Am 28.06.2014 08:01, schrieb Matthieu Moy:
> Karsten Blees <karsten.blees@gmail.com> writes:
> 
>> I still don't like that the invalidation is done in git_config_set, though, as
>> this is also used to write completely unrelated files.
> 
> I don't get it. It is used to write the config files. Yes, we trigger a
> complete reload instead of just changing this particular value in the
> hashmap, but I do not see "unrelated files" in the picture.
> 

git-cherry-pick, revert: writes '.git/sequencer/opts' (sequencer.c:save_opts)
git-mv <submodule>: writes '.gitmodules' (submodule.c:update_path_in_gitmodules)
  ...and the submodule's '.git/config' (submodule.c:connect_work_tree_and_git_dir)


Note that the typical configuration commands git-config/remote/branch seem
to call git_config_set* immediately before exit, so no need to invalidate
the config cache here either.

Which leaves clone, init and push (transport.c:set_upstreams, seems to be
just before exit as well).

I didn't look further after finding the example in cmd_clone, so with the
statistics above, it may well be the only instance of {git_config; 
git_config_set; git_config}, I don't know.

I've attached the reverse call hierarchy as generated by Eclipse - quite
useful for things like this.

>> Wouldn't it be better to have a 'git_config_refresh()' that could be
>> used in place of (or before) current 'git_config(callback)' calls? The
>> initial implementation could just invalidate the config cache. If
>> there's time and energy to spare, a more advanced version could first
>> check if any of the involved config files has changed.
> 
> That would not change the "xstrdup" vs "no xstrdup" issue, right?
> 

Right. (Unless it turns out invalidation isn't needed after all. But even
then using interned strings for the config would be a Good Thing IMO.)

>> The xstrdup() problem could be solved by interning strings (see the
>> attached patch for a trivial implementation). I.e. allocate each distinct
>> string only once (and keep it allocated).
> 
> That's an option. We need to be carefull not to modify any string
> in-place,

Hopefully an illegal non-const cast will be caught in the review process.