增加性能分析使用的运行Cycles记录功能

5 files changed, 68 insertions, 6 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index db8bc9f..f7fe045 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -31,6 +31,7 @@ set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${EXTRA_CXX_FLAGS}")
 # Default Install Destination Directory
 #set(CMAKE_INSTALL_PREFIX /opt/iiesoft/marsio)
 add_definitions(-fPIC)
+add_definitions(-D_GNU_SOURCE)
 
 add_subdirectory(${CMAKE_SOURCE_DIR}/support ${CMAKE_BINARY_DIR}/support)
 add_subdirectory(${CMAKE_SOURCE_DIR}/core)
diff --git a/core/include/mr_stat.h b/core/include/mr_stat.h
index bafcb33..9a20194 100644
--- a/core/include/mr_stat.h
+++ b/core/include/mr_stat.h
@@ -1,6 +1,7 @@
 #pragma once
 
 #include <rte_config.h>
+#include <rte_cycles.h>
 #include <rte_memory.h>
 #include <mr_ctx.h>
 #include <rte_mbuf.h>
@@ -50,9 +51,29 @@ struct mr_app_stat
 	} __per_app[MR_APP_MAX];
 };
 
+struct mr_perf_item
+{
+	volatile uint64_t begin;
+	volatile uint64_t runtime;
+	volatile uint64_t cycles;
+};
+
+struct mr_perf_stat
+{
+	struct
+	{
+		struct mr_perf_item sc_rx_loop_all;
+		struct mr_perf_item sc_rx_loop_eth_rx_burst;
+		struct mr_perf_item sc_rx_loop_hash_calc;
+		struct mr_perf_item sc_rx_loop_dispatch;
+		struct mr_perf_item pag_get_frame;
+		struct mr_perf_item pag_get;
+	} __rte_cache_aligned _per_thread[MR_MAX_GSID];
+};
 
 extern struct mr_vnode_stat * mr_vnode_stat_get();
 extern struct mr_device_stat * mr_device_stat_get();
+extern struct mr_perf_stat * mr_perf_stat_get();
 
 #define UPDATE_DEV_STAT(_s, _p, _n)										\
 do {																	\
@@ -66,6 +87,27 @@ do {																	\
 	if(st) st->_per_thread[mr_thread_id()]._s += _n;					\
 }while(0)																\
 
+#define PERF_BEGIN(_s)													\
+do {																	\
+	struct mr_perf_stat * st = mr_perf_stat_get();						\
+	uint64_t _tsc = rte_rdtsc();										\
+	if(st) st->_per_thread[mr_thread_id()]._s.begin = _tsc;				\
+}while(0)																\
+
+
+#define PERF_END(_s)													\
+do {																	\
+	struct mr_perf_stat * st = mr_perf_stat_get();						\
+	if(st)																\
+	{																	\
+		thread_id_t _tid = mr_thread_id();								\
+		uint64_t _tsc = rte_rdtsc();									\
+		uint64_t _begin = st->_per_thread[_tid]._s.begin;				\
+		uint64_t _delta = _tsc - _begin;								\
+		st->_per_thread[_tid]._s.runtime += _delta;						\
+		st->_per_thread[_tid]._s.cycles++;								\
+	}																	\
+} while(0)																\
 
 // Tools Functions
 static inline int pktslen(struct rte_mbuf * mbufs[], unsigned nr_mbufs)
diff --git a/core/src/stat.c b/core/src/stat.c
index 9295193..cd36101 100644
--- a/core/src/stat.c
+++ b/core/src/stat.c
@@ -17,6 +17,7 @@ struct stat_ctx
 	unsigned int vnode_stat_enable;
 	struct mr_device_stat device_stat_handle;
 	struct mr_vnode_stat vnode_stat_handle;
+	struct mr_perf_stat perf_stat_handle;
 };
 
 static struct stat_ctx * module_ctx;
@@ -34,6 +35,12 @@ struct mr_vnode_stat * mr_vnode_stat_get()
 	return &module_ctx->vnode_stat_handle;
 }
 
+struct mr_perf_stat * mr_perf_stat_get()
+{
+	//TODO: ���ÿ���
+	return &module_ctx->perf_stat_handle;
+}
+
 static int stat_preinit(struct mr_module * m_desc, struct mr_config * g_cfg,
 	struct mr_global_ctx * g_ctx)
 {
diff --git a/pag/libpag.c b/pag/libpag.c
index 6f8aedf..3233949 100644
--- a/pag/libpag.c
+++ b/pag/libpag.c
@@ -6,15 +6,15 @@
 

 #include <unistd.h>

 #include <assert.h>

+#include <netinet/in.h>

 #include <rte_string_fns.h>

 #include <rte_ether.h>

 #include <MESA_prof_load.h>

 #include <mr_common.h>

 #include <mr_rawio.h>

+#include <mr_stat.h>

 #include <mrlib.h>

-

 #include "libpag.h"

-#include <netinet/in.h>

 

 struct pag_instance

 {

@@ -125,6 +125,8 @@ int pag_open()
 	return 0;

 }

 

+//TODO: �ձ��ķ�ʽ�Ľ�����������һ�����ػ�������

+//TODO: ��һ��rtdevice����n���������ϲ�ȡ�á�

 void * pag_get_frame(int sid __rte_unused)

 {

 	if (unlikely(__pag_thread_inited == 0))

@@ -132,7 +134,8 @@ void * pag_get_frame(int sid __rte_unused)
 		mrlib_thread_init();

 		__pag_thread_inited = 1;

 	}

-	

+

+	PERF_BEGIN(pag_get_frame);

 	struct pag_thread_instance * tinstance = &thread_instance_;

 	struct pag_instance * instance = &pag_config_;

 

@@ -149,6 +152,7 @@ void * pag_get_frame(int sid __rte_unused)
 	int ret = marsio_raw_recv(instance->raw_sockets[cur_socket], &tinstance->rxmbuf, 1);

 	tinstance->cur_socket = (cur_socket + 1) % instance->nr_raw_sockets;

 

+	PERF_END(pag_get_frame);

 	if (ret != 1) return NULL;

 	return rte_pktmbuf_mtod(tinstance->rxmbuf, void *);

 }

diff --git a/service/src/rxtx.c b/service/src/rxtx.c
index 303c330..e8b8c68 100644
--- a/service/src/rxtx.c
+++ b/service/src/rxtx.c
@@ -23,13 +23,16 @@ void sc_thread_rx_loop(struct sc_instance * instance,
 	hash_t hash_result[MR_BURST_MAX];
 	int nr_rx_burst = instance->sz_rx_burst;
 
-	for(int i =0; i < tinstance->nr_devices; i++)
+	PERF_BEGIN(sc_rx_loop_all);
+	for(unsigned int i =0; i < tinstance->nr_devices; i++)
 	{
 		mr_devq_t * devq;
 		devq = tinstance->rt_dev[i].devq;
 		
-		int nr_rx_bufs = rte_eth_rx_burst(devq->port_id,
+		PERF_BEGIN(sc_rx_loop_eth_rx_burst);
+		uint16_t nr_rx_bufs = rte_eth_rx_burst(devq->port_id,
 			devq->rx_queue_id, rx_bufs, nr_rx_burst);
+		PERF_END(sc_rx_loop_eth_rx_burst);
 
 		// û�����ݰ�ֱ������
 		if (unlikely(nr_rx_bufs == 0)) continue;
@@ -38,16 +41,21 @@ void sc_thread_rx_loop(struct sc_instance * instance,
 		UPDATE_DEV_STAT(dev_rx_bit, devq->port_id, pktslen(rx_bufs, nr_rx_bufs));
 
 		// �������ݰ����طֵ�ʹ�õ�Hashֵ
+		PERF_BEGIN(sc_rx_loop_hash_calc);
 		distributer_hash_caculate(instance->distmode,
 			rx_bufs, nr_rx_bufs, hash_result);
+		PERF_END(sc_rx_loop_hash_calc);
 
 		// ����VNode�У��ַ�����Ӧ��
 		struct vnodeman_attach_desc * attach_desc;
 		attach_desc = tinstance->rt_dev[i].rxvnode_desc;
+		PERF_BEGIN(sc_rx_loop_dispatch);
 		vnode_enqueue_burst_with_hash(attach_desc->prod, &vnodeman_ops,
 			attach_desc->queue_id, (void **)rx_bufs, hash_result, nr_rx_bufs);
+		PERF_END(sc_rx_loop_dispatch);
 	}
-	
+
+	PERF_END(sc_rx_loop_all);
 	return;
 }