On Wed, Sep 01, 2021 at 05:40:22PM -0400, Taylor Blau wrote: > Hmm. As I was wondering about about, this is more complicated than meets > the eye. Consider t5616.36, which tests that repacking does not loosen > promisor objects. > > In builtin/repack.c:repack_promisor_object(), the repack builtin tells > pack-objects about the pack that it just wrote with `--keep-pack` (and > we rely on that working in order to not loosen all of the objects that > we just wrote). > > Except when we iterate through `get_all_packs()`, we don't see the keep > pack yet, because it is still prefixed with .tmp. > > So, this does get kind of tricky. There are some internal callers that > do want to know about .tmp packs and a whole host of other callers that > don't or shouldn't. Maybe that should point us towards "we should be > more careful about the order we write packs in, even temporary ones". So I happened to be looking at some packing stuff again today, and I realized the situation is much less dire than I made it out to be. On the reading side, we _won't_ look at tmp_pack_*, nor tmp_idx_*, because they don't end in '.pack' or '.idx'. So really, the only confusing case is the ".tmp-$$-etc.idx" that is generated by pack-objects (because it thinks of that as the final name to use itself), and then later rearranged by git-repack. So I think we _are_ OK as long as the correct order is observed during the rename-into-place steps. I.e., your original patch (plus the extra post in index-pack) would mean that we're fully covered. The .tmp-*.idx ones would appear in the correct order due to pack-objects, and the final pack-*.idx ones due to git-repack. -Peff
