Re: git credential cache timeout questions

From: Jeff King <peff@peff.net>
Date: Feb 19, 2021, 17:04 UTC
Message-ID: <YC/vqJBSnDHoLIdJ@coredump.intra.peff.net>
In-Reply-To: <CAP8UukiL0niGSm3o7uYNBzL3FP-UEgfOuq-Tu=fksWJkerT5fg@mail.gmail.com>

On Fri, Feb 19, 2021 at 10:46:48AM -0500, John Ratliff wrote:

Show 14 quoted lines

> I have configured my git to cache my credentials for 12 hours using
> this section in my .gitconfig
> 
> [credential "https://mygithub.example.edu"]
>     username = myuser
>     helper = cache --timeout 43200
> 
> However, the credentials don’t always seem to expire after 12 hours.
> Sometimes I come back the next morning and the credentials still work.
> This is generally after leaving at 5:00 PM and coming back in the next
> day at 9:00 AM, well past the 12 hour timeout.
> 
> Is there any way to see the current timeout value? Is it a rolling
> timeout (i.e. any git action resets the timeout)?

It's the "rolling" thing, though the source is a bit subtle. The credential-cache helper sets an absolute expiration when the value is stored, and it doesn't update it on a "get" request.

However, Git's interaction with the helpers is generally:

  - when we need a credential ask for one

  - when a credential is rejected by a server, tell helpers to erase it

  - when a credential is accepted by a server, tell helpers to store it

And it's that last one that provides the rolling timeout, because we do it even if the credential came from a helper in the first place!

I actually wrote a patch long ago to switch this behavior:

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

But it turned out some people actually rely on it. :)

There's some discussion in that thread about paths forward, and I think I even played around with it back then. But then it sat on my todo list, and now it has been 9 years, so I don't remember if there were good reasons not to push it forward, or if I simply never got around to it (I suspect the latter; nobody had a pressing use case that was solved by avoiding the rolling timeouts, it just seemed to me to be a bit less surprising). I'd be happy if somebody wanted to revisit the topic.

(To your other question, "is there a way to see the timeout value", the answer is "not really, without running it under gdb". I wouldn't be opposed to adding more diagnostic output to the daemon. But you can also see some of what's going on by setting GIT_TRACE=1 in the environment, which will show the extra "store" operation being done by Git).

-Peff