Re: [PATCH v5 1/5] setup: add gentle version of read_gitfile

From: Jeff King <peff@peff.net>
Date: Apr 28, 2015, 20:19 UTC
Message-ID: <20150428201918.GA10902@peff.net>
In-Reply-To: <CAMpP7Nb3aiMoTjtPJNJPv38G54ZawS8B+NDX0x2iNe6FA-L+Lw@mail.gmail.com>

On Tue, Apr 28, 2015 at 10:07:43PM +0200, erik elfström wrote:

Show 25 quoted lines

> On Tue, Apr 28, 2015 at 8:17 AM, Jeff King <peff@peff.net> wrote:
> >
> > There was a discussion not too long ago on strategies for returning
> > errors, and one of the suggestions was to return an "error strbuf"
> > rather than a code[1]. That's less flexible, as the caller can't react
> > differently based on the type of error. But for cases like this, where
> > the only fate for the code is to get converted back into a message,
> > it can reduce the boilerplate.
> >
> > What you have here is OK to me, and I don't want to hold up your patch
> > series in a flamewar about error-reporting techniques. But I think it's
> > an interesting case study.
> >
> > -Peff
> 
> Thanks. I haven't had time to look through that thread yet, I'll try
> to get to that later.
> 
> My initial reaction is a bit skeptical though. For this case we
> currently don't want any error reporting, the NULL return is
> sufficient and even allocating and sending in the int* is pure noise.
> Allocating and releasing a strbuf seems like a lot more overhead for
> this type of caller? The one other potential candidate caller for
> read_gitfile_gently that I have seen (clone.c:get_repo_path) don't
> want any error code or message either as far as i can tell.

I had envisioned that the strbuf would be optional. I.e., you would have:

  /* like error(), but dump the message in a strbuf instead of stderr */
  int error_buf(struct strbuf *buf, const char *fmt, ...)
  {
	if (buf) {
		va_list ap;
		va_start(ap, fmt);
		strbuf_vaddf(buf, fmt, ap);
		va_end(ap);
	}
	return -1;
  }

and then in the error-reporting function:

  const char *read_gitfile_gently(const char *path, struct strbuf *err)
  {
	...
	fd = open(path, O_RDONLY);
	if (fd < 0) {
		error_buf(err, "unable to open %s: %s", path, strerror(errno));
		return NULL; /* or goto cleanup if necessary */
	}
  }

and then one caller can do:

  if (!read_gitfile_gently(path, NULL)) {
	/* we know there was an error, but we did not ask for details */
	...
  }

and the non-gentle read_gitfile() becomes:

  const char *read_gitfile(const char *path)
  {
	struct strbuf err = STRBUF_INIT;
	const char *ret = read_gitfile_gently(path, &err);
	if (!ret)
		die("%s", err.buf);
	/* no need to free err; if there was no error, nothing was written */
	return path;
  }

Note that the "return -1" from error_buf() is not useful here, but it might be used as a shortcut in other situations (e.g., the same places we call "return error()" now).

> Also if it turns out that we actually need to treat the "file too
> large" error differently in clean (as discussed in thread on the file
> size check) then we can no longer communicate that back using the
> strbuf interface.

Yeah, agreed. This system breaks down as soon as you need to programatically know which error happened.

-Peff