fix missing increment of the reference count on every call to startup()

also change how the implicit startup is tracked.
The extra change is needed since the startup() calls in
srt::CUDT::socket()
and
srt::CUDT::createGroup
would endlessly increment the startup counter as sockets were created.

This also adds handling for the case where the user code fails to call
srt_startup when initially using the library, but does later on.
It better matches up the implicit startup() with an implicit cleanup().

fixes #3098

srtcore/api.cpp

@@ -180,6 +180,7 @@ srt::CUDTUnited::CUDTUnited()
     , m_InitLock()
     , m_iInstanceCount(0)
     , m_bGCStatus(false)
+    , m_bImplicitStartup(false)
     , m_ClosedSockets()
 {
     // Socket ID MUST start from a random value
@@ -198,12 +199,15 @@ srt::CUDTUnited::CUDTUnited()
 
 srt::CUDTUnited::~CUDTUnited()
 {
-    // Call it if it wasn't called already.
-    // This will happen at the end of main() of the application,
-    // when the user didn't call srt_cleanup().
-    if (m_bGCStatus)
     {
-        cleanup();
+        // If an implicit startup was called, then call a matching cleanup
+        // This will happen at the end of main() of the application,
+        // when the user didn't call srt_cleanup() before using the library.
+        ScopedLock gcinit(m_InitLock);
+        if (m_bImplicitStartup)
+        {
+            cleanup();
+        }
     }
 
     releaseMutex(m_GlobControlLock);
@@ -233,15 +237,24 @@ string srt::CUDTUnited::CONID(SRTSOCKET sock)
     return os.str();
 }
 
-int srt::CUDTUnited::startup()
+void srt::CUDTUnited::implicitStartup()
 {
     ScopedLock gcinit(m_InitLock);
     if (m_bGCStatus)
-        return 1;
+        return;
+    m_bImplicitStartup = true;
+    startup();
+}
 
+int srt::CUDTUnited::startup()
+{
+    ScopedLock gcinit(m_InitLock);
     if (m_iInstanceCount++ > 0)
         return 1;
 
+    if (m_bGCStatus)
+        return 1;
+
         // Global initialization code
 #ifdef _WIN32
     WORD    wVersionRequested;
@@ -299,6 +312,7 @@ int srt::CUDTUnited::cleanup()
     m_GCThread.join();
 
     m_bGCStatus = false;
+    m_bImplicitStartup = false;
 
     // Global destruction code
 #ifdef _WIN32
@@ -3469,7 +3483,7 @@ int srt::CUDT::cleanup()
 
 SRTSOCKET srt::CUDT::socket()
 {
-    uglobal().startup();
+    uglobal().implicitStartup();
 
     try
     {
@@ -3519,7 +3533,7 @@ srt::CUDTGroup& srt::CUDT::newGroup(const int type)
 SRTSOCKET srt::CUDT::createGroup(SRT_GROUP_TYPE gt)
 {
     // Doing the same lazy-startup as with srt_create_socket()
-    uglobal().startup();
+    uglobal().implicitStartup();
 
     try
     {

srtcore/api.h

@@ -428,6 +428,7 @@ class CUDTUnited
     CUDTSocket* locateSocket_LOCKED(SRTSOCKET u);
     CUDTSocket* locatePeer(const sockaddr_any& peer, const SRTSOCKET id, int32_t isn);
 
+    void implicitStartup();
 #if ENABLE_BONDING
     CUDTGroup* locateAcquireGroup(SRTSOCKET u, ErrorHandling erh = ERH_RETURN);
     CUDTGroup* acquireSocketsGroup(CUDTSocket* s);
@@ -542,6 +543,8 @@ class CUDTUnited
     int         m_iInstanceCount; // number of startup() called by application
     SRT_ATTR_GUARDED_BY(m_InitLock)
     bool        m_bGCStatus;      // if the GC thread is working (true)
+    SRT_ATTR_GUARDED_BY(m_InitLock)
+    bool        m_bImplicitStartup; // if startup was implicitly called
 
     SRT_ATTR_GUARDED_BY(m_InitLock)
     sync::CThread m_GCThread;