ocamlnet: EAFNOSUPPORT on Max OSX

ocamlnet: EAFNOSUPPORT on Max OSX

[Joel Reymont]

Gerd,

Netplex reports EAFNOSUPPORT when starting up on Mac OSX which means  
"Address family not supported by protocol family".

The issue is that in netplex_controller.ml/start_containers you are  
doing

       let (fd_clnt, fd_srv) =
	Unix.socketpair Unix.PF_UNIX Unix.SOCK_STREAM 0 in
       let (sys_fd_clnt, sys_fd_srv) =
	Unix.socketpair Unix.PF_UNIX Unix.SOCK_STREAM 0 in

i.e. creating Unix Domain sockets.

Later on in rpc_transport.ml/stream_rpc_multiplex_controller you are  
doing

     `Sockaddr(Unix.getsockname fd)

which passes, and

     `Sockaddr(Unix.getpeername fd)

which fails with EAFNOSUPPORT.

A Ruby test, on the other hand, ran without a hitch:

irb(main):001:0> require 'socket'
=> true

irb(main):002:0> x = Socket.pair Socket::PF_UNIX, Socket::SOCK_STREAM, 0
=> [#<Socket:0x7daa4>, #<Socket:0x7dab8>]

irb(main):003:0> x.each { |sock| sock.getpeername }
=> [#<Socket:0x7daa4>, #<Socket:0x7dab8>]

	Thanks, Joel

--
http://wagerlabs.com/

Re: ocamlnet: EAFNOSUPPORT on Max OSX

[Joel Reymont]

Here's a summary. Any suggestions?

Mac OSX 10.4.9 Intel

         Objective Caml version 3.10+dev26 (2007-04-25)

# #use "topfind";;
- : unit = ()

# #require "unix";;
/usr/local/lib/ocaml/unix.cma: loaded

# let x, y = Unix.socketpair Unix.PF_UNIX Unix.SOCK_STREAM 0;;
val x : Unix.file_descr = <abstr>
val y : Unix.file_descr = <abstr>

# Unix.getsockname x;;
Exception: Unix.Unix_error (Unix.EAFNOSUPPORT, "", "").

# Unix.getpeername x;;
Exception: Unix.Unix_error (Unix.EAFNOSUPPORT, "", "").

# Unix.getsockname y;;
Exception: Unix.Unix_error (Unix.EAFNOSUPPORT, "", "").

# Unix.getpeername y;;
Exception: Unix.Unix_error (Unix.EAFNOSUPPORT, "", "").

	Thanks, Joel

--
http://wagerlabs.com/

Re: [Caml-list] Re: ocamlnet: EAFNOSUPPORT on Max OSX

[Richard Jones]

On Sat, May 05, 2007 at 01:50:09PM +0100, Joel Reymont wrote:
> # #use "topfind";;
> - : unit = ()
> 
> # #require "unix";;
> /usr/local/lib/ocaml/unix.cma: loaded
> 
> # let x, y = Unix.socketpair Unix.PF_UNIX Unix.SOCK_STREAM 0;;
> val x : Unix.file_descr = <abstr>
> val y : Unix.file_descr = <abstr>
> 
> # Unix.getsockname x;;
> Exception: Unix.Unix_error (Unix.EAFNOSUPPORT, "", "").
> 
> # Unix.getpeername x;;
> Exception: Unix.Unix_error (Unix.EAFNOSUPPORT, "", "").

You might want to run this under ktrace, but the error seems clear
enough from the getsockname man page on Mac OS X:

  BUGS
     Names bound to sockets in the UNIX domain are inaccessible; getsockname
     returns a zero length name.

(That differs from the observed behaviour, but returning an error
seems more probably than returning a zero length name).

In general, I wouldn't expect anonymous Unix domain sockets created by
socketpair to have a sockname (they're anonymous!), but I might expect
them to have some sort of peername - obviously they don't have that
either.

Rich.

-- 
Richard Jones
Red Hat

Mac OSX getsocketpair/getsockname and IPv6

[Joel Reymont]

The test C program [1] runs fine. I also tried extracting the  
relevant C bits from otherlibs/unix/{socket.c, socketaddr.h, etc.}  
and putting them together. Again, everything is fine but the  
following OCaml code still bombs out.

         Objective Caml version 3.10+dev26 (2007-04-25)

# #use "topfind";;
- : unit = ()

# #require "unix";;
/usr/local/lib/ocaml/unix.cma: loaded

# let x, y = Unix.socketpair Unix.PF_UNIX Unix.SOCK_STREAM 0;;
val x : Unix.file_descr = <abstr>
val y : Unix.file_descr = <abstr>

# Unix.getsockname x;;
Exception: Unix.Unix_error (Unix.EAFNOSUPPORT, "", "").

I'm baffled but it definitely has something to do with IPv6 code in  
OCaml.

I'll try to recompile 3.10 without IPv6 support to see if it works  
and then will try to insert printouts into the C wrappers for  
socketpair and getsockname to see what OCaml is giving C with IPv6  
enabled.

	Thanks, Joel

[1] Sample C program:

#include <sys/types.h>
#include <sys/socket.h>
#include <stdio.h>

int main (int argc, char** argv)
{
   int err, sv[2];
   struct sockaddr addr;
   socklen_t addr_len;

   err = socketpair(PF_UNIX, SOCK_STREAM, 0, sv);

   printf("socketpair: %d\n", err);

   addr_len = sizeof(addr);
   err = getsockname(sv[0], &addr, &addr_len);

   printf("getsockname: %d\n", err);

}


--
http://wagerlabs.com/

Re: [Caml-list] Re: ocamlnet: EAFNOSUPPORT on Max OSX

[Joel Reymont]

Rich,

On May 5, 2007, at 1:57 PM, Richard Jones wrote:

>   BUGS
>      Names bound to sockets in the UNIX domain are inaccessible;  
> getsockname
>      returns a zero length name.

The issue is that non-OCaml networking code on my box (C, Ruby) runs  
fine, unlike the OCaml code. Furthermore, even if I extract the C  
code from the OCaml wrappers I still don't get the error.

--
http://wagerlabs.com/

Serious bug in the OCaml FFI?

[Joel Reymont]

I'm using 32-bit Mac OSX 10.4.9 Intel.

I inserted a print-out into my version of socketpair.c [1] and ran my  
OCaml test script

     #use "topfind";;
     #require "unix";;
     let x, y = Unix.socketpair Unix.PF_UNIX Unix.SOCK_STREAM 0;;
     Unix.getpeername y;;

which resulted in

     unix_socketpair: domain = 1, type = 1, proto = 0
     Exception: Unix.Unix_error (Unix.EAFNOSUPPORT, "", "").

Back at the toplevel

     # (Obj.magic Unix.PF_UNIX:int);;
     - : int = 0

     # (Obj.magic Unix.SOCK_STREAM:int);;
     - : int = 0

Now, how exactly does 0 become 1 here? I have no idea but there are  
two tables defined in otherlibs/unix/socket.c:

     int socket_domain_table[] = {
       PF_UNIX, PF_INET,
     #if defined(HAS_IPV6)
       PF_INET6
     #elif defined(PF_UNDEF)
       PF_UNDEF
     #else
       0
     #endif
     };

     int socket_type_table[] = {
       SOCK_STREAM, SOCK_DGRAM, SOCK_RAW, SOCK_SEQPACKET
     };

As you can see, instead of calling socketpair with PF_UNIX and  
SOCK_STREAM), Ocaml is calling it with PF_INET6 and SOCK_DGRAM instead.

This is obviously a major problem. Any suggestions on what may be  
causing it and how to fix it?

	Thanks, Joel

[1] otherlibs/unix/socketpair.c

CAMLprim value unix_socketpair(value domain, value type, value proto)
{
   int sv[2];
   value res;

   printf("unix_socketpair: domain = %d, type = %d, proto = %d\n",
       socket_domain_table[Int_val(domain)],
       socket_type_table[Int_val(type)],
       Int_val(proto));

   if (socketpair(socket_domain_table[Int_val(domain)],
                  socket_type_table[Int_val(type)],
                  Int_val(proto), sv) == -1)
     uerror("socketpair", Nothing);
   res = alloc_small(2, 0);
   Field(res,0) = Val_int(sv[0]);
   Field(res,1) = Val_int(sv[1]);
   return res;
}

--
http://wagerlabs.com/

Re: Mac OSX getsocketpair/getsockname and IPv6

[Joel Reymont]

There are no bugs in the OCaml FFI.

socketpair.c:

   printf("unix_socketpair: domain = %d/%d/%d, type = %d/%d/%d, proto  
= %d\n",
       PF_UNIX,
       Int_val(domain),
       socket_domain_table[Int_val(domain)],
       SOCK_STREAM,
       Int_val(type),
       socket_type_table[Int_val(type)],
       Int_val(proto));

Running the test

     #use "topfind";;
     #require "unix";;
     let x, y = Unix.socketpair Unix.PF_UNIX Unix.SOCK_STREAM 0;;
     Unix.getpeername y;;

produces

     unix_socketpair: domain = 1/0/1, type = 1/0/1, proto = 0
     Exception: Unix.Unix_error (Unix.EAFNOSUPPORT, "", "").

So we _are_ passing the proper values and they become proper values  
as well.

The problem must be somewhere else, possibly in getsockbyname.c.

	Joel

--
http://wagerlabs.com/

Re: Mac OSX getsocketpair/getsockname and IPv6

[Joel Reymont]

As Rich suggested, I ran OCaml with kernel process tracing to see  
what the syscalls return.

28407 ocamlrun CALL  socketpair(0x1,0x1,0,0xbffff268)
28407 ocamlrun RET   socketpair 0
28407 ocamlrun CALL  getsockname(0x5,0xbffff204,0xbffff200)
28407 ocamlrun RET   getsockname 0

28407 ocamlrun GIO   fd 1 wrote 56 bytes
        "Exception: Unix.Unix_error (Unix.EAFNOSUPPORT, "", "").

So both socketpair and getsockname return 0 but OCaml still complains.

My version of getsockname.c below outputs:

     unix_getsockname: sock = 5, len = 0, retcode = 0
     Exception: Unix.Unix_error (Unix.EAFNOSUPPORT, "", "").

Any suggestions?

	Thanks, Joel

---

CAMLprim value unix_getsockname(value sock)
{
   int retcode;
   union sock_addr_union addr;
   socklen_param_type addr_len;

   addr_len = sizeof(addr);
   retcode = getsockname(Int_val(sock), &addr.s_gen, &addr_len);

   printf("unix_getsockname: sock = %d, len = %d, retcode = %d\n",
       Int_val(sock), addr_len, retcode);

   if (retcode == -1) uerror("getsockname", Nothing);
   return alloc_sockaddr(&addr, addr_len, -1);
}



--
http://wagerlabs.com/

Re: [Caml-list] Serious bug in the OCaml FFI?

[Olivier Andrieu]

On 5/5/07, Joel Reymont <joelr1@gmail.com> wrote:
> I'm using 32-bit Mac OSX 10.4.9 Intel.
>
> I inserted a print-out into my version of socketpair.c [1] and ran my
> OCaml test script
>
>      #use "topfind";;
>      #require "unix";;
>      let x, y = Unix.socketpair Unix.PF_UNIX Unix.SOCK_STREAM 0;;
>      Unix.getpeername y;;
>
> which resulted in
>
>      unix_socketpair: domain = 1, type = 1, proto = 0
>      Exception: Unix.Unix_error (Unix.EAFNOSUPPORT, "", "").
>
> Back at the toplevel
>
>      # (Obj.magic Unix.PF_UNIX:int);;
>      - : int = 0
>
>      # (Obj.magic Unix.SOCK_STREAM:int);;
>      - : int = 0
>
> Now, how exactly does 0 become 1 here?

It doesn't, your printf prints the value of the C constants PF_UNIX
and SOCK_STREAM, which apparently is 1.

-- 
  Olivier

Solved! (was Re: Mac OSX getsocketpair/getsockname and IPv6)

[Joel Reymont]

The culprit is in alloc_sockaddr [1].

adr->s_gen.sa_family is 0 which stands for AF_UNSPEC in my /usr/ 
include/sys/socket.h.

alloc_sockaddr does not handle this family type and throws EAFNOSUPPORT.

I'm inclined to think that ocamlnet should be fixed and not the OCaml  
library, although it would be helpful if an error different from  
EAFNOSUPPORT was used.

I would have been most helpful to have a stack trace on exception as  
this would have allowed me to narrow down the source of the problem  
in no time rather than going spelunking through ocamlnet and OCaml  
networking libraries.

	Thanks, Joel

[1] otherlibs/unix/socketaddr.c

value alloc_sockaddr(union sock_addr_union * adr /*in*/,
                      socklen_param_type adr_len, int close_on_error)
{
   value res;
   switch(adr->s_gen.sa_family) {
#ifndef _WIN32
   case AF_UNIX:
     { value n = copy_string(adr->s_unix.sun_path);
       Begin_root (n);
         res = alloc_small(1, 0);
         Field(res,0) = n;
       End_roots();
       break;
     }
#endif
   case AF_INET:
     { value a = alloc_inet_addr(&adr->s_inet.sin_addr);
       Begin_root (a);
         res = alloc_small(2, 1);
         Field(res,0) = a;
         Field(res,1) = Val_int(ntohs(adr->s_inet.sin_port));
       End_roots();
       break;
     }
#ifdef HAS_IPV6
   case AF_INET6:
     { value a = alloc_inet6_addr(&adr->s_inet6.sin6_addr);
       Begin_root (a);
         res = alloc_small(2, 1);
         Field(res,0) = a;
         Field(res,1) = Val_int(ntohs(adr->s_inet6.sin6_port));
       End_roots();
       break;
     }
#endif
   default:
     if (close_on_error != -1) close (close_on_error);
     unix_error(EAFNOSUPPORT, "", Nothing);
   }
   return res;
}

--
http://wagerlabs.com/

Re: [Caml-list] Serious bug in the OCaml FFI?

[Joel Reymont]

On May 5, 2007, at 3:56 PM, Olivier Andrieu wrote:

> It doesn't, your printf prints the value of the C constants PF_UNIX
> and SOCK_STREAM, which apparently is 1.

Yes, mea culpa. The problem turned out to be something else entirely.

	Thanks, Joel

--
http://wagerlabs.com/

Re: [Caml-list] Solved! (was Re: Mac OSX getsocketpair/getsockname and IPv6)

[Gerd Stolpmann]

Am Samstag, den 05.05.2007, 16:16 +0100 schrieb Joel Reymont:
> The culprit is in alloc_sockaddr [1].
> 
> adr->s_gen.sa_family is 0 which stands for AF_UNSPEC in my /usr/ 
> include/sys/socket.h.
> 
> alloc_sockaddr does not handle this family type and throws EAFNOSUPPORT.
> 
> I'm inclined to think that ocamlnet should be fixed and not the OCaml  
> library, although it would be helpful if an error different from  
> EAFNOSUPPORT was used.

Yes, this is true, EAFNOSUPPORT is misleading because it means that the
whole address family is not supported. A better code would be EOPNOTSUPP
which is generally used when a certain socket operation cannot be
performed for the type of socket (and the Single Unix Spec documents it
for getsockname/getpeername). Or Unix.sockaddr is extended by an extra
case for the inaccessible address.


>    default:
>      if (close_on_error != -1) close (close_on_error);
>      unix_error(EAFNOSUPPORT, "", Nothing);
>    }

Gerd
-- 
------------------------------------------------------------
Gerd Stolpmann * Viktoriastr. 45 * 64293 Darmstadt * Germany 
gerd@gerd-stolpmann.de          http://www.gerd-stolpmann.de
Phone: +49-6151-153855                  Fax: +49-6151-997714
------------------------------------------------------------

Re: [Caml-list] Solved! (was Re: Mac OSX getsocketpair/getsockname and IPv6)

[Paul Snively]

So: if a Mac OS X O'Caml user wanted to solve this, what is the  
actual solution? Surely it's not:

case AF_UNSPEC:
case AF_UNIX:

in socketaddr.c?

Many thanks and best regards,
Paul

On May 5, 2007, at 8:16 AM, Joel Reymont wrote:

> The culprit is in alloc_sockaddr [1].
>
> adr->s_gen.sa_family is 0 which stands for AF_UNSPEC in my /usr/ 
> include/sys/socket.h.
>
> alloc_sockaddr does not handle this family type and throws  
> EAFNOSUPPORT.
>
> I'm inclined to think that ocamlnet should be fixed and not the  
> OCaml library, although it would be helpful if an error different  
> from EAFNOSUPPORT was used.
>
> I would have been most helpful to have a stack trace on exception  
> as this would have allowed me to narrow down the source of the  
> problem in no time rather than going spelunking through ocamlnet  
> and OCaml networking libraries.
>
> 	Thanks, Joel
>
> [1] otherlibs/unix/socketaddr.c
>
> value alloc_sockaddr(union sock_addr_union * adr /*in*/,
>                      socklen_param_type adr_len, int close_on_error)
> {
>   value res;
>   switch(adr->s_gen.sa_family) {
> #ifndef _WIN32
>   case AF_UNIX:
>     { value n = copy_string(adr->s_unix.sun_path);
>       Begin_root (n);
>         res = alloc_small(1, 0);
>         Field(res,0) = n;
>       End_roots();
>       break;
>     }
> #endif
>   case AF_INET:
>     { value a = alloc_inet_addr(&adr->s_inet.sin_addr);
>       Begin_root (a);
>         res = alloc_small(2, 1);
>         Field(res,0) = a;
>         Field(res,1) = Val_int(ntohs(adr->s_inet.sin_port));
>       End_roots();
>       break;
>     }
> #ifdef HAS_IPV6
>   case AF_INET6:
>     { value a = alloc_inet6_addr(&adr->s_inet6.sin6_addr);
>       Begin_root (a);
>         res = alloc_small(2, 1);
>         Field(res,0) = a;
>         Field(res,1) = Val_int(ntohs(adr->s_inet6.sin6_port));
>       End_roots();
>       break;
>     }
> #endif
>   default:
>     if (close_on_error != -1) close (close_on_error);
>     unix_error(EAFNOSUPPORT, "", Nothing);
>   }
>   return res;
> }
>
> --
> http://wagerlabs.com/
>
>
>
>
>
> _______________________________________________
> Caml-list mailing list. Subscription management:
> http://yquem.inria.fr/cgi-bin/mailman/listinfo/caml-list
> Archives: http://caml.inria.fr
> Beginner's list: http://groups.yahoo.com/group/ocaml_beginners
> Bug reports: http://caml.inria.fr/bin/caml-bugs