupervisorContext(), &k.vfs, auth.NewRootCredentials(k.rootUserNamespace)) + if err != nil { + return fmt.Errorf("failed to create tmpfs filesystem: %v", err) + } + defer tmpfsFilesystem.DecRef() + defer tmpfsRoot.DecRef() + shmMount, err := k.vfs.NewDisconnectedMount(tmpfsFilesystem, tmpfsRoot, &vfs.MountOptions{}) + if err != nil { + return fmt.Errorf("failed to create tmpfs mount: %v", err) + } + k.shmMount = shmMount + + socketFilesystem, err := sockfs.NewFilesystem(&k.vfs) + if err != nil { + return fmt.Errorf("failed to create sockfs filesystem: %v", err) + } + defer socketFilesystem.DecRef() + socketMount, err := k.vfs.NewDisconnectedMount(socketFilesystem, nil, &vfs.MountOptions{}) + if err != nil { + return fmt.Errorf("failed to create sockfs mount: %v", err) + } + k.socketMount = socketMount + } + return nil } @@ -450,6 +535,11 @@ func (k *Kernel) flushMountSourceRefs() error { // // Precondition: Must be called with the kernel paused. func (ts *TaskSet) forEachFDPaused(f func(*fs.File, *vfs.FileDescription) error) (err error) { + // TODO(gvisor.dev/issue/1663): Add save support for VFS2. + if VFS2Enabled { + return nil + } + ts.mu.RLock() defer ts.mu.RUnlock() for t := range ts.Root.tids { @@ -467,7 +557,7 @@ func (ts *TaskSet) forEachFDPaused(f func(*fs.File, *vfs.FileDescription) error) } func (ts *TaskSet) flushWritesToFiles(ctx context.Context) error { - // TODO(gvisor.dev/issues/1663): Add save support for VFS2. + // TODO(gvisor.dev/issue/1663): Add save support for VFS2. return ts.forEachFDPaused(func(file *fs.File, _ *vfs.FileDescription) error { if flags := file.Flags(); !flags.Write { return nil @@ -516,17 +606,32 @@ func (k *Kernel) invalidateUnsavableMappings(ctx context.Context) error { } func (ts *TaskSet) unregisterEpollWaiters() { + // TODO(gvisor.dev/issue/1663): Add save support for VFS2. + if VFS2Enabled { + return + } + ts.mu.RLock() defer ts.mu.RUnlock() + + // Tasks that belong to the same process could potentially point to the + // same FDTable. So we retain a map of processed ones to avoid + // processing the same FDTable multiple times. + processed := make(map[*FDTable]struct{}) for t := range ts.Root.tids { // We can skip locking Task.mu here since the kernel is paused. - if t.fdTable != nil { - t.fdTable.forEach(func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { - if e, ok := file.FileOperations.(*epoll.EventPoll); ok { - e.UnregisterEpollWaiters() - } - }) + if t.fdTable == nil { + continue } + if _, ok := processed[t.fdTable]; ok { + continue + } + t.fdTable.forEach(func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { + if e, ok := file.FileOperations.(*epoll.EventPoll); ok { + e.UnregisterEpollWaiters() + } + }) + processed[t.fdTable] = struct{}{} } } @@ -534,8 +639,6 @@ func (ts *TaskSet) unregisterEpollWaiters() { func (k *Kernel) LoadFrom(r io.Reader, net inet.Stack, clocks sentrytime.Clocks) error { loadStart := time.Now() - k.networkStack = net - initAppCores := k.applicationCores // Load the pre-saved CPUID FeatureSet. @@ -566,6 +669,10 @@ func (k *Kernel) LoadFrom(r io.Reader, net inet.Stack, clocks sentrytime.Clocks) log.Infof("Kernel load stats: %s", &stats) log.Infof("Kernel load took [%s].", time.Since(kernelStart)) + // rootNetworkNamespace should be populated after loading the state file. + // Restore the root network stack. + k.rootNetworkNamespace.RestoreRootStack(net) + // Load the memory file's state. memoryStart := time.Now() if err := k.mf.LoadFrom(k.SupervisorContext(), r); err != nil { @@ -624,7 +731,7 @@ type CreateProcessArgs struct { // File is a passed host FD pointing to a file to load as the init binary. // // This is checked if and only if Filename is "". - File *fs.File + File fsbridge.File // Argvv is a list of arguments. Argv []string @@ -673,6 +780,13 @@ type CreateProcessArgs struct { // increment it). MountNamespace *fs.MountNamespace + // MountNamespaceVFS2 optionally contains the mount namespace for this + // process. If nil, the init process's mount namespace is used. + // + // Anyone setting MountNamespaceVFS2 must donate a reference (i.e. + // increment it). + MountNamespaceVFS2 *vfs.MountNamespace + // ContainerID is the container that the process belongs to. ContainerID string } @@ -711,13 +825,26 @@ func (ctx *createProcessContext) Value(key interface{}) interface{} { return ctx.args.Credentials case fs.CtxRoot: if ctx.args.MountNamespace != nil { - // MountNamespace.Root() will take a reference on the root - // dirent for us. + // MountNamespace.Root() will take a reference on the root dirent for us. return ctx.args.MountNamespace.Root() } return nil + case vfs.CtxRoot: + if ctx.args.MountNamespaceVFS2 == nil { + return nil + } + // MountNamespaceVFS2.Root() takes a reference on the root dirent for us. + return ctx.args.MountNamespaceVFS2.Root() + case vfs.CtxMountNamespace: + if ctx.k.globalInit == nil { + return nil + } + // MountNamespaceVFS2 takes a reference for us. + return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter + case inet.CtxStack: + return ctx.k.RootNetworkNamespace().Stack() case ktime.CtxRealtimeClock: return ctx.k.RealtimeClock() case limits.CtxLimits: @@ -757,34 +884,77 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, defer k.extMu.Unlock() log.Infof("EXEC: %v", args.Argv) - // Grab the mount namespace. - mounts := args.MountNamespace - if mounts == nil { - mounts = k.GlobalInit().Leader().MountNamespace() - mounts.IncRef() - } - - tg := k.NewThreadGroup(mounts, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits) ctx := args.NewContext(k) - // Get the root directory from the MountNamespace. - root := mounts.Root() - // The call to newFSContext below will take a reference on root, so we - // don't need to hold this one. - defer root.DecRef() - - // Grab the working directory. - remainingTraversals := uint(args.MaxSymlinkTraversals) - wd := root // Default. - if args.WorkingDirectory != "" { - var err error - wd, err = mounts.FindInode(ctx, root, nil, args.WorkingDirectory, &remainingTraversals) - if err != nil { - return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + var ( + opener fsbridge.Lookup + fsContext *FSContext + mntns *fs.MountNamespace + ) + + if VFS2Enabled { + mntnsVFS2 := args.MountNamespaceVFS2 + if mntnsVFS2 == nil { + // MountNamespaceVFS2 adds a reference to the namespace, which is + // transferred to the new process. + mntnsVFS2 = k.GlobalInit().Leader().MountNamespaceVFS2() + } + // Get the root directory from the MountNamespace. + root := args.MountNamespaceVFS2.Root() + // The call to newFSContext below will take a reference on root, so we + // don't need to hold this one. + defer root.DecRef() + + // Grab the working directory. + wd := root // Default. + if args.WorkingDirectory != "" { + pop := vfs.PathOperation{ + Root: root, + Start: wd, + Path: fspath.Parse(args.WorkingDirectory), + FollowFinalSymlink: true, + } + var err error + wd, err = k.VFS().GetDentryAt(ctx, args.Credentials, &pop, &vfs.GetDentryOptions{ + CheckSearchable: true, + }) + if err != nil { + return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + } + defer wd.DecRef() + } + opener = fsbridge.NewVFSLookup(mntnsVFS2, root, wd) + fsContext = NewFSContextVFS2(root, wd, args.Umask) + + } else { + mntns = args.MountNamespace + if mntns == nil { + mntns = k.GlobalInit().Leader().MountNamespace() + mntns.IncRef() } - defer wd.DecRef() + // Get the root directory from the MountNamespace. + root := mntns.Root() + // The call to newFSContext below will take a reference on root, so we + // don't need to hold this one. + defer root.DecRef() + + // Grab the working directory. + remainingTraversals := args.MaxSymlinkTraversals + wd := root // Default. + if args.WorkingDirectory != "" { + var err error + wd, err = mntns.FindInode(ctx, root, nil, args.WorkingDirectory, &remainingTraversals) + if err != nil { + return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + } + defer wd.DecRef() + } + opener = fsbridge.NewFSLookup(mntns, root, wd) + fsContext = newFSContext(root, wd, args.Umask) } + tg := k.NewThreadGroup(mntns, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits) + // Check which file to start from. switch { case args.Filename != "": @@ -805,11 +975,9 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, } // Create a fresh task context. - remainingTraversals = uint(args.MaxSymlinkTraversals) + remainingTraversals := args.MaxSymlinkTraversals loadArgs := loader.LoadArgs{ - Mounts: mounts, - Root: root, - WorkingDirectory: wd, + Opener: opener, RemainingTraversals: &remainingTraversals, ResolveFinal: true, Filename: args.Filename, @@ -834,13 +1002,15 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, Kernel: k, ThreadGroup: tg, TaskContext: tc, - FSContext: newFSContext(root, wd, args.Umask), + FSContext: fsContext, FDTable: args.FDTable, Credentials: args.Credentials, + NetworkNamespace: k.RootNetworkNamespace(), AllowedCPUMask: sched.NewFullCPUSet(k.applicationCores), UTSNamespace: args.UTSNamespace, IPCNamespace: args.IPCNamespace, AbstractSocketNamespace: args.AbstractSocketNamespace, + MountNamespaceVFS2: args.MountNamespaceVFS2, ContainerID: args.ContainerID, } t, err := k.tasks.NewTask(config) @@ -923,9 +1093,15 @@ func (k *Kernel) pauseTimeLocked() { // This means we'll iterate FDTables shared by multiple tasks repeatedly, // but ktime.Timer.Pause is idempotent so this is harmless. if t.fdTable != nil { - t.fdTable.forEach(func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { - if tfd, ok := file.FileOperations.(*timerfd.TimerOperations); ok { - tfd.PauseTimer() + t.fdTable.forEach(func(_ int32, file *fs.File, fd *vfs.FileDescription, _ FDFlags) { + if VFS2Enabled { + if tfd, ok := fd.Impl().(*timerfd.TimerFileDescription); ok { + tfd.PauseTimer() + } + } else { + if tfd, ok := file.FileOperations.(*oldtimerfd.TimerOperations); ok { + tfd.PauseTimer() + } } }) } @@ -953,9 +1129,15 @@ func (k *Kernel) resumeTimeLocked() { } } if t.fdTable != nil { - t.fdTable.forEach(func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { - if tfd, ok := file.FileOperations.(*timerfd.TimerOperations); ok { - tfd.ResumeTimer() + t.fdTable.forEach(func(_ int32, file *fs.File, fd *vfs.FileDescription, _ FDFlags) { + if VFS2Enabled { + if tfd, ok := fd.Impl().(*timerfd.TimerFileDescription); ok { + tfd.ResumeTimer() + } + } else { + if tfd, ok := file.FileOperations.(*oldtimerfd.TimerOperations); ok { + tfd.ResumeTimer() + } } }) } @@ -1100,6 +1282,14 @@ func (k *Kernel) SendExternalSignal(info *arch.SignalInfo, context string) { k.sendExternalSignal(info, context) } +// SendExternalSignalThreadGroup injects a signal into an specific ThreadGroup. +// This function doesn't skip signals like SendExternalSignal does. +func (k *Kernel) SendExternalSignalThreadGroup(tg *ThreadGroup, info *arch.SignalInfo) error { + k.extMu.Lock() + defer k.extMu.Unlock() + return tg.SendSignal(info) +} + // SendContainerSignal sends the given signal to all processes inside the // namespace that match the given container ID. func (k *Kernel) SendContainerSignal(cid string, info *arch.SignalInfo) error { @@ -1178,10 +1368,9 @@ func (k *Kernel) RootAbstractSocketNamespace() *AbstractSocketNamespace { return k.rootAbstractSocketNamespace } -// NetworkStack returns the network stack. NetworkStack may return nil if no -// network stack is available. -func (k *Kernel) NetworkStack() inet.Stack { - return k.networkStack +// RootNetworkNamespace returns the root network namespace, always non-nil. +func (k *Kernel) RootNetworkNamespace() *inet.Namespace { + return k.rootNetworkNamespace } // GlobalInit returns the thread group with ID 1 in the root PID namespace, or @@ -1311,9 +1500,10 @@ func (k *Kernel) SupervisorContext() context.Context { // +stateify savable type SocketEntry struct { socketEntry - k *Kernel - Sock *refs.WeakRef - ID uint64 // Socket table entry number. + k *Kernel + Sock *refs.WeakRef + SockVFS2 *vfs.FileDescription + ID uint64 // Socket table entry number. } // WeakRefGone implements refs.WeakRefUser.WeakRefGone. @@ -1336,7 +1526,30 @@ func (k *Kernel) RecordSocket(sock *fs.File) { k.extMu.Unlock() } +// RecordSocketVFS2 adds a VFS2 socket to the system-wide socket table for +// tracking. +// +// Precondition: Caller must hold a reference to sock. +// +// Note that the socket table will not hold a reference on the +// vfs.FileDescription, because we do not support weak refs on VFS2 files. +func (k *Kernel) RecordSocketVFS2(sock *vfs.FileDescription) { + k.extMu.Lock() + id := k.nextSocketEntry + k.nextSocketEntry++ + s := &SocketEntry{ + k: k, + ID: id, + SockVFS2: sock, + } + k.sockets.PushBack(s) + k.extMu.Unlock() +} + // ListSockets returns a snapshot of all sockets. +// +// Callers of ListSockets() in VFS2 should use SocketEntry.SockVFS2.TryIncRef() +// to get a reference on a socket in the table. func (k *Kernel) ListSockets() []*SocketEntry { k.extMu.Lock() var socks []*SocketEntry @@ -1378,8 +1591,24 @@ func (ctx supervisorContext) Value(key interface{}) interface{} { return ctx.k.globalInit.mounts.Root() } return nil + case vfs.CtxRoot: + if ctx.k.globalInit == nil { + return vfs.VirtualDentry{} + } + mntns := ctx.k.GlobalInit().Leader().MountNamespaceVFS2() + defer mntns.DecRef() + // Root() takes a reference on the root dirent for us. + return mntns.Root() + case vfs.CtxMountNamespace: + if ctx.k.globalInit == nil { + return nil + } + // MountNamespaceVFS2() takes a reference for us. + return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter + case inet.CtxStack: + return ctx.k.RootNetworkNamespace().Stack() case ktime.CtxRealtimeClock: return ctx.k.RealtimeClock() case limits.CtxLimits: @@ -1423,3 +1652,36 @@ func (k *Kernel) EmitUnimplementedEvent(ctx context.Context) { Registers: t.Arch().StateData().Proto(), }) } + +// VFS returns the virtual filesystem for the kernel. +func (k *Kernel) VFS() *vfs.VirtualFilesystem { + return &k.vfs +} + +// SetHostMount sets the hostfs mount. +func (k *Kernel) SetHostMount(mnt *vfs.Mount) { + if k.hostMount != nil { + panic("Kernel.hostMount cannot be set more than once") + } + k.hostMount = mnt +} + +// HostMount returns the hostfs mount. +func (k *Kernel) HostMount() *vfs.Mount { + return k.hostMount +} + +// PipeMount returns the pipefs mount. +func (k *Kernel) PipeMount() *vfs.Mount { + return k.pipeMount +} + +// ShmMount returns the tmpfs mount. +func (k *Kernel) ShmMount() *vfs.Mount { + return k.shmMount +} + +// SocketMount returns the sockfs mount. +func (k *Kernel) SocketMount() *vfs.Mount { + return k.socketMount +} |