XTLS_Xray-core/proxy/tun/tcp_lightweight.go

package tun

import (
	"encoding/binary"
	"io"
	"math"
	"math/rand/v2"
	"slices"
	"sync"
	"time"

	"github.com/xtls/xray-core/common/buf"
	"github.com/xtls/xray-core/common/errors"
	"github.com/xtls/xray-core/common/net"
)

const (
	tcpFin = 1 << iota
	tcpSyn
	tcpRst
	tcpPsh
	tcpAck
)

const (
	// tcpWindow is what a connection takes from the system before its reader has to take it,
	// and the most that it sends to the system before that is acknowledged. More than that, for several
	// connections at once, is more packets than the queue of a tun holds (500 on Linux)
	tcpWindow = 256 << 10
	// tcpLargeWindow is the same for a tun that takes and gives segments that are larger than its MTU: those are
	// few, and a way with a longer round trip, as a router has to what is behind it, is not held back by it
	tcpLargeWindow = 1 << 20
	// tcpWindowShift makes them fit the 16 bits of a window
	tcpWindowShift = 5
	// tcpHeld is how many segments wait for one that was lost before them,
	// and of how many parts what the system says it has is known
	tcpHeld = 1024
	// tcpSmallest is the smallest mss that the system is believed, the one of Linux
	tcpSmallest = 48
	// tcpOpening is how many connections can be waiting for the answer to their SYN-ACK
	tcpOpening = 65535
	// tcpStart is how many segments are sent before any of them is acknowledged
	tcpStart = 10
	// tcpSegment is the largest payload of a segment that the tun splits by itself,
	// tcpParts of how many parts of what was written it is made at most
	tcpSegment = math.MaxUint16 - ipv6HeaderSize - tcpHeaderSize
	tcpParts   = 7

	// tcpTick is how often a connection that waits for something looks at what it waits for
	tcpTick = 100 * time.Millisecond
	// ticks without anything acknowledged before it is sent again, twice as many every next time, a minute at most
	tcpRetransmitTicks     = 3
	tcpRetransmitTicksMost = 600
	// how many times it is sent again before the connection is given up: that is after four minutes, as what is
	// behind a wireless link may be away for a while, and after twenty seconds for one that never answered
	tcpRetransmits    = 10
	tcpSynRetransmits = 5
	// ticks between the probes of a window that stays closed
	tcpProbeTicks = 10
	// tcpLinger is how long a closed connection waits for the system to close its side
	tcpLinger = 30 * time.Second
)

var (
	errTCPReset   = errors.New("connection reset")
	errTCPTimeout = errors.New("connection timed out")
)

// segment is a TCP segment that came from the tun
type segment struct {
	seq, ack uint32
	flags    byte
	window   uint16
	options  []byte
	payload  []byte
}

// nextOption returns the first option of o and what follows it, kind is zero when there is none
func nextOption(o []byte) (kind byte, value, rest []byte) {
	for len(o) > 0 && o[0] == 1 {
		o = o[1:]
	}
	if len(o) < 2 || int(o[1]) < 2 || int(o[1]) > len(o) {
		return 0, nil, nil
	}
	return o[0], o[2:o[1]], o[o[1]:]
}

// held is a segment that came before the one in front of it
type held struct {
	seq  uint32
	data *buf.Buffer
}

// tcpConn is a TCP connection that the system opened through the tun.
// Mostly the two are on the same machine, with nothing but the queue of the tun in between, which drops what
// does not fit. But the system may also pass on what others send, as a router does, and then the link to them
// is in between. So a connection has windows that are small enough, keeps what came after a lost segment,
// tells the system what it has (SACK), sends again what the system misses, and sends no more at once
// than what has arrived before (slow start and congestion avoidance as in RFC 5681)
type tcpConn struct {
	stack    *stack
	flow     flow
	src, dst net.Destination
	// ip is the IP header of what is sent, without the length
	ip [ipv6HeaderSize]byte
	// ipSize is the size of that header
	ipSize int
	// mss is the largest payload of a segment that is sent
	mss int
	// window is tcpWindow or tcpLargeWindow
	window int32
	// sndShift and rcvShift are the window scale of what the system and we advertise
	sndShift, rcvShift uint8
	// sack is set if the system takes selective acknowledgments
	sack bool

	mu       sync.Mutex
	readable sync.Cond
	writable sync.Cond

	established bool
	// closed is set when Close is called
	closed bool
	// err is why the connection is gone
	err error

	rcvNxt uint32
	// rcvAcked is the last acknowledgment number sent
	rcvAcked uint32
	// rcvAdvertised is the last window sent, in bytes
	rcvAdvertised int32
	rcvQueue      buf.MultiBuffer
	rcvLen        int32
	rcvFin        bool
	// rcvEnd is where what the system sends ends, if rcvEnds: its FIN came with a segment that waits in rcvHeld
	rcvEnd  uint32
	rcvEnds bool
	// rcvHeld is what came after something that did not, in the order of seq and without overlaps
	rcvHeld []held
	// rcvLast is the seq of the last one that was put into rcvHeld
	rcvLast uint32

	sndUna, sndNxt uint32
	sndWnd         uint32
	// sndQueue is what was sent and is not acknowledged
	sndQueue buf.MultiBuffer
	// sndBase is the sequence number of the first byte of sndQueue
	sndBase uint32
	sndFin  bool
	// writing is how many of the last of sndQueue WriteMultiBuffer is sending, it is the only one to release them
	writing int
	// blocked is set while WriteMultiBuffer waits for the window
	blocked bool
	// sending is set while WriteMultiBuffer writes a segment without mu, ackWanted if sendAck was called meanwhile
	sending, ackWanted bool
	dupAcks            int
	// cwnd is how much the way to the system takes at once, as far as that is known: it grows while what is sent
	// arrives, by as much as arrives below ssthresh, and is cut when something is lost
	cwnd, ssthresh uint32
	// sacked is what the system says it has past sndUna, in order and apart from each other
	sacked [][2]uint32
	// recovering is set from a sign of loss until what was sent by then, up to recover, is acknowledged
	recovering bool
	recover    uint32
	// lost is the end of what is lost past the last of sacked, resent the end of what was sent again since then
	lost, resent uint32
	// timed is the end of a segment that was sent at timedAt and not again, to see how long its acknowledgment takes.
	// The first one is the SYN-ACK
	timed   uint32
	timedAt time.Time
	// srtt is how long an acknowledgment takes, rttvar how much that varies (RFC 6298)
	srtt, rttvar time.Duration

	timer *time.Timer
	armed bool
	// stalled counts ticks without an acknowledgment of what waits for one
	stalled, retransmitTicks, retransmits int
	lingered                              bool
}

func newTCPConn(s *stack, f flow, syn segment) *tcpConn {
	c := &tcpConn{
		stack:           s,
		flow:            f,
		src:             net.TCPDestination(net.IPAddress(f.src.Addr().AsSlice()), net.Port(f.src.Port())),
		dst:             net.TCPDestination(net.IPAddress(f.dst.Addr().AsSlice()), net.Port(f.dst.Port())),
		rcvNxt:          syn.seq + 1,
		sndWnd:          uint32(syn.window),
		window:          s.window,
		ssthresh:        uint32(s.window),
		timedAt:         time.Now(),
		retransmitTicks: tcpRetransmitTicks,
	}
	c.readable.L = &c.mu
	c.writable.L = &c.mu
	c.ipSize = putIPHeader(c.ip[:], protoTCP, f.dst.Addr(), f.src.Addr())
	c.mss = s.mtu - c.ipSize - tcpHeaderSize
	for kind, value, o := nextOption(syn.options); kind != 0; kind, value, o = nextOption(o) {
		switch {
		case kind == 2 && len(value) == 2:
			c.mss = min(c.mss, int(binary.BigEndian.Uint16(value)))
		case kind == 3 && len(value) == 1:
			c.sndShift = min(value[0], 14)
			c.rcvShift = tcpWindowShift
		case kind == 4:
			c.sack = true
		}
	}
	c.mss = max(c.mss, tcpSmallest)
	c.cwnd = tcpStart * uint32(c.mss)
	c.sndUna = rand.Uint32()
	c.sndNxt = c.sndUna + 1
	return c
}

// header writes the headers of a segment with size bytes of payload into b and returns their size.
// The checksum of the segment is left to write. c.mu is held
func (c *tcpConn) header(b []byte, seq uint32, flags byte, options []byte, size int) int {
	n := copy(b, c.ip[:c.ipSize])
	t := b[n : n+tcpHeaderSize+len(options)]
	binary.BigEndian.PutUint16(t, c.flow.dst.Port())
	binary.BigEndian.PutUint16(t[2:], c.flow.src.Port())
	binary.BigEndian.PutUint32(t[4:], seq)
	binary.BigEndian.PutUint32(t[8:], c.rcvNxt)
	t[12] = byte(len(t)/4) << 4
	t[13] = flags
	window := max(c.window-c.rcvLen, 0)
	if flags&tcpSyn != 0 {
		// the window of a SYN is not scaled
		binary.BigEndian.PutUint16(t[14:], uint16(min(window, math.MaxUint16)))
	} else {
		binary.BigEndian.PutUint16(t[14:], uint16(min(window>>c.rcvShift, math.MaxUint16)))
	}
	t[16], t[17], t[18], t[19] = 0, 0, 0, 0
	copy(t[tcpHeaderSize:], options)
	setIPLength(b[:n], n+len(t)+size)
	c.rcvAcked = c.rcvNxt
	c.rcvAdvertised = window
	return n + len(t)
}

// write completes the headers h with the checksum of the segment and writes it to the tun.
// A segment that is not written is one that is lost: it is sent again when it is not acknowledged
func (c *tcpConn) write(h, payload []byte) {
	t := h[c.ipSize:]
	binary.BigEndian.PutUint16(t[16:], transportChecksum(h[:c.ipSize], protoTCP, t, payload))
	_ = c.stack.tun.WritePacket(h, payload)
}

// send writes a segment to the tun. c.mu is held
func (c *tcpConn) send(seq uint32, flags byte, options, payload []byte) {
	var b [ipv6HeaderSize + tcpHeaderSize + 40]byte
	c.write(b[:c.header(b[:], seq, flags, options, len(payload))], payload)
}

func (c *tcpConn) sendSynAck() {
	options := [12]byte{2, 4}
	binary.BigEndian.PutUint16(options[2:], uint16(c.stack.mtu-c.ipSize-tcpHeaderSize))
	n := 4
	if c.sack {
		n += copy(options[n:], []byte{1, 1, 4, 2})
	}
	if c.rcvShift != 0 {
		n += copy(options[n:], []byte{1, 3, 3, c.rcvShift})
	}
	c.send(c.sndUna, tcpSyn|tcpAck, options[:n], nil)
}

// sendAck acknowledges what was received, and tells what of rcvHeld waits for a lost segment
func (c *tcpConn) sendAck() {
	if c.sending {
		// after the segment that is being written: it was made before, and what the system is told goes in that order
		c.ackWanted = true
		return
	}
	if !c.sack || len(c.rcvHeld) == 0 {
		c.send(c.sndNxt, tcpAck, nil, nil)
		return
	}
	// the block with the segment that came last goes first
	var blocks [4][2]uint32
	n, last := 1, false
	for i := 0; i < len(c.rcvHeld); {
		start, end, recent := c.rcvHeld[i].seq, c.rcvHeld[i].seq, false
		for ; i < len(c.rcvHeld) && c.rcvHeld[i].seq == end; i++ {
			recent = recent || end == c.rcvLast
			end += uint32(c.rcvHeld[i].data.Len())
		}
		if recent {
			blocks[0], last = [2]uint32{start, end}, true
		} else if n < len(blocks) {
			blocks[n] = [2]uint32{start, end}
			n++
		}
	}
	list := blocks[:n]
	if !last {
		list = list[1:]
	}
	options := [36]byte{1, 1, 5, byte(2 + 8*len(list))}
	for i, block := range list {
		binary.BigEndian.PutUint32(options[4+8*i:], block[0])
		binary.BigEndian.PutUint32(options[8+8*i:], block[1])
	}
	c.send(c.sndNxt, tcpAck, options[:4+8*len(list)], nil)
}

// input takes a segment of the connection. It returns false if the connection is gone
// and the segment is the start of another one
func (c *tcpConn) input(s segment) bool {
	c.mu.Lock()
	defer c.mu.Unlock()
	if c.err != nil {
		return s.flags&(tcpSyn|tcpAck|tcpRst) != tcpSyn
	}
	if s.flags&tcpRst != 0 {
		// from where the system is, or was before its FIN, not from someone who knows no more than the addresses
		if s.seq-c.rcvNxt <= uint32(c.window) || c.rcvFin && s.seq+1 == c.rcvNxt {
			c.abort(errTCPReset)
		}
		return true
	}
	if s.flags&tcpSyn != 0 {
		switch {
		case !c.established && s.seq+1 == c.rcvNxt:
			c.sendSynAck()
			c.timedAt = time.Time{}
		case c.closed && s.flags&tcpAck == 0:
			c.abort(errTCPReset)
			return false
		default:
			c.sendAck()
		}
		return true
	}
	if s.flags&tcpAck == 0 {
		return true
	}
	if !c.established {
		if s.ack != c.sndNxt {
			return true
		}
		c.established = true
		c.sndUna = s.ack
		c.sndBase = s.ack
		c.sndWnd = uint32(s.window) << c.sndShift
		c.stalled, c.retransmits = 0, 0
		if !c.timedAt.IsZero() {
			c.clocked()
		}
		c.stack.opening.Add(-1)
		go c.stack.handle(c, c.dst)
	} else {
		c.acknowledge(s)
	}
	c.receive(s)
	if c.sndFin && c.rcvFin && c.sndUna == c.sndNxt {
		// both sides have closed it and nothing waits to be acknowledged
		c.abort(io.ErrClosedPipe)
	}
	return true
}

// acknowledge takes what a segment says about what was sent. c.mu is held
func (c *tcpConn) acknowledge(s segment) {
	acked, sent := s.ack-c.sndUna, c.sndNxt-c.sndUna
	if acked > sent {
		// older than the last one, or of what was not sent
		return
	}
	window := uint32(s.window) << c.sndShift
	if acked != 0 {
		c.sndUna = s.ack
		c.dupAcks, c.stalled, c.retransmits = 0, 0, 0
		c.release()
		n := 0
		for n < len(c.sacked) && int32(c.sacked[n][0]-c.sndUna) <= 0 {
			n++
		}
		c.sacked = slices.Delete(c.sacked, 0, n)
		if int32(c.resent-c.sndUna) < 0 {
			c.resent = c.sndUna
		}
		if !c.timedAt.IsZero() && int32(c.sndUna-c.timed) >= 0 {
			c.clocked()
		}
		switch {
		case c.recovering && int32(c.sndUna-c.recover) >= 0:
			c.recovering = false
		case c.recovering:
			// what was sent again has arrived, and what was sent before it has not: the next segment is lost as well
			if end := c.sndUna + min(uint32(c.mss), c.recover-c.sndUna); int32(end-c.lost) > 0 {
				c.lost = end
			}
		}
		// the way takes more if what it was given has arrived, as long as it was given about as much as it takes
		if sent >= c.cwnd/2 {
			if c.cwnd < c.ssthresh {
				c.cwnd = min(c.cwnd+acked, c.ssthresh)
			} else if !c.recovering {
				// a segment more for every cwnd that arrives
				c.cwnd = min(c.cwnd+max(uint32(uint64(c.mss)*uint64(acked)/uint64(c.cwnd)), 1), uint32(c.window))
			}
		}
	} else if sent != 0 && window == c.sndWnd && len(s.payload) == 0 && s.flags&tcpFin == 0 {
		c.dupAcks++
	}
	if c.sack {
		for kind, value, o := nextOption(s.options); kind != 0; kind, value, o = nextOption(o) {
			for ; kind == 5 && len(value) >= 8; value = value[8:] {
				// of what was sent and is not acknowledged
				start, end := binary.BigEndian.Uint32(value)-c.sndUna, binary.BigEndian.Uint32(value[4:])-c.sndUna
				if start != 0 && start < end && end <= c.sndNxt-c.sndUna {
					c.sackAdd(c.sndUna+start, c.sndUna+end)
				}
			}
		}
	}
	if !c.recovering && c.sndUna != c.sndNxt && (len(c.sacked) > 0 || c.dupAcks >= 3) {
		// the system has something past sndUna, or says the same for the third time: the segment at sndUna is lost
		c.congested()
		c.lost = c.sndUna + min(uint32(c.mss), c.sndNxt-c.sndUna)
	}
	if c.recovering {
		c.resend()
	}
	c.sndWnd = window
	// the writer goes on when it can write a good part of the window, not for every segment that is acknowledged.
	// Unless the way to the system is what holds it back: that takes what the acknowledgments say has left it,
	// and what is sent as they come does not come to it all at once
	limit := min(c.sndWnd, uint32(c.window))
	enough := limit / 2
	if c.cwnd < limit {
		enough = uint32(c.mss)
	}
	if c.blocked && (c.sndUna == c.sndNxt || c.space() >= enough) {
		c.writable.Broadcast()
	}
}

// clocked takes the time since timedAt as how long an acknowledgment takes. The next ones are waited for as long
// as that says and no less than a few ticks. Until there is such a time again, after something was sent again,
// they are waited for as long as the last time. c.mu is held
func (c *tcpConn) clocked() {
	d := time.Since(c.timedAt)
	c.timedAt = time.Time{}
	if c.srtt == 0 {
		c.srtt, c.rttvar = d, d/2
	} else {
		c.rttvar += ((c.srtt - d).Abs() - c.rttvar) / 4
		c.srtt += (d - c.srtt) / 8
	}
	c.retransmitTicks = max(tcpRetransmitTicks, int((c.srtt+4*c.rttvar)/tcpTick)+2)
}

// sackAdd puts what the system says it has into sacked. c.mu is held
func (c *tcpConn) sackAdd(start, end uint32) {
	i := 0
	for i < len(c.sacked) && int32(c.sacked[i][1]-start) < 0 {
		i++
	}
	j := i
	for ; j < len(c.sacked) && int32(c.sacked[j][0]-end) <= 0; j++ {
		if int32(c.sacked[j][0]-start) < 0 {
			start = c.sacked[j][0]
		}
		if int32(c.sacked[j][1]-end) > 0 {
			end = c.sacked[j][1]
		}
	}
	if i == j && len(c.sacked) == tcpHeld {
		return
	}
	c.sacked = slices.Replace(c.sacked, i, j, [2]uint32{start, end})
}

// congested is called when something is lost: the way to the system takes half of what it was given,
// and what is lost is sent again from sndUna. c.mu is held
func (c *tcpConn) congested() {
	c.ssthresh = max(min(c.cwnd, c.sndNxt-c.sndUna)/2, 2*uint32(c.mss))
	c.cwnd = c.ssthresh
	c.recovering, c.recover, c.resent = true, c.sndNxt, c.sndUna
	c.timedAt = time.Time{}
}

// hole returns the first part at or after seq of what is lost: what the system does not have
// before something that it says it has, or before lost. There is none if it ends where it starts. c.mu is held
func (c *tcpConn) hole(seq uint32) (uint32, uint32) {
	end := c.lost
	for _, b := range c.sacked {
		if int32(b[0]-seq) > 0 {
			end = b[0]
			break
		}
		if int32(b[1]-seq) > 0 {
			seq = b[1]
		}
	}
	if int32(end-seq) < 0 {
		end = seq
	}
	return seq, end
}

// flight is how much of what was sent is on its way: not what the system says it has,
// and not what is lost and was not sent again. c.mu is held
func (c *tcpConn) flight() uint32 {
	n := c.sndNxt - c.sndUna
	if !c.recovering {
		return n
	}
	seq := c.resent
	for _, b := range c.sacked {
		n -= b[1] - b[0]
		if int32(b[0]-seq) > 0 {
			n -= b[0] - seq
		}
		if int32(b[1]-seq) > 0 {
			seq = b[1]
		}
	}
	if int32(c.lost-seq) > 0 {
		n -= c.lost - seq
	}
	return n
}

// resend sends again what is lost and was not yet: the segment that the system waits for at once,
// the rest as the way to it takes it. c.mu is held
func (c *tcpConn) resend() {
	for {
		start, end := c.hole(c.resent)
		if start == end {
			return
		}
		room := uint32(c.mss)
		if start != c.sndUna {
			flight := c.flight()
			if flight+room > c.cwnd {
				return
			}
			room = c.cwnd - flight
		}
		end = start + min(end-start, room)
		c.retransmit(start, end)
		c.resent = end
	}
}

// space is how much can be sent now: what the system takes, and what the way to it takes. c.mu is held
func (c *tcpConn) space() uint32 {
	window, sent, flight := min(c.sndWnd, uint32(c.window)), c.sndNxt-c.sndUna, c.flight()
	if sent >= window || flight >= c.cwnd {
		return 0
	}
	return min(window-sent, c.cwnd-flight)
}

// receive takes the payload of a segment. c.mu is held
func (c *tcpConn) receive(s segment) {
	seq, payload, fin := s.seq, s.payload, s.flags&tcpFin != 0
	if len(payload) == 0 && !fin {
		if int32(c.rcvNxt-seq) > 0 {
			// a keepalive, or a probe of a window that the system has as closed: answered with what is expected
			c.sendAck()
		}
		return
	}
	if before := int32(c.rcvNxt - seq); before > 0 {
		if int(before) > len(payload) || int(before) == len(payload) && !fin {
			// sent again, or a keepalive: answered with what is expected
			c.sendAck()
			return
		}
		seq, payload = c.rcvNxt, payload[before:]
	}
	if c.rcvFin {
		c.sendAck()
		return
	}
	if c.closed && len(payload) > 0 {
		// nobody reads it anymore
		c.reset()
		return
	}
	// the system is told at once when something is wrong, or it sends more of the same
	now := len(c.rcvHeld) > 0
	if room := int(c.window - c.rcvLen - int32(seq-c.rcvNxt)); len(payload) > room {
		payload, fin, now = payload[:max(room, 0)], false, true
	}
	if seq != c.rcvNxt {
		// something before it was lost, it waits for that, and so does the end if this is it
		c.hold(seq, payload)
		if fin {
			c.rcvEnd, c.rcvEnds = seq+uint32(len(payload)), true
		}
		c.sendAck()
		return
	}

	c.rcvQueue = buf.MergeBytes(c.rcvQueue, payload)
	c.rcvLen += int32(len(payload))
	c.rcvNxt += uint32(len(payload))
	for !fin && len(c.rcvHeld) > 0 && int32(c.rcvNxt-c.rcvHeld[0].seq) >= 0 {
		h := c.rcvHeld[0]
		if before := int32(c.rcvNxt - h.seq); before < h.data.Len() {
			c.rcvQueue = buf.MergeBytes(c.rcvQueue, h.data.BytesFrom(before))
			c.rcvLen += h.data.Len() - before
			c.rcvNxt += uint32(h.data.Len() - before)
		}
		h.data.Release()
		c.rcvHeld = c.rcvHeld[1:]
	}
	if fin = fin || c.rcvEnds && c.rcvNxt == c.rcvEnd; fin {
		c.rcvFin = true
		c.rcvNxt++
	}
	c.readable.Signal()
	// the reader acknowledges what it takes, with one segment for all that came by then.
	// It is not waited for when it does not read for long
	if fin || now || c.rcvNxt-c.rcvAcked >= tcpWindow/4 {
		c.sendAck()
	} else {
		c.arm(tcpTick)
	}
}

// release releases what of sndQueue is acknowledged. c.mu is held
func (c *tcpConn) release() {
	for len(c.sndQueue) > c.writing {
		end := c.sndBase + uint32(c.sndQueue[0].Len())
		if int32(c.sndUna-end) < 0 {
			break
		}
		c.sndQueue[0].Release()
		c.sndQueue[0] = nil
		c.sndQueue = c.sndQueue[1:]
		c.sndBase = end
	}
}

// hold keeps a segment that is not the next one. c.mu is held
func (c *tcpConn) hold(seq uint32, payload []byte) {
	i := 0
	for i < len(c.rcvHeld) && int32(c.rcvHeld[i].seq-seq) < 0 {
		i++
	}
	// without what the one before and the one after have
	if i > 0 {
		if over := int32(c.rcvHeld[i-1].seq + uint32(c.rcvHeld[i-1].data.Len()) - seq); over > 0 {
			if int(over) >= len(payload) {
				return
			}
			seq, payload = seq+uint32(over), payload[over:]
		}
	}
	if i < len(c.rcvHeld) {
		if over := int32(seq + uint32(len(payload)) - c.rcvHeld[i].seq); over > 0 {
			if int(over) >= len(payload) {
				return
			}
			payload = payload[:len(payload)-int(over)]
		}
	}
	if len(payload) == 0 || len(c.rcvHeld) == tcpHeld {
		return
	}
	b := buf.NewWithSize(int32(len(payload)))
	b.Write(payload)
	c.rcvHeld = slices.Insert(c.rcvHeld, i, held{seq, b})
	c.rcvLast = seq
}

// retransmit sends again what was sent from seq to end. c.mu is held
func (c *tcpConn) retransmit(seq, end uint32) {
	fin := false
	if int32(end-c.sndNxt) >= 0 {
		end, fin = c.sndNxt, c.sndFin
		if fin {
			end--
		}
	}
	for int32(end-seq) > 0 {
		var part [1][]byte
		_, n := c.queued(part[:0], seq, min(c.mss, int(end-seq)))
		if n == 0 {
			break
		}
		c.send(seq, tcpAck, nil, part[0])
		seq += uint32(n)
	}
	if fin && seq == end {
		c.send(end, tcpFin|tcpAck, nil, nil)
	}
}

// queued adds to parts what of sndQueue starts at seq, size bytes at most and as many parts as it has room for,
// and returns them with their size. c.mu is held
func (c *tcpConn) queued(parts [][]byte, seq uint32, size int) ([][]byte, int) {
	skip, n := int32(seq-c.sndBase), 0
	for _, b := range c.sndQueue {
		if skip >= b.Len() {
			skip -= b.Len()
			continue
		}
		if n == size || len(parts) == cap(parts) {
			break
		}
		p := b.BytesFrom(skip)
		skip = 0
		p = p[:min(len(p), size-n)]
		parts = append(parts, p)
		n += len(p)
	}
	return parts, n
}

// arm makes tick run after d, unless it is going to run already. c.mu is held
func (c *tcpConn) arm(d time.Duration) {
	if c.armed {
		return
	}
	c.armed = true
	if c.timer == nil {
		c.timer = time.AfterFunc(d, c.tick)
	} else {
		c.timer.Reset(d)
	}
}

// tick does what was put off and what has to be done again
func (c *tcpConn) tick() {
	c.mu.Lock()
	defer c.mu.Unlock()
	c.armed = false
	if c.err != nil {
		return
	}
	if c.rcvAcked != c.rcvNxt {
		c.sendAck()
	}
	switch {
	case c.sndUna != c.sndNxt:
		if c.stalled++; c.stalled >= c.retransmitTicks {
			if c.retransmits++; c.retransmits > tcpRetransmits || !c.established && c.retransmits > tcpSynRetransmits {
				c.abort(errTCPTimeout)
				return
			}
			c.stalled = 0
			c.retransmitTicks = min(c.retransmitTicks*2, tcpRetransmitTicksMost)
			if !c.established {
				c.sendSynAck()
				c.timedAt = time.Time{}
			} else {
				// what the system does not say it has is lost, and the way to it is tried with a segment again
				c.congested()
				c.cwnd, c.lost = uint32(c.mss), c.sndNxt
				c.resend()
			}
		}
		c.arm(tcpTick)
	case c.blocked:
		// the segment that opens the window may have been lost, this one is answered with the window
		if c.stalled++; c.stalled%tcpProbeTicks == 0 {
			c.send(c.sndUna-1, tcpAck, nil, nil)
		}
		c.arm(tcpTick)
	case c.closed:
		if c.lingered {
			c.reset()
			return
		}
		c.lingered = true
		c.arm(tcpLinger)
	}
}

// reset tells the system that the connection is gone. c.mu is held
func (c *tcpConn) reset() {
	c.send(c.sndNxt, tcpRst|tcpAck, nil, nil)
	c.abort(io.ErrClosedPipe)
}

// abort ends the connection without another word to the system. What it has sent is still there to read.
// c.mu is held
func (c *tcpConn) abort(err error) {
	if c.err != nil {
		return
	}
	c.err = err
	if c.timer != nil {
		c.timer.Stop()
	}
	for _, h := range c.rcvHeld {
		h.data.Release()
	}
	c.rcvHeld = nil
	for _, b := range c.sndQueue[:len(c.sndQueue)-c.writing] {
		b.Release()
	}
	c.sndQueue = nil
	c.readable.Broadcast()
	c.writable.Broadcast()
	c.stack.remove(c)
}

// wait waits for something to read. c.mu is held
func (c *tcpConn) wait() error {
	for c.rcvLen == 0 && !c.rcvFin && !c.closed && c.err == nil {
		c.readable.Wait()
	}
	switch {
	case c.rcvLen > 0:
		return nil
	case c.err != nil && !c.rcvFin:
		return c.err
	}
	return io.EOF
}

// taken is called when size bytes have been read. c.mu is held
func (c *tcpConn) taken(size int32) {
	c.rcvLen -= size
	// the system may also have stopped at the window that it was told
	if c.err == nil && (c.rcvAcked != c.rcvNxt || c.rcvAdvertised < c.window/2) {
		c.sendAck()
	}
}

// ReadMultiBuffer implements buf.Reader
func (c *tcpConn) ReadMultiBuffer() (buf.MultiBuffer, error) {
	c.mu.Lock()
	defer c.mu.Unlock()
	if err := c.wait(); err != nil {
		return nil, err
	}
	mb := c.rcvQueue
	c.rcvQueue = nil
	c.taken(c.rcvLen)
	return mb, nil
}

// Read implements net.Conn
func (c *tcpConn) Read(p []byte) (int, error) {
	c.mu.Lock()
	defer c.mu.Unlock()
	if err := c.wait(); err != nil {
		return 0, err
	}
	var n int
	c.rcvQueue, n = buf.SplitBytes(c.rcvQueue, p)
	c.taken(int32(n))
	return n, nil
}

// WriteMultiBuffer implements buf.Writer. It is not called again before it returns
func (c *tcpConn) WriteMultiBuffer(mb buf.MultiBuffer) error {
	c.mu.Lock()
	defer c.mu.Unlock()
	// without the empty ones
	n, size := 0, uint32(0)
	for _, b := range mb {
		if b.IsEmpty() {
			b.Release()
			continue
		}
		mb[n] = b
		n++
		size += uint32(b.Len())
	}
	mb = mb[:n]
	if c.err != nil || c.closed {
		buf.ReleaseMulti(mb)
		return io.ErrClosedPipe
	}
	// what is written is in sndQueue until it is acknowledged, and is not released before all of it is sent
	c.sndQueue = append(c.sndQueue, mb...)
	c.writing = len(mb)
	// a segment is one part of a Buffer, or as much as there is to send if the tun splits it by itself
	var room [tcpParts][]byte
	limit, parts := c.mss, room[:0:1]
	if c.stack.segmenter != nil {
		limit, parts = tcpSegment, room[:0]
	}
	for end := c.sndNxt + size; c.sndNxt != end; {
		// not a part of a segment when all of it can go later, with what is acknowledged by then
		for c.err == nil {
			if space := c.space(); space >= min(end-c.sndNxt, uint32(c.mss)) || space != 0 && c.sndUna == c.sndNxt {
				break
			}
			c.blocked = true
			c.arm(tcpTick)
			c.writable.Wait()
			c.blocked = false
		}
		if c.err != nil {
			break
		}
		most := min(limit, int(c.space()))
		if most > c.mss {
			// the same for what the tun splits
			most -= most % c.mss
		}
		p, n := c.queued(parts, c.sndNxt, most)
		flags := byte(tcpAck)
		if c.sndNxt+uint32(n) == end {
			flags |= tcpPsh
		}
		var b [ipv6HeaderSize + tcpHeaderSize]byte
		h := b[:c.header(b[:], c.sndNxt, flags, nil, n)]
		if c.sndUna == c.sndNxt {
			// the ticks without an acknowledgment are counted from the first segment that waits for one
			c.stalled = 0
		}
		c.sndNxt += uint32(n)
		if c.timedAt.IsZero() {
			c.timed, c.timedAt = c.sndNxt, time.Now()
		}
		c.arm(tcpTick)
		// segments that come in the meantime are taken while this one is written, and acknowledged after it
		c.sending = true
		c.mu.Unlock()
		if s := c.stack.segmenter; s != nil {
			t := h[c.ipSize:]
			binary.BigEndian.PutUint16(t[16:], ^checksum(pseudoSum(h[:c.ipSize], protoTCP, len(t)+n)))
			_ = s.WriteSegment(h, p, c.mss)
		} else {
			c.write(h, p[0])
		}
		c.mu.Lock()
		c.sending = false
		if c.ackWanted && c.err == nil {
			c.sendAck()
		}
		c.ackWanted = false
	}
	c.writing = 0
	if c.err != nil {
		buf.ReleaseMulti(mb)
		return io.ErrClosedPipe
	}
	c.release()
	return nil
}

// Write implements net.Conn
func (c *tcpConn) Write(p []byte) (int, error) {
	if err := c.WriteMultiBuffer(buf.MergeBytes(nil, p)); err != nil {
		return 0, err
	}
	return len(p), nil
}

// Close implements net.Conn
func (c *tcpConn) Close() error {
	c.mu.Lock()
	defer c.mu.Unlock()
	if c.closed {
		return nil
	}
	c.closed = true
	unread := c.rcvLen > 0
	c.rcvQueue, c.rcvLen = buf.ReleaseMulti(c.rcvQueue), 0
	if c.err != nil {
		return nil
	}
	if unread || c.writing != 0 {
		// something was not read or not written, the system has to know
		c.reset()
		return nil
	}
	c.readable.Broadcast()
	c.send(c.sndNxt, tcpFin|tcpAck, nil, nil)
	c.sndFin = true
	c.sndNxt++
	c.arm(tcpTick)
	return nil
}

func (c *tcpConn) LocalAddr() net.Addr {
	return c.dst.RawNetAddr()
}

func (c *tcpConn) RemoteAddr() net.Addr {
	return c.src.RawNetAddr()
}

func (c *tcpConn) SetDeadline(time.Time) error {
	return nil
}

func (c *tcpConn) SetReadDeadline(time.Time) error {
	return nil
}

func (c *tcpConn) SetWriteDeadline(time.Time) error {
	return nil
}