Re: [PATCH v5 25/27] refs: add LMDB refs storage backend

[David Turner <dturner@twopensource.com>]

On Sat, 2016-02-20 at 09:58 +0700, Duy Nguyen wrote:
> > On Fri, 2016-02-19 at 09:54 +0700, Duy Nguyen wrote:
> > > On Fri, Feb 19, 2016 at 3:23 AM, David Turner <
> > > dturner@twopensource.com> wrote:
> > > > > > +static int read_per_worktree_ref(const char *submodule,
> > > > > > const
> > > > > > char
> > > > > > *refname,
> > > > > > +                            struct MDB_val *val, int
> > > > > > *needs_free)
> > > > > 
> > > > > From what I read, I suspect these _per_worktree functions
> > > > > will be
> > > > > identical for the next backend. Should we just hand over the
> > > > > job
> > > > > for
> > > > > files backend? For all entry points that may deal with per
> > > > > -worktree
> > > > > refs, e.g. lmdb_resolve_ref_unsafe, can we check ref_type()
> > > > > first
> > > > > thing, if it's per-worktree we call
> > > > > refs_be_files.resolve_ref_unsafe()
> > > > > instead?  It could even be done at frontend level,
> > > > > e.g. refs.c:resolve_ref_unsafe().
> > > > > 
> > > > > Though I may be talking rubbish here because I don't know how
> > > > > whether
> > > > > it has anything to do with transactions.
> > > > 
> > > > The reason I did it this way is that some ref chains cross
> > > > backend
> > > > boundaries (e.g. HEAD -> refs/heads/master).  But if we have
> > > > other
> > > > backends later, we could generalize.
> > > 
> > > Crossing backends should go through frontend again, imo. But I
> > > don't
> > > really know if it's efficient.
> > 
> > It's pretty tricky to maintain state (e.g. count of symref
> > redirects)
> > across that barrier.  So I'm not sure how to do it cleanly.
> 
> I notice files backend does pretty much the same thing. "files"
> backend looks more like two backends combined in one, one is files,
> the other is packed-refs. And it looks like we could solve it by
> providing a lower level api, read_raw_ref() or something, that
> retrieves the ref without any validation or link following. More on
> this later.

That basica pproach appears to mostly work.  I'll send another series
with read_raw_ref as soon as I'm done applying all comments on this
series.

> > > > > I'm not sure I get this comment. D/F conflicts are no longer
> > > > > a
> > > > > thing
> > > > > for lmdb backend, right?
> > > > 
> > > > I'm trying to avoid the lmdb backend creating a set of refs
> > > > that
> > > > the
> > > > files backend can't handle.  This would make collaboration with
> > > > other
> > > > versions of git more difficult.
> > > 
> > > It already is. If you create refs "foo" and "FOO" on case
> > > sensitive
> > > file system and clone it on a case-insensitive one, you face the
> > > same
> > > problem. We may have an optional configuration knob to prevent
> > > incompatibilities with files backend, but I think that should be
> > > done
> > > (and enforced if necessary) outside backends.
> > 
> > Sure, the current state isn't perfect, but why make it worse?
> 
> I see it from a different angle. The current state isn't perfect, but
> we will be moving to a better future where "files" backend may
> eventually be deprecated. Why hold back?
> 
> But this line of reasoning only works if we have a new backend
> capable
> of replacing "files" without regressions or introducing new
> dependency. Which is why I suggest a new backend [1] (or implement
> Shawn's RefTree if it's proven as good with small repos)
> 
> I have no problem if you want to stay strictly compatible with
> "files"
> though.
> 
> [1] http://thread.gmane.org/gmane.comp.version-control.git/285893/foc
> us=286654

Won't RefTree have the same d/f conflict issue?

> OK how about we keep resolve_ref_1() whole and split real backend
> code
> out? Something like these three patches (only built, did not test). A
> bit ugly with continue_symlink, but it's just demonstration.

I did this a somewhat different way -- will see what you think when I
send it.