repack: exclude cruft pack(s) from the MIDX where possible

In ddee370 (builtin/repack.c: add cruft packs to MIDX during
geometric repack, 2022-05-20), repack began adding cruft pack(s) to the
MIDX with '--write-midx' to ensure that the resulting MIDX was always
closed under reachability in order to generate reachability bitmaps.

While the previous patch added the '--stdin-packs=follow' option to
pack-objects, it is not yet on by default. Given that, suppose you have
a once-unreachable object packed in a cruft pack, which later becomes
reachable from one or more objects in a geometrically repacked pack.
That once-unreachable object *won't* appear in the new pack, since the
cruft pack was not specified as included or excluded when the
geometrically repacked pack was created with 'pack-objects
--stdin-packs' (*not* '--stdin-packs=follow', which is not on). If that
new pack is included in a MIDX without the cruft pack, then trying to
generate bitmaps for that MIDX may fail. This happens when the bitmap
selection process picks one or more commits which reach the
once-unreachable objects.

To mitigate this failure mode, commit ddee370 ensures that the MIDX
will be closed under reachability by including cruft pack(s). If cruft
pack(s) were not included, we would fail to generate a MIDX bitmap. But
ddee370 alludes to the fact that this is sub-optimal by saying

    [...] it's desirable to avoid including cruft packs in the MIDX
    because it causes the MIDX to store a bunch of objects which are
    likely to get thrown away.

, which is true, but hides an even larger problem. If repositories
rarely prune their unreachable objects and/or have many of them, the
MIDX must keep track of a large number of objects which bloats the MIDX
and slows down object lookup.

This is doubly unfortunate because the vast majority of objects in cruft
pack(s) are unlikely to be read. But any object lookups that go through
the MIDX must binary search over them anyway, slowing down object
lookups using the MIDX.

This patch causes geometrically-repacked packs to contain a copy of any
once-unreachable object(s) with 'git pack-objects --stdin-packs=follow',
allowing us to avoid including any cruft packs in the MIDX. This is
because a sequence of geometrically-repacked packs that were all
generated with '--stdin-packs=follow' are guaranteed to have their union
be closed under reachability.

Note that you cannot guarantee that a collection of packs is closed
under reachability if not all of them were generated with "following" as
above. One tell-tale sign that not all geometrically-repacked packs in
the MIDX were generated with "following" is to see if there is a pack in
the existing MIDX that is not going to be somehow represented (either
verbatim or as part of a geometric rollup) in the new MIDX.

If there is, then starting to generate packs with "following" during
geometric repacking won't work, since it's open to the same race as
described above.

But if you're starting from scratch (e.g., building the first MIDX after
an all-into-one '--cruft' repack), then you can guarantee that the union
of subsequently generated packs from geometric repacking *is* closed
under reachability.

(One exception here is when "starting from scratch" results in a noop
repack, e.g., because the non-cruft pack(s) in a repository already form
a geometric progression. Since we can't tell whether or not those were
generated with '--stdin-packs=follow', they may depend on
once-unreachable objects, so we have to include the cruft pack in the
MIDX in this case.)

Detect when this is the case and avoid including cruft packs in the MIDX
where possible. The existing behavior remains the default, and the new
behavior is available with the config 'repack.midxMustIncludeCruft' set
to 'false'.

Signed-off-by: Taylor Blau <[email protected]>
Signed-off-by: Junio C Hamano <[email protected]>

Author: ttaylorr

Documentation/config/repack.adoc

@@ -39,3 +39,10 @@ repack.cruftThreads::
 	a cruft pack and the respective parameters are not given over
 	the command line. See similarly named `pack.*` configuration
 	variables for defaults and meaning.
+
+repack.midxMustContainCruft::
+	When set to true, linkgit:git-repack[1] will unconditionally include
+	cruft pack(s), if any, in the multi-pack index when invoked with
+	`--write-midx`. When false, cruft packs are only included in the MIDX
+	when necessary (e.g., because they might be required to form a
+	reachability closure with MIDX bitmaps). Defaults to true.

builtin/repack.c

@@ -39,6 +39,7 @@ static int write_bitmaps = -1;
 static int use_delta_islands;
 static int run_update_server_info = 1;
 static char *packdir, *packtmp_name, *packtmp;
+static int midx_must_contain_cruft = 1;
 
 static const char *const git_repack_usage[] = {
 	N_("git repack [-a] [-A] [-d] [-f] [-F] [-l] [-n] [-q] [-b] [-m]\n"
@@ -108,6 +109,10 @@ static int repack_config(const char *var, const char *value,
 		free(cruft_po_args->threads);
 		return git_config_string(&cruft_po_args->threads, var, value);
 	}
+	if (!strcmp(var, "repack.midxmustcontaincruft")) {
+		midx_must_contain_cruft = git_config_bool(var, value);
+		return 0;
+	}
 	return git_default_config(var, value, ctx, cb);
 }
 
@@ -690,6 +695,77 @@ static void free_pack_geometry(struct pack_geometry *geometry)
 	free(geometry->pack);
 }
 
+static int midx_has_unknown_packs(char **midx_pack_names,
+				  size_t midx_pack_names_nr,
+				  struct string_list *include,
+				  struct pack_geometry *geometry,
+				  struct existing_packs *existing)
+{
+	size_t i;
+
+	string_list_sort(include);
+
+	for (i = 0; i < midx_pack_names_nr; i++) {
+		const char *pack_name = midx_pack_names[i];
+
+		/*
+		 * Determine whether or not each MIDX'd pack from the existing
+		 * MIDX (if any) is represented in the new MIDX. For each pack
+		 * in the MIDX, it must either be:
+		 *
+		 *  - In the "include" list of packs to be included in the new
+		 *    MIDX. Note this function is called before the include
+		 *    list is populated with any cruft pack(s).
+		 *
+		 *  - Below the geometric split line (if using pack geometry),
+		 *    indicating that the pack won't be included in the new
+		 *    MIDX, but its contents were rolled up as part of the
+		 *    geometric repack.
+		 *
+		 *  - In the existing non-kept packs list (if not using pack
+		 *    geometry), and marked as non-deleted.
+		 */
+		if (string_list_has_string(include, pack_name)) {
+			continue;
+		} else if (geometry) {
+			struct strbuf buf = STRBUF_INIT;
+			uint32_t j;
+
+			for (j = 0; j < geometry->split; j++) {
+				strbuf_reset(&buf);
+				strbuf_addstr(&buf, pack_basename(geometry->pack[j]));
+				strbuf_strip_suffix(&buf, ".pack");
+				strbuf_addstr(&buf, ".idx");
+
+				if (!strcmp(pack_name, buf.buf)) {
+					strbuf_release(&buf);
+					break;
+				}
+			}
+
+			strbuf_release(&buf);
+
+			if (j < geometry->split)
+				continue;
+		} else {
+			struct string_list_item *item;
+
+			item = string_list_lookup(&existing->non_kept_packs,
+						  pack_name);
+			if (item && !pack_is_marked_for_deletion(item))
+				continue;
+		}
+
+		/*
+		 * If we got to this point, the MIDX includes some pack that we
+		 * don't know about.
+		 */
+		return 1;
+	}
+
+	return 0;
+}
+
 struct midx_snapshot_ref_data {
 	struct tempfile *f;
 	struct oidset seen;
@@ -758,6 +834,8 @@ static void midx_snapshot_refs(struct tempfile *f)
 
 static void midx_included_packs(struct string_list *include,
 				struct existing_packs *existing,
+				char **midx_pack_names,
+				size_t midx_pack_names_nr,
 				struct string_list *names,
 				struct pack_geometry *geometry)
 {
@@ -811,26 +889,56 @@ static void midx_included_packs(struct string_list *include,
 		}
 	}
 
-	for_each_string_list_item(item, &existing->cruft_packs) {
+	if (midx_must_contain_cruft ||
+	    midx_has_unknown_packs(midx_pack_names, midx_pack_names_nr,
+				   include, geometry, existing)) {
 		/*
-		 * When doing a --geometric repack, there is no need to check
-		 * for deleted packs, since we're by definition not doing an
-		 * ALL_INTO_ONE repack (hence no packs will be deleted).
-		 * Otherwise we must check for and exclude any packs which are
-		 * enqueued for deletion.
+		 * If there are one or more unknown pack(s) present (see
+		 * midx_has_unknown_packs() for what makes a pack
+		 * "unknown") in the MIDX before the repack, keep them
+		 * as they may be required to form a reachability
+		 * closure if the MIDX is bitmapped.
 		 *
-		 * So we could omit the conditional below in the --geometric
-		 * case, but doing so is unnecessary since no packs are marked
-		 * as pending deletion (since we only call
-		 * `mark_packs_for_deletion()` when doing an all-into-one
-		 * repack).
+		 * For example, a cruft pack can be required to form a
+		 * reachability closure if the MIDX is bitmapped and one
+		 * or more of the bitmap's selected commits reaches a
+		 * once-cruft object that was later made reachable.
 		 */
-		if (pack_is_marked_for_deletion(item))
-			continue;
+		for_each_string_list_item(item, &existing->cruft_packs) {
+			/*
+			 * When doing a --geometric repack, there is no
+			 * need to check for deleted packs, since we're
+			 * by definition not doing an ALL_INTO_ONE
+			 * repack (hence no packs will be deleted).
+			 * Otherwise we must check for and exclude any
+			 * packs which are enqueued for deletion.
+			 *
+			 * So we could omit the conditional below in the
+			 * --geometric case, but doing so is unnecessary
+			 *  since no packs are marked as pending
+			 *  deletion (since we only call
+			 *  `mark_packs_for_deletion()` when doing an
+			 *  all-into-one repack).
+			 */
+			if (pack_is_marked_for_deletion(item))
+				continue;
 
-		strbuf_reset(&buf);
-		strbuf_addf(&buf, "%s.idx", item->string);
-		string_list_insert(include, buf.buf);
+			strbuf_reset(&buf);
+			strbuf_addf(&buf, "%s.idx", item->string);
+			string_list_insert(include, buf.buf);
+		}
+	} else {
+		/*
+		 * Modern versions of Git (with the appropriate
+		 * configuration setting) will write new copies of
+		 * once-cruft objects when doing a --geometric repack.
+		 *
+		 * If the MIDX has no cruft pack, new packs written
+		 * during a --geometric repack will not rely on the
+		 * cruft pack to form a reachability closure, so we can
+		 * avoid including them in the MIDX in that case.
+		 */
+		;
 	}
 
 	strbuf_release(&buf);
@@ -1145,6 +1253,8 @@ int cmd_repack(int argc,
 	struct tempfile *refs_snapshot = NULL;
 	int i, ext, ret;
 	int show_progress;
+	char **midx_pack_names = NULL;
+	size_t midx_pack_names_nr = 0;
 
 	/* variables to be filled by option parsing */
 	int delete_redundant = 0;
@@ -1361,7 +1471,10 @@ int cmd_repack(int argc,
 		    !(pack_everything & PACK_CRUFT))
 			strvec_push(&cmd.args, "--pack-loose-unreachable");
 	} else if (geometry.split_factor) {
-		strvec_push(&cmd.args, "--stdin-packs");
+		if (midx_must_contain_cruft)
+			strvec_push(&cmd.args, "--stdin-packs");
+		else
+			strvec_push(&cmd.args, "--stdin-packs=follow");
 		strvec_push(&cmd.args, "--unpacked");
 	} else {
 		strvec_push(&cmd.args, "--unpacked");
@@ -1401,8 +1514,25 @@ int cmd_repack(int argc,
 	if (ret)
 		goto cleanup;
 
-	if (!names.nr && !po_args.quiet)
-		printf_ln(_("Nothing new to pack."));
+	if (!names.nr) {
+		if (!po_args.quiet)
+			printf_ln(_("Nothing new to pack."));
+		/*
+		 * If we didn't write any new packs, the non-cruft packs
+		 * may refer to once-unreachable objects in the cruft
+		 * pack(s).
+		 *
+		 * If there isn't already a MIDX, the one we write
+		 * must include the cruft pack(s), in case the
+		 * non-cruft pack(s) refer to once-cruft objects.
+		 *
+		 * If there is already a MIDX, we can punt here, since
+		 * midx_has_unknown_packs() will make the decision for
+		 * us.
+		 */
+		if (!get_local_multi_pack_index(the_repository))
+			midx_must_contain_cruft = 1;
+	}
 
 	if (pack_everything & PACK_CRUFT) {
 		const char *pack_prefix = find_pack_prefix(packdir, packtmp);
@@ -1483,6 +1613,19 @@ int cmd_repack(int argc,
 
 	string_list_sort(&names);
 
+	if (get_local_multi_pack_index(the_repository)) {
+		struct multi_pack_index *m =
+			get_local_multi_pack_index(the_repository);
+
+		ALLOC_ARRAY(midx_pack_names,
+			    m->num_packs + m->num_packs_in_base);
+
+		for (; m; m = m->base_midx)
+			for (uint32_t i = 0; i < m->num_packs; i++)
+				midx_pack_names[midx_pack_names_nr++] =
+					xstrdup(m->pack_names[i]);
+	}
+
 	close_object_store(the_repository->objects);
 
 	/*
@@ -1524,7 +1667,8 @@ int cmd_repack(int argc,
 
 	if (write_midx) {
 		struct string_list include = STRING_LIST_INIT_DUP;
-		midx_included_packs(&include, &existing, &names, &geometry);
+		midx_included_packs(&include, &existing, midx_pack_names,
+				    midx_pack_names_nr, &names, &geometry);
 
 		ret = write_midx_included_packs(&include, &geometry, &names,
 						refs_snapshot ? get_tempfile_path(refs_snapshot) : NULL,
@@ -1575,6 +1719,9 @@ int cmd_repack(int argc,
 	string_list_clear(&names, 1);
 	existing_packs_release(&existing);
 	free_pack_geometry(&geometry);
+	for (size_t i = 0; i < midx_pack_names_nr; i++)
+		free(midx_pack_names[i]);
+	free(midx_pack_names);
 	pack_objects_args_release(&po_args);
 	pack_objects_args_release(&cruft_po_args);