sockfilt: fix handling of ready closed sockets on Windows

Replace the incomplete workaround regarding FD_CLOSE
only signalling once by instead doing a pre-check with
standard select and storing the result for later use.

select keeps triggering on closed sockets on Windows while
WSAEventSelect fires only once with data still available.
By doing the pre-check we do not run in a deadlock
due to waiting forever for another FD_CLOSE event.

1 files changed, 35 insertions, 25 deletions

diff --git a/tests/server/sockfilt.c b/tests/server/sockfilt.c
index 9c43f90e4..26bbf416d 100644
--- a/tests/server/sockfilt.c
+++ b/tests/server/sockfilt.c
@@ -742,6 +742,7 @@ static HANDLE select_ws_wait(HANDLE handle, HANDLE signal,
   return thread;
 }
 struct select_ws_data {
+  WSANETWORKEVENTS pre;  /* the internal select result (indexed by fds/idx) */
   curl_socket_t fd;      /* the original input handle  (indexed by fds/idx) */
   curl_socket_t wsasock; /* the internal socket handle (indexed by wsa) */
   WSAEVENT wsaevent;     /* the internal WINSOCK event (indexed by wsa) */
@@ -752,6 +753,7 @@ static int select_ws(int nfds, fd_set *readfds, fd_set *writefds,
                      fd_set *exceptfds, struct timeval *timeout)
 {
   HANDLE abort, mutex, signal, handle, *handles;
+  fd_set readsock, writesock, exceptsock;
   DWORD milliseconds, wait, idx;
   WSANETWORKEVENTS wsanetevents;
   struct select_ws_data *data;
@@ -811,14 +813,24 @@ static int select_ws(int nfds, fd_set *readfds, fd_set *writefds,
     long networkevents = 0;
     handles[nfd] = 0;
 
-    if(FD_ISSET(fds, readfds))
+    FD_ZERO(&readsock);
+    FD_ZERO(&writesock);
+    FD_ZERO(&exceptsock);
+
+    if(FD_ISSET(fds, readfds)) {
+      FD_SET(fds, &readsock);
       networkevents |= FD_READ|FD_ACCEPT|FD_CLOSE;
+    }
 
-    if(FD_ISSET(fds, writefds))
+    if(FD_ISSET(fds, writefds)) {
+      FD_SET(fds, &writesock);
       networkevents |= FD_WRITE|FD_CONNECT;
+    }
 
-    if(FD_ISSET(fds, exceptfds))
-      networkevents |= FD_OOB|FD_CLOSE;
+    if(FD_ISSET(fds, exceptfds)) {
+      FD_SET(fds, &exceptsock);
+      networkevents |= FD_OOB;
+    }
 
     /* only wait for events for which we actually care */
     if(networkevents) {
@@ -854,6 +866,20 @@ static int select_ws(int nfds, fd_set *readfds, fd_set *writefds,
             handles[nfd] = handle;
             data[wsa].wsasock = curlx_sitosk(fds);
             data[wsa].wsaevent = wsaevent;
+            data[nfd].pre.lNetworkEvents = 0;
+            tv->tv_sec = 0;
+            tv->tv_usec = 0;
+            /* check if the socket is already ready */
+            if(select(fds + 1, &readsock, &writesock, &exceptsock, tv) == 1) {
+              logmsg("[select_ws] socket %d is ready", fds);
+              WSASetEvent(wsaevent);
+              if(FD_ISSET(fds, &readsock))
+                data[nfd].pre.lNetworkEvents |= FD_READ;
+              if(FD_ISSET(fds, &writesock))
+                data[nfd].pre.lNetworkEvents |= FD_WRITE;
+              if(FD_ISSET(fds, &exceptsock))
+                data[nfd].pre.lNetworkEvents |= FD_OOB;
+            }
             wsa++;
           }
           else {
@@ -919,6 +945,9 @@ static int select_ws(int nfds, fd_set *readfds, fd_set *writefds,
         wsanetevents.lNetworkEvents = 0;
         error = WSAEnumNetworkEvents(fds, handle, &wsanetevents);
         if(error != SOCKET_ERROR) {
+          /* merge result from pre-check using select */
+          wsanetevents.lNetworkEvents |= data[idx].pre.lNetworkEvents;
+
           /* remove from descriptor set if not ready for read/accept/close */
           if(!(wsanetevents.lNetworkEvents & (FD_READ|FD_ACCEPT|FD_CLOSE)))
             FD_CLR(sock, readfds);
@@ -927,17 +956,8 @@ static int select_ws(int nfds, fd_set *readfds, fd_set *writefds,
           if(!(wsanetevents.lNetworkEvents & (FD_WRITE|FD_CONNECT)))
             FD_CLR(sock, writefds);
 
-          /* HACK:
-           * use exceptfds together with readfds to signal
-           * that the connection was closed by the client.
-           *
-           * Reason: FD_CLOSE is only signaled once, sometimes
-           * at the same time as FD_READ with data being available.
-           * This means that recv/sread is not reliable to detect
-           * that the connection is closed.
-           */
           /* remove from descriptor set if not exceptional */
-          if(!(wsanetevents.lNetworkEvents & (FD_OOB|FD_CLOSE)))
+          if(!(wsanetevents.lNetworkEvents & (FD_OOB)))
             FD_CLR(sock, exceptfds);
         }
       }
@@ -1061,9 +1081,6 @@ static bool juggle(curl_socket_t *sockfdp,
     else {
       /* there's always a socket to wait for */
       FD_SET(sockfd, &fds_read);
-#ifdef USE_WINSOCK
-      FD_SET(sockfd, &fds_err);
-#endif
       maxfd = (int)sockfd;
     }
     break;
@@ -1074,9 +1091,6 @@ static bool juggle(curl_socket_t *sockfdp,
     /* sockfd turns CURL_SOCKET_BAD when our connection has been closed */
     if(CURL_SOCKET_BAD != sockfd) {
       FD_SET(sockfd, &fds_read);
-#ifdef USE_WINSOCK
-      FD_SET(sockfd, &fds_err);
-#endif
       maxfd = (int)sockfd;
     }
     else {
@@ -1254,11 +1268,7 @@ static bool juggle(curl_socket_t *sockfdp,
       lograw(buffer, nread_socket);
     }
 
-    if(nread_socket <= 0
-#ifdef USE_WINSOCK
-       || FD_ISSET(sockfd, &fds_err)
-#endif
-       ) {
+    if(nread_socket <= 0) {
       logmsg("====> Client disconnect");
       if(!write_stdout("DISC\n", 5))
         return FALSE;