On 2020/04/22 4:49, Joerg Sonnenberger wrote:
On Fri, Apr 17, 2020 at 12:05:25AM +0900, SAITOH Masanobu wrote:Hi, all. On 2020/02/13 20:10, is%netbsd.org@localhost wrote:On Tue, Feb 11, 2020 at 05:34:58PM +0100, Frank Kardel wrote:clockinterrupt be 4 times too slow explains the high TSC frequency(*4) and ACPI-Fast and hpet0 seem the have the same frequency scaling as clockinterrupt. So it loock like the clockinterrupt setup is botched.Possibly related: * after a reboot, our (efi)boot doesn't count down. * after a single powerdown in this condition, boot counts down, netbsd kernel boots, but hangs after the internal ubt (last device) has attached. * needs a 2nd powerdown in this condition to boot again. (after shutdown -p and powerup , boots ok) -isPlease test the following (work-in-progrss) diff: http://www.netbsd.org/~msaitoh/tsc-20200416-2.dif I bought a new Intel NUC(BXNUC10I3FNK (Comet Lake U)) and it had the same problem.This seems to be quite the wrong direction. What seems to be the problem here is that the i8254 is jut fubar. It's what the use to calibrate all other timers again though. I don't like using the MSRs at all, they are very notorious for providing nonsense. I think what we can do is make sure that the ACPI timer or HPET is available much earlier and then use that one.
I agree to detect HPET much earlier. Please someone(TM) do it. I heard that future Intel chips have no legacy device support which include i8254 timer, so we must do it in (near?) future.
It is both faster and has a higher precision than the i8254. Joerg
--
-----------------------------------------------
SAITOH Masanobu (msaitoh%execsw.org@localhost
msaitoh%netbsd.org@localhost)