RFC patchRewriting revs in place in push target repository

Rewrite refs in place in receive-pack & friends

When updating a ref, it would write a new file with the new ref and then rename it, overwriting the original file. The problem is that this destroys permissions and ownership of the original file, which is troublesome especially in multiuser environment, like the one I live in.

This might be controversial, but it's a showbreaker for me wrt. pushing now. Some alternative solution barely solving my particular situation might be surely worked out, but this is more general. The question is:

* Does this break atomicity?

	I think it does not in real setups, since thanks to O_RDWR the
	file should be overwritten only when the write() happens.
	Can a 41-byte write() be non-atomic in any real conditions?

* Does this break with full disk/quota?

	I'm not sure - we are substituting 41 bytes by another 41
	bytes; will the system ever be evil enough to truncate the
	file, then decide the user is over his quota and not write
	the new contents?

Signed-off-by: Petr Baudis <pasky@suse.cz>

On Sat, 13 Aug 2005, Petr Baudis wrote:

That's not the problem.

The problem is that your change means that there is no locking, and you now can have two writers that both update the same file, and they _both_ think that they succeed. They'll both read the old contents, decide that it still is the one from before the push, and then they'll both do the write.

And yes, in most (all?) sane filesystems, the end result is that one of them "wins", and the end result is a nice 41-byte file. But the problem is that the other write just totally got lost, and the person doing the push _thought_ he had updated the thing, but never did.

To make things worse, with NFS and client-side caching, different clients that look at the tree at around that time can literally see _different_ heads winning the race. One of the writers wrote "first", and that client (and other NFS clients doing a read at that time) will see it succeed. But then the other pusher writes, and now people will see _that_ one succeed.

Confusion reigns.

In contrast, with the "create lock-file and rename" thing, if there is a race, somebody will win, and the loser will hopefully know they lost.

Probably not.

But how about you just try to copy the permission/group of the original file before you do the rename? I assume that if you're depending on permissions, it's either a shared group or by having the thing writable by others, so doing a

	if (!fstat(oldfd, &st)) {
		fchown(fd, (uid_t) -1, st.st_gid);
		fchmod(fd, st.st_mode & ALLPERMS);
	}
	.. do rename here ..

which should get you where you want, no?

		Linus

Petr Baudis <pasky@suse.cz> writes:

Hmph. If a repo is _really_ used multiuser then you should not have to care about ownership. If you can write into a repository for a project (implying that you are a member of that project group), and if your umask is set up correctly (meaning it is 002 or looser), and with g+s bit on the directory at the repository root level when it was created, shouldn't your newly created ref file be also writable by others in that project?

On Sat, 13 Aug 2005, Junio C Hamano wrote:

I think Pasky's usage is that different heads are owned by different groups and/or users, and he wants to use the filesystem permissions to determine who gets to update which branch. Which is reasonable in a way.

On the other hand, I don't think filesystem permissions are really very useful. I think it's more appropriate to use triggers to say something like "only allow people in the 'xyz' group to write to this head".

Obviously, triggers aren't about _security_ - somebody who has write permissions to the tree can always screw up others. But triggers are fine for things like branch ownership, where you trust your users, but you just want to avoid mistakes.

So a trigger might be something like

	#!/bin/sh
	. git-sh-setup-script
	branch="$1"
	old="$2"
	new="$3"
	if [ -e $GIT_DIR/permissions/$branch ]; then
		id=$(id -un)
		grep -q "^$id$" $GIT_DIR/permissions/$branch ||
			die "You're not allowed to write to $branch"
	fi
	true

and that would allow you to list all users that are allowed to write to the branch in $GIT_DIR/permissions/<branchname>.

Totally untested, of course. But the concept should work.

		Linus

Dear diary, on Sun, Aug 14, 2005 at 02:55:16AM CEST, I got a letter where Junio C Hamano <junkio@cox.net> told me that...

Hmm, but how do you actually set the umask correctly just for git pushing? I'm sorry but it doesn't occur to me.

I like Linus' solution, but have no time to do the patch now, so I did just a quick dirty fix and added a chmod to hooks/update. ;-)

On Sat, Aug 13, 2005 at 11:47:25PM +0200, Petr Baudis wrote:

yes

if you journal metadata only you can see a file extended w/o having the block flushed

Hi, Chris Wedgwood wrote:

??? but the file is *not* extended. Also, whether or not a block is flushed should only matter if the machine crashes ..?