diff options
Diffstat (limited to 'pkg/sentry/socket/netlink/socket.go')
| -rw-r--r-- | pkg/sentry/socket/netlink/socket.go | 140 |
1 files changed, 137 insertions, 3 deletions
diff --git a/pkg/sentry/socket/netlink/socket.go b/pkg/sentry/socket/netlink/socket.go index d0aab293d..4a1b87a9a 100644 --- a/pkg/sentry/socket/netlink/socket.go +++ b/pkg/sentry/socket/netlink/socket.go @@ -27,7 +27,9 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/safemem" "gvisor.dev/gvisor/pkg/sentry/socket" "gvisor.dev/gvisor/pkg/sentry/socket/netlink/port" "gvisor.dev/gvisor/pkg/sentry/socket/unix" @@ -52,6 +54,8 @@ const ( maxSendBufferSize = 4 << 20 // 4MB ) +var errNoFilter = syserr.New("no filter attached", linux.ENOENT) + // netlinkSocketDevice is the netlink socket virtual device. var netlinkSocketDevice = device.NewAnonDevice() @@ -60,7 +64,7 @@ var netlinkSocketDevice = device.NewAnonDevice() // This implementation only supports userspace sending and receiving messages // to/from the kernel. // -// Socket implements socket.Socket. +// Socket implements socket.Socket and transport.Credentialer. // // +stateify savable type Socket struct { @@ -103,9 +107,19 @@ type Socket struct { // sendBufferSize is the send buffer "size". We don't actually have a // fixed buffer but only consume this many bytes. sendBufferSize uint32 + + // passcred indicates if this socket wants SCM credentials. + passcred bool + + // filter indicates that this socket has a BPF filter "installed". + // + // TODO(gvisor.dev/issue/1119): We don't actually support filtering, + // this is just bookkeeping for tracking add/remove. + filter bool } var _ socket.Socket = (*Socket)(nil) +var _ transport.Credentialer = (*Socket)(nil) // NewSocket creates a new Socket. func NewSocket(t *kernel.Task, skType linux.SockType, protocol Protocol) (*Socket, *syserr.Error) { @@ -171,6 +185,22 @@ func (s *Socket) EventUnregister(e *waiter.Entry) { s.ep.EventUnregister(e) } +// Passcred implements transport.Credentialer.Passcred. +func (s *Socket) Passcred() bool { + s.mu.Lock() + passcred := s.passcred + s.mu.Unlock() + return passcred +} + +// ConnectedPasscred implements transport.Credentialer.ConnectedPasscred. +func (s *Socket) ConnectedPasscred() bool { + // This socket is connected to the kernel, which doesn't need creds. + // + // This is arbitrary, as ConnectedPasscred on this type has no callers. + return false +} + // Ioctl implements fs.FileOperations.Ioctl. func (*Socket) Ioctl(context.Context, *fs.File, usermem.IO, arch.SyscallArguments) (uintptr, error) { // TODO(b/68878065): no ioctls supported. @@ -308,9 +338,20 @@ func (s *Socket) GetSockOpt(t *kernel.Task, level int, name int, outPtr usermem. // We don't have limit on receiving size. return int32(math.MaxInt32), nil + case linux.SO_PASSCRED: + if outLen < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + var passcred int32 + if s.Passcred() { + passcred = 1 + } + return passcred, nil + default: socket.GetSockOptEmitUnimplementedEvent(t, name) } + case linux.SOL_NETLINK: switch name { case linux.NETLINK_BROADCAST_ERROR, @@ -347,6 +388,7 @@ func (s *Socket) SetSockOpt(t *kernel.Task, level int, name int, opt []byte) *sy s.sendBufferSize = size s.mu.Unlock() return nil + case linux.SO_RCVBUF: if len(opt) < sizeOfInt32 { return syserr.ErrInvalidArgument @@ -354,6 +396,52 @@ func (s *Socket) SetSockOpt(t *kernel.Task, level int, name int, opt []byte) *sy // We don't have limit on receiving size. So just accept anything as // valid for compatibility. return nil + + case linux.SO_PASSCRED: + if len(opt) < sizeOfInt32 { + return syserr.ErrInvalidArgument + } + passcred := usermem.ByteOrder.Uint32(opt) + + s.mu.Lock() + s.passcred = passcred != 0 + s.mu.Unlock() + return nil + + case linux.SO_ATTACH_FILTER: + // TODO(gvisor.dev/issue/1119): We don't actually + // support filtering. If this socket can't ever send + // messages, then there is nothing to filter and we can + // advertise support. Otherwise, be conservative and + // return an error. + if s.protocol.CanSend() { + socket.SetSockOptEmitUnimplementedEvent(t, name) + return syserr.ErrProtocolNotAvailable + } + + s.mu.Lock() + s.filter = true + s.mu.Unlock() + return nil + + case linux.SO_DETACH_FILTER: + // TODO(gvisor.dev/issue/1119): See above. + if s.protocol.CanSend() { + socket.SetSockOptEmitUnimplementedEvent(t, name) + return syserr.ErrProtocolNotAvailable + } + + s.mu.Lock() + filter := s.filter + s.filter = false + s.mu.Unlock() + + if !filter { + return errNoFilter + } + + return nil + default: socket.SetSockOptEmitUnimplementedEvent(t, name) } @@ -416,6 +504,24 @@ func (s *Socket) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, have Peek: flags&linux.MSG_PEEK != 0, } + // If MSG_TRUNC is set with a zero byte destination then we still need + // to read the message and discard it, or in the case where MSG_PEEK is + // set, leave it be. In both cases the full message length must be + // returned. However, the memory manager for the destination will not read + // the endpoint if the destination is zero length. + // + // In order for the endpoint to be read when the destination size is zero, + // we must cause a read of the endpoint by using a separate fake zero + // length block sequence and calling the EndpointReader directly. + if trunc && dst.Addrs.NumBytes() == 0 { + // Perform a read to a zero byte block sequence. We can ignore the + // original destination since it was zero bytes. The length returned by + // ReadToBlocks is ignored and we return the full message length to comply + // with MSG_TRUNC. + _, err := r.ReadToBlocks(safemem.BlockSeqOf(safemem.BlockFromSafeSlice(make([]byte, 0)))) + return int(r.MsgSize), linux.MSG_TRUNC, from, fromLen, socket.ControlMessages{}, syserr.FromError(err) + } + if n, err := dst.CopyOutFrom(t, &r); err != syserror.ErrWouldBlock || flags&linux.MSG_DONTWAIT != 0 { var mflags int if n < int64(r.MsgSize) { @@ -464,6 +570,26 @@ func (s *Socket) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, _ }) } +// kernelSCM implements control.SCMCredentials with credentials that represent +// the kernel itself rather than a Task. +// +// +stateify savable +type kernelSCM struct{} + +// Equals implements transport.CredentialsControlMessage.Equals. +func (kernelSCM) Equals(oc transport.CredentialsControlMessage) bool { + _, ok := oc.(kernelSCM) + return ok +} + +// Credentials implements control.SCMCredentials.Credentials. +func (kernelSCM) Credentials(*kernel.Task) (kernel.ThreadID, auth.UID, auth.GID) { + return 0, auth.RootUID, auth.RootGID +} + +// kernelCreds is the concrete version of kernelSCM used in all creds. +var kernelCreds = &kernelSCM{} + // sendResponse sends the response messages in ms back to userspace. func (s *Socket) sendResponse(ctx context.Context, ms *MessageSet) *syserr.Error { // Linux combines multiple netlink messages into a single datagram. @@ -472,10 +598,15 @@ func (s *Socket) sendResponse(ctx context.Context, ms *MessageSet) *syserr.Error bufs = append(bufs, m.Finalize()) } + // All messages are from the kernel. + cms := transport.ControlMessages{ + Credentials: kernelCreds, + } + if len(bufs) > 0 { // RecvMsg never receives the address, so we don't need to send // one. - _, notify, err := s.connection.Send(bufs, transport.ControlMessages{}, tcpip.FullAddress{}) + _, notify, err := s.connection.Send(bufs, cms, tcpip.FullAddress{}) // If the buffer is full, we simply drop messages, just like // Linux. if err != nil && err != syserr.ErrWouldBlock { @@ -499,7 +630,10 @@ func (s *Socket) sendResponse(ctx context.Context, ms *MessageSet) *syserr.Error PortID: uint32(ms.PortID), }) - _, notify, err := s.connection.Send([][]byte{m.Finalize()}, transport.ControlMessages{}, tcpip.FullAddress{}) + // Add the dump_done_errno payload. + m.Put(int64(0)) + + _, notify, err := s.connection.Send([][]byte{m.Finalize()}, cms, tcpip.FullAddress{}) if err != nil && err != syserr.ErrWouldBlock { return err } |