On Thu, Sep 27 2018, Jonathan Tan wrote:
>> > If you wanted to do this, it seems better to me to just declare a "null"
>> > negotiation algorithm that does not perform any negotiation at all.
>>
>> I think such an algorithm is a good idea in general, especially for
>> testing, and yeah, maybe that's the best way out of this, i.e. to do:
>>
>> if git rev-parse {}/HEAD 2>/dev/null
>> then
>> git fetch --negotiation-tip={}/HEAD {}
>> else
>> git -c fetch.negotiationAlgorithm=null fetch {}
>> fi
>>
>> Would such an algorithm be added by overriding default.c's add_tip
>> function to never add anything by calling default_negotiator_init()
>> followed by null_negotiator_init(), which would only override add_tip?
>> (yay C OO)
>>
>> If so from fetch-pack.c it looks like there may be the limitation on the
>> interface that the negotiator can't exit early (in
>> fetch-pack.c:mark_tips). But I've just skimmed this, so maybe I've
>> missed something.
>
> (I was reminded to reply to this offlist - sorry for the late reply.)
>
> I think too many things need to be replaced (known_common, add_tip, and
> ack all need to do nothing), so it's best to start from scratch. That
> way, we also don't need to deal with the subtleties of C OO :-)
>
>> Also, looks like because of the current interface =null and
>> --negotiation-tip=* would (somewhat confusingly) do a "real" negotiation
>> if done that way, since it'll bypass the API and insert tips for it to
>> negotiate, but it looks like overriding next() will get around that.
>
> If you do it as I suggest (in particular, add_tip doing nothing) then
> there is the opposite problem that it won't be easy to inform the user
> that --negotiation-tip does nothing in this case. Maybe there needs to
> be an "accepts_tips" field in struct fetch_negotiator that, if false,
> means that custom tips (or any tips) are not accepted, allowing the
> caller of the negotiator to print a warning message in this case.
Thanks, yeah it seems the interface would need to be tweaked for such a
"null" negotiator.
Some more general questions (which I can turn into docs once I
understand this). If I run this, as a testcase for two random repos
where I "fetch" an unrelated one and use the first ever commit to
git.git as an alias for this "null" negotiatior, i.e. "just present this
one commit":
(
rm -rf /tmp/git &&
git clone https://github.com/git/git.git /tmp/git &&
cd /tmp/git &&
git remote add gitlab-shell https://github.com/cr-marcstevens/sha1collisiondetection &&
GIT_TRACE_PACKET=/tmp/git/packet.trace git fetch --negotiation-tip=$(git log --reverse|head -n 1|cut -d ' ' -f2) gitlab-shell &&
grep -c "fetch-pack> have" /tmp/git/packet.trace
)
I get:
warning: Ignoring --negotiation-tip because the protocol does not support it.
And the grep -c shows we tried to present 55170 commits in "have" lines
to the server. Now, change that to SSH and all is well:
(
rm -rf /tmp/git &&
git clone git@xxxxxxxxxx:git/git.git /tmp/git &&
cd /tmp/git &&
git remote add gitlab-shell git@xxxxxxxxxx:cr-marcstevens/sha1collisiondetection &&
GIT_TRACE_PACKET=/tmp/git/packet.trace git fetch --negotiation-tip=$(git log --reverse|head -n 1|cut -d ' ' -f2) gitlab-shell &&
grep -c "fetch-pack> have" /tmp/git/packet.trace
)
I don't understand this limitation. With the SSH version we skip
straight to saying we "want" with just the 1 "have" line of
"e83c5163316f89bfbde7d9ab23ca2e25604af290".
Why aren't we doing the same over http? I don't get how protocol support
is needed, it's us who decide to send over the "have" lines. Some
variant of this does work over "skipping":
(
rm -rf /tmp/git &&
git clone https://github.com/git/git.git /tmp/git &&
cd /tmp/git &&
git remote add gitlab-shell https://github.com/cr-marcstevens/sha1collisiondetection &&
GIT_TRACE_PACKET=/tmp/git/packet.trace git -c fetch.negotiationAlgorithm=skipping fetch gitlab-shell &&
grep -c "fetch-pack> have" /tmp/git/packet.trace
)
There we send 14002 "have" lines, which seems expected, but then with
the same thing over SSH we don't send any:
(
rm -rf /tmp/git &&
git clone git@xxxxxxxxxx:git/git.git /tmp/git &&
cd /tmp/git &&
git remote add gitlab-shell git@xxxxxxxxxx:cr-marcstevens/sha1collisiondetection &&
GIT_TRACE_PACKET=/tmp/git/packet.trace git -c fetch.negotiationAlgorithm=skipping fetch gitlab-shell &&
grep -c "fetch-pack> have" /tmp/git/packet.trace
)
So that seems like another bug, and as an aside, a "skipping"
implementation that sends ~1/4 of the commits in the repo seems way less
aggressive than it should be. I was expecting something that would
gradually "ramp up" from the tips. Where say starting at master/next/pu
we present every 100th commit as a "have" until the 1000th commit, then
every 1000 commits until 10k and quickly after that step up the size
rapidly.
