add more optim

source/libs/transport/inc/transComm.h

@@ -213,6 +213,12 @@ typedef struct SConnBuffer {
 
 typedef void (*AsyncCB)(uv_async_t* handle);
 
+typedef struct {
+  void*           pThrd;
+  queue           qmsg;
+  pthread_mutex_t mtx;  // protect qmsg;
+} SAsyncItem;
+
 typedef struct {
   int         index;
   int         nAsync;
@@ -221,7 +227,7 @@ typedef struct {
 
 SAsyncPool* transCreateAsyncPool(uv_loop_t* loop, void* arg, AsyncCB cb);
 void        transDestroyAsyncPool(SAsyncPool* pool);
-int         transSendAsync(SAsyncPool* pool);
+int         transSendAsync(SAsyncPool* pool, queue* mq);
 
 int transInitBuffer(SConnBuffer* buf);
 int transClearBuffer(SConnBuffer* buf);

source/libs/transport/src/transCli.c

@@ -432,14 +432,15 @@ static void clientHandleReq(SCliMsg* pMsg, SCliThrdObj* pThrd) {
   }
 }
 static void clientAsyncCb(uv_async_t* handle) {
-  SCliThrdObj* pThrd = handle->data;
+  SAsyncItem*  item = handle->data;
+  SCliThrdObj* pThrd = item->pThrd;
   SCliMsg*     pMsg = NULL;
   queue        wq;
 
   // batch process to avoid to lock/unlock frequently
-  pthread_mutex_lock(&pThrd->msgMtx);
-  QUEUE_MOVE(&pThrd->msg, &wq);
-  pthread_mutex_unlock(&pThrd->msgMtx);
+  pthread_mutex_lock(&item->mtx);
+  QUEUE_MOVE(&item->qmsg, &wq);
+  pthread_mutex_unlock(&item->mtx);
 
   int count = 0;
   while (!QUEUE_IS_EMPTY(&wq)) {
@@ -548,11 +549,11 @@ static void clientSendQuit(SCliThrdObj* thrd) {
   SCliMsg* msg = calloc(1, sizeof(SCliMsg));
   msg->ctx = NULL;  //
 
-  pthread_mutex_lock(&thrd->msgMtx);
-  QUEUE_PUSH(&thrd->msg, &msg->q);
-  pthread_mutex_unlock(&thrd->msgMtx);
+  // pthread_mutex_lock(&thrd->msgMtx);
+  // QUEUE_PUSH(&thrd->msg, &msg->q);
+  // pthread_mutex_unlock(&thrd->msgMtx);
 
-  transSendAsync(thrd->asyncPool);
+  transSendAsync(thrd->asyncPool, &msg->q);
   // uv_async_send(thrd->cliAsync);
 }
 void taosCloseClient(void* arg) {
@@ -598,14 +599,14 @@ void rpcSendRequest(void* shandle, const SEpSet* pEpSet, SRpcMsg* pMsg, int64_t*
 
   SCliThrdObj* thrd = ((SClientObj*)pRpc->tcphandle)->pThreadObj[index % pRpc->numOfThreads];
 
-  pthread_mutex_lock(&thrd->msgMtx);
-  QUEUE_PUSH(&thrd->msg, &cliMsg->q);
-  pthread_mutex_unlock(&thrd->msgMtx);
+  // pthread_mutex_lock(&thrd->msgMtx);
+  // QUEUE_PUSH(&thrd->msg, &cliMsg->q);
+  // pthread_mutex_unlock(&thrd->msgMtx);
 
-  int start = taosGetTimestampUs();
-  transSendAsync(thrd->asyncPool);
+  // int start = taosGetTimestampUs();
+  transSendAsync(thrd->asyncPool, &(cliMsg->q));
   // uv_async_send(thrd->cliAsync);
-  int end = taosGetTimestampUs() - start;
+  // int end = taosGetTimestampUs() - start;
   // tError("client sent to rpc, time cost: %d", (int)end);
 }
 #endif

source/libs/transport/src/transComm.c

@@ -247,7 +247,7 @@ int transDestroyBuffer(SConnBuffer* buf) {
 }
 
 SAsyncPool* transCreateAsyncPool(uv_loop_t* loop, void* arg, AsyncCB cb) {
-  static int sz = 20;
+  static int sz = 10;
 
   SAsyncPool* pool = calloc(1, sizeof(SAsyncPool));
   pool->index = 0;
@@ -257,24 +257,46 @@ SAsyncPool* transCreateAsyncPool(uv_loop_t* loop, void* arg, AsyncCB cb) {
   for (int i = 0; i < pool->nAsync; i++) {
     uv_async_t* async = &(pool->asyncs[i]);
     uv_async_init(loop, async, cb);
-    async->data = arg;
+
+    SAsyncItem* item = calloc(1, sizeof(SAsyncItem));
+    item->pThrd = arg;
+    QUEUE_INIT(&item->qmsg);
+    pthread_mutex_init(&item->mtx, NULL);
+
+    async->data = item;
   }
   return pool;
 }
 void transDestroyAsyncPool(SAsyncPool* pool) {
   for (int i = 0; i < pool->nAsync; i++) {
     uv_async_t* async = &(pool->asyncs[i]);
+
+    SAsyncItem* item = async->data;
+    pthread_mutex_destroy(&item->mtx);
+    free(item);
   }
   free(pool->asyncs);
   free(pool);
 }
-int transSendAsync(SAsyncPool* pool) {
+int transSendAsync(SAsyncPool* pool, queue* q) {
   int idx = pool->index;
   idx = idx % pool->nAsync;
   // no need mutex here
   if (pool->index++ > pool->nAsync) {
     pool->index = 0;
   }
-  return uv_async_send(&(pool->asyncs[idx]));
+  uv_async_t* async = &(pool->asyncs[idx]);
+  SAsyncItem* item = async->data;
+
+  int64_t st = taosGetTimestampUs();
+  pthread_mutex_lock(&item->mtx);
+  QUEUE_PUSH(&item->qmsg, q);
+  pthread_mutex_unlock(&item->mtx);
+  int64_t el = taosGetTimestampUs() - st;
+  if (el > 50) {
+    // tInfo("lock and unlock cost: %d", (int)el);
+  }
+
+  return uv_async_send(async);
 }
 #endif

source/libs/transport/src/transSrv.c

@@ -376,13 +376,15 @@ static void destroySmsg(SSrvMsg* smsg) {
   free(smsg);
 }
 void uvWorkerAsyncCb(uv_async_t* handle) {
-  SWorkThrdObj* pThrd = handle->data;
+  SAsyncItem*   item = handle->data;
+  SWorkThrdObj* pThrd = item->pThrd;
   SSrvConn*     conn = NULL;
   queue         wq;
   // batch process to avoid to lock/unlock frequently
-  pthread_mutex_lock(&pThrd->msgMtx);
-  QUEUE_MOVE(&pThrd->msg, &wq);
-  pthread_mutex_unlock(&pThrd->msgMtx);
+  pthread_mutex_lock(&item->mtx);
+  QUEUE_MOVE(&item->qmsg, &wq);
+  pthread_mutex_unlock(&item->mtx);
+  // pthread_mutex_unlock(&mtx);
 
   while (!QUEUE_IS_EMPTY(&wq)) {
     queue* head = QUEUE_HEAD(&wq);
@@ -539,7 +541,7 @@ static bool addHandleToAcceptloop(void* arg) {
     tError("failed to bind: %s", uv_err_name(err));
     return false;
   }
-  if ((err = uv_listen((uv_stream_t*)&srv->server, 128, uvOnAcceptCb)) != 0) {
+  if ((err = uv_listen((uv_stream_t*)&srv->server, 512, uvOnAcceptCb)) != 0) {
     tError("failed to listen: %s", uv_err_name(err));
     return false;
   }
@@ -671,12 +673,12 @@ void destroyWorkThrd(SWorkThrdObj* pThrd) {
 void sendQuitToWorkThrd(SWorkThrdObj* pThrd) {
   SSrvMsg* srvMsg = calloc(1, sizeof(SSrvMsg));
 
-  pthread_mutex_lock(&pThrd->msgMtx);
-  QUEUE_PUSH(&pThrd->msg, &srvMsg->q);
-  pthread_mutex_unlock(&pThrd->msgMtx);
+  // pthread_mutex_lock(&pThrd->msgMtx);
+  // QUEUE_PUSH(&pThrd->msg, &srvMsg->q);
+  // pthread_mutex_unlock(&pThrd->msgMtx);
   tDebug("send quit msg to work thread");
 
-  transSendAsync(pThrd->asyncPool);
+  transSendAsync(pThrd->asyncPool, &srvMsg->q);
   // uv_async_send(pThrd->workerAsync);
 }
 
@@ -712,12 +714,12 @@ void rpcSendResponse(const SRpcMsg* pMsg) {
   srvMsg->pConn = pConn;
   srvMsg->msg = *pMsg;
 
-  pthread_mutex_lock(&pThrd->msgMtx);
-  QUEUE_PUSH(&pThrd->msg, &srvMsg->q);
-  pthread_mutex_unlock(&pThrd->msgMtx);
+  // pthread_mutex_lock(&pThrd->msgMtx);
+  // QUEUE_PUSH(&pThrd->msg, &srvMsg->q);
+  // pthread_mutex_unlock(&pThrd->msgMtx);
 
   tDebug("conn %p start to send resp", pConn);
-  transSendAsync(pThrd->asyncPool);
+  transSendAsync(pThrd->asyncPool, &srvMsg->q);
   // uv_async_send(pThrd->workerAsync);
 }