Re: [PATCH v3 5/7] fetch: add --negotiation-include option for negotiation

> On 2026-04-22 16:25, Derrick Stolee via GitGitGadget wrote:
>> From: Derrick Stolee <stolee@gmail.com>
>>
>> Add a new --negotiation-include option to 'git fetch', which ensures
>> that certain ref tips are always sent as 'have' lines during fetch
>> negotiation, regardless of what the negotiation algorithm selects.
>>
>> This is useful when the repository has a large number of references, so
>> the normal negotiation algorithm truncates the list. This is especially
>> important in repositories with long parallel commit histories. For
>> example, a repo could have a 'dev' branch for development and a
>> 'release' branch for released versions. If the 'dev' branch isn't
>> selected for negotiation, then it's not a big deal because there are
>> many in-progress development branches with a shared history. However, if
>> 'release' is not selected for negotiation, then the server may think
>> that this is the first time the client has asked for that reference,
>> causing a full download of its parallel commit history (and any extra
>> data that may be unique to that branch). This is based on a real example
>> where certain fetches would grow to 60+ GB when a release branch
>> updated.
>>
>> This option is a complement to --negotiation-restrict, which reduces the
>> negotiation ref set to a specific list. In the earlier example, using
>> --negotiation-restrict to focus the negotiation to 'dev' and 'release'
>> would avoid those problematic downloads, but would still not allow
>> advertising potentially-relevant user brances. In this way, the
>> 'include' version solves the problem I mention while allowing
>> negotiation to pick other references opportunistically. The two options
>> can also be combined to allow the best of both worlds.
> 
> Nice explanation and motivation for the need of such as feature.
> 
> One small typo: s/brances/branches/

>> --- a/Documentation/fetch-options.adoc
>> +++ b/Documentation/fetch-options.adoc
>> @@ -73,6 +73,25 @@ See also the `fetch.negotiationAlgorithm` and `push.negotiate`
>>   configuration variables documented in linkgit:git-config[1], and the
>>   `--negotiate-only` option below.
>> +`--negotiation-include=<revision>`::
>> +    Ensure that the given ref tip is always sent as a "have" line
>> +    during fetch negotiation, regardless of what the negotiation
>> +    algorithm selects.  This is useful to guarantee that common
>> +    history reachable from specific refs is always considered, even
>> +    when `--negotiation-restrict` restricts the set of tips or when
>> +    the negotiation algorithm would otherwise skip them.
>> ++
>> +This option may be specified more than once; if so, each ref is sent
>> +unconditionally.
>> ++
>> +The argument may be an exact ref name (e.g. `refs/heads/release`) or a
>> +glob pattern (e.g. `refs/heads/release/{asterisk}`).  The pattern syntax
>> +is the same as for `--negotiation-restrict`.
>> ++
>> +If `--negotiation-restrict` is used, the have set is first restricted by
>> +that option and then increased to include the tips specified by
>> +`--negotiation-include`.
>> +
> 
> The placeholder `<revision>` and the description in the body of "ref
> name or glob" slightly disagree with each other. The `--negotiation-restrict` 
> docs use `(<commit>|<glob>)` in the syntax definition and
> "a glob on ref names, a ref, or .. SHA-1 of a commit".

>> @@ -1547,10 +1548,14 @@ static void add_negotiation_restrict_tips(struct 
>> git_transport_options *smart_op
>>           int old_nr;
>>           if (!has_glob_specials(s)) {
>>               struct object_id oid;
>> +
>> +            /* Ignore missing reference. */
>>               if (repo_get_oid(the_repository, s, &oid))
>> -                die(_("%s is not a valid object"), s);
>> +                continue;
>> +            /* Fail on missing object pointed by ref. */
>>               if (!odb_has_object(the_repository->objects, &oid, 0))
>>                   die(_("the object %s does not exist"), s);
>> +
>>               oid_array_append(oids, &oid);
>>               continue;
>>           }
> 
> This is the change in behaviour - unresolvable revs were a fatal error
> and are now silently ignored.
> 
> Note that t5510 '--negotiation-tip rejects missing OIDs' still passes
> because it uses an all-zero OID, which parses as a valid hex string,
> and dies on the second check "object does not exist". Using something
> like `--negotiation-tip=notreal` that previously would error will now
> silently be ignored.
> 
> Is it worth another test? (invalid object vs not exists)?

>> @@ -1615,6 +1620,13 @@ static struct transport *prepare_transport(struct 
>> remote *remote, int deepen,
>>               strbuf_release(&config_name);
>>           }
>>       }
>> +    if (negotiation_include.nr) {
>> +        if (transport->smart_options)
>> +            transport->smart_options->negotiation_include = 
>> &negotiation_include;
>> +        else
>> +            warning(_("ignoring %s because the protocol does not support it"),
>> +                "--negotiation-include");
>> +    }
>>       return transport;
>>   }
> 
> There is a difference between the existing `--negotiation-restrict`
> option and the new `--negotiation-include` option. Patch 3's commit
> message says:
> 
>    "The 'tips' part is kept because this is an oid_array in the transport
>    layer. This requires the builtin to handle parsing refs into
>    collections of oids so the transport layer can handle this cleaner
>    form of the data."
> 
> The new option passes the raw `string_list` to the transport layer and
> lets it resolve it instead. If the transport layer now learns how to
> resolve refs to oids, why not for tips/restrict?
> 
> Would it be easier for future readers for these complementary options
> to resolve their inputs at the same layer? Or at least call out why:
> "would prefer raw tips but for back-compat we resolve in the built-in"
> for example.

>> +static void resolve_negotiation_include(const struct string_list 
>> *negotiation_include,
>> +                    struct oidset *result)
>> +{
>> +    struct string_list_item *item;
>> +
>> +    if (!negotiation_include || !negotiation_include->nr)
>> +        return;
>> +
>> +    for_each_string_list_item(item, negotiation_include) {
>> +        if (!has_glob_specials(item->string)) {
>> +            struct object_id oid;
>> +
>> +            /* Ignore missing reference. */
>> +            if (repo_get_oid(the_repository, item->string, &oid))
>> +                continue;
>> +
>> +            /* Fail on missing object pointed by ref. */
>> +            if (!odb_has_object(the_repository->objects, &oid, 0))
>> +                die(_("the object %s does not exist"),
>> +                    item->string);
>> +
>> +            oidset_insert(result, &oid);
>> +        } else {
>> +            struct refs_for_each_ref_options opts = {
>> +                .pattern = item->string,
>> +            };
>> +            refs_for_each_ref_ext(
>> +                get_main_ref_store(the_repository),
>> +                add_oid_to_oidset, result, &opts);
>> +        }
>> +    }
>> +}
>> +
> 
> `resolve_negotiation_include()` is basically doing the same as
> `add_negotiation_restrict_tips()` except outputting to an `oidset`
> vs `oid_array`. This is a result of the difference in ref resolution
> layer between `--negotiation-restrict/tip` and `-include`.

>>   static int find_common(struct fetch_negotiator *negotiator,
>>                  struct fetch_pack_args *args,
>>                  int fd[2], struct object_id *result_oid,
>> @@ -347,6 +390,7 @@ static int find_common(struct fetch_negotiator *negotiator,
>>       struct strbuf req_buf = STRBUF_INIT;
>>       size_t state_len = 0;
>>       struct packet_reader reader;
>> +    struct oidset negotiation_include_oids = OIDSET_INIT;
>>       if (args->stateless_rpc && multi_ack == 1)
>>           die(_("the option '%s' requires '%s'"), "--stateless-rpc", 
>> "multi_ack_detailed");
>> @@ -474,6 +518,33 @@ static int find_common(struct fetch_negotiator *negotiator,
>>       trace2_region_enter("fetch-pack", "negotiation_v0_v1", the_repository);
>>       flushes = 0;
>>       retval = -1;
>> +
>> +    /* Send unconditional haves from --negotiation-include */
>> +    resolve_negotiation_include(args->negotiation_include,
>> +                    &negotiation_include_oids);
>> +    if (oidset_size(&negotiation_include_oids)) {
>> +        struct oidset_iter iter;
>> +        oidset_iter_init(&negotiation_include_oids, &iter);
>> +
>> +        while ((oid = oidset_iter_next(&iter))) {
>> +            struct commit *commit;
>> +            packet_buf_write(&req_buf, "have %s\n",
>> +                     oid_to_hex(oid));
>> +            print_verbose(args, "have %s", oid_to_hex(oid));
>> +            count++;
>> +
>> +            /*
>> +             * If this is a commit, then mark as COMMON to
>> +             * avoid the negotiator also outputting it as
>> +             * a have.
>> +             */
>> +            commit = lookup_commit(the_repository, oid);
>> +            if (commit &&
>> +                !repo_parse_commit(the_repository, commit))
>> +                commit->object.flags |= COMMON;
>> +        }
>> +    }
>> +
> 
> I want to make sure I understand the COMMON pre-marking before
> commenting further on this patch. My understanding is there are actually
> two different COMMON bits in the tree, one defined in fetch-pack.c
> (bit 6) and one in negotiator/default.c (bit 2):
> 
> - fetch-pack.c's COMMON (bit 6) is set after a server ACK confirms an
>    OID is common with us and is read to decide when we've established
>    enough common ground to terminate negotiation. This is not consulted
>    in find_common().
> 
> - negotiator/default.c's COMMON (bit 2) is a book-keeping flag used by
>    `get_rev()` to decide if we skip emitting a commit as a 'have'.
> 
> Since we're in fetch-pack.c here, the `commit->object.flags |= COMMON`
> line is setting bit 6. The `get_rev()` call in negotiator/default.c
> never checks bit 6, only bit 2. As far as I can tell, this mark won't
> suppress the negotiator from emitting another 'have' line in the
> protocol v0/v1 paths in `find_common()`.
> 
> The v2 path doesn't touch the flags.. `add_haves` dedups via `oidset_contains()`:
> 
>    while ((oid = negotiator->next(negotiator))) {
>        if (negotiation_include_oids &&
>            oidset_contains(negotiation_include_oids, oid))
>            continue;
>        packet_buf_write(req_buf, "have %s\n", ...);
>    }
> 
> This works, and is what the new 'avoids duplicates with negotiator' test
> runs against, on protocol v2. If we run on protocol v0/v1, and if my
> assessment is correct, then we'd see a duplicate I think?
> 
> Sorry if I've not understood correctly or am missing something, which is
> entirely possible :-)