[oi-dev] OI project reboot required

[Garrett D'Amore]

On May 12, 2013, at 8:51 AM, Alan Coopersmith <alan.coopersmith at oracle.com> wrote:

> On 05/12/13 05:19 AM, David Höppner wrote:
>> I noticed Oracle upstream moves aggressively to amd64 only;
>> installing amd64 just in bin not in bin/$(MACH64).
> 
> It has been a few years since Oracle upstream dropped 32-bit i386 support,
> so that's just one of the decisions OI has to make - track upstream as is
> or fork/patch as needed to continue to support 32-bit on i386.

Yep.  And that has sweeping consequences; lots of things depend upon it this decision.

I'm of the opinion that enough time has passed that we should seriously consider doing the same.  Its been about a decade since  64-bit  x86 systems came on the scene (Opteron was released in June 2003).  The only 32-bit only variants that I'm aware of that are even vaguely recent are the older Atom processors (still about 4-5 years old), AMD Geode (last released 2009, though I guess they are selling them for embedded customers until 2015 or thereabouts), and certain Via processors (although all current Via products support x86_64).

Of these, I'm only aware of the following 32-bit CPUs that might be out there running illumos code that aren't also truly ancient (a decade or so old):

	* Intel Atom systems -- there were a few of these, but they are really really low performance systems
	* AMD Sempron
	* Pentium 4 (certain systems dating back to about 2004)

There were some older virtualization products where it made sense to run illumos in a 32-bit VM, but all those hypervisors have *long* since been updated to support amd64.

None of the others have, afaik, ever been used to run illumos or Solaris outside of experimental cases just to see if it would work.

I haven't seen a serious use of a 32-bit only CPU in production in over 5 years.  And I think most hobbyists upgrade their kit more frequently as well -- I have to believe almost everyone is on 64-bit kit these days.  Furthermore, most interesting systems (based on illumos) require more memory than is practical with a 32-bit only CPU.

I have to believe we could eliminate a *lot* of baggage by nixing 32-bit support.  I *know* we can, because I've nixed a bunch of system utilities in our DEY environment that were delivered in both 32 and 64 bit variants.

We're going to have to support a 32-bit userland for some time to come, unfortunately, but we should no longer make that the default, and we should deliver all of our system utilities in 64-bit only form, IMO; and we could entirely kill off the 32-bit kernel.

Alternatively, if there is sufficient demand, one could imagine a separate architecture for ia32, that is the 32-bit variant port.  That would not need to carry the 64-bit support.   Admittedly going this route reduces the likelihood of keeping certain bits capable of running 32-bit mode (e.g. device drivers), but one would argue that 32 bit systems are unlikely to adapt new devices, etc.  precisely because they are *so* friggin' old.

So, out of curiosity -- *who* is actively running illumos on 32-bit kit today?  I'm not interested in hypothetical uses or kit that is sitting around in your garage waiting for you to do something with it…. I'm interested in people who would be immediately impacted and severely so if illumos were not available on 32-bit CPUs right now.  (To give a counter example: I have a 32-bit Atom netbook, that I have OpenIndiana on.   I turn it on once every year or so… if that often … so I can't seriously claim that I would be negatively impacted if illumos were to move to 64-bit only.)

	- Garrett