net/http2/server.go

// Copyright 2014 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.
// See https://code.google.com/p/go/source/browse/CONTRIBUTORS
// Licensed under the same terms as Go itself:
// https://code.google.com/p/go/source/browse/LICENSE

// TODO: replace all <-sc.doneServing with reads from the stream's cw
// instead, and make sure that on close we close all open
// streams. then remove doneServing?

// TODO: finish GOAWAY support. Consider each incoming frame type and
// whether it should be ignored during a shutdown race.

// TODO: disconnect idle clients. GFE seems to do 4 minutes. make
// configurable?  or maximum number of idle clients and remove the
// oldest?

// TODO: turn off the serve goroutine when idle, so
// an idle conn only has the readFrames goroutine active. (which could
// also be optimized probably to pin less memory in crypto/tls). This
// would involve tracking when the serve goroutine is active (atomic
// int32 read/CAS probably?) and starting it up when frames arrive,
// and shutting it down when all handlers exit. the occasional PING
// packets could use time.AfterFunc to call sc.wakeStartServeLoop()
// (which is a no-op if already running) and then queue the PING write
// as normal. The serve loop would then exit in most cases (if no
// Handlers running) and not be woken up again until the PING packet
// returns.

// TODO (maybe): add a mechanism for Handlers to going into
// half-closed-local mode (rw.(io.Closer) test?) but not exit their
// handler, and continue to be able to read from the
// Request.Body. This would be a somewhat semantic change from HTTP/1
// (or at least what we expose in net/http), so I'd probably want to
// add it there too. For now, this package says that returning from
// the Handler ServeHTTP function means you're both done reading and
// done writing, without a way to stop just one or the other.

package http2

import (
	"bufio"
	"bytes"
	"crypto/tls"
	"errors"
	"fmt"
	"io"
	"log"
	"net"
	"net/http"
	"net/url"
	"strconv"
	"strings"
	"sync"
	"time"

	"golang.org/x/net/http2/hpack"
)

const (
	prefaceTimeout        = 10 * time.Second
	firstSettingsTimeout  = 2 * time.Second // should be in-flight with preface anyway
	handlerChunkWriteSize = 4 << 10
	defaultMaxStreams     = 250 // TODO: make this 100 as the GFE seems to?
)

var (
	errClientDisconnected = errors.New("client disconnected")
	errClosedBody         = errors.New("body closed by handler")
	errStreamBroken       = errors.New("http2: stream broken")
)

var responseWriterStatePool = sync.Pool{
	New: func() interface{} {
		rws := &responseWriterState{}
		rws.bw = bufio.NewWriterSize(chunkWriter{rws}, handlerChunkWriteSize)
		return rws
	},
}

// Test hooks.
var (
	testHookOnConn        func()
	testHookGetServerConn func(*serverConn)
	testHookOnPanicMu     *sync.Mutex // nil except in tests
	testHookOnPanic       func(sc *serverConn, panicVal interface{}) (rePanic bool)
)

// Server is an HTTP/2 server.
type Server struct {
	// MaxHandlers limits the number of http.Handler ServeHTTP goroutines
	// which may run at a time over all connections.
	// Negative or zero no limit.
	// TODO: implement
	MaxHandlers int

	// MaxConcurrentStreams optionally specifies the number of
	// concurrent streams that each client may have open at a
	// time. This is unrelated to the number of http.Handler goroutines
	// which may be active globally, which is MaxHandlers.
	// If zero, MaxConcurrentStreams defaults to at least 100, per
	// the HTTP/2 spec's recommendations.
	MaxConcurrentStreams uint32

	// MaxReadFrameSize optionally specifies the largest frame
	// this server is willing to read. A valid value is between
	// 16k and 16M, inclusive. If zero or otherwise invalid, a
	// default value is used.
	MaxReadFrameSize uint32

	// PermitProhibitedCipherSuites, if true, permits the use of
	// cipher suites prohibited by the HTTP/2 spec.
	PermitProhibitedCipherSuites bool
}

func (s *Server) maxReadFrameSize() uint32 {
	if v := s.MaxReadFrameSize; v >= minMaxFrameSize && v <= maxFrameSize {
		return v
	}
	return defaultMaxReadFrameSize
}

func (s *Server) maxConcurrentStreams() uint32 {
	if v := s.MaxConcurrentStreams; v > 0 {
		return v
	}
	return defaultMaxStreams
}

// ConfigureServer adds HTTP/2 support to a net/http Server.
//
// The configuration conf may be nil.
//
// ConfigureServer must be called before s begins serving.
func ConfigureServer(s *http.Server, conf *Server) {
	if conf == nil {
		conf = new(Server)
	}
	if s.TLSConfig == nil {
		s.TLSConfig = new(tls.Config)
	}

	// Note: not setting MinVersion to tls.VersionTLS12,
	// as we don't want to interfere with HTTP/1.1 traffic
	// on the user's server. We enforce TLS 1.2 later once
	// we accept a connection. Ideally this should be done
	// during next-proto selection, but using TLS <1.2 with
	// HTTP/2 is still the client's bug.

	// Be sure we advertise tls.TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256
	// at least.
	// TODO: enable PreferServerCipherSuites?
	if s.TLSConfig.CipherSuites != nil {
		const requiredCipher = tls.TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256
		haveRequired := false
		for _, v := range s.TLSConfig.CipherSuites {
			if v == requiredCipher {
				haveRequired = true
				break
			}
		}
		if !haveRequired {
			s.TLSConfig.CipherSuites = append(s.TLSConfig.CipherSuites, requiredCipher)
		}
	}

	haveNPN := false
	for _, p := range s.TLSConfig.NextProtos {
		if p == NextProtoTLS {
			haveNPN = true
			break
		}
	}
	if !haveNPN {
		s.TLSConfig.NextProtos = append(s.TLSConfig.NextProtos, NextProtoTLS)
	}
	// h2-14 is temporary (as of 2015-03-05) while we wait for all browsers
	// to switch to "h2".
	s.TLSConfig.NextProtos = append(s.TLSConfig.NextProtos, "h2-14")

	if s.TLSNextProto == nil {
		s.TLSNextProto = map[string]func(*http.Server, *tls.Conn, http.Handler){}
	}
	protoHandler := func(hs *http.Server, c *tls.Conn, h http.Handler) {
		if testHookOnConn != nil {
			testHookOnConn()
		}
		conf.handleConn(hs, c, h)
	}
	s.TLSNextProto[NextProtoTLS] = protoHandler
	s.TLSNextProto["h2-14"] = protoHandler // temporary; see above.
}

func (srv *Server) handleConn(hs *http.Server, c net.Conn, h http.Handler) {
	sc := &serverConn{
		srv:              srv,
		hs:               hs,
		conn:             c,
		remoteAddrStr:    c.RemoteAddr().String(),
		bw:               newBufferedWriter(c),
		handler:          h,
		streams:          make(map[uint32]*stream),
		readFrameCh:      make(chan readFrameResult),
		wantWriteFrameCh: make(chan frameWriteMsg, 8),
		wroteFrameCh:     make(chan struct{}, 1), // buffered; one send in reading goroutine
		bodyReadCh:       make(chan bodyReadMsg), // buffering doesn't matter either way
		doneServing:      make(chan struct{}),
		advMaxStreams:    srv.maxConcurrentStreams(),
		writeSched: writeScheduler{
			maxFrameSize: initialMaxFrameSize,
		},
		initialWindowSize: initialWindowSize,
		headerTableSize:   initialHeaderTableSize,
		serveG:            newGoroutineLock(),
		pushEnabled:       true,
	}
	sc.flow.add(initialWindowSize)
	sc.inflow.add(initialWindowSize)
	sc.hpackEncoder = hpack.NewEncoder(&sc.headerWriteBuf)
	sc.hpackDecoder = hpack.NewDecoder(initialHeaderTableSize, sc.onNewHeaderField)

	fr := NewFramer(sc.bw, c)
	fr.SetMaxReadFrameSize(srv.maxReadFrameSize())
	sc.framer = fr

	if tc, ok := c.(*tls.Conn); ok {
		sc.tlsState = new(tls.ConnectionState)
		*sc.tlsState = tc.ConnectionState()
		// 9.2 Use of TLS Features
		// An implementation of HTTP/2 over TLS MUST use TLS
		// 1.2 or higher with the restrictions on feature set
		// and cipher suite described in this section. Due to
		// implementation limitations, it might not be
		// possible to fail TLS negotiation. An endpoint MUST
		// immediately terminate an HTTP/2 connection that
		// does not meet the TLS requirements described in
		// this section with a connection error (Section
		// 5.4.1) of type INADEQUATE_SECURITY.
		if sc.tlsState.Version < tls.VersionTLS12 {
			sc.rejectConn(ErrCodeInadequateSecurity, "TLS version too low")
			return
		}

		if sc.tlsState.ServerName == "" {
			// Client must use SNI, but we don't enforce that anymore,
			// since it was causing problems when connecting to bare IP
			// addresses during development.
			//
			// TODO: optionally enforce? Or enforce at the time we receive
			// a new request, and verify the the ServerName matches the :authority?
			// But that precludes proxy situations, perhaps.
			//
			// So for now, do nothing here again.
		}

		if !srv.PermitProhibitedCipherSuites && isBadCipher(sc.tlsState.CipherSuite) {
			// "Endpoints MAY choose to generate a connection error
			// (Section 5.4.1) of type INADEQUATE_SECURITY if one of
			// the prohibited cipher suites are negotiated."
			//
			// We choose that. In my opinion, the spec is weak
			// here. It also says both parties must support at least
			// TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256 so there's no
			// excuses here. If we really must, we could allow an
			// "AllowInsecureWeakCiphers" option on the server later.
			// Let's see how it plays out first.
			sc.rejectConn(ErrCodeInadequateSecurity, fmt.Sprintf("Prohibited TLS 1.2 Cipher Suite: %x", sc.tlsState.CipherSuite))
			return
		}
	}

	if hook := testHookGetServerConn; hook != nil {
		hook(sc)
	}
	sc.serve()
}

// isBadCipher reports whether the cipher is blacklisted by the HTTP/2 spec.
func isBadCipher(cipher uint16) bool {
	switch cipher {
	case tls.TLS_RSA_WITH_RC4_128_SHA,
		tls.TLS_RSA_WITH_3DES_EDE_CBC_SHA,
		tls.TLS_RSA_WITH_AES_128_CBC_SHA,
		tls.TLS_RSA_WITH_AES_256_CBC_SHA,
		tls.TLS_ECDHE_ECDSA_WITH_RC4_128_SHA,
		tls.TLS_ECDHE_ECDSA_WITH_AES_128_CBC_SHA,
		tls.TLS_ECDHE_ECDSA_WITH_AES_256_CBC_SHA,
		tls.TLS_ECDHE_RSA_WITH_RC4_128_SHA,
		tls.TLS_ECDHE_RSA_WITH_3DES_EDE_CBC_SHA,
		tls.TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA,
		tls.TLS_ECDHE_RSA_WITH_AES_256_CBC_SHA:
		// Reject cipher suites from Appendix A.
		// "This list includes those cipher suites that do not
		// offer an ephemeral key exchange and those that are
		// based on the TLS null, stream or block cipher type"
		return true
	default:
		return false
	}
}

func (sc *serverConn) rejectConn(err ErrCode, debug string) {
	sc.vlogf("REJECTING conn: %v, %s", err, debug)
	// ignoring errors. hanging up anyway.
	sc.framer.WriteGoAway(0, err, []byte(debug))
	sc.bw.Flush()
	sc.conn.Close()
}

type serverConn struct {
	// Immutable:
	srv              *Server
	hs               *http.Server
	conn             net.Conn
	bw               *bufferedWriter // writing to conn
	handler          http.Handler
	framer           *Framer
	hpackDecoder     *hpack.Decoder
	doneServing      chan struct{}        // closed when serverConn.serve ends
	readFrameCh      chan readFrameResult // written by serverConn.readFrames
	wantWriteFrameCh chan frameWriteMsg   // from handlers -> serve
	wroteFrameCh     chan struct{}        // from writeFrameAsync -> serve, tickles more frame writes
	bodyReadCh       chan bodyReadMsg     // from handlers -> serve
	testHookCh       chan func(int)       // code to run on the serve loop
	flow             flow                 // conn-wide (not stream-specific) outbound flow control
	inflow           flow                 // conn-wide inbound flow control
	tlsState         *tls.ConnectionState // shared by all handlers, like net/http
	remoteAddrStr    string

	// Everything following is owned by the serve loop; use serveG.check():
	serveG                goroutineLock // used to verify funcs are on serve()
	pushEnabled           bool
	sawFirstSettings      bool // got the initial SETTINGS frame after the preface
	needToSendSettingsAck bool
	unackedSettings       int    // how many SETTINGS have we sent without ACKs?
	clientMaxStreams      uint32 // SETTINGS_MAX_CONCURRENT_STREAMS from client (our PUSH_PROMISE limit)
	advMaxStreams         uint32 // our SETTINGS_MAX_CONCURRENT_STREAMS advertised the client
	curOpenStreams        uint32 // client's number of open streams
	maxStreamID           uint32 // max ever seen
	streams               map[uint32]*stream
	initialWindowSize     int32
	headerTableSize       uint32
	peerMaxHeaderListSize uint32            // zero means unknown (default)
	canonHeader           map[string]string // http2-lower-case -> Go-Canonical-Case
	req                   requestParam      // non-zero while reading request headers
	writingFrame          bool              // started write goroutine but haven't heard back on wroteFrameCh
	needsFrameFlush       bool              // last frame write wasn't a flush
	writeSched            writeScheduler
	inGoAway              bool // we've started to or sent GOAWAY
	needToSendGoAway      bool // we need to schedule a GOAWAY frame write
	goAwayCode            ErrCode
	shutdownTimerCh       <-chan time.Time // nil until used
	shutdownTimer         *time.Timer      // nil until used

	// Owned by the writeFrameAsync goroutine:
	headerWriteBuf bytes.Buffer
	hpackEncoder   *hpack.Encoder
}

func (sc *serverConn) maxHeaderListSize() uint32 {
	n := sc.hs.MaxHeaderBytes
	if n <= 0 {
		n = http.DefaultMaxHeaderBytes
	}
	// http2's count is in a slightly different unit and includes 32 bytes per pair.
	// So, take the net/http.Server value and pad it up a bit, assuming 10 headers.
	const perFieldOverhead = 32 // per http2 spec
	const typicalHeaders = 10   // conservative
	return uint32(n + typicalHeaders*perFieldOverhead)
}

// requestParam is the state of the next request, initialized over
// potentially several frames HEADERS + zero or more CONTINUATION
// frames.
type requestParam struct {
	// stream is non-nil if we're reading (HEADER or CONTINUATION)
	// frames for a request (but not DATA).
	stream            *stream
	header            http.Header
	method, path      string
	scheme, authority string
	sawRegularHeader  bool  // saw a non-pseudo header already
	invalidHeader     bool  // an invalid header was seen
	headerListSize    int64 // actually uint32, but easier math this way
}

// stream represents a stream. This is the minimal metadata needed by
// the serve goroutine. Most of the actual stream state is owned by
// the http.Handler's goroutine in the responseWriter. Because the
// responseWriter's responseWriterState is recycled at the end of a
// handler, this struct intentionally has no pointer to the
// *responseWriter{,State} itself, as the Handler ending nils out the
// responseWriter's state field.
type stream struct {
	// immutable:
	id   uint32
	body *pipe       // non-nil if expecting DATA frames
	cw   closeWaiter // closed wait stream transitions to closed state

	// owned by serverConn's serve loop:
	bodyBytes     int64   // body bytes seen so far
	declBodyBytes int64   // or -1 if undeclared
	flow          flow    // limits writing from Handler to client
	inflow        flow    // what the client is allowed to POST/etc to us
	parent        *stream // or nil
	weight        uint8
	state         streamState
	sentReset     bool // only true once detached from streams map
	gotReset      bool // only true once detacted from streams map
}

func (sc *serverConn) Framer() *Framer  { return sc.framer }
func (sc *serverConn) CloseConn() error { return sc.conn.Close() }
func (sc *serverConn) Flush() error     { return sc.bw.Flush() }
func (sc *serverConn) HeaderEncoder() (*hpack.Encoder, *bytes.Buffer) {
	return sc.hpackEncoder, &sc.headerWriteBuf
}

func (sc *serverConn) state(streamID uint32) (streamState, *stream) {
	sc.serveG.check()
	// http://http2.github.io/http2-spec/#rfc.section.5.1
	if st, ok := sc.streams[streamID]; ok {
		return st.state, st
	}
	// "The first use of a new stream identifier implicitly closes all
	// streams in the "idle" state that might have been initiated by
	// that peer with a lower-valued stream identifier. For example, if
	// a client sends a HEADERS frame on stream 7 without ever sending a
	// frame on stream 5, then stream 5 transitions to the "closed"
	// state when the first frame for stream 7 is sent or received."
	if streamID <= sc.maxStreamID {
		return stateClosed, nil
	}
	return stateIdle, nil
}

func (sc *serverConn) vlogf(format string, args ...interface{}) {
	if VerboseLogs {
		sc.logf(format, args...)
	}
}

func (sc *serverConn) logf(format string, args ...interface{}) {
	if lg := sc.hs.ErrorLog; lg != nil {
		lg.Printf(format, args...)
	} else {
		log.Printf(format, args...)
	}
}

func (sc *serverConn) condlogf(err error, format string, args ...interface{}) {
	if err == nil {
		return
	}
	str := err.Error()
	if err == io.EOF || strings.Contains(str, "use of closed network connection") {
		// Boring, expected errors.
		sc.vlogf(format, args...)
	} else {
		sc.logf(format, args...)
	}
}

func (sc *serverConn) onNewHeaderField(f hpack.HeaderField) {
	sc.serveG.check()
	sc.vlogf("got header field %+v", f)
	switch {
	case !validHeader(f.Name):
		sc.req.invalidHeader = true
	case strings.HasPrefix(f.Name, ":"):
		if sc.req.sawRegularHeader {
			sc.logf("pseudo-header after regular header")
			sc.req.invalidHeader = true
			return
		}
		var dst *string
		switch f.Name {
		case ":method":
			dst = &sc.req.method
		case ":path":
			dst = &sc.req.path
		case ":scheme":
			dst = &sc.req.scheme
		case ":authority":
			dst = &sc.req.authority
		default:
			// 8.1.2.1 Pseudo-Header Fields
			// "Endpoints MUST treat a request or response
			// that contains undefined or invalid
			// pseudo-header fields as malformed (Section
			// 8.1.2.6)."
			sc.logf("invalid pseudo-header %q", f.Name)
			sc.req.invalidHeader = true
			return
		}
		if *dst != "" {
			sc.logf("duplicate pseudo-header %q sent", f.Name)
			sc.req.invalidHeader = true
			return
		}
		*dst = f.Value
	default:
		sc.req.sawRegularHeader = true
		sc.req.header.Add(sc.canonicalHeader(f.Name), f.Value)
		const headerFieldOverhead = 32 // per spec
		sc.req.headerListSize += int64(len(f.Name)) + int64(len(f.Value)) + headerFieldOverhead
		if sc.req.headerListSize > int64(sc.maxHeaderListSize()) {
			sc.hpackDecoder.SetEmitEnabled(false)
		}
	}
}

func (sc *serverConn) canonicalHeader(v string) string {
	sc.serveG.check()
	cv, ok := commonCanonHeader[v]
	if ok {
		return cv
	}
	cv, ok = sc.canonHeader[v]
	if ok {
		return cv
	}
	if sc.canonHeader == nil {
		sc.canonHeader = make(map[string]string)
	}
	cv = http.CanonicalHeaderKey(v)
	sc.canonHeader[v] = cv
	return cv
}

type readFrameResult struct {
	f   Frame // valid until readMore is called
	err error

	// readMore should be called once the consumer no longer needs or
	// retains f. After readMore, f is invalid and more frames can be
	// read.
	readMore func()
}

// readFrames is the loop that reads incoming frames.
// It takes care to only read one frame at a time, blocking until the
// consumer is done with the frame.
// It's run on its own goroutine.
func (sc *serverConn) readFrames() {
	gate := make(gate)
	for {
		f, err := sc.framer.ReadFrame()
		select {
		case sc.readFrameCh <- readFrameResult{f, err, gate.Done}:
		case <-sc.doneServing:
			return
		}
		select {
		case <-gate:
		case <-sc.doneServing:
			return
		}
	}
}

// writeFrameAsync runs in its own goroutine and writes a single frame
// and then reports when it's done.
// At most one goroutine can be running writeFrameAsync at a time per
// serverConn.
func (sc *serverConn) writeFrameAsync(wm frameWriteMsg) {
	err := wm.write.writeFrame(sc)
	if ch := wm.done; ch != nil {
		select {
		case ch <- err:
		default:
			panic(fmt.Sprintf("unbuffered done channel passed in for type %T", wm.write))
		}
	}
	sc.wroteFrameCh <- struct{}{} // tickle frame selection scheduler
}

func (sc *serverConn) closeAllStreamsOnConnClose() {
	sc.serveG.check()
	for _, st := range sc.streams {
		sc.closeStream(st, errClientDisconnected)
	}
}

func (sc *serverConn) stopShutdownTimer() {
	sc.serveG.check()
	if t := sc.shutdownTimer; t != nil {
		t.Stop()
	}
}

func (sc *serverConn) notePanic() {
	if testHookOnPanicMu != nil {
		testHookOnPanicMu.Lock()
		defer testHookOnPanicMu.Unlock()
	}
	if testHookOnPanic != nil {
		if e := recover(); e != nil {
			if testHookOnPanic(sc, e) {
				panic(e)
			}
		}
	}
}

func (sc *serverConn) serve() {
	sc.serveG.check()
	defer sc.notePanic()
	defer sc.conn.Close()
	defer sc.closeAllStreamsOnConnClose()
	defer sc.stopShutdownTimer()
	defer close(sc.doneServing) // unblocks handlers trying to send

	sc.vlogf("HTTP/2 connection from %v on %p", sc.conn.RemoteAddr(), sc.hs)

	sc.writeFrame(frameWriteMsg{
		write: writeSettings{
			{SettingMaxFrameSize, sc.srv.maxReadFrameSize()},
			{SettingMaxConcurrentStreams, sc.advMaxStreams},
			{SettingMaxHeaderListSize, sc.maxHeaderListSize()},

			// TODO: more actual settings, notably
			// SettingInitialWindowSize, but then we also
			// want to bump up the conn window size the
			// same amount here right after the settings
		},
	})
	sc.unackedSettings++

	if err := sc.readPreface(); err != nil {
		sc.condlogf(err, "error reading preface from client %v: %v", sc.conn.RemoteAddr(), err)
		return
	}

	go sc.readFrames() // closed by defer sc.conn.Close above

	settingsTimer := time.NewTimer(firstSettingsTimeout)
	loopNum := 0
	for {
		loopNum++
		select {
		case wm := <-sc.wantWriteFrameCh:
			sc.writeFrame(wm)
		case <-sc.wroteFrameCh:
			if sc.writingFrame != true {
				panic("internal error: expected to be already writing a frame")
			}
			sc.writingFrame = false
			sc.scheduleFrameWrite()
		case res := <-sc.readFrameCh:
			if !sc.processFrameFromReader(res) {
				return
			}
			res.readMore()
			if settingsTimer.C != nil {
				settingsTimer.Stop()
				settingsTimer.C = nil
			}
		case m := <-sc.bodyReadCh:
			sc.noteBodyRead(m.st, m.n)
		case <-settingsTimer.C:
			sc.logf("timeout waiting for SETTINGS frames from %v", sc.conn.RemoteAddr())
			return
		case <-sc.shutdownTimerCh:
			sc.vlogf("GOAWAY close timer fired; closing conn from %v", sc.conn.RemoteAddr())
			return
		case fn := <-sc.testHookCh:
			fn(loopNum)
		}
	}
}

// readPreface reads the ClientPreface greeting from the peer
// or returns an error on timeout or an invalid greeting.
func (sc *serverConn) readPreface() error {
	errc := make(chan error, 1)
	go func() {
		// Read the client preface
		buf := make([]byte, len(ClientPreface))
		if _, err := io.ReadFull(sc.conn, buf); err != nil {
			errc <- err
		} else if !bytes.Equal(buf, clientPreface) {
			errc <- fmt.Errorf("bogus greeting %q", buf)
		} else {
			errc <- nil
		}
	}()
	timer := time.NewTimer(prefaceTimeout) // TODO: configurable on *Server?
	defer timer.Stop()
	select {
	case <-timer.C:
		return errors.New("timeout waiting for client preface")
	case err := <-errc:
		if err == nil {
			sc.vlogf("client %v said hello", sc.conn.RemoteAddr())
		}
		return err
	}
}

var errChanPool = sync.Pool{
	New: func() interface{} { return make(chan error, 1) },
}

// writeDataFromHandler writes the data described in req to stream.id.
//
// The flow control currently happens in the Handler where it waits
// for 1 or more bytes to be available to then write here.  So at this
// point we know that we have flow control. But this might have to
// change when priority is implemented, so the serve goroutine knows
// the total amount of bytes waiting to be sent and can can have more
// scheduling decisions available.
func (sc *serverConn) writeDataFromHandler(stream *stream, writeData *writeData) error {
	ch := errChanPool.Get().(chan error)
	sc.writeFrameFromHandler(frameWriteMsg{
		write:  writeData,
		stream: stream,
		done:   ch,
	})
	select {
	case err := <-ch:
		errChanPool.Put(ch)
		return err
	case <-sc.doneServing:
		return errClientDisconnected
	case <-stream.cw:
		return errStreamBroken
	}
}

// writeFrameFromHandler sends wm to sc.wantWriteFrameCh, but aborts
// if the connection has gone away.
//
// This must not be run from the serve goroutine itself, else it might
// deadlock writing to sc.wantWriteFrameCh (which is only mildly
// buffered and is read by serve itself). If you're on the serve
// goroutine, call writeFrame instead.
func (sc *serverConn) writeFrameFromHandler(wm frameWriteMsg) {
	sc.serveG.checkNotOn() // NOT
	var scheduled bool
	select {
	case sc.wantWriteFrameCh <- wm:
		scheduled = true
	case <-sc.doneServing:
		// Client has closed their connection to the server.
	case <-wm.stream.cw:
		// Stream closed.
	}
	// Don't block writers expecting a reply.
	if !scheduled && wm.done != nil {
		select {
		case wm.done <- errStreamBroken:
		default:
			panic("expected buffered channel")
		}
	}
}

// writeFrame schedules a frame to write and sends it if there's nothing
// already being written.
//
// There is no pushback here (the serve goroutine never blocks). It's
// the http.Handlers that block, waiting for their previous frames to
// make it onto the wire
//
// If you're not on the serve goroutine, use writeFrameFromHandler instead.
func (sc *serverConn) writeFrame(wm frameWriteMsg) {
	sc.serveG.check()
	sc.writeSched.add(wm)
	sc.scheduleFrameWrite()
}

// startFrameWrite starts a goroutine to write wm (in a separate
// goroutine since that might block on the network), and updates the
// serve goroutine's state about the world, updated from info in wm.
func (sc *serverConn) startFrameWrite(wm frameWriteMsg) {
	sc.serveG.check()
	if sc.writingFrame {
		panic("internal error: can only be writing one frame at a time")
	}
	sc.writingFrame = true

	st := wm.stream
	if st != nil {
		switch st.state {
		case stateHalfClosedLocal:
			panic("internal error: attempt to send frame on half-closed-local stream")
		case stateClosed:
			if st.sentReset || st.gotReset {
				// Skip this frame. But fake the frame write to reschedule:
				sc.wroteFrameCh <- struct{}{}
				return
			}
			panic(fmt.Sprintf("internal error: attempt to send a write %v on a closed stream", wm))
		}
	}

	sc.needsFrameFlush = true
	if endsStream(wm.write) {
		if st == nil {
			panic("internal error: expecting non-nil stream")
		}
		switch st.state {
		case stateOpen:
			// Here we would go to stateHalfClosedLocal in
			// theory, but since our handler is done and
			// the net/http package provides no mechanism
			// for finishing writing to a ResponseWriter
			// while still reading data (see possible TODO
			// at top of this file), we go into closed
			// state here anyway, after telling the peer
			// we're hanging up on them.
			st.state = stateHalfClosedLocal // won't last long, but necessary for closeStream via resetStream
			errCancel := StreamError{st.id, ErrCodeCancel}
			sc.resetStream(errCancel)
		case stateHalfClosedRemote:
			sc.closeStream(st, nil)
		}
	}
	go sc.writeFrameAsync(wm)
}

// scheduleFrameWrite tickles the frame writing scheduler.
//
// If a frame is already being written, nothing happens. This will be called again
// when the frame is done being written.
//
// If a frame isn't being written we need to send one, the best frame
// to send is selected, preferring first things that aren't
// stream-specific (e.g. ACKing settings), and then finding the
// highest priority stream.
//
// If a frame isn't being written and there's nothing else to send, we
// flush the write buffer.
func (sc *serverConn) scheduleFrameWrite() {
	sc.serveG.check()
	if sc.writingFrame {
		return
	}
	if sc.needToSendGoAway {
		sc.needToSendGoAway = false
		sc.startFrameWrite(frameWriteMsg{
			write: &writeGoAway{
				maxStreamID: sc.maxStreamID,
				code:        sc.goAwayCode,
			},
		})
		return
	}
	if sc.needToSendSettingsAck {
		sc.needToSendSettingsAck = false
		sc.startFrameWrite(frameWriteMsg{write: writeSettingsAck{}})
		return
	}
	if !sc.inGoAway {
		if wm, ok := sc.writeSched.take(); ok {
			sc.startFrameWrite(wm)
			return
		}
	}
	if sc.needsFrameFlush {
		sc.startFrameWrite(frameWriteMsg{write: flushFrameWriter{}})
		sc.needsFrameFlush = false // after startFrameWrite, since it sets this true
		return
	}
}

func (sc *serverConn) goAway(code ErrCode) {
	sc.serveG.check()
	if sc.inGoAway {
		return
	}
	if code != ErrCodeNo {
		sc.shutDownIn(250 * time.Millisecond)
	} else {
		// TODO: configurable
		sc.shutDownIn(1 * time.Second)
	}
	sc.inGoAway = true
	sc.needToSendGoAway = true
	sc.goAwayCode = code
	sc.scheduleFrameWrite()
}

func (sc *serverConn) shutDownIn(d time.Duration) {
	sc.serveG.check()
	sc.shutdownTimer = time.NewTimer(d)
	sc.shutdownTimerCh = sc.shutdownTimer.C
}

func (sc *serverConn) resetStream(se StreamError) {
	sc.serveG.check()
	sc.writeFrame(frameWriteMsg{write: se})
	if st, ok := sc.streams[se.StreamID]; ok {
		st.sentReset = true
		sc.closeStream(st, se)
	}
}

// curHeaderStreamID returns the stream ID of the header block we're
// currently in the middle of reading. If this returns non-zero, the
// next frame must be a CONTINUATION with this stream id.
func (sc *serverConn) curHeaderStreamID() uint32 {
	sc.serveG.check()
	st := sc.req.stream
	if st == nil {
		return 0
	}
	return st.id
}

// processFrameFromReader processes the serve loop's read from readFrameCh from the
// frame-reading goroutine.
// processFrameFromReader returns whether the connection should be kept open.
func (sc *serverConn) processFrameFromReader(res readFrameResult) bool {
	sc.serveG.check()
	err := res.err
	if err != nil {
		if err == ErrFrameTooLarge {
			sc.goAway(ErrCodeFrameSize)
			return true // goAway will close the loop
		}
		clientGone := err == io.EOF || strings.Contains(err.Error(), "use of closed network connection")
		if clientGone {
			// TODO: could we also get into this state if
			// the peer does a half close
			// (e.g. CloseWrite) because they're done
			// sending frames but they're still wanting
			// our open replies?  Investigate.
			// TODO: add CloseWrite to crypto/tls.Conn first
			// so we have a way to test this? I suppose
			// just for testing we could have a non-TLS mode.
			return false
		}
	} else {
		f := res.f
		sc.vlogf("got %v: %#v", f.Header(), f)
		err = sc.processFrame(f)
		if err == nil {
			return true
		}
	}

	switch ev := err.(type) {
	case StreamError:
		sc.resetStream(ev)
		return true
	case goAwayFlowError:
		sc.goAway(ErrCodeFlowControl)
		return true
	case ConnectionError:
		sc.logf("%v: %v", sc.conn.RemoteAddr(), ev)
		sc.goAway(ErrCode(ev))
		return true // goAway will handle shutdown
	default:
		if res.err != nil {
			sc.logf("disconnecting; error reading frame from client %s: %v", sc.conn.RemoteAddr(), err)
		} else {
			sc.logf("disconnection due to other error: %v", err)
		}
		return false
	}
}

func (sc *serverConn) processFrame(f Frame) error {
	sc.serveG.check()

	// First frame received must be SETTINGS.
	if !sc.sawFirstSettings {
		if _, ok := f.(*SettingsFrame); !ok {
			return ConnectionError(ErrCodeProtocol)
		}
		sc.sawFirstSettings = true
	}

	if s := sc.curHeaderStreamID(); s != 0 {
		if cf, ok := f.(*ContinuationFrame); !ok {
			return ConnectionError(ErrCodeProtocol)
		} else if cf.Header().StreamID != s {
			return ConnectionError(ErrCodeProtocol)
		}
	}

	switch f := f.(type) {
	case *SettingsFrame:
		return sc.processSettings(f)
	case *HeadersFrame:
		return sc.processHeaders(f)
	case *ContinuationFrame:
		return sc.processContinuation(f)
	case *WindowUpdateFrame:
		return sc.processWindowUpdate(f)
	case *PingFrame:
		return sc.processPing(f)
	case *DataFrame:
		return sc.processData(f)
	case *RSTStreamFrame:
		return sc.processResetStream(f)
	case *PriorityFrame:
		return sc.processPriority(f)
	case *PushPromiseFrame:
		// A client cannot push. Thus, servers MUST treat the receipt of a PUSH_PROMISE
		// frame as a connection error (Section 5.4.1) of type PROTOCOL_ERROR.
		return ConnectionError(ErrCodeProtocol)
	default:
		sc.vlogf("Ignoring frame: %v", f.Header())
		return nil
	}
}

func (sc *serverConn) processPing(f *PingFrame) error {
	sc.serveG.check()
	if f.Flags.Has(FlagSettingsAck) {
		// 6.7 PING: " An endpoint MUST NOT respond to PING frames
		// containing this flag."
		return nil
	}
	if f.StreamID != 0 {
		// "PING frames are not associated with any individual
		// stream. If a PING frame is received with a stream
		// identifier field value other than 0x0, the recipient MUST
		// respond with a connection error (Section 5.4.1) of type
		// PROTOCOL_ERROR."
		return ConnectionError(ErrCodeProtocol)
	}
	sc.writeFrame(frameWriteMsg{write: writePingAck{f}})
	return nil
}

func (sc *serverConn) processWindowUpdate(f *WindowUpdateFrame) error {
	sc.serveG.check()
	switch {
	case f.StreamID != 0: // stream-level flow control
		st := sc.streams[f.StreamID]
		if st == nil {
			// "WINDOW_UPDATE can be sent by a peer that has sent a
			// frame bearing the END_STREAM flag. This means that a
			// receiver could receive a WINDOW_UPDATE frame on a "half
			// closed (remote)" or "closed" stream. A receiver MUST
			// NOT treat this as an error, see Section 5.1."
			return nil
		}
		if !st.flow.add(int32(f.Increment)) {
			return StreamError{f.StreamID, ErrCodeFlowControl}
		}
	default: // connection-level flow control
		if !sc.flow.add(int32(f.Increment)) {
			return goAwayFlowError{}
		}
	}
	sc.scheduleFrameWrite()
	return nil
}

func (sc *serverConn) processResetStream(f *RSTStreamFrame) error {
	sc.serveG.check()

	state, st := sc.state(f.StreamID)
	if state == stateIdle {
		// 6.4 "RST_STREAM frames MUST NOT be sent for a
		// stream in the "idle" state. If a RST_STREAM frame
		// identifying an idle stream is received, the
		// recipient MUST treat this as a connection error
		// (Section 5.4.1) of type PROTOCOL_ERROR.
		return ConnectionError(ErrCodeProtocol)
	}
	if st != nil {
		st.gotReset = true
		sc.closeStream(st, StreamError{f.StreamID, f.ErrCode})
	}
	return nil
}

func (sc *serverConn) closeStream(st *stream, err error) {
	sc.serveG.check()
	if st.state == stateIdle || st.state == stateClosed {
		panic(fmt.Sprintf("invariant; can't close stream in state %v", st.state))
	}
	st.state = stateClosed
	sc.curOpenStreams--
	delete(sc.streams, st.id)
	if p := st.body; p != nil {
		p.Close(err)
	}
	st.cw.Close() // signals Handler's CloseNotifier, unblocks writes, etc
	sc.writeSched.forgetStream(st.id)
}

func (sc *serverConn) processSettings(f *SettingsFrame) error {
	sc.serveG.check()
	if f.IsAck() {
		sc.unackedSettings--
		if sc.unackedSettings < 0 {
			// Why is the peer ACKing settings we never sent?
			// The spec doesn't mention this case, but
			// hang up on them anyway.
			return ConnectionError(ErrCodeProtocol)
		}
		return nil
	}
	if err := f.ForeachSetting(sc.processSetting); err != nil {
		return err
	}
	sc.needToSendSettingsAck = true
	sc.scheduleFrameWrite()
	return nil
}

func (sc *serverConn) processSetting(s Setting) error {
	sc.serveG.check()
	if err := s.Valid(); err != nil {
		return err
	}
	sc.vlogf("processing setting %v", s)
	switch s.ID {
	case SettingHeaderTableSize:
		sc.headerTableSize = s.Val
		sc.hpackEncoder.SetMaxDynamicTableSize(s.Val)
	case SettingEnablePush:
		sc.pushEnabled = s.Val != 0
	case SettingMaxConcurrentStreams:
		sc.clientMaxStreams = s.Val
	case SettingInitialWindowSize:
		return sc.processSettingInitialWindowSize(s.Val)
	case SettingMaxFrameSize:
		sc.writeSched.maxFrameSize = s.Val
	case SettingMaxHeaderListSize:
		sc.peerMaxHeaderListSize = s.Val
	default:
		// Unknown setting: "An endpoint that receives a SETTINGS
		// frame with any unknown or unsupported identifier MUST
		// ignore that setting."
	}
	return nil
}

func (sc *serverConn) processSettingInitialWindowSize(val uint32) error {
	sc.serveG.check()
	// Note: val already validated to be within range by
	// processSetting's Valid call.

	// "A SETTINGS frame can alter the initial flow control window
	// size for all current streams. When the value of
	// SETTINGS_INITIAL_WINDOW_SIZE changes, a receiver MUST
	// adjust the size of all stream flow control windows that it
	// maintains by the difference between the new value and the
	// old value."
	old := sc.initialWindowSize
	sc.initialWindowSize = int32(val)
	growth := sc.initialWindowSize - old // may be negative
	for _, st := range sc.streams {
		if !st.flow.add(growth) {
			// 6.9.2 Initial Flow Control Window Size
			// "An endpoint MUST treat a change to
			// SETTINGS_INITIAL_WINDOW_SIZE that causes any flow
			// control window to exceed the maximum size as a
			// connection error (Section 5.4.1) of type
			// FLOW_CONTROL_ERROR."
			return ConnectionError(ErrCodeFlowControl)
		}
	}
	return nil
}

func (sc *serverConn) processData(f *DataFrame) error {
	sc.serveG.check()
	// "If a DATA frame is received whose stream is not in "open"
	// or "half closed (local)" state, the recipient MUST respond
	// with a stream error (Section 5.4.2) of type STREAM_CLOSED."
	id := f.Header().StreamID
	st, ok := sc.streams[id]
	if !ok || st.state != stateOpen {
		// This includes sending a RST_STREAM if the stream is
		// in stateHalfClosedLocal (which currently means that
		// the http.Handler returned, so it's done reading &
		// done writing). Try to stop the client from sending
		// more DATA.
		return StreamError{id, ErrCodeStreamClosed}
	}
	if st.body == nil {
		panic("internal error: should have a body in this state")
	}
	data := f.Data()

	// Sender sending more than they'd declared?
	if st.declBodyBytes != -1 && st.bodyBytes+int64(len(data)) > st.declBodyBytes {
		st.body.Close(fmt.Errorf("sender tried to send more than declared Content-Length of %d bytes", st.declBodyBytes))
		return StreamError{id, ErrCodeStreamClosed}
	}
	if len(data) > 0 {
		// Check whether the client has flow control quota.
		if int(st.inflow.available()) < len(data) {
			return StreamError{id, ErrCodeFlowControl}
		}
		st.inflow.take(int32(len(data)))
		wrote, err := st.body.Write(data)
		if err != nil {
			return StreamError{id, ErrCodeStreamClosed}
		}
		if wrote != len(data) {
			panic("internal error: bad Writer")
		}
		st.bodyBytes += int64(len(data))
	}
	if f.StreamEnded() {
		if st.declBodyBytes != -1 && st.declBodyBytes != st.bodyBytes {
			st.body.Close(fmt.Errorf("request declared a Content-Length of %d but only wrote %d bytes",
				st.declBodyBytes, st.bodyBytes))
		} else {
			st.body.Close(io.EOF)
		}
		st.state = stateHalfClosedRemote
	}
	return nil
}

func (sc *serverConn) processHeaders(f *HeadersFrame) error {
	sc.serveG.check()
	id := f.Header().StreamID
	if sc.inGoAway {
		// Ignore.
		return nil
	}
	// http://http2.github.io/http2-spec/#rfc.section.5.1.1
	if id%2 != 1 || id <= sc.maxStreamID || sc.req.stream != nil {
		// Streams initiated by a client MUST use odd-numbered
		// stream identifiers. [...] The identifier of a newly
		// established stream MUST be numerically greater than all
		// streams that the initiating endpoint has opened or
		// reserved. [...]  An endpoint that receives an unexpected
		// stream identifier MUST respond with a connection error
		// (Section 5.4.1) of type PROTOCOL_ERROR.
		return ConnectionError(ErrCodeProtocol)
	}
	if id > sc.maxStreamID {
		sc.maxStreamID = id
	}
	st := &stream{
		id:    id,
		state: stateOpen,
	}
	if f.StreamEnded() {
		st.state = stateHalfClosedRemote
	}
	st.cw.Init()

	st.flow.conn = &sc.flow // link to conn-level counter
	st.flow.add(sc.initialWindowSize)
	st.inflow.conn = &sc.inflow      // link to conn-level counter
	st.inflow.add(initialWindowSize) // TODO: update this when we send a higher initial window size in the initial settings

	sc.streams[id] = st
	if f.HasPriority() {
		adjustStreamPriority(sc.streams, st.id, f.Priority)
	}
	sc.curOpenStreams++
	sc.req = requestParam{
		stream: st,
		header: make(http.Header),
	}
	sc.hpackDecoder.SetEmitEnabled(true)
	return sc.processHeaderBlockFragment(st, f.HeaderBlockFragment(), f.HeadersEnded())
}

func (sc *serverConn) processContinuation(f *ContinuationFrame) error {
	sc.serveG.check()
	st := sc.streams[f.Header().StreamID]
	if st == nil || sc.curHeaderStreamID() != st.id {
		return ConnectionError(ErrCodeProtocol)
	}
	return sc.processHeaderBlockFragment(st, f.HeaderBlockFragment(), f.HeadersEnded())
}

func (sc *serverConn) processHeaderBlockFragment(st *stream, frag []byte, end bool) error {
	sc.serveG.check()
	if _, err := sc.hpackDecoder.Write(frag); err != nil {
		// TODO: convert to stream error I assume?
		return err
	}
	if !end {
		return nil
	}
	if err := sc.hpackDecoder.Close(); err != nil {
		// TODO: convert to stream error I assume?
		return err
	}
	defer sc.resetPendingRequest()
	if sc.curOpenStreams > sc.advMaxStreams {
		// "Endpoints MUST NOT exceed the limit set by their
		// peer. An endpoint that receives a HEADERS frame
		// that causes their advertised concurrent stream
		// limit to be exceeded MUST treat this as a stream
		// error (Section 5.4.2) of type PROTOCOL_ERROR or
		// REFUSED_STREAM."
		if sc.unackedSettings == 0 {
			// They should know better.
			return StreamError{st.id, ErrCodeProtocol}
		}
		// Assume it's a network race, where they just haven't
		// received our last SETTINGS update. But actually
		// this can't happen yet, because we don't yet provide
		// a way for users to adjust server parameters at
		// runtime.
		return StreamError{st.id, ErrCodeRefusedStream}
	}

	rw, req, err := sc.newWriterAndRequest()
	if err != nil {
		return err
	}
	st.body = req.Body.(*requestBody).pipe // may be nil
	st.declBodyBytes = req.ContentLength

	handler := sc.handler.ServeHTTP
	if !sc.hpackDecoder.EmitEnabled() {
		// Their header list was too long. Send a 431 error.
		handler = handleHeaderListTooLong
	}

	go sc.runHandler(rw, req, handler)
	return nil
}

func (sc *serverConn) processPriority(f *PriorityFrame) error {
	adjustStreamPriority(sc.streams, f.StreamID, f.PriorityParam)
	return nil
}

func adjustStreamPriority(streams map[uint32]*stream, streamID uint32, priority PriorityParam) {
	st, ok := streams[streamID]
	if !ok {
		// TODO: not quite correct (this streamID might
		// already exist in the dep tree, but be closed), but
		// close enough for now.
		return
	}
	st.weight = priority.Weight
	parent := streams[priority.StreamDep] // might be nil
	if parent == st {
		// if client tries to set this stream to be the parent of itself
		// ignore and keep going
		return
	}

	// section 5.3.3: If a stream is made dependent on one of its
	// own dependencies, the formerly dependent stream is first
	// moved to be dependent on the reprioritized stream's previous
	// parent. The moved dependency retains its weight.
	for piter := parent; piter != nil; piter = piter.parent {
		if piter == st {
			parent.parent = st.parent
			break
		}
	}
	st.parent = parent
	if priority.Exclusive && (st.parent != nil || priority.StreamDep == 0) {
		for _, openStream := range streams {
			if openStream != st && openStream.parent == st.parent {
				openStream.parent = st
			}
		}
	}
}

// resetPendingRequest zeros out all state related to a HEADERS frame
// and its zero or more CONTINUATION frames sent to start a new
// request.
func (sc *serverConn) resetPendingRequest() {
	sc.serveG.check()
	sc.req = requestParam{}
}

func (sc *serverConn) newWriterAndRequest() (*responseWriter, *http.Request, error) {
	sc.serveG.check()
	rp := &sc.req
	if rp.invalidHeader || rp.method == "" || rp.path == "" ||
		(rp.scheme != "https" && rp.scheme != "http") {
		// See 8.1.2.6 Malformed Requests and Responses:
		//
		// Malformed requests or responses that are detected
		// MUST be treated as a stream error (Section 5.4.2)
		// of type PROTOCOL_ERROR."
		//
		// 8.1.2.3 Request Pseudo-Header Fields
		// "All HTTP/2 requests MUST include exactly one valid
		// value for the :method, :scheme, and :path
		// pseudo-header fields"
		return nil, nil, StreamError{rp.stream.id, ErrCodeProtocol}
	}
	var tlsState *tls.ConnectionState // nil if not scheme https
	if rp.scheme == "https" {
		tlsState = sc.tlsState
	}
	authority := rp.authority
	if authority == "" {
		authority = rp.header.Get("Host")
	}
	needsContinue := rp.header.Get("Expect") == "100-continue"
	if needsContinue {
		rp.header.Del("Expect")
	}
	// Merge Cookie headers into one "; "-delimited value.
	if cookies := rp.header["Cookie"]; len(cookies) > 1 {
		rp.header.Set("Cookie", strings.Join(cookies, "; "))
	}
	bodyOpen := rp.stream.state == stateOpen
	body := &requestBody{
		conn:          sc,
		stream:        rp.stream,
		needsContinue: needsContinue,
	}
	// TODO: handle asterisk '*' requests + test
	url, err := url.ParseRequestURI(rp.path)
	if err != nil {
		// TODO: find the right error code?
		return nil, nil, StreamError{rp.stream.id, ErrCodeProtocol}
	}
	req := &http.Request{
		Method:     rp.method,
		URL:        url,
		RemoteAddr: sc.remoteAddrStr,
		Header:     rp.header,
		RequestURI: rp.path,
		Proto:      "HTTP/2.0",
		ProtoMajor: 2,
		ProtoMinor: 0,
		TLS:        tlsState,
		Host:       authority,
		Body:       body,
	}
	if bodyOpen {
		body.pipe = &pipe{
			b: buffer{buf: make([]byte, initialWindowSize)}, // TODO: share/remove XXX
		}
		body.pipe.c.L = &body.pipe.m

		if vv, ok := rp.header["Content-Length"]; ok {
			req.ContentLength, _ = strconv.ParseInt(vv[0], 10, 64)
		} else {
			req.ContentLength = -1
		}
	}

	rws := responseWriterStatePool.Get().(*responseWriterState)
	bwSave := rws.bw
	*rws = responseWriterState{} // zero all the fields
	rws.conn = sc
	rws.bw = bwSave
	rws.bw.Reset(chunkWriter{rws})
	rws.stream = rp.stream
	rws.req = req
	rws.body = body

	rw := &responseWriter{rws: rws}
	return rw, req, nil
}

// Run on its own goroutine.
func (sc *serverConn) runHandler(rw *responseWriter, req *http.Request, handler func(http.ResponseWriter, *http.Request)) {
	defer rw.handlerDone()
	// TODO: catch panics like net/http.Server
	handler(rw, req)
}

func handleHeaderListTooLong(w http.ResponseWriter, r *http.Request) {
	// 10.5.1 Limits on Header Block Size:
	// .. "A server that receives a larger header block than it is
	// willing to handle can send an HTTP 431 (Request Header Fields Too
	// Large) status code"
	const statusRequestHeaderFieldsTooLarge = 431 // only in Go 1.6+
	w.WriteHeader(statusRequestHeaderFieldsTooLarge)
	io.WriteString(w, "<h1>HTTP Error 431</h1><p>Request Header Field(s) Too Large</p>")
}

// called from handler goroutines.
// h may be nil.
func (sc *serverConn) writeHeaders(st *stream, headerData *writeResHeaders) {
	sc.serveG.checkNotOn() // NOT on
	var errc chan error
	if headerData.h != nil {
		// If there's a header map (which we don't own), so we have to block on
		// waiting for this frame to be written, so an http.Flush mid-handler
		// writes out the correct value of keys, before a handler later potentially
		// mutates it.
		errc = errChanPool.Get().(chan error)
	}
	sc.writeFrameFromHandler(frameWriteMsg{
		write:  headerData,
		stream: st,
		done:   errc,
	})
	if errc != nil {
		select {
		case <-errc:
			// Ignore. Just for synchronization.
			// Any error will be handled in the writing goroutine.
			errChanPool.Put(errc)
		case <-sc.doneServing:
			// Client has closed the connection.
		case <-st.cw:
			// Client did RST_STREAM, etc. (but conn still alive)
		}
	}
}

// called from handler goroutines.
func (sc *serverConn) write100ContinueHeaders(st *stream) {
	sc.writeFrameFromHandler(frameWriteMsg{
		write:  write100ContinueHeadersFrame{st.id},
		stream: st,
	})
}

// A bodyReadMsg tells the server loop that the http.Handler read n
// bytes of the DATA from the client on the given stream.
type bodyReadMsg struct {
	st *stream
	n  int
}

// called from handler goroutines.
// Notes that the handler for the given stream ID read n bytes of its body
// and schedules flow control tokens to be sent.
func (sc *serverConn) noteBodyReadFromHandler(st *stream, n int) {
	sc.serveG.checkNotOn() // NOT on
	sc.bodyReadCh <- bodyReadMsg{st, n}
}

func (sc *serverConn) noteBodyRead(st *stream, n int) {
	sc.serveG.check()
	sc.sendWindowUpdate(nil, n) // conn-level
	if st.state != stateHalfClosedRemote && st.state != stateClosed {
		// Don't send this WINDOW_UPDATE if the stream is closed
		// remotely.
		sc.sendWindowUpdate(st, n)
	}
}

// st may be nil for conn-level
func (sc *serverConn) sendWindowUpdate(st *stream, n int) {
	sc.serveG.check()
	// "The legal range for the increment to the flow control
	// window is 1 to 2^31-1 (2,147,483,647) octets."
	// A Go Read call on 64-bit machines could in theory read
	// a larger Read than this. Very unlikely, but we handle it here
	// rather than elsewhere for now.
	const maxUint31 = 1<<31 - 1
	for n >= maxUint31 {
		sc.sendWindowUpdate32(st, maxUint31)
		n -= maxUint31
	}
	sc.sendWindowUpdate32(st, int32(n))
}

// st may be nil for conn-level
func (sc *serverConn) sendWindowUpdate32(st *stream, n int32) {
	sc.serveG.check()
	if n == 0 {
		return
	}
	if n < 0 {
		panic("negative update")
	}
	var streamID uint32
	if st != nil {
		streamID = st.id
	}
	sc.writeFrame(frameWriteMsg{
		write:  writeWindowUpdate{streamID: streamID, n: uint32(n)},
		stream: st,
	})
	var ok bool
	if st == nil {
		ok = sc.inflow.add(n)
	} else {
		ok = st.inflow.add(n)
	}
	if !ok {
		panic("internal error; sent too many window updates without decrements?")
	}
}

type requestBody struct {
	stream        *stream
	conn          *serverConn
	closed        bool
	pipe          *pipe // non-nil if we have a HTTP entity message body
	needsContinue bool  // need to send a 100-continue
}

func (b *requestBody) Close() error {
	if b.pipe != nil {
		b.pipe.Close(errClosedBody)
	}
	b.closed = true
	return nil
}

func (b *requestBody) Read(p []byte) (n int, err error) {
	if b.needsContinue {
		b.needsContinue = false
		b.conn.write100ContinueHeaders(b.stream)
	}
	if b.pipe == nil {
		return 0, io.EOF
	}
	n, err = b.pipe.Read(p)
	if n > 0 {
		b.conn.noteBodyReadFromHandler(b.stream, n)
	}
	return
}

// responseWriter is the http.ResponseWriter implementation.  It's
// intentionally small (1 pointer wide) to minimize garbage.  The
// responseWriterState pointer inside is zeroed at the end of a
// request (in handlerDone) and calls on the responseWriter thereafter
// simply crash (caller's mistake), but the much larger responseWriterState
// and buffers are reused between multiple requests.
type responseWriter struct {
	rws *responseWriterState
}

// Optional http.ResponseWriter interfaces implemented.
var (
	_ http.CloseNotifier = (*responseWriter)(nil)
	_ http.Flusher       = (*responseWriter)(nil)
	_ stringWriter       = (*responseWriter)(nil)
)

type responseWriterState struct {
	// immutable within a request:
	stream *stream
	req    *http.Request
	body   *requestBody // to close at end of request, if DATA frames didn't
	conn   *serverConn

	// TODO: adjust buffer writing sizes based on server config, frame size updates from peer, etc
	bw *bufio.Writer // writing to a chunkWriter{this *responseWriterState}

	// mutated by http.Handler goroutine:
	handlerHeader http.Header // nil until called
	snapHeader    http.Header // snapshot of handlerHeader at WriteHeader time
	status        int         // status code passed to WriteHeader
	wroteHeader   bool        // WriteHeader called (explicitly or implicitly). Not necessarily sent to user yet.
	sentHeader    bool        // have we sent the header frame?
	handlerDone   bool        // handler has finished
	curWrite      writeData

	closeNotifierMu sync.Mutex // guards closeNotifierCh
	closeNotifierCh chan bool  // nil until first used
}

type chunkWriter struct{ rws *responseWriterState }

func (cw chunkWriter) Write(p []byte) (n int, err error) { return cw.rws.writeChunk(p) }

// writeChunk writes chunks from the bufio.Writer. But because
// bufio.Writer may bypass its chunking, sometimes p may be
// arbitrarily large.
//
// writeChunk is also responsible (on the first chunk) for sending the
// HEADER response.
func (rws *responseWriterState) writeChunk(p []byte) (n int, err error) {
	if !rws.wroteHeader {
		rws.writeHeader(200)
	}
	if !rws.sentHeader {
		rws.sentHeader = true
		var ctype, clen string // implicit ones, if we can calculate it
		if rws.handlerDone && rws.snapHeader.Get("Content-Length") == "" {
			clen = strconv.Itoa(len(p))
		}
		if rws.snapHeader.Get("Content-Type") == "" {
			ctype = http.DetectContentType(p)
		}
		endStream := rws.handlerDone && len(p) == 0
		rws.conn.writeHeaders(rws.stream, &writeResHeaders{
			streamID:      rws.stream.id,
			httpResCode:   rws.status,
			h:             rws.snapHeader,
			endStream:     endStream,
			contentType:   ctype,
			contentLength: clen,
		})
		if endStream {
			return 0, nil
		}
	}
	if len(p) == 0 && !rws.handlerDone {
		return 0, nil
	}
	curWrite := &rws.curWrite
	curWrite.streamID = rws.stream.id
	curWrite.p = p
	curWrite.endStream = rws.handlerDone
	if err := rws.conn.writeDataFromHandler(rws.stream, curWrite); err != nil {
		return 0, err
	}
	return len(p), nil
}

func (w *responseWriter) Flush() {
	rws := w.rws
	if rws == nil {
		panic("Header called after Handler finished")
	}
	if rws.bw.Buffered() > 0 {
		if err := rws.bw.Flush(); err != nil {
			// Ignore the error. The frame writer already knows.
			return
		}
	} else {
		// The bufio.Writer won't call chunkWriter.Write
		// (writeChunk with zero bytes, so we have to do it
		// ourselves to force the HTTP response header and/or
		// final DATA frame (with END_STREAM) to be sent.
		rws.writeChunk(nil)
	}
}

func (w *responseWriter) CloseNotify() <-chan bool {
	rws := w.rws
	if rws == nil {
		panic("CloseNotify called after Handler finished")
	}
	rws.closeNotifierMu.Lock()
	ch := rws.closeNotifierCh
	if ch == nil {
		ch = make(chan bool, 1)
		rws.closeNotifierCh = ch
		go func() {
			rws.stream.cw.Wait() // wait for close
			ch <- true
		}()
	}
	rws.closeNotifierMu.Unlock()
	return ch
}

func (w *responseWriter) Header() http.Header {
	rws := w.rws
	if rws == nil {
		panic("Header called after Handler finished")
	}
	if rws.handlerHeader == nil {
		rws.handlerHeader = make(http.Header)
	}
	return rws.handlerHeader
}

func (w *responseWriter) WriteHeader(code int) {
	rws := w.rws
	if rws == nil {
		panic("WriteHeader called after Handler finished")
	}
	rws.writeHeader(code)
}

func (rws *responseWriterState) writeHeader(code int) {
	if !rws.wroteHeader {
		rws.wroteHeader = true
		rws.status = code
		if len(rws.handlerHeader) > 0 {
			rws.snapHeader = cloneHeader(rws.handlerHeader)
		}
	}
}

func cloneHeader(h http.Header) http.Header {
	h2 := make(http.Header, len(h))
	for k, vv := range h {
		vv2 := make([]string, len(vv))
		copy(vv2, vv)
		h2[k] = vv2
	}
	return h2
}

// The Life Of A Write is like this:
//
// * Handler calls w.Write or w.WriteString ->
// * -> rws.bw (*bufio.Writer) ->
// * (Handler migth call Flush)
// * -> chunkWriter{rws}
// * -> responseWriterState.writeChunk(p []byte)
// * -> responseWriterState.writeChunk (most of the magic; see comment there)
func (w *responseWriter) Write(p []byte) (n int, err error) {
	return w.write(len(p), p, "")
}

func (w *responseWriter) WriteString(s string) (n int, err error) {
	return w.write(len(s), nil, s)
}

// either dataB or dataS is non-zero.
func (w *responseWriter) write(lenData int, dataB []byte, dataS string) (n int, err error) {
	rws := w.rws
	if rws == nil {
		panic("Write called after Handler finished")
	}
	if !rws.wroteHeader {
		w.WriteHeader(200)
	}
	if dataB != nil {
		return rws.bw.Write(dataB)
	} else {
		return rws.bw.WriteString(dataS)
	}
}

func (w *responseWriter) handlerDone() {
	rws := w.rws
	if rws == nil {
		panic("handlerDone called twice")
	}
	rws.handlerDone = true
	w.Flush()
	w.rws = nil
	responseWriterStatePool.Put(rws)
}