From a72fdb9ca2a71c57ab3f0b3d7aec0608307eea04 Mon Sep 17 00:00:00 2001
From: Igor Kushnir <igorkuo@gmail.com>
Date: Fri, 12 Mar 2021 12:57:25 +0200
Subject: [PATCH] ConcurrentQueue::cancelPending: don't reset jobsLeft to 0

Worker threads may well be executing jobs while this function is being
called. If ConcurrentQueue::waitAll() is called soon enough after
cancelPending(), the worker threads may still be running, but waitAll()
would return immediately as jobsLeft would be nonpositive.

Subtracting _queue.size() from jobsLeft sets this variable to the number
of worker threads that are executing jobs at the moment.

ConcurrentQueueTest::cancelPending1UserThread() passes most of the time
now. But it still fails occasionally because it depends on the timing of
thread scheduling, which is unreliable.
---
 common/concurrent_queue.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/common/concurrent_queue.h b/common/concurrent_queue.h
index c983b277..300f783d 100644
--- a/common/concurrent_queue.h
+++ b/common/concurrent_queue.h
@@ -47,8 +47,8 @@ public:
     {
         std::unique_lock<std::mutex> lockQueue(queueMutex);
         std::unique_lock<std::mutex> lockJobsLeft(jobsLeftMutex);
-        _queue = std::queue<std::function<void(void)>>();
-        jobsLeft = 0;
+        jobsLeft -= _queue.size();
+        _queue = {};
     }
 
     void waitAll()