[PATCH] Fix poll/select signal socket as write ready on connect failure

[Marc Hoersken]

Hi Corinna,

Am 16.07.2020 um 11:25 schrieb Corinna Vinschen:
> Thanks for the patch.  I pushed it.


thanks for pushing it already. Please excuse my delayed response, family 
live kept me busy.


> But then I got second thoughts in terms of how to fix the issue.


Yes, I also got second thoughts yesterday about my initial approach.


> The reason is that the FD_CLOSE problem shouldn't exist,
> simply for the fact that we never remove FD_CLOSE from
> the events mask, see
>
> https://sourceware.org/git/?p=newlib-cygwin.git;a=blob;f=winsup/cygwin/fhandler_socket_inet.cc;hb=HEAD#l377


Thanks, I also understood in the meantime, that some flags/events are 
not removed from wsock_events->event. FD_CLOSE seems not to be affected 
as you described and I was unable to produce an issue in case the 
connection was closed from the server side. So only the FD_CONNECT to 
FD_WRITE handling remained problematic (before the patch).


> So, rather than setting FD_WRITE at some later point in the code, what
> about handling this where the other FD_CONNECT stuff is handled, by
> not erasing the FD_CONNECT bit, just like with FD_CLOSE?


I think this makes more sense, yes. I am just not sure if the socket 
should also be write ready in case of a socket error. Looking at the 
description of EINPROGRESS on the man page of connect [1], it seems like 
writability is given regardless of the connection being successful or 
not, but as soon as the connection attempt is no longer pending. For 
successful connections FD_WRITE will be given already, so we will only 
need to set it for failed connections regardless of a socket error in 
wsock_events->connect_errorcode. Therefore I suggest to move the line 
setting FD_WRITE [2] one level up outside of the else branch.


> diff --git a/winsup/cygwin/fhandler_socket_inet.cc b/winsup/cygwin/fhandler_socket_inet.cc
> index e5b0d2d1443e..b64d96225db1 100644
> --- a/winsup/cygwin/fhandler_socket_inet.cc
> +++ b/winsup/cygwin/fhandler_socket_inet.cc
> @@ -354,7 +354,12 @@ fhandler_socket_wsock::evaluate_events (const long event_mask, long &events,
>   	    }
>   	  else
>   	    wsock_events->events |= FD_WRITE;
> -	  wsock_events->events &= ~FD_CONNECT;
> +	  /* Since FD_CONNECT is only given once, we have to keep FD_CONNECT
> +	     for connection failed sockets to have consistent behaviour in
> +	     programs calling poll/select multiple times.  Example test to
> +	     non-listening port: curl -v 127.0.0.1:47 */
> +	  if (connect_state () != connect_failed)
> +	    wsock_events->events &= ~FD_CONNECT;
>   	  wsock_events->connect_errorcode = 0;
>   	}
>         /* This test makes accept/connect behave as on Linux when accept/connect
> @@ -376,12 +381,6 @@ fhandler_socket_wsock::evaluate_events (const long event_mask, long &events,
>         if (erase)
>   	wsock_events->events &= ~(events & ~(FD_WRITE | FD_CLOSE));
>       }
> -  /* Since FD_CONNECT is only given once, we manually need to set
> -     FD_WRITE for connection failed sockets to have consistent
> -     behaviour in programs calling poll/select multiple times.
> -     Example test to non-listening port: curl -v 127.0.0.1:47 */
> -  if ((connect_state () == connect_failed) && (event_mask & FD_WRITE))
> -    wsock_events->events |= FD_WRITE;
>     UNLOCK_EVENTS;
>   
>     return ret;
>
> What do you think?


I already tested your diff successfully, so this could be an alternative 
approach to the issue. I just think the wsock_events->connect_errorcode 
should also only be reset if FD_CONNECT is removed, right? So the if 
branch would need to be extended to include the second line [3] as well.


Everything together, I think our suggestions together would look like this:


diff --git a/winsup/cygwin/fhandler_socket_inet.cc 
b/winsup/cygwin/fhandler_socket_inet.cc
index e5b0d2d14..84cd63698 100644
--- a/winsup/cygwin/fhandler_socket_inet.cc
+++ b/winsup/cygwin/fhandler_socket_inet.cc
@@ -352,10 +352,15 @@ fhandler_socket_wsock::evaluate_events (const long 
event_mask, long &events,
               WSASetLastError (wsa_err);
               ret = SOCKET_ERROR;
             }
-         else
-           wsock_events->events |= FD_WRITE;
-         wsock_events->events &= ~FD_CONNECT;
-         wsock_events->connect_errorcode = 0;
+         wsock_events->events |= FD_WRITE;
+         /* Since FD_CONNECT is only given once, we have to keep FD_CONNECT
+            for connection failed sockets to have consistent behaviour
+            programs calling poll/select multiple times.  Example test to
+            non-listening port: curl -v 127.0.0.1:47 */
+         if (connect_state () != connect_failed) {
+           wsock_events->events &= ~FD_CONNECT;
+           wsock_events->connect_errorcode = 0;
+         }
         }
        /* This test makes accept/connect behave as on Linux when 
accept/connect
           is called on a socket for which shutdown has been called.  
The second
@@ -376,12 +381,6 @@ fhandler_socket_wsock::evaluate_events (const long 
event_mask, long &events,
        if (erase)
         wsock_events->events &= ~(events & ~(FD_WRITE | FD_CLOSE));
      }
-  /* Since FD_CONNECT is only given once, we manually need to set
-     FD_WRITE for connection failed sockets to have consistent
-     behaviour in programs calling poll/select multiple times.
-     Example test to non-listening port: curl -v 127.0.0.1:47 */
-  if ((connect_state () == connect_failed) && (event_mask & FD_WRITE))
-    wsock_events->events |= FD_WRITE;
    UNLOCK_EVENTS;

    return ret;



Best regards,
Marc

[1] https://www.man7.org/linux/man-pages/man2/connect.2.html#ERRORS

[2] 
https://sourceware.org/git/?p=newlib-cygwin.git;a=blob;f=winsup/cygwin/fhandler_socket_inet.cc;h=e5b0d2d1443ecc4430104f6cfb78bf580a8116e5;hb=aa86784937ec7868c358dd90ea5e5324f0be750d#l356

[3] 
https://sourceware.org/git/?p=newlib-cygwin.git;a=blob;f=winsup/cygwin/fhandler_socket_inet.cc;h=e5b0d2d1443ecc4430104f6cfb78bf580a8116e5;hb=aa86784937ec7868c358dd90ea5e5324f0be750d#l358