Re: Usage of real m68k hardware

[Thorsten Glaser <t.glaser@tarent.de>]

Adrian wrote:

>Yes, of course. But Andreas hit a nerve with this on me. This project
>has cost me lots of blood, tears and sweat and if someone is asking
>for it to be completely thrown out out of nothing, I'm getting a bit
>stressed out.

I completely agree here. While I’m no longer involved with the
m68k port specifically, after having spent THREE YEARS of blood,
sweat and pain to resurrect it, there are several reasons:

• odd architectures *do* help in finding odd bugs, often before
  they hit other architectures where they’re hidden by, for
  example, compiler optimisations (more aggressive inlining)
  or arch-specific asm code, until these hiding things no longer
  appear

  ‣ granted, m68k has this specific “2-byte alignment” thing,
    but then, anything that actually relies on the precise amount
    of struct padding is relying on IB/UB in the first place and,
    with that, buggy

• I have come to actually like that, having been a die-hard 8088
  user in my childhood, and found the people and community very
  interesting
  ‣ there are fun projects like a PCI bridge, which allows using
    a PCI Radeon graphics card with LCDs at 1900x12something
    resolution, currently with GEM/AES only, not yet in Linux

• it sends a signal, and the wrong signal in my eyes, that
  everything not-mainstream is not worth to be supported

  ‣ specialisation is for downstreams, Debian should stay universal

  ‣ read up on monoculture in agriculture and why everyone, by now,
    thinks it’s a bad idea
    ⇒ hint: Meltdown/Spectre…

• I’m running (and helping to work on) x32… another port

• I found Debian ports very useful to gain deep insight on
  how Debian and all of its components work, and can recommend
  porting a new or resurrecting an old architecture to everyone
  wishing to peek below the surface

• I’ve heard someone’s working on making dak dports-capable,
  solving the current worst problem of the fact we use mini-dak
  in that NBS packages are removed from the archive even if
  they’re still Depended upon, which made me really excited
  about dports work


On the more technical side, while Adrian’s buildds are qemu,
I’ve continued running an ARAnyM (also emulation, but different
and thanks to Doko even FPU-complete) buildd for as long as the
system it was hosted on allowed me to do so. (That GPLhost domU
is currently unusable because of spontaneous reboots and other
problems. I might look into running one on some other system;
I have a couple of VMs on my workplace desktop but can’t use
those as they are bridged into the company LAN.)

We also have a number of Amiga and Atari and I believe at least
one or two Macintosh systems which, at one point or the other,
are or were in use as buildds and/or porterboxen.

I’ve also made a point in making ready-to-use ARAnyM images
in the Debian wiki which any maintainer could use to run a
box locally, due to the lack of currently-accessible porterboxen.
This has eroded a bit since I moved away from m68k.

I don’t know how the actual hardware can be helped to become
more usable. I also don’t know if the standard Debian porterbox
setup can be used on/for them. DSA normally does these things;
in dports we want to make things as closely to the main Debian
as possible, but as long as dports are officially unsupported,
it’s hard. (Also, you’d have to talk to Ingo, perhaps Adrian
and ragnar76 about the actual hardware.)

As for the “qemu bug” issue, using an ARAnyM, Amiga, Atari
or Macintosh machine to retry the build (since they all are
slower, although my previous desktop could emulate a 200 MHz
m68k with ARAnyM) before complaining would certainly help.
But this is also not easy, and only a few problems are caused
by qemu issues (I’m actually surprised, I’d have not thought
a qemu-based buildd a viable solution, and I recall Adrian
and me fighting a bit over it initially), so I don’t understand
An3as’ violent reaction.

Contrasted to that, x32 hardware is actually easy: use an amd64
system with “syscall.x32=y” in GRUB_CMDLINE_LINUX then just use
a foreign-arch chroot with pbuilder/cowbuilder or schroot/sbuild
like you would with i386. (I’ve had two cases in which an FTBFS
was actually a hiccup of the buildd, or a difference in host CPU,
and which built just fine on my system, again, in a clean&minimal
environment, so I just did a porter upload.) These are dead useful
to reproduce issues, by the way.


It might also be useful to create one or two buildds with
large hard discs (and possibly RAM) since some of the recent
packages (gcc-*-cross-* most prominently) make Adrian’s
systems explode… especially as his virtual buildds share
(limited) space.


Adrian is currently the single most-involved person driving
debian-ports forwards, on a *lot* of architectures, (not saying
there are no other porters) so I can understand his frustration.

I might even look if I can help any further. Unfortunately, as
I said above, I have no easy solution for running a buildd or
porterbox (company LAN), only for local porter builds (in clean
environments sufficiently suitable for uploadinig to the archive,
of course).

We need people helping with debugging issues all the time, e.g.
the new Qt 5.10 bugs on sparc64 and x32, and this seems to be
more like an upstream code issue than a ports-specific issue, and
I plain down don’t understand C++ well enough, even *if* I recently
have tried to become involved in MuseScore upstream.

So, please DO NOT leave us alone!

Thanks,
//mirabilos (tg@d.o)
-- 
15:41⎜<Lo-lan-do:#fusionforge> Somebody write a testsuite for helloworld :-)