I thought using sizeof() comparison to compare the types was clever,
but it doesn't work on void pointers, as sizeof(void) is illegal.

In this case, "work" means "fail to compile cleanly".

Don't implicitly cast away const in the tests.

In particular, avoid doing arithmetic on NULL; decay the function to a
pointer using a trinary conditional instead.

It's common to want to create a struct tdb_data which contains a const
pointer, so add a helper to do so, and use it to clean up other -Wcast-qual
warnings.

Samba uses -Wcast-qual to warn about casting away const.  The kind of
coders who care about const correctness are likely to turn that option
on, so we should make sure it's compatible with cast_const().

The foreach garbage collection assumed that iterators were all on the
stack, but they could be on the heap or a global (or static) variable.
We can prevent the heap case by tricky use of macros to complain on
any iterator which isn't a single token, but we can't prevent
globals/statics.

So, if an iterator already seems to be "off" the stack, mark it as
such and simply never free it.

David Gibson noted that foreach requires more than one argument.
There's no particularly good reason for this, other than lack of
imagination on my part.

This makes transition from tdb1 much simpler.

As per TDB1, except that the sequence number is 63 bit.

As per TDB1, with one enhancement: a non-NULL name argument passed to
tdb_open() with the TDB_INTERNAL flag is preserved so you can identify
internal TDBs too.

We don't need this now, since we use explicit error numbers.

Everyone writes their own, so provide a standard one.

These are the same as the TDB1 functions; but note that there is no reliable
way to tell if tdb_exists() fails due to an error.  This simplifies the API,
but means you have to use tdb_fetch() if you really care. 

Finally, we enable sharing of the struct tdb_file.  We add a reference count
so we can free the struct tdb_file when the last tdb is closed.

The last step before sharing struct tdb_file is to add a field to
identify which tdb context is holding each lock, so we can fail when
they conflict.

Internal reshuffle which makes things shorter and neater.

It makes sense to share the mmap between multiple openers.

This moves the fd and locking information into a new 'struct tdb_file',
opening the way for it to be shared by multiple tdb_open calls on the
same file.

Simply tidying up.

As detailed in doc/design.lyx section 2.16 "Record Headers Are Not
Expandible", we make sure that if there is padding at the end of a record,
the first byte of padding is a zero.

As detailed in doc/design.lyx section 2.15 "Extending The Header Is
Difficult"; we add features_used and features_offered fields to the
header, so we can identify if we add new features, and then if someone
opens it who doesn't understand that feature.

This make sure that the data being handed to the function in tdb_traverse
is the correct type.

This is vitally important for forwards compatibility.

Two missed cases in b2100462 (tdb2: change API to return the error
value.)