I don't think this was actually a leak (since smb.msgs should be non-zero
any time mail is non-null), but to make Coverity happy (CID 515659) and just
be more fault-tolerant, always free(mail).

Should be possible to implement the hard stuff now, though it won't
actually be testable until the KEX bits exist too.

Smooth sailing after that though.

If the path already exists (is a directory), don't need to check each path
element for existence.

This should help with the speed of saving of large file.ini (Nelgin's) when a
lot of directories have "Check for File Existence" set to "Yes".

(128 bytes)

Fix up the filename in error messages reported during unpacking QWK packets
(MESSAGES.DAT not BBSID.QWK).

Convert more log messages to use de-duplication as solution to issue #619.

A lot more lprintf() -> errprintf() conversions will be needed to fully
realize this solution (e.g. in the mail, ftp, web, services serverss).

In printfile() and printtail() which are used by menu() and many other
functions.

These should be helpful to sysops that just want to know what file is being
displayed under what conditions (e.g. for trouble-shooting or customization
reasons).

... instead of just the file name, as was expected.

Under heavy loads, I've found periodic non-blocking lock attempts just aren't
reliable enough: in particular, on Vertrauen, the guest user account record
in the user.tab file gets read and written-to A LOT (updating stats), and
under heavy loads (especially with hundreds of web client sessions), I'd get
timeouts attempting to lock the guest record in the user.tab (after about a
minute of periodic retries) - so let's just block "forever" to get the user
record lock (or until signaled or notified of a deadlock). With this change,
I've not had any user.tab lock errors or apparent deadlocks.

Let the underlying OS/filesystem handle queuing and deadlock detection,
when supported. lock() still behaves as before: non-blocking lock attempt.

... from ERROR to WARNING, so the repeats won't be logged to error.log file
and won't (normally) be sent as notifications (e.g. emails) to the sysop. The
duplicates are identified as from the same source file and line number as the
immediately previous logged error within the last 12 hours. String comparison
doesn't really work for deduplication since if you look closely, the errors
usually are *not* exact duplicates (i.e. there's a node number or a socket
descriptor or something that uniquely identifies the user/client/session).
Repeated errors don't increment the node's error counter either.

This does not de-duplicate errors logged via other means (e.g. direct calls to
log()/lputs(),lprintf() with a severity of LOG_ERR or higher), but will
solve the majority of duplicate errors that can be logged from the terminal
server.

So this at least partially addresses issue #619.

Next step is to implement a KEX.

With this, SyncTERM and SexPOTS on *nix won't actually need to strip
received bytes themselves.  No idea what Windows does though, so
it'll likely still be needed forever.

The plan is to hide non-standard stuff in here.  Currently, just
has [[maybe-unused]] from C23.

The intent is that this will be a library implementation of SSH
which allows simple extension of the base protocol, and exposes all
defined features of the protocol (including ones nobody uses).

This will support both client and server, with initial focus on the
client implementation for use in SyncTERM.

The goal of the core library is to only support the most secure
standard methods.  When there are multiple methods that are not
clearly of differing levels of security, the most common two will
be implemented (ie: RSA and Ed25519).

Additionally, care will be taken during development to allow for
the OpenSSH Encrypt-then-MAC MAC modes in the expectation that
someone will standardize an ETM mode at some point, and I expect
to write an extension to support it soon after DeuceSSH is
complete.

Finally, if I cannot mitigate Terrapin without non-standard
extensions, the OpenSSH "strict KEX" extension will be supported
until there's a standard defined that solves the issue.

Even though we use O_EXCL, networked file systems still sometimes allow the
secondary open() to succeed. We use record locking in xpdev's sopen()
implementation for *nix, so this lock will insure such opens (and locks) will
now fail.

This is an attempt to address occassional error on Vertrauen:
ERROR 2 (No such file or directory) renaming *.rep to *.bad

Return correct error value upon seek failure.

When OFD locks are available, that's what we should be using.
Otherwise, we suffer the horrible behavior of POSIX file/region locks and
a subsequent open/close of the file releases any/all locks on it.

This is currently in an !BSD block, which appears to include macOS, but
macOS *does* support OFD locks, so I'll be fixing that here shortly.

"performing custom action"

Normally, the sysop would set the NodeActionCustom text.dat string
to override this.

This is to be used (e.g. instead of NODE_BXFR) for sysop-custom node actions.

so that if multiple entities call these functions at the same time, they
don't retry at exactly the same times.

Add/use SOCKET_STRERROR which gets the description of the *last* socket
error.

We should start using socket_errno(true) in place of ERROR_VALUE (horribly
named macro) for comparing last socket error nubers against values/macros
from errno.h.

Eliminate get_socket_errno(), unused and redundant with socket_errno().

Really need to make the option order (arity) neutral