diff --git a/src/counters.c b/src/counters.c
index 45ecd2d0c0..b2ef102c03 100644
--- a/src/counters.c
+++ b/src/counters.c
@@ -471,6 +471,10 @@ static void *SCPerfMgmtThread(void *arg)
         }
     }
 
+    while (!TmThreadsCheckFlag(tv_local, THV_DEINIT)) {
+        usleep(100);
+    }
+
     TmThreadsSetFlag(tv_local, THV_CLOSED);
     return NULL;
 }
@@ -535,11 +539,15 @@ static void *SCPerfWakeupThread(void *arg)
         }
 
         if (TmThreadsCheckFlag(tv_local, THV_KILL)) {
-            TmThreadsSetFlag(tv_local, THV_CLOSED);
             run = 0;
         }
     }
 
+    while (!TmThreadsCheckFlag(tv_local, THV_DEINIT)) {
+        usleep(100);
+    }
+
+    TmThreadsSetFlag(tv_local, THV_CLOSED);
     return NULL;
 }
 
diff --git a/src/cuda-packet-batcher.c b/src/cuda-packet-batcher.c
index 8eb98f6b10..cf8f9ba2b4 100644
--- a/src/cuda-packet-batcher.c
+++ b/src/cuda-packet-batcher.c
@@ -372,6 +372,10 @@ void *SCCudaPBTmThreadsSlot1(void *td)
         }
     }
 
+    while (!TmThreadsCheckFlag(tv, THV_DEINIT)) {
+        usleep(100);
+    }
+
     if (s->SlotThreadExitPrintStats != NULL) {
         s->SlotThreadExitPrintStats(tv, s->slot_data);
     }
diff --git a/src/flow.c b/src/flow.c
index 8d63abdf18..a63d921593 100644
--- a/src/flow.c
+++ b/src/flow.c
@@ -1486,6 +1486,7 @@ void FlowForceReassembly(void)
     }
 
     TmThreadsSetFlag(tv, THV_KILL);
+    TmThreadsSetFlag(tv, THV_DEINIT);
 
     /* be sure it has shut down */
     while (!TmThreadsCheckFlag(tv, THV_CLOSED)) {
@@ -1737,6 +1738,10 @@ void *FlowManagerThread(void *td)
         }
     }
 
+    while (!TmThreadsCheckFlag(th_v, THV_DEINIT)) {
+        usleep(100);
+    }
+
     FlowHashDebugDeinit();
 
     SCLogInfo("%" PRIu32 " new flows, %" PRIu32 " established flows were "
diff --git a/src/threadvars.h b/src/threadvars.h
index 87da5ccfeb..7b3546fae1 100644
--- a/src/threadvars.h
+++ b/src/threadvars.h
@@ -39,6 +39,9 @@ struct TmSlot_;
 #define THV_KILL      0x08 /** thread has been asked to cleanup and exit */
 #define THV_FAILED    0x10 /** thread has encountered an error and failed */
 #define THV_CLOSED    0x20 /** thread done, should be joinable */
+/* used to indicate the thread is going through de-init.  Introduced as more
+ * of a hack for solving stream-timeout-shutdown.  Is set by the main thread. */
+#define THV_DEINIT    0x40
 
 /** Thread flags set and read by threads, to control the threads, when they
  *  encounter certain conditions like failure */
diff --git a/src/tm-threads.c b/src/tm-threads.c
index 6bf69bd47a..da9e9c19b5 100644
--- a/src/tm-threads.c
+++ b/src/tm-threads.c
@@ -177,9 +177,13 @@ void *TmThreadsSlot1NoIn(void *td)
         }
     } /* while (run) */
 
+    while (!TmThreadsCheckFlag(tv, THV_DEINIT)) {
+        usleep(100);
+    }
+
     /* wait for synchronization from master, if this TV has a synchronization
      * point set by this name */
-    TmThreadsMSSlaveHitSyncPt(tv, "ReceiveTMBeforeDeInit");
+    //TmThreadsMSSlaveHitSyncPt(tv, "ReceiveTMBeforeDeInit");
 
     if (s->SlotThreadExitPrintStats != NULL) {
         s->SlotThreadExitPrintStats(tv, s->slot_data);
@@ -251,9 +255,13 @@ void *TmThreadsSlot1NoOut(void *td)
         }
     } /* while (run) */
 
+    while (!TmThreadsCheckFlag(tv, THV_DEINIT)) {
+        usleep(100);
+    }
+
     /* wait for synchronization from master, if this TV has a synchronization
      * point set by this name */
-    TmThreadsMSSlaveHitSyncPt(tv, "ReceiveTMBeforeDeInit");
+    //TmThreadsMSSlaveHitSyncPt(tv, "ReceiveTMBeforeDeInit");
 
     if (s->SlotThreadExitPrintStats != NULL) {
         s->SlotThreadExitPrintStats(tv, s->slot_data);
@@ -320,9 +328,13 @@ void *TmThreadsSlot1NoInOut(void *td)
         }
     } /* while (run) */
 
+    while (!TmThreadsCheckFlag(tv, THV_DEINIT)) {
+        usleep(100);
+    }
+
     /* wait for synchronization from master, if this TV has a synchronization
      * point set by this name */
-    TmThreadsMSSlaveHitSyncPt(tv, "ReceiveTMBeforeDeInit");
+    //TmThreadsMSSlaveHitSyncPt(tv, "ReceiveTMBeforeDeInit");
 
     if (s->SlotThreadExitPrintStats != NULL) {
         s->SlotThreadExitPrintStats(tv, s->slot_data);
@@ -424,9 +436,13 @@ void *TmThreadsSlot1(void *td)
         }
     } /* while (run) */
 
+    while (!TmThreadsCheckFlag(tv, THV_DEINIT)) {
+        usleep(100);
+    }
+
     /* wait for synchronization from master, if this TV has a synchronization
      * point set by this name */
-    TmThreadsMSSlaveHitSyncPt(tv, "ReceiveTMBeforeDeInit");
+    //TmThreadsMSSlaveHitSyncPt(tv, "ReceiveTMBeforeDeInit");
 
     if (s->SlotThreadExitPrintStats != NULL) {
         s->SlotThreadExitPrintStats(tv, s->slot_data);
@@ -581,9 +597,13 @@ void *TmThreadsSlotPktAcqLoop(void *td) {
     }
     SCPerfUpdateCounterArray(tv->sc_perf_pca, &tv->sc_perf_pctx, 0);
 
+    while (!TmThreadsCheckFlag(tv, THV_DEINIT)) {
+        usleep(100);
+    }
+
     /* wait for synchronization from master, if this TV has a synchronization
      * point set by this name */
-    TmThreadsMSSlaveHitSyncPt(tv, "ReceiveTMBeforeDeInit");
+    //TmThreadsMSSlaveHitSyncPt(tv, "ReceiveTMBeforeDeInit");
 
     for (slot = s; slot != NULL; slot = slot->slot_next) {
         if (slot->SlotThreadExitPrintStats != NULL) {
@@ -705,9 +725,13 @@ void *TmThreadsSlotVar(void *td)
     } /* while (run) */
     SCPerfUpdateCounterArray(tv->sc_perf_pca, &tv->sc_perf_pctx, 0);
 
+    while (!TmThreadsCheckFlag(tv, THV_DEINIT)) {
+        usleep(100);
+    }
+
     /* wait for synchronization from master, if this TV has a synchronization
      * point set by this name */
-    TmThreadsMSSlaveHitSyncPt(tv, "ReceiveTMBeforeDeInit");
+    //TmThreadsMSSlaveHitSyncPt(tv, "ReceiveTMBeforeDeInit");
 
     s = (TmSlot *)tv->tm_slots;
 
@@ -1376,6 +1400,7 @@ void TmThreadKillThread(ThreadVars *tv)
     /* set the thread flag informing the thread that it needs to be
      * terminated */
     TmThreadsSetFlag(tv, THV_KILL);
+    TmThreadsSetFlag(tv, THV_DEINIT);
 
     if (tv->inq != NULL) {
         /* signal the queue for the number of users */
@@ -1510,9 +1535,10 @@ void TmThreadKillThreads(void) {
             }
 
             TmThreadsSetFlag(tv, THV_KILL);
+            TmThreadsSetFlag(tv, THV_DEINIT);
             SCLogDebug("told thread %s to stop", tv->name);
 
-            TmThreadsMSMasterDisableSlaveAllSyncPts(tv);
+            //TmThreadsMSMasterDisableSlaveAllSyncPts(tv);
 
             if (tv->inq != NULL) {
                 int i;
diff --git a/src/util-mpm-b2g-cuda.c b/src/util-mpm-b2g-cuda.c
index d8c6e19f9a..83ec65c1a4 100644
--- a/src/util-mpm-b2g-cuda.c
+++ b/src/util-mpm-b2g-cuda.c
@@ -2269,6 +2269,10 @@ void *CudaMpmB2gThreadsSlot1(void *td)
         }
     }
 
+    while (!TmThreadsCheckFlag(tv, THV_DEINIT)) {
+        usleep(100);
+    }
+
     if (s->SlotThreadExitPrintStats != NULL) {
         s->SlotThreadExitPrintStats(tv, s->slot_data);
     }