In the last week-end I've spent many hours trying to get PLIP working
between two out-of-the-box RedHat 6.1 machines. It took a while to
understand (at least partially) what wasn't working, so here is my
diagnosis (which applies also to bug 8523):
apparently the PLIP driver was written thinking that the
fake MAC addresses and ARP are completely useless to connections
between Linux machines, and are to be implemented only for
compatibility with Windows. This is _not true_. Not only PLIP won't
work without ARP (and you must specify it calling ifconfig, as by
default it is turned of by the initialization code), but the PLIP code
is broken and generates the same MAC address for the source and the
target fake ethernet card under certain conditions. The guilty code is
in plip_open() around line 1000 (my source version is 2.2.12):

      struct in_ifaddr *ifa=in_dev->ifa_list;
      if (ifa != NULL) {
         memcpy(dev->dev_addr+2, &ifa->ifa_local, 4);
         memcpy(dev->broadcast+2, &ifa->ifa_address, 4);
      }

I don't know the details of the ifa_addr structure, but in my case
(and at least in another case I discovered on the news) ifa_local and
ifa_address are equal (to the IP address of the interface). Since PLIP
uses dev->dev_broadcast to keep the receiver MAC address, the packets
are sent around with the same source and destination MAC address: as a
result, they travel but they are not seen by inetd.

I patched the kernels of the two machines forcing the last byte of
dev->broadcast to be the right one (i.e., the last byte of the IP
address of the _other_ machine), and everything works fine. Of course,
due to my lack of knowledge of the internals of the networking system
I could be fixing the problem in the wrong place...

All in all, I suggest two fixes:

1) In plip_init_dev(), do NOT set IFF_NOARP
2) Fix the code above so to create correct fake MAC addresses (I
   suppose someone can guess how to do it right)

There are also some other problems. The PLIP.txt file in the
Documentation directory claims that PLIP is able to transfer data with
the port set in "bidirectional" mode. Unfortunately the docs does not
claim _which_ type of bidirectional port (PS/2? EPP? ECP?). From the
suggested cabling it seems like it's not ECP. In any case, no such
support appears in plip.c, so probably it should be eliminated from
the docs, too. BTW it's a pity, because ECP would be 10x faster than
the current implementation.