On Fri, Jul 27, 2018 at 7:31 PM Stefan Beller <sbeller@xxxxxxxxxx> wrote: > > On Fri, Jul 27, 2018 at 10:19 AM Brandon Williams <bmwill@xxxxxxxxxx> wrote: > > > > On 07/27, Duy Nguyen wrote: > > > On Fri, Jul 27, 2018 at 2:40 AM Stefan Beller <sbeller@xxxxxxxxxx> wrote: > > > > > > > > Currently the refs API takes a 'ref_store' as an argument to specify > > > > which ref store to iterate over; however it is more useful to specify > > > > the repository instead (or later a specific worktree of a repository). > > > > > > There is no 'later'. worktrees.c already passes a worktree specific > > > ref store. If you make this move you have to also design a way to give > > > a specific ref store now. > > > > > > Frankly I still dislike the decision to pass repo everywhere, > > > especially when refs code already has a nice ref-store abstraction. > > > Some people frown upon back pointers. But I think adding a back > > > pointer in ref-store, pointing back to the repository is the right > > > move. > > > > I don't quite understand why the refs code would need a whole repository > > and not just the ref-store it self. I thought the refs code was self > > contained enough that all its state was based on the passed in > > ref-store. If its not, then we've done a terrible job at avoiding > > layering violations (well actually we're really really bad at this in > > general, and I *think* we're trying to make this better though the > > object store/index refactoring). > > > > If anything I would expect that the actual ref-store code would remain > > untouched by any refactoring and that instead the higher-level API that > > hasn't already been converted to explicitly use a ref-store (and instead > > just calls the underlying impl with get_main_ref_store()). Am I missing > > something here? > > Then I think we might want to go with the original in Stolees proposal > https://github.com/gitgitgadget/git/pull/11/commits/300db80140dacc927db0d46c804ca0ef4dcc1be1 > but there the call to for_each_replace_ref just looks ugly, as it takes the > repository as both the repository where to obtain the ref store from > as well as the back pointer. > > I anticipate that we need to have a lot of back pointers to the repository > in question, hence I think we should have the repository pointer promoted > to not just a back pointer. I will probably need more time to study that commit and maybe the mail archive for the history of this series. But if I remember correctly some of these for_each_ api is quite a pain (perhaps it's the for_each version of reflog?) and it's probably better to redesign it (again talking without real understanding of the problem). -- Duy
