性能优化，修改了VNODE入队列的实现，循环展开。调整APP入队列哈希操作等。

6 files changed, 52 insertions, 8 deletions

diff --git a/app/src/marsio.c b/app/src/marsio.c
index 7221d03..8ed226e 100644
--- a/app/src/marsio.c
+++ b/app/src/marsio.c
@@ -353,6 +353,7 @@ int marsio_thread_init(struct mr_instance * instance)
     thread_info.instance = instance;
     thread_info.cpu_id = cpu_id;
     thread_info.thread_id = instance->to_suppose_tid++;
+    RTE_PER_LCORE(_lcore_id) = cpu_id;
 
     ret = RT_SUCCESS;
     goto out;
diff --git a/app/src/rawio.c b/app/src/rawio.c
index 0432241..d9a3f87 100644
--- a/app/src/rawio.c
+++ b/app/src/rawio.c
@@ -64,7 +64,7 @@ int marsio_send_burst(struct mr_sendpath * sendpath, queue_id_t sid, marsio_buff
 
     /* 计算哈希，目前全部按SID填 */
     hash_t hash[MR_BURST_MAX];
-    for (int i = 0; i < nr_mbufs; i++) hash[i] = sid;
+    for (int i = 0; i < nr_mbufs; i++) hash[i] = __mbufs[i]->hash.usr;
     
     vnode_mirror_enqueue_bulk(sendpath->target_vdi->vnode_tx_prod, sid, __mbufs, hash, nr_mbufs);
     return RT_SUCCESS;
diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt
index a9e81f7..311215c 100644
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -9,5 +9,8 @@ target_link_libraries(mrtest-l2fwd marsio pthread)
 add_executable(mrtest-fb feedback.c)
 target_link_libraries(mrtest-fb marsio pthread)
 
+add_executable(mrtest-rxonly rxonly.c)
+target_link_libraries(mrtest-rxonly marsio pthread)
+
 #add_executable(pag_rx_sample pag/rx_sample.c)
 #target_link_libraries(pag_rx_sample pag)
\ No newline at end of file
diff --git a/examples/rxonly.c b/examples/rxonly.c
index 9bd2f9c..633cbc4 100644
--- a/examples/rxonly.c
+++ b/examples/rxonly.c
@@ -29,7 +29,7 @@ void * rxonly_loop(void * arg)
     for (;;)
     {
         ret = marsio_recv_burst(dev_handler, sid, rx_buff, nr_burst);
-        marsio_buff_free(dev_handler, rx_buff, nr_burst);
+        marsio_buff_free(dev_handler, rx_buff, ret);
     }
 
     return (void *)NULL;
diff --git a/include/internal/mr_config.h.in b/include/internal/mr_config.h.in
index 009cf7b..9c5b954 100644
--- a/include/internal/mr_config.h.in
+++ b/include/internal/mr_config.h.in
@@ -1,6 +1,7 @@
-#pragma once

+#pragma once

 

-#cmakedefine MR_INSTALL_PREFIX	"@CMAKE_INSTALL_PREFIX@"

-#cmakedefine MR_GLOBAL_CONFFILE	"@MR_GLOBAL_CONFFILE@"

-#cmakedefine MR_LOCAL_CONFFILE	"@MR_LOCAL_CONFFILE@"

-#cmakedefine PAG_LOCAL_CONFFILE "@PAG_LOCAL_CONFFILE@"
\ No newline at end of file
+#cmakedefine MR_INSTALL_PREFIX	            "@CMAKE_INSTALL_PREFIX@"

+#cmakedefine MR_GLOBAL_CONFFILE         	"@MR_GLOBAL_CONFFILE@"

+#cmakedefine MR_LOCAL_CONFFILE	            "@MR_LOCAL_CONFFILE@"

+#cmakedefine PAG_LOCAL_CONFFILE             "@PAG_LOCAL_CONFFILE@"
+#cmakedefine MR_VNODE_MIRROR_BY_REFCNT      "@MR_VNODE_MIRROR_BY_REFCNT@"
\ No newline at end of file
diff --git a/infra/src/vnode_mirror.c b/infra/src/vnode_mirror.c
index 23efe1c..108b23e 100644
--- a/infra/src/vnode_mirror.c
+++ b/infra/src/vnode_mirror.c
@@ -71,7 +71,7 @@ static inline int dist_tunnel_dequeue(struct tunnel_desc * desc, void * obj, int
 static inline void dist_tunnel_enqueue_without_buffer(struct tunnel_desc * desc, void * obj)
 {
 	assert(desc->en_buffer == NULL && desc->sz_en_buffer == 0);
-	int ret = rte_ring_enqueue_burst(desc->tunnel_object, &obj, 1);
+	int ret = rte_ring_sp_enqueue_burst(desc->tunnel_object, &obj, 1);
 	if (unlikely(ret != 1)) rte_pktmbuf_free(obj);
 	return;
 }
@@ -80,6 +80,7 @@ static inline void dist_tunnel_enqueue_without_buffer(struct tunnel_desc * desc,
 // only use by prod.
 // TODO: rewrite in SSE/SSE2/AVX/AVX2 intrinsics
 
+#if 0
 static inline void dist_tunnel_block_enqueue_with_hash(struct tunnel_block * block, 
 	int prodq, struct rte_mbuf * obj[], uint32_t hash[], int nr_obj)
 {
@@ -94,6 +95,44 @@ static inline void dist_tunnel_block_enqueue_with_hash(struct tunnel_block * blo
 	}
 }
 
+#endif
+
+#define __FWDSTEP       4
+
+static inline void dist_tunnel_block_enqueue_with_hash(struct tunnel_block * block,
+    int prodq, struct rte_mbuf * obj[], uint32_t hash[], int nr_obj)
+{
+    struct tunnel_desc * tunnel;
+    unsigned int idx = 0;
+    unsigned int consq;
+
+    switch (nr_obj % 4)
+    {
+    case 0:
+        while (idx != nr_obj) {
+            consq = hash[idx] % block->nr_consq;
+            tunnel = *tunnel_block_locate(block, prodq, consq);
+            dist_tunnel_enqueue(block->prod, block->cons, prodq, consq, tunnel, obj[idx]);
+            idx++;
+    case 3:
+        consq = hash[idx] % block->nr_consq;
+        tunnel = *tunnel_block_locate(block, prodq, consq);
+        dist_tunnel_enqueue(block->prod, block->cons, prodq, consq, tunnel, obj[idx]);
+        idx++;
+    case 2:
+        consq = hash[idx] % block->nr_consq;
+        tunnel = *tunnel_block_locate(block, prodq, consq);
+        dist_tunnel_enqueue(block->prod, block->cons, prodq, consq, tunnel, obj[idx]);
+        idx++;
+    case 1:
+        consq = hash[idx] % block->nr_consq;
+        tunnel = *tunnel_block_locate(block, prodq, consq);
+        dist_tunnel_enqueue(block->prod, block->cons, prodq, consq, tunnel, obj[idx]);
+        idx++;
+        }
+    }
+}
+
 // Tunnel Block Dequeue, dequeue from block, only used by cons.
 // TODO: rewrite in SSE/SSE2/AVX/AVX2 intrinsics