for #320, support macro to disable the complex send algorithm and enable tcp no delay. 2.0.129

-#ifndef _srs_icpp_init_stub
-#define _srs_icpp_init_stub
+#ifndef _srs_upp_icpp_init_stub
+#define _srs_upp_icpp_init_stub
 #endif

@@ -36,8 +36,8 @@ file
 	../../src/kernel/srs_kernel_log.cpp,
 	../../src/kernel/srs_kernel_mp3.hpp,
 	../../src/kernel/srs_kernel_mp3.cpp,
-        ../../src/kernel/srs_rtsp_stack.hpp,
-        ../../src/kernel/srs_rtsp_stack.cpp,
+	../../src/kernel/srs_rtsp_stack.hpp,
+	../../src/kernel/srs_rtsp_stack.cpp,
 	../../src/kernel/srs_kernel_stream.hpp,
 	../../src/kernel/srs_kernel_stream.cpp,
 	../../src/kernel/srs_kernel_ts.cpp,
@@ -45,8 +45,8 @@ file
 	../../src/kernel/srs_kernel_utility.hpp,
 	../../src/kernel/srs_kernel_utility.cpp,
 	protocol readonly separator,
-        ../../src/protocol/srs_raw_avc.hpp,
-        ../../src/protocol/srs_raw_avc.cpp,
+	../../src/protocol/srs_raw_avc.hpp,
+	../../src/protocol/srs_raw_avc.cpp,
 	../../src/protocol/srs_rtmp_amf0.hpp,
 	../../src/protocol/srs_rtmp_amf0.cpp,
 	../../src/protocol/srs_rtmp_buffer.hpp,
@@ -102,8 +102,8 @@ file
 	../../src/app/srs_app_json.cpp,
 	../../src/app/srs_app_kbps.hpp,
 	../../src/app/srs_app_kbps.cpp,
-        ../../src/app/srs_app_listener.hpp,
-        ../../src/app/srs_app_listener.cpp,
+	../../src/app/srs_app_listener.hpp,
+	../../src/app/srs_app_listener.cpp,
 	../../src/app/srs_app_log.hpp,
 	../../src/app/srs_app_log.cpp,
 	../../src/app/srs_app_mpegts_udp.hpp,
@@ -116,8 +116,8 @@ file
 	../../src/app/srs_app_reload.cpp,
 	../../src/app/srs_app_rtmp_conn.hpp,
 	../../src/app/srs_app_rtmp_conn.cpp,
-        ../../src/app/srs_app_rtsp.hpp,
-        ../../src/app/srs_app_rtsp.cpp,
+	../../src/app/srs_app_rtsp.hpp,
+	../../src/app/srs_app_rtsp.cpp,
 	../../src/app/srs_app_pithy_print.hpp,
 	../../src/app/srs_app_pithy_print.cpp,
 	../../src/app/srs_app_security.hpp,

@@ -27,6 +27,7 @@ CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 #include <sys/socket.h>
 #include <netinet/in.h>
 #include <arpa/inet.h>
+#include <netinet/tcp.h>
 
 using namespace std;
 
@@ -624,6 +625,9 @@ int SrsRtmpConn::do_playing(SrsSource* source, SrsConsumer* consumer, SrsQueueRe
     mw_enabled = true;
     change_mw_sleep(_srs_config->get_mw_sleep_ms(req->vhost));
     
+    // set the sock options.
+    play_set_sock_options();
+    
     while (true) {
         // collect elapse for pithy print.
         pprint->elapse();
@@ -1089,7 +1093,7 @@ void SrsRtmpConn::change_mw_sleep(int sleep_ms)
     socklen_t sock_buf_size = sizeof(int);
     getsockopt(fd, SOL_SOCKET, SO_SNDBUF, &onb_sbuf, &sock_buf_size);
     
-#ifdef SRS_PERF_MW_SO_SNDBUF    
+#ifdef SRS_PERF_MW_SO_SNDBUF
     // the bytes:
     //      4KB=4096, 8KB=8192, 16KB=16384, 32KB=32768, 64KB=65536,
     //      128KB=131072, 256KB=262144, 512KB=524288
@@ -1123,6 +1127,29 @@ void SrsRtmpConn::change_mw_sleep(int sleep_ms)
     mw_sleep = sleep_ms;
 }
 
+void SrsRtmpConn::play_set_sock_options()
+{
+    int fd = st_netfd_fileno(stfd);
+    
+#ifdef SRS_PERF_TCP_NODELAY
+    if (true) {
+        socklen_t nb_v = sizeof(int);
+        
+        int ov = 0;
+        getsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &ov, &nb_v);
+        
+        int v = 1;
+        // set the socket send buffer when required larger buffer
+        if (setsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &v, nb_v) < 0) {
+            srs_warn("set sock TCP_NODELAY=%d failed.", v);
+        }
+        getsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &v, &nb_v);
+        
+        srs_trace("set TCP_NODELAY %d=>%d", ov, v);
+    }
+#endif
+}
+
 int SrsRtmpConn::check_edge_token_traverse_auth()
 {
     int ret = ERROR_SUCCESS;

@@ -111,6 +111,7 @@ private:
     virtual int process_publish_message(SrsSource* source, SrsCommonMessage* msg, bool vhost_is_edge);
     virtual int process_play_control_msg(SrsConsumer* consumer, SrsCommonMessage* msg);
     virtual void change_mw_sleep(int sleep_ms);
+    virtual void play_set_sock_options();
 private:
     virtual int check_edge_token_traverse_auth();
     virtual int connect_server(int origin_index, st_netfd_t* pstsock);

@@ -31,7 +31,7 @@ CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 // current release version
 #define VERSION_MAJOR       2
 #define VERSION_MINOR       0
-#define VERSION_REVISION    128
+#define VERSION_REVISION    129
 
 // server info.
 #define RTMP_SIG_SRS_KEY "SRS"

@@ -108,7 +108,8 @@ CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 * whether set the socket send buffer size.
 * @see https://github.com/winlinvip/simple-rtmp-server/issues/251
 */
-#undef SRS_PERF_MW_SO_SNDBUF
+#define SRS_PERF_MW_SO_SNDBUF
+
 /**
 * whether set the socket recv buffer size.
 * @see https://github.com/winlinvip/simple-rtmp-server/issues/251
@@ -154,5 +155,20 @@ CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 // in seconds, the live queue length.
 #define SRS_PERF_PLAY_QUEUE 30
 
+/**
+* whether always use complex send algorithm.
+* for some network does not support the complex send,
+* @see https://github.com/winlinvip/simple-rtmp-server/issues/320
+*/
+//#undef SRS_PERF_COMPLEX_SEND
+#define SRS_PERF_COMPLEX_SEND
+/**
+* whether enable the TCP_NODELAY
+* user maybe need send small tcp packet for some network.
+* @see https://github.com/winlinvip/simple-rtmp-server/issues/320
+*/
+//#define SRS_PERF_TCP_NODELAY
+#undef SRS_PERF_TCP_NODELAY
+
 #endif
 

@@ -762,8 +762,9 @@ int SrsProtocol::do_send_messages(SrsSharedPtrMessage** msgs, int nb_msgs)
 {
     int ret = ERROR_SUCCESS;
     
+#ifdef SRS_PERF_COMPLEX_SEND
     int iov_index = 0;
-    iovec* iov = out_iovs + iov_index;
+    iovec* iovs = out_iovs + iov_index;
     
     int c0c3_cache_index = 0;
     char* c0c3_cache = out_c0c3_caches + c0c3_cache_index;
@@ -796,13 +797,13 @@ int SrsProtocol::do_send_messages(SrsSharedPtrMessage** msgs, int nb_msgs)
             srs_assert(nbh > 0);
             
             // header iov
-            iov[0].iov_base = c0c3_cache;
-            iov[0].iov_len = nbh;
+            iovs[0].iov_base = c0c3_cache;
+            iovs[0].iov_len = nbh;
             
             // payload iov
             int payload_size = srs_min(out_chunk_size, pend - p);
-            iov[1].iov_base = p;
-            iov[1].iov_len = payload_size;
+            iovs[1].iov_base = p;
+            iovs[1].iov_len = payload_size;
             
             // consume sendout bytes.
             p += payload_size;
@@ -822,7 +823,7 @@ int SrsProtocol::do_send_messages(SrsSharedPtrMessage** msgs, int nb_msgs)
             
             // to next pair of iovs
             iov_index += 2;
-            iov = out_iovs + iov_index;
+            iovs = out_iovs + iov_index;
 
             // to next c0c3 header cache
             c0c3_cache_index += nbh;
@@ -849,7 +850,7 @@ int SrsProtocol::do_send_messages(SrsSharedPtrMessage** msgs, int nb_msgs)
                 // reset caches, while these cache ensure 
                 // atleast we can sendout a chunk.
                 iov_index = 0;
-                iov = out_iovs + iov_index;
+                iovs = out_iovs + iov_index;
                 
                 c0c3_cache_index = 0;
                 c0c3_cache = out_c0c3_caches + c0c3_cache_index;
@@ -866,6 +867,61 @@ int SrsProtocol::do_send_messages(SrsSharedPtrMessage** msgs, int nb_msgs)
         nb_msgs, iov_index, SRS_PERF_MW_MSGS, nb_out_iovs);
 
     return do_iovs_send(out_iovs, iov_index);
+#else
+    // try to send use the c0c3 header cache,
+    // if cache is consumed, try another loop.
+    for (int i = 0; i < nb_msgs; i++) {
+        SrsSharedPtrMessage* msg = msgs[i];
+        
+        if (!msg) {
+            continue;
+        }
+    
+        // ignore empty message.
+        if (!msg->payload || msg->size <= 0) {
+            srs_info("ignore empty message.");
+            continue;
+        }
+    
+        // p set to current write position,
+        // it's ok when payload is NULL and size is 0.
+        char* p = msg->payload;
+        char* pend = msg->payload + msg->size;
+        
+        // always write the header event payload is empty.
+        while (p < pend) {
+            // for simple send, send each chunk one by one
+            iovec* iovs = out_iovs;
+            char* c0c3_cache = out_c0c3_caches;
+            int nb_cache = SRS_CONSTS_C0C3_HEADERS_MAX;
+            
+            // always has header
+            int nbh = msg->chunk_header(c0c3_cache, nb_cache, p == msg->payload);
+            srs_assert(nbh > 0);
+            
+            // header iov
+            iovs[0].iov_base = c0c3_cache;
+            iovs[0].iov_len = nbh;
+            
+            // payload iov
+            int payload_size = srs_min(out_chunk_size, pend - p);
+            iovs[1].iov_base = p;
+            iovs[1].iov_len = payload_size;
+            
+            // consume sendout bytes.
+            p += payload_size;
+
+            if ((ret = skt->writev(iovs, 2, NULL)) != ERROR_SUCCESS) {
+                if (!srs_is_client_gracefully_close(ret)) {
+                    srs_error("send packet with writev failed. ret=%d", ret);
+                }
+                return ret;
+            }
+        }
+    }
+    
+    return ret;
+#endif   
 }
 
 int SrsProtocol::do_iovs_send(iovec* iovs, int size)