diff --git a/src/decode.h b/src/decode.h
index 111ab5fcf2..f685e2b268 100644
--- a/src/decode.h
+++ b/src/decode.h
@@ -382,10 +382,8 @@ typedef struct Packet_
     uint64_t pcap_cnt;
 
     /* ready to set verdict counter, only set in root */
-    uint8_t rtv_cnt;
-    /* tunnel packet ref count */
-    uint8_t tpr_cnt;
-    SCMutex mutex_rtv_cnt;
+    SC_ATOMIC_DECLARE(unsigned short, tunnel_rtv_cnt);
+    SC_ATOMIC_DECLARE(unsigned short, tunnel_tpr_cnt);
 
     /* decoder events */
     PacketDecoderEvents events;
@@ -506,14 +504,16 @@ typedef struct DecodeThreadVars_
 #ifndef __SC_CUDA_SUPPORT__
 #define PACKET_INITIALIZE(p) { \
     memset((p), 0x00, SIZE_OF_PACKET); \
-    SCMutexInit(&(p)->mutex_rtv_cnt, NULL); \
+    SC_ATOMIC_INIT(p->tunnel_rtv_cnt); \
+    SC_ATOMIC_INIT(p->tunnel_tpr_cnt); \
     PACKET_RESET_CHECKSUMS((p)); \
     (p)->pkt = ((uint8_t *)(p)) + sizeof(Packet); \
 }
 #else
 #define PACKET_INITIALIZE(p) { \
     memset((p), 0x00, SIZE_OF_PACKET); \
-    SCMutexInit(&(p)->mutex_rtv_cnt, NULL); \
+    SC_ATOMIC_INIT(p->tunnel_rtv_cnt); \
+    SC_ATOMIC_INIT(p->tunnel_tpr_cnt); \
     PACKET_RESET_CHECKSUMS((p)); \
     SCMutexInit(&(p)->cuda_mutex, NULL); \
     SCCondInit(&(p)->cuda_cond, NULL); \
@@ -576,10 +576,8 @@ typedef struct DecodeThreadVars_
         (p)->pktlen = 0;                        \
         (p)->alerts.cnt = 0;                    \
         (p)->pcap_cnt = 0;                      \
-        (p)->rtv_cnt = 0;                       \
-        (p)->tpr_cnt = 0;                       \
-        SCMutexDestroy(&(p)->mutex_rtv_cnt);    \
-        SCMutexInit(&(p)->mutex_rtv_cnt, NULL); \
+        SC_ATOMIC_RESET((p)->tunnel_rtv_cnt);   \
+        SC_ATOMIC_RESET((p)->tunnel_tpr_cnt);   \
         (p)->events.cnt = 0;                    \
         (p)->next = NULL;                       \
         (p)->prev = NULL;                       \
@@ -608,14 +606,16 @@ typedef struct DecodeThreadVars_
         if ((p)->pktvar != NULL) {              \
             PktVarFree((p)->pktvar);            \
         }                                       \
-        SCMutexDestroy(&(p)->mutex_rtv_cnt);    \
+        SC_ATOMIC_DESTROY((p)->tunnel_rtv_cnt); \
+        SC_ATOMIC_DESTROY((p)->tunnel_tpr_cnt); \
     } while (0)
 #else
 #define PACKET_CLEANUP(p) do {                  \
     if ((p)->pktvar != NULL) {                  \
         PktVarFree((p)->pktvar);                \
     }                                           \
-    SCMutexDestroy(&(p)->mutex_rtv_cnt);        \
+    SC_ATOMIC_DESTROY((p)->tunnel_rtv_cnt);     \
+    SC_ATOMIC_DESTROY((p)->tunnel_tpr_cnt);     \
     SCMutexDestroy(&(p)->cuda_mutex);           \
     SCCondDestroy(&(p)->cuda_cond);             \
 } while(0)
@@ -668,31 +668,27 @@ typedef struct DecodeThreadVars_
 } while (0)
 
 #define TUNNEL_INCR_PKT_RTV(p) do {                                                 \
-        SCMutexLock((p)->root ? &(p)->root->mutex_rtv_cnt : &(p)->mutex_rtv_cnt);   \
-        ((p)->root ? (p)->root->rtv_cnt++ : (p)->rtv_cnt++);                        \
-        SCMutexUnlock((p)->root ? &(p)->root->mutex_rtv_cnt : &(p)->mutex_rtv_cnt); \
+        ((p)->root ? SC_ATOMIC_ADD((p)->root->tunnel_rtv_cnt, 1) :                  \
+                     SC_ATOMIC_ADD((p)->tunnel_rtv_cnt, 1));                        \
     } while (0)
 
 #define TUNNEL_INCR_PKT_TPR(p) do {                                                 \
-        SCMutexLock((p)->root ? &(p)->root->mutex_rtv_cnt : &(p)->mutex_rtv_cnt);   \
-        ((p)->root ? (p)->root->tpr_cnt++ : (p)->tpr_cnt++);                        \
-        SCMutexUnlock((p)->root ? &(p)->root->mutex_rtv_cnt : &(p)->mutex_rtv_cnt); \
+        ((p)->root ? SC_ATOMIC_ADD((p)->root->tunnel_tpr_cnt, 1) :                  \
+                     SC_ATOMIC_ADD((p)->tunnel_tpr_cnt, 1));                        \
     } while (0)
 
 #define TUNNEL_DECR_PKT_TPR(p) do {                                                 \
-        SCMutexLock((p)->root ? &(p)->root->mutex_rtv_cnt : &(p)->mutex_rtv_cnt);   \
-        ((p)->root ? (p)->root->tpr_cnt-- : (p)->tpr_cnt--);                        \
-        SCMutexUnlock((p)->root ? &(p)->root->mutex_rtv_cnt : &(p)->mutex_rtv_cnt); \
+        ((p)->root ? SC_ATOMIC_SUB((p)->root->tunnel_tpr_cnt, 1) :                  \
+                     SC_ATOMIC_SUB((p)->tunnel_tpr_cnt, 1));                        \
     } while (0)
 
-#define TUNNEL_DECR_PKT_TPR_NOLOCK(p) do {                   \
-        ((p)->root ? (p)->root->tpr_cnt-- : (p)->tpr_cnt--); \
-    } while (0)
-
-#define TUNNEL_PKT_RTV(p)             ((p)->root ? (p)->root->rtv_cnt : (p)->rtv_cnt)
-#define TUNNEL_PKT_TPR(p)             ((p)->root ? (p)->root->tpr_cnt : (p)->tpr_cnt)
+#define TUNNEL_PKT_RTV(p)           ((p)->root ?                                    \
+        SC_ATOMIC_GET((p)->root->tunnel_rtv_cnt) :                                  \
+        SC_ATOMIC_GET((p)->tunnel_rtv_cnt))
+#define TUNNEL_PKT_TPR(p)           ((p)->root ?                                    \
+        SC_ATOMIC_GET((p)->root->tunnel_tpr_cnt) :                                  \
+        SC_ATOMIC_GET((p)->tunnel_tpr_cnt))
 
-//#define IS_TUNNEL_ROOT_PKT(p)  (((p)->root == NULL && (p)->tunnel_pkt == 1))
 #define IS_TUNNEL_PKT(p)            (((p)->flags & PKT_TUNNEL))
 #define SET_TUNNEL_PKT(p)           ((p)->flags |= PKT_TUNNEL)
 #define IS_TUNNEL_ROOT_PKT(p)       (IS_TUNNEL_PKT(p) && (p)->root == NULL)
diff --git a/src/source-ipfw.c b/src/source-ipfw.c
index cc0af3481b..53c70976f0 100644
--- a/src/source-ipfw.c
+++ b/src/source-ipfw.c
@@ -549,16 +549,15 @@ TmEcode VerdictIPFW(ThreadVars *tv, Packet *p, void *data, PacketQueue *pq, Pack
     if (IS_TUNNEL_PKT(p)) {
         char verdict = 1;
 
-        pthread_mutex_t *m = p->root ? &p->root->mutex_rtv_cnt : &p->mutex_rtv_cnt;
-        SCMutexLock(m);
         /* if there are more tunnel packets than ready to verdict packets,
          * we won't verdict this one
          */
         if (TUNNEL_PKT_TPR(p) > TUNNEL_PKT_RTV(p)) {
-            SCLogDebug("VerdictIPFW: not ready to verdict yet: TUNNEL_PKT_TPR(p) > TUNNEL_PKT_RTV(p) = %" PRId32 " > %" PRId32 "", TUNNEL_PKT_TPR(p), TUNNEL_PKT_RTV(p));
+            SCLogDebug("VerdictIPFW: not ready to verdict yet: "
+                    "TUNNEL_PKT_TPR(p) > TUNNEL_PKT_RTV(p) = %" PRId32
+                    " > %" PRId32 "", TUNNEL_PKT_TPR(p), TUNNEL_PKT_RTV(p));
             verdict = 0;
         }
-        SCMutexUnlock(m);
 
         /* don't verdict if we are not ready */
         if (verdict == 1) {
diff --git a/src/source-nfq.c b/src/source-nfq.c
index 7f864e0d79..e87465f7b8 100644
--- a/src/source-nfq.c
+++ b/src/source-nfq.c
@@ -867,15 +867,14 @@ TmEcode VerdictNFQ(ThreadVars *tv, Packet *p, void *data, PacketQueue *pq, Packe
         char verdict = 1;
         //printf("VerdictNFQ: tunnel pkt: %p %s\n", p, p->root ? "upper layer" : "root");
 
-        SCMutex *m = p->root ? &p->root->mutex_rtv_cnt : &p->mutex_rtv_cnt;
-        SCMutexLock(m);
         /* if there are more tunnel packets than ready to verdict packets,
          * we won't verdict this one */
         if (TUNNEL_PKT_TPR(p) > TUNNEL_PKT_RTV(p)) {
-            //printf("VerdictNFQ: not ready to verdict yet: TUNNEL_PKT_TPR(p) > TUNNEL_PKT_RTV(p) = %" PRId32 " > %" PRId32 "\n", TUNNEL_PKT_TPR(p), TUNNEL_PKT_RTV(p));
+            SCLogDebug("not ready to verdict yet: TUNNEL_PKT_TPR(p) > "
+                    "TUNNEL_PKT_RTV(p) = %" PRId32 " > %" PRId32,
+                    TUNNEL_PKT_TPR(p), TUNNEL_PKT_RTV(p));
             verdict = 0;
         }
-        SCMutexUnlock(m);
 
         /* don't verdict if we are not ready */
         if (verdict == 1) {
diff --git a/src/tmqh-packetpool.c b/src/tmqh-packetpool.c
index d9e169efae..b61f6ce09c 100644
--- a/src/tmqh-packetpool.c
+++ b/src/tmqh-packetpool.c
@@ -139,10 +139,6 @@ void TmqhOutputPacketpool(ThreadVars *t, Packet *p)
         SCLogDebug("Packet %p is a tunnel packet: %s",
             p,p->root ? "upper layer" : "tunnel root");
 
-        /* get a lock */
-        SCMutex *m = p->root ? &p->root->mutex_rtv_cnt : &p->mutex_rtv_cnt;
-        SCMutexLock(m);
-
         if (IS_TUNNEL_ROOT_PKT(p)) {
             SCLogDebug("IS_TUNNEL_ROOT_PKT == TRUE");
             if (TUNNEL_PKT_TPR(p) == 0) {
@@ -158,7 +154,6 @@ void TmqhOutputPacketpool(ThreadVars *t, Packet *p)
                  * by the tunnel packets, and we will enqueue it
                  * when we handle them */
                 SET_TUNNEL_PKT_VERDICTED(p);
-                SCMutexUnlock(m);
                 SCReturn;
             }
         } else {
@@ -175,7 +170,7 @@ void TmqhOutputPacketpool(ThreadVars *t, Packet *p)
                 SCLogDebug("p->root->tunnel_verdicted == 1 && TUNNEL_PKT_TPR(p) == 1");
                 /* the root is ready and we are the last tunnel packet,
                  * lets enqueue them both. */
-                TUNNEL_DECR_PKT_TPR_NOLOCK(p);
+                TUNNEL_DECR_PKT_TPR(p);
 
                 /* handle the root */
                 SCLogDebug("calling PacketEnqueue for root pkt, p->root %p (tunnel packet %p)", p->root, p);
@@ -186,12 +181,11 @@ void TmqhOutputPacketpool(ThreadVars *t, Packet *p)
                 /* root not ready yet, so get rid of the tunnel pkt only */
 
                 SCLogDebug("NOT p->root->tunnel_verdicted == 1 && TUNNEL_PKT_TPR(p) == 1 (%" PRIu32 ")", TUNNEL_PKT_TPR(p));
-                TUNNEL_DECR_PKT_TPR_NOLOCK(p);
+                TUNNEL_DECR_PKT_TPR(p);
 
                  /* fall through */
             }
         }
-        SCMutexUnlock(m);
         SCLogDebug("tunnel stuff done, move on (proot %d)", proot);
     }