From d01e7698e39d0c08456cd3e850217f6e6551d516 Mon Sep 17 00:00:00 2001
From: "mergify[bot]" <37929162+mergify[bot]@users.noreply.github.com>
Date: Thu, 22 Jul 2021 11:14:07 -0500
Subject: [PATCH] Fix timer failure (#27006) (#27017)

Fixes #26205

Fixes deadlocks in heartbeat scheduler.

No changelog needed because this bug only exists in go 1.16+ and v7.14.0 is the first affected beats version

The underlying cause is a likely bug in golang, reported here by myself here: golang/go#47329

(cherry picked from commit 2f62e8fbd7cecd8c2cda3c76bad42136b54caafd)

Co-authored-by: Andrew Cholakian <andrew@andrewvc.com>
---
 heartbeat/scheduler/timerqueue/queue.go      | 19 +++++++----
 heartbeat/scheduler/timerqueue/queue_test.go | 33 +++++++++++++++++---
 2 files changed, 41 insertions(+), 11 deletions(-)

diff --git a/heartbeat/scheduler/timerqueue/queue.go b/heartbeat/scheduler/timerqueue/queue.go
index f545c43be8e..61d4e0ac933 100644
--- a/heartbeat/scheduler/timerqueue/queue.go
+++ b/heartbeat/scheduler/timerqueue/queue.go
@@ -88,7 +88,7 @@ func (tq *TimerQueue) Start() {
 				if tq.th.Len() > 0 {
 					nr := tq.th[0].runAt
 					tq.nextRunAt = &nr
-					tq.timer.Reset(nr.Sub(time.Now()))
+					tq.timer = time.NewTimer(time.Until(nr))
 				} else {
 					tq.timer.Stop()
 					tq.nextRunAt = nil
@@ -104,14 +104,21 @@ func (tq *TimerQueue) pushInternal(tt *timerTask) {
 	heap.Push(&tq.th, tt)
 
 	if tq.nextRunAt == nil || tq.nextRunAt.After(tt.runAt) {
-		// Stop and drain the timer prior to reset per https://golang.org/pkg/time/#Timer.Reset
-		// Only drain if nextRunAt is set, otherwise the timer channel has already been stopped the
-		// channel is empty (and thus would block)
 		if tq.nextRunAt != nil && !tq.timer.Stop() {
 			<-tq.timer.C
 		}
-		tq.timer.Reset(tt.runAt.Sub(time.Now()))
-
+		// Originally the line below this comment was
+		//
+		//  tq.timer.Reset(time.Until(tt.runAt))
+		//
+		// however this broke in go1.16rc1, specifically on the commit b4b014465216790e01aa66f9120d03230e4aff46
+		//, specifically on this line:
+		// https://github.com/golang/go/commit/b4b014465216790e01aa66f9120d03230e4aff46#diff-73699b6edfe5dbb3f6824e66bb3566bce9405e9a8c810cac55c8199459f0ac19R652
+		// where some nice new optimizations don't actually work reliably
+		// This can be worked around by instantiating a new timer rather than resetting the timer.
+		// since that internally calls deltimer in runtime/timer.go rather than modtimer,
+		// I suspect that the problem is in modtimer's setting of &pp.timerModifiedEarliest
+		tq.timer = time.NewTimer(time.Until(tt.runAt))
 		tq.nextRunAt = &tt.runAt
 	}
 }
diff --git a/heartbeat/scheduler/timerqueue/queue_test.go b/heartbeat/scheduler/timerqueue/queue_test.go
index 2e5a2e59eb8..26235390bb4 100644
--- a/heartbeat/scheduler/timerqueue/queue_test.go
+++ b/heartbeat/scheduler/timerqueue/queue_test.go
@@ -20,6 +20,8 @@ package timerqueue
 import (
 	"context"
 	"math/rand"
+	"os"
+	"runtime/pprof"
 	"sort"
 	"testing"
 	"time"
@@ -27,11 +29,32 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
-func TestQueueRunsInOrder(t *testing.T) {
-	t.Skip("flaky test on windows: https://github.com/elastic/beats/issues/26205")
-	// Bugs can show up only occasionally
-	for i := 0; i < 100; i++ {
-		testQueueRunsInOrderOnce(t)
+func TestRunsInOrder(t *testing.T) {
+	testQueueRunsInOrderOnce(t)
+}
+
+// TestStress tries to figure out if we have any deadlocks that show up under concurrency
+func TestStress(t *testing.T) {
+	for i := 0; i < 120000; i++ {
+		failed := make(chan bool)
+		succeeded := make(chan bool)
+
+		watchdog := time.AfterFunc(time.Second*5, func() {
+			failed <- true
+		})
+
+		go func() {
+			testQueueRunsInOrderOnce(t)
+			succeeded <- true
+		}()
+
+		select {
+		case <-failed:
+			pprof.Lookup("goroutine").WriteTo(os.Stdout, 1)
+			require.FailNow(t, "Scheduler test iteration timed out, deadlock issue?")
+		case <-succeeded:
+			watchdog.Stop()
+		}
 	}
 }