Apparently tring to us the same estrdup copy twice could cause a
stale-pointer bug later in initialization.

This had a bad interaction with the async-DNS code that leads to
sporadic crashes. There are some possible fixes that add code
complexity in Classic, but the right thing to do is (a) notice that
this call is pretty useless on modern systems, which swap so seldom
that many people don't bother with swap partitions, and (b) simplify
by dropping the code.

One good reason to retain this might have been crypto security.  But
our crypto geek says: "Some people like to use memory locking to make
sure that intermediate results of crypto computations don't get
written to swap and sit there long-term. But that's only useful if
you're not already storing keys on disk (rather than on, say, an HSM),
and we don't support any alternative to that."

This used t be done because Linux had a memory limit sometimes smaller than
ntpd's working setup, but that is no longer true - probably changed in the
64-bit transition.

The smoke test for this is that local refclock declarations didn't break.

...in the generic driver.

This should have gone away with the clocktypes array.

...it requites a streams module from an obsolete Sun version.

Finger error from the as2201 -> austron name change.

...there was too much peer state that never got updated.  We're going
to have to take more of an outside-view approach, capturing raw packet reads
and circumventing the poll-interval machinery.

This patch begins to clear the decks, getting rid of intercept_receive -
which will be replaced by a packet-read-level intercept.  Fortunately, most
of the other machinery will not need to change.

A step towards complete socket mocking for capture/replay.

This is for the reclock syntax; "mode" is still recognized as a synonym,
abd still used in the NMEA driver where that optoon controls sentence
mix and baud rate.

A good deal of internal complexity (notably the mode 7 code and
autogen) has been shed since that was written.  With refclocks
configured out, the working set is actually pretty modest by 2016
standards.

...it can't find sys/capability.h on linux.

I've turned this into a more general check and moved it to
runtime. This gives operators the ability to recover by leaving off
the -u and -i options.

This commit is a bit of a grab-bag.  First, it adds a build option
--enable-classic-mode that can be used to prevent the new behaviors
the rest of the patch implements.

Then, it modifies some logging functions so that the
clock-identification fields in them become composed from the driver
shortname string and the unit number, rather than being a 127.127.t.u
address.  The --enable-classic-mode option *restores* the old behavior.

Finally, it modifies the documentation to reflect the new behavior.

...that scrolled off my screen before I could see it first time.

We need to back out my logging-related changes until we figure out
what earlier change broke peerstats logging.

We need to back out my logging-related changes until we figure out
what earlier change broke peerstats logging.

Adding "subtype" to the token table as an alias for "mode" apparently confuses
the crap out of the scanner generator.  The symptom of this is strangely
specific - the "peerstats" keyword is no longer recognized, breaking logging.

This reversion restores things to a good state.  Next, figure out out why
the scanner generator is losing its merbles.

Looks like the scanner doesn't like it when you try to alias multiple strings
to one value.

Logging live-tested at tis revision works.

Restores some changes earlier reverted.

...so it can have access to the procptr member of the clock. We'll
use this in a later change.

Live-testing clockstats logging works at this revision.

...so they have access to peer->procptr. We'll use this in a later
behavior change.

Logging of peer, loop, clock, and raw statistics works live at this
revision.

The --enable-classic-mode build switch restores the old behavior.

All log channels work live (with new clock IDs) at this revision.

The changes to banish magic driver addresses made it possible to pass the
generic driver unit numbers above 3.

For some odd reason the code for parsing server clauses created a secondary
FIFO of attribute/value nodes to be composed into flag bits later, rather
than simply composing the flag bits immediately.  It worked, but why?

According to clockctl(4), if the owner of a non-root device has access to
/dev/clocktl, that process will be be able to set and slew the clock
through all the time-related system calls normally availabke to root.

What we do here is presume that the ntp user has that access.  Thus
we refrain from bailing out after dropping root.

...into ntp_config.c.  No code changes, this is just better containment.

Resolved merge conflicts surrounding removal of xleave from the
parser, a change to the signature newpeer(), and removal of the
is_network_packet() function.