[PATCH] Make glibc sunrpc work with 2.4 unconnected UDP port error handling

[Andreas Jaeger]

>>>>> Jakub Jelinek writes:

 > Hi!
 > This seems to work for me on 2.2 (and I tried a short test which did all the
 > syscalls gethostbyname is doing when trying to talk to portmap which is not
 > running on 2.4). 
 > Maybe I've put too many tests in there, the msg.msg_namelen == 0 test is
 > there because of bug in 2.2 as pointed out by Andi Kleen, ret can be 0 e.g
 > if some router on the way sends host unreacheable. If I get payload back, I
 > compare it to what we just sent to make sure it was ICMP for the same
 > message.
 > I'll try that glibc on 2.4 tomorrow.
 > The problem this patch solves can be reproduced e.g. by making sure nis is
 > before dns in nsswitch.conf and doing
 > domainname xyz.foo.com
 > killall portmap
 > then on 2.4 any hostname lookup will take eons.

 > 2001-01-09  Jakub Jelinek  <jakub@redhat.com>

 > 	* sunrpc/clnt_udp.c (clntudp_bufcreate): Set IP_RECVERR on the
 > 	UDP socket.
 > 	(clntudp_call): Handle MSG_ERRQUEUE.
 > 	* sysdeps/generic/bits/errqueue.h: New.
 > 	* sysdeps/unix/sysv/linux/bits/errqueue.h: New.

 > --- libc/sunrpc/clnt_udp.c.jj	Tue Aug  3 19:20:26 1999
 > +++ libc/sunrpc/clnt_udp.c	Tue Jan  9 18:51:32 2001
 > @@ -51,6 +51,11 @@ static char sccsid[] = "@(#)clnt_udp.c 1
 >  #include <rpc/pmap_clnt.h>
 >  #include <net/if.h>
 
 > +#ifdef IP_RECVERR
 > +#include <bits/errqueue.h>
 > +#include <sys/uio.h>
 > +#endif
 > +
 >  extern bool_t xdr_opaque_auth (XDR *, struct opaque_auth *);
 >  extern u_long _create_xid (void);
 
 > @@ -186,6 +191,12 @@ clntudp_bufcreate (struct sockaddr_in *r
 >        (void) bindresvport (*sockp, (struct sockaddr_in *) 0);
 >        /* the sockets rpc controls are non-blocking */
 >        (void) __ioctl (*sockp, FIONBIO, (char *) &dontblock);
 > +#ifdef IP_RECVERR
 > +      {
 > +	int on = 1;
 > +	setsockopt(*sockp, SOL_IP, IP_RECVERR, &on, sizeof(on));
 > +      }
 > +#endif
cu-> cu_closeit = TRUE;
 >      }
 >    else
 > @@ -361,6 +372,47 @@ send_again:
cu-> cu_error.re_errno = errno;
 >  	  return (cu->cu_error.re_status = RPC_CANTRECV);
 >  	}
 > +#ifdef IP_RECVERR
 > +      if (fd.revents & POLLERR)
 > +	{
 > +	  struct msghdr msg;
 > +	  struct cmsghdr *cmsg;
 > +	  struct sock_extended_err *e;
 > +	  struct sockaddr_in err_addr;
 > +	  struct iovec iov;
 > +	  char *cbuf = (char *) alloca (outlen + 256);
 > +	  int ret;
 > +
 > +	  iov.iov_base = cbuf + 256;
 > +	  iov.iov_len = outlen;
 > +	  msg.msg_name = (void *) &err_addr;
 > +	  msg.msg_namelen = sizeof (err_addr);
 > +	  msg.msg_iov = &iov;
 > +	  msg.msg_iovlen = 1;
 > +	  msg.msg_flags = 0;
 > +	  msg.msg_control = cbuf;
 > +	  msg.msg_controllen = 256;
 > +	  ret = recvmsg (cu->cu_sock, &msg, MSG_ERRQUEUE);
 > +	  if ((ret == 0
 > +	       || (ret == outlen
 > +		   && memcmp (cbuf + 256, cu->cu_outbuf, outlen) == 0))

Andi Kleen told me that ret == 0 doesn't make sense since that would
mean that the error queue is empty.  The outlen check is also wrong
since ICMP can shorten packages.

He advices to use:
if ((ret >= SUNRPC_MIN_HDR && memcpy(cbuf+256,cu->cu_outbuf,outlen)  == 0)

where SUNRPC_MIN_HDR is a minimal value we will use for comparison.

Andreas
-- 
 Andreas Jaeger
  SuSE Labs aj@suse.de
   private aj@arthur.inka.de
    http://www.suse.de/~aj