Re: [PATCH 4/8] use child_process member "args" instead of string array variable

["René Scharfe" <l.s.r@web.de>]

Am 27.10.22 um 23:09 schrieb Ævar Arnfjörð Bjarmason:
>
> On Thu, Oct 27 2022, René Scharfe wrote:
>
>> @@ -729,20 +727,22 @@ static int is_expected_rev(const struct object_id *oid)
>>  enum bisect_error bisect_checkout(const struct object_id *bisect_rev,
>>  				  int no_checkout)
>>  {
>> -	char bisect_rev_hex[GIT_MAX_HEXSZ + 1];
>>  	struct commit *commit;
>>  	struct pretty_print_context pp = {0};
>>  	struct strbuf commit_msg = STRBUF_INIT;
>>
>> -	oid_to_hex_r(bisect_rev_hex, bisect_rev);
>>  	update_ref(NULL, "BISECT_EXPECTED_REV", bisect_rev, NULL, 0, UPDATE_REFS_DIE_ON_ERR);
>>
>> -	argv_checkout[2] = bisect_rev_hex;
>>  	if (no_checkout) {
>>  		update_ref(NULL, "BISECT_HEAD", bisect_rev, NULL, 0,
>>  			   UPDATE_REFS_DIE_ON_ERR);
>>  	} else {
>> -		if (run_command_v_opt(argv_checkout, RUN_GIT_CMD))
>> +		struct child_process cmd = CHILD_PROCESS_INIT;
>> +
>> +		cmd.git_cmd = 1;
>> +		strvec_pushl(&cmd.args, "checkout", "-q",
>> +			     oid_to_hex(bisect_rev), "--", NULL);
>> +		if (run_command(&cmd))
>>  			/*
>>  			 * Errors in `run_command()` itself, signaled by res < 0,
>>  			 * and errors in the child process, signaled by res > 0
>
> Perhaps I went overboard with it in my version, but it's probably worth
> mentioning when converting some of these that the reason for the
> pre-image of some is really not like the others.
>
> Now that we're on C99 it perhaps make s no difference, but the pre-image
> here is explicitly trying to avoid dynamic initializer elements, per
> 442c27dde78 (CodingGuidelines: mention dynamic C99 initializer elements,
> 2022-10-10).

True, some cases could be converted to string array initializations,
which also would get rid of magic numbers.  This would make the final
patch to convert them to run_command() longer.

> Well, partially, some of it appears to just be based on a
> misunderstanding of how our own APIs work, i.e. the use of
> oid_to_hex_r() over oid_to_hex().
>
>> diff --git a/builtin/am.c b/builtin/am.c
>> index 39fea24833..20aea0d248 100644
>> --- a/builtin/am.c
>> +++ b/builtin/am.c
>> @@ -2187,14 +2187,12 @@ static int show_patch(struct am_state *state, enum show_patch_type sub_mode)
>>  	int len;
>>
>>  	if (!is_null_oid(&state->orig_commit)) {
>> -		const char *av[4] = { "show", NULL, "--", NULL };
>> -		char *new_oid_str;
>> -		int ret;
>> +		struct child_process cmd = CHILD_PROCESS_INIT;
>>
>> -		av[1] = new_oid_str = xstrdup(oid_to_hex(&state->orig_commit));
>> -		ret = run_command_v_opt(av, RUN_GIT_CMD);
>> -		free(new_oid_str);
>> -		return ret;
>> +		strvec_pushl(&cmd.args, "show", oid_to_hex(&state->orig_commit),
>> +			     "--", NULL);
>> +		cmd.git_cmd = 1;
>> +		return run_command(&cmd);
>>  	}
>
> The same goes for this, FWIW I split this one out into its own commit (I
> left the earlier one alone):
> https://lore.kernel.org/git/patch-v2-04.10-5cfd6a94ce3-20221017T170316Z-avarab@gmail.com/;
> It uses the same pattern

OK, I just chalked that up as "slightly odd" and bulldozed over them
without a second thought.  Hmm.

>
>> diff --git a/builtin/difftool.c b/builtin/difftool.c
>> index 4b10ad1a36..22bcc3444b 100644
>> --- a/builtin/difftool.c
>> +++ b/builtin/difftool.c
>> @@ -360,8 +360,8 @@ static int run_dir_diff(const char *extcmd, int symlinks, const char *prefix,
>>  	struct pair_entry *entry;
>>  	struct index_state wtindex;
>>  	struct checkout lstate, rstate;
>> -	int flags = RUN_GIT_CMD, err = 0;
>> -	const char *helper_argv[] = { "difftool--helper", NULL, NULL, NULL };
>> +	int err = 0;
>> +	struct child_process cmd = CHILD_PROCESS_INIT;
>
> In general, I like the disection of this series, but with this...
>
>>  	struct hashmap wt_modified, tmp_modified;
>>  	int indices_loaded = 0;
>>
>> @@ -563,16 +563,17 @@ static int run_dir_diff(const char *extcmd, int symlinks, const char *prefix,
>>  	}
>>
>>  	strbuf_setlen(&ldir, ldir_len);
>> -	helper_argv[1] = ldir.buf;
>>  	strbuf_setlen(&rdir, rdir_len);
>> -	helper_argv[2] = rdir.buf;
>>
>>  	if (extcmd) {
>> -		helper_argv[0] = extcmd;
>> -		flags = 0;
>> -	} else
>> +		strvec_push(&cmd.args, extcmd);
>> +	} else {
>> +		strvec_push(&cmd.args, "difftool--helper");
>> +		cmd.git_cmd = 1;
>
> ...and the frequent occurance of just e.g. "cmd.git_cmd = 1" and nothing
> else I'm wondering if we're not throwing the baby out with the bath
> water in having no convenience wrappers or macros at all.
>
> A lot of your 3-lines would be 1 lines if we just had e.g. (untested,
> and could be a function not a macro, but you get the idea):
>
> 	#define run_command_git_simple(__VA_ARGS__) \
> 		struct child_process cmd = CHILD_PROCESS_INIT; \
> 		cmd.git_cmd = 1; \
> 		strvec_pushl(&cmd.args, __VA_ARGS__); \
> 		run_command(&cmd);
>
> But maybe nobody except me thinks that's worthwhile...

I have similar temptations; you could see that in my scratch patch
https://lore.kernel.org/git/9d924a5d-5c72-fbe6-270c-a8f6c5fc5850@web.de/
which added run_git_or_die() in builtin/gc.c.  Why, oh why?  Perhaps
because taking a blank form (CHILD_PROCESS_INIT), ticking boxes
(.git_cmd = 1), filling out text fields (strvec_push(...)) and
submitting it (run_command()) feels tedious and bureaucratic, Java-esque
even.  And some patterns appear again and again.

How bad is that?  Is it bad at all?  I think overall we should try to
reduce the number of external calls and make those we have to do
self-documenting and leak-free.  A bit of tedium is OK; this API should
be used rarely and sparingly.  Still I get the urge to search for
patterns and define shortcuts when I see all those similar calls..

run_command_git_simple as defined above wouldn't compile, but I get it.
Reducing the number of lines feels good, but it also makes the code less
flexible -- adding a conditional parameter requires converting back to
run_command().

>
>>  static void read_empty(const struct object_id *oid)
>>  {
>> -	int i = 0;
>> -	const char *args[7];
>> -
>> -	args[i++] = "read-tree";
>> -	args[i++] = "-m";
>> -	args[i++] = "-u";
>> -	args[i++] = empty_tree_oid_hex();
>> -	args[i++] = oid_to_hex(oid);
>> -	args[i] = NULL;
>> +	struct child_process cmd = CHILD_PROCESS_INIT;
>> +
>> +	strvec_pushl(&cmd.args, "read-tree", "-m", "-u", empty_tree_oid_hex(),
>> +		     oid_to_hex(oid), NULL);
>> +	cmd.git_cmd = 1;
>>
>> -	if (run_command_v_opt(args, RUN_GIT_CMD))
>> +	if (run_command(&cmd))
>>  		die(_("read-tree failed"));
>>  }
>>
>>  static void reset_hard(const struct object_id *oid)
>>  {
>> -	int i = 0;
>> -	const char *args[6];
>> -
>> -	args[i++] = "read-tree";
>> -	args[i++] = "-v";
>> -	args[i++] = "--reset";
>> -	args[i++] = "-u";
>> -	args[i++] = oid_to_hex(oid);
>> -	args[i] = NULL;
>> +	struct child_process cmd = CHILD_PROCESS_INIT;
>> +
>> +	strvec_pushl(&cmd.args, "read-tree", "-v", "--reset", "-u",
>> +		     oid_to_hex(oid), NULL);
>> +	cmd.git_cmd = 1;
>>
>> -	if (run_command_v_opt(args, RUN_GIT_CMD))
>> +	if (run_command(&cmd))
>>  		die(_("read-tree failed"));
>>  }
>
> Two perfect examples, e.g. the former would just be:
>
> 	if (run_command_git_simple("read-tree", "-m", "-u", empty_tree_oid_hex(),
> 				   oid_to_hex(oid), NULL))
> 		die(...);