[RFC PATCH 0/2] Alternate ab/valgrind-fixes fix-up

On Mon, May 16 2022, Derrick Stolee wrote:

This whole topic-at-large is a stylistic fix-up for a case where I obviously got it wrong, so take this with a double grain of salt.

Re the "What's Cooking" mention of ds/object-file-unpack-loose-header-fix: I don't mind it being merged down at all. The below is all small potatoes.

I don't think the rationale ("adds complications for future developers") makes sense in this case.

Let's leave aside the question of whether we exhaustively check enum arms as in the pre-image, or check "not ok" as in the post-image.

Surely we can agree that whatever pattern is preferred we're better off consistently picking one or the other?

I think this proposed change would make more sense and be in line with
its commit message if it also proposed this:
	
	diff --git a/streaming.c b/streaming.c
	index fe54665d86e..bb4ed198463 100644
	--- a/streaming.c
	+++ b/streaming.c
	@@ -230,15 +230,10 @@ static int open_istream_loose(struct git_istream *st, struct repository *r,
	 	st->u.loose.mapped = map_loose_object(r, oid, &st->u.loose.mapsize);
	 	if (!st->u.loose.mapped)
	 		return -1;
	-	switch (unpack_loose_header(&st->z, st->u.loose.mapped,
	-				    st->u.loose.mapsize, st->u.loose.hdr,
	-				    sizeof(st->u.loose.hdr), NULL)) {
	-	case ULHR_OK:
	-		break;
	-	case ULHR_BAD:
	-	case ULHR_TOO_LONG:
	+	if (unpack_loose_header(&st->z, st->u.loose.mapped,
	+				st->u.loose.mapsize, st->u.loose.hdr,
	+				sizeof(st->u.loose.hdr), NULL) != ULHR_OK)
	 		goto error;
	-	}
	 	if (parse_loose_header(st->u.loose.hdr, &oi) < 0 || *type < 0)
	 		goto error;

I.e. now we've converted the 2/3 callers of the API that only cared about "not OK", there's a third one that cares about all the enum arms currently, so that one remains a "switch".

The reason I think the rationale doesn't make sense is because of this inconsistency. I.e. if we suppose a developer adds another enum value, they'll then discover those three callers.

Surely whatever our preference for how to handle those 2/3 callers it's less complicated if they don't use different patterns for no obvious reason.

But anyway. Looking a bit deeper at this code again I think these two patches are where we'd eventually want to head with this API. I.e. I think the whole business of making this a tri-state return was premature on my part.

After this RFC unpack_loose_header() is again a function that returns a negative value on error, and the enum is gone. As noted in 2/2 there's a slight trade-off there, which I think is for the better, both in terms of API simplicity, and in the new "error" output we'll omit in these obscure cases. I.e.:

	-				error: header for $bogus_long_sha1 too long, exceeds 32 bytes
	+				error: header too long, exceeds 32 bytes
	+				error: unable to unpack $bogus_long_sha1 header

This whole "switch" complexity was because the old error message wanted to note the OID in the "header too long" message.

Again, I'm perfectly fine with ds/object-file-unpack-loose-header-fix advancing to "next", I can rebase this on top, or drop it depending on the consensus about whether it's worthwile. I did want to un-block that topic one way or the other, so to the extent that it was waiting on my feedback...

Ævar Arnfjörð Bjarmason (2):
  object-file API: fix obscure unpack_loose_header() return
  object-file API: have unpack_loose_header() return "int" again

 cache.h             | 25 +++++-------------------
 object-file.c       | 46 +++++++++++++++++----------------------------
 streaming.c         | 11 +++--------
 t/t1006-cat-file.sh |  6 ++++--
 4 files changed, 29 insertions(+), 59 deletions(-)