Re: timecounter changes break netbsd-9/i386 for NET4501

["John D. Baker" <jdbaker%consolidated.net@localhost>]

On Sun, 26 Jul 2020, John D. Baker wrote:

> On Fri, 24 Jul 2020, John D. Baker wrote:
> 
> > [   1.0000000] NetBSD 9.0_STABLE (NET4501_JDB) #79: Sat Jul 18 08:19:53 CDT 2020
> > [   1.0000000] 	sysop%plex760.technoskunk.fur@localhost:/r0/build/netbsd-9/obj/i386/sys/arch/i386/compile/NET4501_JDB
> > [   1.0000000] total memory = 65148 KB
> > [   1.0000000] avail memory = 60628 KB
> > [   1.0000050] mainbus0 (root)
> > [   1.0000050] cpu0 at mainbus0
> > [   1.0000050] fatal privileged instruction fault in supervisor mode
> > [   1.0000050] trap type 0 code 0 eip 0xc0114c30 cs 0xc0430008 eflags 0x10246 cr2 0 ilevel 0x8 esp 0xc011f0cb
> > [   1.0000050] curlwp 0xc0351960 pid 0 lid 1 lowest kstack 0xc04342c0
> > kernel: supervisor trap privileged instruction fault, code=0
> > Stopped in pid 0.1 (system) at  netbsd:cpu_counter:     rdtsc
> 
> So, the pull-up of x86 TSC changes to netbsd-9 needs some more attention.

I think the problem lies with:

  cvs rdiff -u -r1.171.2.1 -r1.171.2.2 src/sys/arch/x86/x86/cpu.c

The entire code block following the "if (cpu_hascounter())" should
be in braces.  Otherwise when "freq != 0" is found to be false, the
"else" clause unconditionally calls "cpu_counter_serializing()" which
calls "cpu_counter()" which attempts to execute the "rdtsc" instruction
which causes the above fault on Soekris net4501.

"cpu_get_tsc_freq()" in -current is significantly more sophisticated,
but the code paths that could call "cpu_counter()" are protected by
testing the result of "cpu_hascounter()".

-- 
|/"\ John D. Baker, KN5UKS               NetBSD     Darwin/MacOS X
|\ / jdbaker[snail]consolidated[flyspeck]net  OpenBSD            FreeBSD
| X  No HTML/proprietary data in email.   BSD just sits there and works!
|/ \ GPGkeyID:  D703 4A7E 479F 63F8 D3F4  BD99 9572 8F23 E4AD 1645