On Thu, May 16, 2019 at 04:15:09PM -0700, Jonathan Tan wrote:
> > /*
> > * Second pass:
> > * - for all non-delta objects, look if it is used as a base for
> > * deltas;
> > * - if used as a base, uncompress the object and apply all deltas,
> > * recursively checking if the resulting object is used as a base
> > * for some more deltas.
> > */
>
> I haven't seen any code that contradicts this comment. And looking at
> the code, for each non-delta object, I think that all deltas are checked
> - regardless of whether they appear before or after that non-delta
> object. (find_ref_delta() does a binary search from 0 to
> nr_ref_deltas, calculated in parse_pack_objects() which happens before
> any resolution of deltas.)
>
> And find_unresolved_deltas_1() (called from resolve_deltas() indirectly)
> sets the real_type when it resolves a delta, as far as I can tell.
>
> So there is more than one "resolve deltas" step - resolve_deltas() and
> then fix_unresolved_deltas(). The pre-fetching happens only during the
> latter.
OK, that is better than I realized. But I think there is still one
interesting case: what happens if a thin delta is used as the base for a
non-thin delta (even one that is OFS_DELTA)?
We cannot handle that second delta until the third pass in
fix_unresolved_deltas(), because we do not realize we have the base
until we resolve the thin delta.
Specifically, I wonder:
- do we pre-fetch it, not realizing that we will soon have the base?
That's not ideal, but I think is necessary if we pre-fetch (and is
still a worst case even if we did single on-demand fetches).
- will we ever append a presumed-thin base to the pack, only to later
realize that we already have that object, creating a duplicate
object in the pack? If so, do we handle this correctly when
generating the index (I know we've had issues in the past and have
expressly forbidden duplicates from appearing in the index; even
having a duplicate in the pack stream itself is non-ideal, though,
as it screws up things like on-disk size calculations).
Because of the sorting in fix_unresolved_deltas(), I think this
could easily be prevented if the non-thin delta is OFS_DELTA (by
just checking for the base in our already-found list of objects
before we call read_object_file(). But for REF_DELTA, I think we
have no way of knowing that appending is the wrong thing (and no
good way of backing it out afterwards).
-Peff
