Re: PATCH] bisect--helper: plug strvec leak in bisect_start()

>
> On Tue, Oct 04 2022, René Scharfe wrote:
>
>> The strvec "argv" is used to build a command for run_command_v_opt(),
>> but never freed.  Use the "args" strvec of struct child_process and
>> run_command() instead, which releases the allocated memory both on
>> success and on error.  We just also need to set the "git_cmd" bit
>> directly.
>>
>> Signed-off-by: René Scharfe <l.s.r@web.de>
>> ---
>>  builtin/bisect--helper.c | 7 ++++---
>>  1 file changed, 4 insertions(+), 3 deletions(-)
>>
>> diff --git a/builtin/bisect--helper.c b/builtin/bisect--helper.c
>> index 501245fac9..9fe0c08479 100644
>> --- a/builtin/bisect--helper.c
>> +++ b/builtin/bisect--helper.c
>> @@ -765,11 +765,12 @@ static enum bisect_error bisect_start(struct bisect_terms *terms, const char **a
>>  		strbuf_read_file(&start_head, git_path_bisect_start(), 0);
>>  		strbuf_trim(&start_head);
>>  		if (!no_checkout) {
>> -			struct strvec argv = STRVEC_INIT;
>> +			struct child_process cmd = CHILD_PROCESS_INIT;
>>
>> -			strvec_pushl(&argv, "checkout", start_head.buf,
>> +			cmd.git_cmd = 1;
>> +			strvec_pushl(&cmd.args, "checkout", start_head.buf,
>>  				     "--", NULL);
>> -			if (run_command_v_opt(argv.v, RUN_GIT_CMD)) {
>> +			if (run_command(&cmd)) {
>>  				res = error(_("checking out '%s' failed."
>>  						 " Try 'git bisect start "
>>  						 "<valid-branch>'."),
>
> Okey, so we leak the strvec, and instead of adding a strvec_clear()
> you're just switching the lower-level API, which we'd need in some cases
> with this API, and would often be cleaner.
>
> But I don't get it in this case, why not just:
>
> 	diff --git a/builtin/bisect--helper.c b/builtin/bisect--helper.c
> 	index 4e97817fba5..f9645a9d0df 100644
> 	--- a/builtin/bisect--helper.c
> 	+++ b/builtin/bisect--helper.c
> 	@@ -763,11 +763,9 @@ static enum bisect_error bisect_start(struct bisect_terms *terms, const char **a
> 	 		strbuf_read_file(&start_head, git_path_bisect_start(), 0);
> 	 		strbuf_trim(&start_head);
> 	 		if (!no_checkout) {
> 	-			struct strvec argv = STRVEC_INIT;
> 	+			const char *argv[] = { "checkout", start_head.buf, "--", NULL };
>
> 	-			strvec_pushl(&argv, "checkout", start_head.buf,
> 	-				     "--", NULL);
> 	-			if (run_command_v_opt(argv.v, RUN_GIT_CMD)) {
> 	+			if (run_command_v_opt(argv, RUN_GIT_CMD)) {
> 	 				res = error(_("checking out '%s' failed."
> 	 						 " Try 'git bisect start "
> 	 						 "<valid-branch>'."),

> The common pattern for run_command_v_opt() callers that don't need a
> dynamic list is exactly that, e.g.:
>
> 	builtin/difftool.c=static int print_tool_help(void)
> 	builtin/difftool.c-{
> 	builtin/difftool.c-     const char *argv[] = { "mergetool", "--tool-help=diff", NULL };
> 	builtin/difftool.c:     return run_command_v_opt(argv, RUN_GIT_CMD);
> 	builtin/difftool.c-}
>
> And:
>
> 	fsmonitor-ipc.c=static int spawn_daemon(void)
> 	fsmonitor-ipc.c-{
> 	fsmonitor-ipc.c-        const char *args[] = { "fsmonitor--daemon", "start", NULL };
> 	fsmonitor-ipc.c-
> 	fsmonitor-ipc.c:        return run_command_v_opt_tr2(args, RUN_COMMAND_NO_STDIN | RUN_GIT_CMD,
> 	fsmonitor-ipc.c-                                    "fsmonitor");
> 	fsmonitor-ipc.c-}
>
> Here we have the "start_head" which we'll need to strbuf_release(), but
> we're not returning directly, and the function is doing that already.
>
> Your version is slightly more memory efficient, i.e. we'll end up having
> to push this to a "struct child_process"'s argv anyway, but this is less
> code & we don't need to carefully eyeball run_command_v_opt_cd_env_tr2()
> to see that it's correct (which I did, your version does the right thing
> too).