net: allow more generous slop in Fluctuation tests

It appears that at least the OpenBSD kernel gets sloppier the longer
the timeout we give it, up to an observed overhead of around 25%.
Let's give it a little more than that (33%) in the comparison, and
also increase the growth curve to match the actual observed times
instead of exponential initial growth.

Fixes #36108

Change-Id: Id3e54559b7c45b7c8bc0ca07dce74ca60e77e7ed
Reviewed-on: https://go-review.googlesource.com/c/go/+/366176
Trust: Bryan C. Mills <bcmills@google.com>
Run-TryBot: Bryan C. Mills <bcmills@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>

1 files changed, 46 insertions, 15 deletions

diff --git a/src/net/timeout_test.go b/src/net/timeout_test.go
index d345bf85ac..515aa07ec3 100644
--- a/src/net/timeout_test.go
+++ b/src/net/timeout_test.go
@@ -655,11 +655,37 @@ const (
 	// maxDynamicTimeout is the maximum timeout to attempt for
 	// tests that automatically increase timeouts until succeess.
 	//
-	// This should be a strict upper bound on the latency of the timeout: if a
-	// test would increase the timeout beyond this value, the test fails.
-	maxDynamicTimeout = 1 * time.Second
+	// This should be a strict upper bound on the latency required to hit a
+	// timeout accurately, even on a slow or heavily-loaded machine. If a test
+	// would increase the timeout beyond this value, the test fails.
+	maxDynamicTimeout = 4 * time.Second
 )
 
+// timeoutUpperBound returns the maximum time that we expect a timeout of
+// duration d to take to return the caller.
+func timeoutUpperBound(d time.Duration) time.Duration {
+	// In https://storage.googleapis.com/go-build-log/1e637cd3/openbsd-amd64-68_3585d3e7.log,
+	// we observed that an openbsd-amd64-68 builder took 636ms for a 512ms timeout
+	// (24.2% overhead).
+	return d * 4 / 3
+}
+
+// nextTimeout returns the next timeout to try after an operation took the given
+// actual duration with a timeout shorter than that duration.
+func nextTimeout(actual time.Duration) (next time.Duration, ok bool) {
+	if actual >= maxDynamicTimeout {
+		return maxDynamicTimeout, false
+	}
+	// Since the previous attempt took actual, we can't expect to beat that
+	// duration by any significant margin. Try the next attempt with an arbitrary
+	// factor above that, so that our growth curve is at least exponential.
+	next = actual * 5 / 4
+	if next > maxDynamicTimeout {
+		return maxDynamicTimeout, true
+	}
+	return next, true
+}
+
 func TestReadTimeoutFluctuation(t *testing.T) {
 	ln, err := newLocalListener("tcp")
 	if err != nil {
@@ -703,14 +729,15 @@ func TestReadTimeoutFluctuation(t *testing.T) {
 		if t.Failed() {
 			return
 		}
-		if actual > d*11/10 {
-			if actual > maxDynamicTimeout || d > maxDynamicTimeout/2 {
-				t.Fatalf("Read took %s; expected %v", actual, d)
+		if want := timeoutUpperBound(d); actual > want {
+			next, ok := nextTimeout(actual)
+			if !ok {
+				t.Fatalf("Read took %s; expected at most %v", actual, want)
 			}
 			// Maybe this machine is too slow to reliably schedule goroutines within
 			// the requested duration. Increase the timeout and try again.
 			t.Logf("Read took %s (expected %s); trying with longer timeout", actual, d)
-			d *= 2
+			d = next
 			continue
 		}
 
@@ -761,14 +788,15 @@ func TestReadFromTimeoutFluctuation(t *testing.T) {
 		if t.Failed() {
 			return
 		}
-		if actual > d*11/10 {
-			if actual > maxDynamicTimeout || d > maxDynamicTimeout/2 {
-				t.Fatalf("ReadFrom took %s; expected %s", actual, d)
+		if want := timeoutUpperBound(d); actual > want {
+			next, ok := nextTimeout(actual)
+			if !ok {
+				t.Fatalf("ReadFrom took %s; expected at most %s", actual, want)
 			}
 			// Maybe this machine is too slow to reliably schedule goroutines within
 			// the requested duration. Increase the timeout and try again.
 			t.Logf("ReadFrom took %s (expected %s); trying with longer timeout", actual, d)
-			d *= 2
+			d = next
 			continue
 		}
 
@@ -830,7 +858,7 @@ func TestWriteTimeoutFluctuation(t *testing.T) {
 		if t.Failed() {
 			return
 		}
-		if actual > d*11/10 {
+		if want := timeoutUpperBound(d); actual > want {
 			if n > 0 {
 				// SetWriteDeadline specifies a time “after which I/O operations fail
 				// instead of blocking”. However, the kernel's send buffer is not yet
@@ -838,15 +866,18 @@ func TestWriteTimeoutFluctuation(t *testing.T) {
 				// bytes to it without blocking.
 				t.Logf("Wrote %d bytes into send buffer; retrying until buffer is full", n)
 				if d <= maxDynamicTimeout/2 {
+					// We don't know how long the actual write loop would have taken if
+					// the buffer were full, so just guess and double the duration so that
+					// the next attempt can make twice as much progress toward filling it.
 					d *= 2
 				}
-			} else if actual > maxDynamicTimeout || d > maxDynamicTimeout/2 {
-				t.Fatalf("Write took %s; expected %s", actual, d)
+			} else if next, ok := nextTimeout(actual); !ok {
+				t.Fatalf("Write took %s; expected at most %s", actual, want)
 			} else {
 				// Maybe this machine is too slow to reliably schedule goroutines within
 				// the requested duration. Increase the timeout and try again.
 				t.Logf("Write took %s (expected %s); trying with longer timeout", actual, d)
-				d *= 2
+				d = next
 			}
 			continue
 		}