patchdocs: clarify that credential discards unrecognised attributes

From: M Hickford <mirth.hickford@gmail.com>

It was previously unclear how unrecognised attributes are handled.

Signed-off-by: M Hickford <mirth.hickford@gmail.com>
---
    docs: clarify that credential discards unrecognised attributes
    
    It was previously unclear how unrecognised attributes are handled.
    
    Signed-off-by: M Hickford mirth.hickford@gmail.com

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1393%2Fhickford%2Funrecognised-attributes-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1393/hickford/unrecognised-attributes-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/1393

 Documentation/git-credential.txt | 2 ++
 Documentation/gitcredentials.txt | 1 +
 2 files changed, 3 insertions(+)

On Mon, Oct 24, 2022 at 07:57:48AM +0000, M Hickford via GitGitGadget wrote:

Yeah, this was always part of the intended behavior, but I agree we did not say it very explicitly (aside from an in-code comment!). Both the intent and content of your patch look good to me.

We did discuss patches a long time ago that would let Git carry arbitrary keys between helpers, even if Git itself didn't understand it. One of the intended uses was to let helpers talk to each other about TTLs. So if you had say:

  [credential]
  helper = generate-some-token
  helper = cache

where the first helper generates a token, and the second caches it, the first one could shove a "ttl" or "expiration" key into the protocol, which the cache could then learn to respect.

But we never merged such a thing, and in practice I think people would just implement both parts as a single helper for simplicity. And anyway, even if we did want to do that, such a patch would want to modify the documentation to explain the new behavior. :)

-Peff

On Mon, Oct 24, 2022 at 07:59:22PM -0400, Jeff King wrote:

In case anyone is morbidly curious, it was in this thread:

  https://lore.kernel.org/git/20120407033417.GA13914@sigill.intra.peff.net/

-Peff

Thanks Jeff for your reply. This is helpful to understand the background.

(first message using git send-email, hopefully I followed the instructions correctly)

On 10/25/22 08:51, M Hickford wrote:

You messed up the thread (you broke it).

On every message showed on lore.kernel.org/git, there is reply instructions at the bottom of the message's page. Follow it closely.

If you insisted on either using git-send-email(1) or mailto: link to reply, you need to manually quote any portion of original message you wish to give the context. Make sure every line of quoted text starts with `> `.

Better yet, reply using "reply-all" button or key in your mail client. However, make sure your email is text/plain (not HTML) and your mail client doesn't mess up with whitespaces (convert tabs to spaces, word wrapping).

Thanks.

Curious. The thread overview "5+ messages" at https://lore.kernel.org/git/20221025015116.4730-1-mirth.hickford@gmail.com/#related looks okay to me. I followed the git send-email instructions setting the In-Reply-To header. I changed the subject -- maybe that confused some clients? Thanks for the tips.

On Wed, Oct 26, 2022 at 05:39:05AM +0100, M Hickford wrote:

Your reply looked fine to me. The in-reply-to you used is correct. It's OK to change the subject (though usually only done when the conversation drifts to a different topic).

-Peff

M Hickford <mirth.hickford@gmail.com> writes:

Your In-Reply-To: header looked OK to me. Some webmail platforms break the thread when Subject: is edited, but that is their bug. As long as you keep the In-Reply-To: chain correct, retitling the message as necessary is indeed encouraged practice.

On Tue, 25 Oct 2022 at 00:59, Jeff King <peff@peff.net> wrote:

Thanks. What happens next? I should look for this change in the seen branch? https://git-scm.com/docs/MyFirstContribution#after-approval

Composing helpers like this is how I encourage users to configure git-credential-oauth [1][2]. Note that the storage helper should come *before* the generator, so that `credential fill` finds a stored credential before it generates a fresh credential.

Composing helpers has the advantage that the user can choose their preferred storage. Generated credentials aren't necessarily short lived. GitHub OAuth tokens, for example, are good for at least one year [3].

[1] https://lore.kernel.org/git/CAGJzqs=+fCQzkDX53H8Mz-DjXicVVgRmmzPjkatSiOpYO7wGGA@mail.gmail.com/ [2] https://github.com/hickford/git-credential-oauth [3] https://docs.github.com/en/authentication/keeping-your-account-and-data-secure/token-expiration-and-revocation#token-expired-due-to-lack-of-use

On Sat, Nov 12, 2022 at 02:21:24AM +0000, M Hickford wrote:

Usually the maintainer would pick it up, it would end up in seen, then eventually 'next', and then eventually 'master'. You can check the periodic "What's Cooking" messages from the maintainer to see more discussion of various topic branches.

In this case, though, I don't see any indication that the maintainer picked saw it. It sometimes happens that a topic is simply overlooked, even if it received positive reviews.

The usual thing to do is repost it, cc-ing the maintainer. I've also cc'd the interim maintainer here, so that may get things moving. :)

Right, it's been a while since I've constructed an example like this. ;)

What you're doing works fine with the code as-is; you just can't carry extra data (like a ttl) between the two.

The thread I linked earlier also discusses (in the very top-level patch) a change in behavior that would break the flow you're relying on here (because it may unexpectedly propagate credentials between helpers). But I don't think anybody is interested in pursuing that, and it has been 10 years now.

Yeah, the composability was one of the goals of the system. I just think in practice that not many people use it. You can also compose outside of Git (I think the thread I linked earlier has an example of a wrapper that does so), but again, I don't think anybody really does so in practice.

I agree for GitHub's tokens that the times involved make auto-expiration not that important. The example back in that thread was something more time-limited (like minutes or hours). I don't know how often that kind of things is in the wild.

-Peff

On Sat, 12 Nov 2022 at 16:47, Jeff King <peff@peff.net> wrote:

FWIW I have a draft patch that adds password_expiry_utc and oauth_refresh_token attributes to credential https://github.com/gitgitgadget/git/pull/1394 introducing expiry logic in the credential layer. I'll share a RFC sometime in future.

GitLab OAuth tokens expire after 2 hours (the refresh tokens are valid longer). This is a security improvement over long-lived tokens.

On Sat, Nov 12, 2022 at 07:08:42PM +0000, M Hickford wrote:

Neat.

I'm not _totally_ opposed to introducing these as something Git understands, but I think it makes more sense to just teach Git to relay unknown entries between helpers.

The oauth thing is going to be very helper specific, and not something I think Git would ever do anything with itself.

In theory Git might care about expiration, but in practice I think it doesn't. It's very unlikely for a token to expire in the course of Git using it. It's only much later, when we ask for it back, that a helper will notice it's expired. Git could save the helper some work by noticing this on read, but since the helper has to learn to store and report the expiration in the first place, not much is gained.

And in the case of something like credential-cache, we want to do more than just store; we'd actually drop the credential entirely (and maybe even cause the daemon to exit) if it expires before the usual timeout.

-Peff

On Sat, Nov 12, 2022 at 11:47:30AM -0500, Jeff King wrote:

I think that Junio may have missed it when picking up topics. But it happened before I start combing the list more finely, so I had already purged it from my inbox without picking it up, either.

I'll take a look now and queue it to start merging down. Thanks for the nudge.

Thanks, Taylor