Merge 216da0b7 (automated)

8 files changed, 1273 insertions, 0 deletions

diff --git a/pkg/sentry/socket/netlink/message.go b/pkg/sentry/socket/netlink/message.go
new file mode 100644
index 000000000..5bd3b49ce
--- /dev/null
+++ b/pkg/sentry/socket/netlink/message.go
@@ -0,0 +1,159 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package netlink
+
+import (
+	"fmt"
+	"math"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// alignUp rounds a length up to an alignment.
+//
+// Preconditions: align is a power of two.
+func alignUp(length int, align uint) int {
+	return (length + int(align) - 1) &^ (int(align) - 1)
+}
+
+// Message contains a complete serialized netlink message.
+type Message struct {
+	buf []byte
+}
+
+// NewMessage creates a new Message containing the passed header.
+//
+// The header length will be updated by Finalize.
+func NewMessage(hdr linux.NetlinkMessageHeader) *Message {
+	return &Message{
+		buf: binary.Marshal(nil, usermem.ByteOrder, hdr),
+	}
+}
+
+// Finalize returns the []byte containing the entire message, with the total
+// length set in the message header. The Message must not be modified after
+// calling Finalize.
+func (m *Message) Finalize() []byte {
+	// Update length, which is the first 4 bytes of the header.
+	usermem.ByteOrder.PutUint32(m.buf, uint32(len(m.buf)))
+
+	// Align the message. Note that the message length in the header (set
+	// above) is the useful length of the message, not the total aligned
+	// length. See net/netlink/af_netlink.c:__nlmsg_put.
+	aligned := alignUp(len(m.buf), linux.NLMSG_ALIGNTO)
+	m.putZeros(aligned - len(m.buf))
+	return m.buf
+}
+
+// putZeros adds n zeros to the message.
+func (m *Message) putZeros(n int) {
+	for n > 0 {
+		m.buf = append(m.buf, 0)
+		n--
+	}
+}
+
+// Put serializes v into the message.
+func (m *Message) Put(v interface{}) {
+	m.buf = binary.Marshal(m.buf, usermem.ByteOrder, v)
+}
+
+// PutAttr adds v to the message as a netlink attribute.
+//
+// Preconditions: The serialized attribute (linux.NetlinkAttrHeaderSize +
+// binary.Size(v) fits in math.MaxUint16 bytes.
+func (m *Message) PutAttr(atype uint16, v interface{}) {
+	l := linux.NetlinkAttrHeaderSize + int(binary.Size(v))
+	if l > math.MaxUint16 {
+		panic(fmt.Sprintf("attribute too large: %d", l))
+	}
+
+	m.Put(linux.NetlinkAttrHeader{
+		Type:   atype,
+		Length: uint16(l),
+	})
+	m.Put(v)
+
+	// Align the attribute.
+	aligned := alignUp(l, linux.NLA_ALIGNTO)
+	m.putZeros(aligned - l)
+}
+
+// PutAttrString adds s to the message as a netlink attribute.
+func (m *Message) PutAttrString(atype uint16, s string) {
+	l := linux.NetlinkAttrHeaderSize + len(s) + 1
+	m.Put(linux.NetlinkAttrHeader{
+		Type:   atype,
+		Length: uint16(l),
+	})
+
+	// String + NUL-termination.
+	m.Put([]byte(s))
+	m.putZeros(1)
+
+	// Align the attribute.
+	aligned := alignUp(l, linux.NLA_ALIGNTO)
+	m.putZeros(aligned - l)
+}
+
+// MessageSet contains a series of netlink messages.
+type MessageSet struct {
+	// Multi indicates that this a multi-part message, to be terminated by
+	// NLMSG_DONE. NLMSG_DONE is sent even if the set contains only one
+	// Message.
+	//
+	// If Multi is set, all added messages will have NLM_F_MULTI set.
+	Multi bool
+
+	// PortID is the destination port for all messages.
+	PortID int32
+
+	// Seq is the sequence counter for all messages in the set.
+	Seq uint32
+
+	// Messages contains the messages in the set.
+	Messages []*Message
+}
+
+// NewMessageSet creates a new MessageSet.
+//
+// portID is the destination port to set as PortID in all messages.
+//
+// seq is the sequence counter to set as seq in all messages in the set.
+func NewMessageSet(portID int32, seq uint32) *MessageSet {
+	return &MessageSet{
+		PortID: portID,
+		Seq:    seq,
+	}
+}
+
+// AddMessage adds a new message to the set and returns it for further
+// additions.
+//
+// The passed header will have Seq, PortID and the multi flag set
+// automatically.
+func (ms *MessageSet) AddMessage(hdr linux.NetlinkMessageHeader) *Message {
+	hdr.Seq = ms.Seq
+	hdr.PortID = uint32(ms.PortID)
+	if ms.Multi {
+		hdr.Flags |= linux.NLM_F_MULTI
+	}
+
+	m := NewMessage(hdr)
+	ms.Messages = append(ms.Messages, m)
+	return m
+}
diff --git a/pkg/sentry/socket/netlink/netlink_state_autogen.go b/pkg/sentry/socket/netlink/netlink_state_autogen.go
new file mode 100755
index 000000000..59d902798
--- /dev/null
+++ b/pkg/sentry/socket/netlink/netlink_state_autogen.go
@@ -0,0 +1,36 @@
+// automatically generated by stateify.
+
+package netlink
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/state"
+)
+
+func (x *Socket) beforeSave() {}
+func (x *Socket) save(m state.Map) {
+	x.beforeSave()
+	m.Save("SendReceiveTimeout", &x.SendReceiveTimeout)
+	m.Save("ports", &x.ports)
+	m.Save("protocol", &x.protocol)
+	m.Save("ep", &x.ep)
+	m.Save("connection", &x.connection)
+	m.Save("bound", &x.bound)
+	m.Save("portID", &x.portID)
+	m.Save("sendBufferSize", &x.sendBufferSize)
+}
+
+func (x *Socket) afterLoad() {}
+func (x *Socket) load(m state.Map) {
+	m.Load("SendReceiveTimeout", &x.SendReceiveTimeout)
+	m.Load("ports", &x.ports)
+	m.Load("protocol", &x.protocol)
+	m.Load("ep", &x.ep)
+	m.Load("connection", &x.connection)
+	m.Load("bound", &x.bound)
+	m.Load("portID", &x.portID)
+	m.Load("sendBufferSize", &x.sendBufferSize)
+}
+
+func init() {
+	state.Register("netlink.Socket", (*Socket)(nil), state.Fns{Save: (*Socket).save, Load: (*Socket).load})
+}
diff --git a/pkg/sentry/socket/netlink/port/port.go b/pkg/sentry/socket/netlink/port/port.go
new file mode 100644
index 000000000..e9d3275b1
--- /dev/null
+++ b/pkg/sentry/socket/netlink/port/port.go
@@ -0,0 +1,116 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package port provides port ID allocation for netlink sockets.
+//
+// A netlink port is any int32 value. Positive ports are typically equivalent
+// to the PID of the binding process. If that port is unavailable, negative
+// ports are searched to find a free port that will not conflict with other
+// PIDS.
+package port
+
+import (
+	"fmt"
+	"math"
+	"math/rand"
+	"sync"
+)
+
+// maxPorts is a sanity limit on the maximum number of ports to allocate per
+// protocol.
+const maxPorts = 10000
+
+// Manager allocates netlink port IDs.
+//
+// +stateify savable
+type Manager struct {
+	// mu protects the fields below.
+	mu sync.Mutex `state:"nosave"`
+
+	// ports contains a map of allocated ports for each protocol.
+	ports map[int]map[int32]struct{}
+}
+
+// New creates a new Manager.
+func New() *Manager {
+	return &Manager{
+		ports: make(map[int]map[int32]struct{}),
+	}
+}
+
+// Allocate reserves a new port ID for protocol. hint will be taken if
+// available.
+func (m *Manager) Allocate(protocol int, hint int32) (int32, bool) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	proto, ok := m.ports[protocol]
+	if !ok {
+		proto = make(map[int32]struct{})
+		// Port 0 is reserved for the kernel.
+		proto[0] = struct{}{}
+		m.ports[protocol] = proto
+	}
+
+	if len(proto) >= maxPorts {
+		return 0, false
+	}
+
+	if _, ok := proto[hint]; !ok {
+		// Hint is available, reserve it.
+		proto[hint] = struct{}{}
+		return hint, true
+	}
+
+	// Search for any free port in [math.MinInt32, -4096). The positive
+	// port space is left open for pid-based allocations. This behavior is
+	// consistent with Linux.
+	start := int32(math.MinInt32 + rand.Int63n(math.MaxInt32-4096+1))
+	curr := start
+	for {
+		if _, ok := proto[curr]; !ok {
+			proto[curr] = struct{}{}
+			return curr, true
+		}
+
+		curr--
+		if curr >= -4096 {
+			curr = -4097
+		}
+		if curr == start {
+			// Nothing found. We should always find a free port
+			// because maxPorts < -4096 - MinInt32.
+			panic(fmt.Sprintf("No free port found in %+v", proto))
+		}
+	}
+}
+
+// Release frees the specified port for protocol.
+//
+// Preconditions: port is already allocated.
+func (m *Manager) Release(protocol int, port int32) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	proto, ok := m.ports[protocol]
+	if !ok {
+		panic(fmt.Sprintf("Released port %d for protocol %d which has no allocations", port, protocol))
+	}
+
+	if _, ok := proto[port]; !ok {
+		panic(fmt.Sprintf("Released port %d for protocol %d is not allocated", port, protocol))
+	}
+
+	delete(proto, port)
+}
diff --git a/pkg/sentry/socket/netlink/port/port_state_autogen.go b/pkg/sentry/socket/netlink/port/port_state_autogen.go
new file mode 100755
index 000000000..f01d9704f
--- /dev/null
+++ b/pkg/sentry/socket/netlink/port/port_state_autogen.go
@@ -0,0 +1,22 @@
+// automatically generated by stateify.
+
+package port
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/state"
+)
+
+func (x *Manager) beforeSave() {}
+func (x *Manager) save(m state.Map) {
+	x.beforeSave()
+	m.Save("ports", &x.ports)
+}
+
+func (x *Manager) afterLoad() {}
+func (x *Manager) load(m state.Map) {
+	m.Load("ports", &x.ports)
+}
+
+func init() {
+	state.Register("port.Manager", (*Manager)(nil), state.Fns{Save: (*Manager).save, Load: (*Manager).load})
+}
diff --git a/pkg/sentry/socket/netlink/provider.go b/pkg/sentry/socket/netlink/provider.go
new file mode 100644
index 000000000..76cf12fd4
--- /dev/null
+++ b/pkg/sentry/socket/netlink/provider.go
@@ -0,0 +1,105 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package netlink
+
+import (
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/unix/transport"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+)
+
+// Protocol is the implementation of a netlink socket protocol.
+type Protocol interface {
+	// Protocol returns the Linux netlink protocol value.
+	Protocol() int
+
+	// ProcessMessage processes a single message from userspace.
+	//
+	// If err == nil, any messages added to ms will be sent back to the
+	// other end of the socket. Setting ms.Multi will cause an NLMSG_DONE
+	// message to be sent even if ms contains no messages.
+	ProcessMessage(ctx context.Context, hdr linux.NetlinkMessageHeader, data []byte, ms *MessageSet) *syserr.Error
+}
+
+// Provider is a function that creates a new Protocol for a specific netlink
+// protocol.
+//
+// Note that this is distinct from socket.Provider, which is used for all
+// socket families.
+type Provider func(t *kernel.Task) (Protocol, *syserr.Error)
+
+// protocols holds a map of all known address protocols and their provider.
+var protocols = make(map[int]Provider)
+
+// RegisterProvider registers the provider of a given address protocol so that
+// netlink sockets of that type can be created via socket(2).
+//
+// Preconditions: May only be called before any netlink sockets are created.
+func RegisterProvider(protocol int, provider Provider) {
+	if p, ok := protocols[protocol]; ok {
+		panic(fmt.Sprintf("Netlink protocol %d already provided by %+v", protocol, p))
+	}
+
+	protocols[protocol] = provider
+}
+
+// socketProvider implements socket.Provider.
+type socketProvider struct {
+}
+
+// Socket implements socket.Provider.Socket.
+func (*socketProvider) Socket(t *kernel.Task, stype transport.SockType, protocol int) (*fs.File, *syserr.Error) {
+	// Netlink sockets must be specified as datagram or raw, but they
+	// behave the same regardless of type.
+	if stype != transport.SockDgram && stype != transport.SockRaw {
+		return nil, syserr.ErrSocketNotSupported
+	}
+
+	provider, ok := protocols[protocol]
+	if !ok {
+		return nil, syserr.ErrProtocolNotSupported
+	}
+
+	p, err := provider(t)
+	if err != nil {
+		return nil, err
+	}
+
+	s, err := NewSocket(t, p)
+	if err != nil {
+		return nil, err
+	}
+
+	d := socket.NewDirent(t, netlinkSocketDevice)
+	defer d.DecRef()
+	return fs.NewFile(t, d, fs.FileFlags{Read: true, Write: true}, s), nil
+}
+
+// Pair implements socket.Provider.Pair by returning an error.
+func (*socketProvider) Pair(*kernel.Task, transport.SockType, int) (*fs.File, *fs.File, *syserr.Error) {
+	// Netlink sockets never supports creating socket pairs.
+	return nil, nil, syserr.ErrNotSupported
+}
+
+// init registers the socket provider.
+func init() {
+	socket.RegisterProvider(linux.AF_NETLINK, &socketProvider{})
+}
diff --git a/pkg/sentry/socket/netlink/route/protocol.go b/pkg/sentry/socket/netlink/route/protocol.go
new file mode 100644
index 000000000..9f0a81403
--- /dev/null
+++ b/pkg/sentry/socket/netlink/route/protocol.go
@@ -0,0 +1,197 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package route provides a NETLINK_ROUTE socket protocol.
+package route
+
+import (
+	"bytes"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/inet"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/netlink"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+)
+
+// commandKind describes the operational class of a message type.
+//
+// The route message types use the lower 2 bits of the type to describe class
+// of command.
+type commandKind int
+
+const (
+	kindNew commandKind = 0x0
+	kindDel             = 0x1
+	kindGet             = 0x2
+	kindSet             = 0x3
+)
+
+func typeKind(typ uint16) commandKind {
+	return commandKind(typ & 0x3)
+}
+
+// Protocol implements netlink.Protocol.
+//
+// +stateify savable
+type Protocol struct{}
+
+var _ netlink.Protocol = (*Protocol)(nil)
+
+// NewProtocol creates a NETLINK_ROUTE netlink.Protocol.
+func NewProtocol(t *kernel.Task) (netlink.Protocol, *syserr.Error) {
+	return &Protocol{}, nil
+}
+
+// Protocol implements netlink.Protocol.Protocol.
+func (p *Protocol) Protocol() int {
+	return linux.NETLINK_ROUTE
+}
+
+// dumpLinks handles RTM_GETLINK + NLM_F_DUMP requests.
+func (p *Protocol) dumpLinks(ctx context.Context, hdr linux.NetlinkMessageHeader, data []byte, ms *netlink.MessageSet) *syserr.Error {
+	// NLM_F_DUMP + RTM_GETLINK messages are supposed to include an
+	// ifinfomsg. However, Linux <3.9 only checked for rtgenmsg, and some
+	// userspace applications (including glibc) still include rtgenmsg.
+	// Linux has a workaround based on the total message length.
+	//
+	// We don't bother to check for either, since we don't support any
+	// extra attributes that may be included anyways.
+	//
+	// The message may also contain netlink attribute IFLA_EXT_MASK, which
+	// we don't support.
+
+	// The RTM_GETLINK dump response is a set of messages each containing
+	// an InterfaceInfoMessage followed by a set of netlink attributes.
+
+	// We always send back an NLMSG_DONE.
+	ms.Multi = true
+
+	stack := inet.StackFromContext(ctx)
+	if stack == nil {
+		// No network devices.
+		return nil
+	}
+
+	for id, i := range stack.Interfaces() {
+		m := ms.AddMessage(linux.NetlinkMessageHeader{
+			Type: linux.RTM_NEWLINK,
+		})
+
+		m.Put(linux.InterfaceInfoMessage{
+			Family: linux.AF_UNSPEC,
+			Type:   i.DeviceType,
+			Index:  id,
+			Flags:  i.Flags,
+		})
+
+		m.PutAttrString(linux.IFLA_IFNAME, i.Name)
+		m.PutAttr(linux.IFLA_MTU, i.MTU)
+
+		mac := make([]byte, 6)
+		brd := mac
+		if len(i.Addr) > 0 {
+			mac = i.Addr
+			brd = bytes.Repeat([]byte{0xff}, len(i.Addr))
+		}
+		m.PutAttr(linux.IFLA_ADDRESS, mac)
+		m.PutAttr(linux.IFLA_BROADCAST, brd)
+
+		// TODO(b/68878065): There are many more attributes.
+	}
+
+	return nil
+}
+
+// dumpAddrs handles RTM_GETADDR + NLM_F_DUMP requests.
+func (p *Protocol) dumpAddrs(ctx context.Context, hdr linux.NetlinkMessageHeader, data []byte, ms *netlink.MessageSet) *syserr.Error {
+	// RTM_GETADDR dump requests need not contain anything more than the
+	// netlink header and 1 byte protocol family common to all
+	// NETLINK_ROUTE requests.
+	//
+	// TODO(b/68878065): Filter output by passed protocol family.
+
+	// The RTM_GETADDR dump response is a set of RTM_NEWADDR messages each
+	// containing an InterfaceAddrMessage followed by a set of netlink
+	// attributes.
+
+	// We always send back an NLMSG_DONE.
+	ms.Multi = true
+
+	stack := inet.StackFromContext(ctx)
+	if stack == nil {
+		// No network devices.
+		return nil
+	}
+
+	for id, as := range stack.InterfaceAddrs() {
+		for _, a := range as {
+			m := ms.AddMessage(linux.NetlinkMessageHeader{
+				Type: linux.RTM_NEWADDR,
+			})
+
+			m.Put(linux.InterfaceAddrMessage{
+				Family:    a.Family,
+				PrefixLen: a.PrefixLen,
+				Index:     uint32(id),
+			})
+
+			m.PutAttr(linux.IFA_ADDRESS, []byte(a.Addr))
+
+			// TODO(b/68878065): There are many more attributes.
+		}
+	}
+
+	return nil
+}
+
+// ProcessMessage implements netlink.Protocol.ProcessMessage.
+func (p *Protocol) ProcessMessage(ctx context.Context, hdr linux.NetlinkMessageHeader, data []byte, ms *netlink.MessageSet) *syserr.Error {
+	// All messages start with a 1 byte protocol family.
+	if len(data) < 1 {
+		// Linux ignores messages missing the protocol family. See
+		// net/core/rtnetlink.c:rtnetlink_rcv_msg.
+		return nil
+	}
+
+	// Non-GET message types require CAP_NET_ADMIN.
+	if typeKind(hdr.Type) != kindGet {
+		creds := auth.CredentialsFromContext(ctx)
+		if !creds.HasCapability(linux.CAP_NET_ADMIN) {
+			return syserr.ErrPermissionDenied
+		}
+	}
+
+	// TODO(b/68878065): Only the dump variant of the types below are
+	// supported.
+	if hdr.Flags&linux.NLM_F_DUMP != linux.NLM_F_DUMP {
+		return syserr.ErrNotSupported
+	}
+
+	switch hdr.Type {
+	case linux.RTM_GETLINK:
+		return p.dumpLinks(ctx, hdr, data, ms)
+	case linux.RTM_GETADDR:
+		return p.dumpAddrs(ctx, hdr, data, ms)
+	default:
+		return syserr.ErrNotSupported
+	}
+}
+
+// init registers the NETLINK_ROUTE provider.
+func init() {
+	netlink.RegisterProvider(linux.NETLINK_ROUTE, NewProtocol)
+}
diff --git a/pkg/sentry/socket/netlink/route/route_state_autogen.go b/pkg/sentry/socket/netlink/route/route_state_autogen.go
new file mode 100755
index 000000000..8431bb3d5
--- /dev/null
+++ b/pkg/sentry/socket/netlink/route/route_state_autogen.go
@@ -0,0 +1,20 @@
+// automatically generated by stateify.
+
+package route
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/state"
+)
+
+func (x *Protocol) beforeSave() {}
+func (x *Protocol) save(m state.Map) {
+	x.beforeSave()
+}
+
+func (x *Protocol) afterLoad() {}
+func (x *Protocol) load(m state.Map) {
+}
+
+func init() {
+	state.Register("route.Protocol", (*Protocol)(nil), state.Fns{Save: (*Protocol).save, Load: (*Protocol).load})
+}
diff --git a/pkg/sentry/socket/netlink/socket.go b/pkg/sentry/socket/netlink/socket.go
new file mode 100644
index 000000000..afd06ca33
--- /dev/null
+++ b/pkg/sentry/socket/netlink/socket.go
@@ -0,0 +1,618 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package netlink provides core functionality for netlink sockets.
+package netlink
+
+import (
+	"math"
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/netlink/port"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/unix"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/unix/transport"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+const sizeOfInt32 int = 4
+
+const (
+	// minBufferSize is the smallest size of a send buffer.
+	minSendBufferSize = 4 << 10 // 4096 bytes.
+
+	// defaultSendBufferSize is the default size for the send buffer.
+	defaultSendBufferSize = 16 * 1024
+
+	// maxBufferSize is the largest size a send buffer can grow to.
+	maxSendBufferSize = 4 << 20 // 4MB
+)
+
+// netlinkSocketDevice is the netlink socket virtual device.
+var netlinkSocketDevice = device.NewAnonDevice()
+
+// Socket is the base socket type for netlink sockets.
+//
+// This implementation only supports userspace sending and receiving messages
+// to/from the kernel.
+//
+// Socket implements socket.Socket.
+//
+// +stateify savable
+type Socket struct {
+	fsutil.FilePipeSeek             `state:"nosave"`
+	fsutil.FileNotDirReaddir        `state:"nosave"`
+	fsutil.FileNoFsync              `state:"nosave"`
+	fsutil.FileNoMMap               `state:"nosave"`
+	fsutil.FileNoSplice             `state:"nosave"`
+	fsutil.FileNoopFlush            `state:"nosave"`
+	fsutil.FileUseInodeUnstableAttr `state:"nosave"`
+	socket.SendReceiveTimeout
+
+	// ports provides netlink port allocation.
+	ports *port.Manager
+
+	// protocol is the netlink protocol implementation.
+	protocol Protocol
+
+	// ep is a datagram unix endpoint used to buffer messages sent from the
+	// kernel to userspace. RecvMsg reads messages from this endpoint.
+	ep transport.Endpoint
+
+	// connection is the kernel's connection to ep, used to write messages
+	// sent to userspace.
+	connection transport.ConnectedEndpoint
+
+	// mu protects the fields below.
+	mu sync.Mutex `state:"nosave"`
+
+	// bound indicates that portid is valid.
+	bound bool
+
+	// portID is the port ID allocated for this socket.
+	portID int32
+
+	// sendBufferSize is the send buffer "size". We don't actually have a
+	// fixed buffer but only consume this many bytes.
+	sendBufferSize uint32
+}
+
+var _ socket.Socket = (*Socket)(nil)
+
+// NewSocket creates a new Socket.
+func NewSocket(t *kernel.Task, protocol Protocol) (*Socket, *syserr.Error) {
+	// Datagram endpoint used to buffer kernel -> user messages.
+	ep := transport.NewConnectionless()
+
+	// Bind the endpoint for good measure so we can connect to it. The
+	// bound address will never be exposed.
+	if err := ep.Bind(tcpip.FullAddress{Addr: "dummy"}, nil); err != nil {
+		ep.Close()
+		return nil, err
+	}
+
+	// Create a connection from which the kernel can write messages.
+	connection, err := ep.(transport.BoundEndpoint).UnidirectionalConnect()
+	if err != nil {
+		ep.Close()
+		return nil, err
+	}
+
+	return &Socket{
+		ports:          t.Kernel().NetlinkPorts(),
+		protocol:       protocol,
+		ep:             ep,
+		connection:     connection,
+		sendBufferSize: defaultSendBufferSize,
+	}, nil
+}
+
+// Release implements fs.FileOperations.Release.
+func (s *Socket) Release() {
+	s.connection.Release()
+	s.ep.Close()
+
+	if s.bound {
+		s.ports.Release(s.protocol.Protocol(), s.portID)
+	}
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (s *Socket) Readiness(mask waiter.EventMask) waiter.EventMask {
+	// ep holds messages to be read and thus handles EventIn readiness.
+	ready := s.ep.Readiness(mask)
+
+	if mask&waiter.EventOut == waiter.EventOut {
+		// sendMsg handles messages synchronously and is thus always
+		// ready for writing.
+		ready |= waiter.EventOut
+	}
+
+	return ready
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (s *Socket) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	s.ep.EventRegister(e, mask)
+	// Writable readiness never changes, so no registration is needed.
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (s *Socket) EventUnregister(e *waiter.Entry) {
+	s.ep.EventUnregister(e)
+}
+
+// Ioctl implements fs.FileOperations.Ioctl.
+func (s *Socket) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	// TODO(b/68878065): no ioctls supported.
+	return 0, syserror.ENOTTY
+}
+
+// ExtractSockAddr extracts the SockAddrNetlink from b.
+func ExtractSockAddr(b []byte) (*linux.SockAddrNetlink, *syserr.Error) {
+	if len(b) < linux.SockAddrNetlinkSize {
+		return nil, syserr.ErrBadAddress
+	}
+
+	var sa linux.SockAddrNetlink
+	binary.Unmarshal(b[:linux.SockAddrNetlinkSize], usermem.ByteOrder, &sa)
+
+	if sa.Family != linux.AF_NETLINK {
+		return nil, syserr.ErrInvalidArgument
+	}
+
+	return &sa, nil
+}
+
+// bindPort binds this socket to a port, preferring 'port' if it is available.
+//
+// port of 0 defaults to the ThreadGroup ID.
+//
+// Preconditions: mu is held.
+func (s *Socket) bindPort(t *kernel.Task, port int32) *syserr.Error {
+	if s.bound {
+		// Re-binding is only allowed if the port doesn't change.
+		if port != s.portID {
+			return syserr.ErrInvalidArgument
+		}
+
+		return nil
+	}
+
+	if port == 0 {
+		port = int32(t.ThreadGroup().ID())
+	}
+	port, ok := s.ports.Allocate(s.protocol.Protocol(), port)
+	if !ok {
+		return syserr.ErrBusy
+	}
+
+	s.portID = port
+	s.bound = true
+	return nil
+}
+
+// Bind implements socket.Socket.Bind.
+func (s *Socket) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error {
+	a, err := ExtractSockAddr(sockaddr)
+	if err != nil {
+		return err
+	}
+
+	// No support for multicast groups yet.
+	if a.Groups != 0 {
+		return syserr.ErrPermissionDenied
+	}
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	return s.bindPort(t, int32(a.PortID))
+}
+
+// Connect implements socket.Socket.Connect.
+func (s *Socket) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error {
+	a, err := ExtractSockAddr(sockaddr)
+	if err != nil {
+		return err
+	}
+
+	// No support for multicast groups yet.
+	if a.Groups != 0 {
+		return syserr.ErrPermissionDenied
+	}
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	if a.PortID == 0 {
+		// Netlink sockets default to connected to the kernel, but
+		// connecting anyways automatically binds if not already bound.
+		if !s.bound {
+			// Pass port 0 to get an auto-selected port ID.
+			return s.bindPort(t, 0)
+		}
+		return nil
+	}
+
+	// We don't support non-kernel destination ports. Linux returns EPERM
+	// if applications attempt to do this without NL_CFG_F_NONROOT_SEND, so
+	// we emulate that.
+	return syserr.ErrPermissionDenied
+}
+
+// Accept implements socket.Socket.Accept.
+func (s *Socket) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (kdefs.FD, interface{}, uint32, *syserr.Error) {
+	// Netlink sockets never support accept.
+	return 0, nil, 0, syserr.ErrNotSupported
+}
+
+// Listen implements socket.Socket.Listen.
+func (s *Socket) Listen(t *kernel.Task, backlog int) *syserr.Error {
+	// Netlink sockets never support listen.
+	return syserr.ErrNotSupported
+}
+
+// Shutdown implements socket.Socket.Shutdown.
+func (s *Socket) Shutdown(t *kernel.Task, how int) *syserr.Error {
+	// Netlink sockets never support shutdown.
+	return syserr.ErrNotSupported
+}
+
+// GetSockOpt implements socket.Socket.GetSockOpt.
+func (s *Socket) GetSockOpt(t *kernel.Task, level int, name int, outLen int) (interface{}, *syserr.Error) {
+	switch level {
+	case linux.SOL_SOCKET:
+		switch name {
+		case linux.SO_SNDBUF:
+			if outLen < sizeOfInt32 {
+				return nil, syserr.ErrInvalidArgument
+			}
+			s.mu.Lock()
+			defer s.mu.Unlock()
+			return int32(s.sendBufferSize), nil
+
+		case linux.SO_RCVBUF:
+			if outLen < sizeOfInt32 {
+				return nil, syserr.ErrInvalidArgument
+			}
+			// We don't have limit on receiving size.
+			return int32(math.MaxInt32), nil
+
+		default:
+			socket.GetSockOptEmitUnimplementedEvent(t, name)
+		}
+	case linux.SOL_NETLINK:
+		switch name {
+		case linux.NETLINK_BROADCAST_ERROR,
+			linux.NETLINK_CAP_ACK,
+			linux.NETLINK_DUMP_STRICT_CHK,
+			linux.NETLINK_EXT_ACK,
+			linux.NETLINK_LIST_MEMBERSHIPS,
+			linux.NETLINK_NO_ENOBUFS,
+			linux.NETLINK_PKTINFO:
+
+			t.Kernel().EmitUnimplementedEvent(t)
+		}
+	}
+	// TODO(b/68878065): other sockopts are not supported.
+	return nil, syserr.ErrProtocolNotAvailable
+}
+
+// SetSockOpt implements socket.Socket.SetSockOpt.
+func (s *Socket) SetSockOpt(t *kernel.Task, level int, name int, opt []byte) *syserr.Error {
+	switch level {
+	case linux.SOL_SOCKET:
+		switch name {
+		case linux.SO_SNDBUF:
+			if len(opt) < sizeOfInt32 {
+				return syserr.ErrInvalidArgument
+			}
+			size := usermem.ByteOrder.Uint32(opt)
+			if size < minSendBufferSize {
+				size = minSendBufferSize
+			} else if size > maxSendBufferSize {
+				size = maxSendBufferSize
+			}
+			s.mu.Lock()
+			s.sendBufferSize = size
+			s.mu.Unlock()
+			return nil
+		case linux.SO_RCVBUF:
+			if len(opt) < sizeOfInt32 {
+				return syserr.ErrInvalidArgument
+			}
+			// We don't have limit on receiving size. So just accept anything as
+			// valid for compatibility.
+			return nil
+		default:
+			socket.SetSockOptEmitUnimplementedEvent(t, name)
+		}
+
+	case linux.SOL_NETLINK:
+		switch name {
+		case linux.NETLINK_ADD_MEMBERSHIP,
+			linux.NETLINK_BROADCAST_ERROR,
+			linux.NETLINK_CAP_ACK,
+			linux.NETLINK_DROP_MEMBERSHIP,
+			linux.NETLINK_DUMP_STRICT_CHK,
+			linux.NETLINK_EXT_ACK,
+			linux.NETLINK_LISTEN_ALL_NSID,
+			linux.NETLINK_NO_ENOBUFS,
+			linux.NETLINK_PKTINFO:
+
+			t.Kernel().EmitUnimplementedEvent(t)
+		}
+
+	}
+	// TODO(b/68878065): other sockopts are not supported.
+	return syserr.ErrProtocolNotAvailable
+}
+
+// GetSockName implements socket.Socket.GetSockName.
+func (s *Socket) GetSockName(t *kernel.Task) (interface{}, uint32, *syserr.Error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	sa := linux.SockAddrNetlink{
+		Family: linux.AF_NETLINK,
+		PortID: uint32(s.portID),
+	}
+	return sa, uint32(binary.Size(sa)), nil
+}
+
+// GetPeerName implements socket.Socket.GetPeerName.
+func (s *Socket) GetPeerName(t *kernel.Task) (interface{}, uint32, *syserr.Error) {
+	sa := linux.SockAddrNetlink{
+		Family: linux.AF_NETLINK,
+		// TODO(b/68878065): Support non-kernel peers. For now the peer
+		// must be the kernel.
+		PortID: 0,
+	}
+	return sa, uint32(binary.Size(sa)), nil
+}
+
+// RecvMsg implements socket.Socket.RecvMsg.
+func (s *Socket) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (int, int, interface{}, uint32, socket.ControlMessages, *syserr.Error) {
+	from := linux.SockAddrNetlink{
+		Family: linux.AF_NETLINK,
+		PortID: 0,
+	}
+	fromLen := uint32(binary.Size(from))
+
+	trunc := flags&linux.MSG_TRUNC != 0
+
+	r := unix.EndpointReader{
+		Endpoint: s.ep,
+		Peek:     flags&linux.MSG_PEEK != 0,
+	}
+
+	if n, err := dst.CopyOutFrom(t, &r); err != syserror.ErrWouldBlock || flags&linux.MSG_DONTWAIT != 0 {
+		var mflags int
+		if n < int64(r.MsgSize) {
+			mflags |= linux.MSG_TRUNC
+		}
+		if trunc {
+			n = int64(r.MsgSize)
+		}
+		return int(n), mflags, from, fromLen, socket.ControlMessages{}, syserr.FromError(err)
+	}
+
+	// We'll have to block. Register for notification and keep trying to
+	// receive all the data.
+	e, ch := waiter.NewChannelEntry(nil)
+	s.EventRegister(&e, waiter.EventIn)
+	defer s.EventUnregister(&e)
+
+	for {
+		if n, err := dst.CopyOutFrom(t, &r); err != syserror.ErrWouldBlock {
+			var mflags int
+			if n < int64(r.MsgSize) {
+				mflags |= linux.MSG_TRUNC
+			}
+			if trunc {
+				n = int64(r.MsgSize)
+			}
+			return int(n), mflags, from, fromLen, socket.ControlMessages{}, syserr.FromError(err)
+		}
+
+		if err := t.BlockWithDeadline(ch, haveDeadline, deadline); err != nil {
+			if err == syserror.ETIMEDOUT {
+				return 0, 0, nil, 0, socket.ControlMessages{}, syserr.ErrTryAgain
+			}
+			return 0, 0, nil, 0, socket.ControlMessages{}, syserr.FromError(err)
+		}
+	}
+}
+
+// Read implements fs.FileOperations.Read.
+func (s *Socket) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, _ int64) (int64, error) {
+	if dst.NumBytes() == 0 {
+		return 0, nil
+	}
+	return dst.CopyOutFrom(ctx, &unix.EndpointReader{
+		Endpoint: s.ep,
+	})
+}
+
+// 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.
+	bufs := make([][]byte, 0, len(ms.Messages))
+	for _, m := range ms.Messages {
+		bufs = append(bufs, m.Finalize())
+	}
+
+	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{})
+		// If the buffer is full, we simply drop messages, just like
+		// Linux.
+		if err != nil && err != syserr.ErrWouldBlock {
+			return err
+		}
+		if notify {
+			s.connection.SendNotify()
+		}
+	}
+
+	// N.B. multi-part messages should still send NLMSG_DONE even if
+	// MessageSet contains no messages.
+	//
+	// N.B. NLMSG_DONE is always sent in a different datagram. See
+	// net/netlink/af_netlink.c:netlink_dump.
+	if ms.Multi {
+		m := NewMessage(linux.NetlinkMessageHeader{
+			Type:   linux.NLMSG_DONE,
+			Flags:  linux.NLM_F_MULTI,
+			Seq:    ms.Seq,
+			PortID: uint32(ms.PortID),
+		})
+
+		_, notify, err := s.connection.Send([][]byte{m.Finalize()}, transport.ControlMessages{}, tcpip.FullAddress{})
+		if err != nil && err != syserr.ErrWouldBlock {
+			return err
+		}
+		if notify {
+			s.connection.SendNotify()
+		}
+	}
+
+	return nil
+}
+
+// processMessages handles each message in buf, passing it to the protocol
+// handler for final handling.
+func (s *Socket) processMessages(ctx context.Context, buf []byte) *syserr.Error {
+	for len(buf) > 0 {
+		if len(buf) < linux.NetlinkMessageHeaderSize {
+			// Linux ignores messages that are too short. See
+			// net/netlink/af_netlink.c:netlink_rcv_skb.
+			break
+		}
+
+		var hdr linux.NetlinkMessageHeader
+		binary.Unmarshal(buf[:linux.NetlinkMessageHeaderSize], usermem.ByteOrder, &hdr)
+
+		if hdr.Length < linux.NetlinkMessageHeaderSize || uint64(hdr.Length) > uint64(len(buf)) {
+			// Linux ignores malformed messages. See
+			// net/netlink/af_netlink.c:netlink_rcv_skb.
+			break
+		}
+
+		// Data from this message.
+		data := buf[linux.NetlinkMessageHeaderSize:hdr.Length]
+
+		// Advance to the next message.
+		next := alignUp(int(hdr.Length), linux.NLMSG_ALIGNTO)
+		if next >= len(buf)-1 {
+			next = len(buf) - 1
+		}
+		buf = buf[next:]
+
+		// Ignore control messages.
+		if hdr.Type < linux.NLMSG_MIN_TYPE {
+			continue
+		}
+
+		// TODO(b/68877377): ACKs not supported yet.
+		if hdr.Flags&linux.NLM_F_ACK == linux.NLM_F_ACK {
+			return syserr.ErrNotSupported
+		}
+
+		ms := NewMessageSet(s.portID, hdr.Seq)
+		if err := s.protocol.ProcessMessage(ctx, hdr, data, ms); err != nil {
+			return err
+		}
+
+		if err := s.sendResponse(ctx, ms); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+// sendMsg is the core of message send, used for SendMsg and Write.
+func (s *Socket) sendMsg(ctx context.Context, src usermem.IOSequence, to []byte, flags int, controlMessages socket.ControlMessages) (int, *syserr.Error) {
+	dstPort := int32(0)
+
+	if len(to) != 0 {
+		a, err := ExtractSockAddr(to)
+		if err != nil {
+			return 0, err
+		}
+
+		// No support for multicast groups yet.
+		if a.Groups != 0 {
+			return 0, syserr.ErrPermissionDenied
+		}
+
+		dstPort = int32(a.PortID)
+	}
+
+	if dstPort != 0 {
+		// Non-kernel destinations not supported yet. Treat as if
+		// NL_CFG_F_NONROOT_SEND is not set.
+		return 0, syserr.ErrPermissionDenied
+	}
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	// For simplicity, and consistency with Linux, we copy in the entire
+	// message up front.
+	if src.NumBytes() > int64(s.sendBufferSize) {
+		return 0, syserr.ErrMessageTooLong
+	}
+
+	buf := make([]byte, src.NumBytes())
+	n, err := src.CopyIn(ctx, buf)
+	if err != nil {
+		// Don't partially consume messages.
+		return 0, syserr.FromError(err)
+	}
+
+	if err := s.processMessages(ctx, buf); err != nil {
+		return 0, err
+	}
+
+	return n, nil
+}
+
+// SendMsg implements socket.Socket.SendMsg.
+func (s *Socket) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, haveDeadline bool, deadline ktime.Time, controlMessages socket.ControlMessages) (int, *syserr.Error) {
+	return s.sendMsg(t, src, to, flags, controlMessages)
+}
+
+// Write implements fs.FileOperations.Write.
+func (s *Socket) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, _ int64) (int64, error) {
+	n, err := s.sendMsg(ctx, src, nil, 0, socket.ControlMessages{})
+	return int64(n), err.ToError()
+}