[9fans] 386

[9fans] 386

[Russ Cox]

Is anyone out there using Plan 9 on 386 machines?
I'm just wondering whether it's going to affect anyone
if the default kernels assume they can use some of
the 486 instructions (specifically INVLPG).

Russ

Re: [9fans] 386

[Lucio De Re]

> Is anyone out there using Plan 9 on 386 machines?
> I'm just wondering whether it's going to affect anyone
> if the default kernels assume they can use some of
> the 486 instructions (specifically INVLPG).

I think that's a slippery slope.  My AUTH server already crashes
whenever I forget that AWK uses floating point arithmetic.

Think of it this way: if one of Plan 9's strength is portability
across different architectures, it would be nice if this was
consistent.  By all means add the 486 instructions, but somehow allow
some documented way around it.

++L

Re: [9fans] 386

[William Josephson]

On Sat, Oct 29, 2005 at 11:08:18AM -0400, Russ Cox wrote:
> Is anyone out there using Plan 9 on 386 machines?
> I'm just wondering whether it's going to affect anyone
> if the default kernels assume they can use some of
> the 486 instructions (specifically INVLPG).

I am not and for some things I'm working on, I've given
up and assumed a PentiumPro.  That said, it would be
nice to retain support for older machines.

It has been a while since I've looked at Plan 9's VM
system.  How hard would it be to select an implementation
at run time?  It isn't too hard to probe for these sorts
of things at boot.  I suppose one could also make it a
compile-time option, but I'd be inclined to avoid that
since it could lead to a real mess if people decided that
they wanted to use other features down the road.

Why the sudden interest in invlpg?

Re: [9fans] 386

[jmk]

We're testing a rewrite of the x86 VM system and invlpg
is much less expensive than the alternative of flushing
the entire TLB for temporary mappings. I've been using
invlpg in the old VM system for a while but the code has
to have tests for processor type.

I'd actually like to go further and only support processors
that have the CPUID instruction (that includes the plain
Pentium otherwise my home machine would need to be replaced).

We could provide instructions for what to do if someone
desperately wanted to run the system on deprecated hardware,
it's not hard but I feel it's not worth keeping the support
in the main tree. Also, I can't find the old AT&T Safari
and Gateway Nomad laptops I kept around for testing 386
and 486 compatibility so we've no way to test on those
any more.

--jim

On Sat Oct 29 15:08:57 EDT 2005, jkw@eecs.harvard.edu wrote:
> On Sat, Oct 29, 2005 at 11:08:18AM -0400, Russ Cox wrote:
> > Is anyone out there using Plan 9 on 386 machines?
> > I'm just wondering whether it's going to affect anyone
> > if the default kernels assume they can use some of
> > the 486 instructions (specifically INVLPG).
> 
> I am not and for some things I'm working on, I've given
> up and assumed a PentiumPro.  That said, it would be
> nice to retain support for older machines.
> 
> It has been a while since I've looked at Plan 9's VM
> system.  How hard would it be to select an implementation
> at run time?  It isn't too hard to probe for these sorts
> of things at boot.  I suppose one could also make it a
> compile-time option, but I'd be inclined to avoid that
> since it could lead to a real mess if people decided that
> they wanted to use other features down the road.
> 
> Why the sudden interest in invlpg?

Re: [9fans] 386

[Lucio De Re]

> We could provide instructions for what to do if someone
> desperately wanted to run the system on deprecated hardware,

I'm sure that would be adequate.  The difficult part is knowing which
hardware is deprecated.  One discovers these things the hard way.

++L

Re: [9fans] 386

[Christopher Nielsen]

I am not, and I know at least FreeBSD cut mainstream support
for the 386 a while back, amid much gnashing of teeth. There
might be someone, someday that might want to use Plan 9 on
386, but I say support what makes sense; we don't have the
resources to support legacy hardware too far back.

On Sat, Oct 29, 2005 at 11:08:18AM -0400, Russ Cox wrote:
> Is anyone out there using Plan 9 on 386 machines?
> I'm just wondering whether it's going to affect anyone
> if the default kernels assume they can use some of
> the 486 instructions (specifically INVLPG).
> 
> Russ
> 

-- 
Christopher Nielsen
"They who can give up essential liberty for temporary
safety, deserve neither liberty nor safety." --Benjamin Franklin

Re: [9fans] 386

[Uriel]

On Sat, Oct 29, 2005 at 04:01:36PM -0400, jmk@plan9.bell-labs.com wrote:
> We're testing a rewrite of the x86 VM system and invlpg
Will that fix the issues with certain BIOS/mobos that wont boot with AGP
video cards? That would be very nice, I got a couple of such boards.

> I'd actually like to go further and only support processors
> that have the CPUID instruction (that includes the plain
> Pentium otherwise my home machine would need to be replaced).
That sounds reasonable. We have had people in #plan9 ask if they can't
run on a 486 or such, but we had to turn them down because they can't
run the installer on such system(not enough RAM).

As auth or fs servers it might make sense, but people that would use
them as such should have no problem getting an old/custom kernel built
for it.

This days it's not hard to find >P100 in junkyards if you look for them.

> We could provide instructions for what to do if someone
> desperately wanted to run the system on deprecated hardware,
> it's not hard but I feel it's not worth keeping the support
> in the main tree.
Agreed.

> Also, I can't find the old AT&T Safari and Gateway Nomad laptops I
> kept around for testing 386 and 486 compatibility so we've no way to
> test on those any more.
If you want I got an old 486 thinkpad lying around I never had time to
do anything with it, it's pretty cool, but probably shipping would be
more expensive than getting a new one from Ebay.

uriel

Re: [9fans] 386

[jmk]

On Sat Oct 29 17:08:01 EDT 2005, uriell@binarydream.org wrote:
> ...
> That sounds reasonable. We have had people in #plan9 ask if they can't
> run on a 486 or such, but we had to turn them down because they can't
> run the installer on such system(not enough RAM).
> ...

You can put lots of memory on a 486 if you want, depends on
the motherboard I suppose. We had a 486 (AMI Enterprise III
motherboard) with 128MB memory which was the video server for
Inferno.

Re: [9fans] 386

[Ronald G Minnich]

jmk@plan9.bell-labs.com wrote:

> I'd actually like to go further and only support processors
> that have the CPUID instruction (that includes the plain
> Pentium otherwise my home machine would need to be replaced).

I'm for it. The 386 is dead or should be. Even the crappy embedded stuff 
nowadays is at minimum a 486, and usually newer.

Old hardware, let it RIP. Or use it for door stops, or performance art.

ron

Re: [9fans] 386

[Francisco J Ballesteros]

I´d say that it´d be even reasonable to assume a pentium at least.


On 10/29/05, Ronald G Minnich <rminnich@lanl.gov> wrote:
> jmk@plan9.bell-labs.com wrote:
>
> > I'd actually like to go further and only support processors
> > that have the CPUID instruction (that includes the plain
> > Pentium otherwise my home machine would need to be replaced).
>
> I'm for it. The 386 is dead or should be. Even the crappy embedded stuff
> nowadays is at minimum a 486, and usually newer.
>
> Old hardware, let it RIP. Or use it for door stops, or performance art.
>
>

Re: [9fans] 386

[John DeGood]

 > Old hardware, let it RIP. Or use it for door stops, or performance art.

Here in the US, vintage Pentium systems are essentially free.  Several 
months ago I purchased a bundle consisting of a 300Mhz Pentium II, 440LX 
motherboard, and 64MBytes SDRAM on eBay for a penny (shipping was $8.)

The 60 MHz Pentium was introduced in March 1993.  If we drop 386/486 and 
require at least a Pentium, Plan 9 will still run on virtually every 
desktop and laptop sold in the past 10 years.

John

Re: [9fans] 386

[William Josephson]

On Sat, Oct 29, 2005 at 07:45:50PM -0400, John DeGood wrote:
> > Old hardware, let it RIP. Or use it for door stops, or performance art.
> 
> The 60 MHz Pentium was introduced in March 1993.  If we drop 386/486 and 
> require at least a Pentium, Plan 9 will still run on virtually every 
> desktop and laptop sold in the past 10 years.

The concern has for sometime been embedded hardware, not
desktop hardware.  But my impression is that most embedded
applications have shifted away even from the 486 as the
various Pentium clones have become more popular.

Re: [9fans] 386

[Ronald G Minnich]

Francisco J Ballesteros wrote:
> I´d say that it´d be even reasonable to assume a pentium at least.

well, you'll break all the amd sc520's out there, and geodes (maybe), so 
that's my only worry with that one.

ron

Re: [9fans] 386

[geoff]

I haven't run Plan 9 on 386s in a long time.  Before I retired my 1994
486 (and I can't recall precisely when that was), I was having trouble
running Plan 9 on it, though that might have been due to problems with
VLB support when trying to use it as a terminal.

Re: [9fans] 386

[Jack Johnson]

On 10/29/05, Ronald G Minnich <rminnich@lanl.gov> wrote:
> Francisco J Ballesteros wrote:
> > I´d say that it´d be even reasonable to assume a pentium at least.
>
> well, you'll break all the amd sc520's out there, and geodes (maybe), so
> that's my only worry with that one.

It would also be good to ensure that we don't break all the Via C3s
and the (possibly incomplete) emulated Pentium in Bochs, QEMU, etc.

-Jack

Re: [9fans] 386

[Brantley Coile]

> Old hardware, let it RIP. Or use it for door stops, or performance art.
> 
> ron

my door stop is a Sparc Java Station.  i never got java to work on it,
but i did get inferno to boot and used it for that for a while.  makes
a humorour item that most guest notice. gray and purple ½ shoe
box with that  coffee cup logo.

we've nothing that runs 386 here.

  bc
  1011 1100

Re: [9fans] 386

[Ronald G Minnich]

Jack Johnson wrote:

> 
> It would also be good to ensure that we don't break all the Via C3s
> and the (possibly incomplete) emulated Pentium in Bochs, QEMU, etc.

oh yes,  C3s don't have big pages.

ron

Re: [9fans] 386

[jmk]

Seems to me the concensus on this list is that we
should keep the support so that Plan 9 will continue
to run on all the broken non-Intel hardware clones
and emulators.

Re: [9fans] 386

[Uriel]

On Sun, Oct 30, 2005 at 03:02:34PM -0500, jmk@plan9.bell-labs.com wrote:
> Seems to me the concensus on this list is that we
> should keep the support so that Plan 9 will continue
> to run on all the broken non-Intel hardware clones
> and emulators.
Do really those emulators and clones require 386 support? From this
thread my conclusion was that it is not an issue.

I seriously doubt requiring Pentium would be a problem for most
emulators or recent embedded hardware.

Hell, qemu(the only emulator that matters this days as far as Plan 9 is
concerned) can even emulate x86_64:

http://fabrice.bellard.free.fr/qemu/status.html

uriel

Re: [9fans] 386

[Devon H. O'Dell]

2005/10/30, Uriel <uriell@binarydream.org>:
> On Sun, Oct 30, 2005 at 03:02:34PM -0500, jmk@plan9.bell-labs.com wrote:
> > Seems to me the concensus on this list is that we
> > should keep the support so that Plan 9 will continue
> > to run on all the broken non-Intel hardware clones
> > and emulators.
> Do really those emulators and clones require 386 support? From this
> thread my conclusion was that it is not an issue.

They don't. FreeBSD without 386 support runs in VMWare and qemu. Bochs
is dying as far as I can tell. I feel that removing support for 386 is
fine, and my understanding after reading this thread is that we're
worried about embedded platforms, and it seems that most of these that
we run on are already 486 or higher.

With regard to the instruction set clones, I've run post-386-debacle
FreeBSD (and seen it run) on several clone boards including the Via
embedded stuff and Geode stuff.

My understanding of how this got into the embedded chipset stuff was
that a suggestion was made to deprecate 486 as well, which I don't
think we should do at all. Get rid of 386, support 486+.

> I seriously doubt requiring Pentium would be a problem for most
> emulators or recent embedded hardware.

It is a big problem for recent embedded hardware. Until the
pentium-based embedded boards get cheaper, it's still nice to have
cheap-ish 486 boards, and these will _not_ go anywhere for a while,
especially since their price will fall more if the pentium boards get
cheaper.

> Hell, qemu(the only emulator that matters this days as far as Plan 9 is
> concerned) can even emulate x86_64:
>
> http://fabrice.bellard.free.fr/qemu/status.html
>
> uriel

--Devon

Re: [9fans] 386

[jmk]

On Sun Oct 30 15:34:52 EST 2005, devon.odell@gmail.com wrote:
> ...
> My understanding of how this got into the embedded chipset stuff was
> that a suggestion was made to deprecate 486 as well, which I don't
> think we should do at all. Get rid of 386, support 486+.
> ...

If you keep the 486 then there is no point in removing the 386,
you would gain nothing.

Re: [9fans] 386

[Devon H. O'Dell]

2005/10/30, jmk@plan9.bell-labs.com <jmk@plan9.bell-labs.com>:
> On Sun Oct 30 15:34:52 EST 2005, devon.odell@gmail.com wrote:
> > ...
> > My understanding of how this got into the embedded chipset stuff was
> > that a suggestion was made to deprecate 486 as well, which I don't
> > think we should do at all. Get rid of 386, support 486+.
> > ...
>
> If you keep the 486 then there is no point in removing the 386,
> you would gain nothing.

Huh? I thought we would gain the use of the page invalidation instruction?

Re: [9fans] 386

[Francisco J Ballesteros]

Why not keep the 386 as a separate architecture?

On 10/30/05, Ronald G Minnich <rminnich@lanl.gov> wrote:
> jmk@plan9.bell-labs.com wrote:
>
> > If you keep the 486 then there is no point in removing the 386,
> > you would gain nothing.
>
> Sorry, I was not looking carefully enough. I guess we're stuck ... I
> don't think we want to lose the embedded world.
>
>

Re: [9fans] 386

[Mike Haertel]

jmk wrote:
>If you keep the 486 then there is no point in removing the 386,
>you would gain nothing.

Just to clarify exactly what the differences are, the 486 added:

* Integrated floating point (and better IEEE 754 conformance)

* CR0.WP control bit -- allows copy-on-write paging strategies
  even for accesses from ring 0  (386 didn't honor page write
  protection when CPU was in kernel mode)

* New instructions:
  * BSWAP
  * CMPXCHG
  * XADD
  * INVD
  * WBINVD

I agree with jmk that Plan 9 would be unlikely to gain much, since:

* Plan 9 on the 386 already requires a 387, and Plan 9 is not
  pedantic about IEEE 754

* Historically Plan 9 has not been concerned with the kind
  of pedal-to-the-metal optimization that would benefit
  from CR0.WP, BSWAP, CMPXCHG, or XADD.

* There is rarely any good reason to use WBINVD, and almost
  never a good reason to use INVD (except maybe in the BIOS),
  since x86 systems are cache coherent.

Re: [9fans] 386

[Ronald G Minnich]

jmk@plan9.bell-labs.com wrote:
> Seems to me the concensus on this list is that we
> should keep the support so that Plan 9 will continue
> to run on all the broken non-Intel hardware clones
> and emulators.

how depressing!

I hope that I did not help cause that decision. I would go for requiring 
invlpg and cpuid support and see if anything breaks, and take it from there?

ron

Re: [9fans] 386

[Ronald G Minnich]

jmk@plan9.bell-labs.com wrote:

> If you keep the 486 then there is no point in removing the 386,
> you would gain nothing.

Sorry, I was not looking carefully enough. I guess we're stuck ... I 
don't think we want to lose the embedded world.

ron

Re: [9fans] 386

[jmk]

Sorry, I rounded the effect down. You get to remove the
following from the kernel if you drop the 386:

	{ 3,	-1,	32,	"386", },	/* family defaults */
and 
	if(X86FAMILY(m->cpuidax) == 3)
		conf.copymode = 1;

and to simplify this:

		if(read && conf.copymode == 0 && s->ref == 1) {

but you get to add code to check and whine if you try to
run the kernel on a 386.

On Sun Oct 30 16:16:32 EST 2005, devon.odell@gmail.com wrote:
> 2005/10/30, jmk@plan9.bell-labs.com <jmk@plan9.bell-labs.com>:
> > On Sun Oct 30 15:34:52 EST 2005, devon.odell@gmail.com wrote:
> > > ...
> > > My understanding of how this got into the embedded chipset stuff was
> > > that a suggestion was made to deprecate 486 as well, which I don't
> > > think we should do at all. Get rid of 386, support 486+.
> > > ...
> >
> > If you keep the 486 then there is no point in removing the 386,
> > you would gain nothing.
> 
> Huh? I thought we would gain the use of the page invalidation instruction?

Re: [9fans] 386

[Ronald G Minnich]

jmk@plan9.bell-labs.com wrote:

> but you get to add code to check and whine if you try to
> run the kernel on a 386.

well, hell, what's wrong with a good whine now and then ...

I think the 'new architecture' maybe was the right way to go? I'm just 
not sure it's a good idea to screw the 486 systems over. I wish they 
would all go away, but they won't; I found to my horror I was using one 
and didnt' even know it!

ron

Re: [9fans] 386

[Russ Cox]

On reflection it's just not that hard to write a C invlpg
that calls putcr3 if it's running on a 386.  We use
LOCK INCL to do ref counts so that it works on the 386,
we never use BSWAP, and we use CR0.WP by checking
the cpuid.

For actually running a system, VMware and Qemu are
pretty much neck and neck and Bochs is far behind.

For debugging a system, Bochs is slightly ahead of
Qemu and VMware is far behind.  Bochs has a reasonable
debugger interface that knows about 16-bit mode and
segments and offsets and page tables and everything.
Qemu is much faster but has no debugger.  You can attach
to it with gdb but gdb only knows about 32-bit mode user-level
sorts of things, so usually I have to set a break point using gdb
and then dump register info back at qemu's console.
I lost count of the number of times one or the other got me out
of a jam while I was writing the new VM code or the realmode
code that made vesa-in-the-kernel possible.

Russ

Re: [9fans] 386

[Charles Forsyth]

>Just to clarify exactly what the differences are, the 486 added:
> * Integrated floating point (and better IEEE 754 conformance)

except for 486sx!

Re: [9fans] 386

[Dave Eckhardt]

> For debugging a system, Bochs is slightly ahead of
> Qemu and VMware is far behind.

There is also simics, a very faithful but proprietary
multi-platform simulator with an integrated debugger.
Academic licenses are essentially free, though.

Dave Eckhardt