1 files changed, 235 insertions, 42 deletions

diff --git a/pkg/tcpip/stack/conntrack.go b/pkg/tcpip/stack/conntrack.go
index af9c325ca..d39baf620 100644
--- a/pkg/tcpip/stack/conntrack.go
+++ b/pkg/tcpip/stack/conntrack.go
@@ -15,9 +15,12 @@
 package stack
 
 import (
+	"encoding/binary"
 	"sync"
+	"time"
 
 	"gvisor.dev/gvisor/pkg/tcpip"
+	"gvisor.dev/gvisor/pkg/tcpip/hash/jenkins"
 	"gvisor.dev/gvisor/pkg/tcpip/header"
 	"gvisor.dev/gvisor/pkg/tcpip/transport/tcpconntrack"
 )
@@ -30,6 +33,10 @@ import (
 //
 // 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 direction int
 
@@ -48,7 +55,12 @@ const (
 
 // 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
+
 	tupleID
 
 	// conn is the connection tracking entry this tuple belongs to.
@@ -61,6 +73,8 @@ type tuple struct {
 // 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
@@ -83,6 +97,8 @@ func (ti tupleID) reply() tupleID {
 }
 
 // conn is a tracked connection.
+//
+// +stateify savable
 type conn struct {
 	// original is the tuple in original direction. It is immutable.
 	original tuple
@@ -98,22 +114,67 @@ type conn struct {
 	tcbHook Hook
 
 	// mu protects tcb.
-	mu sync.Mutex
+	mu sync.Mutex `state:"nosave"`
 
 	// tcb is TCB control block. It is used to keep track of states
 	// of tcp connection and is protected by mu.
 	tcb tcpconntrack.TCB
+
+	// 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)"`
+}
+
+// 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 {
-	// mu protects conns.
-	mu sync.RWMutex
+	// seed is a one-time random value initialized at stack startup
+	// and is used in the calculation of hash keys for the list of buckets.
+	// It is immutable.
+	seed uint32
 
-	// conns maintains a map of tuples needed for connection tracking for
-	// iptables NAT rules. It is protected by mu.
-	conns map[tupleID]tuple
+	// 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
+}
+
+// +stateify savable
+type bucket struct {
+	// mu protects tuples.
+	mu     sync.Mutex `state:"nosave"`
+	tuples tupleList
 }
 
 // packetToTupleID converts packet to a tuple ID. It fails when pkt lacks a valid
@@ -143,8 +204,9 @@ func packetToTupleID(pkt *PacketBuffer) (tupleID, *tcpip.Error) {
 // newConn creates new connection.
 func newConn(orig, reply tupleID, manip manipType, hook Hook) *conn {
 	conn := conn{
-		manip:   manip,
-		tcbHook: hook,
+		manip:    manip,
+		tcbHook:  hook,
+		lastUsed: time.Now(),
 	}
 	conn.original = tuple{conn: &conn, tupleID: orig}
 	conn.reply = tuple{conn: &conn, tupleID: reply, direction: dirReply}
@@ -162,14 +224,28 @@ func (ct *ConnTrack) connFor(pkt *PacketBuffer) (*conn, direction) {
 		return nil, dirOriginal
 	}
 
-	ct.mu.Lock()
-	defer ct.mu.Unlock()
-
-	tuple, ok := ct.conns[tid]
-	if !ok {
-		return nil, dirOriginal
+	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 tid == other.tupleID {
+			return other.conn, other.direction
+		}
 	}
-	return tuple.conn, tuple.direction
+
+	return nil, dirOriginal
 }
 
 // createConnFor creates a new conn for pkt.
@@ -197,13 +273,31 @@ func (ct *ConnTrack) createConnFor(pkt *PacketBuffer, hook Hook, rt RedirectTarg
 	}
 	conn := newConn(tid, replyTID, manip, hook)
 
-	// Add the changed tuple to the map.
-	// TODO(gvisor.dev/issue/170): Need to support collisions using linked
-	// list.
-	ct.mu.Lock()
-	defer ct.mu.Unlock()
-	ct.conns[tid] = conn.original
-	ct.conns[replyTID] = conn.reply
+	// 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()
+	}
+
+	// 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
 }
@@ -297,35 +391,134 @@ func (ct *ConnTrack) handlePacket(pkt *PacketBuffer, hook Hook, gso *GSO, r *Rou
 	// other tcp states.
 	conn.mu.Lock()
 	defer conn.mu.Unlock()
-	var st tcpconntrack.Result
-	tcpHeader := header.TCP(pkt.TransportHeader)
-	if conn.tcb.IsEmpty() {
+
+	// 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)
 	}
+}
+
+// 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)
+}
 
-	// Delete conn if tcp connection is closed.
-	if st == tcpconntrack.ResultClosedByPeer || st == tcpconntrack.ResultClosedBySelf || st == tcpconntrack.ResultReset {
-		ct.deleteConn(conn)
+// 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
 }
 
-// deleteConn deletes the connection.
-func (ct *ConnTrack) deleteConn(conn *conn) {
-	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
 	}
 
-	ct.mu.Lock()
-	defer ct.mu.Unlock()
+	// 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
+	}
+
+	// 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.conns, conn.original.tupleID)
-	delete(ct.conns, conn.reply.tupleID)
+	return true
 }