summaryrefslogtreecommitdiffhomepage
path: root/pkg/tcpip/stack
diff options
context:
space:
mode:
Diffstat (limited to 'pkg/tcpip/stack')
-rw-r--r--pkg/tcpip/stack/BUILD18
-rw-r--r--pkg/tcpip/stack/conntrack.go604
-rw-r--r--pkg/tcpip/stack/forwarder_test.go6
-rw-r--r--pkg/tcpip/stack/iptables.go65
-rw-r--r--pkg/tcpip/stack/iptables_state.go40
-rw-r--r--pkg/tcpip/stack/iptables_targets.go23
-rw-r--r--pkg/tcpip/stack/iptables_types.go42
-rw-r--r--pkg/tcpip/stack/nic.go7
-rw-r--r--pkg/tcpip/stack/nic_test.go5
-rw-r--r--pkg/tcpip/stack/packet_buffer.go14
-rw-r--r--pkg/tcpip/stack/registration.go7
-rw-r--r--pkg/tcpip/stack/stack.go31
-rw-r--r--pkg/tcpip/stack/stack_options.go106
-rw-r--r--pkg/tcpip/stack/stack_test.go80
-rw-r--r--pkg/tcpip/stack/transport_demuxer.go65
15 files changed, 795 insertions, 318 deletions
diff --git a/pkg/tcpip/stack/BUILD b/pkg/tcpip/stack/BUILD
index 24f52b735..6b9a6b316 100644
--- a/pkg/tcpip/stack/BUILD
+++ b/pkg/tcpip/stack/BUILD
@@ -27,6 +27,18 @@ go_template_instance(
},
)
+go_template_instance(
+ name = "tuple_list",
+ out = "tuple_list.go",
+ package = "stack",
+ prefix = "tuple",
+ template = "//pkg/ilist:generic_list",
+ types = {
+ "Element": "*tuple",
+ "Linker": "*tuple",
+ },
+)
+
go_library(
name = "stack",
srcs = [
@@ -35,6 +47,7 @@ go_library(
"forwarder.go",
"icmp_rate_limit.go",
"iptables.go",
+ "iptables_state.go",
"iptables_targets.go",
"iptables_types.go",
"linkaddrcache.go",
@@ -48,7 +61,9 @@ go_library(
"route.go",
"stack.go",
"stack_global_state.go",
+ "stack_options.go",
"transport_demuxer.go",
+ "tuple_list.go",
],
visibility = ["//visibility:public"],
deps = [
@@ -78,6 +93,7 @@ go_test(
"transport_demuxer_test.go",
"transport_test.go",
],
+ shard_count = 20,
deps = [
":stack",
"//pkg/rand",
@@ -93,7 +109,7 @@ go_test(
"//pkg/tcpip/transport/icmp",
"//pkg/tcpip/transport/udp",
"//pkg/waiter",
- "@com_github_google_go-cmp//cmp:go_default_library",
+ "@com_github_google_go_cmp//cmp:go_default_library",
],
)
diff --git a/pkg/tcpip/stack/conntrack.go b/pkg/tcpip/stack/conntrack.go
index 05bf62788..d39baf620 100644
--- a/pkg/tcpip/stack/conntrack.go
+++ b/pkg/tcpip/stack/conntrack.go
@@ -26,28 +26,25 @@ import (
)
// Connection tracking is used to track and manipulate packets for NAT rules.
-// The connection is created for a packet if it does not exist. Every connection
-// contains two tuples (original and reply). The tuples are manipulated if there
-// is a matching NAT rule. The packet is modified by looking at the tuples in the
-// Prerouting and Output hooks.
+// The connection is created for a packet if it does not exist. Every
+// connection contains two tuples (original and reply). The tuples are
+// manipulated if there is a matching NAT rule. The packet is modified by
+// looking at the tuples in the Prerouting and Output hooks.
+//
+// Currently, only TCP tracking is supported.
+
+// Our hash table has 16K buckets.
+// TODO(gvisor.dev/issue/170): These should be tunable.
+const numBuckets = 1 << 14
// Direction of the tuple.
-type ctDirection int
+type direction int
const (
- dirOriginal ctDirection = iota
+ dirOriginal direction = iota
dirReply
)
-// Status of connection.
-// TODO(gvisor.dev/issue/170): Add other states of connection.
-type connStatus int
-
-const (
- connNew connStatus = iota
- connEstablished
-)
-
// Manipulation type for the connection.
type manipType int
@@ -56,250 +53,258 @@ const (
manipDstOutput
)
-// connTrackMutable is the manipulatable part of the tuple.
-type connTrackMutable struct {
- // addr is source address of the tuple.
- addr tcpip.Address
-
- // port is source port of the tuple.
- port uint16
-
- // protocol is network layer protocol.
- protocol tcpip.NetworkProtocolNumber
-}
-
-// connTrackImmutable is the non-manipulatable part of the tuple.
-type connTrackImmutable struct {
- // addr is destination address of the tuple.
- addr tcpip.Address
+// tuple holds a connection's identifying and manipulating data in one
+// direction. It is immutable.
+//
+// +stateify savable
+type tuple struct {
+ // tupleEntry is used to build an intrusive list of tuples.
+ tupleEntry
- // direction is direction (original or reply) of the tuple.
- direction ctDirection
+ tupleID
- // port is destination port of the tuple.
- port uint16
+ // conn is the connection tracking entry this tuple belongs to.
+ conn *conn
- // protocol is transport layer protocol.
- protocol tcpip.TransportProtocolNumber
+ // direction is the direction of the tuple.
+ direction direction
}
-// connTrackTuple represents the tuple which is created from the
-// packet.
-type connTrackTuple struct {
- // dst is non-manipulatable part of the tuple.
- dst connTrackImmutable
-
- // src is manipulatable part of the tuple.
- src connTrackMutable
+// tupleID uniquely identifies a connection in one direction. It currently
+// contains enough information to distinguish between any TCP or UDP
+// connection, and will need to be extended to support other protocols.
+//
+// +stateify savable
+type tupleID struct {
+ srcAddr tcpip.Address
+ srcPort uint16
+ dstAddr tcpip.Address
+ dstPort uint16
+ transProto tcpip.TransportProtocolNumber
+ netProto tcpip.NetworkProtocolNumber
}
-// connTrackTupleHolder is the container of tuple and connection.
-type ConnTrackTupleHolder struct {
- // conn is pointer to the connection tracking entry.
- conn *connTrack
-
- // tuple is original or reply tuple.
- tuple connTrackTuple
+// reply creates the reply tupleID.
+func (ti tupleID) reply() tupleID {
+ return tupleID{
+ srcAddr: ti.dstAddr,
+ srcPort: ti.dstPort,
+ dstAddr: ti.srcAddr,
+ dstPort: ti.srcPort,
+ transProto: ti.transProto,
+ netProto: ti.netProto,
+ }
}
-// connTrack is the connection.
-type connTrack struct {
- // originalTupleHolder contains tuple in original direction.
- originalTupleHolder ConnTrackTupleHolder
+// conn is a tracked connection.
+//
+// +stateify savable
+type conn struct {
+ // original is the tuple in original direction. It is immutable.
+ original tuple
- // replyTupleHolder contains tuple in reply direction.
- replyTupleHolder ConnTrackTupleHolder
+ // reply is the tuple in reply direction. It is immutable.
+ reply tuple
- // status indicates connection is new or established.
- status connStatus
+ // manip indicates if the packet should be manipulated. It is immutable.
+ manip manipType
- // timeout indicates the time connection should be active.
- timeout time.Duration
+ // tcbHook indicates if the packet is inbound or outbound to
+ // update the state of tcb. It is immutable.
+ tcbHook Hook
- // manip indicates if the packet should be manipulated.
- manip manipType
+ // mu protects tcb.
+ mu sync.Mutex `state:"nosave"`
// tcb is TCB control block. It is used to keep track of states
- // of tcp connection.
+ // of tcp connection and is protected by mu.
tcb tcpconntrack.TCB
- // tcbHook indicates if the packet is inbound or outbound to
- // update the state of tcb.
- tcbHook Hook
+ // lastUsed is the last time the connection saw a relevant packet, and
+ // is updated by each packet on the connection. It is protected by mu.
+ lastUsed time.Time `state:".(unixTime)"`
}
-// ConnTrackTable contains a map of all existing connections created for
-// NAT rules.
-type ConnTrackTable struct {
- // connMu protects connTrackTable.
- connMu sync.RWMutex
-
- // connTrackTable maintains a map of tuples needed for connection tracking
- // for iptables NAT rules. The key for the map is an integer calculated
- // using seed, source address, destination address, source port and
- // destination port.
- CtMap map[uint32]ConnTrackTupleHolder
+// timedOut returns whether the connection timed out based on its state.
+func (cn *conn) timedOut(now time.Time) bool {
+ const establishedTimeout = 5 * 24 * time.Hour
+ const defaultTimeout = 120 * time.Second
+ cn.mu.Lock()
+ defer cn.mu.Unlock()
+ if cn.tcb.State() == tcpconntrack.ResultAlive {
+ // Use the same default as Linux, which doesn't delete
+ // established connections for 5(!) days.
+ return now.Sub(cn.lastUsed) > establishedTimeout
+ }
+ // Use the same default as Linux, which lets connections in most states
+ // other than established remain for <= 120 seconds.
+ return now.Sub(cn.lastUsed) > defaultTimeout
+}
+// ConnTrack tracks all connections created for NAT rules. Most users are
+// expected to only call handlePacket and createConnFor.
+//
+// ConnTrack keeps all connections in a slice of buckets, each of which holds a
+// linked list of tuples. This gives us some desirable properties:
+// - Each bucket has its own lock, lessening lock contention.
+// - The slice is large enough that lists stay short (<10 elements on average).
+// Thus traversal is fast.
+// - During linked list traversal we reap expired connections. This amortizes
+// the cost of reaping them and makes reapUnused faster.
+//
+// Locks are ordered by their location in the buckets slice. That is, a
+// goroutine that locks buckets[i] can only lock buckets[j] s.t. i < j.
+//
+// +stateify savable
+type ConnTrack struct {
// seed is a one-time random value initialized at stack startup
- // and is used in calculation of hash key for connection tracking
- // table.
- Seed uint32
+ // and is used in the calculation of hash keys for the list of buckets.
+ // It is immutable.
+ seed uint32
+
+ // mu protects the buckets slice, but not buckets' contents. Only take
+ // the write lock if you are modifying the slice or saving for S/R.
+ mu sync.RWMutex `state:"nosave"`
+
+ // buckets is protected by mu.
+ buckets []bucket
}
-// packetToTuple converts packet to a tuple in original direction.
-func packetToTuple(pkt *PacketBuffer, hook Hook) (connTrackTuple, *tcpip.Error) {
- var tuple connTrackTuple
+// +stateify savable
+type bucket struct {
+ // mu protects tuples.
+ mu sync.Mutex `state:"nosave"`
+ tuples tupleList
+}
- netHeader := header.IPv4(pkt.NetworkHeader)
+// packetToTupleID converts packet to a tuple ID. It fails when pkt lacks a valid
+// TCP header.
+func packetToTupleID(pkt *PacketBuffer) (tupleID, *tcpip.Error) {
// TODO(gvisor.dev/issue/170): Need to support for other
// protocols as well.
+ netHeader := header.IPv4(pkt.NetworkHeader)
if netHeader == nil || netHeader.TransportProtocol() != header.TCPProtocolNumber {
- return tuple, tcpip.ErrUnknownProtocol
+ return tupleID{}, tcpip.ErrUnknownProtocol
}
tcpHeader := header.TCP(pkt.TransportHeader)
if tcpHeader == nil {
- return tuple, tcpip.ErrUnknownProtocol
+ return tupleID{}, tcpip.ErrUnknownProtocol
}
- tuple.src.addr = netHeader.SourceAddress()
- tuple.src.port = tcpHeader.SourcePort()
- tuple.src.protocol = header.IPv4ProtocolNumber
-
- tuple.dst.addr = netHeader.DestinationAddress()
- tuple.dst.port = tcpHeader.DestinationPort()
- tuple.dst.protocol = netHeader.TransportProtocol()
-
- return tuple, nil
+ return tupleID{
+ srcAddr: netHeader.SourceAddress(),
+ srcPort: tcpHeader.SourcePort(),
+ dstAddr: netHeader.DestinationAddress(),
+ dstPort: tcpHeader.DestinationPort(),
+ transProto: netHeader.TransportProtocol(),
+ netProto: header.IPv4ProtocolNumber,
+ }, nil
}
-// getReplyTuple creates reply tuple for the given tuple.
-func getReplyTuple(tuple connTrackTuple) connTrackTuple {
- var replyTuple connTrackTuple
- replyTuple.src.addr = tuple.dst.addr
- replyTuple.src.port = tuple.dst.port
- replyTuple.src.protocol = tuple.src.protocol
- replyTuple.dst.addr = tuple.src.addr
- replyTuple.dst.port = tuple.src.port
- replyTuple.dst.protocol = tuple.dst.protocol
- replyTuple.dst.direction = dirReply
-
- return replyTuple
-}
-
-// makeNewConn creates new connection.
-func makeNewConn(tuple, replyTuple connTrackTuple) connTrack {
- var conn connTrack
- conn.status = connNew
- conn.originalTupleHolder.tuple = tuple
- conn.originalTupleHolder.conn = &conn
- conn.replyTupleHolder.tuple = replyTuple
- conn.replyTupleHolder.conn = &conn
-
- return conn
-}
-
-// getTupleHash returns hash of the tuple. The fields used for
-// generating hash are seed (generated once for stack), source address,
-// destination address, source port and destination ports.
-func (ct *ConnTrackTable) getTupleHash(tuple connTrackTuple) uint32 {
- h := jenkins.Sum32(ct.Seed)
- h.Write([]byte(tuple.src.addr))
- h.Write([]byte(tuple.dst.addr))
- portBuf := make([]byte, 2)
- binary.LittleEndian.PutUint16(portBuf, tuple.src.port)
- h.Write([]byte(portBuf))
- binary.LittleEndian.PutUint16(portBuf, tuple.dst.port)
- h.Write([]byte(portBuf))
-
- return h.Sum32()
+// newConn creates new connection.
+func newConn(orig, reply tupleID, manip manipType, hook Hook) *conn {
+ conn := conn{
+ manip: manip,
+ tcbHook: hook,
+ lastUsed: time.Now(),
+ }
+ conn.original = tuple{conn: &conn, tupleID: orig}
+ conn.reply = tuple{conn: &conn, tupleID: reply, direction: dirReply}
+ return &conn
}
-// connTrackForPacket returns connTrack for packet.
-// TODO(gvisor.dev/issue/170): Only TCP packets are supported. Need to support other
-// transport protocols.
-func (ct *ConnTrackTable) connTrackForPacket(pkt *PacketBuffer, hook Hook, createConn bool) (*connTrack, ctDirection) {
- var dir ctDirection
- tuple, err := packetToTuple(pkt, hook)
+// connFor gets the conn for pkt if it exists, or returns nil
+// if it does not. It returns an error when pkt does not contain a valid TCP
+// header.
+// TODO(gvisor.dev/issue/170): Only TCP packets are supported. Need to support
+// other transport protocols.
+func (ct *ConnTrack) connFor(pkt *PacketBuffer) (*conn, direction) {
+ tid, err := packetToTupleID(pkt)
if err != nil {
- return nil, dir
+ return nil, dirOriginal
}
- ct.connMu.Lock()
- defer ct.connMu.Unlock()
-
- connTrackTable := ct.CtMap
- hash := ct.getTupleHash(tuple)
-
- var conn *connTrack
- switch createConn {
- case true:
- // If connection does not exist for the hash, create a new
- // connection.
- replyTuple := getReplyTuple(tuple)
- replyHash := ct.getTupleHash(replyTuple)
- newConn := makeNewConn(tuple, replyTuple)
- conn = &newConn
-
- // Add tupleHolders to the map.
- // TODO(gvisor.dev/issue/170): Need to support collisions using linked list.
- ct.CtMap[hash] = conn.originalTupleHolder
- ct.CtMap[replyHash] = conn.replyTupleHolder
- default:
- tupleHolder, ok := connTrackTable[hash]
- if !ok {
- return nil, dir
- }
-
- // If this is the reply of new connection, set the connection
- // status as ESTABLISHED.
- conn = tupleHolder.conn
- if conn.status == connNew && tupleHolder.tuple.dst.direction == dirReply {
- conn.status = connEstablished
+ bucket := ct.bucket(tid)
+ now := time.Now()
+
+ ct.mu.RLock()
+ defer ct.mu.RUnlock()
+ ct.buckets[bucket].mu.Lock()
+ defer ct.buckets[bucket].mu.Unlock()
+
+ // Iterate over the tuples in a bucket, cleaning up any unused
+ // connections we find.
+ for other := ct.buckets[bucket].tuples.Front(); other != nil; other = other.Next() {
+ // Clean up any timed-out connections we happen to find.
+ if ct.reapTupleLocked(other, bucket, now) {
+ // The tuple expired.
+ continue
}
- if tupleHolder.conn == nil {
- panic("tupleHolder has null connection tracking entry")
+ if tid == other.tupleID {
+ return other.conn, other.direction
}
-
- dir = tupleHolder.tuple.dst.direction
}
- return conn, dir
+
+ return nil, dirOriginal
}
-// SetNatInfo will manipulate the tuples according to iptables NAT rules.
-func (ct *ConnTrackTable) SetNatInfo(pkt *PacketBuffer, rt RedirectTarget, hook Hook) {
- // Get the connection. Connection is always created before this
- // function is called.
- conn, _ := ct.connTrackForPacket(pkt, hook, false)
- if conn == nil {
- panic("connection should be created to manipulate tuples.")
+// createConnFor creates a new conn for pkt.
+func (ct *ConnTrack) createConnFor(pkt *PacketBuffer, hook Hook, rt RedirectTarget) *conn {
+ tid, err := packetToTupleID(pkt)
+ if err != nil {
+ return nil
+ }
+ if hook != Prerouting && hook != Output {
+ return nil
+ }
+
+ // Create a new connection and change the port as per the iptables
+ // rule. This tuple will be used to manipulate the packet in
+ // handlePacket.
+ replyTID := tid.reply()
+ replyTID.srcAddr = rt.MinIP
+ replyTID.srcPort = rt.MinPort
+ var manip manipType
+ switch hook {
+ case Prerouting:
+ manip = manipDstPrerouting
+ case Output:
+ manip = manipDstOutput
}
- replyTuple := conn.replyTupleHolder.tuple
- replyHash := ct.getTupleHash(replyTuple)
-
- // TODO(gvisor.dev/issue/170): Support only redirect of ports. Need to
- // support changing of address for Prerouting.
-
- // Change the port as per the iptables rule. This tuple will be used
- // to manipulate the packet in HandlePacket.
- conn.replyTupleHolder.tuple.src.addr = rt.MinIP
- conn.replyTupleHolder.tuple.src.port = rt.MinPort
- newHash := ct.getTupleHash(conn.replyTupleHolder.tuple)
-
- // Add the changed tuple to the map.
- ct.connMu.Lock()
- defer ct.connMu.Unlock()
- ct.CtMap[newHash] = conn.replyTupleHolder
- if hook == Output {
- conn.replyTupleHolder.conn.manip = manipDstOutput
+ conn := newConn(tid, replyTID, manip, hook)
+
+ // Lock the buckets in the correct order.
+ tupleBucket := ct.bucket(tid)
+ replyBucket := ct.bucket(replyTID)
+ ct.mu.RLock()
+ defer ct.mu.RUnlock()
+ if tupleBucket < replyBucket {
+ ct.buckets[tupleBucket].mu.Lock()
+ ct.buckets[replyBucket].mu.Lock()
+ } else if tupleBucket > replyBucket {
+ ct.buckets[replyBucket].mu.Lock()
+ ct.buckets[tupleBucket].mu.Lock()
+ } else {
+ // Both tuples are in the same bucket.
+ ct.buckets[tupleBucket].mu.Lock()
}
- // Delete the old tuple.
- delete(ct.CtMap, replyHash)
+ // Add the tuple to the map.
+ ct.buckets[tupleBucket].tuples.PushFront(&conn.original)
+ ct.buckets[replyBucket].tuples.PushFront(&conn.reply)
+
+ // Unlocking can happen in any order.
+ ct.buckets[tupleBucket].mu.Unlock()
+ if tupleBucket != replyBucket {
+ ct.buckets[replyBucket].mu.Unlock()
+ }
+
+ return conn
}
// handlePacketPrerouting manipulates ports for packets in Prerouting hook.
-// TODO(gvisor.dev/issue/170): Change address for Prerouting hook..
-func handlePacketPrerouting(pkt *PacketBuffer, conn *connTrack, dir ctDirection) {
+// TODO(gvisor.dev/issue/170): Change address for Prerouting hook.
+func handlePacketPrerouting(pkt *PacketBuffer, conn *conn, dir direction) {
netHeader := header.IPv4(pkt.NetworkHeader)
tcpHeader := header.TCP(pkt.TransportHeader)
@@ -308,13 +313,13 @@ func handlePacketPrerouting(pkt *PacketBuffer, conn *connTrack, dir ctDirection)
// modified.
switch dir {
case dirOriginal:
- port := conn.replyTupleHolder.tuple.src.port
+ port := conn.reply.srcPort
tcpHeader.SetDestinationPort(port)
- netHeader.SetDestinationAddress(conn.replyTupleHolder.tuple.src.addr)
+ netHeader.SetDestinationAddress(conn.reply.srcAddr)
case dirReply:
- port := conn.originalTupleHolder.tuple.dst.port
+ port := conn.original.dstPort
tcpHeader.SetSourcePort(port)
- netHeader.SetSourceAddress(conn.originalTupleHolder.tuple.dst.addr)
+ netHeader.SetSourceAddress(conn.original.dstAddr)
}
netHeader.SetChecksum(0)
@@ -322,7 +327,7 @@ func handlePacketPrerouting(pkt *PacketBuffer, conn *connTrack, dir ctDirection)
}
// handlePacketOutput manipulates ports for packets in Output hook.
-func handlePacketOutput(pkt *PacketBuffer, conn *connTrack, gso *GSO, r *Route, dir ctDirection) {
+func handlePacketOutput(pkt *PacketBuffer, conn *conn, gso *GSO, r *Route, dir direction) {
netHeader := header.IPv4(pkt.NetworkHeader)
tcpHeader := header.TCP(pkt.TransportHeader)
@@ -331,13 +336,13 @@ func handlePacketOutput(pkt *PacketBuffer, conn *connTrack, gso *GSO, r *Route,
// modified. For prerouting redirection, we only reach this point
// when replying, so packet sources are modified.
if conn.manip == manipDstOutput && dir == dirOriginal {
- port := conn.replyTupleHolder.tuple.src.port
+ port := conn.reply.srcPort
tcpHeader.SetDestinationPort(port)
- netHeader.SetDestinationAddress(conn.replyTupleHolder.tuple.src.addr)
+ netHeader.SetDestinationAddress(conn.reply.srcAddr)
} else {
- port := conn.originalTupleHolder.tuple.dst.port
+ port := conn.original.dstPort
tcpHeader.SetSourcePort(port)
- netHeader.SetSourceAddress(conn.originalTupleHolder.tuple.dst.addr)
+ netHeader.SetSourceAddress(conn.original.dstAddr)
}
// Calculate the TCP checksum and set it.
@@ -356,9 +361,9 @@ func handlePacketOutput(pkt *PacketBuffer, conn *connTrack, gso *GSO, r *Route,
netHeader.SetChecksum(^netHeader.CalculateChecksum())
}
-// HandlePacket will manipulate the port and address of the packet if the
+// handlePacket will manipulate the port and address of the packet if the
// connection exists.
-func (ct *ConnTrackTable) HandlePacket(pkt *PacketBuffer, hook Hook, gso *GSO, r *Route) {
+func (ct *ConnTrack) handlePacket(pkt *PacketBuffer, hook Hook, gso *GSO, r *Route) {
if pkt.NatDone {
return
}
@@ -367,21 +372,9 @@ func (ct *ConnTrackTable) HandlePacket(pkt *PacketBuffer, hook Hook, gso *GSO, r
return
}
- conn, dir := ct.connTrackForPacket(pkt, hook, false)
- // Connection or Rule not found for the packet.
+ conn, dir := ct.connFor(pkt)
if conn == nil {
- return
- }
-
- netHeader := header.IPv4(pkt.NetworkHeader)
- // TODO(gvisor.dev/issue/170): Need to support for other transport
- // protocols as well.
- if netHeader == nil || netHeader.TransportProtocol() != header.TCPProtocolNumber {
- return
- }
-
- tcpHeader := header.TCP(pkt.TransportHeader)
- if tcpHeader == nil {
+ // Connection not found for the packet or the packet is invalid.
return
}
@@ -396,39 +389,136 @@ func (ct *ConnTrackTable) HandlePacket(pkt *PacketBuffer, hook Hook, gso *GSO, r
// Update the state of tcb.
// TODO(gvisor.dev/issue/170): Add support in tcpcontrack to handle
// other tcp states.
- var st tcpconntrack.Result
- if conn.tcb.IsEmpty() {
+ conn.mu.Lock()
+ defer conn.mu.Unlock()
+
+ // Mark the connection as having been used recently so it isn't reaped.
+ conn.lastUsed = time.Now()
+ // Update connection state.
+ if tcpHeader := header.TCP(pkt.TransportHeader); conn.tcb.IsEmpty() {
conn.tcb.Init(tcpHeader)
conn.tcbHook = hook
+ } else if hook == conn.tcbHook {
+ conn.tcb.UpdateStateOutbound(tcpHeader)
} else {
- switch hook {
- case conn.tcbHook:
- st = conn.tcb.UpdateStateOutbound(tcpHeader)
- default:
- st = conn.tcb.UpdateStateInbound(tcpHeader)
- }
+ conn.tcb.UpdateStateInbound(tcpHeader)
}
+}
- // Delete conntrack if tcp connection is closed.
- if st == tcpconntrack.ResultClosedByPeer || st == tcpconntrack.ResultClosedBySelf || st == tcpconntrack.ResultReset {
- ct.deleteConnTrack(conn)
+// bucket gets the conntrack bucket for a tupleID.
+func (ct *ConnTrack) bucket(id tupleID) int {
+ h := jenkins.Sum32(ct.seed)
+ h.Write([]byte(id.srcAddr))
+ h.Write([]byte(id.dstAddr))
+ shortBuf := make([]byte, 2)
+ binary.LittleEndian.PutUint16(shortBuf, id.srcPort)
+ h.Write([]byte(shortBuf))
+ binary.LittleEndian.PutUint16(shortBuf, id.dstPort)
+ h.Write([]byte(shortBuf))
+ binary.LittleEndian.PutUint16(shortBuf, uint16(id.transProto))
+ h.Write([]byte(shortBuf))
+ binary.LittleEndian.PutUint16(shortBuf, uint16(id.netProto))
+ h.Write([]byte(shortBuf))
+ ct.mu.RLock()
+ defer ct.mu.RUnlock()
+ return int(h.Sum32()) % len(ct.buckets)
+}
+
+// reapUnused deletes timed out entries from the conntrack map. The rules for
+// reaping are:
+// - Most reaping occurs in connFor, which is called on each packet. connFor
+// cleans up the bucket the packet's connection maps to. Thus calls to
+// reapUnused should be fast.
+// - Each call to reapUnused traverses a fraction of the conntrack table.
+// Specifically, it traverses len(ct.buckets)/fractionPerReaping.
+// - After reaping, reapUnused decides when it should next run based on the
+// ratio of expired connections to examined connections. If the ratio is
+// greater than maxExpiredPct, it schedules the next run quickly. Otherwise it
+// slightly increases the interval between runs.
+// - maxFullTraversal caps the time it takes to traverse the entire table.
+//
+// reapUnused returns the next bucket that should be checked and the time after
+// which it should be called again.
+func (ct *ConnTrack) reapUnused(start int, prevInterval time.Duration) (int, time.Duration) {
+ // TODO(gvisor.dev/issue/170): This can be more finely controlled, as
+ // it is in Linux via sysctl.
+ const fractionPerReaping = 128
+ const maxExpiredPct = 50
+ const maxFullTraversal = 60 * time.Second
+ const minInterval = 10 * time.Millisecond
+ const maxInterval = maxFullTraversal / fractionPerReaping
+
+ now := time.Now()
+ checked := 0
+ expired := 0
+ var idx int
+ ct.mu.RLock()
+ defer ct.mu.RUnlock()
+ for i := 0; i < len(ct.buckets)/fractionPerReaping; i++ {
+ idx = (i + start) % len(ct.buckets)
+ ct.buckets[idx].mu.Lock()
+ for tuple := ct.buckets[idx].tuples.Front(); tuple != nil; tuple = tuple.Next() {
+ checked++
+ if ct.reapTupleLocked(tuple, idx, now) {
+ expired++
+ }
+ }
+ ct.buckets[idx].mu.Unlock()
+ }
+ // We already checked buckets[idx].
+ idx++
+
+ // If half or more of the connections are expired, the table has gotten
+ // stale. Reschedule quickly.
+ expiredPct := 0
+ if checked != 0 {
+ expiredPct = expired * 100 / checked
+ }
+ if expiredPct > maxExpiredPct {
+ return idx, minInterval
}
+ if interval := prevInterval + minInterval; interval <= maxInterval {
+ // Increment the interval between runs.
+ return idx, interval
+ }
+ // We've hit the maximum interval.
+ return idx, maxInterval
}
-// deleteConnTrack deletes the connection.
-func (ct *ConnTrackTable) deleteConnTrack(conn *connTrack) {
- if conn == nil {
- return
+// reapTupleLocked tries to remove tuple and its reply from the table. It
+// returns whether the tuple's connection has timed out.
+//
+// Preconditions: ct.mu is locked for reading and bucket is locked.
+func (ct *ConnTrack) reapTupleLocked(tuple *tuple, bucket int, now time.Time) bool {
+ if !tuple.conn.timedOut(now) {
+ return false
}
- tuple := conn.originalTupleHolder.tuple
- hash := ct.getTupleHash(tuple)
- replyTuple := conn.replyTupleHolder.tuple
- replyHash := ct.getTupleHash(replyTuple)
+ // To maintain lock order, we can only reap these tuples if the reply
+ // appears later in the table.
+ replyBucket := ct.bucket(tuple.reply())
+ if bucket > replyBucket {
+ return true
+ }
- ct.connMu.Lock()
- defer ct.connMu.Unlock()
+ // Don't re-lock if both tuples are in the same bucket.
+ differentBuckets := bucket != replyBucket
+ if differentBuckets {
+ ct.buckets[replyBucket].mu.Lock()
+ }
+
+ // We have the buckets locked and can remove both tuples.
+ if tuple.direction == dirOriginal {
+ ct.buckets[replyBucket].tuples.Remove(&tuple.conn.reply)
+ } else {
+ ct.buckets[replyBucket].tuples.Remove(&tuple.conn.original)
+ }
+ ct.buckets[bucket].tuples.Remove(tuple)
+
+ // Don't re-unlock if both tuples are in the same bucket.
+ if differentBuckets {
+ ct.buckets[replyBucket].mu.Unlock()
+ }
- delete(ct.CtMap, hash)
- delete(ct.CtMap, replyHash)
+ return true
}
diff --git a/pkg/tcpip/stack/forwarder_test.go b/pkg/tcpip/stack/forwarder_test.go
index a6546cef0..eefb4b07f 100644
--- a/pkg/tcpip/stack/forwarder_test.go
+++ b/pkg/tcpip/stack/forwarder_test.go
@@ -22,6 +22,7 @@ import (
"gvisor.dev/gvisor/pkg/tcpip"
"gvisor.dev/gvisor/pkg/tcpip/buffer"
+ "gvisor.dev/gvisor/pkg/tcpip/header"
)
const (
@@ -301,6 +302,11 @@ func (e *fwdTestLinkEndpoint) WriteRawPacket(vv buffer.VectorisedView) *tcpip.Er
// Wait implements stack.LinkEndpoint.Wait.
func (*fwdTestLinkEndpoint) Wait() {}
+// ARPHardwareType implements stack.LinkEndpoint.ARPHardwareType.
+func (*fwdTestLinkEndpoint) ARPHardwareType() header.ARPHardwareType {
+ panic("not implemented")
+}
+
func fwdTestNetFactory(t *testing.T, proto *fwdTestNetworkProtocol) (ep1, ep2 *fwdTestLinkEndpoint) {
// Create a stack with the network protocol and two NICs.
s := New(Options{
diff --git a/pkg/tcpip/stack/iptables.go b/pkg/tcpip/stack/iptables.go
index 4e9b404c8..f846ea2e5 100644
--- a/pkg/tcpip/stack/iptables.go
+++ b/pkg/tcpip/stack/iptables.go
@@ -16,6 +16,7 @@ package stack
import (
"fmt"
+ "time"
"gvisor.dev/gvisor/pkg/tcpip"
"gvisor.dev/gvisor/pkg/tcpip/header"
@@ -41,6 +42,9 @@ const (
// underflow.
const HookUnset = -1
+// reaperDelay is how long to wait before starting to reap connections.
+const reaperDelay = 5 * time.Second
+
// DefaultTables returns a default set of tables. Each chain is set to accept
// all packets.
func DefaultTables() *IPTables {
@@ -111,10 +115,10 @@ func DefaultTables() *IPTables {
Prerouting: []string{TablenameMangle, TablenameNat},
Output: []string{TablenameMangle, TablenameNat, TablenameFilter},
},
- connections: ConnTrackTable{
- CtMap: make(map[uint32]ConnTrackTupleHolder),
- Seed: generateRandUint32(),
+ connections: ConnTrack{
+ seed: generateRandUint32(),
},
+ reaperDone: make(chan struct{}, 1),
}
}
@@ -170,17 +174,16 @@ func (it *IPTables) GetTable(name string) (Table, bool) {
func (it *IPTables) ReplaceTable(name string, table Table) {
it.mu.Lock()
defer it.mu.Unlock()
+ // If iptables is being enabled, initialize the conntrack table and
+ // reaper.
+ if !it.modified {
+ it.connections.buckets = make([]bucket, numBuckets)
+ it.startReaper(reaperDelay)
+ }
+ it.modified = true
it.tables[name] = table
}
-// ModifyTables acquires write-lock and calls fn with internal name-to-table
-// map. This function can be used to update multiple tables atomically.
-func (it *IPTables) ModifyTables(fn func(map[string]Table)) {
- it.mu.Lock()
- defer it.mu.Unlock()
- fn(it.tables)
-}
-
// GetPriorities returns slice of priorities associated with hook.
func (it *IPTables) GetPriorities(hook Hook) []string {
it.mu.RLock()
@@ -209,9 +212,18 @@ const (
//
// Precondition: pkt.NetworkHeader is set.
func (it *IPTables) Check(hook Hook, pkt *PacketBuffer, gso *GSO, r *Route, address tcpip.Address, nicName string) bool {
+ // Many users never configure iptables. Spare them the cost of rule
+ // traversal if rules have never been set.
+ it.mu.RLock()
+ if !it.modified {
+ it.mu.RUnlock()
+ return true
+ }
+ it.mu.RUnlock()
+
// Packets are manipulated only if connection and matching
// NAT rule exists.
- it.connections.HandlePacket(pkt, hook, gso, r)
+ it.connections.handlePacket(pkt, hook, gso, r)
// Go through each table containing the hook.
for _, tablename := range it.GetPriorities(hook) {
@@ -248,6 +260,35 @@ func (it *IPTables) Check(hook Hook, pkt *PacketBuffer, gso *GSO, r *Route, addr
return true
}
+// beforeSave is invoked by stateify.
+func (it *IPTables) beforeSave() {
+ // Ensure the reaper exits cleanly.
+ it.reaperDone <- struct{}{}
+ // Prevent others from modifying the connection table.
+ it.connections.mu.Lock()
+}
+
+// afterLoad is invoked by stateify.
+func (it *IPTables) afterLoad() {
+ it.startReaper(reaperDelay)
+}
+
+// startReaper starts a goroutine that wakes up periodically to reap timed out
+// connections.
+func (it *IPTables) startReaper(interval time.Duration) {
+ go func() { // S/R-SAFE: reaperDone is signalled when iptables is saved.
+ bucket := 0
+ for {
+ select {
+ case <-it.reaperDone:
+ return
+ case <-time.After(interval):
+ bucket, interval = it.connections.reapUnused(bucket, interval)
+ }
+ }
+ }()
+}
+
// CheckPackets runs pkts through the rules for hook and returns a map of packets that
// should not go forward.
//
diff --git a/pkg/tcpip/stack/iptables_state.go b/pkg/tcpip/stack/iptables_state.go
new file mode 100644
index 000000000..529e02a07
--- /dev/null
+++ b/pkg/tcpip/stack/iptables_state.go
@@ -0,0 +1,40 @@
+// Copyright 2020 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 stack
+
+import (
+ "time"
+)
+
+// +stateify savable
+type unixTime struct {
+ second int64
+ nano int64
+}
+
+// saveLastUsed is invoked by stateify.
+func (cn *conn) saveLastUsed() unixTime {
+ return unixTime{cn.lastUsed.Unix(), cn.lastUsed.UnixNano()}
+}
+
+// loadLastUsed is invoked by stateify.
+func (cn *conn) loadLastUsed(unix unixTime) {
+ cn.lastUsed = time.Unix(unix.second, unix.nano)
+}
+
+// beforeSave is invoked by stateify.
+func (ct *ConnTrack) beforeSave() {
+ ct.mu.Lock()
+}
diff --git a/pkg/tcpip/stack/iptables_targets.go b/pkg/tcpip/stack/iptables_targets.go
index 92e31643e..d43f60c67 100644
--- a/pkg/tcpip/stack/iptables_targets.go
+++ b/pkg/tcpip/stack/iptables_targets.go
@@ -24,7 +24,7 @@ import (
type AcceptTarget struct{}
// Action implements Target.Action.
-func (AcceptTarget) Action(*PacketBuffer, *ConnTrackTable, Hook, *GSO, *Route, tcpip.Address) (RuleVerdict, int) {
+func (AcceptTarget) Action(*PacketBuffer, *ConnTrack, Hook, *GSO, *Route, tcpip.Address) (RuleVerdict, int) {
return RuleAccept, 0
}
@@ -32,7 +32,7 @@ func (AcceptTarget) Action(*PacketBuffer, *ConnTrackTable, Hook, *GSO, *Route, t
type DropTarget struct{}
// Action implements Target.Action.
-func (DropTarget) Action(*PacketBuffer, *ConnTrackTable, Hook, *GSO, *Route, tcpip.Address) (RuleVerdict, int) {
+func (DropTarget) Action(*PacketBuffer, *ConnTrack, Hook, *GSO, *Route, tcpip.Address) (RuleVerdict, int) {
return RuleDrop, 0
}
@@ -41,7 +41,7 @@ func (DropTarget) Action(*PacketBuffer, *ConnTrackTable, Hook, *GSO, *Route, tcp
type ErrorTarget struct{}
// Action implements Target.Action.
-func (ErrorTarget) Action(*PacketBuffer, *ConnTrackTable, Hook, *GSO, *Route, tcpip.Address) (RuleVerdict, int) {
+func (ErrorTarget) Action(*PacketBuffer, *ConnTrack, Hook, *GSO, *Route, tcpip.Address) (RuleVerdict, int) {
log.Debugf("ErrorTarget triggered.")
return RuleDrop, 0
}
@@ -52,7 +52,7 @@ type UserChainTarget struct {
}
// Action implements Target.Action.
-func (UserChainTarget) Action(*PacketBuffer, *ConnTrackTable, Hook, *GSO, *Route, tcpip.Address) (RuleVerdict, int) {
+func (UserChainTarget) Action(*PacketBuffer, *ConnTrack, Hook, *GSO, *Route, tcpip.Address) (RuleVerdict, int) {
panic("UserChainTarget should never be called.")
}
@@ -61,7 +61,7 @@ func (UserChainTarget) Action(*PacketBuffer, *ConnTrackTable, Hook, *GSO, *Route
type ReturnTarget struct{}
// Action implements Target.Action.
-func (ReturnTarget) Action(*PacketBuffer, *ConnTrackTable, Hook, *GSO, *Route, tcpip.Address) (RuleVerdict, int) {
+func (ReturnTarget) Action(*PacketBuffer, *ConnTrack, Hook, *GSO, *Route, tcpip.Address) (RuleVerdict, int) {
return RuleReturn, 0
}
@@ -92,7 +92,7 @@ type RedirectTarget struct {
// TODO(gvisor.dev/issue/170): Parse headers without copying. The current
// implementation only works for PREROUTING and calls pkt.Clone(), neither
// of which should be the case.
-func (rt RedirectTarget) Action(pkt *PacketBuffer, ct *ConnTrackTable, hook Hook, gso *GSO, r *Route, address tcpip.Address) (RuleVerdict, int) {
+func (rt RedirectTarget) Action(pkt *PacketBuffer, ct *ConnTrack, hook Hook, gso *GSO, r *Route, address tcpip.Address) (RuleVerdict, int) {
// Packet is already manipulated.
if pkt.NatDone {
return RuleAccept, 0
@@ -150,12 +150,11 @@ func (rt RedirectTarget) Action(pkt *PacketBuffer, ct *ConnTrackTable, hook Hook
return RuleAccept, 0
}
- // Set up conection for matching NAT rule.
- // Only the first packet of the connection comes here.
- // Other packets will be manipulated in connection tracking.
- if conn, _ := ct.connTrackForPacket(pkt, hook, true); conn != nil {
- ct.SetNatInfo(pkt, rt, hook)
- ct.HandlePacket(pkt, hook, gso, r)
+ // Set up conection for matching NAT rule. Only the first
+ // packet of the connection comes here. Other packets will be
+ // manipulated in connection tracking.
+ if conn := ct.createConnFor(pkt, hook, rt); conn != nil {
+ ct.handlePacket(pkt, hook, gso, r)
}
default:
return RuleDrop, 0
diff --git a/pkg/tcpip/stack/iptables_types.go b/pkg/tcpip/stack/iptables_types.go
index 4a6a5c6f1..eb70e3104 100644
--- a/pkg/tcpip/stack/iptables_types.go
+++ b/pkg/tcpip/stack/iptables_types.go
@@ -78,12 +78,14 @@ const (
)
// IPTables holds all the tables for a netstack.
+//
+// +stateify savable
type IPTables struct {
- // mu protects tables and priorities.
+ // mu protects tables, priorities, and modified.
mu sync.RWMutex
- // tables maps table names to tables. User tables have arbitrary names. mu
- // needs to be locked for accessing.
+ // tables maps table names to tables. User tables have arbitrary names.
+ // mu needs to be locked for accessing.
tables map[string]Table
// priorities maps each hook to a list of table names. The order of the
@@ -91,11 +93,21 @@ type IPTables struct {
// hook. mu needs to be locked for accessing.
priorities map[Hook][]string
- connections ConnTrackTable
+ // modified is whether tables have been modified at least once. It is
+ // used to elide the iptables performance overhead for workloads that
+ // don't utilize iptables.
+ modified bool
+
+ connections ConnTrack
+
+ // reaperDone can be signalled to stop the reaper goroutine.
+ reaperDone chan struct{}
}
// A Table defines a set of chains and hooks into the network stack. It is
-// really just a list of rules with some metadata for entrypoints and such.
+// really just a list of rules.
+//
+// +stateify savable
type Table struct {
// Rules holds the rules that make up the table.
Rules []Rule
@@ -110,10 +122,6 @@ type Table struct {
// UserChains holds user-defined chains for the keyed by name. Users
// can give their chains arbitrary names.
UserChains map[string]int
-
- // Metadata holds information about the Table that is useful to users
- // of IPTables, but not to the netstack IPTables code itself.
- metadata interface{}
}
// ValidHooks returns a bitmap of the builtin hooks for the given table.
@@ -125,20 +133,12 @@ func (table *Table) ValidHooks() uint32 {
return hooks
}
-// Metadata returns the metadata object stored in table.
-func (table *Table) Metadata() interface{} {
- return table.metadata
-}
-
-// SetMetadata sets the metadata object stored in table.
-func (table *Table) SetMetadata(metadata interface{}) {
- table.metadata = metadata
-}
-
// A Rule is a packet processing rule. It consists of two pieces. First it
// contains zero or more matchers, each of which is a specification of which
// packets this rule applies to. If there are no matchers in the rule, it
// applies to any packet.
+//
+// +stateify savable
type Rule struct {
// Filter holds basic IP filtering fields common to every rule.
Filter IPHeaderFilter
@@ -151,6 +151,8 @@ type Rule struct {
}
// IPHeaderFilter holds basic IP filtering data common to every rule.
+//
+// +stateify savable
type IPHeaderFilter struct {
// Protocol matches the transport protocol.
Protocol tcpip.TransportProtocolNumber
@@ -258,5 +260,5 @@ type Target interface {
// Action takes an action on the packet and returns a verdict on how
// traversal should (or should not) continue. If the return value is
// Jump, it also returns the index of the rule to jump to.
- Action(packet *PacketBuffer, connections *ConnTrackTable, hook Hook, gso *GSO, r *Route, address tcpip.Address) (RuleVerdict, int)
+ Action(packet *PacketBuffer, connections *ConnTrack, hook Hook, gso *GSO, r *Route, address tcpip.Address) (RuleVerdict, int)
}
diff --git a/pkg/tcpip/stack/nic.go b/pkg/tcpip/stack/nic.go
index afb7dfeaf..7b80534e6 100644
--- a/pkg/tcpip/stack/nic.go
+++ b/pkg/tcpip/stack/nic.go
@@ -1358,16 +1358,19 @@ func (n *NIC) DeliverTransportPacket(r *Route, protocol tcpip.TransportProtocolN
// TransportHeader is nil only when pkt is an ICMP packet or was reassembled
// from fragments.
if pkt.TransportHeader == nil {
- // TODO(gvisor.dev/issue/170): ICMP packets don't have their
- // TransportHeader fields set. See icmp/protocol.go:protocol.Parse for a
+ // TODO(gvisor.dev/issue/170): ICMP packets don't have their TransportHeader
+ // fields set yet, parse it here. See icmp/protocol.go:protocol.Parse for a
// full explanation.
if protocol == header.ICMPv4ProtocolNumber || protocol == header.ICMPv6ProtocolNumber {
+ // ICMP packets may be longer, but until icmp.Parse is implemented, here
+ // we parse it using the minimum size.
transHeader, ok := pkt.Data.PullUp(transProto.MinimumPacketSize())
if !ok {
n.stack.stats.MalformedRcvdPackets.Increment()
return
}
pkt.TransportHeader = transHeader
+ pkt.Data.TrimFront(len(pkt.TransportHeader))
} else {
// This is either a bad packet or was re-assembled from fragments.
transProto.Parse(pkt)
diff --git a/pkg/tcpip/stack/nic_test.go b/pkg/tcpip/stack/nic_test.go
index 31f865260..3bc9fd831 100644
--- a/pkg/tcpip/stack/nic_test.go
+++ b/pkg/tcpip/stack/nic_test.go
@@ -84,6 +84,11 @@ func (e *testLinkEndpoint) WriteRawPacket(buffer.VectorisedView) *tcpip.Error {
return tcpip.ErrNotSupported
}
+// ARPHardwareType implements stack.LinkEndpoint.ARPHardwareType.
+func (*testLinkEndpoint) ARPHardwareType() header.ARPHardwareType {
+ panic("not implemented")
+}
+
var _ NetworkEndpoint = (*testIPv6Endpoint)(nil)
// An IPv6 NetworkEndpoint that throws away outgoing packets.
diff --git a/pkg/tcpip/stack/packet_buffer.go b/pkg/tcpip/stack/packet_buffer.go
index 1b5da6017..e3556d5d2 100644
--- a/pkg/tcpip/stack/packet_buffer.go
+++ b/pkg/tcpip/stack/packet_buffer.go
@@ -14,6 +14,7 @@
package stack
import (
+ "gvisor.dev/gvisor/pkg/sync"
"gvisor.dev/gvisor/pkg/tcpip"
"gvisor.dev/gvisor/pkg/tcpip/buffer"
)
@@ -24,7 +25,7 @@ import (
// multiple endpoints. Clone() should be called in such cases so that
// modifications to the Data field do not affect other copies.
type PacketBuffer struct {
- _ noCopy
+ _ sync.NoCopy
// PacketBufferEntry is used to build an intrusive list of
// PacketBuffers.
@@ -102,14 +103,3 @@ func (pk *PacketBuffer) Clone() *PacketBuffer {
NatDone: pk.NatDone,
}
}
-
-// noCopy may be embedded into structs which must not be copied
-// after the first use.
-//
-// See https://golang.org/issues/8005#issuecomment-190753527
-// for details.
-type noCopy struct{}
-
-// Lock is a no-op used by -copylocks checker from `go vet`.
-func (*noCopy) Lock() {}
-func (*noCopy) Unlock() {}
diff --git a/pkg/tcpip/stack/registration.go b/pkg/tcpip/stack/registration.go
index 5cbc946b6..f260eeb7f 100644
--- a/pkg/tcpip/stack/registration.go
+++ b/pkg/tcpip/stack/registration.go
@@ -18,6 +18,7 @@ import (
"gvisor.dev/gvisor/pkg/sleep"
"gvisor.dev/gvisor/pkg/tcpip"
"gvisor.dev/gvisor/pkg/tcpip/buffer"
+ "gvisor.dev/gvisor/pkg/tcpip/header"
"gvisor.dev/gvisor/pkg/waiter"
)
@@ -436,6 +437,12 @@ type LinkEndpoint interface {
// Wait will not block if the endpoint hasn't started any goroutines
// yet, even if it might later.
Wait()
+
+ // ARPHardwareType returns the ARPHRD_TYPE of the link endpoint.
+ //
+ // See:
+ // https://github.com/torvalds/linux/blob/aa0c9086b40c17a7ad94425b3b70dd1fdd7497bf/include/uapi/linux/if_arp.h#L30
+ ARPHardwareType() header.ARPHardwareType
}
// InjectableLinkEndpoint is a LinkEndpoint where inbound packets are
diff --git a/pkg/tcpip/stack/stack.go b/pkg/tcpip/stack/stack.go
index 51abe32a7..2b7ece851 100644
--- a/pkg/tcpip/stack/stack.go
+++ b/pkg/tcpip/stack/stack.go
@@ -425,6 +425,7 @@ type Stack struct {
handleLocal bool
// tables are the iptables packet filtering and manipulation rules.
+ // TODO(gvisor.dev/issue/170): S/R this field.
tables *IPTables
// resumableEndpoints is a list of endpoints that need to be resumed if the
@@ -471,6 +472,14 @@ type Stack struct {
// randomGenerator is an injectable pseudo random generator that can be
// used when a random number is required.
randomGenerator *mathrand.Rand
+
+ // sendBufferSize holds the min/default/max send buffer sizes for
+ // endpoints other than TCP.
+ sendBufferSize SendBufferSizeOption
+
+ // receiveBufferSize holds the min/default/max receive buffer sizes for
+ // endpoints other than TCP.
+ receiveBufferSize ReceiveBufferSizeOption
}
// UniqueID is an abstract generator of unique identifiers.
@@ -683,6 +692,16 @@ func New(opts Options) *Stack {
tempIIDSeed: opts.TempIIDSeed,
forwarder: newForwardQueue(),
randomGenerator: mathrand.New(randSrc),
+ sendBufferSize: SendBufferSizeOption{
+ Min: MinBufferSize,
+ Default: DefaultBufferSize,
+ Max: DefaultMaxBufferSize,
+ },
+ receiveBufferSize: ReceiveBufferSizeOption{
+ Min: MinBufferSize,
+ Default: DefaultBufferSize,
+ Max: DefaultMaxBufferSize,
+ },
}
// Add specified network protocols.
@@ -1076,6 +1095,11 @@ type NICInfo struct {
// Context is user-supplied data optionally supplied in CreateNICWithOptions.
// See type NICOptions for more details.
Context NICContext
+
+ // ARPHardwareType holds the ARP Hardware type of the NIC. This is the
+ // value sent in haType field of an ARP Request sent by this NIC and the
+ // value expected in the haType field of an ARP response.
+ ARPHardwareType header.ARPHardwareType
}
// HasNIC returns true if the NICID is defined in the stack.
@@ -1107,6 +1131,7 @@ func (s *Stack) NICInfo() map[tcpip.NICID]NICInfo {
MTU: nic.linkEP.MTU(),
Stats: nic.stats,
Context: nic.context,
+ ARPHardwareType: nic.linkEP.ARPHardwareType(),
}
}
return nics
@@ -1408,6 +1433,12 @@ func (s *Stack) RegisterTransportEndpoint(nicID tcpip.NICID, netProtos []tcpip.N
return s.demux.registerEndpoint(netProtos, protocol, id, ep, flags, bindToDevice)
}
+// CheckRegisterTransportEndpoint checks if an endpoint can be registered with
+// the stack transport dispatcher.
+func (s *Stack) CheckRegisterTransportEndpoint(nicID tcpip.NICID, netProtos []tcpip.NetworkProtocolNumber, protocol tcpip.TransportProtocolNumber, id TransportEndpointID, flags ports.Flags, bindToDevice tcpip.NICID) *tcpip.Error {
+ return s.demux.checkEndpoint(netProtos, protocol, id, flags, bindToDevice)
+}
+
// UnregisterTransportEndpoint removes the endpoint with the given id from the
// stack transport dispatcher.
func (s *Stack) UnregisterTransportEndpoint(nicID tcpip.NICID, netProtos []tcpip.NetworkProtocolNumber, protocol tcpip.TransportProtocolNumber, id TransportEndpointID, ep TransportEndpoint, flags ports.Flags, bindToDevice tcpip.NICID) {
diff --git a/pkg/tcpip/stack/stack_options.go b/pkg/tcpip/stack/stack_options.go
new file mode 100644
index 000000000..0b093e6c5
--- /dev/null
+++ b/pkg/tcpip/stack/stack_options.go
@@ -0,0 +1,106 @@
+// Copyright 2020 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 stack
+
+import "gvisor.dev/gvisor/pkg/tcpip"
+
+const (
+ // MinBufferSize is the smallest size of a receive or send buffer.
+ MinBufferSize = 4 << 10 // 4 KiB
+
+ // DefaultBufferSize is the default size of the send/recv buffer for a
+ // transport endpoint.
+ DefaultBufferSize = 212 << 10 // 212 KiB
+
+ // DefaultMaxBufferSize is the default maximum permitted size of a
+ // send/receive buffer.
+ DefaultMaxBufferSize = 4 << 20 // 4 MiB
+)
+
+// SendBufferSizeOption is used by stack.(Stack*).Option/SetOption to
+// get/set the default, min and max send buffer sizes.
+type SendBufferSizeOption struct {
+ Min int
+ Default int
+ Max int
+}
+
+// ReceiveBufferSizeOption is used by stack.(Stack*).Option/SetOption to
+// get/set the default, min and max receive buffer sizes.
+type ReceiveBufferSizeOption struct {
+ Min int
+ Default int
+ Max int
+}
+
+// SetOption allows setting stack wide options.
+func (s *Stack) SetOption(option interface{}) *tcpip.Error {
+ switch v := option.(type) {
+ case SendBufferSizeOption:
+ // Make sure we don't allow lowering the buffer below minimum
+ // required for stack to work.
+ if v.Min < MinBufferSize {
+ return tcpip.ErrInvalidOptionValue
+ }
+
+ if v.Default < v.Min || v.Default > v.Max {
+ return tcpip.ErrInvalidOptionValue
+ }
+
+ s.mu.Lock()
+ s.sendBufferSize = v
+ s.mu.Unlock()
+ return nil
+
+ case ReceiveBufferSizeOption:
+ // Make sure we don't allow lowering the buffer below minimum
+ // required for stack to work.
+ if v.Min < MinBufferSize {
+ return tcpip.ErrInvalidOptionValue
+ }
+
+ if v.Default < v.Min || v.Default > v.Max {
+ return tcpip.ErrInvalidOptionValue
+ }
+
+ s.mu.Lock()
+ s.receiveBufferSize = v
+ s.mu.Unlock()
+ return nil
+
+ default:
+ return tcpip.ErrUnknownProtocolOption
+ }
+}
+
+// Option allows retrieving stack wide options.
+func (s *Stack) Option(option interface{}) *tcpip.Error {
+ switch v := option.(type) {
+ case *SendBufferSizeOption:
+ s.mu.RLock()
+ *v = s.sendBufferSize
+ s.mu.RUnlock()
+ return nil
+
+ case *ReceiveBufferSizeOption:
+ s.mu.RLock()
+ *v = s.receiveBufferSize
+ s.mu.RUnlock()
+ return nil
+
+ default:
+ return tcpip.ErrUnknownProtocolOption
+ }
+}
diff --git a/pkg/tcpip/stack/stack_test.go b/pkg/tcpip/stack/stack_test.go
index 5aacbf53e..7657a4101 100644
--- a/pkg/tcpip/stack/stack_test.go
+++ b/pkg/tcpip/stack/stack_test.go
@@ -3338,3 +3338,83 @@ func TestDoDADWhenNICEnabled(t *testing.T) {
t.Fatalf("got stack.GetMainNICAddress(%d, %d) = (%s, nil), want = (%s, nil)", nicID, header.IPv6ProtocolNumber, got, addr.AddressWithPrefix)
}
}
+
+func TestStackReceiveBufferSizeOption(t *testing.T) {
+ const sMin = stack.MinBufferSize
+ testCases := []struct {
+ name string
+ rs stack.ReceiveBufferSizeOption
+ err *tcpip.Error
+ }{
+ // Invalid configurations.
+ {"min_below_zero", stack.ReceiveBufferSizeOption{Min: -1, Default: sMin, Max: sMin}, tcpip.ErrInvalidOptionValue},
+ {"min_zero", stack.ReceiveBufferSizeOption{Min: 0, Default: sMin, Max: sMin}, tcpip.ErrInvalidOptionValue},
+ {"default_below_min", stack.ReceiveBufferSizeOption{Min: sMin, Default: sMin - 1, Max: sMin - 1}, tcpip.ErrInvalidOptionValue},
+ {"default_above_max", stack.ReceiveBufferSizeOption{Min: sMin, Default: sMin + 1, Max: sMin}, tcpip.ErrInvalidOptionValue},
+ {"max_below_min", stack.ReceiveBufferSizeOption{Min: sMin, Default: sMin + 1, Max: sMin - 1}, tcpip.ErrInvalidOptionValue},
+
+ // Valid Configurations
+ {"in_ascending_order", stack.ReceiveBufferSizeOption{Min: sMin, Default: sMin + 1, Max: sMin + 2}, nil},
+ {"all_equal", stack.ReceiveBufferSizeOption{Min: sMin, Default: sMin, Max: sMin}, nil},
+ {"min_default_equal", stack.ReceiveBufferSizeOption{Min: sMin, Default: sMin, Max: sMin + 1}, nil},
+ {"default_max_equal", stack.ReceiveBufferSizeOption{Min: sMin, Default: sMin + 1, Max: sMin + 1}, nil},
+ }
+ for _, tc := range testCases {
+ t.Run(tc.name, func(t *testing.T) {
+ s := stack.New(stack.Options{})
+ defer s.Close()
+ if err := s.SetOption(tc.rs); err != tc.err {
+ t.Fatalf("s.SetOption(%#v) = %v, want: %v", tc.rs, err, tc.err)
+ }
+ var rs stack.ReceiveBufferSizeOption
+ if tc.err == nil {
+ if err := s.Option(&rs); err != nil {
+ t.Fatalf("s.Option(%#v) = %v, want: nil", rs, err)
+ }
+ if got, want := rs, tc.rs; got != want {
+ t.Fatalf("s.Option(..) returned unexpected value got: %#v, want: %#v", got, want)
+ }
+ }
+ })
+ }
+}
+
+func TestStackSendBufferSizeOption(t *testing.T) {
+ const sMin = stack.MinBufferSize
+ testCases := []struct {
+ name string
+ ss stack.SendBufferSizeOption
+ err *tcpip.Error
+ }{
+ // Invalid configurations.
+ {"min_below_zero", stack.SendBufferSizeOption{Min: -1, Default: sMin, Max: sMin}, tcpip.ErrInvalidOptionValue},
+ {"min_zero", stack.SendBufferSizeOption{Min: 0, Default: sMin, Max: sMin}, tcpip.ErrInvalidOptionValue},
+ {"default_below_min", stack.SendBufferSizeOption{Min: 0, Default: sMin - 1, Max: sMin - 1}, tcpip.ErrInvalidOptionValue},
+ {"default_above_max", stack.SendBufferSizeOption{Min: 0, Default: sMin + 1, Max: sMin}, tcpip.ErrInvalidOptionValue},
+ {"max_below_min", stack.SendBufferSizeOption{Min: sMin, Default: sMin + 1, Max: sMin - 1}, tcpip.ErrInvalidOptionValue},
+
+ // Valid Configurations
+ {"in_ascending_order", stack.SendBufferSizeOption{Min: sMin, Default: sMin + 1, Max: sMin + 2}, nil},
+ {"all_equal", stack.SendBufferSizeOption{Min: sMin, Default: sMin, Max: sMin}, nil},
+ {"min_default_equal", stack.SendBufferSizeOption{Min: sMin, Default: sMin, Max: sMin + 1}, nil},
+ {"default_max_equal", stack.SendBufferSizeOption{Min: sMin, Default: sMin + 1, Max: sMin + 1}, nil},
+ }
+ for _, tc := range testCases {
+ t.Run(tc.name, func(t *testing.T) {
+ s := stack.New(stack.Options{})
+ defer s.Close()
+ if err := s.SetOption(tc.ss); err != tc.err {
+ t.Fatalf("s.SetOption(%+v) = %v, want: %v", tc.ss, err, tc.err)
+ }
+ var ss stack.SendBufferSizeOption
+ if tc.err == nil {
+ if err := s.Option(&ss); err != nil {
+ t.Fatalf("s.Option(%+v) = %v, want: nil", ss, err)
+ }
+ if got, want := ss, tc.ss; got != want {
+ t.Fatalf("s.Option(..) returned unexpected value got: %#v, want: %#v", got, want)
+ }
+ }
+ })
+ }
+}
diff --git a/pkg/tcpip/stack/transport_demuxer.go b/pkg/tcpip/stack/transport_demuxer.go
index 118b449d5..b902c6ca9 100644
--- a/pkg/tcpip/stack/transport_demuxer.go
+++ b/pkg/tcpip/stack/transport_demuxer.go
@@ -221,6 +221,18 @@ func (epsByNIC *endpointsByNIC) registerEndpoint(d *transportDemuxer, netProto t
return multiPortEp.singleRegisterEndpoint(t, flags)
}
+func (epsByNIC *endpointsByNIC) checkEndpoint(d *transportDemuxer, netProto tcpip.NetworkProtocolNumber, transProto tcpip.TransportProtocolNumber, flags ports.Flags, bindToDevice tcpip.NICID) *tcpip.Error {
+ epsByNIC.mu.RLock()
+ defer epsByNIC.mu.RUnlock()
+
+ multiPortEp, ok := epsByNIC.endpoints[bindToDevice]
+ if !ok {
+ return nil
+ }
+
+ return multiPortEp.singleCheckEndpoint(flags)
+}
+
// unregisterEndpoint returns true if endpointsByNIC has to be unregistered.
func (epsByNIC *endpointsByNIC) unregisterEndpoint(bindToDevice tcpip.NICID, t TransportEndpoint, flags ports.Flags) bool {
epsByNIC.mu.Lock()
@@ -289,6 +301,17 @@ func (d *transportDemuxer) registerEndpoint(netProtos []tcpip.NetworkProtocolNum
return nil
}
+// checkEndpoint checks if an endpoint can be registered with the dispatcher.
+func (d *transportDemuxer) checkEndpoint(netProtos []tcpip.NetworkProtocolNumber, protocol tcpip.TransportProtocolNumber, id TransportEndpointID, flags ports.Flags, bindToDevice tcpip.NICID) *tcpip.Error {
+ for _, n := range netProtos {
+ if err := d.singleCheckEndpoint(n, protocol, id, flags, bindToDevice); err != nil {
+ return err
+ }
+ }
+
+ return nil
+}
+
// multiPortEndpoint is a container for TransportEndpoints which are bound to
// the same pair of address and port. endpointsArr always has at least one
// element.
@@ -380,7 +403,7 @@ func (ep *multiPortEndpoint) singleRegisterEndpoint(t TransportEndpoint, flags p
ep.mu.Lock()
defer ep.mu.Unlock()
- bits := flags.Bits()
+ bits := flags.Bits() & ports.MultiBindFlagMask
if len(ep.endpoints) != 0 {
// If it was previously bound, we need to check if we can bind again.
@@ -395,6 +418,22 @@ func (ep *multiPortEndpoint) singleRegisterEndpoint(t TransportEndpoint, flags p
return nil
}
+func (ep *multiPortEndpoint) singleCheckEndpoint(flags ports.Flags) *tcpip.Error {
+ ep.mu.RLock()
+ defer ep.mu.RUnlock()
+
+ bits := flags.Bits() & ports.MultiBindFlagMask
+
+ if len(ep.endpoints) != 0 {
+ // If it was previously bound, we need to check if we can bind again.
+ if ep.flags.TotalRefs() > 0 && bits&ep.flags.IntersectionRefs() == 0 {
+ return tcpip.ErrPortInUse
+ }
+ }
+
+ return nil
+}
+
// unregisterEndpoint returns true if multiPortEndpoint has to be unregistered.
func (ep *multiPortEndpoint) unregisterEndpoint(t TransportEndpoint, flags ports.Flags) bool {
ep.mu.Lock()
@@ -406,7 +445,7 @@ func (ep *multiPortEndpoint) unregisterEndpoint(t TransportEndpoint, flags ports
ep.endpoints[len(ep.endpoints)-1] = nil
ep.endpoints = ep.endpoints[:len(ep.endpoints)-1]
- ep.flags.DropRef(flags.Bits())
+ ep.flags.DropRef(flags.Bits() & ports.MultiBindFlagMask)
break
}
}
@@ -439,6 +478,28 @@ func (d *transportDemuxer) singleRegisterEndpoint(netProto tcpip.NetworkProtocol
return epsByNIC.registerEndpoint(d, netProto, protocol, ep, flags, bindToDevice)
}
+func (d *transportDemuxer) singleCheckEndpoint(netProto tcpip.NetworkProtocolNumber, protocol tcpip.TransportProtocolNumber, id TransportEndpointID, flags ports.Flags, bindToDevice tcpip.NICID) *tcpip.Error {
+ if id.RemotePort != 0 {
+ // SO_REUSEPORT only applies to bound/listening endpoints.
+ flags.LoadBalanced = false
+ }
+
+ eps, ok := d.protocol[protocolIDs{netProto, protocol}]
+ if !ok {
+ return tcpip.ErrUnknownProtocol
+ }
+
+ eps.mu.RLock()
+ defer eps.mu.RUnlock()
+
+ epsByNIC, ok := eps.endpoints[id]
+ if !ok {
+ return nil
+ }
+
+ return epsByNIC.checkEndpoint(d, netProto, protocol, flags, bindToDevice)
+}
+
// unregisterEndpoint unregisters the endpoint with the given id such that it
// won't receive any more packets.
func (d *transportDemuxer) unregisterEndpoint(netProtos []tcpip.NetworkProtocolNumber, protocol tcpip.TransportProtocolNumber, id TransportEndpointID, ep TransportEndpoint, flags ports.Flags, bindToDevice tcpip.NICID) {
generated by cgit v1.2.3 (git 2.39.1) at 2025-04-10 06:04:05 +0000