The Linux Kernel Disabling HPET For More Platforms - Including Ice Lake

[phoronix]

Phoronix: The Linux Kernel Disabling HPET For More Platforms - Including Ice Lake

Reported on earlier this month is the decision by Linux kernel developers to disable HPET for Intel Coffee Lake systems. The High Precision Event Timer was being disabled since on some Coffee Lake systems at least this timer skews when entering the PC10 power state and that makes the time-stamp counter unstable...

http://www.phoronix.com/scan.php?page=news_item&px=Linux-Disabling-HPET-H-Icelake

[polarathene]

This is believed to be a firmware issue rather than a hardware problem, but the best course of action was just disabling HPET for Coffee Lake systems to avoid the problem.

Reminds me of my system I built in 2016, it's using Skylake, but whenever I did a suspend and resumed, the TSC clocksource became unstable iirc, and this caused big perf issues in my VFIO VM with Windows becoming a snail, can't recall if it was falling back to HPET or something else(HPET might have been disabled). I contacted ASRock about it and they sent me a beta firmware update(no idea if it ever got officially released), fixed the issue for me. I just assumed it wasn't a common issue.

[ThoreauHD]

Drink!

[jbennett]

What happened to Intel? Have they always had these problems, or has their bug count actually skyrocketed recently?

[duby229]

What happened to Intel? Have they always had these problems, or has their bug count actually skyrocketed recently?

all ICs have bugs, it's just that Intel makes some of the most complex ones in the world and they're getting more complex each new generation. AMD and ARM cores have bugs too, remember the first thoroubred or the first bulldozer, or on the graphics side the first GCN or the first vega. Recalls are extremely rare, so the common course of action is to patch firmware for affected revisions and then release fixed revisions.

[r08z]

To honest HPET has been problematic since the days of Intel Core 2 Duo CPUs. If enabled at kernel an annoying bug is often guaranteed slowing down your machine when doing a system suspend, hibernation, and resume - or can just occur if a game is left running on idle and the screen saver kicks in - dmesg reports something like:

hpet increased min_delta_ns to 20115 nsec

When the kernel is configured with 1000HZ/PREEMPT the effects of the bug are reduced but I've personally disabled this feature all together for years now. If you don't want to recompile your kernel just disable it at kernel boot line with:

hpet=disable

[milkylainen]

When the kernel is configured with 1000HZ/PREEMPT the effects of the bug are reduced but I've personally disabled this feature all together for years now. If you don't want to recompile your kernel just disable it at kernel boot line with:

hpet=disable

Well. Depends on what you're looking for. To me, this looks like a blanket statement.
HPET has more overhead and is generally slower than RDTSC.
But depending on the precision on the required event interrupt, the HPET interface might be the only sane choice.
So. Again. Depends what you want to do.
A broken HPET is a bad thing.

Edit: I should also add that the kernel usually selects TSC over HPET. If it defaults to HPET it probably depends on a funky TSC.