A temporary signal handler for SIGTERM is set up in server_setup()
that calls exit() from within a pure signal handler. This causes a
race condition where it's possible that if the SSSD is restarted
immediately while it is still initializing the sysdb caches for
the first time, it can leave the cache in a corrupt and unusable
state.

https://bugzilla.redhat.com/show_bug.cgi?id=658444

There is a race condition where if we send a SIGTERM before the
kernel has scheduled the child, it may be lost, and the child will
not terminate and will leave the monitor hung in wait().

This patch alters this behavior so that we will send the SIGTERM
again every 10ms and check the wait() in a nonblocking manner.

Previously, there was a race-condition where the monitor might
terminate before its children.

In several places, we were creating a new timer and assigning it
to the tev variable, but then we were checking for NULL from the
te variable (which, incidentally, is guaranteed never to be NULL
in this situation)

https://fedorahosted.org/sssd/ticket/523

We need to guarantee at all times that reads and writes complete
successfully. This means that they must be checked for returning
EINTR and EAGAIN, and all writes must be wrapped in a loop to
ensure that they do not truncate their output.

If read() returns with errno set to EINTR -1 is added to total_len.

This patch removes some tab-indentations from pamsrv.c, too.