diff --git a/src/core/thread/osthread.d b/src/core/thread/osthread.d
index 1470b41e1e..c4cab8acf1 100644
--- a/src/core/thread/osthread.d
+++ b/src/core/thread/osthread.d
@@ -3805,7 +3805,8 @@ private
     version (CRuntime_Microsoft)
         extern(C) extern __gshared ubyte msvcUsesUCRT; // from rt/msvc.c
 
-    package(core) __gshared bool thread_DLLProcessDetaching;
+    /// set during termination of a DLL on Windows, i.e. while executing DllMain(DLL_PROCESS_DETACH)
+    public __gshared bool thread_DLLProcessDetaching;
 
     __gshared HMODULE ll_dllModule;
     __gshared ThreadID ll_dllMonitorThread;
diff --git a/src/gc/impl/conservative/gc.d b/src/gc/impl/conservative/gc.d
index 2f1673b76b..e8607243a4 100644
--- a/src/gc/impl/conservative/gc.d
+++ b/src/gc/impl/conservative/gc.d
@@ -2597,6 +2597,15 @@ struct Gcx
         //printf("\tpool address range = %p .. %p\n", minAddr, maxAddr);
 
         {
+            version (COLLECT_PARALLEL)
+            {
+                bool doParallel = config.parallel > 0;
+                if (doParallel && !scanThreadData)
+                    startScanThreads();
+            }
+            else
+                enum doParallel = false;
+
             // lock roots and ranges around suspending threads b/c they're not reentrant safe
             rangesLock.lock();
             rootsLock.lock();
@@ -2615,11 +2624,6 @@ struct Gcx
             prepTime += (stop - start);
             start = stop;
 
-            version (COLLECT_PARALLEL)
-                bool doParallel = config.parallel > 0;
-            else
-                enum doParallel = false;
-
             if (doParallel)
             {
                 version (COLLECT_PARALLEL)
@@ -2744,9 +2748,6 @@ struct Gcx
         void** pbot = toscanRoots._p;
         void** ptop = toscanRoots._p + toscanRoots._length;
 
-        if (!scanThreadData)
-            startScanThreads();
-
         debug(PARALLEL_PRINTF) printf("markParallel\n");
 
         size_t pointersPerThread = toscanRoots._length / (numScanThreads + 1);
@@ -2869,8 +2870,12 @@ struct Gcx
             if (scanThreadData[idx].tid != scanThreadData[idx].tid.init)
                 startedThreads++;
 
+        version (Windows)
+            alias allThreadsDead = thread_DLLProcessDetaching;
+        else
+            enum allThreadsDead = false;
         stopGC = true;
-        while (atomicLoad(stoppedThreads) < startedThreads)
+        while (atomicLoad(stoppedThreads) < startedThreads && !allThreadsDead)
         {
             evStart.set();
             evDone.wait(dur!"msecs"(1));
diff --git a/test/gc/Makefile b/test/gc/Makefile
index 7a81cfd49f..935d046ce0 100644
--- a/test/gc/Makefile
+++ b/test/gc/Makefile
@@ -1,6 +1,6 @@
 include ../common.mak
 
-TESTS := sentinel printf memstomp invariant logging precise precisegc forkgc forkgc2 sigmaskgc
+TESTS := sentinel printf memstomp invariant logging precise precisegc forkgc forkgc2 sigmaskgc startbackgc
 
 SRC_GC = ../../src/gc/impl/conservative/gc.d
 SRC = $(SRC_GC) ../../src/rt/lifetime.d
@@ -46,5 +46,8 @@ $(ROOT)/forkgc2: forkgc2.d
 $(ROOT)/sigmaskgc: sigmaskgc.d
 	$(DMD) $(UDFLAGS) -of$@ sigmaskgc.d
 
+$(ROOT)/startbackgc: startbackgc.d
+	$(DMD) $(UDFLAGS) -of$@ sigmaskgc.d
+
 clean:
 	rm -rf $(ROOT)
diff --git a/test/gc/startbackgc.d b/test/gc/startbackgc.d
new file mode 100644
index 0000000000..bf51fbdbef
--- /dev/null
+++ b/test/gc/startbackgc.d
@@ -0,0 +1,22 @@
+// https://issues.dlang.org/show_bug.cgi?id=20270
+import core.sys.posix.sys.wait : waitpid;
+import core.sys.posix.unistd : fork, _exit;
+import core.thread : Thread;
+
+void main()
+{
+    foreach (t; 0 .. 10)
+        new Thread({
+            foreach (n; 0 .. 100)
+            {
+                foreach (x; 0 .. 100)
+                    new ubyte[x];
+                auto f = fork();
+                assert(f >= 0);
+                if (f == 0)
+                    _exit(0);
+                else
+                    waitpid(f, null, 0);
+            }
+        }).start();
+}