14 March 1999

Hi,

This is a patch for dhcpcd-1.3.16 that prevents it from hanging the Linux
PCMCIA "cardmgr" when cardmgr uses it to bring up a new network
interface.

I am using v3.0.9 of the Linux PCMCIA package on a Toshiba Libretto 70
running RedHat 5.2, plus linux kernel 2.2.3, and the half-a-dozen or
so packages that RedHat released to make RH 5.2 work under Linux
2.2.3.  Those RedHat upgrades include dhcpcd v1.3.16.  (As an aside --
the old dhcpcd, version 0.70, no longer works under Linux 2.1/2.2
because those kernels do not allow network interfaces bound to
0.0.0.0).

I noticed that cardmgr was hanging every time I inserted a network
card that was using DHCP to get an IP address.  When I say "hanging",
I mean cardmgr was unresponsive to all future events -- both a removal
of the card that caused the hang, and insertions of other cards in the
same slot or other slots.  Finally, I tracked this down to the
following:

1- cardmgr uses popen() to run the card's initialization script, such
   as "network start eth0".  It does this so it can read the output of
   the script and copy it to the syslog.
2- cardmgr keeps reading from the pipe until it receives an EOF on it.
   In other words, cardmgr keeps reading the output of the "network"
   script, and does not respond to any other card events, until the
   pipe closes.
3- For interfaces that use DHCP to get an IP address, the "network"
   script launches dhcpcd.  dhcpcd forks and turns itself into a
   daemon that runs forever, so it can renew your IP lease.  The
   dhcpcd inherits the pipe that network has open to cardmgr.
4- dhcpcd never closes its standard output, so cardmgr never sees an
   EOF on its pipe.  Even though the "network" script itself terminates,
   the forked dhcpcd still has an open pipe to cardmgr.

Since cardmgr never sees EOF, it hangs -- blocked forever waiting for
output from a process that will never generate any.  You can unhang it
by killing dhcpcd manually, but this is annoying.

The patch below is a simple fix to dhcpcd that has dhcpcd close stdin,
stdout, and stderr after it forks.  The patch works in the simple case
where dhcpcd succeeds in getting your IP address, forks off a daemon,
and then exits.  But if you can't see a DHCP server, dhcpcd never
forks and the cardmgr will hang until dhcpcd times out and dies.

This solution isn't perfect - my next project is to patch cardmgr so
that it forks every time it needs to run a script, so that it doesn't
let a hung script block it from providing service to other sockets.

The patch is against dhcpcd 1.3.16; or, if you don't want to compile
it (and trust me to compile a safe root-run binary for you), I put an
i386 binary on my ftp site at

ftp://ftp.circlemud.org/pub/jelson/dhcpcd.


Regards,
Jeremy

diff -u dhcpcd-1.3.16/dhcpcd.c dhcpcd-1.3.16-patched/dhcpcd.c
--- dhcpcd-1.3.16/dhcpcd.c	Sat Nov  7 18:07:43 1998
+++ dhcpcd-1.3.16-patched/dhcpcd.c	Sun Mar 14 21:41:24 1999
@@ -178,6 +178,13 @@
 #ifndef DEBUG
   if ( fork() ) exit(0); /* got into bound state. */
 #endif
+  /* patched by Jeremy Elson <jelson@circlemud.org> 14 Mar 1999 - 
+   * close stdin, stdout, and stderr so that programs such as the PCMCIA
+   * cardmgr who are running us with a pipe don't read from the daemon
+   */
+  close(STDIN_FILENO);
+  close(STDOUT_FILENO);
+  close(STDERR_FILENO);
   chdir("/");
   writePidFile();
   do currState=(int (*)())currState(); while ( currState );