commit.c: don't persist substituted parents when unshallowing
commitce16364e897e70a17bd1864b6007719eeec959f3
authorTaylor Blau <me@ttaylorr.com>
Wed, 8 Jul 2020 21:10:53 +0000 (8 17:10 -0400)
committerJunio C Hamano <gitster@pobox.com>
Wed, 8 Jul 2020 23:13:46 +0000 (8 16:13 -0700)
tree1d3d8c13dbb374921d02926647f6b4671fc58319
parent37b9dcabfc48b0cbce638140279878dac37aec73
commit.c: don't persist substituted parents when unshallowing

Since 37b9dcabfc (shallow.c: use '{commit,rollback}_shallow_file',
2020-04-22), Git knows how to reset stat-validity checks for the
$GIT_DIR/shallow file, allowing it to change between a shallow and
non-shallow state in the same process (e.g., in the case of 'git fetch
--unshallow').

However, when $GIT_DIR/shallow changes, Git does not alter or remove any
grafts (nor substituted parents) in memory.

This comes up in a "git fetch --unshallow" with fetch.writeCommitGraph
set to true. Ordinarily in a shallow repository (and before 37b9dcabfc,
even in this case), commit_graph_compatible() would return false,
indicating that the repository should not be used to write a
commit-graphs (since commit-graph files cannot represent a shallow
history). But since 37b9dcabfc, in an --unshallow operation that check
succeeds.

Thus even though the repository isn't shallow any longer (that is, we
have all of the objects), the in-core representation of those objects
still has munged parents at the shallow boundaries.  When the
commit-graph write proceeds, we use the incorrect parentage, producing
wrong results.

There are two ways for a user to work around this: either (1) set
'fetch.writeCommitGraph' to 'false', or (2) drop the commit-graph after
unshallowing.

One way to fix this would be to reset the parsed object pool entirely
(flushing the cache and thus preventing subsequent reads from modifying
their parents) after unshallowing. That would produce a problem when
callers have a now-stale reference to the old pool, and so this patch
implements a different approach. Instead, attach a new bit to the pool,
'substituted_parent', which indicates if the repository *ever* stored a
commit which had its parents modified (i.e., the shallow boundary
prior to unshallowing).

This bit needs to be sticky because all reads subsequent to modifying a
commit's parents are unreliable when unshallowing. Modify the check in
'commit_graph_compatible' to take this bit into account, and correctly
avoid generating commit-graphs in this case, thus solving the bug.

Helped-by: Derrick Stolee <dstolee@microsoft.com>
Helped-by: Jonathan Nieder <jrnieder@gmail.com>
Reported-by: Jay Conrod <jayconrod@google.com>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
commit-graph.c
commit.c
object.h
t/t5537-fetch-shallow.sh