implement a new backoff abstraction (#75)

2026-07-26 10:38:47 +00:00 · 2026-04-14 16:11:03 -03:00
parent 191b8ee4cc
commit 5bde7a9979
5 changed files with 100 additions and 86 deletions
@@ -1,5 +1,7 @@
 package lneto

+import "time"
+
 // StackNode is an abstraction of a packet exchanging protocol controller. This is the building block for all protocols,
 // from Ethernet to IP to TCP, practically any protocol can be expressed as a StackNode and function completely.
 // Today protocols represented by StackNode also include NTP, DNS, DHCP, ARP, ICMP, UDP, mDNS.
@@ -37,12 +39,18 @@ type StackNode interface {
 	// SetFlagPending(flagPending func(numPendingEncapsulations int))
 }

-//go:generate stringer -type=IPProto,errGeneric -linecomment -output stringers.go .
+// BackoffStrategy is the abstraction of a backoff strategy for retrying an operation.
+// It returns the amount of time to sleep for. If returned value is 0 then runtime.Gosched is called.
+// A negative value results in no action taken.
+// consecutiveBackoffs starts at 1 and increments by 1 every time the operation is retried.
+type BackoffStrategy func(consecutiveBackoffs int) (sleep time.Duration)

 // IPProto represents the IP protocol number.
 type IPProto uint8

 // IP protocol numbers.
+//
+//go:generate stringer -type=IPProto,errGeneric -linecomment -output stringers.go .
 const (
 	IPProtoHopByHop        IPProto = 0   // IPv6 Hop-by-Hop Option [RFC8200]
 	IPProtoICMP            IPProto = 1   // ICMP [RFC792]
@@ -2,61 +2,16 @@ package internal

 import "time"

-type BackoffFlags uint8
-
+// ConnRWBackoff implements exponential backoff suitable for TCP connection
+// read/write polling. It starts at 1us and caps at 5ms, doubling on each consecutive backoff.
+func ConnRWBackoff(consecutiveBackoffs int) {
 	const (
-	BackoffHasPriority BackoffFlags = 1 << iota
-	BackoffCriticalPath
-	BackoffTCPConn
+		minWait = time.Microsecond >> 1
+		maxWait = 5 * time.Millisecond
 	)
-
-const backoffMinWait = time.Microsecond
-
-func backoffMaxWait(priority BackoffFlags) time.Duration {
-	switch {
-	case priority&BackoffCriticalPath != 0:
-		return 1 * time.Millisecond
-	case priority&BackoffTCPConn != 0:
-		return 5 * time.Millisecond
-	default:
-		return time.Second >> (priority & BackoffHasPriority)
-	}
-}
-
-func NewBackoff(priority BackoffFlags) Backoff {
-	return Backoff{
-		wait:      uint32(backoffMinWait),
-		maxWait:   uint32(backoffMaxWait(priority)),
-		startWait: uint32(backoffMinWait),
-	}
-}
-
-// A Backoff with a non-zero MaxWait is ready for use.
-type Backoff struct {
-	// wait defines the amount of time that Miss will wait on next call.
-	wait uint32
-	// Maximum allowable value for Wait.
-	maxWait uint32
-	// startWait is the intial Wait value, as well as the value that Wait takes after a call to Hit.
-	startWait uint32
-}
-
-// Hit sets eb.Wait to the StartWait value.
-func (eb *Backoff) Hit() {
-	if eb.maxWait == 0 {
-		panic("MaxWait cannot be zero")
-	}
-	eb.wait = eb.startWait
-}
-
-// Miss sleeps for eb.Wait and increases eb.Wait exponentially.
-func (eb *Backoff) Miss() {
-	if eb.maxWait == 0 {
-		panic("MaxWait cannot be zero")
-	}
-	time.Sleep(time.Duration(eb.wait))
-	eb.wait *= 2
-	if eb.wait > eb.maxWait {
-		eb.wait = eb.maxWait
+	wait := minWait << consecutiveBackoffs
+	if wait > maxWait {
+		wait = maxWait
 	}
+	time.Sleep(wait)
 }
@@ -29,6 +29,7 @@ type Conn struct {
 	h          Handler
 	remoteAddr []byte

+	_backoff lneto.BackoffStrategy
 	rdead    time.Time
 	wdead    time.Time
 	abortErr error
@@ -54,6 +55,9 @@ type ConnConfig struct {
 	RxBuf             []byte
 	TxBuf             []byte
 	TxPacketQueueSize int
+	// RWBackoff sets the backoff policy for backoff when data unavailable on Read or buffer full on Write.
+	// If not set a default backoff strategy will be used. See [internal.ConnRWBackoff].
+	RWBackoff lneto.BackoffStrategy
 	Logger    *slog.Logger
 }

@@ -64,6 +68,7 @@ func (conn *Conn) Configure(config ConnConfig) (err error) {
 	if err != nil {
 		return err
 	}
+	conn._backoff = config.RWBackoff
 	conn.logger.log = config.Logger
 	return nil
 }
@@ -206,8 +211,8 @@ func (conn *Conn) Write(b []byte) (int, error) {
 	} else if plen == 0 {
 		return 0, nil
 	}
-	backoff := internal.NewBackoff(internal.BackoffTCPConn)
 	n := 0
+	backoffs := 0
 	for {
 		if err := conn.checkPipe(connid, &conn.wdead); err != nil {
 			return 0, err
@@ -221,10 +226,11 @@ func (conn *Conn) Write(b []byte) (int, error) {
 		if (err != nil && err != internal.ErrRingBufferFull) || n == plen {
 			break
 		} else if ngot > 0 {
-			backoff.Hit()
+			backoffs = 0
 			runtime.Gosched() // Do a little yield since we won't have data for sure otherwise.
 		} else {
-			backoff.Miss()
+			backoffs++
+			conn.backoff(backoffs)
 		}
 		conn.trace("TCPConn.Write:insuf-buf", slog.Int("missing", plen-n), slog.Uint64("lport", uint64(lport)), slog.Uint64("rport", uint64(rport)))
 		if conn.deadlineExceeded(&conn.wdead) {
@@ -239,17 +245,13 @@ func (conn *Conn) Flush() error {
 	if err != nil {
 		return err
 	}
-	if conn.deadlineExceeded(&conn.wdead) {
-		return errDeadlineExceeded
-	} else if conn.BufferedUnsent() == 0 {
-		return nil
-	}
-	backoff := internal.NewBackoff(internal.BackoffTCPConn)
+	backoffs := 0
 	for conn.BufferedUnsent() != 0 {
 		if err := conn.checkPipe(connid, &conn.wdead); err != nil {
 			return err
 		}
-		backoff.Miss()
+		backoffs++
+		conn.backoff(backoffs)
 	}
 	return nil
 }
@@ -268,7 +270,7 @@ func (conn *Conn) Read(b []byte) (int, error) {
 	lport := conn.LocalPort()
 	rport := conn.RemotePort()
 	conn.trace("TCPConn.Read:start", slog.Uint64("lport", uint64(lport)), slog.Uint64("rport", uint64(rport)))
-	backoff := internal.NewBackoff(internal.BackoffTCPConn)
+	backoffs := 0
 	for conn.BufferedInput() == 0 {
 		state := conn.State()
 		if !state.RxDataOpen() {
@@ -280,7 +282,8 @@ func (conn *Conn) Read(b []byte) (int, error) {
 			}
 			return 0, err
 		}
-		backoff.Miss()
+		backoffs++
+		conn.backoff(backoffs)
 	}
 	return conn.handlerRead(b)
 }
@@ -446,3 +449,11 @@ func (conn *Conn) deadlineExceeded(deadline *time.Time) bool {
 func (conn *Conn) ConnectionID() *uint64 {
 	return conn.h.ConnectionID()
 }
+
+func (conn *Conn) backoff(consecutiveBackoffs int) {
+	if conn._backoff != nil {
+		conn._backoff(consecutiveBackoffs)
+	} else {
+		internal.ConnRWBackoff(consecutiveBackoffs)
+	}
+}
@@ -22,6 +22,7 @@ type Conn struct {

 	remoteAddr []byte

+	_backoff lneto.BackoffStrategy
 	rdead    time.Time
 	wdead    time.Time

@@ -38,6 +39,10 @@ type ConnConfig struct {
 	RxQueueSize int
 	// TxQueueSize is the maximum number of outgoing datagrams that can be queued.
 	TxQueueSize int
+	// RWBackoff sets the backoff policy for backoff when data unavailable on Read or buffer full on Write.
+	// This field is ineffective on configuring a [Handler].
+	// If not set a default backoff strategy will be used. See [internal.ConnRWBackoff].
+	RWBackoff lneto.BackoffStrategy
 }

 // Configure initializes the connection with the given buffer and queue configuration.
@@ -50,6 +55,7 @@ func (conn *Conn) Configure(cfg ConnConfig) error {
 	if err != nil {
 		return err
 	}
+	conn._backoff = cfg.RWBackoff
 	return nil
 }

@@ -120,13 +126,14 @@ func (conn *Conn) Write(b []byte) (int, error) {
 	if len(b) == 0 {
 		return 0, nil
 	}
-	backoff := internal.NewBackoff(internal.BackoffTCPConn)
-	for {
-		if conn.deadlineExceeded(&conn.wdead) {
-			return 0, os.ErrDeadlineExceeded
+	connID, err := conn.lockPipeConnID()
+	if err != nil {
+		return 0, err
 	}
+	backoffs := 0
+	for {
 		conn.mu.Lock()
-		if conn.h.closeCalled {
+		if conn.h.closeCalled || connID != conn.h.connid {
 			conn.mu.Unlock()
 			return 0, net.ErrClosed
 		}
@@ -135,20 +142,25 @@ func (conn *Conn) Write(b []byte) (int, error) {
 		if n > 0 {
 			return n, err
 		}
-		backoff.Miss()
+		if conn.deadlineExceeded(&conn.wdead) {
+			return 0, os.ErrDeadlineExceeded
+		}
+		backoffs++
+		conn.backoff(backoffs)
 	}
 }

 // Read dequeues a single datagram. If the buffer is smaller than the datagram,
 // the remaining bytes are discarded (SOCK_DGRAM semantics).
 func (conn *Conn) Read(b []byte) (int, error) {
-	backoff := internal.NewBackoff(internal.BackoffTCPConn)
-	for {
-		if conn.deadlineExceeded(&conn.rdead) {
-			return 0, os.ErrDeadlineExceeded
+	connID, err := conn.lockPipeConnID()
+	if err != nil {
+		return 0, err
 	}
+	backoffs := 0
+	for {
 		conn.mu.Lock()
-		if conn.h.closeCalled && conn.h.BufferedInput() == 0 {
+		if conn.h.closeCalled && conn.h.BufferedInput() == 0 || connID != conn.h.connid {
 			conn.mu.Unlock()
 			return 0, net.ErrClosed
 		}
@@ -157,7 +169,11 @@ func (conn *Conn) Read(b []byte) (int, error) {
 		if n > 0 {
 			return n, err
 		}
-		backoff.Miss()
+		if conn.deadlineExceeded(&conn.rdead) {
+			return 0, os.ErrDeadlineExceeded
+		}
+		backoffs++
+		conn.backoff(backoffs)
 	}
 }

@@ -284,3 +300,20 @@ func (conn *Conn) FreeInput() int {
 	defer conn.mu.Unlock()
 	return conn.h.FreeInput()
 }
+
+func (conn *Conn) backoff(consecutiveBackoffs int) {
+	if conn._backoff != nil {
+		conn._backoff(consecutiveBackoffs)
+	} else {
+		internal.ConnRWBackoff(consecutiveBackoffs)
+	}
+}
+
+func (conn *Conn) lockPipeConnID() (uint64, error) {
+	conn.mu.Lock()
+	defer conn.mu.Unlock()
+	if conn.h.closeCalled && len(conn.h.rxDgrams) == 0 {
+		return 0, net.ErrClosed
+	}
+	return conn.h.connid, nil
+}
@@ -53,6 +53,9 @@ type TCPPoolConfig struct {
 	ClosingTimeout time.Duration
 	// NewUserData is used to create user data used for each individual TCP connection and returned on GetTCP.
 	NewUserData func() any
+	// NewBackoff returns the backoff to use for every newly configured TCP connection.
+	// This should always return a static(non-method) function unless you know what you are doing.
+	NewBackoff func() lneto.BackoffStrategy
 }

 func NewTCPPool(cfg TCPPoolConfig) (*TCPPool, error) {
@@ -76,12 +79,16 @@ func NewTCPPool(cfg TCPPoolConfig) (*TCPPool, error) {
 	for i := range pool.conns {
 		bufoff := i * allocPerConn
 		txOff := bufoff + cfg.RxBufSize
-		err := pool.conns[i].Configure(tcp.ConnConfig{
+		conncfg := tcp.ConnConfig{
 			RxBuf:             bufSpace[bufoff:txOff],
 			TxBuf:             bufSpace[txOff : txOff+cfg.TxBufSize],
 			TxPacketQueueSize: cfg.QueueSize,
 			Logger:            cfg.ConnLogger,
-		})
+		}
+		if cfg.NewBackoff != nil {
+			conncfg.RWBackoff = cfg.NewBackoff()
+		}
+		err := pool.conns[i].Configure(conncfg)
 		if err != nil {
 			return nil, err
 		}