rproxy: cut off an answer that stalls after arriving before the whole request (closes #345)

When the app's answer arrived before the app had been sent the whole
request, SWRP_MAX_RESPONSE_TIME_MS had no timer: only the end of the
request started one. If the app then stalled, only a write could fail at
the deadline. With nothing written, the connection stayed open until the
request time limit.

modifyResponse now starts the response timer when it ends the request.
Once the timer cuts off the app's answer, nothing is left to flush, so the
last flush in finish succeeded and Go's server read the rest of the body.
finish now also ends that reading at once after a refusal of an answer
already passed on.

Model: opus-5-5

internal/rproxy/limits_test.go

@@ -12,7 +12,6 @@ import (
 	"os"
 	"strconv"
 	"strings"
-	"sync"
 	"sync/atomic"
 	"testing"
 	"time"
@@ -918,49 +917,48 @@ func TestAVisitorThatStopsALittleShortHasTheConnectionClosedWhenTheAnswerIsCutOf
 ) {
 	t.Parallel()
 
-	const limit = 300 * time.Millisecond
+	// A busy host passes on a part sent at once well within this limit.
+	const limit = 2 * time.Second
 
-	// With two parts, rproxy fails to send the first and then waits for the
-	// second, which the app sends only once rproxy has ended the request to it.
+	// The app sends the parts in atOnce with its status and headers. It sends the
+	// parts in late only once rproxy has ended the request to it, which rproxy
+	// does when it cuts off the answer, so that those never reach the visitor.
 	for _, tc := range []struct {
-		name  string
-		parts []string
+		name         string
+		atOnce, late []string
 	}{
-		{"with the body in one part", []string{greeting}},
-		{"with the body in two parts", []string{greeting[:5], greeting[5:]}},
+		{"with the body in one part", nil, []string{greeting}},
+		{"with the body in two parts", nil, []string{greeting[:5], greeting[5:]}},
+		{
+			"with the first part of the body at once",
+			[]string{greeting[:5]}, []string{greeting[5:]},
+		},
 	} {
 		t.Run(tc.name, func(t *testing.T) {
 			t.Parallel()
 
-			headersRead := make(chan struct{})
-			// letAppSendBody also runs when the test ends, as the app's server waits
-			// for the app.
-			letAppSendBody := sync.OnceFunc(func() { close(headersRead) })
 			standIn := startWAF(t, allow)
-			// The app sends its body once the visitor has the status and headers and
-			// the response time limit has run out since, so that rproxy cannot send
-			// it. After each part, it reads the request's body, which ends when rproxy
-			// ends the request to it.
 			app := startApp(t, func(w http.ResponseWriter, r *http.Request) {
 				rc := http.NewResponseController(w)
 				_ = rc.EnableFullDuplex()
 
 				w.Header().Set(contentLengthHeader, strconv.Itoa(len(greeting)))
 
+				for _, part := range tc.atOnce {
+					_, _ = io.WriteString(w, part)
+				}
+
 				_ = rc.Flush()
 
-				<-headersRead
-				time.Sleep(limit)
+				// The request's body ends when rproxy ends the request to the app.
+				_, _ = io.Copy(io.Discard, r.Body)
 
-				for _, part := range tc.parts {
+				for _, part := range tc.late {
 					_, _ = io.WriteString(w, part)
 					_ = rc.Flush()
-					_, _ = io.Copy(io.Discard, r.Body)
 				}
 			})
 
-			t.Cleanup(letAppSendBody)
-
 			// The request time limit is far longer than dial's deadline.
 			addr, _ := startRProxy(t, standIn.URL, app.URL, map[string]string{
 				maxRequestTime:  "120000",
@@ -975,15 +973,14 @@ func TestAVisitorThatStopsALittleShortHasTheConnectionClosedWhenTheAnswerIsCutOf
 				t.Fatalf("read the answer: %v", err)
 			}
 
-			letAppSendBody()
-
 			// If the connection stays open, dial's deadline ends this read with a
 			// timeout.
 			body, err := io.ReadAll(res.Body)
 			_ = res.Body.Close()
 
-			if len(body) != 0 || !errors.Is(err, io.ErrUnexpectedEOF) {
-				t.Errorf("read %q (%v), want the connection closed", body, err)
+			want := strings.Join(tc.atOnce, "")
+			if string(body) != want || !errors.Is(err, io.ErrUnexpectedEOF) {
+				t.Errorf("read %q (%v), want %q and the connection closed", body, err, want)
 			}
 		})
 	}

internal/rproxy/request.go

@@ -272,6 +272,10 @@ func (rq *request) modifyResponse(res *http.Response) error {
 
 	rq.closeAfterAnswer()
 
+	// An answer that arrives before the app has been sent the whole request ends
+	// the request, and SWRP_MAX_RESPONSE_TIME_MS starts.
+	rq.startResponseTimer(requestEnd)
+
 	res.Body = &responseBody{body: limitBody(res.Body, maxBytes), rq: rq}
 	rq.out.afterStatus = rq.setResponseDeadline
 
@@ -386,14 +390,17 @@ func (rq *request) finish(ctx context.Context) {
 
 	// After a cut-off answer, the reading of a body that has not arrived whole
 	// ends now. Go's server would otherwise keep the connection open while it
-	// reads the rest. A flush fails if the time to send the answer ran out, and
+	// reads the rest. The app's answer is cut off by a refusal once it is passed
+	// on, as afterStatus shows, since a refusal ends the request to the app. Any
+	// answer is cut off when the time to send it runs out: a flush then fails, and
 	// every flush after a failed one fails too. ReverseProxy ignores the failure,
 	// but panics if it was still waiting for more of the app's answer; finish runs
 	// either way. Before a status, a flush would send a 200; after a 101,
 	// ReverseProxy has taken over the connection.
 	switch {
-	case rq.out.status != 0 && rq.out.status != http.StatusSwitchingProtocols &&
-		errors.Is(rq.rc.Flush(), os.ErrDeadlineExceeded):
+	case rq.out.afterStatus != nil && rq.refused.Load() != 0,
+		rq.out.status != 0 && rq.out.status != http.StatusSwitchingProtocols &&
+			errors.Is(rq.rc.Flush(), os.ErrDeadlineExceeded):
 		rq.stopPassingOnBody(time.Now())
 	case rq.refused.Load() == 0:
 		rq.stopPassingOnBody(rq.requestDeadline())
@@ -553,19 +560,28 @@ func (rq *request) wroteRequest(info httptrace.WroteRequestInfo) {
 	requestEnd := rq.endRequest()
 
 	rq.mu.Lock()
-	defer rq.mu.Unlock()
+	stopTimer(rq.requestTimer)
+	rq.mu.Unlock()
 
-	if rq.timersStopped {
+	rq.startResponseTimer(requestEnd)
+}
+
+// startResponseTimer starts SWRP_MAX_RESPONSE_TIME_MS, which runs from
+// requestEnd, unless it has started already.
+func (rq *request) startResponseTimer(requestEnd time.Time) {
+	limit := rq.h.config.MaxResponseTime
+	if limit == 0 {
 		return
 	}
 
-	stopTimer(rq.requestTimer)
+	rq.mu.Lock()
+	defer rq.mu.Unlock()
 
-	limit := rq.h.config.MaxResponseTime
-	if limit > 0 {
-		rq.responseTimer = time.AfterFunc(time.Until(requestEnd.Add(limit)),
-			rq.responseTimedOut)
+	if rq.timersStopped || rq.responseTimer != nil {
+		return
 	}
+
+	rq.responseTimer = time.AfterFunc(time.Until(requestEnd.Add(limit)), rq.responseTimedOut)
 }
 
 // responseTimedOut is called when SWRP_MAX_RESPONSE_TIME_MS runs out before the