RFC: Another proposed hash function transition plan

Hi,

This past week we came up with this idea for what a transition to a new hash function for Git would look like. I'd be interested in your thoughts (especially if you can make them as comments on the document, which makes it easier to address them and update the document).

This document is still in flux but I thought it best to send it out early to start getting feedback.

We tried to incorporate some thoughts from the thread
http://public-inbox.org/git/[email protected]
but it is a little long so it is easy to imagine we've missed some things already discussed there.

You can use the doc URL

https://goo.gl/gh2Mzc

to view the latest version and comment.

Thoughts welcome, as always.

Git hash function transition
============================
Status: Draft
Last Updated: 2017-03-03

Objective
---------
Migrate Git from SHA-1 to a stronger hash function.

Background
----------
The Git version control system can be thought of as a content addressable filesystem. It uses the SHA-1 hash function to name content. For example, files, trees, commits are referred to by hash values unlike in other traditional version control systems where files or versions are referred to via sequential numbers. The use of a hash function to address its content delivers a few advantages:

* Integrity checking is easy. Bit flips, for example, are easily detected, as the hash of corrupted content does not match its name. Lookup of objects is fast.

Using a cryptographically secure hash function brings additional advantages:

* Object names can be signed and third parties can trust the hash to address the signed object and all objects it references.
* Communication using Git protocol and out of band communication methods have a short reliable string that can be used to reliably address stored content.

Over time some flaws in SHA-1 have been discovered by security researchers. https://shattered.io demonstrated a practical SHA-1 hash collision. As a result, SHA-1 cannot be considered cryptographically secure any more. This impacts the communication of hash values because we cannot trust that a given hash value represents the known good version of content that the speaker intended.

SHA-1 still possesses the other properties such as fast object lookup and safe error checking, but other hash functions are equally suitable that are believed to be cryptographically secure.

Goals
-----
1. The transition to SHA256 can be done one local repository at a time. a. Requiring no action by any other party.
b. A SHA256 repository can communicate with SHA-1 Git servers and
clients (push/fetch).
c. Users can use SHA-1 and SHA256 identifiers for objects
interchangeably.
d. New signed objects make use of a stronger hash function than
SHA-1 for their security guarantees.
2. Allow a complete transition away from SHA-1.
a. Local metadata for SHA-1 compatibility can be dropped in a
repository if compatibility with SHA-1 is no longer needed.
3. Maintainability throughout the process.
a. The object format is kept simple and consistent.
b. Creation of a generalized repository conversion tool.

Non-Goals
---------
1. Add SHA256 support to Git protocol. This is valuable and the logical next step but it is out of scope for this initial design.
2. Transparently improving the security of existing SHA-1 signed objects.
3. Intermixing objects using multiple hash functions in a single repository.
4. Taking the opportunity to fix other bugs in git's formats and protocols.
5. Shallow clones and fetches into a SHA256 repository. (This will change when we add SHA256 support to Git protocol.)
6. Skip fetching some submodules of a project into a SHA256 repository. (This also depends on SHA256 support in Git protocol.)

Overview
--------
We introduce a new repository format extension sha256. Repositories with this extension enabled use SHA256 instead of SHA-1 to name their objects. This affects both object names and object content --- both the names of objects and all references to other objects within an object are switched to the new hash function.

sha256 repositories cannot be read by older versions of Git.

Alongside the packfile, a sha256 stores a bidirectional mapping between sha256 and sha1 object names. The mapping is generated locally and can be verified using "git fsck". Object lookups use this mapping to allow naming objects using either their sha1 and sha256 names interchangeably.

"git cat-file" and "git hash-object" gain options to display a sha256 object in its sha1 form and write a sha256 object given its sha1 form. This requires all objects referenced by that object to be present in the object database so that they can be named using the appropriate name (using the bidirectional hash mapping).

Fetches from a SHA-1 based server convert the fetched objects into sha256 form and record the mapping in the bidirectional mapping table (see below for details). Pushes to a SHA-1 based server convert the objects being pushed into sha1 form so the server does not have to be aware of the hash function the client is using.

Detailed Design
---------------
Object names
~~~~~~~~~~~~
Objects can be named by their 40 hexadecimal digit sha1-name or 64 hexadecimal digit sha256-name, plus names derived from those (see gitrevisions(7)).

The sha1-name of an object is the SHA-1 of the concatenation of its type, length, a nul byte, and the object's sha1-content. This is the traditional <sha1> used in Git to name objects.

The sha256-name of an object is the SHA-256 of the concatenation of its type, length, a nul byte, and the object's sha256-content.

Object format
~~~~~~~~~~~~~
Objects are stored using a compressed representation of their sha256-content. The sha256-content of an object is the same as its sha1-content, except that:
* objects referenced by the object are named using their sha256-names instead of sha1-names
* signed tags, commits, and merges of signed tags get some additional fields (see below)

The format allows round-trip conversion between sha256-content and sha1-content.

Loose objects use zlib compression and packed objects use the packed format described in Documentation/technical/pack-format.txt, just like today.

Translation table
~~~~~~~~~~~~~~~~~
A fast bidirectional mapping between sha1-names and sha256-names of all local objects in the repository is kept on disk. The exact format of that mapping is to be determined.

All operations that make new objects (e.g., "git commit") add the new objects to the translation table.

Reading an object's sha1-content
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The sha1-content of an object can be read by converting all sha256-names its sha256-content references to sha1-names using the translation table. There is an additional minor transformation needed for signed tags, commits, and merges (see below).

Fetch
~~~~~
Fetching from a SHA-1 based server requires translating between SHA-1 and SHA-256 based representations on the fly.

SHA-1s named in the ref advertisement can be translated to SHA-256 and looked up as local objects using the translation table.

Negotiation proceeds as today. Any "have"s or "want"s generated locally are converted to SHA-1 before being sent to the server, and SHA-1s mentioned by the server are converted to SHA-256 when looking them up locally.

After negotiation, the server sends a packfile containing the requested objects. We convert the packfile to SHA-256 format using the following steps:

1. index-pack: inflate each object in the packfile and compute its SHA-1. Objects can contain deltas in OBJ_REF_DELTA format against objects the client has locally. These objects can be looked up using the translation table and their sha1-content read as described above to resolve the deltas.
2. topological sort: starting at the "want"s from the negotiation phase, walk through objects in the pack and emit a list of them in topologically sorted order. (This list only contains objects reachable from the "wants". If the pack from the server contained additional extraneous objects, then they will be discarded.)
3. convert to sha256: open a new (sha256) packfile. Read the topologically sorted list just generated in reverse order. For each object, inflate its sha1-content, convert to sha256-content, and write it to the sha256 pack. Write an idx file for this pack and include the new sha1<->sha256 mapping entry in the translation table.
4. clean up: remove the SHA-1 based pack file, index, and topologically sorted list obtained from the server and steps 1 and 2.

Step 3 requires every object referenced by the new object to be in the translation table. This is why the topological sort step is necessary.

As an optimization, step 1 can write a file describing what objects each object it has inflated from the packfile references. This makes the topological sort in step 2 possible without inflating the objects in the packfile for a second time. The objects need to be inflated again in step 3, for a total of two inflations.

Push
~~~~
Push is simpler than fetch because the objects referenced by the pushed objects are already in the translation table. The sha1-content of each object being pushed can be read as described in the "Reading an object's sha1-content" section to generate the pack written by git send-pack.

Signed Objects
~~~~~~~~~~~~~~
Commits
^^^^^^^
Commits currently have the following sequence of header lines:

"tree" SP object-name
("parent" SP object-name)*
"author" SP ident
"committer" SP ident
("mergetag" SP object-content)?
("gpgsig" SP pgp-signature)?

We introduce new header lines "hash" and "nohash" that come after the "gpgsig" field. No "hash" lines may appear unless the "gpgsig" field is present.

Hash lines have the form

"hash" SP hash-function SP field SP alternate-object-name

Nohash lines have the form

"nohash" SP hash-function

There are only two recognized values of hash-function: "sha1" and "sha256". "git fsck" will tolerate values of hash-function it does not recognize, as long as they do not come before either of those two. All "nohash" lines come before all "hash" lines. Any "hash sha1" lines must come before all "hash sha256" lines, and likewise for nohash. The Git project determines any future supported hash-functions that can come after those two and their order.

There can be at most one "nohash <hash-function>" for one hash function, indicating that this hash function should not be used when checking the commit's signature.

There is one "hash <hash-function>" line for each tree or parent field in the commit object header. The hash lines record object names for those trees and parents using the indicated hash function, to be used when checking the commit's signature.

TODO: simplify signature rules, handle the mergetag field better.

sha256-content of signed commits ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
The sha256-content of a commit with a "gpgsig" header can include no hash and nohash lines, a "nohash sha256" line and "hash sha1", or just a "hash sha1" line.

Examples:
1. tree 98ea6e4f216f2fb4b69fff9b3a44842c38686ca685f3f55dc48c5d3fb1107be4 parent e094bc809626f0a401a40d75c56df478e546902ff812772c4594265203b23980 parent 1059dab4748aa33b86dad5ca97357bd322abaa558921255623fbddd066bb3315 author A U Thor <[email protected]> 1465982009 +0000 committer C O Mitter <[email protected]> 1465982009 +0000 gpgsig ...
2. tree 98ea6e4f216f2fb4b69fff9b3a44842c38686ca685f3f55dc48c5d3fb1107be4 parent e094bc809626f0a401a40d75c56df478e546902ff812772c4594265203b23980 parent 1059dab4748aa33b86dad5ca97357bd322abaa558921255623fbddd066bb3315 author A U Thor <[email protected]> 1465982009 +0000 committer C O Mitter <[email protected]> 1465982009 +0000 gpgsig ...
nohash sha256
hash sha1 tree c7b1cff039a93f3600a1d18b82d26688668c7dea hash sha1 parent c33429be94b5f2d3ee9b0adad223f877f174b05d hash sha1 parent 04b871796dc0420f8e7561a895b52484b701d51a
3. tree 98ea6e4f216f2fb4b69fff9b3a44842c38686ca685f3f55dc48c5d3fb1107be4 parent e094bc809626f0a401a40d75c56df478e546902ff812772c4594265203b23980 parent 1059dab4748aa33b86dad5ca97357bd322abaa558921255623fbddd066bb3315 author A U Thor <[email protected]> 1465982009 +0000 committer C O Mitter <[email protected]> 1465982009 +0000 gpgsig ...
hash sha1 tree c7b1cff039a93f3600a1d18b82d26688668c7dea hash sha1 parent c33429be94b5f2d3ee9b0adad223f877f174b05d hash sha1 parent 04b871796dc0420f8e7561a895b52484b701d51a

sha1-content of signed commits ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
The sha1-content of a commit with a "gpgsig" header can contain a "nohash sha1" and "hash sha256" line, no hash or nohash lines, or just a "hash sha256" line.

Examples:
1. tree c7b1cff039a93f3600a1d18b82d26688668c7dea
parent c33429be94b5f2d3ee9b0adad223f877f174b05d
parent 04b871796dc0420f8e7561a895b52484b701d51a
author A U Thor <[email protected]> 1465982009 +0000
committer C O Mitter <[email protected]> 1465982009 +0000
gpgsig ...
nohash sha1
hash sha256 tree 98ea6e4f216f2fb4b69fff9b3a44842c38686ca685f3f55dc48c5d3fb1107be4 hash sha256 parent e094bc809626f0a401a40d75c56df478e546902ff812772c4594265203b23980 hash sha256 parent 1059dab4748aa33b86dad5ca97357bd322abaa558921255623fbddd066bb3315
2. tree c7b1cff039a93f3600a1d18b82d26688668c7dea
parent c33429be94b5f2d3ee9b0adad223f877f174b05d
parent 04b871796dc0420f8e7561a895b52484b701d51a
author A U Thor <[email protected]> 1465982009 +0000
committer C O Mitter <[email protected]> 1465982009 +0000
gpgsig ...
3. tree c7b1cff039a93f3600a1d18b82d26688668c7dea
parent c33429be94b5f2d3ee9b0adad223f877f174b05d
parent 04b871796dc0420f8e7561a895b52484b701d51a
author A U Thor <[email protected]> 1465982009 +0000
committer C O Mitter <[email protected]> 1465982009 +0000
gpgsig ...
hash sha256 tree 98ea6e4f216f2fb4b69fff9b3a44842c38686ca685f3f55dc48c5d3fb1107be4 hash sha256 parent e094bc809626f0a401a40d75c56df478e546902ff812772c4594265203b23980 hash sha256 parent 1059dab4748aa33b86dad5ca97357bd322abaa558921255623fbddd066bb3315

Converting signed commits
^^^^^^^^^^^^^^^^^^^^^^^^^
To convert the sha1-content of a signed commit to its sha256-content:

1. Change "tree" and "parent" lines to use the sha256-names of referenced objects, as with unsigned commits.
2. If there is a "mergetag" field, convert it from sha1-content to sha256-content, as with unsigned commits with a mergetag (see the "Mergetag" section below).
3. Unless there is a "nohash sha1" line, add a full set of "hash sha1 <field> <sha1>" lines indicating the sha1-names of the tree and parents.
4. Remove any "hash sha256 <field> <sha256>" lines. If no such lines were present, add a "nohash sha256" line.

Converting the sha256-content of a signed commit to sha1-content uses the same process with sha1 and sha256 switched.

Verifying signed commit signatures ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
If the commit has a "hash sha1" line (or is sha1-content without a "nohash sha1" line): check that the signature matches the sha1-content with gpgsig field stripped out.

Otherwise: check that the signature matches the sha1-content with gpgsig, nohash, tree, and parents fields stripped out.

With the examples above, the signed payloads are
1. author A U Thor <[email protected]> 1465982009 +0000
committer C O Mitter <[email protected]> 1465982009 +0000
hash sha256 tree 98ea6e4f216f2fb4b69fff9b3a44842c38686ca685f3f55dc48c5d3fb1107be4 hash sha256 parent e094bc809626f0a401a40d75c56df478e546902ff812772c4594265203b23980 hash sha256 parent 1059dab4748aa33b86dad5ca97357bd322abaa558921255623fbddd066bb3315
2. tree c7b1cff039a93f3600a1d18b82d26688668c7dea
parent c33429be94b5f2d3ee9b0adad223f877f174b05d
parent 04b871796dc0420f8e7561a895b52484b701d51a
author A U Thor <[email protected]> 1465982009 +0000
committer C O Mitter <[email protected]> 1465982009 +0000
3. tree c7b1cff039a93f3600a1d18b82d26688668c7dea
parent c33429be94b5f2d3ee9b0adad223f877f174b05d
parent 04b871796dc0420f8e7561a895b52484b701d51a
author A U Thor <[email protected]> 1465982009 +0000
committer C O Mitter <[email protected]> 1465982009 +0000
hash sha1
hash sha256 tree 98ea6e4f216f2fb4b69fff9b3a44842c38686ca685f3f55dc48c5d3fb1107be4 hash sha256 parent e094bc809626f0a401a40d75c56df478e546902ff812772c4594265203b23980 hash sha256 parent 1059dab4748aa33b86dad5ca97357bd322abaa558921255623fbddd066bb3315

Current versions of "git verify-commit" can verify examples (2) and (3) (but not (1)).

Tags
~~~~
Tags currently have the following sequence of header lines:

"object" SP object-name
"type" SP type
"tag" SP identifier
"tagger" SP ident

A tag's signature, if it exists, is in the message body.

We introduce new header lines "nohash" and "hash" that come after the "tagger" field. No "nohash" or "hash" lines may appear unless the message body contains a PGP signature.

As with commits, "nohash" lines have the form "nohash <hash-function>", indicating that this hash function should not be used when checking the tag's signature.

"hash" lines have the form

"hash" SP hash-function SP alternate-object-name

This records the pointed-to object name using the indicated hash function, to be used when checking the tag's signature.

As with commits, "sha1" and "sha256" are the only permitted values of hash-function and can only appear in that order for a field when they appear. There can be at most one "nohash" line, and it comes before any "hash" lines. There can be only one "hash" line for a given hash function.

sha256-content of signed tags ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
The sha256-content of a signed tag can include no "hash" or "nohash" lines, a "nohash sha256" and "hash sha1 <sha1>" line, or just a "hash sha1 <sha1>" line.

Examples:
1. object 98ea6e4f216f2fb4b69fff9b3a44842c38686ca685f3f55dc48c5d3fb1107be4 type tree
tag v1.0
tagger C O Mitter <[email protected]> 1465981006 +0000

Tag Demo v1.0
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1

iQEcBAABAgAGBQJXYRhOAAoJEGEJLoW3InGJklkIAIcnhL7RwEb/+QeX9enkXhxn
...
2. object 98ea6e4f216f2fb4b69fff9b3a44842c38686ca685f3f55dc48c5d3fb1107be4 type tree
tag v1.0
tagger C O Mitter <[email protected]> 1465981006 +0000 nohash sha256
hash sha1 c7b1cff039a93f3600a1d18b82d26688668c7dea

Tag Demo v1.0
-----BEGIN PGP SIGNATURE-----
...
3. object 98ea6e4f216f2fb4b69fff9b3a44842c38686ca685f3f55dc48c5d3fb1107be4 type tree
tag v1.0
tagger C O Mitter <[email protected]> 1465981006 +0000 hash sha1 c7b1cff039a93f3600a1d18b82d26688668c7dea

Tag Demo v1.0
...

sha1-content of signed tags
^^^^^^^^^^^^^^^^^^^^^^^^^^^
The sha1-content of a signed tag can include a "nohash sha1" and "hash sha256" line, no "nohash" or "hash" lines, or just a "hash sha256 <sha256>" line.

Examples:
1. object c7b1cff039a93f3600a1d18b82d26688668c7dea
...
tagger C O Mitter <[email protected]> 1465981006 +0000 nohash sha1
hash sha256 98ea6e4f216f2fb4b69fff9b3a44842c38686ca685f3f55dc48c5d3fb1107be4

Tag Demo v1.0
-----BEGIN PGP SIGNATURE-----
...
2. object c7b1cff039a93f3600a1d18b82d26688668c7dea
...
tagger C O Mitter <[email protected]> 1465981006 +0000

Tag Demo v1.0
-----BEGIN PGP SIGNATURE-----
...
3. object c7b1cff039a93f3600a1d18b82d26688668c7dea
...
tagger C O Mitter <[email protected]> 1465981006 +0000
hash sha256 98ea6e4f216f2fb4b69fff9b3a44842c38686ca685f3f55dc48c5d3fb1107be4

Tag Demo v1.0
-----BEGIN PGP SIGNATURE-----
...

Signed tags can be converted between sha1-content and sha256-content using the same process as signed commits.

Verifying signed tags
^^^^^^^^^^^^^^^^^^^^^
As with commits, if the tag has a "hash sha1" (or is sha1-content without a "nohash sha1" line): check that the signature matches the sha1-content with PGP signature stripped out.

Otherwise: check that the signature matches the sha1-content with nohash and object fields and PGP signature stripped out.

Mergetag signatures
~~~~~~~~~~~~~~~~~~~
The mergetag field in the sha1-content of a commit contains the sha1-content of a tag that was merged by that commit.

The mergetag field in the sha256-content of the same commit contains the sha256-content of the same tag.

Submodules
~~~~~~~~~~
To convert recorded submodule pointers, you need to have the converted submodule repository in place. The bidirectional mapping of the submodule can be used to look up the new hash.

Caveats
-------
Shallow clone and submodules
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Because this requires all referenced objects to be available in the locally generated translation table, this design does not support shallow clone or unfetched submodules.

Protocol improvements might allow lifting this restriction.

Alternatives considered
-----------------------
Upgrading everyone working on a particular project on a flag day
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Projects like the Linux kernel are large and complex enough that flipping the switch for all projects based on the repository at once is infeasible.

Not only would all developers and server operators supporting developers have to switch on the same flag day, but supporting tooling (continuous integration, code review, bug trackers, etc) would have to be adapted as well. This also makes it difficult to get early feedback from some project participants testing before it is time for mass adoption.

Using hash functions in parallel
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
(e.g. https://public-inbox.org/git/[email protected]/ ) Objects newly created would be addressed by the new hash, but inside such an object (e.g. commit) it is still possible to address objects using the old hash function.

* You cannot trust its history (needed for bisectability) in the future without further work
* Maintenance burden as the number of supported hash functions grows (they will never go away, so they accumulate). In this proposal, by comparison, converted objects lose all references to SHA-1 except where needed to verify signatures.

This actually looks very reasonable if you can implement it cleanly enough. In many ways the "convert entirely to a new 256-bit hash" is the cleanest model, and interoperability was at least my personal concern. Maybe your model solves it (devil in the details), in which case I really like it.

I do think that if you end up essentially converting the objects without really having any true backwards compatibility at the object layer (just the translation code), you should seriously look at doing some other changes at the same time. Like not using zlib compression, it really is very slow.

Btw, I do think the particular choice of hash should still be on the table. sha-256 may be the obvious first choice, but there are definitely a few reasons to consider alternatives, especially if it's a complete switch-over like this.

One is large-file behavior - a parallel (or tree) mode could improve on that noticeably. BLAKE2 does have special support for that, for example. And SHA-256 does have known attacks compared to SHA-3-256 or BLAKE2 - whether that is due to age or due to more effort, I can't really judge. But if we're switching away from SHA1 due to known attacks, it does feel like we should be careful.

Linus

If you think you can do it, I'm all for it.

I agree with Linus on this. SHA-256 is the slowest option, and it's the one with the most advanced cryptanalysis. SHA-3-256 is faster on 64-bit machines (which, as we've seen on the list, is the overwhelming majority of machines using Git), and even BLAKE2b-256 is stronger.

Doing this all over again in another couple years should also be a non-goal.

brian m. carlson / brian with sandals: Houston, Texas, US
+1 832 623 2791 | https://www.crustytoothpaste.net/~bmc | My opinion only
OpenPGP: https://keybase.io/bk2204

I agree that when we decide to move to a new algorithm that we should select one which we plan on using for as long as possible (much longer than a couple years). While writing the document we simply used "sha256" because it was more tangible and easier to reference.

Brandon Williams

Hi,

I thought it better to revive this old thread rather than start a new thread, so as to automatically reach everybody who chimed in originally.

The SHA-1 transition *requires* a knob telling Git that the current repository uses a hash function different from SHA-1.

It would make *a whole of a lot of sense* to make that knob *not* Boolean, but to specify *which* hash function is in use.

That way, it will be easier to switch another time when it becomes necessary.

And it will also make it easier for interested parties to use a different hash function in their infrastructure if they want.

And it lifts part of that burden that we have to consider *very carefully* which function to pick. We still should be more careful than in 2005, when Git was born, and when, incidentally, when the first attacks on SHA-1 became known, of course. We were just lucky for almost 12 years.

Now, with Dunning-Kruger in mind, I feel that my degree in mathematics equips me with *just enough* competence to know just how little *even I* know about cryptography.

The smart thing to do, hence, was to get involved in this discussion and act as Lt Tawney Madison between us Git developers and experts in cryptography.

It just so happens that I work at a company with access to excellent cryptographers, and as we own the largest Git repository on the planet, we have a vested interest in ensuring Git's continued success.

After a couple of conversations with a couple of experts who I cannot thank enough for their time and patience, let alone their knowledge about this matter, it would appear that we may not have had a complete enough picture yet to even start to make the decision on the hash function to use.

From what I read, pretty much everybody who participated in the discussion was aware that the essential question is: performance vs security.

It turns out that we can have essentially both.

SHA-256 is most likely the best-studied hash function we currently know about (*maybe* SHA3-256 has been studied slightly more, but only slightly). All the experts in the field banged on it with multiple sticks and other weapons. And so far, they only found one weakness that does not even apply to Git's usage [*1*]. For cryptography experts, this is the ultimate measure of security: if something has been attacked that intensely, by that many experts, for that long, with that little effect, it is the best we got at the time.

And since SHA-256 has become the standard, and more importantly: since SHA-256 was explicitly designed to allow for relatively inexpensive hardware acceleration, this is what we will soon have: hardware support in the form of, say, special CPU instructions. (That is what I meant by: we can have performance *and* security.)

This is a rather important point to stress, by the way: BLAKE's design is apparently *not* friendly to CPU instruction implementations. Meaning that SHA-256 will be faster than BLAKE (and even than BLAKE2) once the Intel and AMD CPUs with hardware support for SHA-256 become common.

I also heard something really worrisome about BLAKE2 that makes me want to stay away from it (in addition to the difficulty it poses for hardware acceleration): to compete in the SHA-3 contest, BLAKE added complexity so that it would be roughly on par with its competitors. To allow for faster execution in software, this complexity was *removed* from BLAKE to create BLAKE2, making it weaker than SHA-256.

Another important point to consider is that SHA-256 implementations are everywhere. Think e.g. how difficult we would make it on, say, JGit or go-git if we chose a less common hash function.

As to KangarooTwelve: it has seen substantially less cryptanalysis than SHA-256 and SHA3-256. That does not necessarily mean that it is weaker, but it means that we simply cannot know whether it is as strong. On that basis alone, I would already reject it, and then there are far fewer implementations, too.

When it comes to choosing SHA-256 vs SHA3-256, I would like to point out that hardware acceleration is a lot farther in the future than SHA-256 support. And according to the experts I asked, they are roughly equally secure as far as Git's usage is concerned, even if the SHA-3 contest provided SHA3-256 with even fiercer cryptanalysis than SHA-256.

In short: my takeaway from the conversations with cryptography experts was that SHA-256 would be the best choice for now, and that we should make sure that the next switch is not as painful as this one (read: we should not repeat the mistake of hard-coding the new hash function into Git as much as we hard-coded SHA-1 into it).

Ciao,
Johannes

Footnote *1*: SHA-256, as all hash functions whose output is essentially the entire internal state, are susceptible to a so-called "length extension attack", where the hash of a secret+message can be used to generate the hash of secret+message+piggyback without knowing the secret. This is not the case for Git: only visible data are hashed. The type of attacks Git has to worry about is very different from the length extension attacks, and it is highly unlikely that that weakness of SHA-256 leads to, say, a collision attack.

What do the experts think or SHA512/256, which completely removes the concerns over length extension attack? (which I'd argue is better than sweeping them under the carpet)

Mike

I don't think it's sweeping them under the carpet. Git does not use the hash as a MAC, so length extension attacks aren't a thing (and even if we later wanted to use the same algorithm as a MAC, the HMAC construction is a well-studied technique for dealing with it).

That said, SHA-512 is typically a little faster than SHA-256 on 64-bit platforms. I don't know if that will change with the advent of hardware instructions oriented towards SHA-256.

-Peff

On Thu, Jun 15 2017, Jeff King jotted:

Quoting my own CACBZZX7JRA2niwt9wsGAxnzS+gWS8hTUgzWm8NaY1gs87o8xVQ@mail.gmail.com sent ~2 weeks ago to the list:

On Fri, Jun 2, 2017 at 7:54 PM, Jonathan Nieder <[email protected]> wrote:
[...]
> 4. When choosing a hash function, people may argue about performance.
> It would be useful for run some benchmarks for git (running
> the test suite, t/perf tests, etc) using a variety of hash
> functions as input to such a discussion.

To the extent that such benchmarks matter, it seems prudent to heavily weigh them in favor of whatever seems to be likely to be the more common hash function going forward, since those are likely to get faster through future hardware acceleration.

E.g. Intel announced Goldmont last year which according to one SHA-1 implementation improved from 9.5 cycles per byte to 2.7 cpb[1]. They only have acceleration for SHA-1 and SHA-256[2]

1. https://github.com/weidai11/cryptopp/issues/139#issuecomment-264283385

2. https://en.wikipedia.org/wiki/Goldmont

Maybe someone else knows of better numbers / benchmarks, but such a reduction in CBP likely makes it faster than SHA-512.

Hi,

I really tried to drive that point home, as it had been made very clear to me that the length extension attack is something that Git need not concern itself.

The length extension attack *only* comes into play when there are secrets that are hashed. In that case, one would not want others to be able to produce a valid hash *without* knowing the secrets. And SHA-256 allows to "reconstruct" the internal state (which is the hash value) in order to continue at any point, i.e. if the hash for secret+message is known, it is easy to calculate the hash for secret+message+addition, without knowing the secret at all.

That is exactly *not* the case with Git. In Git, what we want to hash is known in its entirety. If the hash value were not identical to the internal state, it would be easy enough to reconstruct, because *there are no secrets*.

So please understand that even the direction that the length extension attack takes is completely different than the direction any attack would have to take that weakens SHA-256 for Git's purposes. As far as Git's usage is concerned, SHA-256 has no known weaknesses.

It is *really, really, really* important to understand this before going on to suggest another hash function such as SHA-512/256 (i.e. SHA-512 truncated to 256 bits), based only on that perceived weakness of SHA-256.

Very, very likely faster than SHA-512.

I'd like to stress explicitly that the Intel SHA extensions do *not* cover SHA-512:

https://en.wikipedia.org/wiki/Intel_SHA_extensions

In other words, once those extensions become commonplace, SHA-256 will be faster than SHA-512, hands down.

Ciao,
Dscho

(I was asked to comment a few points in public by Jonathan.)

I think this group can safely assume that SHA-256, SHA-512, BLAKE2, K12, etc are all secure to the extent that I don't believe that making comparisons between them on that axis is meaningful. Thus I think the question is primarily concerned with performance and implementation availability.

I think any of the above would be reasonable choices. I don't believe that length-extension is a concern here.

SHA-512/256 will be faster than SHA-256 on 64-bit systems in software. The graph at https://blake2.net/ suggests a 50% speedup on Skylake. On my Ivy Bridge system, it's about 20%.

(SHA-512/256 does not enjoy the same availability in common libraries however.)

Both Intel and ARM have SHA-256 instructions defined. I've not seen good benchmarks of them yet, but they will make SHA-256 faster than SHA-512 when available. However, it's very possible that something like BLAKE2bp will still be faster. Of course, BLAKE2bp does not enjoy the ubiquity of SHA-256, but nor do you have to wait years for the CPU population to advance for high performance.

So, overall, none of these choices should obviously be excluded. The considerations at this point are not cryptographic and the tradeoff between implementation ease and performance is one that the git community would have to make.

Cheers

AGL

SHA-256 acceleration exists for some existing Intel platforms already. However, they're not practically present on anything but servers at the moment, and so I don't think the acceleration of SHA-256 is a something we should consider.

The SUPERCOP benchmarks tell me that generally, on 64-bit systems where acceleration is not available, SHA-256 is the slowest, followed by SHA3-256. BLAKE2b is the fastest.

If our goal is performance, then I would argue BLAKE2b-256 is the best choice. It is secure and extremely fast. It does have the benefit that we get to tell people that by moving away from SHA-1, they will get a performance boost, pretty much no matter what the system.

BLAKE2bp may be faster, but it introduces additional implementation complexity. I'm not sure crypto libraries will implement it, but then again, OpenSSL only implements BLAKE2b-512 at the moment. I don't care much either way, but we should add good tests to exercise the implementation thoroughly. We're generally going to need to ship our own implementation anyway.

I've argued that SHA3-256 probably has the longest life and good unaccelerated performance, and for that reason, I've preferred it. But if AGL says that they're all secure (and I generally think he knows what he's talking about), we could consider performance more.

brian m. carlson / brian with sandals: Houston, Texas, US
https://www.crustytoothpaste.net/~bmc | My opinion only
OpenPGP: https://keybase.io/bk2204

Whatever next-gen hash Git ends up with is going to be in use for decades, so what hardware acceleration exists in consumer products right now is practically irrelevant, but what acceleration is likely to exist for the lifetime of the hash existing *is* relevant.

So I don't follow the argument that we shouldn't weigh future HW acceleration highly just because you can't easily buy a laptop today with these features.

Aside from that I think you've got this backwards, it's AMD that's adding SHA acceleration to their high-end Ryzen chips[1] but Intel is starting at the lower end this year with Goldmont which'll be in lower-end consumer devices[2]. If you read the github issue I linked to upthread[3] you can see that the cryptopp devs already tested their SHA accelerated code on a consumer Celeron[4] recently.

I don't think Intel has announced the SHA extensions for future Xeon releases, but it seems given that they're going to have it there as well. Have there every been x86 extensions that aren't eventually portable across the entire line, or that they've ended up removing from x86 once introduced?

In any case, I think by the time we're ready to follow-up the current hash refactoring efforts with actually changing the hash implementation many of us are likely to have laptops with these extensions, making this easy to test.

1. https://en.wikipedia.org/wiki/Intel_SHA_extensions
2. https://en.wikipedia.org/wiki/Goldmont
3. https://github.com/weidai11/cryptopp/issues/139#issuecomment-264283385
4. https://ark.intel.com/products/95594/Intel-Celeron-Processor-J3455-2M-Cache-up-to-2_3-GHz

The life of MD5 was about 23 years (introduction to first document collision). SHA-1 had about 22. Decades, yes, but just barely. SHA-2 was introduced in 2001, and by the same estimate, we're a little over halfway through its life.

I think you underestimate the life of hardware and software. I have servers running KVM development instances that have been running since at least 2012. Those machines are not scheduled for replacement anytime soon.

Whatever we deploy within the next year is going to run on existing hardware for probably a decade, whether we want it to or not. Most of those machines don't have acceleration.

Furthermore, you need a reasonably modern crypto library to get hardware acceleration. OpenSSL has only recently gained support for it. RHEL 7 does not currently support it, and probably never will. That OS is going to be around for the next 6 years.

If we're optimizing for performance, I don't want to optimize for the latest, greatest machines. Those machines are going to outperform everything else either way. I'd rather optimize for something which performs well on the whole everywhere. There are a lot of developers who have older machines, for cost reasons or otherwise.

Here are some stats (cycles/byte for long messages):

                   SHA-256    BLAKE2b
Ryzen                 1.89       3.06
Knight's Landing     19.00       5.65
Cortex-A72            1.99       5.48
Cortex-A57           11.81       5.47

Cortex-A7 28.19 15.16

In other words, BLAKE2b performs well uniformly across a wide variety of architectures even without acceleration. I'd rather tell people that upgrading to a new hash algorithm is a performance win either way, not just if they have the latest hardware.

brian m. carlson / brian with sandals: Houston, Texas, US
https://www.crustytoothpaste.net/~bmc | My opinion only
OpenPGP: https://keybase.io/bk2204

On Fri, Jun 16 2017, brian m. carlson jotted:

I'm talking about the lifetime of SHA-1 or $newhash's use in Git. As our continued use of SHA-1 demonstrates the window of practical hash function use extends well beyond the window from introduction to published breakage.

It's also telling that SHA-1, which any cryptographer would have waived you off from since around 2011, is just getting widely deployed HW acceleration now in 2017. The practical use of hash functions far exceeds their recommended use in new projects.

To clarify, I'm not dismissing the need to consider existing hardware without these acceleration functions or future processors without them. I don't think that makes any sense, we need to keep those in mind.

I was replying to a bit in your comment where you (it seems to me) were making the claim that we shouldn't consider the HW acceleration of certain hash functions either.

Clearly both need to be considered.

We have real data showing that the intersection between people who care about the hash slowing down and those who can't afford the latest hardware is pretty much nil.

I.e. in 2.13.0 SHA-1 got slower, and pretty much nobody noticed or cared except Johannes Schindelin, myself & Christian Couder. This is because in practice hashing only becomes a bottleneck on huge monorepos that need to e.g. re-hash the contents of a huge index.

Yup, all of those need to be considered, although given my comment above about big repos a 40% improvement on Ryzen (a processor likely to be used for big repos) stands out, where are those numbers from, and is that with or without HW accel for SHA-256 on Ryzen?

Hi,

Yes, I also had the impression that it stressed the status quo quite a bit too much.

We know for a fact that SHA-256 acceleration is coming to consumer CPUs. We know of no plans for any of the other mentioned hash functions to hardware-accelerate them in consumer CPUs.

And remember: for those who are affected most (humongous monorepos, source code hosters), upgrading hardware is less of an issue than having a secure hash function for the rest of us.

And while I am really thankful that Adam chimed in, I think he would agree that BLAKE2 is a purposefully weakened version of BLAKE, for the benefit of speed (with the caveat that one of my experts disagrees that BLAKE2b would be faster than hardware-accelerated SHA-256). And while BLAKE has seen roughly equivalent cryptanalysis as Keccak (which became SHA-3), BLAKE2 has not.

That makes me *very* uneasy about choosing BLAKE2.

Indeed. I am still concerned about that. As you mention, though, it really only affects users of ginormous monorepos, and of course source code hosters.

The jury's still out on how much it impacts my colleagues, by the way.

I have no doubt that Visual Studio Team Services, GitHub and Atlassian will eventually end up with FPGAs for hash computation. So that's that.

Side note: BLAKE is actually *not* friendly to hardware acceleration, I have been told by one cryptography expert. In contrast, the Keccak team claims SHA3-256 to be the easiest to hardware-accelerate, making it "a green cryptographic primitive":
http://keccak.noekeon.org/is_sha3_slow.html

When it comes to BLAKE2, I would actually strongly suggest to consider the amount of attempts to break it. Or rather, how much less attention it got than, say, SHA-256.

In any case, I have been encouraged to stress the importance of "crypto-agility", i.e. the ability to switch to another algorithm when the current one gets broken "enough".

And I am delighted that that is exactly the direction we are going. In other words, even if I still think (backed up by the experts on whose knowledge I lean heavily to form my opinions) that SHA-256 would be the best choice for now, it should be relatively easy to offer BLAKE2b support for (and by [*1*]) those who want it.

Ciao,
Dscho

Footnote *1*: I say that the support for BLAKE2b should come from those parties who desire it also because it is not as ubiquituous as SHA-256. Hence, it would add the burden of having a performant and reasonably bug-free implementation in Git's source tree. IIUC OpenSSL added BLAKE2b support only in OpenSSL 1.1.0, the 1.0.2 line (which is still in use in many places, e.g. Git for Windows' SDK) does not, meaning: Git's implementation would be the one *everybody* relies on, with *no* fall-back.

That is correct.

Although worth keeping in mind that the analysis results from the SHA-3 process informed this rebalancing. Indeed, NIST proposed[1] to do the same with Keccak before stamping it as SHA-3 (although ultimately did not in the context of public feeling in late 2013). The Keccak team have essentially done the same with K12. Thus there is evidence of a fairly widespread belief that the SHA-3 parameters were excessively cautious.

[1] https://docs.google.com/file/d/0BzRYQSHuuMYOQXdHWkRiZXlURVE/edit, slide 48

The numbers given above for SHA-256 on Ryzen and Cortex-A72 must be with hardware acceleration and I thank Brian Carlson for digging them up as I hadn't seen them before.

I suggested above that BLAKE2bp (note the p at the end) might be faster than hardware SHA-256 and that appears to be plausible based on benchmarks[2] of that function. (With the caveat those numbers are for Haswell and Skylake and so cannot be directly compared with Ryzen.)

K12 reports similar speeds on Skylake[3] and thus is also plausibly faster than hardware SHA-256.

[2] https://github.com/sneves/blake2-avx2
[3] http://keccak.noekeon.org/KangarooTwelve.pdf

However, as I'm not a git developer, I've no opinion on whether the cost of carrying implementations of these functions is worth the speed vs using SHA-256, which can be assumed to be supported everywhere already.

Cheers

AGL

Thanks.

My impression from this thread is that even though fast may be better than slow, ubiquity trumps it for our use case, as long as the thing is not absurdly and unusably slow, of course. Which makes me lean towards something older/more established like SHA-256, and it would be a very nice bonus if it gets hardware acceleration more widely than others ;-)

Ah, I recall one thing that was mentioned but not discussed much in the thread: possible use of tree-hashing to exploit multiple cores hashing a large-ish payload. As long as it is OK to pick a sound tree hash coding on top of any (secure) underlying hash function, I do not think the use of tree-hashing should not affect which exact underlying hash function is to be used, and I also am not convinced if we really want tree hashing (some codepaths that deal with a large payload wants to stream the data in single pass from head to tail) in the context of Git, but I am not a crypto person, so ...

Tree hashing also affects single-core performance because of the availability of SIMD instructions.

That is how software implementations of e.g. blake2bp-256 and SHA-256x16[1] are able to have competitive performance with (slightly better performance than, at least in some cases) hardware implementations of SHA-256.

It is also satisfying that we have options like these that are faster than SHA-1.

All that said, SHA-256 seems like a fine choice, despite its worse performance. The wide availability of reasonable-quality implementations (e.g. in Java you can use 'MessageDigest.getInstance("SHA-256")') makes it a very tempting one.

Part of the reason I suggested previously that it would be helpful to try to benchmark Git with various hash functions (which didn't go over well, for some reason) is that it makes these comparisons more concrete. Without measuring, it is hard to get a sense of the distribution of input sizes and how much practical effect the differences we are talking about have.

Thanks,
Jonathan

[1] https://eprint.iacr.org/2012/476.pdf

On Fri, Jun 16 2017, Jonathan Nieder jotted:

It would be great to have such benchmarks (I probably missed the "didn't go over well" part), but FWIW you can get pretty close to this right now in git by running various t/perf benchmarks with BLKSHA1/OPENSSL/SHA1DC.

Between the three of those (particularly SHA1DC being slower than OpenSSL) you get a similar performance difference as some SHA-1 v.s. SHA-256 benchmarks I've seen, so to the extent that we have existing performance tests it's revealing to see what's slower & faster.

It makes a particularly big difference for e.g. p3400-rebase.sh.

I actually doubt this from the GitHub side. Hash performance is not even on our radar as a bottleneck. In most cases the problem is touching uncompressed data _at all_, not computing the hash over it (so things like reusing on-disk deltas are really important).

-Peff

Hi Peff,

Thanks for pointing that out! As a mainly client-side person, I rarely get insights into the server side...

Ciao,
Dscho

AIUI, length extension does make brute force collision attacks (which, really Shattered was) cheaper by allowing one to create the collision with a small message and extend it later.

This might not be a credible thread against git, but if we go by that standard, post-shattered Sha-1 is still fine for git. As a matter of fact, MD5 would also be fine: there is still, to this day, no preimage attack against them.

Mike

I think collision attacks are of interest to Git. But I would think 2^128 would be enough (TBH, 2^80 probably would have been enough for SHA-1; it was the weaknesses that brought that down by a factor of a million that made it a problem).

-Peff

100% agree on this point. I believe the current plan is to have the hashing function used for a repository be a repository format extension which would be a value (most likely a string like 'sha1', 'sha256', 'black2', etc) stored in a repository's .git/config. This way, upon startup git will die or ignore a repository which uses a hashing function which it does not recognize or does not compiled to handle.

I hope (and expect) that the end produce of this transition is a nice, clean hashing API and interface with sufficient abstractions such that if I wanted to switch to a different hashing function I would just need to implement the interface with the new hashing function and ensure that 'verify_repository_format' allows the new function.

Brandon Williams

Yup. I thought that part has already been agreed upon, but it is a good thing that somebody is writing it down (perhaps "again", if not "for the first time").

Thanks.

Hi Dscho,

I don't completely agree with this framing. The essential question is:
how to get the right security properties without abysmal performance.

[... etc ...]

Thanks for a thoughtful restart to the discussion. This is much more concrete than your previous objections about process, and that is very helpful.

In the interest of transparency: here are my current questions for cryptographers to whom I have forwarded this thread. Several of these questions involve predictions or opinions, so in my ideal world we'd want multiple, well reasoned answers to them. Please feel free to forward them to appropriate people or add more.

1. Now it sounds like SHA-512/256 is the safest choice (see also Mike Hommey's response to Dscho's message). Please poke holes in my understanding.

2. Would you be willing to weigh in publicly on the mailing list? I think that would be the most straightforward way to move this forward (and it would give you a chance to ask relevant questions, etc). Feel free to contact me privately if you have any questions about how this particular mailing list works.

3. On the speed side, Dscho states "SHA-256 will be faster than BLAKE (and even than BLAKE2) once the Intel and AMD CPUs with hardware support for SHA-256 become common." Do you agree?

4. On the security side, Dscho states "to compete in the SHA-3 contest, BLAKE added complexity so that it would be roughly on par with its competitors. To allow for faster execution in software, this complexity was *removed* from BLAKE to create BLAKE2, making it weaker than SHA-256." Putting aside the historical questions, do you agree with this "weaker than" claim?

5. On the security side, Dscho states, "The type of attacks Git has to worry about is very different from the length extension attacks, and it is highly unlikely that that weakness of SHA-256 leads to, say, a collision attack", and Jeff King states, "Git does not use the hash as a MAC, so length extension attacks aren't a thing (and even if we later wanted to use the same algorithm as a MAC, the HMAC construction is a well-studied technique for dealing with it)." Is this correct in spirit? Is SHA-256 equally strong to SHA-512/256 for Git's purposes, or are the increased bits of internal state (or other differences) relevant? How would you compare the two functions' properties?

6. On the speed side, Jeff King states "That said, SHA-512 is typically a little faster than SHA-256 on 64-bit platforms. I don't know if that will change with the advent of hardware instructions oriented towards SHA-256." Thoughts?

7. If the answer to (2) is "no", do I have permission to quote or paraphrase your replies that were given here?

Thanks, sincerely,
Jonathan

Thanks for the kind words on what had quite a few flaws still. Here's a new draft. I think the next version will be a patch against Documentation/technical/.

As before, comments welcome, both here and inline at

https://goo.gl/gh2Mzc

Changes since v2:

Use SHA3-256 instead of SHA2 (thanks, Linus and brian m. carlson).[1][2]

Make sha3-based signatures a separate field, avoiding the need for "hash" and "nohash" fields (thanks to peff[3]).

Add a sorting phase to fetch (thanks to Junio for noticing the need for this).

Omit blobs from the topological sort during fetch (thanks to peff).

Discuss alternates, git notes, and git servers in the caveats section (thanks to Junio Hamano, brian m. carlson[4], and Shawn Pearce).

Clarify language throughout (thanks to various commenters, especially Junio).

Sincerely,
Jonathan

Git hash function transition
============================
Status: Draft
Last Updated: 2017-03-06

Objective
---------
Migrate Git from SHA-1 to a stronger hash function.

Background
----------
At its core, the Git version control system is a content addressable filesystem. It uses the SHA-1 hash function to name content. For example, files, directories, and revisions are referred to by hash values unlike in other traditional version control systems where files or versions are referred to via sequential numbers. The use of a hash function to address its content delivers a few advantages:

* Integrity checking is easy. Bit flips, for example, are easily detected, as the hash of corrupted content does not match its name.
* Lookup of objects is fast.

Using a cryptographically secure hash function brings additional advantages:

* Object names can be signed and third parties can trust the hash to address the signed object and all objects it references.
* Communication using Git protocol and out of band communication methods have a short reliable string that can be used to reliably address stored content.

Over time some flaws in SHA-1 have been discovered by security researchers. https://shattered.io demonstrated a practical SHA-1 hash collision. As a result, SHA-1 cannot be considered cryptographically secure any more. This impacts the communication of hash values because we cannot trust that a given hash value represents the known good version of content that the speaker intended.

SHA-1 still possesses the other properties such as fast object lookup and safe error checking, but other hash functions are equally suitable that are believed to be cryptographically secure.

Goals
-----
1. The transition to SHA3-256 can be done one local repository at a time. a. Requiring no action by any other party.
b. A SHA3-256 repository can communicate with SHA-1 Git servers

      (push/fetch).

c. Users can use SHA-1 and SHA3-256 identifiers for objects
interchangeably.
d. New signed objects make use of a stronger hash function than
SHA-1 for their security guarantees.
2. Allow a complete transition away from SHA-1.
a. Local metadata for SHA-1 compatibility can be removed from a
repository if compatibility with SHA-1 is no longer needed.
3. Maintainability throughout the process.
a. The object format is kept simple and consistent. b. Creation of a generalized repository conversion tool.

Non-Goals
---------
1. Add SHA3-256 support to Git protocol. This is valuable and the logical next step but it is out of scope for this initial design.
2. Transparently improving the security of existing SHA-1 signed objects.
3. Intermixing objects using multiple hash functions in a single repository.
4. Taking the opportunity to fix other bugs in git's formats and protocols.
5. Shallow clones and fetches into a SHA3-256 repository. (This will change when we add SHA3-256 support to Git protocol.)
6. Skip fetching some submodules of a project into a SHA3-256 repository. (This also depends on SHA3-256 support in Git protocol.)

Overview
--------
We introduce a new repository format extension sha3. Repositories with this extension enabled use SHA3-256 instead of SHA-1 to name their objects. This affects both object names and object content --- both the names of objects and all references to other objects within an object are switched to the new hash function.

sha3 repositories cannot be read by older versions of Git.

Alongside the packfile, a sha3 repository stores a bidirectional mapping between sha3 and sha1 object names. The mapping is generated locally and can be verified using "git fsck". Object lookups use this mapping to allow naming objects using either their sha1 and sha3 names interchangeably.

"git cat-file" and "git hash-object" gain options to display an object in its sha1 form and write an object given its sha1 form. This requires all objects referenced by that object to be present in the object database so that they can be named using the appropriate name (using the bidirectional hash mapping).

Fetches from a SHA-1 based server convert the fetched objects into sha3 form and record the mapping in the bidirectional mapping table (see below for details). Pushes to a SHA-1 based server convert the objects being pushed into sha1 form so the server does not have to be aware of the hash function the client is using.

Detailed Design
---------------
Object names
~~~~~~~~~~~~
Objects can be named by their 40 hexadecimal digit sha1-name or 64 hexadecimal digit sha3-name, plus names derived from those (see gitrevisions(7)).

The sha1-name of an object is the SHA-1 of the concatenation of its type, length, a nul byte, and the object's sha1-content. This is the traditional <sha1> used in Git to name objects.

The sha3-name of an object is the SHA3-256 of the concatenation of its type, length, a nul byte, and the object's sha3-content.

Object format
~~~~~~~~~~~~~
The content as a byte sequence of a tag, commit, or tree object named by sha1 and sha3 differ because an object named by sha3-name refers to other objects by their sha3-names and an object named by sha1-name refers to other objects by their sha1-names.

The sha3-content of an object is the same as its sha1-content, except that objects referenced by the object are named using their sha3-names instead of sha1-names. Because a blob object does not refer to any other object, its sha1-content and sha3-content are the same.

The format allows round-trip conversion between sha3-content and sha1-content.

Object storage
~~~~~~~~~~~~~~
Loose objects use zlib compression and packed objects use the packed format described in Documentation/technical/pack-format.txt, just like today. The content that is compressed and stored uses sha3-content instead of sha1-content.

Translation table
~~~~~~~~~~~~~~~~~
A fast bidirectional mapping between sha1-names and sha3-names of all local objects in the repository is kept on disk. The exact format of that mapping is to be determined.

All operations that make new objects (e.g., "git commit") add the new objects to the translation table.

(This work could have been deferred to push time, but that would significantly complicate and slow down pushes. Calculating the sha1-name at object creation time at the same time it is being streamed to disk and having its sha3-name calculated should be an acceptable cost.)

Reading an object's sha1-content
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The sha1-content of an object can be read by converting all sha3-names its sha3-content references to sha1-names using the translation table.

Fetch
~~~~~
Fetching from a SHA-1 based server requires translating between SHA-1 and SHA3-256 based representations on the fly.

SHA-1s named in the ref advertisement that are present on the client can be translated to SHA3-256 and looked up as local objects using the translation table.

Negotiation proceeds as today. Any "have"s generated locally are converted to SHA-1 before being sent to the server, and SHA-1s mentioned by the server are converted to SHA3-256 when looking them up locally.

After negotiation, the server sends a packfile containing the requested objects. We convert the packfile to SHA3-256 format using the following steps:

1. index-pack: inflate each object in the packfile and compute its SHA-1. Objects can contain deltas in OBJ_REF_DELTA format against objects the client has locally. These objects can be looked up using the translation table and their sha1-content read as described above to resolve the deltas.
2. topological sort: starting at the "want"s from the negotiation phase, walk through objects in the pack and emit a list of them, excluding blobs, in reverse topologically sorted order, with each object coming later in the list than all objects it references. (This list only contains objects reachable from the "wants". If the pack from the server contained additional extraneous objects, then they will be discarded.)
3. convert to sha3: open a new (sha3) packfile. Read the topologically sorted list just generated. For each object, inflate its sha1-content, convert to sha3-content, and write it to the sha3 pack. Include the new sha1<->sha3 mapping entry in the translation table.
4. sort: reorder entries in the new pack to match the order of objects in the pack the server generated and include blobs. Write a sha3 idx file.
5. clean up: remove the SHA-1 based pack file, index, and topologically sorted list obtained from the server and steps 1 and 2.

Step 3 requires every object referenced by the new object to be in the translation table. This is why the topological sort step is necessary.

As an optimization, step 1 could write a file describing what non-blob objects each object it has inflated from the packfile references. This makes the topological sort in step 2 possible without inflating the objects in the packfile for a second time. The objects need to be inflated again in step 3, for a total of two inflations.

Step 4 is probably necessary for good read-time performance. "git pack-objects" on the server optimizes the pack file for good data locality (see Documentation/technical/pack-heuristics.txt).

Details of this process are likely to change. It will take some experimenting to get this to perform well.

Push
~~~~
Push is simpler than fetch because the objects referenced by the pushed objects are already in the translation table. The sha1-content of each object being pushed can be read as described in the "Reading an object's sha1-content" section to generate the pack written by git send-pack.

Signed Commits
~~~~~~~~~~~~~~
We add a new field "gpgsig-sha3" to the commit object format to allow signing commits without relying on SHA-1. It is similar to the existing "gpgsig" field. Its signed payload is the sha3-content of the commit object with any "gpgsig" and "gpgsig-sha3" fields removed.

This means commits can be signed
1. using SHA-1 only, as in existing signed commit objects
2. using both SHA-1 and SHA3-256, by using both gpgsig-sha3 and gpgsig fields.
3. using only SHA3-256, by only using the gpgsig-sha3 field.

Old versions of "git verify-commit" can verify the gpgsig signature in cases (1) and (2) without modifications and view case (3) as an ordinary unsigned commit.

Signed Tags
~~~~~~~~~~~
We add a new field "gpgsig-sha3" to the tag object format to allow signing tags without relying on SHA-1. Its signed payload is the sha3-content of the tag with its gpgsig-sha3 field and "-----BEGIN PGP SIGNATURE-----" delimited in-body signature removed.

This means tags can be signed
1. using SHA-1 only, as in existing signed tag objects
2. using both SHA-1 and SHA3-256, by using gpgsig-sha3 and an in-body signature.
3. using only SHA3-256, by only using the gpgsig-sha3 field.

Mergetag embedding
~~~~~~~~~~~~~~~~~~
The mergetag field in the sha1-content of a commit contains the sha1-content of a tag that was merged by that commit.

The mergetag field in the sha3-content of the same commit contains the sha3-content of the same tag.

Submodules
~~~~~~~~~~
To convert recorded submodule pointers, you need to have the converted submodule repository in place. The translation table of the submodule can be used to look up the new hash.

Caveats
-------
Invalid objects
~~~~~~~~~~~~~~~
The conversion from sha1-content to sha3-content retains any brokenness in the original object (e.g., tree entry modes encoded with leading 0, tree objects whose paths are not sorted correctly, and commit objects without an author or committer). This is a deliberate feature of the design to allow the conversion to round-trip.

More profoundly broken objects (e.g., a commit with a truncated "tree" header line) cannot be converted but were not usable by current Git anyway.

Shallow clone and submodules
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Because it requires all referenced objects to be available in the locally generated translation table, this design does not support shallow clone or unfetched submodules. Protocol improvements might allow lifting this restriction.

Alternates
~~~~~~~~~~
For the same reason, a sha3 repository cannot borrow objects from a sha1 repository using objects/info/alternates or $GIT_ALTERNATE_OBJECT_REPOSITORIES.

git notes
~~~~~~~~~
The "git notes" tool annotates objects using their sha1-name as key. This design does not describe a way to migrate notes trees to use sha3-names. That migration is expected to happen separately (for example using a file at the root of the notes tree to describe which hash it uses).

Server-side cost
~~~~~~~~~~~~~~~~
Until Git protocol gains SHA3-256 support, using sha3 based storage on public-facing Git servers is strongly discouraged. Once Git protocol gains SHA3-256 support, sha3 based servers are likely not to support sha1 compatibility, to avoid what may be a very expensive hash reencode during clone and to encourage peers to modernize.

The design described here allows fetches by SHA-1 clients of a personal SHA256 repository because it's not much more difficult than allowing pushes from that repository. This support needs to be guarded by a configuration option --- servers like git.kernel.org that serve a large number of clients would not be expected to bear that cost.

Meaning of signatures
~~~~~~~~~~~~~~~~~~~~~
The signed payload for signed commits and tags does not explicitly name the hash used to identify objects. If some day Git adopts a new hash function with the same length as the current SHA-1 (40 hexadecimal digit) or SHA2-256 (64 hexadecimal digit) objects then the intent behind the PGP signed payload in an object signature is unclear:

	object e7e07d5a4fcc2a203d9873968ad3e6bd4d7419d7
	type commit
	tag v2.12.0
	tagger Junio C Hamano <[email protected]> 1487962205 -0800

	Git 2.12

Does this mean Git v2.12.0 is the commit with sha1-name e7e07d5a4fcc2a203d9873968ad3e6bd4d7419d7 or the commit with new-40-digit-hash-name e7e07d5a4fcc2a203d9873968ad3e6bd4d7419d7?

Fortunately SHA3-256 and SHA-1 have different lengths. If Git starts using another hash with the same length to name objects, then it will need to change the format of signed payloads using that hash to address this issue.

Alternatives considered
-----------------------
Upgrading everyone working on a particular project on a flag day
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Projects like the Linux kernel are large and complex enough that flipping the switch for all projects based on the repository at once is infeasible.

Not only would all developers and server operators supporting developers have to switch on the same flag day, but supporting tooling (continuous integration, code review, bug trackers, etc) would have to be adapted as well. This also makes it difficult to get early feedback from some project participants testing before it is time for mass adoption.

Using hash functions in parallel
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
(e.g. https://public-inbox.org/git/[email protected]/ ) Objects newly created would be addressed by the new hash, but inside such an object (e.g. commit) it is still possible to address objects using the old hash function.
* You cannot trust its history (needed for bisectability) in the
future without further work
* Maintenance burden as the number of supported hash functions grows (they will never go away, so they accumulate). In this proposal, by comparison, converted objects lose all references to SHA-1.

Signed objects with multiple hashes
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Instead of introducing the gpgsig-sha3 field in commit and tag objects for sha3-content based signatures, an earlier version of this design added "hash sha3 <sha3-name>" fields to strengthen the existing sha1-content based signatures.

In other words, a single signature was used to attest to the object content using both hash functions. This had some advantages:
* Using one signature instead of two speeds up the signing process.
* Having one signed payload with both hashes allows the signer to attest to the sha1-name and sha3-name referring to the same object.
* All users consume the same signature. Broken signatures are likely to be detected quickly using current versions of git.

However, it also came with disadvantages:
* Verifying a signed object requires access to the sha1-names of all objects it references, even after the transition is complete and translation table is no longer needed for anything else. To support this, the design added fields such as "hash sha1 tree <sha1-name>" and "hash sha1 parent <sha1-name>" to the sha3-content of a signed commit, complicating the conversion process.
* Allowing signed objects without a sha1 (for after the transition is complete) complicated the design further, requiring a "nohash sha1" field to suppress including "hash sha1" fields in the sha3-content and signed payload.

Document History
----------------

2017-03-03
[email protected], [email protected], [email protected], [email protected]

Initial version sent to
http://public-inbox.org/git/[email protected]

2017-03-03 [email protected]
Incorporated suggestions from jonathantanmy and sbeller:
* describe purpose of signed objects with each hash type
* redefine signed object verification using object content under the first hash function

2017-03-06 [email protected]
* Use SHA3-256 instead of SHA2 (thanks, Linus and brian m. carlson).[1][2]
* Make sha3-based signatures a separate field, avoiding the need for "hash" and "nohash" fields (thanks to peff[3]).
* Add a sorting phase to fetch (thanks to Junio for noticing the need for this).
* Omit blobs from the topological sort during fetch (thanks to peff).
* Discuss alternates, git notes, and git servers in the caveats section (thanks to Junio Hamano, brian m. carlson[4], and Shawn Pearce).
* Clarify language throughout (thanks to various commenters, especially Junio).

[1] http://public-inbox.org/git/CA+55aFzJtejiCjV0e43+9oR3QuJK2PiFiLQemytoLpyJWe6P9w@mail.gmail.com/
[2] http://public-inbox.org/git/CA+55aFz+gkAsDZ24zmePQuEs1XPS9BP_s8O7Q4wQ7LV7X5-oDA@mail.gmail.com/
[3] http://public-inbox.org/git/[email protected]/
[4] http://public-inbox.org/git/[email protected]

FWIW, I like this approach.

I saw some discussion about using LevelDB for this mapping table. I think any existing database may be overkill.

For packs, you may be able to simplify by having only one file (pack-*.msha1) that maps SHA-1 to pack offset; idx v2. The CRC32 table in v2 is unnecessary, but you need the 64 bit offset support.

SHA-1 to SHA-3: lookup SHA-1 in .msha1, reverse .idx, find offset to read the SHA-3.
SHA-3 to SHA-1: lookup SHA-3 in .idx, and reverse the .msha1 file to translate offset to SHA-1.

For loose objects, the loose object directories should have only O(4000) entries before auto gc is strongly encouraging packing/pruning. With 256 shards, each given directory has O(16) loose objects in it. When writing a SHA-3 loose object, Git could also append a line "$sha3 $sha1\n" to objects/${first_byte}/sha1, which GC/prune rewrites to remove entries. With O(16) objects in a directory, these files should only have O(16) entries in them.

SHA-3 to SHA-1: open objects/${sha3_first_byte}/sha1 and scan until a match is found.
SHA-1 to SHA-3: brute force read 256 files. Callers performing this mapping may load all 256 files into a table in memory.

Hi,

Thanks for this suggestion. I was initially vaguely nervous about lookup times in an idx-style file, but as you say, object reads from a packfile already have to deal with this kind of lookup and work fine.

Insertion time is what worries me. When writing a small number of objects using a command like "git commit", I don't want to have to regenerate an entire idx file. I don't want to move the pain to O(loose objects) work at read time, either --- some people disable auto gc, and others have a large number of loose objects due to gc ejecting unreachable objects.

But some kind of simplification along these lines should be possible. I'll experiment.

Jonathan

Not exactly. The "reverse .idx" step has to build the reverse mapping on the fly, and it's non-trivial. For instance, try:

  sha1=$(git rev-parse HEAD)
  time echo $sha1 | git cat-file --batch-check='%(objectsize)'
  time echo $sha1 | git cat-file --batch-check='%(objectsize:disk)'

on a large repo (where HEAD is in a big pack). The on-disk size is conceptually simpler, as we only need to look at the offset of the object versus the offset of the object after it. But in practice it takes much longer, because it has to build the revindex on the fly (I get 7ms versus 179ms on linux.git).

The effort is linear in the number of objects (we create the revindex with a radix sort).

The reachability bitmaps suffer from this, too, as they need the revindex to know which object is at which bit position. At GitHub we added an extension to the .bitmap files that stores this "bit cache". Here are timings before and after on linux.git:

  $ time git rev-list --use-bitmap-index --count master
  659371

real 0m0.182s
user 0m0.136s
sys 0m0.044s

  $ time git.gh rev-list --use-bitmap-index --count master
  659371

real 0m0.016s
user 0m0.008s
sys 0m0.004s

It's not a full revindex, but it's enough for bitmap use. You can also use it to generate the revindex slightly more quickly, because you can skip the sorting step (you just insert the entries in the correct order by walking the bit cache and dereferencing the offsets from the .idx portion). So it's still linear, but with a smaller constant factor.

I think for the purposes here, though, we don't actually care about the offsets. For the cost of one uint32_t per object, you can keep a list mapping positions in the sha1 index into the sha3 index. So then you do the log-n binary search to find the sha1, a constant-time lookup in the mapping array, and that gives you the position in the sha3 index, from which you can then access the sha3 (or the actual pack offset, for that matter).

So I think it's solvable, but I suspect we would want an extension to the .idx format to store the mapping array, in order to keep it log-n.

-Peff

Sure. To be clear, I was handwaving over that since adding an on-disk reverse .idx is a relatively small change.

[...]

i.e., this.

The loose object side is the more worrying bit, since we currently don't have any practical bound on the number of loose objects.

One way to deal with that is to disallow loose objects completely. Use packfiles for new objects, batching the objects produced by a single process into a single packfile. Teach "git gc --auto" a behavior similar to Martin Fick's "git exproll" to combine packfiles between full gcs to maintain reasonable performance. For unreachable objects, instead of using loose objects, use "unreachable garbage" packs explicitly labeled as such, with similar semantics to what JGit's DfsRepository backend uses (described in the discussion at
https://git.eclipse.org/r/89455).

That's a direction that I want in the long term anyway. I was hoping not to couple such changes with the hash transition but it might be one of the simpler ways to go.

Jonathan

This document describes what a transition to a new hash function for Git would look like. Add it to Documentation/technical/ as the plan of record so that future changes can be recorded as patches.

---

Okay, here goes.

Instead of sharding the loose object translation tables by first byte, we went for a single table. It simplifies the design and we need to keep the number of loose objects under control anyway.

We also included a description of the transition plan and tried to include a summary of what has been agreed upon so far about the choice of hash function.

Thanks to Junio for reviving the discussion and in particular to Dscho for pushing this forward and making the missing pieces clearer.

Thoughts of all kinds welcome, as always.

 Documentation/Makefile                             |   1 +
 .../technical/hash-function-transition.txt         | 797 +++++++++++++++++++++
 2 files changed, 798 insertions(+)
 create mode 100644 Documentation/technical/hash-function-transition.txt

index 2415e0d657..471bb29725 100644
@@ -67,6 +67,7 @@ SP_ARTICLES += howto/maintain-git
 API_DOCS = $(patsubst %.txt,%,$(filter-out technical/api-index-skel.txt technical/api-index.txt, $(wildcard technical/api-*.txt)))
 SP_ARTICLES += $(API_DOCS)
 
+TECH_DOCS += technical/hash-function-transition
 TECH_DOCS += technical/http-protocol
 TECH_DOCS += technical/index-format
 TECH_DOCS += technical/pack-format

2.14.2.822.g60be5d43e6-goog

Shoudln't these all be s-o-b: (with a note immediately before that to say all four contributed equally or something)?

I'd stop here and continue in a separate message. Thanks for a detailed write-up.

Let's keep it simple by ignoring all of the above. Even though leaving the sha1-gpgsig and other crufts would etch these compatibility metadata in objects forever, these remain only in objects that originate from SHA-1 world, or in objects created in the NewHash world only while the project participants still care about SHA-1 compatibility. Strictly speaking, it would be super nice if we can do without contaminating these newly created objects with SHA-1 compatibility headers, just like we wish to be able to drop the SHA-1 vs NewHash mapping table after projects participants stop careing about SHA-1 compatiblity, it may not be worth it. Of course, if we decide to spend a bit more brain cycle to design how we push these out of the object proper, the same solution would automatically allow us to omit SHA-1 compatibility headers from the objects that were converted from SHA-1 world.

Thinking about this a bit more, I think a single table per .idx file would be the right way to go, not a checksum immediately after or before the object data that is embedded in the pack stream. In the NewHash world (after this initial migration), we would want to be able to stream NewHash packstream that comes from the network straight to disk, which would mean these in-line CRC32 data would need to be sent over the wire (i.e. 4-byte per object sent); that is an unneeded overhead, as the packstream has its trailing checksum to protect the whole thing anyway.

I don't want to get lost in the weeds in the question of how to represent such a collaborative effort in git's metadata.

You're right that I should collect their sign-offs! Your approach of using text instead of machine-readable data for common authorship also seems okay.

In any event, this is indeed

(I just checked :)).

Good point. The wording you suggested will work well.

:) Yep, these are nice too, especially that second one.

It also is how we make diff-ing fast.

We were looking for a phrase meaning that it should be a cryptographic hash function in good standing, which SHA-1 is at least approaching not being.

"more secure" should work fine. Let's go with that.

Will add a note about this to "Object names on the command line". Stefan had already pointed out that that section should really be renamed to something like "Object names in input and output".

That would break round-tripping and would mean that multiple SHA-1 objects could have the same NewHash name. In other words, from my point of view there is something that says that such data must be preserved.

Another way to put it: even after removing all SHA-1 compatibility metadata, one nice feature of this design is that it can be recovered if I change my mind, from data in the NewHash based repository alone.

[...]

It refers to that issue.

[...]

I think we can delete this paragraph. It was written before the "Object names on the command line" section that goes into such issues in more detail.

[...]

Do you mean that adding the hashname into the computation that produces the object name would help in some use case?

Or do you mean storing the hashname on disk somewhere, even if it doesn't enter into the object name? For the latter, we store the hashname in the .git/config extensions.* configuration and the pack index files. You also suggested storing the hash name in .git/objects/loose-object-idx, which seems to me like a good idea.

We didn't touch on the .pack format but we probably need to (if only because of the size of REF_DELTAs and the cksum trailer), and it would also need to name what object format it is using.

For loose objects, it would be nice to name the hash in the file, so that "file" can understand what is happening if someone accidentally mixes types using "cp". The only downside is losing the ability to copy blobs (which have the same content despite being named using different hashes) between repositories after determining their new names. That doesn't seem like a strong downside --- it's pretty harmless to include the hash type in loose object files, too. I think I would prefer this to be a "magic number" instead of part of the zlib-deflated payload, since this way "file" can discover it more easily.

I don't consider that desirable.

A SHA-1 based signature is still of historical interest even if my centuries-newer version of Git is not able to verify it.

[...]

Yes. Will clarify.

Yes. The idx v2 format has a similar design.

As you mentioned under separate cover, it is useful for derived data like this to be outside the packfile.

It's actually only present for the first object format. Will find a better way to describe this.

An earlier version of the design also had a pack-order-to-pack-offset table, but we weren't able to think of any cases where that would be used without also looking up the object name that can be used to verify the integrity of the inflated object.

Do you have an application in mind?

[...]

Yes, good catch.

[...]

Good idea!

[...]

Yes, I think we do need to fsync. :/

[...]

Yes.

Thanks for looking it over.

Jonathan

We may want to find a way to make it clear, then.

What I mean is that for SHA-1 objects we keep the object header to be "<type> <length> NUL". For objects in newer world, use the object header to "<type> <hash> <length> NUL", and include the hashname in the object name computation.

Yeah, thanks for doing pros-and-cons for me ;-)

Agreed. Let's not go there.

I see. One way to do so is to have it upfront before the "after this point, these tables repeat for each of the hashes" part of the file.

The primary thing I was interested in knowing was if we tried to think of any case where it may be useful and then didn't think of any---I couldn't but I know I am not imaginative enough, and I wanted to know you guys didn't, either.

...

Agreed, even a signature made by a now exposed and revoked key still has validity. Especially in a commit or merge. We know it was made prior to the key being compromised / revoked.

This is assuming that the keyholder can definitively say "Don't trust signatures from this key after this date/time+0000". And the signature in question is in the git history prior to that cut off.

Tags are a different animal because they can be added at any time and aren't directly incorporated into the history.

thx,

Jason.

This message stops here for now.

Hi Jonathan,

...

...

...

This message stops here...

...

I do hope we adhere to "no mixed repository" rule. Or, at least, "no mixing of hash types". Ambiguity opens cracks for uncertainty to creep in.

For our case, where we counter-hash the sha1 commits, and counter-sign the sha1-based signatures, we intend to include the relevant sha1<->newhash lookups in the newhash signature body. afaict, the git sha1<->newhash table is not cryptographically secured underneath signatures, and thus can't be used in the verification of objects.

The advantage to this approach is that we can be as explicit as necessary with "SHA-1 -> SHA-512/256" or "SHA-1 -> SHA3-256" in the body of the message.

thx,

Jason.

Can we reboot the discussion and advance this to v4 state?

I think what you have over there looks pretty-much ready as the final outline.

One thing I still do not know how I feel about after re-reading the thread, and I didn't find the above doc, is Linus's suggestion to use the objects themselves as NewHash-to-SHA-1 mapper [*1*].

It does not help the reverse mapping that is needed while pushing things out (the SHA-1 receiver tells us what they have in terms of SHA-1 names; we need to figure out where we stop sending based on that). While it does help maintaining itself (while constructing SHA3-content, we'd be required to find out its SHA1 name but the SHA3 objects that we refer to all know their SHA-1 names), if it is not useful otherwise, then that does not count as a plus. Also having to bake corresponding SHA-1 name in the object would mean mistakes can easily propagate and cannot be corrected without rewriting the history, which would be a huge downside. So perhaps we are better off without it, I guess.

[Reference]

*1* <CA+55aFxj7Vtwac64RfAz_u=U4tob4Xg+2pDBDFNpJdmgaTCmxA@mail.gmail.com>

I think this falls into the same category as the often-talked-about addition of the "generation number" field. It is very tempting to add these "mechanically derivable but expensive to compute" pieces of information to the sha3-content while converting from sha1-content and creating anew.

Because the "sha1-name" or the "generation number" can mechanically be computed, as long as everybody agrees to _always_ place them in the sha3-content, the same sha1-content will be converted into exactly the same sha3-content without ambiguity, and converting them back to sha1-content while pushing to an older repository will correctly produce the original sha1-content, as it would just be the matter of simply stripping these extra pieces of information.

The reason why I still feel a bit uneasy about adding these things (aside from the fact that sha1-name thing will be a baggage we would need to carry forever even after we completely wean ourselves off of the old hash) is because I am not sure what we should do when we encounter sha3-content in the wild that has these things _wrong_. An object that exists today in the SHA-1 world is fetched into the new repository and converted to SHA-3 contents, and Linus's extra "original SHA-1 name" field is added to the object's header while recording the SHA-3 content. But for whatever reason, the original SHA-1 name is recorded incorrectly in the resulting SHA-3 object.

The same thing could happen if we decide to bake "generation number" in the SHA-3 commit objects. One possible definition would be that a root commit will have gen #0; a commit with 1 or more parents will get max(parents' gen numbers) + 1 as its gen number. But somebody may botch the counting and records sum(parents' gen numbers) as its gen number.

In these cases, not just the SHA3-content but also the resulting SHA-3 object name would be different from the name of the object that would have recorded the same contents correctly. So converting back to SHA-1 world from these botched SHA-3 contents may produce the original contents, but we may end up with multiple "plausibly looking" set of SHA-3 objects that (clain to) correspond to a single SHA-1 object, only one of which is a valid one.

Our "git fsck" already treats certain brokenness (like a tree whose entry has mode that is 0-padded to the left) as broken but still tolerate them. I am not sure if it is sufficient to diagnose and declare broken and invalid when we see sha3-content that records these "mechanically derivable but expensive to compute" pieces of information incorrectly.

I am leaning towards saying "yes, catching in fsck is enough" and suggesting to add generation number to sha3-content of the commit objects, and to add even the "original sha1 name" thing if we find good use of it. But I cannot shake this nagging feeling off that I am missing some huge problems that adding these fields and opening ourselves to more classes of broken objects.

Thoughts?

I share your nagging feeling.

I have two thoughts on the "fsck can catch it" line of reasoning.

1. It's harder to fsck generation numbers than other syntactic elements of an object, because it inherently depends on the links. So I can't fsck a commit object in isolation. I have to open its parents and check _their_ generation numbers.

In some sense that isn't a big deal. A real fsck wants to know that we _have_ the parents in the first place. But traditionally we've separated "is this syntactically valid" from "do we have full connectivity". And features like shallow clones rely on us fudging the latter but not the former. A shallow history could never properly fsck the generation numbers.

A multiple-hash field doesn't have this problem. It's purely a function of the bytes in the object.

2. I wouldn't classify the current fsck checks as a wild success in containing breakages. If a buggy implementation produces invalid objects, the same buggy implementation generally lets people (and their colleagues) unwittingly build on top of those objects. It's only later (sometimes much later) that they interact with a non-buggy implementation whose fsck complains.

And what happens then? If they're lucky, the invalid objects haven't spread far, and the worst thing is that they have to learn to use filter-branch (which itself is punishment enough). But sometimes a significant bit of history has been built on top, and it's awkward or impossible to rewrite it.

That puts the burden on whoever is running the non-buggy implementation that wants to reject the objects. Do they accept these broken objects? If so, what do they do to mitigate the wrong answers that Git will return?

I'm much more in favor of keeping that data outside the object-hash computation, and caching the pre-computed results as necessary. Those cache can disagree with the objects, of course, but the cost to dropping and re-building them is much lower than a history rewrite.

I'm speaking primarily to the generation-number thing, where I really don't think there's any benefit to embedding it in the object beyond the obvious "well, it has to go _somewhere_, and this saves us implementing a local cache layer". I haven't thought hard enough on the multiple-hash thing to know if there's some other benefit to having it inside the objects.

-Peff

We didn't discuss that in the doc since this particular transition plan we made uses an external NewHash-to-SHA1 map instead of an internal one because we believe that at some point we would be able to drop compatibility with SHA1. Now I suspect that wont happen for a long time but I think it would be preferable over carrying the SHA1 luggage indefinitely. At some point, then, we would be able to stop hashing objects twice (once with SHA1 and once with NewHash) instead of always requiring that we hash them with each hash function which was used historically.

This wasn't one of the issues that I thought of but it just makes the argument against adding sha1's to the sha3 content stronger.

Brandon Williams

Hi Brandon,

Is there even a question about that? I mean, why would *any* project that switches entirely to SHA-256 want to carry the SHA-1 baggage around?

So even if the code to generate a bidirectional old <-> new hash mapping might be with us forever, it *definitely* should be optional ("optional" at least as in "config setting"), allowing developers who only work with new-hash repositories to save the time and electrons.

It should be possible to push back the SHA-1 ginny into a small gin bottle inside Git's source code, so to say, i.e. encapsulate it to the point where it is a compile-time option, in addition to a runtime option.

Of course, that's only unless the SHA-1 calculation is made mandatory as suggested above. I really shudder at the idea of requiring SHA-1 to be required forever. We ignored advice in 2005 against making ourselves too dependent on SHA-1, and I would hope that we would learn from this.

Yes, please.

... as long as a shallow clone you do not have, of course...

... or Git would simply handle the absence of the generation number header gracefully, so that sha1-content == sha3-content...

Seeing as current Git versions would always ignore the generation number (and therefore work perfectly even with erroneous baked-in generation numbers), and seeing as it would be easy to add a config option to force Git to ignore the embedded generation numbers, I would consider fsck catching those problems the best idea.

It seems that every major Git hoster already has some sort of fsck on the fly for newly-pushed objects, so that would be another "line of defense".

Taking a step back, though, it may be a good idea to leave the generation number business for later, as much fun as it is to get side tracked and focus on relatively trivial stuff instead of the far more difficult and complex task to get the transition plan to a new hash ironed out.

For example, I am still in favor of SHA-256 over SHA3-256, after learning some background details from in-house cryptographers: it provides essentially the same level of security, according to my sources, while hardware support seems to be coming to SHA-256 a lot sooner than to SHA3-256.

Which hash algorithm to choose is a tough question to answer, and discussing generation numbers will sadly not help us answer it any quicker.

Ciao,
Dscho

FWIW, and I know it is not worth much, as far as I can tell there is at least some security/math basis to prefer SHA3-256 to SHA-256.

The SHA1 and SHA-256 hash functions, (iirc along with their older cousins MD5 and MD2) all have a common design feature where they mix a relatively large block size into a much smaller state *each block*. So for instance SHA-256 mixes a 512 bit block into a 256 bit state with a 2:1 "leverage" between the block being read and the state. In SHA1 this was worse, mixing a 512 bit block into a 160 bit state, closer to 3:1 leverage.

SHA3 however uses a completely different design where it mixes a 1088 bit block into a 1600 bit state, for a leverage of 2:3, and the excess is *preserved between each block*.

Assuming everything else is equal between SHA-256 and SHA3 this difference alone would seem to justify choosing SHA3 over SHA-256. We know that there MUST be collisions when compressing a 512 bit block into a 256 bit space, however one cannot say the same about mixing 1088 bits into a 1600 bit state. The excess state which is not directly modified by the input block makes a big difference when reading the next block.

Of course in both cases we end up compressing the entire source document down to the same number of bits, however SHA3 does that *once*, in finalization only, whereas SHA-256 does it *every* block read. So it seems to me that the opportunity for collisions is *much* higher in SHA-256 than it is in SHA3-256. (Even if they should be vanishingly rare regardless.)

For this reason if I had a vote I would definitely vote SHA3-256, or even for SHA3-512. The latter has an impressive 1:2 leverage between block and state, and much better theoretical security levels.

cheers,
Yves
Note: I am not a cryptographer, although I am probably pretty well informed as far hobby-hash-function-enthusiasts go.

perl -Mre=debug -e "/just|another|perl|hacker/"

Hi,

Thanks for spelling this out. From my (very cursory) understanding of the math, what you are saying makes sense. I think there were some hints of this topic on-list before, but not made so explicit before.

Here's my summary of the discussion of other aspects of the choice of hash functions so far:

My understanding from asking cryptographers matches what Dscho said. One of the lessons of the history of hash functions is that some kinds of attempts to improve the security margin of a hash function do not help as much as expected once a function is broken.

In practice, what we are looking for is

- is the algorithm broken, or likely to be broken soon
- do the algorithm's guarantees match the application
- is the algorithm fast enough
- are high quality implementations widely available

On that first question, every well informed person I have asked has assured me that SHA-256, SHA-512, SHA-512/256, SHA-256x16, SHA3-256, K12, BLAKE2bp-256, etc are equally likely to be broken in the next 10 years. The main difference for the longevity question is that some of those algorithms have had more scrutiny than others, but all have had significant scrutiny. See [1] and the surrounding thread for more discussion on that.

On the second question, SHA-256 is vulnerable to length extension attacks, which means it would not be usable as a MAC directly (instead of using the HMAC construction). Fortunately Git doesn't use its hash function that way.

On the third question, SHA-256 is one of the slower ones, even with hardware accelaration, but it should be fast enough.

On the fourth question, SHA-256 shines. See [2]. That is where I had thought the conversation ended up.

For what it's worth, I'm pretty happy both with the level of scrutiny we've given to this question and SHA-256 as an answer. Luckily even if at the last minute we learn something that changes the choice of hash function, that would not significantly affect the transition plan, so we have a chance to learn more.

See also [3].

Thanks,
Jonathan

[1] https://public-inbox.org/git/CAL9PXLzhPyE+geUdcLmd=pidT5P8eFEBbSgX_dS88knz2q_LSw@mail.gmail.com/#t
[2] https://public-inbox.org/git/[email protected]/
[3] https://www.imperialviolet.org/2017/05/31/skipsha3.html,

    https://news.ycombinator.com/item?id=14453622

Hi Jonathan,

I had read this short after it was published, and had missed the updates. One link in particular caught my eye:

	https://eprint.iacr.org/2012/476

Essentially, the authors demonstrate that using SIMD technology can speed up computation by factor 2 for longer messages (2kB being considered "long" already). It is a little bit unclear to me from a cursory look whether their fast algorithm computes SHA-256, or something similar.

As the author of that paper is also known to have contributed to OpenSSL, I had a quick look and it would appear that a comment in crypto/sha/asm/sha256-mb-x86_64.pl speaking about "lanes" suggests that OpenSSL uses the ideas from the paper, even if b783858654 (x86_64 assembly
pack: add multi-block AES-NI, SHA1 and SHA256., 2013-10-03) does not talk about the paper specifically.

The numbers shown in
https://github.com/openssl/openssl/blob/master/crypto/sha/asm/keccak1600-x86_64.pl#L28
and in
https://github.com/openssl/openssl/blob/master/crypto/sha/asm/sha256-mb-x86_64.pl#L17
are sufficiently satisfying.

Ciao,
Dscho

Hi,

The latter: that paper is about a variant on SHA-256 called SHA-256x4 (or SHA-256x16 to take advantage of newer instructions). It's a different hash function. This is what I was alluding to at [1].

This one is about actual SHA-256, but computing the hash of multiple streams in a single funtion call. The paper to read is [2]. We could probably take advantage of it for e.g. bulk-checkin and index-pack. Most other code paths that compute hashes wouldn't be able to benefit from it.

Thanks,
Jonathan

[1] https://public-inbox.org/git/[email protected]/
[2] https://eprint.iacr.org/2012/371

Hi Jonathan,

Thanks for the explanation!

Again, thanks for the explanation.

Ciao,
Dscho

Yes. And considering that the SHA1 attack was actually predicated on the fact that each block was independent (no extra state between), I do think SHA3 is a better model.

So I'd rather see SHA3-256 than SHA256.

Linus

Hi Linus,

SHA-256 got much more cryptanalysis than SHA3-256, and apart from the length-extension problem that does not affect Git's usage, there are no known weaknesses so far.

It would seem that the experts I talked to were much more concerned about that amount of attention than the particulars of the algorithm. My impression was that the new features of SHA3 were less studied than the well-known features of SHA2, and that the new-ness of SHA3 is not necessarily a good thing.

You will have to deal with the fact that I trust the crypto experts' opinion on this a lot more than your opinion. Sure, you learned from the fact that you had been warned about SHA-1 already seeing theoretical attacks in 2005 and still choosing to hard-wire it into Git. And yet, you are still no more of a cryptography expert than I am.

Ciao,
Dscho

Hi Johannes,

I do not think this is true. Keccak/SHA-3 actually got (and is still getting) a lot of cryptanalysis, with papers published at renowned crypto conferences [1].

Keccak/SHA-3 is recognized to have a significant safety margin. E.g., one can cut the number of rounds in half (as in Keyak or KangarooTwelve) and still get a very strong function. I don't think we could say the same for SHA-256 or SHA-512…

Kind regards,
Gilles, for the Keccak team

[1] https://keccak.team/third_party.html

Hi Gilles,

Please read what I said again: SHA-256 got much more cryptanalysis than SHA3-256.

I never said that SHA3-256 got little cryptanalysis. Personally, I think that SHA3-256 got a ton more cryptanalysis than SHA-1, and that SHA-256 *still* got more cryptanalysis. But my opinion does not count, really. However, the two experts I pestered with questions over questions left me with that strong impression, and their opinion does count.

Again, I do not want to criticize SHA3/Keccak. Personally, I have a lot of respect for Keccak.

I also have a lot of respect for everybody who scrutinized the SHA2 family of algorithms.

I also respect the fact that there are more implementations of SHA-256, and thanks to everybody seeming to demand SHA-256 checksums instead of SHA-1 or MD5 for downloads, bugs in those implementations are probably discovered relatively quickly, and I also cannot ignore the prospect of hardware support for SHA-256.

In any case, having SHA3 as a fallback in case SHA-256 gets broken seems like a very good safety net to me.

Ciao,
Johannes

Hi Johannes,

Thanks for your feedback.

Indeed. What I meant is that SHA3-256 got at least as much cryptanalysis as SHA-256. :-)

OK, I respect your opinion and that of your two experts. Yet, the "much more" part of your statement, in particular, is something that may require a bit more explanations.

Kind regards,
Gilles

Hi Gilles,

Oh? I got the opposite impression... I got the impression that *everybody* in the field banged on all the SHA-2 candidates because everybody was worried that SHA-1 would be utterly broken soon, and I got the impression that after this SHA-2 competition, people were less worried?

Besides, I would expect that the difference in age (at *least* 7 years by my humble arithmetic skills) to make a difference...

I would also like to point out the ubiquitousness of SHA-256. I have been asked to provide SHA-256 checksums for the downloads of Git for Windows, but not SHA3-256...

And this is a practically-relevant thing: the more users of an algorithm there are, the more high-quality implementations you can choose from. And this becomes relevant, say, when you have to switch implementations due to license changes (*cough, cough looking in OpenSSL's direction*). Or when you have to support the biggest Git repository on this planet and have to eek out 5-10% more performance using the latest hardware. All of a sudden, your consideration cannot only be "security of the algorithm" any longer.

Having said that, I am *really* happy to have SHA3-256 as a valid fallback option in case SHA-256 should be broken.

Ciao,
Johannes

Dear Johannes,

if ever there was a SHA-2 competition, it must have been held inside NSA:-) But maybe you are confusing with the SHA-3 competition. In any case, when considering SHA-2 vs SHA-3 for usage in git, you may have a look at arguments we give in the following blogpost:

https://keccak.team/2017/open_source_crypto.html

Kind regards,

Joan Daemen

Hi Joan,

Oops. My bad, I indeed got confused about that, as you suggest below (I actually thought of the AES competition, but that was obviously not about SHA-2). Sorry.

Thanks for the pointer!

Small nit: the post uses "its" in place of "it's", twice.

It does have a good point, of course: the scientific exchange (which you call "open-source" in spirit) makes tons of sense.

As far as Git is concerned, we not only care about the source code of the hash algorithm we use, we need to care even more about what you call "executable": ready-to-use, high quality, well-tested implementations.

We carry source code for SHA-1 as part of Git's source code, which was hand-tuned to be as fast as Linus could get it, which was tricky given that the tuning should be general enough to apply to all common intel CPUs.

This hand-crafted code was blown out of the water by OpenSSL's SHA-1 in our tests here at Microsoft, thanks to the fact that OpenSSL does vectorized SHA-1 computation now.

To me, this illustrates why it is not good enough to have only a reference implementation available at our finger tips. Of course, above-mentioned OpenSSL supports SHA-256 and SHA3-256, too, and at least recent versions vectorize those, too.

Also, ARM processors have become a lot more popular, so we'll want to have high-quality implementations of the hash algorithm also for those processors.

Likewise, in contrast to 2005, nowadays implementations of Git in languages as obscure as Javascript are not only theoretical but do exist in practice (https://github.com/creationix/js-git). I had a *very* quick look for libraries providing crypto in Javascript and immediately found the Standford Javascript Crypto library
(https://github.com/bitwiseshiftleft/sjcl/) which seems to offer SHA-256 but not SHA3-256 computation.

Back to Intel processors: I read some vague hints about extensions accelerating SHA-256 computation on future Intel processors, but not SHA3-256.

It would make sense, of course, that more crypto libraries and more hardware support would be available for SHA-256 than for SHA3-256 given the time since publication: 16 vs 5 years (I am playing it loose here, taking just the year into account, not the exact date, so please treat that merely as a ballpark figure).

So from a practical point of view, I wonder what your take is on, say, hardware support for SHA3-256. Do you think this will become a focus soon?

Also, what is your take on the question whether SHA-256 is good enough? SHA-1 was broken theoretically already 10 years after it was published (which unfortunately did not prevent us from baking it into Git), after all, while SHA-256 is 16 years old and the only known weakness does not apply to Git's usage?

Also, while I have the attention of somebody who knows a heck more about cryptography than Git's top 10 committers combined: how soon do you expect practical SHA-1 attacks that are much worse than what we already have seen? I am concerned that if we do not move fast enough to a new hash algorithm, and somebody finds a way in the meantime to craft arbitrary messages given a prefix and an SHA-1, then we have a huge problem on our hands.

Ciao,
Johannes

Dear Johannes,

thanks for your response and taking the effort to express your concerns. Please see below for some feedback.

Thanks, we'll correct that.

There is a lot of high-quality optimized code for all SHA-3 functions and many CPUs in the Keccak code package https://github.com/gvanas/KeccakCodePackage but also OpenSSL contains some good SHA-3 code and then there are all those related to Ethereum.

By the way, you speak about SHA3-256, but the right choice would be to use SHAKE128. Well, what is exactly the right choice depends on what you want. If you want to have a function in the SHA3 standard (FIPS 202), it is SHAKE128. You can boost performance on high-end CPUs by adopting Parallelhash from NIST SP 800-185, still a NIST standard. You can multiply that performance again by a factor of 2 by adopting KangarooTwelve. This is our (Keccak team) proposal for a parallelizable Keccak-based hash function that has a safety margin comparable to that of the SHA-2 functions. See https://keccak.team/kangarootwelve.html
May I also suggest you read https://keccak.team/2017/is_sha3_slow.html

I think this is a chicken-and-egg problem. In any case, hardware support for one SHA3-256 will also work for the other SHA3 and SHAKE functions as they all use the same underlying primitive: the Keccak-f permutation. This is not the case for SHA2 because SHA224 and SHA256 use a different compression function than SHA384, SHA512, SHA512/224 and SHA512/256.

SHA-256 is more conservative than SHA-1 and I don't expect it to be broken in the coming decades (unless NSA inserted a backdoor but I don't think that is likely). But looking at the existing cryptanalysis, I think it is even less likely that I SHAKE128, ParallelHash or KangarooTwelve will be broken anytime.

This is hard to say. To be honest, when witnessing the first MD5 collisions I did not expect them to lead to some real world attacks and just a few years later we saw real-world forged certificates based on MD5 collisions. And SHA-1 has a lot in common with MD5...

But let me end with a philosophical note. Independent of all the arguments for and against, I think this is ultimately about doing the right thing. The choice is here between SHA1/SHA2 on the one hand and SHA3/Keccak on the other. The former standards are imposed on us by NSA and the latter are the best that came out of an open competition involving all experts in the field worldwide. What would be closest to the philosophy of Git (and by extension Linux or open-source in general)?

Kind regards,

Joan

Begin forwarded message:

From: Gilles Van Assche <[email protected]>
Subject: Re: RFC v3: Another proposed hash function transition plan
Date: 30 Sep 2017 22:20:42 CEST
To: Joan Daemen <[email protected]>, [email protected]

Dag Joan,

About the implementations, there are many high-quality implementations of Keccak besides the KCP that you could also mention. E.g., those in OpenSSL are very good. And there are all those related to Ethereum.

I tend to agree with Guido regarding SHA-1, even if you are right, there is no need to reduce/excuse too much the impact of collisions, there could be unexpected use cases. And it's not clean. (And don't underestimate the probability to be quoted on this.)

Finally, just to say that I like your last paragraph.

Kind regards,
Gilles

Joan Daemen <[email protected]> wrote:
what about replying with something like this (please have a critical look). I sent this from my Radboud account as I have problems with my Thunderbird settings. When trying to send a mail, it sometimes works and sometimes it says “An error occurred while sending mail: Outgoing server (SMTP) error. The server responded: 4.7.1 <joans-mbp.home>: Helo command rejected: Host not found." Dear Johannes,
thanks for your response and taking the effort to express your concerns. Please see below for some feedback.
On 30/09/17 00:33, Johannes Schindelin wrote:
Hi Joan,

On Fri, 29 Sep 2017, Joan Daemen wrote:

if ever there was a SHA-2 competition, it must have been held inside NSA:-) Oops. My bad, I indeed got confused about that, as you suggest below (I actually thought of the AES competition, but that was obviously not about SHA-2). Sorry.

But maybe you are confusing with the SHA-3 competition. In any case, when considering SHA-2 vs SHA-3 for usage in git, you may have a look at arguments we give in the following blogpost:

https://keccak.team/2017/open_source_crypto.html
Thanks for the pointer!

Small nit: the post uses "its" in place of "it's", twice. Thanks, we'll correct that.

It does have a good point, of course: the scientific exchange (which you call "open-source" in spirit) makes tons of sense.

As far as Git is concerned, we not only care about the source code of the hash algorithm we use, we need to care even more about what you call "executable": ready-to-use, high quality, well-tested implementations.

We carry source code for SHA-1 as part of Git's source code, which was hand-tuned to be as fast as Linus could get it, which was tricky given that the tuning should be general enough to apply to all common intel CPUs.

This hand-crafted code was blown out of the water by OpenSSL's SHA-1 in our tests here at Microsoft, thanks to the fact that OpenSSL does vectorized SHA-1 computation now.

To me, this illustrates why it is not good enough to have only a reference
implementation available at our finger tips. Of course, above-mentioned
OpenSSL supports SHA-256 and SHA3-256, too, and at least recent versions
vectorize those, too.
There is a lot of high-quality optimized code for all SHA-3 functions and many CPUs in the Keccak code package https://github.com/gvanas/KeccakCodePackage

By the way, you speak about SHA3-256, but the right choice would be to use SHAKE128. Well, what is exactly the right choice depends on what you want. If you want to have a function in the SHA3 standard (FIPS 202), it is SHAKE128. You can boost performance on high-end CPUs by adopting Parallelhash from NIST SP 800-185, still a NIST standard. You can multiply that performance again by a factor of 2 by adopting KangarooTwelve. This is our (Keccak team) proposal for a parallelizable Keccak-based hash function that has a safety margin comparable to that of the SHA-2 functions. See https://keccak.team/kangarootwelve.html
May I also suggest you to read https://keccak.team/2017/is_sha3_slow.html

Also, ARM processors have become a lot more popular, so we'll want to have high-quality implementations of the hash algorithm also for those processors.

Likewise, in contrast to 2005, nowadays implementations of Git in languages as obscure as Javascript are not only theoretical but do exist in practice (https://github.com/creationix/js-git). I had a *very* quick look for libraries providing crypto in Javascript and immediately found the Standford Javascript Crypto library
(https://github.com/bitwiseshiftleft/sjcl/) which seems to offer SHA-256 but not SHA3-256 computation.

Back to Intel processors: I read some vague hints about extensions accelerating SHA-256 computation on future Intel processors, but not SHA3-256.

It would make sense, of course, that more crypto libraries and more hardware support would be available for SHA-256 than for SHA3-256 given the time since publication: 16 vs 5 years (I am playing it loose here, taking just the year into account, not the exact date, so please treat that merely as a ballpark figure).

So from a practical point of view, I wonder what your take is on, say,
hardware support for SHA3-256. Do you think this will become a focus soon?
I think this is a chicken-and-egg problem. In any case, hardware support for one SHA3-256 will also work for the other SHA3 and SHAKE functions as they all use the same underlying primitive: the Keccak-f permutation. This is not the case for SHA2 because SHA224 and SHA256 use a different compression function than SHA384, SHA512, SHA512/224 and SHA512/256.

Also, what is your take on the question whether SHA-256 is good enough?
SHA-1 was broken theoretically already 10 years after it was published
(which unfortunately did not prevent us from baking it into Git), after
all, while SHA-256 is 16 years old and the only known weakness does not
apply to Git's usage?
I think even the weakness of SHA-1 will be hard to exploit to do something bad in Git. SHA-256 is more conservative than SHA-1 and I don't expect it to be broken (unless NSA inserted a backdoor but I don't think that is likely). But I also don't expect SHAKE128, ParallelHash or KangarooTwelve to be broken, looking at the existing cryptanalysis.
Also, while I have the attention of somebody who knows a heck more about
cryptography than Git's top 10 committers combined: how soon do you expect
practical SHA-1 attacks that are much worse than what we already have
seen? I am concerned that if we do not move fast enough to a new hash
algorithm, and somebody finds a way in the meantime to craft arbitrary
messages given a prefix and an SHA-1, then we have a huge problem on
our hands.
As said, I don't expect practical SHA-1 attacks soon. But let me end with a philosophical note. Independent of all the arguments for and against, I think this is about doing the right thing. The choice is here between SHA1/SHA2 on the one hand and SHA3/Keccak on the other. The former standards are imposed on us by NSA and the latter are the best that came out of an open competition involving all experts worldwide. What would be closest to the philosophy of Git (and by extension Linux or open-source in general)?

Kind regards,

Joan

Hi Joan,

Thanks.

I have to admit that all those names that do not start with SHA and do not end in 256 make me a bit dizzy.

Okay.

So given that Git does not exactly have a big sway on hardware vendors, we would have to hope that some other chicken lays that egg.

That's reassuring! ;-)

Oh, okay. I did not realize that MD5 and SHA-1 are so similar in design, thank you for educating me!

Heh. Do you realize that you are talking to a Microsoftie, i.e. one of the "evil company"? ;-)

So philosophically, I am much more pragmatic. Or maybe I am not, after all, I joined a company at a time when it is arguably going through one of the most dramatic cultural changes any company has seen lately (a year ago, we became #1 contributor on GitHub according to Business Insider, and as far as I can tell, we're not willing to pass that belt to anyone else).

But when it comes to the philosophy of Git, I fear I have to disappoint
you: Git's fundamental concepts were not developed in an open process. Git even so much as rejected professional advice *not* to bake SHA-1 into everything.

Of course, we are undoing this damage right now, and your input helps greatly, I would think.

While I feel reassured by your response that SHA-256 would be "good enough" and would have some real-life benefits of announced hardware support, I would now also feel comfortable if my preference was overruled in the end, in favor of a hash from the Keccak family. I would understand, for example, if the parallel option turned out to be enticing enough for other core Git contributors to aim for, say, K12).

Again, thank you very much for chiming in,
Johannes

Hi,

I just wanted to thank you for paying attention to this conversation and weighing in.

Most of the regulars in the git project are not crypto experts. This kind of extra information (and e.g. [2]) is very useful to us.

Thanks,
Jonathan

[2] https://public-inbox.org/git/[email protected]/

Hi all,

Sorry for late commentary...

Well, for what it's worth, we need to be aware that SHA3 is *different*. In crypto, "different" = "bugs haven't been found yet". :-P

And SHA2 is *known*. So we have a pretty good handle on how it'll weaken over time.

While I think that statement is true on it's face (particularly when including post-competition analysis), I don't think it's sufficient justification to chose one over the other.

The only thing I really object to here is the abstract "experts". We're talking about cryptography and integrity here. It's no longer sufficient to cite anonymous experts. Either they can put their thoughts, opinions and analysis on record here, or it shouldn't be considered. Sorry.

Other than their anonymity, though, I do agree with your experts assessments.

However, whether we chose SHA2 or SHA3 doesn't matter. Moving away from SHA1 does. Once the object_id code is in place to facilitate that transition, the problem is solved from git's perspective.

If SHA3 is chosen as the successor, it's going to get a *lot* more adoption, and thus, a lot more analysis. If cracks start to show, the hard work of making git flexible is already done. We can migrate to SHA4/5/whatever in an orderly fashion with far less effort than the transition away from SHA1.

For my use cases, as a user of git, I have a plan to maintain provable integrity of existing objects stored in git under sha1 while migrating away from sha1. The same plan works for migrating away from SHA2 or SHA3 when the time comes.

thx,

Jason.

Hi Jason,

Here, you seem to agree with me.

And here you don't.

I find that very confusing.

Sorry, you are asking cryptography experts to spend their time on the Git mailing list. I tried to get them to speak out on the Git mailing list. They respectfully declined.

I can't fault them, they have real jobs to do, and none of their managers would be happy for them to educate the Git mailing list on matters of cryptography, not after what happened in 2005.

I know what our in-house cryptography experts have to prove to start working at Microsoft. Forgive me, but you are not a known entity to me.

To you, it does not matter.

To me, it matters. To the several thousand developers working on Windows, probably the largest Git repository in active use, it matters. It matters because the speed difference that has little impact on you has a lot more impact on us.

Uh oh. You forgot the mapping. And the protocol. And pretty much everything except the oid.

Sure. And if XYZ789 is chosen, it's going to get a *lot* more adoption, too.

We think.

Let's be realistic. Git is pretty important to us, but it is not important enough to sway, say, Intel into announcing hardware support for SHA3.

And if you try to force through *any* hash function only so that it gets more adoption and hence more support, in the short run you will make life harder for developers on more obscure platforms, who may not easily get high-quality, high-speed implementations of anything but the very mainstream (which is, let's face it, MD5, SHA-1 and SHA-256). I know I would have cursed you for such a decision back when I had to work on AIX and IRIX.

Please do not make the mistake of taking your use case to be a template for everybody's use case.

Migrating a large team away from any hash function to another one *will* be painful, and costly.

Migrating will be very costly for hosting companies like GitHub, Microsoft and BitBucket, too.

Ciao,
Johannes

From: Jonathan Nieder <[email protected]>

This is "RFC v3: Another proposed hash function transition plan" from the git mailing list.

---

This takes the original Google Doc[1] and adds it to our history, such that the discussion can be on on list and in the commit messages.

* replaced SHA3-256 with NEWHASH, sha3 with newhash
* added section 'Implementation plan'
* added section 'Future work'
* added section 'Agreed-upon criteria for selecting NewHash'

As the discussion restarts again, here is our attempt
to add value to the discussion, we planned to polish it more, but as the discussion is restarting, we might just post it as-is.

Thanks.

[1] https://docs.google.com/document/d/18hYAQCTsDgaFUo-VJGhT0UqyetL2LbAzkWNK1fYS8R0/edit

 Documentation/Makefile                             |   1 +
 .../technical/hash-function-transition.txt         | 571 +++++++++++++++++++++
 2 files changed, 572 insertions(+)
 create mode 100644 Documentation/technical/hash-function-transition.txt

index 2415e0d657..471bb29725 100644
@@ -67,6 +67,7 @@ SP_ARTICLES += howto/maintain-git
 API_DOCS = $(patsubst %.txt,%,$(filter-out technical/api-index-skel.txt technical/api-index.txt, $(wildcard technical/api-*.txt)))
 SP_ARTICLES += $(API_DOCS)
 
+TECH_DOCS += technical/hash-function-transition
 TECH_DOCS += technical/http-protocol
 TECH_DOCS += technical/index-format
 TECH_DOCS += technical/pack-format

2.14.0.rc0.3.g6c2e499285

Hi,

I go by [email protected] upstream. :)

I hadn't signed-off on this version, but it's not a big deal.

[...]

Thanks for sending this out. I had let it stall too long.

As a tiny nit, I think NewHash is easier to read than NEWHASH. Not a big deal. More importantly, we need some text describing it and saying it's a placeholder.

The implementation plan included here is out of date. It comes from an email where I was answering a question about what people can do to make progress, before this design had been agreed on. In the context of this design there are other steps we'd want to describe (having to do with implementing the translation table, etc).

I also planned to add a description of the translation table based on what was discussed previously in this thread.

Jonathan

Hi,

Fortunately we have had a few public comments from crypto specialists:

https://public-inbox.org/git/[email protected]/
https://public-inbox.org/git/CAL9PXLzhPyE+geUdcLmd=pidT5P8eFEBbSgX_dS88knz2q_LSw@mail.gmail.com/
https://public-inbox.org/git/CAL9PXLxMHG1nP5_GQaK_WSJTNKs=_qbaL6V5v2GzVG=9VU2+gA@mail.gmail.com/
https://public-inbox.org/git/[email protected]/
https://public-inbox.org/git/[email protected]/

[...]

Yes, I agree with this. (Adoption by Git could lead to adoption by some other projects, leading to more work on high quality software implementations in projects like OpenSSL, but I am not convinced that that would be a good thing for the world anyway. There are downsides to a proliferation of too many crypto primitives. This is the basic argument described in more detail at [1].)

[...]

That said, I'm curious at what plan you are alluding to. Is it something that could benefit others on the list?

Thanks,
Jonathan

[1] https://www.imperialviolet.org/2017/05/31/skipsha3.html

Hi Jonathan,

Well, it's just a plan at this point. As there's a lot of other work to do in the mean-time, and there's no possibility of transitioning until the dust has settled on NEWHASH. :-)

Given an existing repository that needs to migrate from SHA1 to NEWHASH, and maintain backwards compatibility with clients that haven't migrated yet, how do we

a) perform that migration,
b) allow non-updated clients to use the data prior to the switch, and
c) maintain provable integrity of the old objects as well as the new.

The primary method is counter-hashing, which re-uses the blobs, and creates parallel, deterministic tree, commit, and tag objects using NEWHASH for everything up to flag day. post-flag-day only uses NEWHASH. A PGP "transition" key is used to counter-sign the NEWHASH version of the old signed tags. The transition key is not required to be different than the existing maintainers key.

A critical feature is the ability of entities other than the maintainer to migrate to NEWHASH. For example, let's say that git has fully implemented and tested NEWHASH. linux.git intends to migrate, but it's going to take several months (get all the developers herded up).

In the interim, a security company, relying on Linux for it's products can counter-hash Linus' repo, and continue to do so every time he updates his tree. This shrinks the attack window for an entity (with an undisclosed break of SHA1) down to a few minutes to an hour. Otherwise, a check of the counter hashes in the future would reveal the substitution.

The deterministic feature is critical here because there is valuable integrity and trust built by counter-hashing quickly after publication. So once Linux migrates to NEWHASH, the hashes calculated by the security company should be identical. IOW, use the timestamps that are in the SHA1 commit objects for the NEWHASH objects. Which should be obvious, but it's worth explicitly mentioning that determinism provides great value.

We're in the process of writing this up formally, which will provide a lot more detail and rationale that this quick stream of thought. :-)

I'm sure a lot of this has already been discussed on the list. If so, I apologize for being repetitive. Unfortunately, I'm not able to keep up with the MLs like I used to.

thx,

Jason.

Given the interests that you've expressed here I'd recommend taking a look at
https://public-inbox.org/git/[email protected]/
which is the current version of the transition plan that the community has settled on
(https://public-inbox.org/git/[email protected]/
shows that it should be merged to 'next' soon). Once neat aspect of this transition plan is that it doesn't require a flag day but rather anyone can migrate to the new hash function and still interact with repositories (via the wire) which are still running SHA1.

Brandon Williams

Hi Johannes,

Thanks for the response. Sorry for the delay. Had a large deadline for $dayjob.

Yep.

What I'm saying is that there is more to selecting a hash function for git than just the cryptographic assessment. In fact I would argue that the primary cryptographic concern for git is "What is the likelihood that we'll wake up one day to full collisions with no warning?"

To that, I'd argue that SHA-256's time in the field and SHA3-256's competition give them both passing marks in that regard. fwiw, I'd also put Blake and Skein in there as well.

The chance that any of those will suffer sudden, catastrophic failure is minimal. IOW, we'll have warnings, and time to migrate to the next function.

None of us can predict the future, but having a significant amount of vetting reduces the chances of catastrophic failure.

Ok, fair enough. Just please understand that it's difficult to place much weight on statements that we can't discuss with the person who made them.

Well, I'd say it does not matter for *most* users.

Ahhh, so if I understand you correctly, you'd prefer SHA-256 over SHA3-256 because it's more performant for your usecase? Well, that's a completely different animal that cryptographic suitability.

Have you been able to crunch numbers yet? Will you be able to share some empirical data? I'd love to see some comparisons between SHA1, SHA-256, SHA512-256, and SHA3-256 for different git operations under your work load.

That's quite a jump from what I was saying. I would never advise using code in a production setting just to increase adoption.

What I /was/ saying: Let's say you don't get what you want, and SHA3-256 is chosen. It's not the end of the world from a cryptographic PoV. The hard work of making the git (and libgit2) codebases hash-flexible is already done. So, if you're correct, and SHA3 was too immature, the increased visibility will help us discover that more quickly. And, the code will already be in a position to conduct an orderly migration.

Will it still be costly? Yes. But I would argue that it's naive to think that we will be using git/sha3-256 or git/sha-256 10 to 15 years from now. It might be git, it might not. But there *will* be another migration of existing data (code, history, etc) from one object storage model to another. It might be git/SHA4-512, or hg/sha4-384.

So, we aren't trying to find the perfect hash function so that we naively think we'll never have to change again. Rather, we're choosing the next hash function so that we can hold off another migration for as long as possible. After all, SHA4-512 doesn't exist yet. ;-)

I think you're assuming that all developers on obscure platforms have a similar git usecase to your current one. I've not heard of that being the case.

I wasn't. But I will argue that my usecase is valid. Just as yours is.

Assuming that it will never happen again would make that doubly costly.

<with_my_business_hat_on>
GitHub and BitBucket have git as the core of their business model. If they aren't keeping an eye on the future path of git and maintaining migration plans, shame on them.
</with_my_business_hat_on>

Thanks,

Jason.

In almost all loads I've seen, zlib inflate() cost is a bigger deal than the crypto load. The crypto people talk about cycles per byte, but the deflate code is what usually takes the page faults and cache misses etc, and has bad branch prediction. That ends up easily being tens or thousands of cycles, even for small data.

But it does obviously depend on exactly what you do. The Windows people saw SHA1 as costly mainly due to the index file (which is just a "fancy crc", and not even cryptographically important, and where the cache misses actually happen when doing crypto, not decompressing the data).

And fsck and big initial checkins can have a very different profile than most "regular use" profiles. Again, there the crypto happens first, and takes the cache misses. And the crypto is almost certainly _much_ cheaper than just the act of loading the index file contents in the first place. It may show up on profiles fairly clearly, but that's mostly because crypto is *intensive*, not because crypto takes up most of the cycles.

End result: honestly, the real cost on almost any load is not crypto or necessarily even (de)compression, even if those are the things that show up. It's the cache misses and the "get data into user space" (whether using "read()" or page faulting). Worrying about cycles per byte of compression speed is almost certainly missing the real issue.

The people who benchmark cryptography tend to intentionally avoid the actual real work, because they just want to know the crypto costs. So when you see numbers like "9 cycles per byte" vs "12 cycles per byte" and think that it's a big deal - 30% performance difference! - it's almost certainly complete garbage. It may be 30%, but it is likely 30% out of 10% total, meaning that it's almost in the noise for any but some very special case.

Linus

If anyone is interested in the user-visible effects of slower crypto, I think, there are some numbers in 8325e43b82 (Makefile: add DC_SHA1 knob, 2017-03-16). I don't know how SHA-256 compares to sha1dc exactly, but certainly the latter is a lot slower than normal sha1.

The only real-world case I found with a noticeable slowdown was index-pack. Which in the worst case is roughly the same operation as "git fsck" (inflate and compute the sha1 on every byte), but people tend to actually do it a lot more often.

And it really _is_ slower for real-world operations; the CPU for computing the sha1 of an incoming clone of linux.git jumped from ~3 minutes to ~6 minutes. But I don't think we've seen a lot of complaints, probably because that time is lumped in with "time to transfer a gigabyte of data", so unless you're on a slow machine on fast connection, you don't even really notice.

For day-to-day operations in a repository, I never came up with a good example where the speed difference mattered. I think Dscho's giant-index example is an outlier and the right answer there is not "pick a fast crypto algorithm" but "stop using a slow crypto algorithm as a checksum" (and also, stop routinely reading and writing 400MB for day-to-day operations).

-Peff

Hi Dscho,

Agreed. This is a good reason not to store the sha1 inside the sha256-encoded objects. I think that is exactly what Brandon was saying in response to Junio --- did you read it differently?

[...]

Part of the sha1-content is references to other objects using their sha1-name, so it is not possible to have sha1-content == sha3-content.

That said, I am also leaning against including generation numbers as part of this design.

There is an argument for including generation numbers. It is much simpler to have generation numbers in *all* commit objects than only in some, since it means the slop-based heuristics for faking generation numbers using commit timestamp can be completely avoided for a repository using such a format. Including generation numbers in all commit objects is a painless thing to do during a format change, since it can happen without harming round-tripping.

Treating generation numbers as derived data (as in Jeff King's preferred design, if I have understood his replies correctly) would also be possible but it does not interact well with shallow clone or narrow clone.

All that said, for simplicity I still lean against including generation numbers as part of a hash function transition. Nothing stops us from having another format change later.

This is a particularly hard decision because I don't have a strong preference. That leads me to err on the side of simplicity.

I will make sure to discuss this issue in my patch to Documentation/technical/, so we don't have to repeat the same conversations again and again.

[...]

This is unrelated to Brandon's message, except for his use of SHA3 as a placeholder for "the next hash function".

My assumption based on previous conversations (and other external conversations like [1]) is that we are going to use SHA2-256 and have a pretty strong consensus for that. Don't worry!

As a side note, I am probably misreading, but I found this set of paragraphs a bit condescending. It sounds to me like you are saying "You are making the wrong choice of hash function and everything else you are describing is irrelevant when compared to that monumental mistake. Please stop working on things I don't consider important". With that reading it is quite demotivating to read.

An alternative reading is that you are saying that the transition plan described in this thread is not ironed out. Can you spell that out more? What particular aspect of the transition plan (which is of course orthogonal to the choice of hash function) are you discontent with?

Thanks and hope that helps,
Jonathan

[1] https://www.imperialviolet.org/2017/05/31/skipsha3.html

Just like we have skewed committer timestamps, there is no reason to believe that generation numbers embedded in objects are trustable, and there is no way for narrow clients to even verify their correctness.

So I agree with Peff that having generation numbers in object is pointless; I agree any other derivables like corresponding sha-1 name is also pointless to have.

This is a tangent, but it may be fine for a shallow clone to treat the cut-off points in the history as if they are root commits and compute generation numbers locally, just like everybody else does. As generation numbers won't have to be global (because we will not be embedding them in objects), nobody gets hurt if they do not match across repositories---just like often-mentioned rename detection cache, it can be kept as a mere local performance aid and does not have to participate in the object model.

Good.

Hmph, I actually re-read the thread recently, and my impression was that we didn't quite have a consensus but were leaning towards SHA3-256.

I do not personally have a strong preference myself and I would say that anything will do as long as it is with good longevity and availability. SHA2 family would be a fine choice due to its age on both counts, being scrutinized longer and having a chance to be implemented in many places, even though its age itself may have to be subtracted from the longevity factor.

Thanks.

Locally it helps for some operations such as correct walks. For the network case however, it doesn't really help either.

If we had global generation numbers, one could imagine that they are used in the pack negotiation (server advertises the maximum generation number or even gen number per branch; client could binary search in there for the fork point)

I wonder if locally generated generation numbers (for the shallow case) could be used somehow to still improve network operations.

If we'd get the transition somewhat right, the next transition will be easier than the current transition, such that I am not that concerned about longevity. I am rather concerned about the complexity that is added to the code base (whilst accumulating technical debt instead of clearer abstraction layers)

Hi,

[...]

I have a different concern about locally generated generation numbers in a shallow clone. My concern is that it is slow to recompute them when deepening the shallow clone.

However:

1. That only affects performance and for some use cases could be mitigated e.g. by introducing some laziness, and, more convincingly,

2. With a small protocol change, the server could communicate the generation numbers for commit objects at the edge of a shallow clone, avoiding this trouble.

So I am not too concerned.

More generally, unless there is a very very compelling reason to, I don't want to couple other changes into the hash function transition. If they're worthwhile enough to do, they're worthwhile enough to do whether we're transitioning to a new hash function or not: I have not heard a convincing example yet of a "while at it" that is worth the complexity of such coupling.

(That said, if two format changes are worth doing and happen to be implemented at the same time, then we can save users the trouble of experiencing two format change transitions. That is a kind of coupling from the end user's point of view. But from the perspective of someone writing the code, there is no need to count on that, and it is not likely to happen anyway.)

During the transition, users have to suffer reencoding overhead, so it is not good for such transitions to need to happen very often. If the new hash function breaks early, then we have to cope with it and as you say, having the framework in place means we'd be ready for that. But I still don't want the chosen hash function to break early.

In other words, a long lifetime for the hash absolutely is a design goal. Coping well with an unexpectedly short lifetime for the hash is also a design goal.

If the hash function lasts 10 years then I am happy.

Thanks,
Jonathan

Absolutely. When two functions have similar expected remaining life and are equally widely supported, then faster is better than slower. Otherwise our primary goal when picking the function from candidates should be to optimize for its remaining life and wider availability.

Thanks.

Hi Junio,

SHA-256 has been hammered on a lot more than SHA3-256.

That would be a strong point in favor of SHA2.

Ciao,
Dscho

Last year that was even more true of SHA1 than it is true of SHA-256 today.

Anyway,
Yves

perl -Mre=debug -e "/just|another|perl|hacker/"

Hi,

I hope you are not deliberately trying to annoy me. I say that because you seemed to be interested enough in cryptography to know that the known attacks on SHA-256 *today* are unlikely to extend to Git's use case, whereas the known attacks on SHA-1 *in 2005* were already raising doubts.

So while SHA-1 has been hammered on for longer than SHA-256, the latter came out a lot less scathed than the former.

Besides, you are totally missing the point here that the choice is *not* between SHA-1 and SHA-256, but between SHA-256 and SHA3-256.

After all, we would not consider any hash algorithm with known problems (as far as Git's usage is concerned). The amount of scrutiny with which the algorithm was investigated would only be a deciding factor among the remaining choices, yes?

In any case, don't trust me on cryptography (just like I do not trust you on that matter). Trust the cryptographers. I contacted some of my colleagues who are responsible for crypto, and the two who seem to disagree on pretty much everything agreed on this one thing: that SHA-256 would be a good choice for Git (and one of them suggested that it would be much better than SHA3-256, because SHA-256 saw more cryptanalysis).

Ciao,
Johannes

In the proposed transition plan, the treatment of various signatures (deliberately) makes the conversion not quite roundtrip.

When existing SHA-1 history in individual clones are converted to NewHash, we obviously cannot re-sign the corresponding NewHash contents with the same PGP key, so these converted objects will carry only signature on SHA-1 contents. They can still be validated when they are exported back to SHA-1 world via the fetch/push protocol, and can be validated locally by converting them back to SHA-1 contents and then passing the result to gpgv.

The plan also states, if I remember what I read correctly, that newly created and signed objects (this includes signed commits and signed tags; mergetags merely carry over what the tag object that was merged was signed with, so we do not have to worry about them unless the resulting commit that has mergetag is signed itself, but that is already covered by how we handle signed commits) would be signed both for NewHash contents and its corresponding SHA-1 contents (after internally convering it to SHA-1 contents). That would allow us to strip the signature over NewHash contents and derive the SHA-1 contents to be shown to the outside world while migration is going on and I'd imagine it would be a good practice; it would allow us to sign something that allows everybody to verify, when some participants of the project are not yet NewHash capable.

But the signing over SHA-1 contents has to stop at some point, when everybody's Git becomes completely unaware of SHA-1. We may want to have a guideline in the transition plan to (1) encourage signing for both for quite some time, and (2) the criteria for us to decide when to stop.

Thanks.

That's not precisely true. Details below.

Correct.

Also correct.

The NewHash-based signature is included in the SHA-1 content as well, for the sake of round-tripping. It is not stripped out.

Yes, spelling out a rough schedule is a good idea. I'll add that.

A version of Git that is aware of NewHash should be able to verify NewHash signatures even for users that are using SHA-1 locally for the sake of faster fetches and pushes to SHA-1 based peers.

In addition to a new enough Git, this requires the translation table to translate to NewHash to be present.

So the criterion (2) is largely based on how up-to-date the Git used by users wanting to verify signatures is and whether they are willing to tolerate the performance implications of having a translation table. My hope is that when communicating with peers using the same hash function, the translation table will not add too much performance overhead.

Thank you,
Jonathan

Ah, OK, that allays my worries. We rely on the fact that unknown object headers from the future are ignored. We use something other than "gpgsig" header (say, "gpgsigN") to store NewHash based signature on a commit object created in the NewHash world, so that SHA-1 clients will ignore it but still include in the signature computation---is that the idea?

Existing versions of Git that live in the SHA-1 world may still need to learn to ignore/drop "gpgsigN" while amending a commit that originally was created in the NewHash world. Or to force upgrade we may freeze the SHA-1 only versions of Git and stop updating them altogether. I dunno.

We also need to use something other than "mergetag" when carrying over the contents of a tag being merged in the NewHash world, but I'd imagine that you've thought about this already.

Thanks.

Hi Jonathan,

I am sorry you read it that way. I did not feel condescending when I wrote that mail, I felt annoyed by the side track, and anxious. In my mind, the transition is too important for side tracking, and I worry that we are not fast enough (imagine what would happen if a better attack was discovered that is not as easily detected as the one we know about?).

My impression from reading Junio's mail was that he does not consider the transition plan ironed out yet, and that he wants to spend time on discussing generation numbers right now.

I was in particularly frightened by the suggestion to "reboot" [*1*]. Hopefully I misunderstand and he meant "finishing touches" instead.

As to *my* opinion: after reading https://goo.gl/gh2Mzc (is it really correct that its last update has been on March 6th?), my only concern is really that it still talks about SHA3-256 when I think that the performance benefits of SHA-256 (think: "Git at scale", and also hardware support) really make the latter a better choice.

In order to be "ironed out", I think we need to talk about the implementation detail "Translation table". This is important. It needs to be *fast*.

Speaking of *fast*, I could imagine that it would make sense to store the SHA-1 objects on disk, still, instead of converting them on the fly. I am not sure whether this is something we need to define in the document, though, as it may very well be premature optimization; Maybe mention that we could do this if necessary?

Apart from that, I would *love* to see this document as The Official Plan that I can Show To The Manager so that I can ask to Allocate Time.

Ciao,
Dscho

Footnote *1*:
https://public-inbox.org/git/[email protected]/

Agreed, when that document was written it was hand waved as an implementation detail but once we should probably stare ironing out those details soon so that we have a concrete plan in place.

Speaking of having a concrete plan, we discussed in office the other day about finally converting the doc into a Documentation patch. That was always are intention but after writing up the doc we got busy working on other projects. Getting it in as a patch (with a more concrete road map) is probably the next step we'd need to take.

I do want to echo what jonathan has said in other parts of this thread, that the transition plan itself doesn't depend on which hash function we end up going with in the end. I fully expect that for the transition plan to succeed that we'll have infrastructure for dropping in different hash functions so that we can do some sort of benchmarking before selecting one to use. This would also give us the ability to more easily transition to another hash function when the time comes.

Brandon Williams

Thanks for clarifying. That makes sense.

[...]

Sounds promising!

Thanks much for this feedback. This is very helpful for knowing what v4 of the doc needs.

The discussion of the translation table in [1] didn't make it to the doc. You're right that it needs to.

Caching SHA-1 objects (and the pros and cons involved) makes sense to mention in an "ideas for future work" section.

An implementation plan with well-defined pieces for people to take on and estimates of how much work each involves may be useful for Showing To The Manager. So I'll include a sketch of that for reviewers to poke holes in, too.

Another thing the doc doesn't currently describe is how Git protocol would work. That's worth sketching in a "future work" section as well.

Sorry it has been taking so long to get this out. I think we should have something ready to send on Monday.

Thanks,
Jonathan

[1] https://public-inbox.org/git/CAJo=hJtoX9=AyLHHpUJS7fueV9ciZ_MNpnEPHUz8Whui6g9F0A@mail.gmail.com/

Hi Jonathan,

"Jonathan Nieder" <[email protected]> wrote;

I had a look at the current doc https://goo.gl/gh2Mzc and thought that the selection of the "NewHash" should be separated out into a section of it's own as a 'separation of concerns', so that the general transition plan only refers to the "NewHash", so as not to accidentally pre-judge that selection.

I did look up the arguments regarding sha2 (sha256) versus sha3-256 and found these two Q&A items

https://security.stackexchange.com/questions/152360/should-we-be-using-sha3-2017

https://security.stackexchange.com/questions/86283/how-does-sha3-keccak-shake-compare-to-sha2-should-i-use-non-shake-parameter

with an onward link to this:
https://www.imperialviolet.org/2012/10/21/nist.html

"NIST may not have you in mind (21 Oct 2012)"

"A couple of weeks back, NIST announced that Keccak would be SHA-3. Keccak has somewhat disappointing software performance but is a gift to hardware implementations."

which does appear to cover some of the concerns that dscho had noted, and speed does appear to be a core Git selling point.

It would be worth at least covering these trade offs in the "select a NewHash" section of the document, as at the end of the day it will be a political judgement about what the future might hold regarding the contenders.

What may also be worth noting is the fall back plan should the chosen NewHash be the first to fail, perhaps spectacularly, as having a ready plan could support the choice at risk.

--
Philip

This seems like a rather nontrival thing to design. It will need to hold millions of mappings, and be quickly searchable from either direction (sha1->new and new->sha1) while still be fairly fast to insert new records into.

For Linux, just the list of hashes recording the commits is going to be in the millions, whiel the list of hashes of individual files for all those commits is going to be substantially larger.

David Lang

No, for a few reasons:

1. Most of these objects will not be in the filesystem at all, but rather in a packfile.

2. It's not just a different hash over the same bytes. The sha256-name is taken over the sha256-content (which refers to other objects using sha256). So they really are different objects. You probably wouldn't keep the sha1 version around separately, but rather generate it on the fly during a push to a sha1 server.

3. You really need to be able to take a sha256 name and convert it to a sha1 and vice versa. Hardlinks don't help with that, because they only point in one direction. That get you to the same _content_, but not the other name (and I guess this is where your "look up the name and then compute the other digest comes in, but that's probably too expensive to be workable).

I do think updating the mapping could potentially be deferred until interacting with a sha1 server. But because it needs to be generated in reverse-topological order, it's conceptually easier to do it one object at a time.

-Peff

I am currently thinking of using LevelDB, since it has the advantages of being simple, already existing, and having already been ported to Java (allowing JGit can read and write the same format).

If that doesn't work, we'd try some other key-value store like Samba's tdb or Kyoto Cabinet.

Jonathan

FWIW, I'm using notes-like data to store mercurial->git mappings in git-cinnabar, (ab)using the commit type in tree items. It's fast enough.

Mike

Overall it's an interesting idea. I thought at first that you were suggesting servers do on-the-fly conversion, but after a more careful reading that isn't the case. And I don't think that would work, because the conversion is expensive.

So this pushes the conversion cost onto the clients who decide to move to SHA-256. That may be a problem for sites which have a lot of clients (like CI hosts). But I guess they would just stick with SHA-1 as long as possible, until the upstream repo switches (and that _is_ a per-repo flag day, because the upstream host isn't going to convert back to SHA-1 on the fly to serve the old clients).

I'd encourage anybody following along to follow that link. I almost didn't, but there are a ton of comments there (I'm not sure how I feel about splitting the discussion off the list, though).

I suspect we'll never get away from keeping the mapping table. You'll need at least the sha1->sha256 table if you want to look up names found in historic commit messages, mailing list posts, etc.

And you'll need the sha256->sha1 table if you want to verify the gpg signatures on old tags and commits. That might be something people are willing to drop, though.

I don't think we do this right now, but you can actually find the entry (and exit) points of a pack during the index-pack step. Basically:

1. Keep a hashmap of objects mentioned in the pack.

2. When we process an object's content (i.e., compute its hash), also parse it for any object references. Add entries in the hashmap for any object mentioned this way. Mark the entry for the object we processed with a "HAVE" bit, and mark any referenced object with a "REF" bit.

3. After processing all objects, anything with a "HAVE" but no "REF" is an entry point to the pack (i.e., something that we should have asked for with a want). Anything with a "REF" but not a "HAVE" is an exit point (i.e., an object that we are expected to already have in our repo).

(I've thought about this before because we could possibly shortcut the connectivity check using the exit points. It's complicated by the fact that we don't assume the transitive presence of objects unless they are reachable).

I don't think using the "want"s as the entry points is unreasonable, though. The server _shouldn't_ generally be sending us other cruft.

I do wonder if you might be able to omit the extra object-graph walk from your step 2, if you could assign "depths" to each object during step 1 instead of HAVE/REF bits. The trouble, of course, is that you're not visiting the nodes in the right order (so given two trees, you're not sure if one might eventually be a child of the other; how do you assign their depths?). I have a feeling there's a proof that it's impossible, but I might just not be clever enough.

Overall the basics of the conversion seem sound to me. The "nohash" things seems more complicated than I think it ought to be, which probably just means I'm missing something. I left a few related comments on the google doc, so I won't repeat them here.

-Peff

I am sure how I feel about it---we should really discourage it, unless it is an effort to help polishing an early draft for wider distribution and discussion.

We have code to do the "entry point" computation in index-pack already, I think, in 81a04b01 ("index-pack: --clone-bundle option", 2016-03-03).

That's true.

I think "nohash" can be explained in 2 points:
1. When creating signed objects, "nohash" is almost never written. Just create the object as usual and add "hash" lines for every other hash function that you want the signature to cover.
2. When converting from function A to function B, add "nohash B" if there were no "hash B" lines in the original object.

The "nohash" thing was in the hope of requiring only one signature to sign all the hashes (in all the functions) that the user wants, while preserving round-tripping ability.

Maybe some examples would help to address the apparent complexity. These examples are the same as those in the document. I'll also show future compatibility with a hypothetical NEW hash function, and extend the rule about signing/verification to 'sign in the earliest supported hash function in ({object's hash function} + {functions in "hash" lines} - {function in "nohash" line})'.

Example 1 (existing signed commit)

<sha-1 object stuff>  <sha256 object stuff>  <NEW object stuff>
                      nohash sha256          nohash new
                      hash sha1 ...          hash sha1 ...

This object was probably created in a SHA-1 repository with no knowledge that we were going to transition to SHA256 (but there is nothing preventing us from creating the middle or right object and then translating it to the other functions).

Example 2 (recommended way to sign a commit in a SHA256 repo)

<sha-1 object stuff>  <sha256 object stuff>  <NEW object stuff>
hash sha256 ...       hash sha1 ...          nohash new
                                             hash sha1 ...
                                             hash sha256 ...

This is the recommended way to create a SHA256 object in a SHA256 repo. The rule about signing/verification (as stated above) is to sign in SHA-1, so when signing or verifying, we convert the object to SHA-1 and use that as the payload. Note that the signature covers both the SHA-1 and SHA256 hashes, and that existing Git implementations can verify the signature.

Example 3 (a signer that does not care about SHA-1 anymore)

<sha-1 object stuff>  <sha256 object stuff>  <NEW object stuff>
nohash sha1                                  nohash new
hash sha256 ...                              hash sha256 ...

If we were to create a SHA256 object without any mentions of SHA-1, the rule about signing/verification (as stated above) states that the signature payload is the SHA256 object. This means that existing Git implementations cannot verify the signature, but we can still round-trip to SHA-1 and back without losing any information (as far as I can tell).

I do think that that was my least favorite part of the suggestion. Not just "nohash", but all the special "hash" lines too.

I would honestly hope that the design should not be about "other hashes". If you plan your expectations around the new hash being broken, something is wrong to begin with.

I do wonder if things wouldn't be simpler if the new format just included the SHA1 object name in the new object. Put it in the "header" line of the object, so that every time you look up an object, you just _see_ the SHA1 of that object. You can even think of it as an additional protection.

Btw, the multi-collision attack referenced earlier does _not_ work for an iterated hash that has a bigger internal state than the final hash. Which is actually a real argument against sha-256: the internal state of sha-256 is 256 bits, so if an attack can find collisions due to some weakness, you really can then generate exponential collisions by chaining a linear collision search together.

But for sha3-256 or blake2, the internal hash state is larger than the final hash, so now you need to generate collisions not in the 256 bits, but in the much larger search space of the internal hash space if you want to generate those exponential collisions.

So *if* the new object format uses a git header line like

    "blob <size> <sha1>\0"

then it would inherently contain that mapping from 256-bit hash to the SHA1, but it would actually also protect against attacks on the new hash. In fact, in particular for objects with internal format that differs between the two hashing models (ie trees and commits which to some degree are higher-value targets), it would make attacks really quite complicated, I suspect.

And you wouldn't need those "hash" or "nohash" things at all. The old SHA1 would simply always be there, and cheap to look up (ie you wouldn't have to unpack the whole object).

Hmm?

Linus

I'll agree that the "hash" "nohash" bit isn't my favorite and is really only there to address the signing of tags/commits in this new non-sha1 world. I'm inclined to take a closer look at Jeff's suggestion which simply has a signature for the hash that the signer cares about.

I don't know if keeping around the SHA1 for every object buys you all that much. It would add an additional layer of protection but you would also need to compute the SHA1 for each object indefinitely (assuming you include the SHA1 in new objects and not just converted objects). The hope would be that at some point you could not worry about SHA1 at all. That may be difficult for projects with long history with commit msgs which reference SHA1's of other commits (if you wanted to look up the referenced commit, for example), but projects started in the new non-sha1 world shouldn't have to ever compute a sha1.

Also, during this transition phase you would still need to maintain the sha1<->sha256 translation table to make looking up objects by their sha1 name in a sha256 repo fast. Otherwise I think it would take a non-trivial amount of time to search a sha256 repo for a sha1 name. So if you do include the sha1 in the new object format then you would end up with some duplicate information, which isn't the end of the world.

Brandon Williams

This is easy for blobs as you only need to hash twice. I am not sure if you can do the same for trees, though. For that <sha1> to be useful, the hash needs to be over the tree contents whose references are expressed in <sha1>, which in turn would mean...

... ah, you would read these <sha1> off of the object header in the new world and you do not need to expand the whole thing. OK, I see how it could work.

Thanks, this explained it very well.

I understand the tradeoff now, though I am still of the opinion that simplicity is probably a more important goal.

In practice I'd imagine that anybody doing commit-signing would just sign the more-secure hash, and people doing tag releases would probably do a dual-sign to be verifiable by both old and new clients. Those are infrequent enough that the extra computation probably doesn't matter. But that's just my gut feeling.

-Peff

Jonathan Nieder writes ("RFC: Another proposed hash function transition plan"):

Thanks for this.

This is a reasonable plan. It corresponds to approaches (2) and (B) of my survey mail from the other day. Ie, two parallel homogeneous hash trees, rather than a unified but heterogeneous hash tree, with old vs new object names distinguished by length.

I still prefer my proposal with the mixed hash tree, mostly because the handling of signatures here is very awkward, and because my proposal does not involve altering object ids stored other than in the git object graph (eg CI system databases, etc.)

One thing you've missed, I think, is notes: notes have to be dealt with in a more complicated way. Do you intend to rewrite the tree objects for notes commits so that the notes are annotations for the new names for the annotated objects ? And if so, when ?

Also I think you need to specify how abbreviated object names are interpreted.

Regards,
Ian.

One option might be to not use hex for the new hash, but base64 encoding.

That would make the full size ASCII hash encoding length roughly similar (43 base64 characters rather than 40), which would offset some of the new costs (longer filenames in the loose format, for example).

Also, since 256 isn't evenly divisible by 6, and because you'd want some way to explictly disambiguate the new hashes, the rule *could* be that the ASCII representation of a new hash is the base64 encoding of the 258-bit value that has "10" prepended to it as padding.

That way the first character of the hash would be guaranteed to not be a hex digit, because it would be in the range [g-v] (indexes 32..47).

Of course, the downside is that base64 encoded hashes can also end up looking very much like real words, and now case would matter too.

The "use base64 with a "10" two-bit padding prepended" also means that the natural loose format radix format would remain the first 2 characters of the hash, but due to the first character containing the padding, it would be a fan-out of 2**10 rather than 2**12.

Of course, having written that, I now realize how it would cause problems for the usual shit-for-brains case-insensitive filesystems. So I guess base64 encoding doesn't work well for that reason.

Linus

Linus Torvalds writes ("Re: RFC: Another proposed hash function transition plan"):

We should arrange for this to be an uppercase, not a lowercase, letter, for the reasons I explained in my own proposal. To summarise:
It would be undesirable to further increase the overlap between object names and ref names. Few people use uppercase in ref names because of the case-insensitive filesystem problem; so object names starting with uppercase ascii are distinct from most object names.

AFAIAA object names occur in publicly-visible filenames only in notes tree objects, which are manipulated by git internally and do not necessarily need to appear in the filesystem.

The filenames in .git/objects/ can be in whatever encoding we like, so are not an obstacle.

Ian.

Ian Jackson <[email protected]>   These opinions are my own.

If I emailed you from an address @fyvzl.net or @evade.org.uk, that is
a private address which bypasses my fierce spamfilter.

Hi Ian,

Not true.

Ciao,
Johannes

Hi Ian,

Given that the idea was to encode the new hash in base64 or base85, we *are* talking about an encoding. In that respect, yes, it can be whatever encoding we like, and Linus just made a good point (with unnecessary foul language) of explaining why base64/base85 is not that encoding.

Ciao,
Johannes

Since the hash format is switching anyway, how about using base32 instead of hex?

Still get a 20% space savings over hex (minus a little for padding), and it's guaranteed to be a single case.
Jason