(switched to email.  Please respond via emailed reply-to-all, not via the
bugzilla web interface).

On Wed, 1 Jul 2009 18:39:07 GMT
bugzilla-daemon@bugzilla.kernel.org wrote:

> http://bugzilla.kernel.org/show_bug.cgi?id=13688
> 
>            Summary: backlog parameter of listen is not working with TCP
>            Product: Networking
>            Version: 2.5
>     Kernel Version: 2.6
>           Platform: All
>         OS/Version: Linux
>               Tree: Mainline
>             Status: NEW
>           Severity: normal
>           Priority: P1
>          Component: IPV4
>         AssignedTo: shemminger@linux-foundation.org
>         ReportedBy: florent.cloirec@free.fr
>         Regression: No
> 
> 
> Created an attachment (id=22171)
>  --> (http://bugzilla.kernel.org/attachment.cgi?id=22171)
> little test server
> 
> When creating a server I usually create a socket, bind it to any address and
> call listen to limit the queue size, then I call accept in a loop to handle
> clients one by one.
> 
> In a particular case I wanted to serve only 1 client and refuse the other
> connection attempts. I tried any value for the backlog parameter, whatever I
> use the server is not refusing any connection, though the queue should be
> full.
> The SYN+ACK handshake is completed everytime.
> 
> To be sure this is a bug I checked this with BSD and Windows, I can confirm
> that it refuses connections if the max queue size specified with backlog is
> reached (Note that backlog is not exactly the queue size with these systems,
> there is a 3/2 factor, I don't know why).
> 
> Maybe this implementation for linux kernel is wanted even if it is not posix
> compliant, but I did not find anywhere why this is working like this. If this
> is not a bug, could someone explain me the usefullness of the backlog
> parameter?
> 
> I attached a little test program with which anyone should be able to test
> this
> by connecting multiple telnet to.
> 

#include <stdio.h>
#include <stdlib.h>
#include <sys/types.h>
#include <sys/socket.h>
#include <unistd.h>
#include <fcntl.h>
#include <netinet/in.h>
#include <netinet/ip.h>


int main(void)
{
   int servsock = 0;
   int clientsock = 0;
   struct sockaddr_in servaddr;
   struct sockaddr_in clientaddr;
   socklen_t clientaddrlen;
   struct in_addr client_addr;

   servsock = socket(AF_INET, SOCK_STREAM, 0);

   //fcntl(servsock, F_SETFL, O_NONBLOCK);

   if(servsock != -1)
   {
      servaddr.sin_family = AF_INET;
      servaddr.sin_port = htons(6256);
      inet_aton("0.0.0.0", &client_addr);
      servaddr.sin_addr = client_addr;

      if(bind(servsock, &servaddr, sizeof(servaddr)) != -1)
      {
         if(listen(servsock, 0) != -1)
         {
            while(1)
            {
               printf("waiting for clients\n");
               clientsock = accept(servsock, &clientaddr, &clientaddrlen);
               if(clientsock != -1)
               {
                  printf("client accepted, now sleep 60 sec\n");
                  sleep(60);
                  printf("60 sec have elapsed, closing connection\n");
                  close(clientsock);
               }
               else
               {
                  printf("accept error: retrying in 1 sec...\n");
                  sleep(1);
               }
            }
         }
         else
         {
            printf("listen failed\n");
         }
      }
      else
      {
         printf("bind failed\n");
      }

   }
   else
   {
      printf("socket creation failed\n");
   }


   return 0;
}

Reply-To: fw@strlen.de

Andrew Morton <akpm@linux-foundation.org> wrote:
> On Wed, 1 Jul 2009 18:39:07 GMT
> bugzilla-daemon@bugzilla.kernel.org wrote:
> 
> > http://bugzilla.kernel.org/show_bug.cgi?id=13688
> >            Summary: backlog parameter of listen is not working with TCP
> >            Product: Networking
[..]
> > When creating a server I usually create a socket, bind it to any address
> and
> > call listen to limit the queue size, then I call accept in a loop to handle
> > clients one by one.
> > 
> > In a particular case I wanted to serve only 1 client and refuse the other
> > connection attempts. I tried any value for the backlog parameter, whatever
> I
> > use the server is not refusing any connection, though the queue should be
> full.
> > The SYN+ACK handshake is completed everytime.

The behaviour depends on net.ipv4.tcp_abort_on_overflow sysctl.
If its on, new connections are reset once the backlog is exhausted.

I have tested this and you are right. I think listen man page should be updated because this is not documented. I close the bug.