Re: [PATCH v2] add test to demonstrate that shallow recursive clones fail

> On Mon, Nov 16, 2015 at 10:59 AM, Jens Lehmann <Jens.Lehmann@web.de> wrote:
>> Am 14.11.2015 um 01:10 schrieb Stefan Beller:
>>> Thanks for pointing out that we already have some kind of server support.
>>>
>>> I wonder if we should add an additional way to make fetching only some
>>> sha1s possible. ("I don't want users to fetch any sha1, but only those
>>> where superprojects point{ed} to", even if you force push a superproject,
>>> you want to want to only allow fetching all sha1s which exist in the
>>> current
>>> superprojects branch.)
>>
>>
>> Me thinks the restrictions for sha1-fetching could come from the branches
>> these sha1s are found in the upstream submodule: if the client is allowed
>> to fetch a branch, it should be able to fetch any sha1 on that branch.
>
> I'd agree on that. The server side even with uploadpack.allowTipSHA1InWant
> set, is not sufficient though.
>
> To fetch an arbitrary sha1, you would need to check if that sha1 is part
> of the history of any advertised branch and then allow fetching serverside,
> which sounds like some work for the server, which we may want to avoid
> by having smarter data structures there.
>
> Instead of having to search all branches for the requested sha1, we could have
> some sort of data structure to make it not an O(n) operation (n being
> all objects
> in the repo).
>
> Maybe I overestimate the work which needs to be done, because the server has
> bitmaps nowadays.
>
> Maybe a lazy reverse-pointer graph can be established on the serverside.
> So I guess when we add the feature to fetch arbitrary sha1s, reachable from
> any branch, people using submodules will make use of the feature. (such as with
> git fetch --recurse --depth 1 or via a new `git fetch --recursive
> --up-to-submodule-tip-only`)
>
> So once the server is asked for a certain sha1, it will do the
> reachability check,
> which takes some effort, but then stores the result in the form:
> "If ${current tip sha} of ${branch} is reachable, so is requested $sha1."
>
> So when the next fetch request for $sha1 arrives, the server only needs to
> check for ${current tip sha} to be part of $branch, which is expected to be
> a shorter revwalk from the tip. (Because it is nearer to the tip, a bitmap could
> just tell you or at least shorten the walk even more)
> If the ${branch} has changed, the next evaluation for $sha1 can update
> the cache,
> such that the reverse lookup is not expensive on expectation.

>>> Maybe our emails crossed, but in the other mail I pointed out we could use
>>> some sort of hidden ref (refs/superprojects/*) for that, which are
>>> allowed to mark
>>> any sort of sha1, which are allowed in the superproject/submodule context
>>> to be fetched.
>>>
>>> So whenever you push to a superproject (a project that has a gitlink),
>>> we would need to check serverside if that submodule is at us and mark the
>>> correct sha1s in the submodule. Then you can disallow fetching most of the
>>> sha1s
>>> but still could have a correctly working submodule update mechanism.
>>
>>
>> And what happens if the submodule isn't at us? Involving the serverside of
>> a superproject in submodule fetching sounds wrong to me. Me thinks that
>> the upstream of the submodule should always control if a sha1 is allowed
>> to be fetched. Or did I understand you wrong?
>
> Yes and no.
> The serverside submodule repository should be responsible for the ultimate
> decision if you are allowed to fetch that sha1. But maybe on pushing the
> superproject, we can store a hint in the submodule, that this sha1 is legit.
> Although I may be missguided in my thinking here as the superproject
> should have no influence on the submodule.