diff --git a/src/cuda-packet-batcher.c b/src/cuda-packet-batcher.c
index 0e8c7fa8b4..4bc2fddb2b 100644
--- a/src/cuda-packet-batcher.c
+++ b/src/cuda-packet-batcher.c
@@ -129,15 +129,9 @@ static void SCCudaPBSetBatcherAlarmTimeHandler()
  */
 static SigGroupHead *SCCudaPBGetSgh(DetectEngineCtx *de_ctx, Packet *p)
 {
-    int ds, f;
+    int f;
     SigGroupHead *sgh = NULL;
 
-    /* select the dsize_gh */
-    if (p->payload_len <= 100)
-        ds = 0;
-    else
-        ds = 1;
-
     /* select the flow_gh */
     if (p->flowflags & FLOW_PKT_TOCLIENT)
         f = 0;
@@ -145,7 +139,7 @@ static SigGroupHead *SCCudaPBGetSgh(DetectEngineCtx *de_ctx, Packet *p)
         f = 1;
 
     /* find the right mpm instance */
-    DetectAddress *ag = DetectAddressLookupInHead(de_ctx->dsize_gh[ds].flow_gh[f].src_gh[p->proto], &p->src);
+    DetectAddress *ag = DetectAddressLookupInHead(de_ctx->flow_gh[f].src_gh[p->proto], &p->src);
     if (ag != NULL) {
         /* source group found, lets try a dst group */
         ag = DetectAddressLookupInHead(ag->dst_gh,&p->dst);
@@ -222,7 +216,7 @@ static void SCCudaPBQueueBuffer(SCCudaPBThreadCtx *tctx)
         SCMutexLock(&dq_inq->mutex_q);
         if (dq_inq->len == 0) {
             /* if we have no data in queue, wait... */
-            SCondWait(&dq_inq->cond_q, &dq_inq->mutex_q);
+            SCCondWait(&dq_inq->cond_q, &dq_inq->mutex_q);
         }
 
         if (run_batcher == 0) {
@@ -279,7 +273,8 @@ void *SCCudaPBTmThreadsSlot1(void *td)
             pthread_exit((void *) -1);
         }
     }
-    memset(&s->s.slot_pq, 0, sizeof(PacketQueue));
+    memset(&s->s.slot_pre_pq, 0, sizeof(PacketQueue));
+    memset(&s->s.slot_post_pq, 0, sizeof(PacketQueue));
 
     TmThreadsSetFlag(tv, THV_INIT_DONE);
     while(run) {
@@ -299,12 +294,11 @@ void *SCCudaPBTmThreadsSlot1(void *td)
              * the Batcher TM(which is waiting on a cond from the previous
              * feeder TM).  Please handler the NULL packet case in the
              * function that you now call */
-            r = s->s.SlotFunc(tv, p, s->s.slot_data, &s->s.slot_pq);
+            r = s->s.SlotFunc(tv, p, s->s.slot_data, NULL, NULL);
         } else {
-            r = s->s.SlotFunc(tv, p, s->s.slot_data, &s->s.slot_pq);
+            r = s->s.SlotFunc(tv, p, s->s.slot_data, NULL, NULL);
             /* handle error */
             if (r == TM_ECODE_FAILED) {
-                TmqhReleasePacketsToPacketPool(&s->s.slot_pq);
                 TmqhOutputPacketpool(tv, p);
                 TmThreadsSetFlag(tv, THV_FAILED);
                 break;
@@ -498,7 +492,7 @@ TmEcode SCCudaPBThreadInit(ThreadVars *tv, void *initdata, void **data)
  * \retval TM_ECODE_OK     On success.
  * \retval TM_ECODE_FAILED On failure.
  */
-TmEcode SCCudaPBBatchPackets(ThreadVars *tv, Packet *p, void *data, PacketQueue *pq)
+TmEcode SCCudaPBBatchPackets(ThreadVars *tv, Packet *p, void *data, PacketQueue *pq, PacketQueue *post_pq)
 {
 #define ALIGN_UP(offset, alignment) \
     (offset) = ((offset) + (alignment) - 1) & ~((alignment) - 1)
@@ -971,7 +965,7 @@ int SCCudaPBTest01(void)
 
     p.payload = (uint8_t *)strings[0];
     p.payload_len = strlen(strings[0]);
-    SCCudaPBBatchPackets(NULL, &p, tctx, NULL);
+    SCCudaPBBatchPackets(NULL, &p, tctx, NULL, NULL);
     dq = &data_queues[tmq_outq->id];
     result &= (dq->len == 0);
     dq = &data_queues[tmq_inq->id];
@@ -979,7 +973,7 @@ int SCCudaPBTest01(void)
 
     p.payload = (uint8_t *)strings[1];
     p.payload_len = strlen(strings[1]);
-    SCCudaPBBatchPackets(NULL, &p, tctx, NULL);
+    SCCudaPBBatchPackets(NULL, &p, tctx, NULL, NULL);
     dq = &data_queues[tmq_outq->id];
     result &= (dq->len == 0);
     dq = &data_queues[tmq_inq->id];
@@ -987,7 +981,7 @@ int SCCudaPBTest01(void)
 
     p.payload = (uint8_t *)strings[2];
     p.payload_len = strlen(strings[2]);
-    SCCudaPBBatchPackets(NULL, &p, tctx, NULL);
+    SCCudaPBBatchPackets(NULL, &p, tctx, NULL, NULL);
     dq = &data_queues[tmq_outq->id];
     result &= (dq->len == 0);
     dq = &data_queues[tmq_inq->id];
@@ -995,7 +989,7 @@ int SCCudaPBTest01(void)
 
     p.payload = (uint8_t *)strings[3];
     p.payload_len = strlen(strings[3]);
-    SCCudaPBBatchPackets(NULL, &p, tctx, NULL);
+    SCCudaPBBatchPackets(NULL, &p, tctx, NULL, NULL);
     dq = &data_queues[tmq_outq->id];
     result &= (dq->len == 0);
     dq = &data_queues[tmq_inq->id];
@@ -1003,7 +997,7 @@ int SCCudaPBTest01(void)
 
     p.payload = (uint8_t *)strings[4];
     p.payload_len = strlen(strings[4]);
-    SCCudaPBBatchPackets(NULL, &p, tctx, NULL);
+    SCCudaPBBatchPackets(NULL, &p, tctx, NULL, NULL);
     dq = &data_queues[tmq_outq->id];
     result &= (dq->len == 0);
     dq = &data_queues[tmq_inq->id];
@@ -1011,7 +1005,7 @@ int SCCudaPBTest01(void)
 
     p.payload = (uint8_t *)strings[5];
     p.payload_len = strlen(strings[5]);
-    SCCudaPBBatchPackets(NULL, &p, tctx, NULL);
+    SCCudaPBBatchPackets(NULL, &p, tctx, NULL, NULL);
     dq = &data_queues[tmq_outq->id];
     result &= (dq->len == 0);
     dq = &data_queues[tmq_inq->id];
@@ -1019,7 +1013,7 @@ int SCCudaPBTest01(void)
 
     p.payload = (uint8_t *)strings[6];
     p.payload_len = strlen(strings[6]);
-    SCCudaPBBatchPackets(NULL, &p, tctx, NULL);
+    SCCudaPBBatchPackets(NULL, &p, tctx, NULL, NULL);
     dq = &data_queues[tmq_outq->id];
     result &= (dq->len == 0);
     dq = &data_queues[tmq_inq->id];
@@ -1027,7 +1021,7 @@ int SCCudaPBTest01(void)
 
     p.payload = (uint8_t *)strings[7];
     p.payload_len = strlen(strings[7]);
-    SCCudaPBBatchPackets(NULL, &p, tctx, NULL);
+    SCCudaPBBatchPackets(NULL, &p, tctx, NULL, NULL);
     dq = &data_queues[tmq_outq->id];
     result &= (dq->len == 0);
     dq = &data_queues[tmq_inq->id];
@@ -1035,7 +1029,7 @@ int SCCudaPBTest01(void)
 
     p.payload = (uint8_t *)strings[8];
     p.payload_len = strlen(strings[8]);
-    SCCudaPBBatchPackets(NULL, &p, tctx, NULL);
+    SCCudaPBBatchPackets(NULL, &p, tctx, NULL, NULL);
     dq = &data_queues[tmq_outq->id];
     result &= (dq->len == 0);
     dq = &data_queues[tmq_inq->id];
@@ -1043,7 +1037,7 @@ int SCCudaPBTest01(void)
 
     p.payload = (uint8_t *)strings[9];
     p.payload_len = strlen(strings[9]);
-    SCCudaPBBatchPackets(NULL, &p, tctx, NULL);
+    SCCudaPBBatchPackets(NULL, &p, tctx, NULL, NULL);
     dq = &data_queues[tmq_outq->id];
     result &= (dq->len == 1);
     dq = &data_queues[tmq_inq->id];
@@ -1198,7 +1192,7 @@ int SCCudaPBTest02(void)
     string = "test_one";
     p.payload = (uint8_t *)string;
     p.payload_len = strlen(string);
-    SCCudaPBBatchPackets(NULL, &p, tctx, NULL);
+    SCCudaPBBatchPackets(NULL, &p, tctx, NULL, NULL);
     dq = &data_queues[tmq_outq->id];
     result &= (dq->len == 0);
     dq = &data_queues[tmq_inq->id];
diff --git a/src/cuda-packet-batcher.h b/src/cuda-packet-batcher.h
index 60586fb6a3..eb0579f0fe 100644
--- a/src/cuda-packet-batcher.h
+++ b/src/cuda-packet-batcher.h
@@ -124,7 +124,7 @@ void SCCudaPBCleanUpQueuesAndBuffers(void);
 void SCCudaPBSetUpQueuesAndBuffers(void);
 void SCCudaPBKillBatchingPackets(void);
 
-TmEcode SCCudaPBBatchPackets(ThreadVars *, Packet *, void *, PacketQueue *);
+TmEcode SCCudaPBBatchPackets(ThreadVars *, Packet *, void *, PacketQueue *, PacketQueue *);
 TmEcode SCCudaPBThreadInit(ThreadVars *, void *, void **);
 TmEcode SCCudaPBThreadDeInit(ThreadVars *, void *);
 void SCCudaPBThreadExitStats(ThreadVars *, void *);
diff --git a/src/decode.h b/src/decode.h
index 40cd8b3005..7c783e5096 100644
--- a/src/decode.h
+++ b/src/decode.h
@@ -247,10 +247,6 @@ struct PacketQueue_;
  */
 typedef struct Packet_
 {
-    /* double linked list ptrs */
-    struct Packet_ *next;
-    struct Packet_ *prev;
-
     /* Addresses, Ports and protocol
      * these are on top so we can use
      * the Packet as a hash key */
@@ -360,6 +356,10 @@ typedef struct Packet_
     /* decoder events */
     PacketDecoderEvents events;
 
+    /* double linked list ptrs */
+    struct Packet_ *next;
+    struct Packet_ *prev;
+
     /* tunnel/encapsulation handling */
     struct Packet_ *root; /* in case of tunnel this is a ptr
                            * to the 'real' packet, the one we
diff --git a/src/runmodes.c b/src/runmodes.c
index 4effce00db..e08bcc4916 100644
--- a/src/runmodes.c
+++ b/src/runmodes.c
@@ -2312,7 +2312,7 @@ int RunModeFilePcapAuto(DetectEngineCtx *de_ctx, char *file) {
 #if defined(__SC_CUDA_SUPPORT__)
     if (PatternMatchDefaultMatcher() == MPM_B2G_CUDA) {
         ThreadVars *tv_decode1 = TmThreadCreatePacketHandler("Decode",
-                                                             "pickup-queue", "simple",
+                                                             "pickup-queue", "ringbuffer_srsw",
                                                              "decode-queue1", "simple",
                                                              "1slot");
         if (tv_decode1 == NULL) {
@@ -2363,7 +2363,7 @@ int RunModeFilePcapAuto(DetectEngineCtx *de_ctx, char *file) {
 
         ThreadVars *tv_stream1 = TmThreadCreatePacketHandler("Stream1",
                                                              "cuda-pb-queue1", "simple",
-                                                             "stream-queue1", "simple",
+                                                             "stream-queue1", "ringbuffer_mrsw",
                                                              "1slot");
         if (tv_stream1 == NULL) {
             printf("ERROR: TmThreadsCreate failed for Stream1\n");
@@ -2386,8 +2386,8 @@ int RunModeFilePcapAuto(DetectEngineCtx *de_ctx, char *file) {
         }
     } else {
         ThreadVars *tv_decode1 = TmThreadCreatePacketHandler("Decode & Stream",
-                                                             "pickup-queue", "simple",
-                                                             "stream-queue1", "simple",
+                                                             "pickup-queue", "ringbuffer_srsw",
+                                                             "stream-queue1", "ringbuffer_mrsw",
                                                              "varslot");
         if (tv_decode1 == NULL) {
             printf("ERROR: TmThreadsCreate failed for Decode1\n");
diff --git a/src/tm-threads.h b/src/tm-threads.h
index 20f0b5b879..91a8b0346b 100644
--- a/src/tm-threads.h
+++ b/src/tm-threads.h
@@ -34,10 +34,9 @@ enum {
     TVT_MAX,
 };
 
-
 typedef struct TmSlot_ {
     /* function pointers */
-    TmEcode (*SlotFunc)(ThreadVars *, Packet *, void *, PacketQueue *);
+    TmEcode (*SlotFunc)(ThreadVars *, Packet *, void *, PacketQueue *, PacketQueue *);
 
     TmEcode (*SlotThreadInit)(ThreadVars *, void *, void **);
     void (*SlotThreadExitPrintStats)(ThreadVars *, void *);
@@ -46,10 +45,22 @@ typedef struct TmSlot_ {
     /* data storage */
     void *slot_initdata;
     void *slot_data;
-    PacketQueue slot_pq;
+
+    /**< queue filled by the SlotFunc with packets that will
+     *   be processed futher _before_ the current packet.
+     *   The locks in the queue are NOT used */
+    PacketQueue slot_pre_pq;
+
+    /**< queue filled by the SlotFunc with packets that will
+     *   be processed futher _after_ the current packet. The
+     *   locks in the queue are NOT used */
+    PacketQueue slot_post_pq;
 
     /* linked list, only used by TmVarSlot */
     struct TmSlot_ *slot_next;
+
+    int id; /**< slot id, only used my TmVarSlot to know what the first
+             *   slot is. */
 } TmSlot;
 
 /* 1 function slot */
@@ -61,7 +72,6 @@ typedef struct Tm1Slot_ {
 typedef struct TmVarSlot_ {
     TmSlot *s;
 } TmVarSlot;
-
 extern ThreadVars *tv_root[TVT_MAX];
 
 extern SCMutex tv_root_lock;
diff --git a/src/util-mpm-b2g-cuda.c b/src/util-mpm-b2g-cuda.c
index 24b36651c1..87b973f070 100644
--- a/src/util-mpm-b2g-cuda.c
+++ b/src/util-mpm-b2g-cuda.c
@@ -79,7 +79,7 @@ int B2gCudaAddPatternCI(MpmCtx *, uint8_t *, uint16_t, uint16_t, uint16_t,
 int B2gCudaAddPatternCS(MpmCtx *, uint8_t *, uint16_t, uint16_t, uint16_t,
                         uint32_t, uint32_t, uint8_t);
 int B2gCudaPreparePatterns(MpmCtx *mpm_ctx);
-inline uint32_t B2gCudaSearchWrap(MpmCtx *, MpmThreadCtx *,
+uint32_t B2gCudaSearchWrap(MpmCtx *, MpmThreadCtx *,
                                   PatternMatcherQueue *, uint8_t *,
                                   uint16_t);
 uint32_t B2gCudaSearch1(MpmCtx *, MpmThreadCtx *, PatternMatcherQueue *,
@@ -441,8 +441,8 @@ void MpmB2gCudaRegister(void)
 
 void B2gCudaPrintInfo(MpmCtx *mpm_ctx)
 {
+#ifdef DEBUG
     B2gCudaCtx *ctx = (B2gCudaCtx *)mpm_ctx->ctx;
-
     SCLogDebug("MPM B2g Cuda Information:");
     SCLogDebug("Memory allocs:    %" PRIu32, mpm_ctx->memory_cnt);
     SCLogDebug("Memory alloced:   %" PRIu32, mpm_ctx->memory_size);
@@ -456,6 +456,7 @@ void B2gCudaPrintInfo(MpmCtx *mpm_ctx)
     SCLogDebug("Smallest:         %" PRIu32, mpm_ctx->minlen);
     SCLogDebug("Largest:          %" PRIu32, mpm_ctx->maxlen);
     SCLogDebug("Hash size:        %" PRIu32, ctx->hash_size);
+#endif
 
     return;
 }
@@ -1940,7 +1941,7 @@ TmEcode B2gCudaMpmDispThreadDeInit(ThreadVars *tv, void *data)
  * \retval TM_ECODE_OK Always.
  */
 TmEcode B2gCudaMpmDispatcher(ThreadVars *tv, Packet *incoming_buffer,
-                             void *data, PacketQueue *pq)
+                             void *data, PacketQueue *pq, PacketQueue *post_pq)
 {
     SCCudaPBPacketsBuffer *pb = (SCCudaPBPacketsBuffer *)incoming_buffer;
     B2gCudaMpmThreadCtxData *tctx = data;
@@ -2044,7 +2045,7 @@ int B2gCudaResultsPostProcessing(Packet *p, MpmCtx *mpm_ctx,
             SCMutexUnlock(&p->cuda_mutex);
             break;
         } else {
-            SCondWait(&p->cuda_cond, &p->cuda_mutex);
+            SCCondWait(&p->cuda_cond, &p->cuda_mutex);
             SCMutexUnlock(&p->cuda_mutex);
         }
     }
@@ -2154,7 +2155,8 @@ void *CudaMpmB2gThreadsSlot1(void *td)
             pthread_exit((void *) -1);
         }
     }
-    memset(&s->s.slot_pq, 0, sizeof(PacketQueue));
+    memset(&s->s.slot_pre_pq, 0, sizeof(PacketQueue));
+    memset(&s->s.slot_post_pq, 0, sizeof(PacketQueue));
 
     TmThreadsSetFlag(tv, THV_INIT_DONE);
     while(run) {
@@ -2166,7 +2168,7 @@ void *CudaMpmB2gThreadsSlot1(void *td)
         if (data == NULL) {
             //printf("%s: TmThreadsSlot1: p == NULL\n", tv->name);
         } else {
-            r = s->s.SlotFunc(tv, (Packet *)data, s->s.slot_data, &s->s.slot_pq);
+            r = s->s.SlotFunc(tv, (Packet *)data, s->s.slot_data, NULL, NULL);
             /* handle error */
 
             /* output the packet */
@@ -2339,7 +2341,7 @@ static int B2gCudaTest01(void)
     pb->packets_address_buffer[0] = &p;
     p.payload_len = strlen(string);
 
-    B2gCudaMpmDispatcher(NULL, (Packet *)pb, tctx, NULL);
+    B2gCudaMpmDispatcher(NULL, (Packet *)pb, tctx, NULL, NULL);
 
     result &= (p.mpm_offsets[0] == 4);
     result &= (p.mpm_offsets[1] == 1);
@@ -2533,7 +2535,7 @@ static int B2gCudaTest02(void)
     for (i = 0; i < no_of_pkts; i++) {
         p[i]->payload = (uint8_t *)strings[i];
         p[i]->payload_len = strlen(strings[i]);
-        SCCudaPBBatchPackets(NULL, p[i], pb_tctx, NULL);
+        SCCudaPBBatchPackets(NULL, p[i], pb_tctx, NULL, NULL);
     }
 
     dq = &data_queues[tmq_outq->id];
@@ -2556,7 +2558,7 @@ static int B2gCudaTest02(void)
     if (b2g_tctx->b2g_cuda_search_kernel == 0)
         goto end;
 
-    B2gCudaMpmDispatcher(NULL, (Packet *)pb, b2g_tctx, NULL);
+    B2gCudaMpmDispatcher(NULL, (Packet *)pb, b2g_tctx, NULL, NULL);
 
     for (i = 0; i < no_of_pkts; i++) {
         for (j = 0; j < p[i]->mpm_offsets[0]; j++)
@@ -2832,7 +2834,7 @@ static int B2gCudaTest03(void)
     for (i = 0; i < no_of_pkts; i++) {
         p[i]->payload = (uint8_t *)strings[i];
         p[i]->payload_len = strlen(strings[i]);
-        SCCudaPBBatchPackets(NULL, p[i], pb_tctx, NULL);
+        SCCudaPBBatchPackets(NULL, p[i], pb_tctx, NULL, NULL);
     }
 
     dq = &data_queues[tmq_outq->id];
@@ -2855,7 +2857,7 @@ static int B2gCudaTest03(void)
     if (b2g_tctx->b2g_cuda_search_kernel == 0)
         goto end;
 
-    B2gCudaMpmDispatcher(NULL, (Packet *)pb, b2g_tctx, NULL);
+    B2gCudaMpmDispatcher(NULL, (Packet *)pb, b2g_tctx, NULL, NULL);
 
     for (i = 0; i < 10; i++)
         SigMatchSignatures(&de_tv, de_ctx, det_ctx, p[i]);