diff --git a/src/app-layer-detect-proto.c b/src/app-layer-detect-proto.c
index 2dc4008435..21281bc45c 100644
--- a/src/app-layer-detect-proto.c
+++ b/src/app-layer-detect-proto.c
@@ -304,6 +304,8 @@ void *AppLayerDetectProtoThread(void *td)
     StreamMsgQueueSetMinInitChunkLen(FLOW_PKT_TOSERVER, INSPECT_BYTES);
     StreamMsgQueueSetMinInitChunkLen(FLOW_PKT_TOCLIENT, INSPECT_BYTES);
 
+    tv->flags |= THV_INIT_DONE;
+
     /* main loop */
     while(run) {
         TmThreadTestThreadUnPaused(tv);
@@ -361,8 +363,10 @@ void *AppLayerDetectProtoThread(void *td)
             StreamMsgReturnToPool(smsg);
         }
 
-        if (tv->flags & THV_KILL)
+        if (tv->flags & THV_KILL) {
+            PerfUpdateCounterArray(tv->pca, &tv->pctx, 0);
             run = 0;
+        }
     }
 
     pthread_exit((void *) 0);
diff --git a/src/counters.c b/src/counters.c
index 009f06e47f..2baf35aafd 100644
--- a/src/counters.c
+++ b/src/counters.c
@@ -144,6 +144,7 @@ void * PerfMgmtThread(void *arg)
         return NULL;
     }
 
+    tv_local->flags |= THV_INIT_DONE;
     while (run) {
         TmThreadTestThreadUnPaused(tv_local);
 
@@ -183,6 +184,7 @@ void * PerfWakeupThread(void *arg)
 
     printf("PerfWakeupThread: spawned\n");
 
+    tv_local->flags |= THV_INIT_DONE;
     while (run) {
         TmThreadTestThreadUnPaused(tv_local);
 
diff --git a/src/eidps.c b/src/eidps.c
index 6fa677a768..5a32d89011 100644
--- a/src/eidps.c
+++ b/src/eidps.c
@@ -1594,12 +1594,15 @@ int main(int argc, char **argv)
     /* Spawn the L7 App Detect thread */
     AppLayerDetectProtoThreadSpawn();
 
-    /* Spawn the perf counter threads */
+    /* Spawn the perf counter threads.  Let these be the last one spawned */
     PerfSpawnThreads();
 
     /* Check if the alloted queues have at least 1 reader and writer */
     TmValidateQueueState();
 
+    /* Waits till all the threads have been initialized */
+    TmThreadWaitOnThreadInit();
+
     /* Un-pause all the paused threads */
     TmThreadContinueThreads();
 
diff --git a/src/flow.c b/src/flow.c
index e110090a78..21c9496cb4 100644
--- a/src/flow.c
+++ b/src/flow.c
@@ -495,6 +495,7 @@ void *FlowManagerThread(void *td)
 
     printf("%s started...\n", th_v->name);
 
+    th_v->flags |= THV_INIT_DONE;
     while (1)
     {
         TmThreadTestThreadUnPaused(th_v);
@@ -553,6 +554,7 @@ void *FlowManagerThread(void *td)
         }
 
         if (th_v->flags & THV_KILL) {
+            PerfUpdateCounterArray(th_v->pca, &th_v->pctx, 0);
             break;
         }
 
diff --git a/src/threadvars.h b/src/threadvars.h
index d8add8727a..d0d161aed3 100644
--- a/src/threadvars.h
+++ b/src/threadvars.h
@@ -8,11 +8,12 @@
 #include "counters.h"
 
 /** Thread flags set and read by threads to control the threads */
-#define THV_USE     0x01 /** thread is in use */
-#define THV_PAUSE   0x02 /** thread has been paused */
-#define THV_KILL    0x04 /** thread has been asked to cleanup and exit */
-#define THV_FAILED  0x08 /** thread has encountered an error and failed */
-#define THV_CLOSED  0x10 /** thread done, should be joinable */
+#define THV_USE       0x01 /** thread is in use */
+#define THV_INIT_DONE 0x02 /** thread initialization done */
+#define THV_PAUSE     0x04 /** thread has been paused */
+#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 */
 
 /** 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 8196600dfe..e05c170440 100644
--- a/src/tm-threads.c
+++ b/src/tm-threads.c
@@ -81,6 +81,7 @@ void *TmThreadsSlot1NoIn(void *td) {
     }
     memset(&s->s.slot_pq, 0, sizeof(PacketQueue));
 
+    tv->flags |= THV_INIT_DONE;
     while(run) {
         TmThreadTestThreadUnPaused(tv);
 
@@ -101,8 +102,10 @@ void *TmThreadsSlot1NoIn(void *td) {
 
         tv->tmqh_out(tv, p);
 
-        if (tv->flags & THV_KILL)
+        if (tv->flags & THV_KILL) {
+            PerfUpdateCounterArray(tv->pca, &tv->pctx, 0);
             run = 0;
+        }
     }
 
     if (s->s.SlotThreadExitPrintStats != NULL) {
@@ -142,6 +145,7 @@ void *TmThreadsSlot1NoOut(void *td) {
     }
     memset(&s->s.slot_pq, 0, sizeof(PacketQueue));
 
+    tv->flags |= THV_INIT_DONE;
     while(run) {
         TmThreadTestThreadUnPaused(tv);
 
@@ -156,8 +160,10 @@ void *TmThreadsSlot1NoOut(void *td) {
             break;
         }
 
-        if (tv->flags & THV_KILL)
+        if (tv->flags & THV_KILL) {
+            PerfUpdateCounterArray(tv->pca, &tv->pctx, 0);
             run = 0;
+        }
     }
 
     if (s->s.SlotThreadExitPrintStats != NULL) {
@@ -198,6 +204,7 @@ void *TmThreadsSlot1NoInOut(void *td) {
     }
     memset(&s->s.slot_pq, 0, sizeof(PacketQueue));
 
+    tv->flags |= THV_INIT_DONE;
     while(run) {
         TmThreadTestThreadUnPaused(tv);
 
@@ -212,6 +219,7 @@ void *TmThreadsSlot1NoInOut(void *td) {
 
         if (tv->flags & THV_KILL) {
             //printf("%s: TmThreadsSlot1NoInOut: KILL is set\n", tv->name);
+            PerfUpdateCounterArray(tv->pca, &tv->pctx, 0);
             run = 0;
         }
     }
@@ -256,6 +264,7 @@ void *TmThreadsSlot1(void *td) {
     }
     memset(&s->s.slot_pq, 0, sizeof(PacketQueue));
 
+    tv->flags |= THV_INIT_DONE;
     while(run) {
         TmThreadTestThreadUnPaused(tv);
 
@@ -289,6 +298,7 @@ void *TmThreadsSlot1(void *td) {
 
         if (tv->flags & THV_KILL) {
             //printf("%s: TmThreadsSlot1: KILL is set\n", tv->name);
+            PerfUpdateCounterArray(tv->pca, &tv->pctx, 0);
             run = 0;
         }
     }
@@ -375,6 +385,7 @@ void *TmThreadsSlotVar(void *td) {
         memset(&slot->slot_pq, 0, sizeof(PacketQueue));
     }
 
+    tv->flags |= THV_INIT_DONE;
     while(run) {
         TmThreadTestThreadUnPaused(tv);
 
@@ -400,6 +411,7 @@ void *TmThreadsSlotVar(void *td) {
 
         if (tv->flags & THV_KILL) {
             //printf("%s: TmThreadsSlot1: KILL is set\n", tv->name);
+            PerfUpdateCounterArray(tv->pca, &tv->pctx, 0);
             run = 0;
         }
     }
@@ -1033,3 +1045,24 @@ void TmThreadCheckThreadState(void)
 
     return;
 }
+
+/** \brief Used to check if all threads have finished their initialization.  On
+ *         finding an un-initialized thread, it waits till that thread completes
+ *         its initialization, before proceeding to the next thread.
+ */
+void TmThreadWaitOnThreadInit(void)
+{
+    ThreadVars *tv = NULL;
+    int i = 0;
+
+    for (i = 0; i < TVT_MAX; i++) {
+        tv = tv_root[i];
+        while (tv != NULL) {
+            while (!(tv->flags & THV_INIT_DONE))
+                ;
+            tv = tv->next;
+        }
+    }
+
+    return;
+}
diff --git a/src/tm-threads.h b/src/tm-threads.h
index b990847a81..6fb21e879c 100644
--- a/src/tm-threads.h
+++ b/src/tm-threads.h
@@ -50,5 +50,7 @@ void TmThreadPauseThreads(void);
 
 void TmThreadCheckThreadState(void);
 
+void TmThreadWaitOnThreadInit(void);
+
 #endif /* __TM_THREADS_H__ */