Re: [GIT PULL 1/2] Kbuild updates for v5.6-rc1

[Masahiro Yamada]

Hi Geert,

On Mon, Feb 3, 2020 at 8:00 PM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
>
> Hi Yamada-san,
>
> On Mon, Feb 3, 2020 at 2:28 AM Masahiro Yamada <masahiroy@xxxxxxxxxx> wrote:
> > On Sun, Feb 2, 2020 at 3:45 AM Linus Torvalds
> > <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> > > On Fri, Jan 31, 2020 at 8:06 PM Masahiro Yamada <masahiroy@xxxxxxxxxx> wrote:
> > > >
> > > >  - simplify built-in initramfs creation
> > >
> > > Hmm.
> > >
> > > This may simplify it from a _technical_ angle, but it seems to be a
> > > fairly annoying step backwards from a UI perspective.
> > >
> > > Now Kconfig asks a completely pointless question that most people have
> > > absolutely zero interest in. The old situation was better, I feel.
> > >
> > > Basically, I feel that from a "get normal users to test development
> > > kernels", our Kconfig pain ends up being the biggest hurdle by far.
> > >
> > > The kernel is easy to build and doesn't really require all that much
> > > infrastructure, but generating the config - particularly when it
> > > changes over time and you can't just say "just use the distro config"
> > > - is a big step for people.
> > >
> > > So honestly, while I've pulled this, I feel that this kind of change
> > > is going _exactly_ the wrong way when it asks people questions that
> > > they don't care one whit about.
> > >
> > > If I as a kernel developer can't find it in myself to care and go "why
> > > does it ask this new question", then that should tell you something.
> > >
> > > Why do we have this choice in the first place?
> >
> > Generally, initramfs is passed from a boot-loader,
> > but some architectures embed initramfs into vmlinux
> > (perhaps due to poor boot-loader support??)
> >
> > arch/arc/configs/tb10x_defconfig:CONFIG_INITRAMFS_SOURCE="../tb10x-rootfs.cpio"
> > arch/unicore32/configs/defconfig:#CONFIG_INITRAMFS_SOURCE="arch/unicore/ramfs/ramfs_config"
> > arch/xtensa/configs/cadence_csp_defconfig:CONFIG_INITRAMFS_SOURCE="$$KERNEL_INITRAMFS_SOURCE"
>
> Note that the above are examples that do not actually work, as the files
> referred to are not present in mainline (read below[1] why I have just
> checked that ;-).
>
> > So, data-compression is useful - that's is what I understand.
>
> Yes it is, depending on your config.
>
> > For major architectures, vmlinux embeds a tiny initramfs,
> > which is generated based on usr/default_cpio_list.
> >
> > We do not need data-compression for such a small cpio,
> > but handling it in a consistent way is sensible.
> > This is annoying from the users' PoV, I admit.
>
> I was also confused by this question, as by default you have
>     CONFIG_RD_GZIP=y
>     CONFIG_RD_BZIP2=y
>     CONFIG_RD_LZMA=y
>     CONFIG_RD_XZ=y
>     CONFIG_RD_LZO=y
>     CONFIG_RD_LZ4=y
> so the old 'default ".gz" if RD_GZIP' looked like it would use gzip.
> However, the tiny default initramfs ended up being uncompressed anyway
> before, as until commit ddd09bcc899fd374 ("initramfs: make compression
> options not depend on INITRAMFS_SOURCE"), INITRAMFS_COMPRESSION wasn't
> taken into account for the default tiny initramfs...


Because there was a bug in old usr/Kconfig.


See the following dependency in the old code:

config INITRAMFS_COMPRESSION
        depends on INITRAMFS_SOURCE!=""


In the default .config, INITRAMFS_SOURCE is "".
Hence, CONFIG_INITRAMFS_COMPRESSION is undefined.
So, suffix_y gets empty in usr/Makefile.

That is why the cpio was not compressed
even though CONFIG_RD_GZIP=y



> So INITRAMFS_COMPRESSION_NONE is the right answer to retain the
> old behavior?

Yes, INITRAMFS_COMPRESSION_NONE retains the previous behavior.

But, as far as he I understood what Linus said,
"we do not care".


> One might question why not to use gzip anyway, as
> CONFIG_RD_GZIP=y is enabled by default, and would give a (small)
> improvement of ca. 350 bytes ;-)
> Hence there is some area for improvement...


GZIP is not the only compression algorithm.

Somebody may want to disable RD_GZIP,
then enable RD_XZ.
If we allow the data compression,
Kconfig must ask "which compression algorithm".


So, if Kconfig would siltently choose something
as default, INITRAMFS_COMPRESSION_NONE would be the best.




>
> Thanks!
>
> [1] I'm still carrying a local patch for handling relative initramfs
>     paths with O=, but it's been a while I actually used it.  Due to your
>     recent changes, it no longer applies, and needs to be updated.
>     But of course that is only useful if there are some real users...
>     https://lore.kernel.org/lkml/1384467283-14806-1-git-send-email-geert@xxxxxxxxxxxxxx/
>
> Gr{oetje,eeting}s,
>
>                         Geert
>
> --
> Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx
>
> In personal conversations with technical people, I call myself a hacker. But
> when I'm talking to journalists I just say "programmer" or something like that.
>                                 -- Linus Torvalds



-- 
Best Regards
Masahiro Yamada