issue: 3844385 Fix new TCP timers registration lock contention

Avoid calling register_socket_timer_event when a socket is already registered (TIME-WAIT).
Although there is no functionality issue with that, it produces too high rate of posting events for internal-thread.
This leads to lock contantion inside internal-thread and degraded performance of HTTP CPS.

Signed-off-by: Alexander Grissik <agrissik@nvidia.com>

src/core/sock/sockinfo_tcp.cpp

@@ -2482,10 +2482,15 @@ ssize_t sockinfo_tcp::rx(const rx_call_t call_type, iovec *p_iov, ssize_t sz_iov
 
 void sockinfo_tcp::register_timer()
 {
-    si_tcp_logdbg("Registering TCP socket timer: socket: %p, thread-col: %p, global-col: %p", this,
-                  get_tcp_timer_collection(), g_tcp_timers_collection);
+    // A reused time-wait socket wil try to add a timer although it is already registered.
+    // We should avoid calling register_socket_timer_event unnecessarily because it introduces
+    // internal-thread locks contention.
+    if (!is_timer_registered()) {
+        si_tcp_logdbg("Registering TCP socket timer: socket: %p, thread-col: %p, global-col: %p",
+                      this, get_tcp_timer_collection(), g_tcp_timers_collection);
 
-    get_event_mgr()->register_socket_timer_event(this);
+        get_event_mgr()->register_socket_timer_event(this);
+    }
 }
 
 void sockinfo_tcp::queue_rx_ctl_packet(struct tcp_pcb *pcb, mem_buf_desc_t *p_desc)
@@ -6017,15 +6022,16 @@ void tcp_timers_collection::add_new_timer(sockinfo_tcp *sock)
         return;
     }
 
-    // A reused time-wait socket wil try to add a timer although it is already registered.
-    if (m_sock_remove_map.find(sock) != m_sock_remove_map.end()) {
+    if (sock->is_timer_registered()) {
+        __log_warn("Trying to add timer twice for TCP socket %p", sock);
         return;
     }
 
     sock_list &bucket = m_p_intervals[m_n_next_insert_bucket];
     bucket.emplace_back(sock);
     m_sock_remove_map.emplace(sock, std::make_tuple(m_n_next_insert_bucket, --(bucket.end())));
     m_n_next_insert_bucket = (m_n_next_insert_bucket + 1) % m_n_intervals_size;
+    sock->set_timer_registered(true);
 
     if (0 == m_n_count++) {
         m_timer_handle = get_event_mgr()->register_timer_event(safe_mce_sys().timer_resolution_msec,
@@ -6041,6 +6047,7 @@ void tcp_timers_collection::remove_timer(sockinfo_tcp *sock)
     if (node != m_sock_remove_map.end()) {
         m_p_intervals[std::get<0>(node->second)].erase(std::get<1>(node->second));
         m_sock_remove_map.erase(node);
+        sock->set_timer_registered(false);
 
         if (!(--m_n_count)) {
             if (m_timer_handle) {

src/core/sock/sockinfo_tcp.h

@@ -320,6 +320,8 @@ class sockinfo_tcp : public sockinfo {
     }
 
     bool is_incoming() override { return m_b_incoming; }
+    bool is_timer_registered() const { return m_timer_registered; }
+    void set_timer_registered(bool v) { m_timer_registered = v; }
 
     bool is_connected() { return m_sock_state == TCP_SOCK_CONNECTED_RDWR; }
 
@@ -618,6 +620,7 @@ class sockinfo_tcp : public sockinfo {
     bool m_xlio_thr;
     bool m_b_incoming;
     bool m_b_attached;
+    bool m_timer_registered = false;
     /* connection state machine */
     int m_conn_timeout;
     /* RCVBUF acconting */