Re: [PATCH 5/6] submodule: improve submodule_has_commits

[Stefan Beller <sbeller@google.com>]

On Fri, Apr 28, 2017 at 4:54 PM, Brandon Williams <bmwill@google.com> wrote:
> Teach 'submodule_has_commits()' to ensure that if a commit exists in a
> submodule, that it is also reachable from a ref.
>
> This is a prepritory step prior to merging the logic which checkes for

s/prepritory/preparatory/
s/checkes/checks/

This is the first commit in the series that changes user observable behavior,
I guess there will be tests in a later patch? Can you elaborate in this commit
message more about why it is useful (or at least harmless for performing
this check in the case of fetch/push)?

> diff --git a/submodule.c b/submodule.c
> index 3bcf44521..100d31d39 100644
> --- a/submodule.c
> +++ b/submodule.c
> @@ -648,6 +648,30 @@ static int submodule_has_commits(const char *path, struct oid_array *commits)
>                 return 0;
>
>         oid_array_for_each_unique(commits, check_has_commit, &has_commit);
> +
> +       if (has_commit) {
> +               /*
> +                * Even if the submodule is checked out and the commit is
> +                * present, make sure it is reachable from a ref.
> +                */
> +               struct child_process cp = CHILD_PROCESS_INIT;
> +               struct strbuf out = STRBUF_INIT;
> +
> +               argv_array_pushl(&cp.args, "rev-list", "-n", "1", NULL);
> +               oid_array_for_each_unique(commits, append_oid_to_argv, &cp.args);
> +               argv_array_pushl(&cp.args, "--not", "--all", NULL);
> +
> +               prepare_submodule_repo_env(&cp.env_array);
> +               cp.git_cmd = 1;
> +               cp.no_stdin = 1;
> +               cp.dir = path;
> +
> +               if (capture_command(&cp, &out, 1024) || out.len)
> +                       has_commit = 0;

So if we fail to launch capture_command, we assume we do not
have the commits?

capture_command can fail for reasons that are hard to track down
or even spurious (OOM due to excessive output, disk failure,
corrupt repo, error in executing the process, getting a signal and
so on), some of them are ok to ignore, others should never be ignored.

So I'd rather die on capture_command, and inspect out.len only
in case of successful capturing.

In addition to that we're only interested if there is any output,
such that we can optimize further:
c.f. http://public-inbox.org/git/20170324223848.GH31294@aiede.mtv.corp.google.com/

    if (start_command(&cp))
        die("cannot start git-rev-list in submodule'%s', sub->path);

    /* read only one character, if any */
    if (xread(cp.out, &tmp, 1);
        has_commit = 0;

    /*
     * close cp.out, such that the child may get SIGPIPE, upon which
     * it dies (silently, maybe we need to suppress cp.err ?)
     */
    close(cp.out);

    /*
     * Though we need to be nitpicky about finish_command IIUC by now:
     * TODO(sbeller): if this turns out to be true, fixup is_submodule_modified
    */
    int code = finish_command(&cp);
    if (!code && code != 128 + SIGPIPE)
        die("git rev-list failed in submodule'%s'", sub->path);

Upon rereading the patch, I notice the '-n 1', which would make the
optimized code above useless, so just consider it food for thought.

Thanks,
Stefan