blob: a87cac232ea16e3a4a46efc21b47a8ac0f1c611c [file] [log] [blame]
// Copyright 2023 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
//go:build go1.21
package quic
import (
"crypto/tls"
"errors"
"time"
)
// maybeSend sends datagrams, if possible.
//
// If sending is blocked by pacing, it returns the next time
// a datagram may be sent.
//
// If sending is blocked indefinitely, it returns the zero Time.
func (c *Conn) maybeSend(now time.Time) (next time.Time) {
// Assumption: The congestion window is not underutilized.
// If congestion control, pacing, and anti-amplification all permit sending,
// but we have no packet to send, then we will declare the window underutilized.
underutilized := false
defer func() {
c.loss.cc.setUnderutilized(c.log, underutilized)
}()
// Send one datagram on each iteration of this loop,
// until we hit a limit or run out of data to send.
//
// For each number space where we have write keys,
// attempt to construct a packet in that space.
// If the packet contains no frames (we have no data in need of sending),
// abandon the packet.
//
// Speculatively constructing packets means we don't need
// separate code paths for "do we have data to send?" and
// "send the data" that need to be kept in sync.
for {
limit, next := c.loss.sendLimit(now)
if limit == ccBlocked {
// If anti-amplification blocks sending, then no packet can be sent.
return next
}
if !c.sendOK(now) {
return time.Time{}
}
// We may still send ACKs, even if congestion control or pacing limit sending.
// Prepare to write a datagram of at most maxSendSize bytes.
c.w.reset(c.loss.maxSendSize())
dstConnID, ok := c.connIDState.dstConnID()
if !ok {
// It is currently not possible for us to end up without a connection ID,
// but handle the case anyway.
return time.Time{}
}
// Initial packet.
pad := false
var sentInitial *sentPacket
if c.keysInitial.canWrite() {
pnumMaxAcked := c.loss.spaces[initialSpace].maxAcked
pnum := c.loss.nextNumber(initialSpace)
p := longPacket{
ptype: packetTypeInitial,
version: quicVersion1,
num: pnum,
dstConnID: dstConnID,
srcConnID: c.connIDState.srcConnID(),
extra: c.retryToken,
}
c.w.startProtectedLongHeaderPacket(pnumMaxAcked, p)
c.appendFrames(now, initialSpace, pnum, limit)
if logPackets {
logSentPacket(c, packetTypeInitial, pnum, p.srcConnID, p.dstConnID, c.w.payload())
}
if c.logEnabled(QLogLevelPacket) && len(c.w.payload()) > 0 {
c.logPacketSent(packetTypeInitial, pnum, p.srcConnID, p.dstConnID, c.w.packetLen(), c.w.payload())
}
sentInitial = c.w.finishProtectedLongHeaderPacket(pnumMaxAcked, c.keysInitial.w, p)
if sentInitial != nil {
// Client initial packets and ack-eliciting server initial packaets
// need to be sent in a datagram padded to at least 1200 bytes.
// We can't add the padding yet, however, since we may want to
// coalesce additional packets with this one.
if c.side == clientSide || sentInitial.ackEliciting {
pad = true
}
}
}
// Handshake packet.
if c.keysHandshake.canWrite() {
pnumMaxAcked := c.loss.spaces[handshakeSpace].maxAcked
pnum := c.loss.nextNumber(handshakeSpace)
p := longPacket{
ptype: packetTypeHandshake,
version: quicVersion1,
num: pnum,
dstConnID: dstConnID,
srcConnID: c.connIDState.srcConnID(),
}
c.w.startProtectedLongHeaderPacket(pnumMaxAcked, p)
c.appendFrames(now, handshakeSpace, pnum, limit)
if logPackets {
logSentPacket(c, packetTypeHandshake, pnum, p.srcConnID, p.dstConnID, c.w.payload())
}
if c.logEnabled(QLogLevelPacket) && len(c.w.payload()) > 0 {
c.logPacketSent(packetTypeHandshake, pnum, p.srcConnID, p.dstConnID, c.w.packetLen(), c.w.payload())
}
if sent := c.w.finishProtectedLongHeaderPacket(pnumMaxAcked, c.keysHandshake.w, p); sent != nil {
c.packetSent(now, handshakeSpace, sent)
if c.side == clientSide {
// "[...] a client MUST discard Initial keys when it first
// sends a Handshake packet [...]"
// https://www.rfc-editor.org/rfc/rfc9001.html#section-4.9.1-2
c.discardKeys(now, initialSpace)
}
}
}
// 1-RTT packet.
if c.keysAppData.canWrite() {
pnumMaxAcked := c.loss.spaces[appDataSpace].maxAcked
pnum := c.loss.nextNumber(appDataSpace)
c.w.start1RTTPacket(pnum, pnumMaxAcked, dstConnID)
c.appendFrames(now, appDataSpace, pnum, limit)
if pad && len(c.w.payload()) > 0 {
// 1-RTT packets have no length field and extend to the end
// of the datagram, so if we're sending a datagram that needs
// padding we need to add it inside the 1-RTT packet.
c.w.appendPaddingTo(paddedInitialDatagramSize)
pad = false
}
if logPackets {
logSentPacket(c, packetType1RTT, pnum, nil, dstConnID, c.w.payload())
}
if c.logEnabled(QLogLevelPacket) && len(c.w.payload()) > 0 {
c.logPacketSent(packetType1RTT, pnum, nil, dstConnID, c.w.packetLen(), c.w.payload())
}
if sent := c.w.finish1RTTPacket(pnum, pnumMaxAcked, dstConnID, &c.keysAppData); sent != nil {
c.packetSent(now, appDataSpace, sent)
}
}
buf := c.w.datagram()
if len(buf) == 0 {
if limit == ccOK {
// We have nothing to send, and congestion control does not
// block sending. The congestion window is underutilized.
underutilized = true
}
return next
}
if sentInitial != nil {
if pad {
// Pad out the datagram with zeros, coalescing the Initial
// packet with invalid packets that will be ignored by the peer.
// https://www.rfc-editor.org/rfc/rfc9000.html#section-14.1-1
for len(buf) < paddedInitialDatagramSize {
buf = append(buf, 0)
// Technically this padding isn't in any packet, but
// account it to the Initial packet in this datagram
// for purposes of flow control and loss recovery.
sentInitial.size++
sentInitial.inFlight = true
}
}
// If we're a client and this Initial packet is coalesced
// with a Handshake packet, then we've discarded Initial keys
// since constructing the packet and shouldn't record it as in-flight.
if c.keysInitial.canWrite() {
c.packetSent(now, initialSpace, sentInitial)
}
}
c.endpoint.sendDatagram(datagram{
b: buf,
peerAddr: c.peerAddr,
})
}
}
func (c *Conn) packetSent(now time.Time, space numberSpace, sent *sentPacket) {
c.idleHandlePacketSent(now, sent)
c.loss.packetSent(now, c.log, space, sent)
}
func (c *Conn) appendFrames(now time.Time, space numberSpace, pnum packetNumber, limit ccLimit) {
if c.lifetime.localErr != nil {
c.appendConnectionCloseFrame(now, space, c.lifetime.localErr)
return
}
shouldSendAck := c.acks[space].shouldSendAck(now)
if limit != ccOK {
// ACKs are not limited by congestion control.
if shouldSendAck && c.appendAckFrame(now, space) {
c.acks[space].sentAck()
}
return
}
// We want to send an ACK frame if the ack controller wants to send a frame now,
// OR if we are sending a packet anyway and have ack-eliciting packets which we
// have not yet acked.
//
// We speculatively add ACK frames here, to put them at the front of the packet
// to avoid truncation.
//
// After adding all frames, if we don't need to send an ACK frame and have not
// added any other frames, we abandon the packet.
if c.appendAckFrame(now, space) {
defer func() {
// All frames other than ACK and PADDING are ack-eliciting,
// so if the packet is ack-eliciting we've added additional
// frames to it.
if !shouldSendAck && !c.w.sent.ackEliciting {
// There's nothing in this packet but ACK frames, and
// we don't want to send an ACK-only packet at this time.
// Abandoning the packet means we wrote an ACK frame for
// nothing, but constructing the frame is cheap.
c.w.abandonPacket()
return
}
// Either we are willing to send an ACK-only packet,
// or we've added additional frames.
c.acks[space].sentAck()
if !c.w.sent.ackEliciting && c.shouldMakePacketAckEliciting() {
c.w.appendPingFrame()
}
}()
}
if limit != ccOK {
return
}
pto := c.loss.ptoExpired
// TODO: Add all the other frames we can send.
// CRYPTO
c.crypto[space].dataToSend(pto, func(off, size int64) int64 {
b, _ := c.w.appendCryptoFrame(off, int(size))
c.crypto[space].sendData(off, b)
return int64(len(b))
})
// Test-only PING frames.
if space == c.testSendPingSpace && c.testSendPing.shouldSendPTO(pto) {
if !c.w.appendPingFrame() {
return
}
c.testSendPing.setSent(pnum)
}
if space == appDataSpace {
// HANDSHAKE_DONE
if c.handshakeConfirmed.shouldSendPTO(pto) {
if !c.w.appendHandshakeDoneFrame() {
return
}
c.handshakeConfirmed.setSent(pnum)
}
// NEW_CONNECTION_ID, RETIRE_CONNECTION_ID
if !c.connIDState.appendFrames(c, pnum, pto) {
return
}
// PATH_RESPONSE
if pad, ok := c.appendPathFrames(); !ok {
return
} else if pad {
defer c.w.appendPaddingTo(smallestMaxDatagramSize)
}
// All stream-related frames. This should come last in the packet,
// so large amounts of STREAM data don't crowd out other frames
// we may need to send.
if !c.appendStreamFrames(&c.w, pnum, pto) {
return
}
if !c.appendKeepAlive(now) {
return
}
}
// If this is a PTO probe and we haven't added an ack-eliciting frame yet,
// add a PING to make this an ack-eliciting probe.
//
// Technically, there are separate PTO timers for each number space.
// When a PTO timer expires, we MUST send an ack-eliciting packet in the
// timer's space. We SHOULD send ack-eliciting packets in every other space
// with in-flight data. (RFC 9002, section 6.2.4)
//
// What we actually do is send a single datagram containing an ack-eliciting packet
// for every space for which we have keys.
//
// We fill the PTO probe packets with new or unacknowledged data. For example,
// a PTO probe sent for the Initial space will generally retransmit previously
// sent but unacknowledged CRYPTO data.
//
// When sending a PTO probe datagram containing multiple packets, it is
// possible that an earlier packet will fill up the datagram, leaving no
// space for the remaining probe packet(s). This is not a problem in practice.
//
// A client discards Initial keys when it first sends a Handshake packet
// (RFC 9001 Section 4.9.1). Handshake keys are discarded when the handshake
// is confirmed (RFC 9001 Section 4.9.2). The PTO timer is not set for the
// Application Data packet number space until the handshake is confirmed
// (RFC 9002 Section 6.2.1). Therefore, the only times a PTO probe can fire
// while data for multiple spaces is in flight are:
//
// - a server's Initial or Handshake timers can fire while Initial and Handshake
// data is in flight; and
//
// - a client's Handshake timer can fire while Handshake and Application Data
// data is in flight.
//
// It is theoretically possible for a server's Initial CRYPTO data to overflow
// the maximum datagram size, but unlikely in practice; this space contains
// only the ServerHello TLS message, which is small. It's also unlikely that
// the Handshake PTO probe will fire while Initial data is in flight (this
// requires not just that the Initial CRYPTO data completely fill a datagram,
// but a quite specific arrangement of lost and retransmitted packets.)
// We don't bother worrying about this case here, since the worst case is
// that we send a PTO probe for the in-flight Initial data and drop the
// Handshake probe.
//
// If a client's Handshake PTO timer fires while Application Data data is in
// flight, it is possible that the resent Handshake CRYPTO data will crowd
// out the probe for the Application Data space. However, since this probe is
// optional (recall that the Application Data PTO timer is never set until
// after Handshake keys have been discarded), dropping it is acceptable.
if pto && !c.w.sent.ackEliciting {
c.w.appendPingFrame()
}
}
// shouldMakePacketAckEliciting is called when sending a packet containing nothing but an ACK frame.
// It reports whether we should add a PING frame to the packet to make it ack-eliciting.
func (c *Conn) shouldMakePacketAckEliciting() bool {
if c.keysAppData.needAckEliciting() {
// The peer has initiated a key update.
// We haven't sent them any packets yet in the new phase.
// Make this an ack-eliciting packet.
// Their ack of this packet will complete the key update.
return true
}
if c.loss.consecutiveNonAckElicitingPackets >= 19 {
// We've sent a run of non-ack-eliciting packets.
// Add in an ack-eliciting one every once in a while so the peer
// lets us know which ones have arrived.
//
// Google QUICHE injects a PING after sending 19 packets. We do the same.
//
// https://www.rfc-editor.org/rfc/rfc9000#section-13.2.4-2
return true
}
// TODO: Consider making every packet sent when in PTO ack-eliciting to speed up recovery.
return false
}
func (c *Conn) appendAckFrame(now time.Time, space numberSpace) bool {
seen, delay := c.acks[space].acksToSend(now)
if len(seen) == 0 {
return false
}
d := unscaledAckDelayFromDuration(delay, ackDelayExponent)
return c.w.appendAckFrame(seen, d)
}
func (c *Conn) appendConnectionCloseFrame(now time.Time, space numberSpace, err error) {
c.sentConnectionClose(now)
switch e := err.(type) {
case localTransportError:
c.w.appendConnectionCloseTransportFrame(e.code, 0, e.reason)
case *ApplicationError:
if space != appDataSpace {
// "CONNECTION_CLOSE frames signaling application errors (type 0x1d)
// MUST only appear in the application data packet number space."
// https://www.rfc-editor.org/rfc/rfc9000#section-12.5-2.2
c.w.appendConnectionCloseTransportFrame(errApplicationError, 0, "")
} else {
c.w.appendConnectionCloseApplicationFrame(e.Code, e.Reason)
}
default:
// TLS alerts are sent using error codes [0x0100,0x01ff).
// https://www.rfc-editor.org/rfc/rfc9000#section-20.1-2.36.1
var alert tls.AlertError
switch {
case errors.As(err, &alert):
// tls.AlertError is a uint8, so this can't exceed 0x01ff.
code := errTLSBase + transportError(alert)
c.w.appendConnectionCloseTransportFrame(code, 0, "")
default:
c.w.appendConnectionCloseTransportFrame(errInternal, 0, "")
}
}
}