▲ 111 points
back
47 comments
Or, alternatively:
1. Tag your release with `git tag -s`
2. `git push --follow-tags` (or `git push remotename tagname` in git <1.8.3)
3. Upload your pubkey to github.
Much less effort…
Exactly. It has the additional advantage of signing the history, not just one source dump.
Do the above steps generate an actual signed release, or just a signed tag?
A tag, not a release. I didn't realize there was a difference, but in hindsight, I'd argue it's superior anyway: I'm much more comfortable to sign my local repository copy than signing an archive generated by github in a non-reproducible way.
I can verify the contents of github's archives, but I cannot, in good faith, say I've verified that the archives do not contain malicious code that triggers vulnerabilities in zip/tar/gzip.
I'm much more comfortable to sign my local repository copy than signing an archive generated by github
Both can be part of the process.
However, signing the tag doesn't really help someone who isn't cloning the repository. For those people that are just downloading the source tarball directly, having a GPG signature for the release can ensure that github(or other attacker) hasn't modified the contents of the tarball.
Sort of.
I mean if github wants to modify the contents of the tarball, they'll also modify the signature. So unless you're also confirming that it's the right signature, and not just that it's signed, what exactly are you doing?
That is where signing other people's OpenPGP keys comes in, and the web of trust for those who haven't met their upstreams.
github can also modify signed tags.
Reproducing GitHub's archives locally is trivial. They just use a standard git command to do it anyway.
git archive -o project-0.5.tar.gz --prefix project-0.5/ v0.5
I think we have different standards for "reproducible". You can replicate the archive structure and naming, but the archive metadata – and thus checksums – will be different:
$ curl -sL https://github.com/creshal/yspave/archive/0.1.0.zip | sha512sum
375564f127b46f3d1208c30b4eb5c39725985ca09d7bacf940d346727e8bcf9362966f8523c39e6c4f6f1a86b52ceba8823c2f4b016fdedb5947cd6a6bafc930 -
$ git clone git@github.com:creshal/yspave.git
…
$ cd yspave
$ git archive -o 0.1.0.zip --prefix yspave-0.1.0/ 0.1.0
$ sha512sum 0.1.0.zip
aca1163d5270e8ad6796f10511f49b0cf5fe61007fed0786a6643ae522e59a6aa7d6b0cddf4e5b94a435da8d99aa09b2f2ebef7788a82d4c2e7d8d4ad79e0634 0.1.0.zip
–––––
So, I can verify that the contents weren't tampered with, but I cannot verify the archives themselves unless I happen to understand the zip/tar/gz… specifications better than most people implementing it.
you can verify by downloading the archive and comparing the content of the archive with the content of the archive produced by git archive. The issue seems to be that the zip file contains the version of the zip utility used to create the file, so you'd need the exact same version to make a reproducible build.
That's interesting. I'm not able to replicate your results. Cloning that respository and doing the same git archive command produces an identical file to what GitHub serves.
I'm on a fully updated Arch Linux with the latest Git. If you have an older version, that might be making the difference.
I can reproduce the result: The content of the file is identical, the checksum differs.
> I'm on a fully updated Arch Linux with the latest Git.
Same, actually.
According to diffoscope, the (only) differences are in the archive member modification times:
--- ../0.1.0.zip
+++ 0.1.0.zip
├── zipinfo -v {}
│ @@ -28,15 +28,15 @@
│ file system or operating system of origin: MS-DOS, OS/2 or NT FAT
│ version of encoding software: 0.0
│ minimum file system compatibility required: MS-DOS, OS/2 or NT FAT
│ minimum software version required to extract: 1.0
│ compression method: none (stored)
│ file security status: not encrypted
│ extended local header: no
│ - file last modified on (DOS date/time): 2016 Apr 8 06:33:12
│ + file last modified on (DOS date/time): 2016 Apr 8 14:33:12
│ file last modified on (UT extra field modtime): 2016 Apr 8 13:33:13 local
│ file last modified on (UT extra field modtime): 2016 Apr 8 13:33:13 UTC
│ 32-bit CRC value (hex): 00000000
│ compressed size: 0 bytes
│ uncompressed size: 0 bytes
│ length of filename: 13 characters
│ length of extra field: 9 bytes
│ @@ -63,15 +63,15 @@
│ version of encoding software: 0.0
│ minimum file system compatibility required: MS-DOS, OS/2 or NT FAT
│ minimum software version required to extract: 1.0
│ compression method: deflated
│ compression sub-type (deflation): normal
│ file security status: not encrypted
│ extended local header: no
│ - file last modified on (DOS date/time): 2016 Apr 8 06:33:12
│ + file last modified on (DOS date/time): 2016 Apr 8 14:33:12
│ file last modified on (UT extra field modtime): 2016 Apr 8 13:33:13 local
│ file last modified on (UT extra field modtime): 2016 Apr 8 13:33:13 UTC
│ 32-bit CRC value (hex): 570b4391
│ compressed size: 52 bytes
│ uncompressed size: 78 bytes
│ length of filename: 23 characters
│ length of extra field: 9 bytes
│ @@ -97,15 +97,15 @@
│ file system or operating system of origin: Unix
│ version of encoding software: 2.3
│ minimum file system compatibility required: MS-DOS, OS/2 or NT FAT
│ minimum software version required to extract: 1.0
│ compression method: none (stored)
│ file security status: not encrypted
│ extended local header: no
│ - file last modified on (DOS date/time): 2016 Apr 8 06:33:12
│ + file last modified on (DOS date/time): 2016 Apr 8 14:33:12
│ file last modified on (UT extra field modtime): 2016 Apr 8 13:33:13 local
│ file last modified on (UT extra field modtime): 2016 Apr 8 13:33:13 UTC
│ 32-bit CRC value (hex): bc890775
│ compressed size: 59 bytes
│ uncompressed size: 59 bytes
│ length of filename: 20 characters
│ length of extra field: 9 bytes
Timezone difference, with extra DST madness mixed in? 13:33 is UTC, 14:33 is UTC+DST, 06:33 is Pacific Time+DST
Ahh, never mind. The tar.gz is indeed identical, "only" the zip differs.
That does not nearly have the same security level. Signing a tag in git is like signing a commit. A commit is nothing else than a reference to a tree SHA1 (and parent, message and author). That tree SHA1 contains SHA1 of the files and trees in the directory.
Even assuming your client checks all the SHA1 and does not just trust the server (which is not the default), and that you are verifying the git tag signature correctly (which is harder than just checking a tarball), SHA1 collisions are considered practical.
Git is essentially a Merkle DAG, thus a single reference to the root object suffices.
Also collisions are only practical with large amounts of padding which would be visible. That's why nobody really cares too much about them for source control systems.
They missed the bit about verifying the validity of the content of the tarball... Given that you're essentially guarding against GitHub going rogue, that's half the point right?
Verifying the content is one thing, what about making sure the archive itself isn't malicious and triggering tar/gzip/zip vulnerabilities? In my experience, github's archives are not generated in a reproducible way.
Never mind that comment, as established in a different comment chain, it's an inherent limitation of zip archives. .tar.gz archives are fully reproducible.
github archives are definitely generated in a reproducible way.
They are now. They weren't 2 or 3 years ago.
> 4. Go back to your "Releases" section and download the tarball mysoftware-0.4.tar.gz automatically generated by GitHub.
If only. That's by far my biggest irk with GitHub releases because they only use the tag as name for the tarball (e.g. v0.4.tar.gz). Building RPM packages with those tarballs is a bit of a pain because RPM building tools expect all tarballs in one directory and you don't know which tarball belongs to which software. Maybe I'm just being ignorant here.
Other problems I encounter with Github‐generated tarballs: most projects that use Autoconf don’t keep the generated configure script in the repository, so the Github tarball has a build dependency on autoconf (as opposed to tarballs generated with “make dist”).
Also, many people don’t realize that Github tarballs don’t contain submodules. This is a big hassle when upstreams don’t adequately test their releases.
Providing a working tarball is becoming less and less common due to Github’s (and other sites’) UI. Taking the steps in the article is very helpful to packagers all over, but the above factors can be a problem for upstreams who aren’t aware of the problems.
It's not github's fault that RPM's tooling is stuck in the 90s, to be honest. makepkg and others support renaming tarballs during download to cover this use case.
I don't disagree, but it just feels off to tell a packaging tool to rename a tarball because the default name is not verbose enough to even identify the tarball.
But even so it is designed to be likely to get accidental collisions.
Yes, you're being ignorant. Use '--content-disposition' with wget, or '--remote-name' in curl. The top level directory in the tarball corresponds to this name and everything works out as expected.
It's not only the name of the tarball after downloading it. RPM SPEC files require me to pass a full URL to the tarball's upstream source[1]:
> Source0: The full URL for the compressed archive containing the (original) pristine source code, as upstream released it. "Source" is synonymous with "Source0". If you give a full URL (and you should), its basename will be used when looking in the SOURCES directory.
But the only URL that is available is <Project URL>/archive/v0.1.1.tar.gz. The SPEC file is therefore defining a file named v0.1.1.tar.gz although I downloaded my tarball as foo-0.1.1.tar.gz. The packaging process looks for v0.1.1.tar.gz, fails to find it in my SOURCES directory and aborts.
I'm not blaming GitHub here in any way, I just wish I had a bit more flexibility for the tarball URL. Would be neat.
[1] https://fedoraproject.org/wiki/How_to_create_an_RPM_package#...
Specify
https://github.com/username/repo/archive/v0.1.1/foo-0.1.1.tar.gzMy god, that's the solution to my problem. Thank you very much. I did not realize that url was available.
I recommend using some tooling that wraps the original rpm tooling, such as fpm or fpm-cookery. It does a fair job of covering up the worst warts.
What's the threat model you are defending against here?
GitHub changing the binary being downloaded? They could serve a different signature.
Another committer changing the tag? They could equally change the signature....
> GitHub changing the binary being downloaded? They could serve a different signature.
> Another committer changing the tag? They could equally change the signature....
No they couldn't, not unless they have your private key.
(Debian does actually require all keys used to sign debian packages to have a trust path back to them - if you want to become Debian maintainer for a project you have to meet up with a Debian developer and have them sign your GPG key - so GitHub or another committer couldn't just use a different key to sign with).
Github doesn't have your private key, so they won't be able to serve a valid signature.
The only thing they could do is to substitute an older signed version when you request a newer version.
They could create a new private key using your name. Most people will trust 'Verified' on a web page more than the OpenPGP web of trust.
Most people will, but this is written by a Debian developer and they're one of the few groups of people who actually make use of the OpenPGP web of trust.
To emphasise this, there is an evil duplicate of the OpenPGP "strong set" here:
Not sure why you're being downvoted, this "what is the threat model here" was my first thought, too, and this comment is also correct, as well.
Probably my "Most people" quip.
The xkcd trust model https://xkcd.com/1181/
Signed releases should be required for any package management system. I'm finding Ruby to be lacking in this area:
`bundle install --trust-policy HighSecurity`
There appear to be a lot of shots taken behind the basket - even as far back at the 3 pt line. Is this a mistake in measurement?