Is fetch.writeCommitGraph (and thus features.experimental) meant to work in the presence of shallow clones?

Is fetch.writeCommitGraph (and thus features.experimental) meant to work in the presence of shallow clones?

[Elijah Newren]

Hi,

I was building a version of git for internal use, and thought I'd try
turning on features.experimental to get more testing of it.  The
following test error in the testsuite scared me, though:

t5537.9 (fetch --update-shallow):

...
+ git fetch --update-shallow ../shallow/.git refs/heads/*:refs/remotes/shallow/*
remote: Enumerating objects: 18, done.
remote: Counting objects: 100% (18/18), done.
remote: Compressing objects: 100% (6/6), done.
remote: Total 16 (delta 0), reused 6 (delta 0), pack-reused 0
Unpacking objects: 100% (16/16), 1.16 KiB | 1.17 MiB/s, done.
From ../shallow/
 * [new branch]      master     -> shallow/master
 * [new tag]         heavy-tag  -> heavy-tag
 * [new tag]         light-tag  -> light-tag
error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
fatal: unable to parse commit ac67d3021b4319951fb176469d7732e6914530c5

Passing -c fetch.writeCommitGraph=false to the fetch command in that
test makes it pass.

There were also a couple other tests that failed with
features.experimental=true (in t5500), but those weren't scary -- they
were just checking exact want/have lines and features.experimental is
intended to change those.  This test from t5537 was the only one that
showed some unexpected fatal error.

Thanks,
Elijah

Re: Is fetch.writeCommitGraph (and thus features.experimental) meant to work in the presence of shallow clones?

[Taylor Blau]

On Tue, Apr 14, 2020 at 01:22:45PM -0700, Elijah Newren wrote:
> Hi,
>
> I was building a version of git for internal use, and thought I'd try
> turning on features.experimental to get more testing of it.  The
> following test error in the testsuite scared me, though:
>
> t5537.9 (fetch --update-shallow):
>
> ...
> + git fetch --update-shallow ../shallow/.git refs/heads/*:refs/remotes/shallow/*
> remote: Enumerating objects: 18, done.
> remote: Counting objects: 100% (18/18), done.
> remote: Compressing objects: 100% (6/6), done.
> remote: Total 16 (delta 0), reused 6 (delta 0), pack-reused 0
> Unpacking objects: 100% (16/16), 1.16 KiB | 1.17 MiB/s, done.
> From ../shallow/
>  * [new branch]      master     -> shallow/master
>  * [new tag]         heavy-tag  -> heavy-tag
>  * [new tag]         light-tag  -> light-tag
> error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> fatal: unable to parse commit ac67d3021b4319951fb176469d7732e6914530c5
>
> Passing -c fetch.writeCommitGraph=false to the fetch command in that
> test makes it pass.

This failure makes sense, since taking the reachability closure over
the ref tips will fail in shallow clones of repositories that have more
than one commit.

I wonder if fetch should be taught to avoid generating commit-graphs
(and ignore 'fetch.writeCommitGraph') when in a shallow clone.

> There were also a couple other tests that failed with
> features.experimental=true (in t5500), but those weren't scary -- they
> were just checking exact want/have lines and features.experimental is
> intended to change those.  This test from t5537 was the only one that
> showed some unexpected fatal error.
>
> Thanks,
> Elijah

Thanks,
Taylor

Re: Is fetch.writeCommitGraph (and thus features.experimental) meant to work in the presence of shallow clones?

[Derrick Stolee]

On 4/14/2020 4:22 PM, Elijah Newren wrote:
> Hi,
> 
> I was building a version of git for internal use, and thought I'd try
> turning on features.experimental to get more testing of it.  The
> following test error in the testsuite scared me, though:
> 
> t5537.9 (fetch --update-shallow):
> 
> ...
> + git fetch --update-shallow ../shallow/.git refs/heads/*:refs/remotes/shallow/*
> remote: Enumerating objects: 18, done.
> remote: Counting objects: 100% (18/18), done.
> remote: Compressing objects: 100% (6/6), done.
> remote: Total 16 (delta 0), reused 6 (delta 0), pack-reused 0
> Unpacking objects: 100% (16/16), 1.16 KiB | 1.17 MiB/s, done.
> From ../shallow/
>  * [new branch]      master     -> shallow/master
>  * [new tag]         heavy-tag  -> heavy-tag
>  * [new tag]         light-tag  -> light-tag
> error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> fatal: unable to parse commit ac67d3021b4319951fb176469d7732e6914530c5
> 
> Passing -c fetch.writeCommitGraph=false to the fetch command in that
> test makes it pass.
> 
> There were also a couple other tests that failed with
> features.experimental=true (in t5500), but those weren't scary -- they
> were just checking exact want/have lines and features.experimental is
> intended to change those.  This test from t5537 was the only one that
> showed some unexpected fatal error.

Well, commit-graphs are not supposed to do anything if we have
shallow clones. We definitely don't load a commit-graph in that
case. Seems like we need an extra check in write_commit_graph()
to stop writing in the presence of shallow commits.

-Stolee

Re: Is fetch.writeCommitGraph (and thus features.experimental) meant to work in the presence of shallow clones?

[Taylor Blau]

On Tue, Apr 14, 2020 at 04:31:19PM -0400, Derrick Stolee wrote:
> On 4/14/2020 4:22 PM, Elijah Newren wrote:
> > Hi,
> >
> > I was building a version of git for internal use, and thought I'd try
> > turning on features.experimental to get more testing of it.  The
> > following test error in the testsuite scared me, though:
> >
> > t5537.9 (fetch --update-shallow):
> >
> > ...
> > + git fetch --update-shallow ../shallow/.git refs/heads/*:refs/remotes/shallow/*
> > remote: Enumerating objects: 18, done.
> > remote: Counting objects: 100% (18/18), done.
> > remote: Compressing objects: 100% (6/6), done.
> > remote: Total 16 (delta 0), reused 6 (delta 0), pack-reused 0
> > Unpacking objects: 100% (16/16), 1.16 KiB | 1.17 MiB/s, done.
> > From ../shallow/
> >  * [new branch]      master     -> shallow/master
> >  * [new tag]         heavy-tag  -> heavy-tag
> >  * [new tag]         light-tag  -> light-tag
> > error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> > error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> > error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> > fatal: unable to parse commit ac67d3021b4319951fb176469d7732e6914530c5
> >
> > Passing -c fetch.writeCommitGraph=false to the fetch command in that
> > test makes it pass.
> >
> > There were also a couple other tests that failed with
> > features.experimental=true (in t5500), but those weren't scary -- they
> > were just checking exact want/have lines and features.experimental is
> > intended to change those.  This test from t5537 was the only one that
> > showed some unexpected fatal error.
>
> Well, commit-graphs are not supposed to do anything if we have
> shallow clones. We definitely don't load a commit-graph in that
> case. Seems like we need an extra check in write_commit_graph()
> to stop writing in the presence of shallow commits.

This rang a bell to me, too. There's a bug, but it's due to the mutative
side-effects of 'is_repository_shallow' along with '--update-shallow' (a
normal 'git fetch' works fine here, with or without
fetch.writeCommitGraph).

Here's a patch that I didn't sign-off on that fixes the problem for me.

--- >8 ---

Subject: [PATCH] shallow.c: use 'reset_repository_shallow' when appropriate

In bd0b42aed3 (fetch-pack: do not take shallow lock unnecessarily,
2019-01-10), the author noted that 'is_repository_shallow' produces
visible side-effect(s) by setting 'is_shallow' and 'shallow_stat'.

This is a problem for e.g., fetching with '--update-shallow' in a
shallow repsoitory with 'fetch.writeCommitGraph' enabled, since the
update to '.git/shallow' will cause Git to think that the repository
*isn't* shallow when it is, thereby circumventing the commit-graph
compatability check.

This causes problems in shallow repositories with at least shallow refs
that have at least one ancestor (since the client won't have those
object(s), and therefore can't take the reachability closure over
commits to be written to the commit-graph).

Address this by introducing 'reset_repository_shallow()', and calling it
when the shallow file is updated, forcing 'is_repository_shallow' to
re-evaluate whether the repository is still shallow after fetching in
the above scenario.

Signed-off-by: Taylor Blau <me@ttaylorr.com>
---
 commit.h     |  1 +
 fetch-pack.c |  1 +
 shallow.c    | 15 ++++++++-------
 3 files changed, 10 insertions(+), 7 deletions(-)

diff --git a/commit.h b/commit.h
index 008a0fa4a0..ee1ba139d4 100644
--- a/commit.h
+++ b/commit.h
@@ -251,6 +251,7 @@ int register_shallow(struct repository *r, const struct object_id *oid);
 int unregister_shallow(const struct object_id *oid);
 int for_each_commit_graft(each_commit_graft_fn, void *);
 int is_repository_shallow(struct repository *r);
+void reset_repository_shallow(struct repository *r);
 struct commit_list *get_shallow_commits(struct object_array *heads,
 					int depth, int shallow_flag, int not_shallow_flag);
 struct commit_list *get_shallow_commits_by_rev_list(
diff --git a/fetch-pack.c b/fetch-pack.c
index 1734a573b0..051902ef6d 100644
--- a/fetch-pack.c
+++ b/fetch-pack.c
@@ -1630,6 +1630,7 @@ static void update_shallow(struct fetch_pack_args *args,
 		if (*alternate_shallow_file == '\0') { /* --unshallow */
 			unlink_or_warn(git_path_shallow(the_repository));
 			rollback_lock_file(&shallow_lock);
+			reset_repository_shallow(the_repository);
 		} else
 			commit_lock_file(&shallow_lock);
 		alternate_shallow_file = NULL;
diff --git a/shallow.c b/shallow.c
index 7fd04afed1..fac383dec9 100644
--- a/shallow.c
+++ b/shallow.c
@@ -40,13 +40,6 @@ int register_shallow(struct repository *r, const struct object_id *oid)

 int is_repository_shallow(struct repository *r)
 {
-	/*
-	 * NEEDSWORK: This function updates
-	 * r->parsed_objects->{is_shallow,shallow_stat} as a side effect but
-	 * there is no corresponding function to clear them when the shallow
-	 * file is updated.
-	 */
-
 	FILE *fp;
 	char buf[1024];
 	const char *path = r->parsed_objects->alternate_shallow_file;
@@ -79,6 +72,12 @@ int is_repository_shallow(struct repository *r)
 	return r->parsed_objects->is_shallow;
 }

+void reset_repository_shallow(struct repository *r)
+{
+	r->parsed_objects->is_shallow = -1;
+	stat_validity_clear(r->parsed_objects->shallow_stat);
+}
+
 /*
  * TODO: use "int" elemtype instead of "int *" when/if commit-slab
  * supports a "valid" flag.
@@ -362,6 +361,7 @@ void setup_alternate_shallow(struct lock_file *shallow_lock,
 		 * shallow file".
 		 */
 		*alternate_shallow_file = "";
+	reset_repository_shallow(the_repository);
 	strbuf_release(&sb);
 }

@@ -411,6 +411,7 @@ void prune_shallow(unsigned options)
 			die_errno("failed to write to %s",
 				  get_lock_file_path(&shallow_lock));
 		commit_lock_file(&shallow_lock);
+		reset_repository_shallow(the_repository);
 	} else {
 		unlink(git_path_shallow(the_repository));
 		rollback_lock_file(&shallow_lock);
--
2.26.0.106.g9fadedd637

Re: Is fetch.writeCommitGraph (and thus features.experimental) meant to work in the presence of shallow clones?

[Taylor Blau]

On Tue, Apr 14, 2020 at 05:50:57PM -0600, Taylor Blau wrote:
> On Tue, Apr 14, 2020 at 04:31:19PM -0400, Derrick Stolee wrote:
> > On 4/14/2020 4:22 PM, Elijah Newren wrote:
> > > Hi,
> > >
> > > I was building a version of git for internal use, and thought I'd try
> > > turning on features.experimental to get more testing of it.  The
> > > following test error in the testsuite scared me, though:
> > >
> > > t5537.9 (fetch --update-shallow):
> > >
> > > ...
> > > + git fetch --update-shallow ../shallow/.git refs/heads/*:refs/remotes/shallow/*
> > > remote: Enumerating objects: 18, done.
> > > remote: Counting objects: 100% (18/18), done.
> > > remote: Compressing objects: 100% (6/6), done.
> > > remote: Total 16 (delta 0), reused 6 (delta 0), pack-reused 0
> > > Unpacking objects: 100% (16/16), 1.16 KiB | 1.17 MiB/s, done.
> > > From ../shallow/
> > >  * [new branch]      master     -> shallow/master
> > >  * [new tag]         heavy-tag  -> heavy-tag
> > >  * [new tag]         light-tag  -> light-tag
> > > error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> > > error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> > > error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> > > fatal: unable to parse commit ac67d3021b4319951fb176469d7732e6914530c5
> > >
> > > Passing -c fetch.writeCommitGraph=false to the fetch command in that
> > > test makes it pass.
> > >
> > > There were also a couple other tests that failed with
> > > features.experimental=true (in t5500), but those weren't scary -- they
> > > were just checking exact want/have lines and features.experimental is
> > > intended to change those.  This test from t5537 was the only one that
> > > showed some unexpected fatal error.
> >
> > Well, commit-graphs are not supposed to do anything if we have
> > shallow clones. We definitely don't load a commit-graph in that
> > case. Seems like we need an extra check in write_commit_graph()
> > to stop writing in the presence of shallow commits.
>
> This rang a bell to me, too. There's a bug, but it's due to the mutative
> side-effects of 'is_repository_shallow' along with '--update-shallow' (a
> normal 'git fetch' works fine here, with or without
> fetch.writeCommitGraph).
>
> Here's a patch that I didn't sign-off on that fixes the problem for me.

Oh, apparently I did sign this one off ;). I'll blame that 'git commit
--amend -vs' is muscle memory for me. In either case, this probably
needs some work (if it's even the right approach) before queueing it.

> --- >8 ---
>
> ...

Thanks,
Taylor

Re: Is fetch.writeCommitGraph (and thus features.experimental) meant to work in the presence of shallow clones?

[Derrick Stolee]

On 4/14/2020 7:50 PM, Taylor Blau wrote:
> On Tue, Apr 14, 2020 at 04:31:19PM -0400, Derrick Stolee wrote:
>> On 4/14/2020 4:22 PM, Elijah Newren wrote:
>>> Hi,
>>>
>>> I was building a version of git for internal use, and thought I'd try
>>> turning on features.experimental to get more testing of it.  The
>>> following test error in the testsuite scared me, though:
>>>
>>> t5537.9 (fetch --update-shallow):
>>>
>>> ...
>>> + git fetch --update-shallow ../shallow/.git refs/heads/*:refs/remotes/shallow/*
[snip]
>> Well, commit-graphs are not supposed to do anything if we have
>> shallow clones. We definitely don't load a commit-graph in that
>> case. Seems like we need an extra check in write_commit_graph()
>> to stop writing in the presence of shallow commits.

Here, I assumed that the commit-graph wasn't checking the shallow status
appropriately, but...

> This rang a bell to me, too. There's a bug, but it's due to the mutative
> side-effects of 'is_repository_shallow' along with '--update-shallow' (a
> normal 'git fetch' works fine here, with or without
> fetch.writeCommitGraph).

...this makes sense as to why this particular case failed.

(Please allow me this brief moment to communicate my extreme dislike
of shallow clones. There, I'm done.)
 
> Here's a patch that I didn't sign-off on that fixes the problem for me.
> 
> --- >8 ---
> 
> Subject: [PATCH] shallow.c: use 'reset_repository_shallow' when appropriate
> 
> In bd0b42aed3 (fetch-pack: do not take shallow lock unnecessarily,
> 2019-01-10), the author noted that 'is_repository_shallow' produces
> visible side-effect(s) by setting 'is_shallow' and 'shallow_stat'.
> 
> This is a problem for e.g., fetching with '--update-shallow' in a
> shallow repsoitory with 'fetch.writeCommitGraph' enabled, since the

repository

> update to '.git/shallow' will cause Git to think that the repository
> *isn't* shallow when it is, thereby circumventing the commit-graph
> compatability check.

compatibility

> This causes problems in shallow repositories with at least shallow refs
> that have at least one ancestor (since the client won't have those
> object(s), and therefore can't take the reachability closure over
> commits to be written to the commit-graph).
> 
> Address this by introducing 'reset_repository_shallow()', and calling it
> when the shallow file is updated, forcing 'is_repository_shallow' to
> re-evaluate whether the repository is still shallow after fetching in
> the above scenario.
> 
> Signed-off-by: Taylor Blau <me@ttaylorr.com>
> ---
>  commit.h     |  1 +
>  fetch-pack.c |  1 +
>  shallow.c    | 15 ++++++++-------
>  3 files changed, 10 insertions(+), 7 deletions(-)
> 
> diff --git a/commit.h b/commit.h
> index 008a0fa4a0..ee1ba139d4 100644
> --- a/commit.h
> +++ b/commit.h
> @@ -251,6 +251,7 @@ int register_shallow(struct repository *r, const struct object_id *oid);
>  int unregister_shallow(const struct object_id *oid);
>  int for_each_commit_graft(each_commit_graft_fn, void *);
>  int is_repository_shallow(struct repository *r);
> +void reset_repository_shallow(struct repository *r);
>  struct commit_list *get_shallow_commits(struct object_array *heads,
>  					int depth, int shallow_flag, int not_shallow_flag);
>  struct commit_list *get_shallow_commits_by_rev_list(
> diff --git a/fetch-pack.c b/fetch-pack.c
> index 1734a573b0..051902ef6d 100644
> --- a/fetch-pack.c
> +++ b/fetch-pack.c
> @@ -1630,6 +1630,7 @@ static void update_shallow(struct fetch_pack_args *args,
>  		if (*alternate_shallow_file == '\0') { /* --unshallow */
>  			unlink_or_warn(git_path_shallow(the_repository));
>  			rollback_lock_file(&shallow_lock);
> +			reset_repository_shallow(the_repository);
>  		} else
>  			commit_lock_file(&shallow_lock);
>  		alternate_shallow_file = NULL;
> diff --git a/shallow.c b/shallow.c
> index 7fd04afed1..fac383dec9 100644
> --- a/shallow.c
> +++ b/shallow.c
> @@ -40,13 +40,6 @@ int register_shallow(struct repository *r, const struct object_id *oid)
> 
>  int is_repository_shallow(struct repository *r)
>  {
> -	/*
> -	 * NEEDSWORK: This function updates
> -	 * r->parsed_objects->{is_shallow,shallow_stat} as a side effect but
> -	 * there is no corresponding function to clear them when the shallow
> -	 * file is updated.
> -	 */
> -

It's always good to remove these NEEDSWORK comments. I wonder if the
problem is more complicated than your patch makes it seem, or else
the original author would have done it correctly at first.

But you are definitely closing out one dangling side-effect, which is
an improvement.

>  	FILE *fp;
>  	char buf[1024];
>  	const char *path = r->parsed_objects->alternate_shallow_file;
> @@ -79,6 +72,12 @@ int is_repository_shallow(struct repository *r)
>  	return r->parsed_objects->is_shallow;
>  }
> 
> +void reset_repository_shallow(struct repository *r)
> +{
> +	r->parsed_objects->is_shallow = -1;
> +	stat_validity_clear(r->parsed_objects->shallow_stat);
> +}
> +
>  /*
>   * TODO: use "int" elemtype instead of "int *" when/if commit-slab
>   * supports a "valid" flag.
> @@ -362,6 +361,7 @@ void setup_alternate_shallow(struct lock_file *shallow_lock,
>  		 * shallow file".
>  		 */
>  		*alternate_shallow_file = "";
> +	reset_repository_shallow(the_repository);
>  	strbuf_release(&sb);
>  }
> 
> @@ -411,6 +411,7 @@ void prune_shallow(unsigned options)
>  			die_errno("failed to write to %s",
>  				  get_lock_file_path(&shallow_lock));
>  		commit_lock_file(&shallow_lock);
> +		reset_repository_shallow(the_repository);
>  	} else {

These are likely good places to call reset_repository_shallow(),
but perhaps we should also call it here in commit-graph.c?

 static int commit_graph_compatible(struct repository *r)
 {
        if (!r->gitdir)
                return 0;
 
        if (read_replace_refs) {
                prepare_replace_object(r);
                if (hashmap_get_size(&r->objects->replace_map->map))
                        return 0;
        }
 
        prepare_commit_graft(r);
        if (r->parsed_objects && r->parsed_objects->grafts_nr)
                return 0;
+       reset_repository_shallow(r);
        if (is_repository_shallow(r))
                return 0;
 
        return 1;
 }

Or am I misunderstanding the state that reset_repository_shallow()
puts us in? My expectation is that is_repository_shallow() will
act as if the shallow variables had never been set and will look
for shallow data from disk.

Thanks,
-Stolee

Re: Is fetch.writeCommitGraph (and thus features.experimental) meant to work in the presence of shallow clones?

[Taylor Blau]

On Wed, Apr 15, 2020 at 07:55:19AM -0400, Derrick Stolee wrote:
> On 4/14/2020 7:50 PM, Taylor Blau wrote:
> > On Tue, Apr 14, 2020 at 04:31:19PM -0400, Derrick Stolee wrote:
> >> On 4/14/2020 4:22 PM, Elijah Newren wrote:
> >>> Hi,
> >>>
> >>> I was building a version of git for internal use, and thought I'd try
> >>> turning on features.experimental to get more testing of it.  The
> >>> following test error in the testsuite scared me, though:
> >>>
> >>> t5537.9 (fetch --update-shallow):
> >>>
> >>> ...
> >>> + git fetch --update-shallow ../shallow/.git refs/heads/*:refs/remotes/shallow/*
> [snip]
> >> Well, commit-graphs are not supposed to do anything if we have
> >> shallow clones. We definitely don't load a commit-graph in that
> >> case. Seems like we need an extra check in write_commit_graph()
> >> to stop writing in the presence of shallow commits.
>
> Here, I assumed that the commit-graph wasn't checking the shallow status
> appropriately, but...
>
> > This rang a bell to me, too. There's a bug, but it's due to the mutative
> > side-effects of 'is_repository_shallow' along with '--update-shallow' (a
> > normal 'git fetch' works fine here, with or without
> > fetch.writeCommitGraph).
>
> ...this makes sense as to why this particular case failed.
>
> (Please allow me this brief moment to communicate my extreme dislike
> of shallow clones. There, I'm done.)

Noted ;).

> > Here's a patch that I didn't sign-off on that fixes the problem for me.
> >
> > --- >8 ---
> >
> > Subject: [PATCH] shallow.c: use 'reset_repository_shallow' when appropriate
> >
> > In bd0b42aed3 (fetch-pack: do not take shallow lock unnecessarily,
> > 2019-01-10), the author noted that 'is_repository_shallow' produces
> > visible side-effect(s) by setting 'is_shallow' and 'shallow_stat'.
> >
> > This is a problem for e.g., fetching with '--update-shallow' in a
> > shallow repsoitory with 'fetch.writeCommitGraph' enabled, since the
>
> repository
>
> > update to '.git/shallow' will cause Git to think that the repository
> > *isn't* shallow when it is, thereby circumventing the commit-graph
> > compatability check.
>
> compatibility

Whoops, thanks for pointing these out.

> > This causes problems in shallow repositories with at least shallow refs
> > that have at least one ancestor (since the client won't have those
> > object(s), and therefore can't take the reachability closure over
> > commits to be written to the commit-graph).
> >
> > Address this by introducing 'reset_repository_shallow()', and calling it
> > when the shallow file is updated, forcing 'is_repository_shallow' to
> > re-evaluate whether the repository is still shallow after fetching in
> > the above scenario.
> >
> > Signed-off-by: Taylor Blau <me@ttaylorr.com>
> > ---
> >  commit.h     |  1 +
> >  fetch-pack.c |  1 +
> >  shallow.c    | 15 ++++++++-------
> >  3 files changed, 10 insertions(+), 7 deletions(-)
> >
> > diff --git a/commit.h b/commit.h
> > index 008a0fa4a0..ee1ba139d4 100644
> > --- a/commit.h
> > +++ b/commit.h
> > @@ -251,6 +251,7 @@ int register_shallow(struct repository *r, const struct object_id *oid);
> >  int unregister_shallow(const struct object_id *oid);
> >  int for_each_commit_graft(each_commit_graft_fn, void *);
> >  int is_repository_shallow(struct repository *r);
> > +void reset_repository_shallow(struct repository *r);
> >  struct commit_list *get_shallow_commits(struct object_array *heads,
> >  					int depth, int shallow_flag, int not_shallow_flag);
> >  struct commit_list *get_shallow_commits_by_rev_list(
> > diff --git a/fetch-pack.c b/fetch-pack.c
> > index 1734a573b0..051902ef6d 100644
> > --- a/fetch-pack.c
> > +++ b/fetch-pack.c
> > @@ -1630,6 +1630,7 @@ static void update_shallow(struct fetch_pack_args *args,
> >  		if (*alternate_shallow_file == '\0') { /* --unshallow */
> >  			unlink_or_warn(git_path_shallow(the_repository));
> >  			rollback_lock_file(&shallow_lock);
> > +			reset_repository_shallow(the_repository);
> >  		} else
> >  			commit_lock_file(&shallow_lock);
> >  		alternate_shallow_file = NULL;
> > diff --git a/shallow.c b/shallow.c
> > index 7fd04afed1..fac383dec9 100644
> > --- a/shallow.c
> > +++ b/shallow.c
> > @@ -40,13 +40,6 @@ int register_shallow(struct repository *r, const struct object_id *oid)
> >
> >  int is_repository_shallow(struct repository *r)
> >  {
> > -	/*
> > -	 * NEEDSWORK: This function updates
> > -	 * r->parsed_objects->{is_shallow,shallow_stat} as a side effect but
> > -	 * there is no corresponding function to clear them when the shallow
> > -	 * file is updated.
> > -	 */
> > -
>
> It's always good to remove these NEEDSWORK comments. I wonder if the
> problem is more complicated than your patch makes it seem, or else
> the original author would have done it correctly at first.
>
> But you are definitely closing out one dangling side-effect, which is
> an improvement.

Yeah, I have no idea if there are other spots that I'm missing (I only
spent a few minutes on this patch before leaving for the day to go eat
dinner), hence why I CC'd Jonathan Tan to see if he had anything else in
mind when he wrote this 'NEEDSWORK' comment.

If he feels that this is a generally good direction, I'm more than happy
to look for other spots myself, address them, and then submit this as a
real patch.

> >  	FILE *fp;
> >  	char buf[1024];
> >  	const char *path = r->parsed_objects->alternate_shallow_file;
> > @@ -79,6 +72,12 @@ int is_repository_shallow(struct repository *r)
> >  	return r->parsed_objects->is_shallow;
> >  }
> >
> > +void reset_repository_shallow(struct repository *r)
> > +{
> > +	r->parsed_objects->is_shallow = -1;
> > +	stat_validity_clear(r->parsed_objects->shallow_stat);
> > +}
> > +
> >  /*
> >   * TODO: use "int" elemtype instead of "int *" when/if commit-slab
> >   * supports a "valid" flag.
> > @@ -362,6 +361,7 @@ void setup_alternate_shallow(struct lock_file *shallow_lock,
> >  		 * shallow file".
> >  		 */
> >  		*alternate_shallow_file = "";
> > +	reset_repository_shallow(the_repository);
> >  	strbuf_release(&sb);
> >  }
> >
> > @@ -411,6 +411,7 @@ void prune_shallow(unsigned options)
> >  			die_errno("failed to write to %s",
> >  				  get_lock_file_path(&shallow_lock));
> >  		commit_lock_file(&shallow_lock);
> > +		reset_repository_shallow(the_repository);
> >  	} else {
>
> These are likely good places to call reset_repository_shallow(),
> but perhaps we should also call it here in commit-graph.c?
>
>  static int commit_graph_compatible(struct repository *r)
>  {
>         if (!r->gitdir)
>                 return 0;
>
>         if (read_replace_refs) {
>                 prepare_replace_object(r);
>                 if (hashmap_get_size(&r->objects->replace_map->map))
>                         return 0;
>         }
>
>         prepare_commit_graft(r);
>         if (r->parsed_objects && r->parsed_objects->grafts_nr)
>                 return 0;
> +       reset_repository_shallow(r);
>         if (is_repository_shallow(r))
>                 return 0;
>
>         return 1;
>  }
>
> Or am I misunderstanding the state that reset_repository_shallow()
> puts us in? My expectation is that is_repository_shallow() will
> act as if the shallow variables had never been set and will look
> for shallow data from disk.

You're not misunderstanding it at all. I don't think that this location
is strictly necessary, since all of the other locations that change
'.git/shallow' are already invaliding the shallow-ness checks, so by the
time we get to this point 'is_repository_shallow' has to take the slow
path and redetermine whether or not we are actually shallow.

So, I don't think this location is strictly necessary, and it's
potentially slowing us down a little bit, but it is hardening us against
other parts of the code that may not call 'reset_repository_shallow'
when they should have.

> Thanks,
> -Stolee

Thanks for the review.

Thanks,
Taylor

Re: Is fetch.writeCommitGraph (and thus features.experimental) meant to work in the presence of shallow clones?

[Elijah Newren]

On Wed, Apr 15, 2020 at 4:55 AM Derrick Stolee <stolee@gmail.com> wrote:

> (Please allow me this brief moment to communicate my extreme dislike
> of shallow clones. There, I'm done.)

I share your extreme dislike, but I've merely succeeded in
discouraging some people from using them, not in stopping their use
altogether.


(I don't have anything to add to the review of the patches, other than
I'm happy to see folks trying to address the issue.)

Re: Is fetch.writeCommitGraph (and thus features.experimental) meant to work in the presence of shallow clones?

[Jonathan Nieder]

Hi,

Elijah Newren wrote:

> I was building a version of git for internal use, and thought I'd try
> turning on features.experimental to get more testing of it.  The
> following test error in the testsuite scared me, though:
>
> t5537.9 (fetch --update-shallow):
>
> ...
> + git fetch --update-shallow ../shallow/.git refs/heads/*:refs/remotes/shallow/*
> remote: Enumerating objects: 18, done.
> remote: Counting objects: 100% (18/18), done.
> remote: Compressing objects: 100% (6/6), done.
> remote: Total 16 (delta 0), reused 6 (delta 0), pack-reused 0
> Unpacking objects: 100% (16/16), 1.16 KiB | 1.17 MiB/s, done.
> From ../shallow/
>  * [new branch]      master     -> shallow/master
>  * [new tag]         heavy-tag  -> heavy-tag
>  * [new tag]         light-tag  -> light-tag
> error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> fatal: unable to parse commit ac67d3021b4319951fb176469d7732e6914530c5
>
> Passing -c fetch.writeCommitGraph=false to the fetch command in that
> test makes it pass.

Oh!  Thanks for checking this.  At $DAYJOB this was the week we were
going to roll out features.experimental.  Time to roll that back...

How did you go about the experiment?  Does Taylor's patch make it pass?

(I'm thinking it would be nice to have a
GIT_TEST_EXPERIMENTAL_FEATURES setting that uses test_config_global at
test init time.)

Thanks,
Jonathan

Re: Is fetch.writeCommitGraph (and thus features.experimental) meant to work in the presence of shallow clones?

[Elijah Newren]

On Wed, Apr 15, 2020 at 1:54 PM Jonathan Nieder <jrnieder@gmail.com> wrote:
>
> Hi,
>
> Elijah Newren wrote:
>
> > I was building a version of git for internal use, and thought I'd try
> > turning on features.experimental to get more testing of it.  The
> > following test error in the testsuite scared me, though:
> >
> > t5537.9 (fetch --update-shallow):
> >
> > ...
> > + git fetch --update-shallow ../shallow/.git refs/heads/*:refs/remotes/shallow/*
> > remote: Enumerating objects: 18, done.
> > remote: Counting objects: 100% (18/18), done.
> > remote: Compressing objects: 100% (6/6), done.
> > remote: Total 16 (delta 0), reused 6 (delta 0), pack-reused 0
> > Unpacking objects: 100% (16/16), 1.16 KiB | 1.17 MiB/s, done.
> > From ../shallow/
> >  * [new branch]      master     -> shallow/master
> >  * [new tag]         heavy-tag  -> heavy-tag
> >  * [new tag]         light-tag  -> light-tag
> > error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> > error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> > error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> > fatal: unable to parse commit ac67d3021b4319951fb176469d7732e6914530c5
> >
> > Passing -c fetch.writeCommitGraph=false to the fetch command in that
> > test makes it pass.
>
> Oh!  Thanks for checking this.  At $DAYJOB this was the week we were
> going to roll out features.experimental.  Time to roll that back...
>
> How did you go about the experiment?  Does Taylor's patch make it pass?

Yes, Taylor's patch makes the experiment pass.  My experiment was
pretty simple; modify the code so that features.experimental defaults
to true:

diff --git a/repo-settings.c b/repo-settings.c
index a703e407a3..b39bc21b99 100644
--- a/repo-settings.c
+++ b/repo-settings.c
@@ -8,6 +8,7 @@ void prepare_repo_settings(struct repository *r)
 {
        int value;
        char *strval;
+       int feature_experimental;

        if (r->settings.initialized)
                return;
@@ -51,7 +54,10 @@ void prepare_repo_settings(struct repository *r)
        }
        if (!repo_config_get_bool(r, "fetch.writecommitgraph", &value))
                r->settings.fetch_write_commit_graph = value;
-       if (!repo_config_get_bool(r, "feature.experimental", &value) && value) {
+       feature_experimental = 1;
+       if (!repo_config_get_bool(r, "feature.experimental", &value))
+               feature_experimental = value;
+       if (feature_experimental) {
                UPDATE_DEFAULT_BOOL(r->settings.pack_use_sparse, 1);

UPDATE_DEFAULT_BOOL(r->settings.fetch_negotiation_algorithm,
FETCH_NEGOTIATION_SKIPPING);
                UPDATE_DEFAULT_BOOL(r->settings.fetch_write_commit_graph, 1);

(based on 2.26.0; Stolee has since made pack_use_sparse default to
true anyway) and then run the test suite and see what breaks.  For the
tests that fail, check if the failure is expected based on different
defaults and the code still functions correctly with the new defaults
(and update the test manually if so), or whether the test failure
represents some potentially scary problem.  This was the only bug
exposed by this little experiment.

> (I'm thinking it would be nice to have a
> GIT_TEST_EXPERIMENTAL_FEATURES setting that uses test_config_global at
> test init time.)

Re: Is fetch.writeCommitGraph (and thus features.experimental) meant to work in the presence of shallow clones?

[Taylor Blau]

On Wed, Apr 15, 2020 at 03:54:09PM -0700, Elijah Newren wrote:
> On Wed, Apr 15, 2020 at 1:54 PM Jonathan Nieder <jrnieder@gmail.com> wrote:
> >
> > Hi,
> >
> > Elijah Newren wrote:
> >
> > > I was building a version of git for internal use, and thought I'd try
> > > turning on features.experimental to get more testing of it.  The
> > > following test error in the testsuite scared me, though:
> > >
> > > t5537.9 (fetch --update-shallow):
> > >
> > > ...
> > > + git fetch --update-shallow ../shallow/.git refs/heads/*:refs/remotes/shallow/*
> > > remote: Enumerating objects: 18, done.
> > > remote: Counting objects: 100% (18/18), done.
> > > remote: Compressing objects: 100% (6/6), done.
> > > remote: Total 16 (delta 0), reused 6 (delta 0), pack-reused 0
> > > Unpacking objects: 100% (16/16), 1.16 KiB | 1.17 MiB/s, done.
> > > From ../shallow/
> > >  * [new branch]      master     -> shallow/master
> > >  * [new tag]         heavy-tag  -> heavy-tag
> > >  * [new tag]         light-tag  -> light-tag
> > > error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> > > error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> > > error: Could not read ac67d3021b4319951fb176469d7732e6914530c5
> > > fatal: unable to parse commit ac67d3021b4319951fb176469d7732e6914530c5
> > >
> > > Passing -c fetch.writeCommitGraph=false to the fetch command in that
> > > test makes it pass.
> >
> > Oh!  Thanks for checking this.  At $DAYJOB this was the week we were
> > going to roll out features.experimental.  Time to roll that back...
> >
> > How did you go about the experiment?  Does Taylor's patch make it pass?
>
> Yes, Taylor's patch makes the experiment pass.  My experiment was
> pretty simple; modify the code so that features.experimental defaults
> to true:

I'm glad to hear that it is passing now. I would like to have Jonathan
Tan chime in on whether or not this patch is making sense, but if so,
I'll prepare it for real and send it out.

Thanks,
Taylor

Re: Is fetch.writeCommitGraph (and thus features.experimental) meant to work in the presence of shallow clones?

[Jonathan Tan]

> Here's a patch that I didn't sign-off on that fixes the problem for me.
> 
> --- >8 ---
> 
> Subject: [PATCH] shallow.c: use 'reset_repository_shallow' when appropriate
> 
> In bd0b42aed3 (fetch-pack: do not take shallow lock unnecessarily,
> 2019-01-10), the author noted that 'is_repository_shallow' produces
> visible side-effect(s) by setting 'is_shallow' and 'shallow_stat'.

Thanks for the reference to my commit. When I wrote that, if I remember
correctly, I found it difficult to be able to guarantee that the
clearing of is_shallow and shallow_stat was performed upon every
commit-lock (which creates a new shallow file) and
<unlink+rollback-lock> (which removes the shallow file, potentially
changing the repo state from shallow to none), so I added the NEEDSWORK
instead. I see that this patch makes some progress in solving that.

> @@ -1630,6 +1630,7 @@ static void update_shallow(struct fetch_pack_args *args,
>  		if (*alternate_shallow_file == '\0') { /* --unshallow */
>  			unlink_or_warn(git_path_shallow(the_repository));
>  			rollback_lock_file(&shallow_lock);
> +			reset_repository_shallow(the_repository);
>  		} else
>  			commit_lock_file(&shallow_lock);
>  		alternate_shallow_file = NULL;

Here, do you need a reset in the "else" branch as well? (I.e. in both
branches, so you can put it outside the "if" block.) I didn't look at it
too closely, but I would think that when the shallow_lock file is
committed, there might have been no shallow beforehand, changing the
state from no-shallow to shallow.

> @@ -411,6 +411,7 @@ void prune_shallow(unsigned options)
>  			die_errno("failed to write to %s",
>  				  get_lock_file_path(&shallow_lock));
>  		commit_lock_file(&shallow_lock);
> +		reset_repository_shallow(the_repository);
>  	} else {
>  		unlink(git_path_shallow(the_repository));
>  		rollback_lock_file(&shallow_lock);

And the opposite case here - reset in the "else" branch because the
state could have changed from shallow to no-shallow.

In any case, I think the commit message should discuss why
reset_repository_shallow() is added only on the unlink+rollback side in
one "if" statement, but only on the opposite "commit" side in another
"if" statement.