Re: [PATCH 13/22] pack-bitmap: write multi-pack bitmaps

On Mon, May 03, 2021 at 10:02:30PM -0700, Jonathan Tan wrote:

quoted

quoted

+static void prepare_midx_packing_data(struct packing_data *pdata,
+				      struct write_midx_context *ctx)
+{
+	uint32_t i;
+
+	memset(pdata, 0, sizeof(struct packing_data));
+	prepare_packing_data(the_repository, pdata);
+
+	for (i = 0; i < ctx->entries_nr; i++) {
+		struct pack_midx_entry *from = &ctx->entries[ctx->pack_order[i]];
+		struct object_entry *to = packlist_alloc(pdata, &from->oid);
+
+		oe_set_in_pack(pdata, to,
+			       ctx->info[ctx->pack_perm[from->pack_int_id]].p);
+	}
+}

It is surprising to see this right at the top. Scrolling down, I guess
that there is more information needed than just the packing_data struct.

Hmm, which part is surprising to you? This function is setting up the
packing_data structure that I mentioned in the commit message, which
happens in two steps. First, we allocate and call
prepare_packing_data(). And then we call packlist_alloc() for each
object in the MIDX, setting up some information about each object
(like its OID and which physical pack it came from).

But if any of this is unclear, let me know which part and I'd be happy
to add a clarifying comment.

Ah, I think I was unclear. I was thinking that the commit message led me
to believe that all information needed for creating a bitmap lies in the
packing_data struct, so I would have expected several helper functions
followed by a function that actually writes the packing_data struct.
Maybe the commit message could be rewritten to avoid that confusion, but
it's probably not a big deal.

quoted

quoted

+static struct commit **find_commits_for_midx_bitmap(uint32_t *indexed_commits_nr_p,
+						    struct write_midx_context *ctx)
+{
+	struct rev_info revs;
+	struct bitmap_commit_cb cb;
+
+	memset(&cb, 0, sizeof(struct bitmap_commit_cb));
+	cb.ctx = ctx;
+
+	repo_init_revisions(the_repository, &revs, NULL);
+	for_each_ref(add_ref_to_pending, &revs);
+
+	fetch_if_missing = 0;
+	revs.exclude_promisor_objects = 1;

I think that the MIDX bitmap requires all objects be present? If yes, we
should omit these 2 lines.

It does require that all objects are present, but if we fetched any
promisor objects at this stage it would be too late. That's because by
the time we're in this function, all of the packs that are to be
included in the MIDX should already exist on disk.

Skipping promisor objects here is intentional, since it only excludes
them from the list of reachable commits that we want to select from when
computing the selection of MIDX'd commits to receive bitmaps.

But, if one of those promisor objects is reachable from another object
that is included in the bitmap, then we will complain later on that we
couldn't find a reachability closure (and fail appropriately).

That said, I'm not sure any of that is obvious from reading this code,
so I'll add a comment to that effect around these lines.

So you're saying that if we have missing promisor commits as in the
following graph:

   A
  / \
 B   C
 |   |
 .   .
 .   .
 .   .

where B is missing but promised, and only C is NEEDS_BITMAP, then the
MIDX bitmap write will still work? (So the rev walk is intended to walk
through A and C but not B, and because we are only building bitmaps for
C and potentially its ancestors, we only need the objects in C's
transitive closure.) Even if this is true, "exclude_promisor_objects" is
the wrong option here, because it will exclude all commits that came
from a promisor remote (regardless of whether it is present locally).
(That's how "promisor object" is defined in partial-clone.txt.) What we
need would be an option that permits missing links.

And even if we go with that option that permits missing links, it still
remains that we have very little support for missing promisor commits in
Git right now.

It might be better to just assume that MIDX will only be used for full
clones. If you want, you can add a NEEDSWORK explaining the above case.

quoted

quoted

+
+	if (prepare_revision_walk(&revs))
+		die(_("revision walk setup failed"));
+
+	traverse_commit_list(&revs, bitmap_show_commit, NULL, &cb);
+	if (indexed_commits_nr_p)
+		*indexed_commits_nr_p = cb.commits_nr;
+
+	return cb.commits;
+}

Hmm...I might be missing something obvious, but this function and its
callbacks seem to be written like this in order to put the returned
commits in a certain order. But later on in write_midx_bitmap(), the
return value of this function is passed to
bitmap_writer_select_commits(), which resorts the list anyway?

It isn't intentional, but rather just to build up the list in topo
order. In fact, the order we build it up in isn't quite the same as how
the pack bitmap code generates it (it is in true topo order, at least on
GitHub's servers, as a side effect of using delta islands).

The fact that we resort according to date_compare makes me wonder why
changing that seemed to make such a difference for us. The whole
selection code is a mystery to me.

But no, the order shouldn't matter since we QSORT it later. Any code
here that looks like it's putting it in a certain order has much more to
do with convenience than anything else.

If the order doesn't matter, why don't you just copy one-by-one from
data->ctx->entries into data->commits? (Unless data->ctx->entries has
extra commits?)

quoted

quoted

+	ret = bitmap_writer_build(&pdata);
+	if (!ret)
+		goto cleanup;
+
+	bitmap_writer_set_checksum(midx_hash);
+	bitmap_writer_finish(index, pdata.nr_objects, bitmap_name, 0);

So bitmap_writer_build_type_index() and bitmap_writer_finish() are
called with 2 different orders of commits. Is this expected? If yes,
maybe this is worth a comment.

Confusingly so, but yes, these two do expect different orders. You can
see the same re-sorting going on much more subtly in
pack-write.c:write_idx_file(), which is called by
builtin/pack-objects.c:finish_tmp_packfile(), which happens between
bitmap_writer_build_type_index() and bitmap_writer_finish().

Definitely worth adding a comment.

Ah, I see. Thanks for your explanation.

quoted

quoted

@@ -930,9 +1073,16 @@ static int write_midx_internal(const char *object_dir, struct multi_pack_index *
 		for (i = 0; i < ctx.m->num_packs; i++) {
 			ALLOC_GROW(ctx.info, ctx.nr + 1, ctx.alloc);

+			if (prepare_midx_pack(the_repository, ctx.m, i)) {
+				error(_("could not load pack %s"),
+				      ctx.m->pack_names[i]);
+				result = 1;
+				goto cleanup;
+			}
+
 			ctx.info[ctx.nr].orig_pack_int_id = i;
 			ctx.info[ctx.nr].pack_name = xstrdup(ctx.m->pack_names[i]);
-			ctx.info[ctx.nr].p = NULL;
+			ctx.info[ctx.nr].p = ctx.m->packs[i];
 			ctx.info[ctx.nr].expired = 0;
 			ctx.nr++;
 		}

Why is this needed now and not before? From what I see in this function,
nothing seems to happen to this .p pack except that they are closed
later.

These are used by prepare_midx_packing_data().

Ah, thanks.

quoted

quoted

@@ -1096,6 +1264,15 @@ static int write_midx_internal(const char *object_dir, struct multi_pack_index *
 		if (ctx.info[i].p) {
 			close_pack(ctx.info[i].p);
 			free(ctx.info[i].p);
+			if (ctx.m) {
+				/*
+				 * Destroy a stale reference to the pack in
+				 * 'ctx.m'.
+				 */
+				uint32_t orig = ctx.info[i].orig_pack_int_id;
+				if (orig < ctx.m->num_packs)
+					ctx.m->packs[orig] = NULL;
+			}
 		}
 		free(ctx.info[i].pack_name);
 	}

Is this hunk needed? "ctx" is a local variable and will not outlast this
function.

I can't remember exactly why I added this. I'll play around with it and
either remove it or add a comment why it's necessary before the next
reroll.

OK.

quoted

I'll review the rest tomorrow. It seems like I've gotten over the most
difficult patches.

Thanks, and sorry that this took me a few days to get back to. I
appreciate your review immensely.

No worries, and thanks for these patches.