patchcommit.c: don't persist substituted parents when unshallowing

In 37b9dcabfc (shallow.c: use '{commit,rollback}_shallow_file', 2020-04-22), Git learned 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, 37b9dcabfc does not alter or remove any grafts nor substituted parents. This produces a problem when unshallowing if another part of the code relies on the un-grafted and/or non-substituted parentage for commits after, say, fetching.

This can arise 'fetch.writeCommitGraph' is true. Ordinarily (and certainly previous to 37b9dcabfc), 'commit_graph_compatible()' would return 0, indicating that the repository should not generate commit-graphs (since at one point in the same process it was shallow). But with 37b9dcabfc, that check succeeds.

This is where the bug occurs: 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. If a commit-graph write proceeds, we will use the incorrect parentage, producing wrong results.

(Prior to this patch, there were two ways of fixing 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. This can produce a problem when callers have a now-stale reference to the old pool, and so this patch implements a different approach. Instead, we 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 *before* unshallowing).

This bit is sticky, since all subsequent reads after modifying a commit's parent are unreliable when unshallowing. This patch modifies the check in 'commit_graph_compatible' to take this bit into account, and correctly avoid generating commit-graphs in this case.

Helped-by: Derrick Stolee <dstolee@microsoft.com>
Helped-by: Jonathan Nieder <jrnieder@gmail.com>
Reported-by: Jay Conrod <jayconrod@google.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
---
This is a follow-up to Jonathan Nieder's recent message; this patch
fixes the persistent-shallow issue originally reported by Jay Conrod in:

  https://lore.kernel.org/git/20200603034213.GB253041@google.com/

Like Jonathan, I am also late to send this with -rc0 so close around the corner. I think that this *could* wait until v2.28.1 or v2.29.0 since fetch.writeCommitGraph is no longer implied by feature.experimental, but I figure that it is probably better to get this into v2.28.0 since it fixes the issue once and for all, so long as there is consensus that the patch is good.

Thanks in advance for a review.

 commit-graph.c           |  3 ++-
 commit.c                 |  2 ++
 object.h                 |  1 +
 t/t5537-fetch-shallow.sh | 14 ++++++++++++++
 4 files changed, 19 insertions(+), 1 deletion(-)

...oops; I was scripting around my git-who script this morning, and clearly munged a few of the recipient's emails. They are fixed in this email.

On Tue, Jul 07, 2020 at 10:42:22AM -0400, Taylor Blau wrote:

Thanks, Taylor

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>
---
The patch contents are unchanged from v1, but the patch's message is
wordsmithed with jrn's help. This should be ready for -rc0 now.

Range-diff below for easier review:

Range-diff against v1:
1:  82939831ad ! 1:  1f769bbbb4 commit.c: don't persist substituted parents when unshallowing
    @@ Metadata
      ## Commit message ##
         commit.c: don't persist substituted parents when unshallowing

    -    In 37b9dcabfc (shallow.c: use '{commit,rollback}_shallow_file',
    -    2020-04-22), Git learned how to reset stat-validity checks for the
    -    '$GIT_DIR/shallow' file, allowing it to change between a shallow and
    +    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, 37b9dcabfc does not alter or remove any grafts nor substituted
    -    parents. This produces a problem when unshallowing if another part of
    -    the code relies on the un-grafted and/or non-substituted parentage for
    -    commits after, say, fetching.
    +    However, when $GIT_DIR/shallow changes, Git does not alter or remove any
    +    grafts (nor substituted parents) in memory.

    -    This can arise 'fetch.writeCommitGraph' is true. Ordinarily (and
    -    certainly previous to 37b9dcabfc), 'commit_graph_compatible()' would
    -    return 0, indicating that the repository should not generate
    -    commit-graphs (since at one point in the same process it was shallow).
    -    But with 37b9dcabfc, that check succeeds.
    +    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.

    -    This is where the bug occurs: 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. If a commit-graph write proceeds, we will use the incorrect
    -    parentage, producing wrong results.
    +    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.

    -    (Prior to this patch, there were two ways of fixing this: either (1)
    -    set 'fetch.writeCommitGraph' to 'false', or (2) drop the commit-graph
    -    after unshallowing).
    +    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. This can produce a problem when
    +    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, we 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 *before* unshallowing).
    +    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 is sticky, since all subsequent reads after modifying a
    -    commit's parent are unreliable when unshallowing. This patch modifies
    -    the check in 'commit_graph_compatible' to take this bit into account,
    -    and correctly avoid generating commit-graphs in this case.
    +    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>

      ## commit-graph.c ##

 commit-graph.c           |  3 ++-
 commit.c                 |  2 ++
 object.h                 |  1 +
 t/t5537-fetch-shallow.sh | 14 ++++++++++++++
 4 files changed, 19 insertions(+), 1 deletion(-)

Taylor Blau wrote:

This is indeed
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>

Thanks.

Jonathan Nieder <jrnieder@gmail.com> writes:

Thanks, both.

Taylor Blau wrote:

[...]

nit: It wasn't obvious to me on first reading which patch "this patch" refers to --- does it mean 37b9dcabfc or the patch I am reading?

The approach described in Documentation/SubmittingPatches is to treat the commit message kind of like a bug report:

 . start with the problem the change tries to solve, i.e. what is wrong
   with the current code without the change.

 . request the change, justifying the way it solves the problem, i.e.
   why the result with the change is better.

 . then move on to alternate solutions considered but discarded, if any.

Here, I think that would be similar to what you have written:

 Since 37b9dcabfc (shallow.c: [...]), Git knows how to reset
 stat-validity checks for the $GIT_DIR/shallow file, allowing it to
 [etc].  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
 [...].  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.

nit: seems easier to read with s/This can/That would/, since it's describing the road not taken

nit: s/we attach/attach/, to avoid ambiguity with the previous part of the commit message that describes the current, unpatched behavior.

likewise:

 This bit needs to be sticky because [...]

I'm of course inclined to like it, so before reading through the rest my preference would be to include both patches (this fix and the patch I sent to reduce user exposure) in -rc0. But let's see if the patch changes my opinion. ;-)

[...]

A subtlety: we can't *just* check for substituted parents, because there could be a graft affecting a commit we haven't parsed yet. So we'd have to check for both grafts yet to be applied (grafts_nr) and grafts already applied (substituted_parent).

Good.

[...]

This applies right away at the only place in commit parsing where grafts can be discovered.

The only other callers to lookup_commit_graft are get_shallow_commits and prepare_shallow_info. Those are protocol code and they don't affect the in-core history.

[...]

parsed_object_pool is zero-initialized on creation (in parsed_object_pool_new), so this gets correctly initialized to false.

[...]

The indentation is odd here, but it's consistent with the rest of the file.

A fairly straightforward test that demonstrates the bug being fixed and only relies on what was set up in the initial 'setup' test (4 commits). Thanks for tying that loose end.

With or without commit message tweaks along the lines described above,
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>

Thanks for your thoughtful work, as always.

Jonathan Nieder <jrnieder@gmail.com> writes:

Thanks for a nicely written review. It would be very nice if we can see a reroll before I have to tag -rc0.

Thanks.

On Wed, Jul 08, 2020 at 01:55:28PM -0700, Junio C Hamano wrote:

Of course! I just sent one here[1]. Sorry for the last-minute addition to -rc0, and thanks for fast-tracking it down.

Thanks, Taylor

[1]: https://lore.kernel.org/git/1f769bbbb4f1b4ad67d29ee7b3e5282446e4cc0f.1594242582.git.me@ttaylorr.com/