threads / discuss / 50023

2.20.0 - Undocumented change in submodule update wrt # parallel jobs

Subject: 2.20.0 - Undocumented change in submodule update wrt # parallel jobs

## tl;dr

7 messages between Dec 13, 2018 and Dec 14, 2018.

replies: 6people: 5as markdown or json

Sjon Hortensius· Dec 13, 2018, 09:15 UTC · lore

When switching to 2.20 our `git submodule update' (which clones through ssh) broke because our ssh-server rejected the ~20 simultaneous connections the git-client makes. This seems to be caused by a (possibly unintended) change in https://github.com/git/git/commit/90efe595c53f4bb1851371344c35eff71f604d2b which removed the default of max_jobs=1

While this can easily be fixed by configuring submodule.fetchJobs I think this change should be documented - or reverted back to it's previous default of 1

-- Kind regards,

Sjon Hortensius
| Parse Software Development B.V.
| http://www.parse.nl
Ævar Arnfjörð Bjarmason· Dec 13, 2018, 14:02 UTC · re: Sjon Hortensius · lore

Re: 2.20.0 - Undocumented change in submodule update wrt # parallel jobs

On Thu, Dec 13 2018, Sjon Hortensius wrote:
Show 10 quoted lines
> When switching to 2.20 our `git submodule update' (which clones
> through ssh) broke because our ssh-server rejected the ~20
> simultaneous connections the git-client makes. This seems to be caused
> by a (possibly unintended) change in
> https://github.com/git/git/commit/90efe595c53f4bb1851371344c35eff71f604d2b
> which removed the default of max_jobs=1
>
> While this can easily be fixed by configuring submodule.fetchJobs I
> think this change should be documented - or reverted back to it's
> previous default of 1

Just to add to this. SSH-ing with -j<ncores> (which I assume that ~20 is) is going to run into MaxStartups in sshd_config, which is especially annoying as it's pre-ssh-auth, so the server doesn't even get "so and so tried to login", it's just dropped as a potential DoS attack.

Junio C Hamano· Dec 13, 2018, 14:17 UTC · re: Sjon Hortensius · lore

Re: 2.20.0 - Undocumented change in submodule update wrt # parallel jobs

Sjon Hortensius <sjon@parse.nl> writes:
Show 10 quoted lines
> When switching to 2.20 our `git submodule update' (which clones
> through ssh) broke because our ssh-server rejected the ~20
> simultaneous connections the git-client makes. This seems to be caused
> by a (possibly unintended) change in
> https://github.com/git/git/commit/90efe595c53f4bb1851371344c35eff71f604d2b
> which removed the default of max_jobs=1
>
> While this can easily be fixed by configuring submodule.fetchJobs I
> think this change should be documented - or reverted back to it's
> previous default of 1

The commit in question does not look like it _wanted_ to change the default; rather, it appears to me that it wanted to be bug-to-bug compatible with the original, and any such change of behaviour is entirely unintended.

I think the attached may be sufficient to change the default max_jobs back to 1.

By the way, is there a place where we document that the default value for fetchjobs, when unconfigured, is 1? If we are not making such a concrete promise, then I would think it is OK to update the default without any fanfare, as long as we have good reasons to do so. For this particular one, however, as I already said, I do not think we wanted to change the default to unlimited or anything like that, so...

 builtin/submodule--helper.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/builtin/submodule--helper.c b/builtin/submodule--helper.c
index 789d00d87d..e8cdf84f1c 100644
--- a/builtin/submodule--helper.c
+++ b/builtin/submodule--helper.c
@@ -1552,7 +1552,7 @@ struct submodule_update_clone {
 #define SUBMODULE_UPDATE_CLONE_INIT {0, MODULE_LIST_INIT, 0, \
 	SUBMODULE_UPDATE_STRATEGY_INIT, 0, 0, -1, STRING_LIST_INIT_DUP, 0, \
 	NULL, NULL, NULL, \
-	NULL, 0, 0, 0, NULL, 0, 0, 0}
+	NULL, 0, 0, 0, NULL, 0, 0, 1}
 
 
 static void next_submodule_warn_missing(struct submodule_update_clone *suc,
Stefan Beller· Dec 13, 2018, 18:50 UTC · re: Junio C Hamano · lore

Re: 2.20.0 - Undocumented change in submodule update wrt # parallel jobs

On Thu, Dec 13, 2018 at 6:17 AM Junio C Hamano <gitster@pobox.com> wrote:
Show 18 quoted lines
>
> Sjon Hortensius <sjon@parse.nl> writes:
>
> > When switching to 2.20 our `git submodule update' (which clones
> > through ssh) broke because our ssh-server rejected the ~20
> > simultaneous connections the git-client makes. This seems to be caused
> > by a (possibly unintended) change in
> > https://github.com/git/git/commit/90efe595c53f4bb1851371344c35eff71f604d2b
> > which removed the default of max_jobs=1
> >
> > While this can easily be fixed by configuring submodule.fetchJobs I
> > think this change should be documented - or reverted back to it's
> > previous default of 1
>
> The commit in question does not look like it _wanted_ to change the
> default; rather, it appears to me that it wanted to be bug-to-bug
> compatible with the original, and any such change of behaviour is
> entirely unintended.
Indeed.
> I think the attached may be sufficient to change the default
> max_jobs back to 1.
I think so, too. I can wrap it into a commit with a proper message.
>
> By the way, is there a place where we document that the default
> value for fetchjobs, when unconfigured, is 1?
`man git config`
    submodule.fetchJobs
           Specifies how many submodules are fetched/cloned at the
           same time. A positive integer allows up to that number of
           submodules fetched in parallel. A value of 0 will give some
           reasonable default. If unset, it defaults to 1.

and that seems to be the only place, other places only reference this place:

    Documentation$ git grep submodule.fetch
    config/submodule.txt:66:submodule.fetchJobs::
    git-clone.txt:259:      Defaults to the `submodule.fetchJobs` option.
    git-submodule.txt:408:  Defaults to the `submodule.fetchJobs` option.

The behavior of that seems to have been there since the beginning of a028a1930c (fetching submodules: respect `submodule.fetchJobs` config option, 2016-02-29)

Show 6 quoted lines
> If we are not making
> such a concrete promise, then I would think it is OK to update the
> default without any fanfare, as long as we have good reasons to do
> so.  For this particular one, however, as I already said, I do not
> think we wanted to change the default to unlimited or anything like
> that, so...
We definitely want the diff below as a proper patch.
Show 16 quoted lines
>  builtin/submodule--helper.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/builtin/submodule--helper.c b/builtin/submodule--helper.c
> index 789d00d87d..e8cdf84f1c 100644
> --- a/builtin/submodule--helper.c
> +++ b/builtin/submodule--helper.c
> @@ -1552,7 +1552,7 @@ struct submodule_update_clone {
>  #define SUBMODULE_UPDATE_CLONE_INIT {0, MODULE_LIST_INIT, 0, \
>         SUBMODULE_UPDATE_STRATEGY_INIT, 0, 0, -1, STRING_LIST_INIT_DUP, 0, \
>         NULL, NULL, NULL, \
> -       NULL, 0, 0, 0, NULL, 0, 0, 0}
> +       NULL, 0, 0, 0, NULL, 0, 0, 1}
>
>
>  static void next_submodule_warn_missing(struct submodule_update_clone *suc,
Stefan Beller· Dec 13, 2018, 19:02 UTC · re: Stefan Beller · lore

[PATCH] submodule update: run at most one fetch job unless otherwise set

From: Junio C Hamano <gitster@pobox.com>

In a028a1930c (fetching submodules: respect `submodule.fetchJobs` config option, 2016-02-29), we made sure to keep the default behavior of a fetching at most one submodule at once when not setting the newly introduced `submodule.fetchJobs` config.

This regressed in 90efe595c5 (builtin/submodule--helper: factor out submodule updating, 2018-08-03). Fix it.

Reported-by: Sjon Hortensius <sjon@parse.nl>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Signed-off-by: Stefan Beller <sbeller@google.com>
---
 builtin/submodule--helper.c | 2 +-
 t/t5526-fetch-submodules.sh | 2 ++
 2 files changed, 3 insertions(+), 1 deletion(-)
diff --git a/builtin/submodule--helper.c b/builtin/submodule--helper.c
index d38113a31a..1f8a4a9d52 100644
--- a/builtin/submodule--helper.c
+++ b/builtin/submodule--helper.c
@@ -1551,7 +1551,7 @@ struct submodule_update_clone {
 #define SUBMODULE_UPDATE_CLONE_INIT {0, MODULE_LIST_INIT, 0, \
 	SUBMODULE_UPDATE_STRATEGY_INIT, 0, 0, -1, STRING_LIST_INIT_DUP, 0, \
 	NULL, NULL, NULL, \
-	NULL, 0, 0, 0, NULL, 0, 0, 0}
+	NULL, 0, 0, 0, NULL, 0, 0, 1}
 
 
 static void next_submodule_warn_missing(struct submodule_update_clone *suc,
diff --git a/t/t5526-fetch-submodules.sh b/t/t5526-fetch-submodules.sh
index 6c2f9b2ba2..a0317556c6 100755
--- a/t/t5526-fetch-submodules.sh
+++ b/t/t5526-fetch-submodules.sh
@@ -524,6 +524,8 @@ test_expect_success 'fetching submodules respects parallel settings' '
 	git config fetch.recurseSubmodules true &&
 	(
 		cd downstream &&
+		GIT_TRACE=$(pwd)/trace.out git fetch &&
+		grep "1 tasks" trace.out &&
 		GIT_TRACE=$(pwd)/trace.out git fetch --jobs 7 &&
 		grep "7 tasks" trace.out &&
 		git config submodule.fetchJobs 8 &&

-- 
2.20.0.rc2.230.gc28305e538
Eric Sunshine· Dec 13, 2018, 19:04 UTC · re: Stefan Beller · lore

Re: [PATCH] submodule update: run at most one fetch job unless otherwise set

On Thu, Dec 13, 2018 at 2:03 PM Stefan Beller <sbeller@google.com> wrote:
> In a028a1930c (fetching submodules: respect `submodule.fetchJobs`
> config option, 2016-02-29), we made sure to keep the default behavior
> of a fetching at most one submodule at once when not setting the
s/of a/of/
Show 8 quoted lines
> newly introduced `submodule.fetchJobs` config.
>
> This regressed in 90efe595c5 (builtin/submodule--helper: factor
> out submodule updating, 2018-08-03). Fix it.
>
> Reported-by: Sjon Hortensius <sjon@parse.nl>
> Signed-off-by: Junio C Hamano <gitster@pobox.com>
> Signed-off-by: Stefan Beller <sbeller@google.com>
Junio C Hamano· Dec 14, 2018, 02:53 UTC · re: Stefan Beller · lore

Re: [PATCH] submodule update: run at most one fetch job unless otherwise set

Stefan Beller <sbeller@google.com> writes:
Show 14 quoted lines
> From: Junio C Hamano <gitster@pobox.com>
>
> In a028a1930c (fetching submodules: respect `submodule.fetchJobs`
> config option, 2016-02-29), we made sure to keep the default behavior
> of a fetching at most one submodule at once when not setting the
> newly introduced `submodule.fetchJobs` config.
>
> This regressed in 90efe595c5 (builtin/submodule--helper: factor
> out submodule updating, 2018-08-03). Fix it.
>
> Reported-by: Sjon Hortensius <sjon@parse.nl>
> Signed-off-by: Junio C Hamano <gitster@pobox.com>
> Signed-off-by: Stefan Beller <sbeller@google.com>
> ---
Thanks for tying the loose ends.

We may want to convert the _INIT macro to use the designated initializers; I had to count the fields twice to make sure I was tweaking the right one.

It did not help that I saw, before looking at the current code, that 90efe595c5 added one field at the end to the struct but did not touch _INIT macro at all. That made me guess that max_jobs=0 was due to _missing_ initialization value, leading me to an incorrect fix to append an extra 1 at the end, but that was bogus. The missing initialization left by 90efe595 ("builtin/submodule--helper: factor out submodule updating", 2018-08-03) was silently fixed by f1d15713 ("builtin/submodule--helper: store update_clone information in a struct", 2018-08-03), I think, so replacing the 0 at the end is 1 happens to be the right fix, but with designated initializers, all these confusions are more easily avoided.

Show 30 quoted lines
>  builtin/submodule--helper.c | 2 +-
>  t/t5526-fetch-submodules.sh | 2 ++
>  2 files changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/builtin/submodule--helper.c b/builtin/submodule--helper.c
> index d38113a31a..1f8a4a9d52 100644
> --- a/builtin/submodule--helper.c
> +++ b/builtin/submodule--helper.c
> @@ -1551,7 +1551,7 @@ struct submodule_update_clone {
>  #define SUBMODULE_UPDATE_CLONE_INIT {0, MODULE_LIST_INIT, 0, \
>  	SUBMODULE_UPDATE_STRATEGY_INIT, 0, 0, -1, STRING_LIST_INIT_DUP, 0, \
>  	NULL, NULL, NULL, \
> -	NULL, 0, 0, 0, NULL, 0, 0, 0}
> +	NULL, 0, 0, 0, NULL, 0, 0, 1}
>  
>  
>  static void next_submodule_warn_missing(struct submodule_update_clone *suc,
> diff --git a/t/t5526-fetch-submodules.sh b/t/t5526-fetch-submodules.sh
> index 6c2f9b2ba2..a0317556c6 100755
> --- a/t/t5526-fetch-submodules.sh
> +++ b/t/t5526-fetch-submodules.sh
> @@ -524,6 +524,8 @@ test_expect_success 'fetching submodules respects parallel settings' '
>  	git config fetch.recurseSubmodules true &&
>  	(
>  		cd downstream &&
> +		GIT_TRACE=$(pwd)/trace.out git fetch &&
> +		grep "1 tasks" trace.out &&
>  		GIT_TRACE=$(pwd)/trace.out git fetch --jobs 7 &&
>  		grep "7 tasks" trace.out &&
>  		git config submodule.fetchJobs 8 &&

← back to recent threads