fsck: lazily load types under --connectivity-onlyjk/fsck-connectivity-check-fix

The recent fixes to "fsck --connectivity-only" load all of
the objects with their correct types. This keeps the
connectivity-only code path close to the regular one, but it
also introduces some unnecessary inefficiency. While getting
the type of an object is cheap compared to actually opening
and parsing the object (as the non-connectivity-only case
would do), it's still not free.

For reachable non-blob objects, we end up having to parse
them later anyway (to see what they point to), making our
type lookup here redundant.

For unreachable objects, we might never hit them at all in
the reachability traversal, making the lookup completely
wasted. And in some cases, we might have quite a few
unreachable objects (e.g., when alternates are used for
shared object storage between repositories, it's normal for
there to be objects reachable from other repositories but
not the one running fsck).

The comment in mark_object_for_connectivity() claims two
benefits to getting the type up front:

  1. We need to know the types during fsck_walk(). (And not
     explicitly mentioned, but we also need them when
     printing the types of broken or dangling commits).

     We can address this by lazy-loading the types as
     necessary. Most objects never need this lazy-load at
     all, because they fall into one of these categories:

       a. Reachable from our tips, and are coerced into the
	  correct type as we traverse (e.g., a parent link
	  will call lookup_commit(), which converts OBJ_NONE
	  to OBJ_COMMIT).

       b. Unreachable, but not at the tip of a chunk of
          unreachable history. We only mention the tips as
	  "dangling", so an unreachable commit which links
	  to hundreds of other objects needs only report the
	  type of the tip commit.

  2. It serves as a cross-check that the coercion in (1a) is
     correct (i.e., we'll complain about a parent link that
     points to a blob). But we get most of this for free
     already, because right after coercing, we'll parse any
     non-blob objects. So we'd notice then if we expected a
     commit and got a blob.

     The one exception is when we expect a blob, in which
     case we never actually read the object contents.

     So this is a slight weakening, but given that the whole
     point of --connectivity-only is to sacrifice some data
     integrity checks for speed, this seems like an
     acceptable tradeoff.

Here are before and after timings for an extreme case with
~5M reachable objects and another ~12M unreachable (it's the
torvalds/linux repository on GitHub, connected to shared
storage for all of the other kernel forks):

  [before]
  $ time git fsck --no-dangling --connectivity-only
  real	3m4.323s
  user	1m25.121s
  sys	1m38.710s

  [after]
  $ time git fsck --no-dangling --connectivity-only
  real	0m51.497s
  user	0m49.575s
  sys	0m1.776s

Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>

1 files changed, 4 insertions, 0 deletions

diff --git a/fsck.c b/fsck.c
index 4a3069e204..939792752b 100644
--- a/fsck.c
+++ b/fsck.c
@@ -458,6 +458,10 @@ int fsck_walk(struct object *obj, void *data, struct fsck_options *options)
 {
 	if (!obj)
 		return -1;
+
+	if (obj->type == OBJ_NONE)
+		parse_object(obj->oid.hash);
+
 	switch (obj->type) {
 	case OBJ_BLOB:
 		return 0;