Re: amd ryzen PR 52966

[Lars Reichardt <lars%paradoxon.info@localhost>]

* Maxime Villard <max%m00nbsd.net@localhost>:
> Le 08/08/2018 à 10:31, Lars Reichardt a écrit :
> > I've run an updated kernel yesterday with the new (AMD) errata fixes.
> > (src/sys/arch/x86/x86/errata.c rev 1.24)
> > No patches got applied at least nothing was indicated via dmesg I
> > assume the firmware did apply those fixes or the zen+ has them fixed.
> 
> To see the messages you need to boot with "boot netbsd -v -x". Not all
> errata can be fixed by the bios, some are marked in the spec as "no
> fix planned".
 
right, my blindness ;-)
1021, 1049, 1091 and 1095 show up as "testing for erratum" but not as
"erratum patched" seems those bits were already setup and those tested
for are right for a "zen+" cpu.

> > src/sys/arch/x86/x86/identcpu.c got patched to use FPU_SAVE_XSAVEOPT if
> > present and if present switch to eager fpu context switching.
> > 
> > With this change present I run the stream benchmark noted in
> > PR port-amd64/52966
> > without any problems on a Ryzen 2700.
> > Several runs while doing a release build as well as on an otherwise
> > idle machine. (All resulting in: "Solution validates")
> 
> Can you test with EagerFPU disabled? Originally, in the PR, Michael was
> using !EagerFPU.

With EagerFPU disable via sysctl I did four runs on a otherwise idle
machine and four runs while a release build was running.
All runs resulted in: "Solution validates"