Hi, Junio C Hamano wrote: > Jonathan Nieder <jrnieder@xxxxxxxxx> writes: >> Marc Branchaud wrote: >>> The new workdir is empty before the checkout, so attempts to recurse into >>> a non-existent submodule directory fail. >>> >>> Signed-off-by: Marc Branchaud <marcnarc@xxxxxxxxxxx> >>> --- >> >> Thanks for reporting. Can you describe the error message when it fails >> here? [...] > IOW, at the point in that script where we call "git checkout -f", if > we changed it to "git checkout --recurse-submodules -f", what breaks > and why? Shouldn't it succeed instead? I think that's a similar question to the one I asked. But I have a good guess about what goes wrong. It's related to the same issue as the "git worktree" problem Marc described. Inside the superproject's $GIT_DIR, we see config modules/ a/ config b/ config ... The question is what to do with the modules/ directory when you have multiple working directories making use of the refs and objects from this $GIT_DIR. In general, the most useful answer is that the additional working directories should make use for modules/ from this $GIT_DIR as well. After all, each submodule has its own refs and objects, and the same motivation that pushes us to share the refs and objects from the superproject would drive us to share them in the submodules as well. However, if you do this in the most naive way, it will not work. In the config file, there is a core.worktree setting that ensures that commands run from a submodule affect the correct working directory. Which worktree should it point to? All of them. There's still an obvious "most useful" answer: each submodule should contain its own worktrees/ directory with metadata specific to each worktree. This should work fine and is the future work that Marc and I alluded to. Let me call it (*), for later reference. Anything done today is papering over the sad truth that that future work (*) has not been done yet. contrib/workdir is currently naive about all this: it does *not* symlink across the modules/ directory, so each workdir gets its own independent copy of all the submodules. Which kind of defeats the point of this kind of setup. That said, it's better than nothing at all, which is why Marc proposes making it not attempt to check out the submodules right away, instead permitting the user to make the best of things. I suppose another thing that is missing is a warning message to ensure the user knows that once (*) arrives, they need to be ready for it. (Or not: this is contrib/workdir, and there would be no need to make use of it in place of "git worktree" once that moment arrives.) To reiterate, this is all about papering over (*) not having been done. Marc, did I understand correctly? Thanks, Jonathan
