diff --git a/src/stream-tcp-reassemble.c b/src/stream-tcp-reassemble.c
index e3ee004875..cebdd53544 100644
--- a/src/stream-tcp-reassemble.c
+++ b/src/stream-tcp-reassemble.c
@@ -466,6 +466,19 @@ int StreamTcpReassemblyConfig(char quiet)
     if (!quiet)
         SCLogInfo("stream.reassembly \"chunk-prealloc\": %u", stream_chunk_prealloc);
     StreamMsgQueuesInit(stream_chunk_prealloc);
+
+    intmax_t zero_copy_size = 128;
+    if (ConfGetInt("stream.reassembly.zero-copy-size", &zero_copy_size) == 1) {
+        if (zero_copy_size < 0 || zero_copy_size > 0xffff) {
+            SCLogError(SC_ERR_INVALID_ARGUMENT, "stream.reassembly.zero-copy-size of "
+                    "%"PRIiMAX" is invalid: valid values are 0 to 65535", zero_copy_size);
+            return -1;
+        }
+    }
+    stream_config.zero_copy_size = (uint16_t)zero_copy_size;
+    if (!quiet)
+        SCLogInfo("stream.reassembly \"zero-copy-size\": %u", stream_config.zero_copy_size);
+
     return 0;
 }
 
@@ -2577,7 +2590,10 @@ static inline int DoReassemble(ThreadVars *tv, TcpReassemblyThreadCtx *ra_ctx,
                  Packet *p)
 {
     /* fast path 1: segment is exactly what we need */
-    if (likely(rd->data_len == 0 && SEQ_EQ(seg->seq, rd->ra_base_seq+1) && SEQ_EQ(stream->last_ack, (seg->seq + seg->payload_len)))) {
+    if (likely(rd->data_len == 0 &&
+               SEQ_EQ(seg->seq, rd->ra_base_seq+1) &&
+               SEQ_EQ(stream->last_ack, (seg->seq + seg->payload_len))))
+    {
         /* process single segment directly */
         AppLayerHandleTCPData(tv, ra_ctx, p, p->flow, ssn, stream,
                 seg->payload, seg->payload_len,
@@ -2596,7 +2612,11 @@ static inline int DoReassemble(ThreadVars *tv, TcpReassemblyThreadCtx *ra_ctx,
         }
         return 1;
     /* fast path 2: segment acked completely, meets minimal size req for 0copy processing */
-    } else if (rd->data_len == 0 && SEQ_EQ(seg->seq, rd->ra_base_seq+1) && SEQ_GT(stream->last_ack, (seg->seq + seg->payload_len)) && seg->payload_len >= 128) {
+    } else if (rd->data_len == 0 &&
+               SEQ_EQ(seg->seq, rd->ra_base_seq+1) &&
+               SEQ_GT(stream->last_ack, (seg->seq + seg->payload_len)) &&
+               seg->payload_len >= stream_config.zero_copy_size)
+    {
         /* process single segment directly */
         AppLayerHandleTCPData(tv, ra_ctx, p, p->flow, ssn, stream,
                 seg->payload, seg->payload_len,
diff --git a/src/stream-tcp.h b/src/stream-tcp.h
index b8a31855af..113e2e88a1 100644
--- a/src/stream-tcp.h
+++ b/src/stream-tcp.h
@@ -51,6 +51,9 @@ typedef struct TcpStreamCnf_ {
     uint32_t ssn_init_flags; /**< new ssn flags will be initialized to this */
     uint8_t segment_init_flags; /**< new seg flags will be initialized to this */
 
+    uint16_t zero_copy_size;    /**< use zero copy for app layer above segments
+                                 *   of this size */
+
     uint32_t prealloc_sessions; /**< ssns to prealloc per stream thread */
     int midstream;
     int async_oneside;
diff --git a/suricata.yaml.in b/suricata.yaml.in
index 04cb3b2285..2bb208381e 100644
--- a/suricata.yaml.in
+++ b/suricata.yaml.in
@@ -785,6 +785,11 @@ flow-timeouts:
 #       - size: 4               # Size of the (data)segment for a pool
 #         prealloc: 256         # Number of segments to prealloc and keep
 #                               # in the pool.
+#     zero-copy-size: 128       # This option sets in bytes the value at
+#                               # which segment data is passed to the app
+#                               # layer API directly. Data sizes equal to
+#                               # and higher than the value set are passed
+#                               # on directly.
 #
 stream:
   memcap: 32mb
@@ -816,6 +821,7 @@ stream:
     #    prealloc: 1024
     #  - size: 65535
     #    prealloc: 128
+    #zero-copy-size: 128
 
 # Host table:
 #