With earlier versions of RPC and TIRPC it seems that svctcp_create()
calles listen() on the socket (as seen by debugger and strace).
Tooltalk expects this behavior.
However, with newer systems (ArchLinux 5/18+ and similar bleeding edge
versions of SuSE's equivalent: Tumbleweed), this behavior seems to
have changed.
ttsession goes into an infinite loop trying to accept() a connection
in the TIRPC library (via svc_getreqset()).
It appears listen() is no longer called on the socket when it is
created via svctcp_create(). The hack in this commit, always causes
listen() to be called on the socket, and seems to resolve the problem.
But it is a hack I think. I don't know if this is the correct
behavior of svctcp_create() or we were just lucky before.
In init() there was code iterating over all of the possible file
descriptors in a svc_fdset. fdsets are limited to FD_SETSIZE. This
caused coredumps on FreeBSD 10, and possibly other hidden issues.
Moving to poll(), rather than select() would be better, but is a bigger
job. For now, just limit to the FD_SETSIZE that select() requires.
Currently, mp_rpc_server.C tries 538 million ports to acquire an
available transient rpcbind port number. This is bad when rpcbind is
running in secure mode (and you are not using tirpc) - Xsession will
'hang' at the dthello (blue) screen filling up your error logs with
RPC errors.
Now, just try +- 50 (for a total of 100 ports) before bailing. The
dthello 'blue screen of death' is the most common problem in starting
CDE when rpcbind isn't set up properly. This should at least not
cause the appearance of a 'hang'.
This reverts commit 44e384aedb.
This code is actually needed. If svcfd_create() is not available, it
should be fixed only for those systems that it affects.
