Re: [BUG] Git 2.38.0-rc1 t1800 message text comparison

From: René Scharfe <l.s.r@web.de>
Date: Jun 4, 2023, 12:13 UTC
Message-ID: <f5519f93-bca7-ecbc-9da3-b25ce8aef618@web.de>
In-Reply-To: <013601d98cf2$392153c0$ab63fb40$@nexbridge.com>

Am 22.05.23 um 23:13 schrieb rsbecker@nexbridge.com:

Show 17 quoted lines

> On Monday, May 22, 2023 4:49 PM, I wrote:
>> On Monday, May 22, 2023 4:39 PM, René Scharfe wrote:
>>> So trying to execute an executable file consisting only of the line
>>> "#!/bad/path/no/spaces" causes NonStop to report "Permission denied"?
>>> That message text belongs to error code EACCES, not to EPERM
>>> ("Operation not permitted"), right?
>>
>> That should be correct, although the OS Devs I spoke to about this said "EPERM". I
>> am experimenting.
>>
>>> POSIX allows execve to return EACCES if the file to execute is not a
>>> regular file and executing that file type is not supported or if
>>> permissions are missing to one of its path components.
>>
>> Part of the OS Dev's response was that POSIX is actually ambiguous on this point.
>> Linux made the decision to use ENOENT. NonStop decided to use EPERM (although it
>> may actually be EACCESS - I will report back.

In https://pubs.opengroup.org/onlinepubs/9699919799/functions/exec.html I don't see how EACCES would be a valid response in this case. And it makes no intuitive sense -- which permission is missing in this situation?

That doesn't change the fact that this issue needs to be dealt with in Git somehow.

> NonStop actually does report EACCES, not EPERM. The comment at the
> front of run-command.c does describe the situation.

Do you mean the one about EACCES being reported if a directory in $PATH is inaccessible in sane_execvp()? That does not apply here because "#!/bad/path/no/spaces" specifies an absolute path, so $PATH is not searched.

Show 17 quoted lines

> The following is a potential fix:
>
> diff --git a/run-command.c b/run-command.c
> index 60c9419866..b76e117d35 100644
> --- a/run-command.c
> +++ b/run-command.c
> @@ -846,7 +846,7 @@ int start_command(struct child_process *cmd)
>                         execve(argv.v[0], (char *const *) argv.v,
>                                (char *const *) childenv);
>
> -               if (errno == ENOENT) {
> +               if (errno == ENOENT || errno == EACCES) {
>                         if (cmd->silent_exec_failure)
>                                 child_die(CHILD_ERR_SILENT);
>                         child_die(CHILD_ERR_ENOENT);
>
> This does pass and should cover all POSIX interpretations.

That would misreport execution failures due to missing permissions (think e.g. "chmod 0 file") as "No such file or directory". So this should be only done if really needed, perhaps guarded by an access(2) check for verifying that EACCES is bogus, and only on affected platforms.

René