On Mon, May 4, 2020 at 8:03 PM Han-Wen Nienhuys [off-list ref] wrote:

quoted

All this to say - it's hard to convince myself this is a safe change,
and I'd really like to read more to understand why you made it.

I'll be the first to admit that I'm on shaky ground here. However,
given that the test suite passes, if this is breaking some behavior,
it's probably not very well tested behavior.

Here is what I know:

Git stores refs in multiple places:
- normal refs, packed: .git/packed-refs
- normal refs, loose: under refs/*/
- special refs: HEAD, ORIG_HEAD, REBASE_HEAD.

Currently, the special refs can only be read with refs_read_raw_ref().
If you iterate over the refs in the files/packed backend, you can
never find HEAD, ORIG_HEAD etc.

A large source of test failures of the reftable patch series is the
handling of pseudo refs (HEAD, CHERRY_PICK_HEAD etc.)

These are written as

  if (ref_type(refname) == REF_TYPE_PSEUDOREF) {
    assert(refs == get_main_ref_store(the_repository));
    ret = write_pseudoref(refname, new_oid, old_oid, &err);
  } else {
    t = ref_store_transaction_begin(refs, &err);
    ..

write_pseudoref writes a file directly to the .git/ directory.

However, on read (for example, CHERRY_PICK_HEAD), the pseudo ref is
given to lookup_commit_reference_by_name, which in the end calls
refs_resolve_ref_unsafe(). If reftable is active, the pseudo ref isn't
read from .git/CHERRY_PICK_HEAD.

If the pseudo refs are read through a ref backend, shouldn't they be
written through a ref backend too?

This causes all rebase and all cherry-pick tests to fail with reftable.

-- 
Han-Wen Nienhuys - Google Munich
I work 80%. Don't expect answers from me on Fridays.
--
Google Germany GmbH, Erika-Mann-Strasse 33, 80636 Munich
Registergericht und -nummer: Hamburg, HRB 86891
Sitz der Gesellschaft: Hamburg
Geschäftsführer: Paul Manicle, Halimah DeLaine Prado