> > @@ -230,7 +246,7 @@ static int find_common(struct fetch_negotiator
> > *negotiator,
> > if (args->stateless_rpc && multi_ack == 1)
> > die(_("--stateless-rpc requires multi_ack_detailed"));
> >
> > - for_each_ref(rev_list_insert_ref_oid, negotiator);
> > + mark_tips(negotiator, args->negotiation_tips);
> > for_each_cached_alternate(negotiator, insert_one_alternate_object);
> >
> > fetching = 0;
>
> Here we blindly add objects found in an alternate repo. I found and
> debugged this with this:
>
> diff --git a/fetch-negotiator.h b/fetch-negotiator.h
> index 9e3967ce66..cbe71c9c8d 100644
> --- a/fetch-negotiator.h
> +++ b/fetch-negotiator.h
> @@ -33,2 +33,3 @@ struct fetch_negotiator {
> void (*add_tip)(struct fetch_negotiator *, struct commit *);
> + int done_adding;
>
> diff --git a/fetch-pack.c b/fetch-pack.c
> index 3f24d0c8a6..6b43b4f8f1 100644
> --- a/fetch-pack.c
> +++ b/fetch-pack.c
> @@ -238,2 +238,3 @@ static void mark_tips(struct fetch_negotiator
> *negotiator,
> &negotiation_tips->oid[i]);
> + negotiator->done_adding = 1;
> return;
> diff --git a/negotiator/default.c b/negotiator/default.c
> index 4b78f6bf36..4e45f05f25 100644
> --- a/negotiator/default.c
> +++ b/negotiator/default.c
> @@ -137,2 +137,4 @@ static void add_tip(struct fetch_negotiator *n,
> struct commit *c)
> {
> + if (n->done_adding)
> + return;
> n->known_common = NULL;
> @@ -166,2 +168,3 @@ void default_negotiator_init(struct fetch_negotiator
> *negotiator)
> negotiator->add_tip = add_tip;
> + negotiator->done_adding = 0;
> negotiator->next = next;
>
> Perhaps something like that with an assert() is a good idea for the
> negotiation backend code in general? It seems rather fragile to depend
> on there being no other codepath that calls add_tip() again after some
> other code (--negotiation-tip=*) that expects it not to be called again.
Thanks for spotting this bug.
There is already some defense from add_tip() not being called
unexpectedly - see negotiator/default.c and negotiator.skipping.c, which
sets add_tip to NULL when next() is called.
I can see that this doesn't help in this case, when we want to declare
done_adding but we haven't called next() yet, but I don't think that
this API layer is the right place to prevent that.
