[Buildroot] [PATCH v2 1/1] erlang: remove non-SMP build option

[Frank Hunleth]

Yann,

On Wed, Dec 20, 2017 at 9:10 AM, Yann E. MORIN <yann.morin.1998 at free.fr> wrote:
> Frank, All,
>
> On 2017-12-20 08:52 -0500, Frank Hunleth spake thusly:
>> On Wed, Dec 20, 2017 at 8:28 AM, Yann E. MORIN <yann.morin.1998 at free.fr> wrote:
>> > On 2017-12-20 11:26 +0100, Thomas Petazzoni spake thusly:
>> >> On Tue, 19 Dec 2017 13:45:44 -0500, Frank Hunleth wrote:
>> >> > The non-SMP scheduler was deprecated with the Erlang/OTP 20.0 release and
>> >> > slated for removal with the next major Erlang release. Since the non-SMP
>> >> > scheduler isn't even built anymore, this option no longer has the
>> >> > intended effect of saving space or compile time. The SMP scheduler
>> >> > supports both SMP and non-SMP processors, so removing the option will
>> >> > not break any platforms.
>> >> >
>> >> > Signed-off-by: Frank Hunleth <fhunleth at troodon-software.com>
>> >> > ---
>> >> > Changes v1 -> v2:
>> >> >   - Added text to Config.in.legacy to explain the removal of the option
>> >> >     (suggested by Yann - I added to Config.in.legacy after thinking about
>> >> >      it, since anyone who previously selected this option may be surprised
>> >> >      to see the SMP scheduler in their process list.)
>> >> >
>> >> >  Config.in.legacy         | 10 ++++++++++
>> >> >  package/erlang/Config.in | 10 ----------
>> >> >  package/erlang/erlang.mk |  4 ----
>> >> >  3 files changed, 10 insertions(+), 14 deletions(-)
>> >> >
>> >> > diff --git a/Config.in.legacy b/Config.in.legacy
>> >> > index decbace..1607df6 100644
>> >> > --- a/Config.in.legacy
>> >> > +++ b/Config.in.legacy
>> >> > @@ -153,6 +153,16 @@ config BR2_PACKAGE_GNUPG2_GPGV2
>> >> >       The gpgv2 executable is now named gpgv. The config option
>> >> >       has been renamed accordingly.
>> >> >
>> >> > +config BR2_PACKAGE_ERLANG_SMP
>> >> > +   bool "erlang smp option removed"
>> >> > +   select BR2_LEGACY
>> >> > +   help
>> >> > +     This option used to disable Erlang's SMP scheduler to save
>> >> > +     space. Since Erlang 20.0 deprecated the non-SMP scheduler and
>> >> > +     no longer builds it, it is no longer relevant. See the Erlang
>> >> > +     20.0 release notes for more information. Note that the SMP
>> >> > +     scheduler runs on both SMP and non-SMP CPUs.
>> >>
>> >> As Yann explained, we do not want a Config.in.legacy option. Indeed,
>> >> the new behavior is identical to having the option enabled, so there is
>> >> no point in having a Config.in.legacy entry.
>> >
>> > Yes, indeed.
>> >
>> > Frank, what I meant was  that, usually we have to add a legacy entry to
>> > tell users what is happenning. However, in this particular case, it is
>> > not needed because a build with this option enabled previously will now
>> > yield the same result as  build now with this option removed.
>>
>> That's not true. Before if you enabled the option, you'd get the
>> uniprocessor scheduler. With the change, you get the SMP scheduler.
>>
>> In the .mk file, if you were to enable BR2_PACKAGE_ERLANG_SMP, that
>> would pass "--disable-smp-support" to the configure script.
>
> That's wrong, because the code was:
>
>     ifeq ($(BR2_PACKAGE_ERLANG_SMP),)
>     ERLANG_CONF_OPTS += --disable-smp-support
>     endif
>
> If BR2_PACKAGE_ERLANG_SMP was set, it would not fulfill the condition,
> and we would *not* pass --disable-smp-support. But if it was unset, then
> the condition would be true, and we'd pass --disable-smp-support.

Ahh, thanks! I think that I've been out of working with Makefiles for
too long since now that you point it out, this is completely obvious.
Thanks for bearing with me.

> Now, with your change, we will no longer pass --disable-smp-support,
> ever, so this is back to option enabled.
>
> (note that we can't account for the case where the option was disabled;
> that's a limitation, but we don't really care).
>
>> That
>> forces the uniprocessor scheduler. With this patch, you only can get
>> the SMP scheduler. I'm not sure how many people would do this, but
>> operationally, you'd see a process called "beam.smp" running on your
>> system now. The uniprocessor scheduler made a process called "beam".
>> That seems like a different result to me?
>>
>> Looking over the Erlang release notes, the situation may even be worse
>> now since they've added a --enable-plain-scheduler option to build the
>> uniprocessor scheduler, so the current option BR2_PACKAGE_ERLANG_SMP
>> option may not even work. I don't know, but the uniprocessor scheduler
>> is not worth spending time on any more since it's going to be removed
>> soon.
>
> Well, depends. Are there erlang packages that require the UP scheduler?

Not that I'm aware of. We've been using the SMP scheduler for at least
a year across all platforms on the Nerves project and I haven't
received a report of anyone having an issue with it. On the other
hand, the library that wraps sqlite doesn't work with the UP scheduler
and people used to run into that frequently.

> From your commit log, it looks like the SMP scheduler can cope with
> non-SMP stuff as well, so there is no risk of breakage.
>
> Thus, I'm fine with not considering a UP option.
>
>> > But this it is usualy to add a legacy entry, for the reviewr's sake, it
>> > is important that the commit log explains why a legacy entry is not
>> > needed.
>>
>> I'm fine either way. I'm stuck in the thinking that this change isn't
>> identical to the previous behavior, so I'm not sure how to explain it
>> in the commit message. If someone indulges me with text, I'll gladly
>> send a v3 or feel free to merge one of my other versions.
>
> What about:
>
>     erlang: remove non-SMP build option
>
>     [your commit log]
>
>     We do not need to add a legacy entry, because the new behaviour is
>     the same as with the option previously set (i.e. SMP enabled).
>
>     Signed-off-by: You.

Thank you! Patch coming!

Frank