table; use a safer way of pulling out the lrevid.

pack(): The previous logic was broken, now we first gather up all the
oids of the objects affected by packable transactions, then we cruise
through the metadata records of each of those in a separate loop.

Watch out for the call to c.set_range() on the txnoids table to raise
a DBNotFoundError.  This just means we're trying to access a time
stamp past the last one in this table, so just use c.last() if that
happens, and pack all transactions.

When cruising through the txnoids table, stop when you see the first
PROTECTED_TRANSACTION.  That means that we've already packed back to
this transaction (Jim, is this right?  Would we ever want to re-pack a
transaction that's already been packed?)

fails this.

_zaprevision(): Fixed some silly typos.

pack(): Adopt from FileStorage, the conversion scheme of time floats
to TimeStamps.  Also use set_range() on the txnoids cursor because
it's not likely that the timestamp (i.e. revid) will actually exist,
and we just want the one just before the specified time.  Use prev()
instead of prev_dup() on the cursor.

object of its choosing, so we can retrieve the output of RESPONSE.write().

out duplicate oids.  If a pickle holds multiple references to an
object, that object's oids will show up multiple times in the returned
list.

implementation of the GCable `interface' for integration with a cyclic
garbage detector.

Many small structured text fixes, typos fixes, and small wording and formatting corrections. Some of these were pointed out by Tom Deprez.

context check if there is no user context for some reason)

err_exit():  Allow the return code to be passed in, with a default of 2
    (the usual Unix return code for command-line errors).

main():  Remove the try/except that caused a second err_exit() call after
    the first call to sys.exit().  The try/except also prevented the
    first sys.exit() from actually doing so (only the second, unintended
    call caused the script to exit).

    When the users *asks* for help (-h), use a return code of 0 since
    there was not an error that triggered the help message.

matching and resolve one of PJE's issues in the UserProgrammableSecurityObjects
proposal.

try clause. :(

object, don't rely on its prevrevid pointing to the previous metadata
record.  That's because if the last revision was an undo record, it
will have a prevrevid pointing to the revision previous to the undone
one (phew! :).  Instead, use a cursor to find the previous metadata
record to restore.

supportsTransactionalUndo(): Added.

creating test classes.

    - DNE='nonexist' (it must be 8-bytes and this is easier to pick out)

    - _setupDBs(): use self._nextserial to choose the next serial
      number; this is easier to follow than the timestamps the base
      class uses.

    - _begin(): increment the next serial counter and assign that to
      self._serial

Also,

ObjectDoesNotExist: Removed, we use a straight KeyError instead since
undoInfo() obviates the need for this hack.

_vote(): Removed, base class implements this.

_finish(): At the point we increment the refcounts of the objects
refered to by the pickle, add a FIXME note.  Since objects can be
revised multiple times in a single transaction, we need to make sure
not to increment the refcounts too many times (not tested, so fix is
deferred for later).

abortVersion(): The API requires a VersionError if we try to abort the
empty version.

load(): The API requires that if we try to load an object on a
non-existant version, or a version on which the object hasn't been
changed, that we return the non-version revision of the object instead
of raising an exception.

transactionalUndo(): Fixed (as yet untested) bug in extraction of
pack-protection flag from transaction's metadata.

Rewrote the "revid <> tid" branch to check the commit log for any
promised revisions to the object committed earlier in this
transaction.  It's possible to revise an object multiple times in the
same transaction.  Along those lines, record modified oids in a set
(i.e. dictionary) so we don't get duplicates in the return value.

Use write_object_undo() instead of write_moved_object() so previous
revisions in the same transaction get recorded by the commit log.

undoLog(): The API defines default arguments, and their semantics, so
use them!  Also fix typos in cursor method calls and in unpacking the
user/desc/ext information in the transaction metadata.

storages, so it makes sense to move it from there to the base
class.  _vote() now puts the CommitLog into the promised() state.

the infrastructure.