repo-settings: add fetch.haveRefs config

In some large repos and strange commit history data shapes, it can be
very expensive to not advertise certain references. For example, a
parallel 'release' branch that occasionally merges the 'main' branch and
augments the content with release metadata could create a long
commit history that is not reachable from any other user development
branches. If the client fails to advertise its copy of the 'release'
branch, then it could end up redownloading that entire history in the
range 'main..origin/release' instead of only 'release..origin/release'.

Add a new multi-valued config key 'fetch.haveRefs' that forces certain
references to be advertised during negotiation. This change focuses on
safely parsing and storing those strings in the repo_settings struct.

Some things that I needed to be careful about: the
repo_config_get_string_multi() method fills the given list with a
pointer to a list that is managed by the config subsystem. Holding a
pointer to that list is not safe, so it must be copied before moving on
to other config reads.

Since we have a deep copy of these values, we need to free them as the
settings are cleared.

Signed-off-by: Derrick Stolee <stolee@gmail.com>

Author: derrickstolee

Documentation/config/fetch.adoc

@@ -79,6 +79,25 @@
 See also the `--negotiate-only` and `--negotiation-tip` options to
 linkgit:git-fetch[1].
 
+`fetch.haveRefs`::
+	When negotiating with a remote during `git fetch` and `git push`,
+	the client advertises a list of commits that exist locally. In
+	repos with many references, this list of "haves" can be truncated.
+	Depending on data shape, dropping certain references may be
+	expensive. This multi-valued config option specifies ref patterns
+	whose tips should always be sent as "have" commits during fetch
+	negotiation.
++
+Each value is either an exact ref name (e.g. `refs/heads/release`) or a
+glob pattern (e.g. `refs/heads/release/*`). The pattern syntax is the same
+as for `--negotiation-tip`.
++
+This option is additive with the normal negotiation process: the
+negotiation algorithm still runs and advertises its own selected commits,
+but the refs matching `fetch.haveRefs` are sent unconditionally on top of
+those heuristically selected commits. This option is also used during push
+negotiation when `push.negotiate` is enabled.
+
 `fetch.showForcedUpdates`::
 	Set to `false` to enable `--no-show-forced-updates` in
 	linkgit:git-fetch[1] and linkgit:git-pull[1] commands.

repo-settings.c

@@ -35,6 +35,7 @@ void prepare_repo_settings(struct repository *r)
 	int manyfiles;
 	int read_changed_paths;
 	unsigned long ulongval;
+	const struct string_list *have_refs_config;
 
 	if (!r->gitdir)
 		BUG("Cannot add settings for uninitialized repository");
@@ -131,6 +132,16 @@ void prepare_repo_settings(struct repository *r)
 			die("unknown fetch negotiation algorithm '%s'", strval);
 	}
 
+	if (!repo_config_get_string_multi(r, "fetch.haverefs",
+					  &have_refs_config)) {
+		struct string_list_item *item;
+		r->settings.fetch_have_refs = xmalloc(sizeof(*r->settings.fetch_have_refs));
+		string_list_init_dup(r->settings.fetch_have_refs);
+		for_each_string_list_item(item, have_refs_config)
+			string_list_append(r->settings.fetch_have_refs,
+					   item->string);
+	}
+
 	/*
 	 * This setting guards all index reads to require a full index
 	 * over a sparse index. After suitable guards are placed in the
@@ -161,6 +172,10 @@ void repo_settings_clear(struct repository *r)
 	struct repo_settings empty = REPO_SETTINGS_INIT;
 	FREE_AND_NULL(r->settings.fsmonitor);
 	FREE_AND_NULL(r->settings.hooks_path);
+	if (r->settings.fetch_have_refs) {
+		string_list_clear(r->settings.fetch_have_refs, 0);
+		FREE_AND_NULL(r->settings.fetch_have_refs);
+	}
 	r->settings = empty;
 }
 

repo-settings.h

@@ -3,6 +3,7 @@
 
 struct fsmonitor_settings;
 struct repository;
+struct string_list;
 
 enum untracked_cache_setting {
 	UNTRACKED_CACHE_KEEP,
@@ -58,6 +59,7 @@ struct repo_settings {
 	int pack_use_sparse;
 	int pack_use_path_walk;
 	enum fetch_negotiation_setting fetch_negotiation_algorithm;
+	struct string_list *fetch_have_refs;
 
 	int core_multi_pack_index;
 	int warn_ambiguous_refs; /* lazily loaded via accessor */