[PATCH] Support aborting the atomic context

[Floris Bruynooghe]

On Mon 15 Jun 2020 at 07:35 -0300, David Bremner wrote:

> Floris Bruynooghe <flub at devork.be> writes:
>
>> This is an implementation of what was suggested in
>> id:87v9k96xtl.fsf at powell.devork.be It closes the database as that is
>> the only safe way to do this afaik.
>>
>> Currently when the database is closed there are still a bunch of
>> operations which can result in segfaults.

I realise that this is perhaps not a very helpful comment.  I'll see if
I can narrow this down into a proper bug report.

>> Yet the API also promises
>> that some operations are still valid, basically those which only
>> access already previously retrieved information.  It would be nice to
>> find a good solution for this in the python bindings, but I don't yet
>> know what this would be.
>
> There is a Xapian method WritableDatabase::cancel_transaction(), but it
> is not currently exposed by notmuch.  I think it could be added, but
> getting the testing working right is not something I want to tackle at
> this point in the release cycle.  I guess this should be upwardly
> compatible, as all code correct for your current implementation should
> still work if we replace notmuch_database_close with
> notmuch_cancel_atomic. Thoughts?

I agree with your reasoning here that such a change would be upwardly
compatible.  Though I can think of some really convoluted scenarios
where this could be false, e.g. after aborting a transaction code could
rely on handling NOTMUCH_STATUS_XAPIAN_EXCEPTION it gets from a the
closed database.  This seems so much a corner case though, that I'd be
willing to ignore it.