On Thu, Mar 29, 2018 at 7:50 PM, Jeff King <peff@xxxxxxxx> wrote: > On Thu, Mar 29, 2018 at 07:02:21PM +0200, Duy Nguyen wrote: > >> On Wed, Mar 28, 2018 at 7:42 PM, Jeff King <peff@xxxxxxxx> wrote: >> > When we change to the top of the working tree, we manually >> > re-adjust $GIT_DIR and call set_git_dir() again, in order to >> > update any relative git-dir we'd compute earlier. >> >> Another way to approach this problem is not delaying chdir() at all. >> We have to delay calling setup_work_tree() and not do it in >> setup_git_directory() because it can die() when chdir() fails. But in >> many cases, the command does not actually need the worktree and does >> not deserve to die. But what if we make setup_work_tree be gentle? >> >> If it successfully chdir() at the end of setup_git_directory() (and >> perhaps before the first set_git_dir call), great! The problem we're >> dealing here vanishes. If it fails, don't die, just set a flag. Later >> on when a command requests a worktree, we can check this flag and now >> can die(). >> >> It's less code this way, but it uses up more of your (or my) time >> because even though the first set_git_dir() call actually happens at 8 >> places. Is it worth trying ? > > I do kind of like that. I'm reasonably happy with the chdir_notify() > interface, but it would be nicer still if we could get rid of it in the > first place. It's true that we _could_ chdir from other places, but There's another place we do, that I should mention and keep forgetting. Our run-command.c code allows to switch cwd, and if $GIT_DIR and stuff is relative then we should reparent them too just like we do with setup_work_tree(). Your chdir-notify makes it super easy to support that, we just need to move the prep_childenv() down below chdir(). But since nobody has complaint, I suppose that feature is not really popular (or at least not used to launch another git process anyway) > realistically speaking, we do our one big chdir as part of the setup > process. -- Duy
