diff mbox series

[v3,07/25] midx: infer preferred pack when not given one

Message ID 53ef0a6d67469078f237e2719052ab05f642cf2a.1627420428.git.me@ttaylorr.com (mailing list archive)
State Superseded
Headers show
Series multi-pack reachability bitmaps | expand

Commit Message

Taylor Blau July 27, 2021, 9:19 p.m. UTC
In 9218c6a40c (midx: allow marking a pack as preferred, 2021-03-30), the
multi-pack index code learned how to select a pack which all duplicate
objects are selected from. That is, if an object appears in multiple
packs, select the copy in the preferred pack before breaking ties
according to the other rules like pack mtime and readdir() order.

Not specifying a preferred pack can cause serious problems with
multi-pack reachability bitmaps, because these bitmaps rely on having at
least one pack from which all duplicates are selected. Not having such a
pack causes problems with the code in pack-objects to reuse packs
verbatim (e.g., that code assumes that a delta object in a chunk of pack
sent verbatim will have its base object sent from the same pack).

So why does not marking a pack preferred cause problems here? The reason
is roughly as follows:

  - Ties are broken (when handling duplicate objects) by sorting
    according to midx_oid_compare(), which sorts objects by OID,
    preferred-ness, pack mtime, and finally pack ID (more on that
    later).

  - The psuedo pack-order (described in
    Documentation/technical/pack-format.txt under the section
    "multi-pack-index reverse indexes") is computed by
    midx_pack_order(), and sorts by pack ID and pack offset, with
    preferred packs sorting first.

  - But! Pack IDs come from incrementing the pack count in
    add_pack_to_midx(), which is a callback to
    for_each_file_in_pack_dir(), meaning that pack IDs are assigned in
    readdir() order.

When specifying a preferred pack, all of that works fine, because
duplicate objects are correctly resolved in favor of the copy in the
preferred pack, and the preferred pack sorts first in the object order.

"Sorting first" is critical, because the bitmap code relies on finding
out which pack holds the first object in the MIDX's pseudo pack-order to
determine which pack is preferred.

But if we didn't specify a preferred pack, and the pack which comes
first in readdir() order does not also have the lowest timestamp, then
it's possible that that pack (the one that sorts first in pseudo-pack
order, which the bitmap code will treat as the preferred one) did *not*
have all duplicate objects resolved in its favor, resulting in breakage.

The fix is simple: pick a (semi-arbitrary, non-empty) preferred pack
when none was specified. This forces that pack to have duplicates
resolved in its favor, and (critically) to sort first in pseudo-pack
order.  Unfortunately, testing this behavior portably isn't possible,
since it depends on readdir() order which isn't guaranteed by POSIX.

(Note that multi-pack reachability bitmaps have yet to be implemented;
so in that sense this patch is fixing a bug which does not yet exist.
But by having this patch beforehand, we can prevent the bug from ever
materializing.)

Signed-off-by: Taylor Blau <me@ttaylorr.com>
---
 midx.c | 50 ++++++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 44 insertions(+), 6 deletions(-)
diff mbox series

Patch

diff --git a/midx.c b/midx.c
index 092dbf45b6..8956492b9c 100644
--- a/midx.c
+++ b/midx.c
@@ -969,15 +969,57 @@  static int write_midx_internal(const char *object_dir, struct multi_pack_index *
 	if (ctx.m && ctx.nr == ctx.m->num_packs && !packs_to_drop)
 		goto cleanup;
 
-	ctx.preferred_pack_idx = -1;
 	if (preferred_pack_name) {
+		int found = 0;
 		for (i = 0; i < ctx.nr; i++) {
 			if (!cmp_idx_or_pack_name(preferred_pack_name,
 						  ctx.info[i].pack_name)) {
 				ctx.preferred_pack_idx = i;
+				found = 1;
 				break;
 			}
 		}
+
+		if (!found)
+			warning(_("unknown preferred pack: '%s'"),
+				preferred_pack_name);
+	} else if (ctx.nr && (flags & MIDX_WRITE_REV_INDEX)) {
+		struct packed_git *oldest = ctx.info[ctx.preferred_pack_idx].p;
+		ctx.preferred_pack_idx = 0;
+
+		if (packs_to_drop && packs_to_drop->nr)
+			BUG("cannot write a MIDX bitmap during expiration");
+
+		/*
+		 * set a preferred pack when writing a bitmap to ensure that
+		 * the pack from which the first object is selected in pseudo
+		 * pack-order has all of its objects selected from that pack
+		 * (and not another pack containing a duplicate)
+		 */
+		for (i = 1; i < ctx.nr; i++) {
+			struct packed_git *p = ctx.info[i].p;
+
+			if (!oldest->num_objects || p->mtime < oldest->mtime) {
+				oldest = p;
+				ctx.preferred_pack_idx = i;
+			}
+		}
+
+		if (!oldest->num_objects) {
+			/*
+			 * If all packs are empty; unset the preferred index.
+			 * This is acceptable since there will be no duplicate
+			 * objects to resolve, so the preferred value doesn't
+			 * matter.
+			 */
+			ctx.preferred_pack_idx = -1;
+		}
+	} else {
+		/*
+		 * otherwise don't mark any pack as preferred to avoid
+		 * interfering with expiration logic below
+		 */
+		ctx.preferred_pack_idx = -1;
 	}
 
 	if (ctx.preferred_pack_idx > -1) {
@@ -1058,11 +1100,7 @@  static int write_midx_internal(const char *object_dir, struct multi_pack_index *
 						      ctx.info, ctx.nr,
 						      sizeof(*ctx.info),
 						      idx_or_pack_name_cmp);
-
-		if (!preferred)
-			warning(_("unknown preferred pack: '%s'"),
-				preferred_pack_name);
-		else {
+		if (preferred) {
 			uint32_t perm = ctx.pack_perm[preferred->orig_pack_int_id];
 			if (perm == PACK_EXPIRED)
 				warning(_("preferred pack '%s' is expired"),