Re: [PATCH 2/2] archive: avoid spawning `gzip`

[Johannes Schindelin <Johannes.Schindelin@gmx.de>]

[-- Attachment #1: Type: text/plain, Size: 4555 bytes --]

Hi Peff,

On Fri, 12 Apr 2019, Jeff King wrote:

> On Fri, Apr 12, 2019 at 04:04:40PM -0700, Rohit Ashiwal via GitGitGadget wrote:
>
> > From: Rohit Ashiwal <rohit.ashiwal265@gmail.com>
> >
> > As we already link to the zlib library, we can perform the compression
> > without even requiring gzip on the host machine.
>
> Very cool. It's nice to drop a dependency, and this should be a bit more
> efficient, too.

Sadly, no, as René intuited and as your testing shows: there seems to be a
~15% penalty for compressing in the same thread as producing the data to
be compressed.

Given that it reduces the number of dependencies, and given that it might
be better to rely on the external command `pigz -cn` if speed is really a
matter, I still think it makes sense to switch the default, though.

> > diff --git a/archive-tar.c b/archive-tar.c
> > index ba37dad27c..5979ed14b7 100644
> > --- a/archive-tar.c
> > +++ b/archive-tar.c
> > @@ -466,18 +466,34 @@ static int write_tar_filter_archive(const struct archiver *ar,
> >  	filter.use_shell = 1;
> >  	filter.in = -1;
> >
> > -	if (start_command(&filter) < 0)
> > -		die_errno(_("unable to start '%s' filter"), argv[0]);
> > -	close(1);
> > -	if (dup2(filter.in, 1) < 0)
> > -		die_errno(_("unable to redirect descriptor"));
> > -	close(filter.in);
> > +	if (!strcmp("gzip -cn", ar->data)) {
>
> I wondered how you were going to kick this in, since users can define
> arbitrary filters. I think it's kind of neat to automagically convert
> "gzip -cn" (which also happens to be the default). But I think we should
> mention that in the Documentation, in case somebody tries to use a
> custom version of gzip and wonders why it isn't kicking in.
>
> Likewise, it might make sense in the tests to put a poison gzip in the
> $PATH so that we can be sure we're using our internal code, and not just
> calling out to gzip (on platforms that have it, of course).
>
> The alternative is that we could use a special token like ":zlib" or
> something to indicate that the internal implementation should be used
> (and then tweak the baked-in default, too). That might be less
> surprising for users, but most people would still get the benefit since
> they'd be using the default config.

I went with `:zlib`.

> > +		char outmode[4] = "wb\0";
>
> This looks sufficiently magical that it might merit a comment. I had to
> look in the zlib header file to learn that this is just a normal
> stdio-style mode. But we can't just do:
>
>   gzip = gzdopen(fd, "wb");
>
> because we want to (maybe) append a compression level. It's also
> slightly confusing that it explicitly includes a NUL, but later:
>
> > +		if (args->compression_level >= 0 && args->compression_level <= 9)
> > +			outmode[2] = '0' + args->compression_level;
>
> we may overwrite that and assume that outmode[3] is also a NUL. Which it
> is, because of how C initialization works. But that means we also do not
> need the "\0" in the initializer.
>
> Dropping that may make it slightly less jarring (any time I see a
> backslash escape in an initializer, I assume I'm in for some binary
> trickery, but this turns out to be much more mundane).
>
> I'd also consider just using a strbuf:
>
>   struct strbuf outmode = STRBUF_INIT;
>
>   strbuf_addstr(&outmode, "wb");
>   if (args->compression_level >= 0 && args->compression_level <= 9)
> 	strbuf_addch(&outmode, '0' + args->compression_level);
>
> That's overkill in a sense, but it saves us having to deal with
> manually-counted offsets, and this code is only run once per program
> invocation, so the efficiency shouldn't matter.

I'll change that, too, as it seems that `pigz` allows compression levels
higher than 9, in which case we need `strbuf_addf()` anyway. I will not
adjust the condition `<= 9`, of course, as zlib is still limited that way.

> > +		gzip = gzdopen(fileno(stdout), outmode);
> > +		if (!gzip)
> > +			die(_("Could not gzdopen stdout"));
>
> Is there a way to get a more specific error from zlib? I'm less
> concerned about gzdopen here (which should never fail), and more about
> the writing and closing steps. I don't see anything good for gzdopen(),
> but...

Sadly, I did not find anything there.

> > +	if (gzip) {
> > +		if (gzclose(gzip) != Z_OK)
> > +			die(_("gzclose failed"));
>
> ...according to zlib.h, here the returned int is meaningful. And if
> Z_ERRNO, we should probably use die_errno() to give a better message.

Okay.

Thanks,
Dscho