Re: NVMM not working, NetBSD 9x amd64

[Chavdar Ivanov <ci4ic4%gmail.com@localhost>]

Hi,

I decided to reuse this thread; nvmm again ceased to work from yesterday.

On

# uname -a
NetBSD ymir 9.99.69 NetBSD 9.99.69 (GENERIC) #15: Tue Jul 14 11:07:52
BST 2020  sysbuild@ymir:/home/sysbuild/amd64/obj/home/sysbuild/src/sys
/arch/amd64/compile/GENERIC amd64

I can 'modload nvmm', but when I try to start a vm with nvmm
acceleration, I get a hard lock, immediately after the message about
the interface being initialized. I cannot break into the debugger to
trace and I don't get a dump on reboot. It appears the machine is in a
deep CPU loop, although it doesn't appear too  hot.

I then tried booting onetbsd, which is from the 12th of July and on
which nvmm used to work just fine. It is also the same micro version -
9.99.59, so n theory should work - but in this case I get a panic when
I 'modload nvmm' - again, I see the short panic message on the screen
and the machine apparently gets into another loop here, which I cannot
break the usual way into the debugger and the only thing I can do is
hit the power button. There weren't that many kernel changes in this
period, most notably the per-CPU IDT patch, but I don't know if it is
relevant.

Chavdar

On Wed, 20 May 2020 at 22:09, Maxime Villard <max%m00nbsd.net@localhost> wrote:
>
> Le 09/05/2020 à 10:54, Maxime Villard a écrit :
> > Le 01/05/2020 à 19:13, Chavdar Ivanov a écrit :
> >> On Fri, 1 May 2020 at 13:59, Rhialto <rhialto%falu.nl@localhost> wrote:
> >>>
> >>> On Sun 26 Apr 2020 at 21:39:12 +0200, Maxime Villard wrote:
> >>>> Maybe I should add a note in the man page to say that you cannot expect a CPU
> >>>> from before ~2010 to have virtualization support.
> >>>
> >>> Or even better, what one should look for in the output of, for example,
> >>> "cpuctl identify 0". Since I didn't exactly know, I made some guesses
> >>> and assumed that my cpu ("Intel(R) Core(TM) i3-2120 CPU @ 3.30GHz")
> >>> did't have the required features (it is from 2009 or so).  But this
> >>> thread inspired me to modload nvmm, which actually helped, so I found
> >>> out that it even works on this cpu.
> >
> > On Intel CPUs the information is hidden in privileged registers that cpuctl
> > cannot access, so no, it won't be possible.
> >
> > However the day before I had added clear warnings:
> >
> >      https://mail-index.netbsd.org/source-changes/2020/04/30/msg116878.html
> >
> > So now it will tell you what's missing.
> >
> >>> Of course I immediately tried it with Haiku (the BeOS clone) from
> >>> https://download.haiku-os.org/nightly-images/x86_64/ and I got mixed
> >>> results. Once it manages to boot it works fine and nicely fast (much
> >>> better than without nvmm), but quite often it crashes into its kernel
> >>> debugger during the first 10 seconds of booting, with different messages
> >>> (I have seen "General Protection Exception" and "ASSERT failed ...
> >>> fCPUCount >= 0").  ("qemu-system-x86_64 -accel nvmm -m 2G -cdrom
> >>> haiku-master-hrev54106-x86_64-anyboot.iso" on a 9.0 GENERIC kernel)
> >
> > This was a missing filtering in the CPU identification, on CPUs that have SMT,
> > leading Haiku to believe it had SMT threads that it didn't.
> >
> >      https://mail-index.netbsd.org/source-changes/2020/05/09/msg117188.html
> >
> > As far as I can tell, your CPU has SMT.
> >
> >> I've never used Haiku so far; upon reading this I decided to try it on
> >> my NetBSD-current laptop with nvmm.
> >>
> >> So far, with several attempts, it works with no problem whatsoever,
> >> directly booting the newest image on the site pointed above.
> >>
> >> Another OS to play with...
> >>
> >> The host cpu is Intel(R) Core(TM) i7-3820QM CPU @ 2.70GHz, id 0x306a9.
> >
> > This CPU too has SMT.
> >
> > Le 01/05/2020 à 20:10, Rhialto a écrit :
> >> There might well be an improvement between 9.0 and -current, of course.
> >> It's good to hear that it works for you; I might upgrade to a -current
> >> kernel.
> >
> > Overall, no, each improvement in -current is propagated to 9, so you should
> > get the same results on both (modulo kernel bugs added in places not
> > related to NVMM).
> >
> > Le 01/05/2020 à 20:52, Chavdar Ivanov a écrit :
> >> Earlier I had similar issues with OmniOS under qemu-nvmm - sometimes
> >> it worked without a problem, sometimes I couldn't even boot. I still
> >> have no idea why.
> >
> > Maybe that's the same problem, I'll test.
>
> I tested the other day, and I saw no problem. With debugging I noticed that
> OmniOS, too, uses the CPU information that used to be mis-reported by NVMM,
> so probably my fix must have helped.
>
> Please confirm the issues are fixed (HaikuOS+OmniOS).



--
----