On Fri, Dec 2, 2016 at 10:36 AM, Brandon Williams <bmwill@xxxxxxxxxx> wrote: > On 12/01, Jeff King wrote: >> On Thu, Dec 01, 2016 at 01:56:32PM -0800, Stefan Beller wrote: >> >> > > Bleh. Looks like it happens as part of the recently-added >> > > get_common_dir(). I'm not sure if that is ever relevant for submodules, >> > > but I guess in theory you could have a submodule clone that is part of a >> > > worktree? >> > >> > Sure we can, for a test that we don't have that, see the embedgitdirs series. ;) >> > >> > For now each submodule has its own complete git dir, but the vision >> > would be to have a common git dir for submodules in the common >> > superprojects git dir as well, such that objects are shared actually. :) >> >> Fair enough. Given that it seems to behave OK even in error cases, the >> simple stat() test may be the best option, then. I do think we should >> consider adding a few test cases to make sure it continues to behave in >> the error cases (just because we are relying partially on what git's >> setup code happens to do currently, and we'd want to protect ourselves >> against regressions). > > For the naive (ie me), is there a reason why real_path() couldn't be > re-implemented to avoid using chdir? I tried looking into the history of > the function but couldn't find anything explaining why it was done that > way. I assume it has to do with symlinks, but I thought there was a > syscall (readlink?) that could do the resolution. > > -- > Brandon Williams The reason as far as I understand it, is that it uses chdir() to guarantee that it follows symlinks correctly and then looks up the resulting path after the chdir(). I do not think there is a syscall that actually correctly works like real_path() does. You *could* re-write real_path() to do the symlink lookups itself, but as Jeff recently pointed out, that way lies madness. Thanks, Jake