Elijah Newren [off-list ref] writes:
unpack_trees. In fact, it traces back to (and before)
fcc387db9b ("read-tree -m -u: do not overwrite or remove untracked
working tree files.", 2006-05-17)
which has additional commentary over at
https://lore.kernel.org/git/7v8xp1jc9h.fsf_-_@assigned-by-dhcp.cox.net/.
It appears that before this time, git happily nuked untracked files
and considered them expendable, in basically all cases. However, this
patch continued considering them as expendable whenever opts->reset
was true.
Thanks for digging. Yes, the 'reset' bit was treated as the license
to kill untracked working tree files and directories that get in the
way in order to carry out the unpack_trees operation.
So, then...should we preserve untracked (and non-ignored) files in all
these cases? This rebase case seems clear, but others might be less
clear....
In short, the guiding principle ought to be that "checkout --force"
and anything that is given "force" as a stronger override should be
allowed to do whatever minimally necessary to match the end result
in the working tree to what the command wants to show in the absense
of these untracked paths. And without being forced, untracked and
unignored paths are precious and should cause commands to fail, if
they need to be touched for the commands to complete what they are
asked to do [*].
"reset --hard HEAD" is an oddball.
Naïvely, because it is often used as the way to tell Git to "no
matter what, match the working tree to HEAD", even though it does
not have an explicit "--force" on the command line, it feels that it
also should be allowed to do whatever necessary to the working tree
files. And historically, that is what we wanted to implement. If
we suddenly made it "safer", I am sure a lot of existing things will
break.
But unfortunately, "--hard" means a bit more than that in the
context of "reset", in that we want to reset in a way that is
different from "--mixed" (reset the index only without touching the
working tree) and "--soft" (do not touch the index or the working
tree), more specifically, with "--hard", we want to reset both the
index and the working tree to match the given committish (often
"HEAD"). From that point of view, "reset --hard" that tries to
preserve untracked and unignored paths, and "reset --force --hard"
that does whatever necessary to untracked and unignored paths to
match the working tree files, when they reset the index and the
working tree to the named committish, may have made sense. If we
were designing the feature without any existing users, it is no
brainer to imagine that our design would: (1) call the three 'reset'
modes as "both", "index-only" and "neither", instead of "hard",
"mixed" and "soft", and (2) require "--force" to touch untracked and
unignored paths.
And I think that may be a reasonable longer-term goal, but since we
have existing users and scripts, we cannot go there overnight without
devising a migration path.
Thanks.
[Footnote]
* We sometimes talk about adding a precious category of paths that
are "ignored", but this discussion is orthogonal to that. Things
that are not tracked and not ignored ought to be precious in
principle.
On Sun, Sep 5, 2021 at 3:31 PM Junio C Hamano [off-list ref] wrote:
Elijah Newren [off-list ref] writes:
quoted
unpack_trees. In fact, it traces back to (and before)
fcc387db9b ("read-tree -m -u: do not overwrite or remove untracked
working tree files.", 2006-05-17)
which has additional commentary over at
https://lore.kernel.org/git/7v8xp1jc9h.fsf_-_@assigned-by-dhcp.cox.net/.
It appears that before this time, git happily nuked untracked files
and considered them expendable, in basically all cases. However, this
patch continued considering them as expendable whenever opts->reset
was true.
Thanks for digging. Yes, the 'reset' bit was treated as the license
to kill untracked working tree files and directories that get in the
way in order to carry out the unpack_trees operation.
quoted
So, then...should we preserve untracked (and non-ignored) files in all
these cases? This rebase case seems clear, but others might be less
clear....
In short, the guiding principle ought to be that "checkout --force"
and anything that is given "force" as a stronger override should be
allowed to do whatever minimally necessary to match the end result
in the working tree to what the command wants to show in the absense
of these untracked paths. And without being forced, untracked and
unignored paths are precious and should cause commands to fail, if
they need to be touched for the commands to complete what they are
asked to do [*].
"reset --hard HEAD" is an oddball.
Naïvely, because it is often used as the way to tell Git to "no
matter what, match the working tree to HEAD", even though it does
not have an explicit "--force" on the command line, it feels that it
also should be allowed to do whatever necessary to the working tree
files. And historically, that is what we wanted to implement. If
we suddenly made it "safer", I am sure a lot of existing things will
break.
But unfortunately, "--hard" means a bit more than that in the
context of "reset", in that we want to reset in a way that is
different from "--mixed" (reset the index only without touching the
working tree) and "--soft" (do not touch the index or the working
tree), more specifically, with "--hard", we want to reset both the
index and the working tree to match the given committish (often
"HEAD"). From that point of view, "reset --hard" that tries to
preserve untracked and unignored paths, and "reset --force --hard"
that does whatever necessary to untracked and unignored paths to
match the working tree files, when they reset the index and the
working tree to the named committish, may have made sense. If we
were designing the feature without any existing users, it is no
brainer to imagine that our design would: (1) call the three 'reset'
modes as "both", "index-only" and "neither", instead of "hard",
"mixed" and "soft", and (2) require "--force" to touch untracked and
unignored paths.
And I think that may be a reasonable longer-term goal, but since we
have existing users and scripts, we cannot go there overnight without
devising a migration path.
Sounds reasonable. I'm presuming that `read-tree --reset` is also an
oddball, but as plumbing it's better to keep its behavior as-is, i.e.
let it nuke untracked files/directories too. Let me know if you would
prefer otherwise.
I have a bunch of relatively small patches
(https://github.com/git/git/pull/1085, if you want a preview), which
fix most of the problems I found. I'm going to split it up into five
patch series, the first three of which are independent and shorter
than the remaining two.