[Request] Git reset should be able to ignore file permissions

Recently I had to write some automation scripts and I found that git reset --hard actually restores each file's permissions.

That is causing both the created and the last-modified dates of the file to get changed to the time of the git reset.

This behavior is easy to demonstrate:

echo "test" > myfile chmod 777 myfile git add myfile && git commit -m "Test" && git push chmod 775 myfile git reset --hard origin/master

After the git reset --hard command, the entire file was checkout-ed. Instead, git should be able to check if the content of the file changed and only if it did, check it out.

I do realize that checking the content of each file in a big repo could result in a slow operation, but there should be a switch/argument/option to make git reset actually check the content of each file instead of blindly replacing it.

After reading man a few times I didn't saw any option that'd let me do this; the only solution I'm able to think about is actually restoring the permissions of each file to the ones git thinks they should have before doing the git reset.

Maybe I'm wrong and there is a way for doing what I want, if so, please correct me. But if there isn't, should this be implemented? Are there any reasons for not doing it?

Thank you for your attention Regards

-- alexandernst

Alexander Nestorov <alexandernst@gmail.com> writes:

This doesn't tell what the permissions are in origin/master.

If the last line was "git reset --hard HEAD", then it wouldn't touch myfile (it's executable in the worktree and in HEAD, so Git doesn't need to change it). Neither the x bit, nor the ctime or mtime.

If you reset the file to a point where it was not executable, then Git changes its executable bit, and I don't see why it would do otherwise: Git tracks the executable bit, so when you say "reset the file to how it was in this revision", this includes the content and executability.

Reading your message, I don't understand why you need to be able to ignore the x bit.

Git does preserve file permissions, that is, git is aware of the permissions you can set with chmod.

I'm not trying to ignore the x bit, what I'm trying to do is make "git reset" checkout only the files that actually changed instead of checking out all the files with different permissions than the ones git thinks they should have.

Said with other word: when you run "git reset", git does a "status" and checkouts all the files that showed up from the "status". That's exactly what I'm trying to avoid, as "status" is aware of both content changes and permissions changes.

Alexander Nestorov <alexandernst@gmail.com> writes:

Ah, OK, you want "git reset --hard" to just do a chmod, which would not touch the file's mtime (but only the ctime).

Then, it's even easier to demonstrate: just "touch" instead of chmod. Indeed:

$ touch myfile; sleep 2 $ strace -f git reset --hard 2>&1 | grep myfile lstat("myfile", {st_mode=S_IFREG|0755, st_size=5, ...}) = 0 lstat("myfile", {st_mode=S_IFREG|0755, st_size=5, ...}) = 0 unlink("myfile") = 0 open("myfile", O_WRONLY|O_CREAT|O_EXCL, 0777) = 4

(sleep 2 is needed in the demonstration to avoid the "racy git" safeties, but it's not really important)

Git doesn't even try to read the file content: once it detected that the stat information changed, it rewrite the file without looking at its content. It's faster this way for files that actually changed.

No, it's indeed the opposite: "status" re-checks the content of changed files, and update the stat-cache in the index accordingly if the content actually didn't change.

Runing "git status" before "git reset --hard" should solve your problem. The part of "git status" of interest is "git update-index --refresh":

$ touch myfile; sleep 2 $ git update-index --refresh $ strace -f git reset --hard 2>&1 | grep myfile lstat("myfile", {st_mode=S_IFREG|0755, st_size=5, ...}) = 0 $

Indeed, "git update-index --refresh" before "git reset" did the trick :) Anyways, what about the proposal? Should it be implemented?

Thank you

[ The rule here is to keep everyone Cced, not just git@vger ]

Alexander Nestorov <alexandernst@gmail.com> writes:

I'd say the current behavior is OK by default (it's not so common to have stat-only changes, and re-checking the content would slow down the other operations).

We could add a flag to "git reset" asking it to refresh the index before continuing, but that would essentially end-up doing "git update-index --refresh; git reset".

It would make more sense to me to add something about update-index in the documentation of git reset. Patches welcome ;-).

Sorry for not keeping everyone Cced, I wasn't aware of the rules.

Yes, writing about that in the docs seems more reasonable than patching reset, as as you said, that'd just run update-index before the reset. Let me get at home and I'll try to push a change :)

Regards

I'm home, https://github.com/alexandernst/git/commit/61f0a7d558e3cbae308fabdff66bd87569d6aa18 Is that good? Should I PR?

Alexander Nestorov <alexandernst@gmail.com> writes:

Please, post your patches inline, it eases review. More generally, read Documentation/SubmittingPatches.

+Ignore file permissions::

It's not only about permissions, and it does not "ignore" them, it just notices when there's actually no change although the mtime has changed.

How about that:

+Reset only files who's content changed (instead of mtime modification):: ++ +------------ +$ git update-index --refresh <1> +$ git reset --hard <2> +------------ ++ +<1> Make git realize which files actually changed instead of +checking out all files whether their content changed or only +their mtime changed. +<2> Now git reset --hard will checkout only the files that +actually changed.

Alexander Nestorov <alexandernst@gmail.com> writes:

Much better, yes. I'd say "stat information" instead of mtime (that's what used in the description of update-index --refresh, and is a bit more accurate since Git also checks the inode number IIRC).

s/git/Git/ when talking about "The Git system" (as opposed to "The git command"). git-gui or other Git clients would also see the index change.

Ok, this is how it looks. If everything is ok, I'm sending it to the ML

From: Alexander Nestorov <alexandernst@gmail.com>
Date: Wed, 19 Jun 2013 09:55:42 +0200
Subject: [PATCH] Add example for reseting based on content changes instead of
 stat changes

---
 Documentation/git-reset.txt | 12 ++++++++++++
 1 file changed, 12 insertions(+)

Alexander Nestorov <alexandernst@gmail.com> writes:

Please, read Documentation/SubmittingPatches (you lack a sign-off and if you think the patch is ready, you should Cc Junio). Also, it's better to have the commit headers directly as mail headers (git send-email is your friend).

That's still not 100% accurate. Actual mode changes would trigger a rewrite of the file. Perhaps stg like

Reset only files which actually changed (not those with only stat information change)::

(Sorry for nitpicking so much)

Matthieu Moy <Matthieu.Moy@grenoble-inp.fr> writes:

I do not think the above clarifies anything to be of much help.

If you _know_ what "actually changed" means, i.e. either contents or the executable-ness changed, then "(not those ...)" does not help you at all.

If you don't, then "only stat information change" will invite "I did chmod -x and the file does not have any actual change." confusion.

If this addition is to help people who do not know what "actually changed" means, that part needs to be clarified, no?

@Matthieu
 Ok, I'm replacing with "Reset only files which actually changed (not
those with only stat information change)"

@Junio
 I'm not sure what you're asking, sorry, I'm not able to understand
your question.

On 19 June 2013 01:00, Alexander Nestorov <alexandernst@gmail.com> wrote:

You should use "whose" here instead of "who's".

On Tue, Jun 18, 2013 at 03:25:22PM +0200, Alexander Nestorov wrote:

Does "git reset --keep" behave in the same way? I would expect it to leave permissions as they were.

Git reset --keep is not an option as it will abort the operation if there are local changes, which is exactly what I want to do: replace files with local changes but leave file permissions as they are.