Merge branch 'master' into iptables

12 files changed, 143 insertions, 42 deletions

diff --git a/pkg/sentry/kernel/fd_table.go b/pkg/sentry/kernel/fd_table.go
index 23b88f7a6..58001d56c 100644
--- a/pkg/sentry/kernel/fd_table.go
+++ b/pkg/sentry/kernel/fd_table.go
@@ -296,6 +296,50 @@ func (f *FDTable) NewFDs(ctx context.Context, fd int32, files []*fs.File, flags
 	return fds, nil
 }
 
+// NewFDVFS2 allocates a file descriptor greater than or equal to minfd for
+// the given file description. If it succeeds, it takes a reference on file.
+func (f *FDTable) NewFDVFS2(ctx context.Context, minfd int32, file *vfs.FileDescription, flags FDFlags) (int32, error) {
+	if minfd < 0 {
+		// Don't accept negative FDs.
+		return -1, syscall.EINVAL
+	}
+
+	// Default limit.
+	end := int32(math.MaxInt32)
+
+	// Ensure we don't get past the provided limit.
+	if limitSet := limits.FromContext(ctx); limitSet != nil {
+		lim := limitSet.Get(limits.NumberOfFiles)
+		if lim.Cur != limits.Infinity {
+			end = int32(lim.Cur)
+		}
+		if minfd >= end {
+			return -1, syscall.EMFILE
+		}
+	}
+
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	// From f.next to find available fd.
+	fd := minfd
+	if fd < f.next {
+		fd = f.next
+	}
+	for fd < end {
+		if d, _, _ := f.get(fd); d == nil {
+			f.setVFS2(fd, file, flags)
+			if fd == f.next {
+				// Update next search start position.
+				f.next = fd + 1
+			}
+			return fd, nil
+		}
+		fd++
+	}
+	return -1, syscall.EMFILE
+}
+
 // NewFDAt sets the file reference for the given FD. If there is an active
 // reference for that FD, the ref count for that existing reference is
 // decremented.
@@ -316,9 +360,6 @@ func (f *FDTable) newFDAt(ctx context.Context, fd int32, file *fs.File, fileVFS2
 		return syscall.EBADF
 	}
 
-	f.mu.Lock()
-	defer f.mu.Unlock()
-
 	// Check the limit for the provided file.
 	if limitSet := limits.FromContext(ctx); limitSet != nil {
 		if lim := limitSet.Get(limits.NumberOfFiles); lim.Cur != limits.Infinity && uint64(fd) >= lim.Cur {
@@ -327,6 +368,8 @@ func (f *FDTable) newFDAt(ctx context.Context, fd int32, file *fs.File, fileVFS2
 	}
 
 	// Install the entry.
+	f.mu.Lock()
+	defer f.mu.Unlock()
 	f.setAll(fd, file, fileVFS2, flags)
 	return nil
 }
diff --git a/pkg/sentry/kernel/fs_context.go b/pkg/sentry/kernel/fs_context.go
index 7218aa24e..47f78df9a 100644
--- a/pkg/sentry/kernel/fs_context.go
+++ b/pkg/sentry/kernel/fs_context.go
@@ -244,6 +244,28 @@ func (f *FSContext) SetRootDirectory(d *fs.Dirent) {
 	old.DecRef()
 }
 
+// SetRootDirectoryVFS2 sets the root directory. It takes a reference on vd.
+//
+// This is not a valid call after free.
+func (f *FSContext) SetRootDirectoryVFS2(vd vfs.VirtualDentry) {
+	if !vd.Ok() {
+		panic("FSContext.SetRootDirectoryVFS2 called with zero-value VirtualDentry")
+	}
+
+	f.mu.Lock()
+
+	if !f.rootVFS2.Ok() {
+		f.mu.Unlock()
+		panic(fmt.Sprintf("FSContext.SetRootDirectoryVFS2(%v)) called after destroy", vd))
+	}
+
+	old := f.rootVFS2
+	vd.IncRef()
+	f.rootVFS2 = vd
+	f.mu.Unlock()
+	old.DecRef()
+}
+
 // Umask returns the current umask.
 func (f *FSContext) Umask() uint {
 	f.mu.Lock()
diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go
index 7da0368f1..8b76750e9 100644
--- a/pkg/sentry/kernel/kernel.go
+++ b/pkg/sentry/kernel/kernel.go
@@ -111,7 +111,7 @@ 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
@@ -247,6 +247,10 @@ type Kernel struct {
 
 	// VFS keeps the filesystem state used across the kernel.
 	vfs vfs.VirtualFilesystem
+
+	// 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.
@@ -260,8 +264,9 @@ 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
@@ -320,7 +325,10 @@ 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
@@ -543,8 +551,6 @@ 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.
@@ -575,6 +581,10 @@ 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 {
@@ -905,6 +915,7 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID,
 		FSContext:               fsContext,
 		FDTable:                 args.FDTable,
 		Credentials:             args.Credentials,
+		NetworkNamespace:        k.RootNetworkNamespace(),
 		AllowedCPUMask:          sched.NewFullCPUSet(k.applicationCores),
 		UTSNamespace:            args.UTSNamespace,
 		IPCNamespace:            args.IPCNamespace,
@@ -1255,10 +1266,9 @@ 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
diff --git a/pkg/sentry/kernel/rseq.go b/pkg/sentry/kernel/rseq.go
index 18416643b..ded95f532 100644
--- a/pkg/sentry/kernel/rseq.go
+++ b/pkg/sentry/kernel/rseq.go
@@ -304,7 +304,7 @@ func (t *Task) rseqAddrInterrupt() {
 	}
 
 	var cs linux.RSeqCriticalSection
-	if _, err := cs.CopyIn(t, critAddr); err != nil {
+	if err := cs.CopyIn(t, critAddr); err != nil {
 		t.Debugf("Failed to copy critical section from %#x for rseq: %v", critAddr, err)
 		t.forceSignal(linux.SIGSEGV, false /* unconditional */)
 		t.SendSignal(SignalInfoPriv(linux.SIGSEGV))
diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go
index a3443ff21..2cee2e6ed 100644
--- a/pkg/sentry/kernel/task.go
+++ b/pkg/sentry/kernel/task.go
@@ -486,13 +486,10 @@ type Task struct {
 	numaPolicy   int32
 	numaNodeMask uint64
 
-	// If netns is true, the task is in a non-root network namespace. Network
-	// namespaces aren't currently implemented in full; being in a network
-	// namespace simply prevents the task from observing any network devices
-	// (including loopback) or using abstract socket addresses (see unix(7)).
+	// netns is the task's network namespace. netns is never nil.
 	//
-	// netns is protected by mu. netns is owned by the task goroutine.
-	netns bool
+	// netns is protected by mu.
+	netns *inet.Namespace
 
 	// If rseqPreempted is true, before the next call to p.Switch(),
 	// interrupt rseq critical regions as defined by rseqAddr and
@@ -792,6 +789,15 @@ func (t *Task) NewFDFrom(fd int32, file *fs.File, flags FDFlags) (int32, error)
 	return fds[0], nil
 }
 
+// NewFDFromVFS2 is a convenience wrapper for t.FDTable().NewFDVFS2.
+//
+// This automatically passes the task as the context.
+//
+// Precondition: same as FDTable.Get.
+func (t *Task) NewFDFromVFS2(fd int32, file *vfs.FileDescription, flags FDFlags) (int32, error) {
+	return t.fdTable.NewFDVFS2(t, fd, file, flags)
+}
+
 // NewFDAt is a convenience wrapper for t.FDTable().NewFDAt.
 //
 // This automatically passes the task as the context.
@@ -801,6 +807,15 @@ func (t *Task) NewFDAt(fd int32, file *fs.File, flags FDFlags) error {
 	return t.fdTable.NewFDAt(t, fd, file, flags)
 }
 
+// NewFDAtVFS2 is a convenience wrapper for t.FDTable().NewFDAtVFS2.
+//
+// This automatically passes the task as the context.
+//
+// Precondition: same as FDTable.
+func (t *Task) NewFDAtVFS2(fd int32, file *vfs.FileDescription, flags FDFlags) error {
+	return t.fdTable.NewFDAtVFS2(t, fd, file, flags)
+}
+
 // WithMuLocked executes f with t.mu locked.
 func (t *Task) WithMuLocked(f func(*Task)) {
 	t.mu.Lock()
diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go
index ba74b4c1c..78866f280 100644
--- a/pkg/sentry/kernel/task_clone.go
+++ b/pkg/sentry/kernel/task_clone.go
@@ -17,6 +17,7 @@ package kernel
 import (
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/bpf"
+	"gvisor.dev/gvisor/pkg/sentry/inet"
 	"gvisor.dev/gvisor/pkg/syserror"
 	"gvisor.dev/gvisor/pkg/usermem"
 )
@@ -54,8 +55,7 @@ type SharingOptions struct {
 	NewUserNamespace bool
 
 	// If NewNetworkNamespace is true, the task should have an independent
-	// network namespace. (Note that network namespaces are not really
-	// implemented; see comment on Task.netns for details.)
+	// network namespace.
 	NewNetworkNamespace bool
 
 	// If NewFiles is true, the task should use an independent file descriptor
@@ -199,6 +199,11 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) {
 		ipcns = NewIPCNamespace(userns)
 	}
 
+	netns := t.NetworkNamespace()
+	if opts.NewNetworkNamespace {
+		netns = inet.NewNamespace(netns)
+	}
+
 	// TODO(b/63601033): Implement CLONE_NEWNS.
 	mntnsVFS2 := t.mountNamespaceVFS2
 	if mntnsVFS2 != nil {
@@ -268,7 +273,7 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) {
 		FDTable:                 fdTable,
 		Credentials:             creds,
 		Niceness:                t.Niceness(),
-		NetworkNamespaced:       t.netns,
+		NetworkNamespace:        netns,
 		AllowedCPUMask:          t.CPUMask(),
 		UTSNamespace:            utsns,
 		IPCNamespace:            ipcns,
@@ -283,9 +288,6 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) {
 	} else {
 		cfg.InheritParent = t
 	}
-	if opts.NewNetworkNamespace {
-		cfg.NetworkNamespaced = true
-	}
 	nt, err := t.tg.pidns.owner.NewTask(cfg)
 	if err != nil {
 		if opts.NewThreadGroup {
@@ -482,7 +484,7 @@ func (t *Task) Unshare(opts *SharingOptions) error {
 			t.mu.Unlock()
 			return syserror.EPERM
 		}
-		t.netns = true
+		t.netns = inet.NewNamespace(t.netns)
 	}
 	if opts.NewUTSNamespace {
 		if !haveCapSysAdmin {
diff --git a/pkg/sentry/kernel/task_context.go b/pkg/sentry/kernel/task_context.go
index 2be982684..0158b1788 100644
--- a/pkg/sentry/kernel/task_context.go
+++ b/pkg/sentry/kernel/task_context.go
@@ -140,7 +140,7 @@ func (k *Kernel) LoadTaskImage(ctx context.Context, args loader.LoadArgs) (*Task
 	}
 
 	// Prepare a new user address space to load into.
-	m := mm.NewMemoryManager(k, k)
+	m := mm.NewMemoryManager(k, k, k.SleepForAddressSpaceActivation)
 	defer m.DecUsers(ctx)
 	args.MemoryManager = m
 
diff --git a/pkg/sentry/kernel/task_exec.go b/pkg/sentry/kernel/task_exec.go
index 8f57a34a6..00c425cca 100644
--- a/pkg/sentry/kernel/task_exec.go
+++ b/pkg/sentry/kernel/task_exec.go
@@ -220,7 +220,7 @@ func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState {
 	t.mu.Unlock()
 	t.unstopVforkParent()
 	// NOTE(b/30316266): All locks must be dropped prior to calling Activate.
-	t.MemoryManager().Activate()
+	t.MemoryManager().Activate(t)
 
 	t.ptraceExec(oldTID)
 	return (*runSyscallExit)(nil)
diff --git a/pkg/sentry/kernel/task_log.go b/pkg/sentry/kernel/task_log.go
index 6d737d3e5..eeccaa197 100644
--- a/pkg/sentry/kernel/task_log.go
+++ b/pkg/sentry/kernel/task_log.go
@@ -32,21 +32,21 @@ const (
 // Infof logs an formatted info message by calling log.Infof.
 func (t *Task) Infof(fmt string, v ...interface{}) {
 	if log.IsLogging(log.Info) {
-		log.Infof(t.logPrefix.Load().(string)+fmt, v...)
+		log.InfofAtDepth(1, t.logPrefix.Load().(string)+fmt, v...)
 	}
 }
 
 // Warningf logs a warning string by calling log.Warningf.
 func (t *Task) Warningf(fmt string, v ...interface{}) {
 	if log.IsLogging(log.Warning) {
-		log.Warningf(t.logPrefix.Load().(string)+fmt, v...)
+		log.WarningfAtDepth(1, t.logPrefix.Load().(string)+fmt, v...)
 	}
 }
 
 // Debugf creates a debug string that includes the task ID.
 func (t *Task) Debugf(fmt string, v ...interface{}) {
 	if log.IsLogging(log.Debug) {
-		log.Debugf(t.logPrefix.Load().(string)+fmt, v...)
+		log.DebugfAtDepth(1, t.logPrefix.Load().(string)+fmt, v...)
 	}
 }
 
diff --git a/pkg/sentry/kernel/task_net.go b/pkg/sentry/kernel/task_net.go
index 172a31e1d..f7711232c 100644
--- a/pkg/sentry/kernel/task_net.go
+++ b/pkg/sentry/kernel/task_net.go
@@ -22,14 +22,23 @@ import (
 func (t *Task) IsNetworkNamespaced() bool {
 	t.mu.Lock()
 	defer t.mu.Unlock()
-	return t.netns
+	return !t.netns.IsRoot()
 }
 
 // NetworkContext returns the network stack used by the task. NetworkContext
 // may return nil if no network stack is available.
+//
+// TODO(gvisor.dev/issue/1833): Migrate callers of this method to
+// NetworkNamespace().
 func (t *Task) NetworkContext() inet.Stack {
-	if t.IsNetworkNamespaced() {
-		return nil
-	}
-	return t.k.networkStack
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.netns.Stack()
+}
+
+// NetworkNamespace returns the network namespace observed by the task.
+func (t *Task) NetworkNamespace() *inet.Namespace {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.netns
 }
diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go
index f9236a842..a5035bb7f 100644
--- a/pkg/sentry/kernel/task_start.go
+++ b/pkg/sentry/kernel/task_start.go
@@ -17,6 +17,7 @@ package kernel
 import (
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/sentry/arch"
+	"gvisor.dev/gvisor/pkg/sentry/inet"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/futex"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/sched"
@@ -65,9 +66,8 @@ type TaskConfig struct {
 	// Niceness is the niceness of the new task.
 	Niceness int
 
-	// If NetworkNamespaced is true, the new task should observe a non-root
-	// network namespace.
-	NetworkNamespaced bool
+	// NetworkNamespace is the network namespace to be used for the new task.
+	NetworkNamespace *inet.Namespace
 
 	// AllowedCPUMask contains the cpus that this task can run on.
 	AllowedCPUMask sched.CPUSet
@@ -133,7 +133,7 @@ func (ts *TaskSet) newTask(cfg *TaskConfig) (*Task, error) {
 		allowedCPUMask:     cfg.AllowedCPUMask.Copy(),
 		ioUsage:            &usage.IO{},
 		niceness:           cfg.Niceness,
-		netns:              cfg.NetworkNamespaced,
+		netns:              cfg.NetworkNamespace,
 		utsns:              cfg.UTSNamespace,
 		ipcns:              cfg.IPCNamespace,
 		abstractSockets:    cfg.AbstractSocketNamespace,
diff --git a/pkg/sentry/kernel/task_usermem.go b/pkg/sentry/kernel/task_usermem.go
index 2bf3ce8a8..b02044ad2 100644
--- a/pkg/sentry/kernel/task_usermem.go
+++ b/pkg/sentry/kernel/task_usermem.go
@@ -30,7 +30,7 @@ var MAX_RW_COUNT = int(usermem.Addr(math.MaxInt32).RoundDown())
 // Activate ensures that the task has an active address space.
 func (t *Task) Activate() {
 	if mm := t.MemoryManager(); mm != nil {
-		if err := mm.Activate(); err != nil {
+		if err := mm.Activate(t); err != nil {
 			panic("unable to activate mm: " + err.Error())
 		}
 	}