Re: [PATCHv14 5/7] git submodule update: have a dedicated helper for cloning

From: Stefan Beller <sbeller@google.com>
Date: Feb 19, 2016, 23:49 UTC
Message-ID: <CAGZ79kZjD_hRBFEVpj-80NpaYR2qRvnLbBb+9kR4x7MKoRX+UQ@mail.gmail.com>
In-Reply-To: <20160219230740.GM28749@google.com>

On Fri, Feb 19, 2016 at 3:07 PM, Jonathan Nieder <jrnieder@gmail.com> wrote:

>
> I still have trouble reading this patch (patch 5).  Some musing below
> to figure out what could change to make it more readable (perhaps in a
> patch on top).

After this review (I added replies from bottom to top, the same way I write my code ;) I think we need another reroll at least.

>> +#define SUBMODULE_UPDATE_CLONE_INIT {0, 0, 0, NULL, NULL, NULL, NULL, MODULE_LIST_INIT, STRING_LIST_INIT_DUP, SUBMODULE_UPDATE_STRATEGY_INIT}
>
> I think these struct fields need some comments.  It's not clear what
> most of them represent.

ok.

Show 12 quoted lines

>
> [...]
>> +static int update_clone_inspect_next_task(struct child_process *cp,
>
> What is being inspected here?  What does the return value represent?
>
> Would a name like 'prepare_to_clone_next_submodule' make sense?  A
> comment could say that 'ce' points to the candidate submodule to clone,
> that the return value represents whether we want to clone it, and
> that the first parameter is an output parameter representing a command
> to run to carry out the clone.
>

ok.

Show 6 quoted lines

>
> It's common to call submodule_from_path with null_sha1 as a parameter
> but I have trouble continuing to remember what that means.  Maybe
> there should be a separate function that handles that?  As a
> side-effect, the name and docstring of that function could explain
> what it means, which I still am not sure about. :)

Good idea!

Show 5 quoted lines

>         if (pp->update.type == SM_UPDATE_NONE
>             || (pp->update.type == SM_UPDATE_UNSPECIFIED
>                 && sub->update_strategy.type == SM_UPDATE_NONE)) {
>
> What does pp stand for?

parallel processes similar to cp standing for child process.

>
> Is the --update parameter ever set to 'none'?  What does it mean
> when someone sets it to none?

In the initialization. This is to preserve precedence. Yeah there is a test for --update=none.

Show 11 quoted lines

>
>> +             strbuf_addf(err, "Skipping submodule '%s'\n",
>> +                         displaypath);
>
> Does the caller expect a newline at the end of err?
>
> In the refs code that uses an err strbuf, the convention is to
> not end the message with a newline.  That way, a function like
> die() can insert a newline and messages are guaranteed to be
> newline-terminated even if someone is sloppy and does the wrong thing
> when generating an error.

Oh! So we need to add new lines ourselves here.

Show 6 quoted lines

>> +             if (pp->reference)
>> +                     argv_array_push(&cp->args, pp->reference);
>> +             if (pp->depth)
>> +                     argv_array_push(&cp->args, pp->depth);
>
> What does 'cp' stand for mean?  Would a name like 'child' work?

child process ? (any code which had a struct child_process referred to it as *cp)

Show 9 quoted lines

>
> The 'count' variable is more of a cursor than a count.  Would a name +
> comment like
>
>         /* index into 'list' representing the next submodule to consider cloning */
>         int current;
>
> work?
>

I think that's better.

Show 59 quoted lines

>
> [...]
>> +static int update_clone(int argc, const char **argv, const char *prefix)
>> +{
>> +     const char *update = NULL;
>> +     struct string_list_item *item;
>> +     struct submodule_update_clone pp = SUBMODULE_UPDATE_CLONE_INIT;
>> +
>> +     struct option module_list_options[] = {
>
> The name looks stale.
>
>> +             OPT_STRING(0, "prefix", &prefix,
>> +                        N_("path"),
>> +                        N_("path into the working tree")),
>> +             OPT_STRING(0, "recursive_prefix", &pp.recursive_prefix,
>> +                        N_("path"),
>> +                        N_("path into the working tree, across nested "
>> +                           "submodule boundaries")),
>
> What do these options represent?  I'm used to the 'prefix' parameter to
> a command coming from git machinery that remembers what the path to the
> original cwd was relative to the worktree or repository root.  Here
> there's an option to set it --- is that intentional?  Would setting the
> environment variable GIT_PREFIX (that git already knows how to respect)
> work in its place?
>
> What is recursive_prefix relative to?
>
> Nit: it would be more idiomatic to use a dash in place of an underscore
> in the second one's name.  But this is an internal interface so it
> doesn't matter much.
>
>> +             OPT_STRING(0, "update", &update,
>> +                        N_("string"),
>> +                        N_("update command for submodules")),
>
> This one is confusing to me because while the script supports --rebase /
> --merge / --checkout this option seems to be more general.
>
> If the help string said '(rebase, merge, or checkout)' then I wouldn't
> mind.
>
>> +             OPT_STRING(0, "reference", &pp.reference, "<repository>",
>> +                        N_("Use the local reference repository "
>> +                           "instead of a full clone")),
>
> Is this allowed to be relative?  If so, what is it relative to?
>
> [...]
>> +             OPT__QUIET(&pp.quiet, N_("do't print cloning progress")),
>
> Typo.
>
> [...]
>> +     const char *const git_submodule_helper_usage[] = {
>> +             N_("git submodule--helper list [--prefix=<path>] [<path>...]"),
>
> Is this the 'list' subcommand?

no. :(

Show 17 quoted lines

>
> [...]
>> +     gitmodules_config();
>> +     /* Overlay the parsed .gitmodules file with .git/config */
>> +     git_config(submodule_config, NULL);
>> +     run_processes_parallel(1, update_clone_get_next_task,
>> +                               update_clone_start_failure,
>> +                               update_clone_task_finished,
>> +                               &pp);
>
> Neat. \o/
>
>> +     if (pp.print_unmatched) {
>> +             printf("#unmatched\n");
>
> I'm still confused about this.  I think a comment where
> 'print_unmatched' is declared would probably clear it up.

Probably this is a too literal translation from shell to C. By printing #unmatched the shell on the other end of the pipe of this helper program knows to just stop and error out.

So this is telling the downstream program to stop.

Maybe we can name the variable 'quickstop' ? 'abort_and_exit' ?

Show 7 quoted lines

>
> [...]
>> +     for_each_string_list_item(item, &pp.projectlines)
>> +             utf8_fprintf(stdout, "%s", item->string);
>
> (just curious) why are these saved up and printed all at once instead
> of being printed as it goes?

To have a clean abort path, i.e. we need to be done with all the parallel stuff before we start on doing local on-disk stuff. So we need to be sure we do not print "#unmatched\n" but rather all the repositories.

In a next step, we can do the checkout in parallel if there is nothing to assume to lead to trouble. i.e. update strategy is checkout and the submodule is in a clean state at the tip of a branch.

All problems which need to be solved by the user should be presented in a sequential way, i.e. present one problem and then full stop. That seems to be more encouraging as if we'd throw a "Here are a dozen broken submodule repositories which we expect the user fixes up".

Show 9 quoted lines

>
> [...]
>> --- a/git-submodule.sh
>> +++ b/git-submodule.sh
>
> Very nice.
>
> Thanks,
> Jonathan