IRC channel logs

<damo22> what about #ifdef APIC to read hpet for fine tuning?

<damo22>i think we can decide if we expect a HPET at compile time

<damo22>thats what i was getting at in my mail

<damo22>if APIC is defined and there is no HPET, none of the clock interrupts will even work for APs

<damo22>and it will break on an assert anyway

<damo22>hmm we use PIT to calibrate LAPIC timer, and then the LAPIC timers are enabled on all APs, but i think we are still using PIT to make clock interrupts on BSP

<damo22>and the HPET is enabled just for high resolution timer scheduling

<damo22>?

<damo22>or for delays in kernel

<damo22>set_timeout does not use HPET

<damo22>so HPET is only being used in kernel for hpet_udelay etc

<damo22>hpet is still critical to smp timing

<damo22>so yes, we assume its present on smp systems

<ZhaoM>morning

<sneek>Welcome back ZhaoM, you have 2 messages!

<sneek>ZhaoM, youpi says: it's not really useful to look at tsc. Nowadays most machines to have an hpet, so we will be fine with just that

<sneek>ZhaoM, azert says: reading the hpet takes ~100x more cycles then reading the tsc (source: https://docs.redhat.com/en/documentation/red_hat_enterprise_linux_for_real_time/7/html/reference_guide/chap-timestamping#Reading_hardware_clock_sources ). That’s why I mentioned the tsc. That said youpi is always right and possibly it’s a premature

<ZhaoM>sneek later tell azert thanks for the documentation

<sneek>Okay.

<ZhaoM>ok a small conclusion

<ZhaoM>tsc is indeed more efficient than hpet but I think hpet is still needed as an option since it seems some CPU may not have a TSC https://wiki.osdev.org/TSC#:~:text=Portability

<ZhaoM>I may give TSC a try if I'm motivated in the future

<ZhaoM>damo22: if the detection of HPET can be done automatically without any overhead with the help of the branch predictor, that will be very nice

<ZhaoM>I will do a simple benchmark just like the url shared by azert to see if that `if` affects the performance

<damo22>ZhaoM: my point was we may not need to do a runtime check at all, if we assume every APIC machine has HPET

<damo22>and just short circuit the code in the non-APIC case

<damo22>at compile time

<ZhaoM>about `APIC machine has HPET`, is it always the case? I just found this https://wiki.osdev.org/HPET#Detecting_HPET_using_ACPI

<ZhaoM>and yes we indeed init_hpet when APIC is available

<ZhaoM>but I think if runtime check is more flexible and does not cause significant overhead we can give it a go :)

<ZhaoM>conclusion ned

<ZhaoM>s/ned/end/

<damo22>ZhaoM: osdev has some useful info, but in practice every machine with acpi tables since 2005 has a HPET

<damo22>and i wrote init_hpet with that in mind

<damo22>why add unnecessary run time check every time you fine tune the clock?

<damo22>btw, i looked into using TSC but its a pain to calibrate that because it depends on cpu frequency

<damo22>there is a method to the madness

<ZhaoM>well I think you convinced me

<ZhaoM>ok I will use #ifdef APIC

IRC channel logs

2025-03-24.log