In D35735#810600, @mav wrote:

On old systems when this was written, having global eventtimer and slow timecounter, ET_HW_LOCK was very congested, that is why I was trying to move sbinuptime() and everything else I could out of it. It should be less critical these days.

Are there any popular platforms without a per-CPU eventtimer?

If lock contention is still a problem somehow, there are some measures we could take to reduce the et_hw lock scope. For instance, cpu_idleclock() can do this:

spinlock_enter();
now = sbinuptime();
ET_HW_LOCK();
spinlock_exit();
...

though this will hurt single-threaded performance. It may also be possible to split getnextcpuevent() so that only callout events are synchronized. And, we should try to eliminate sources of frequent cross-CPU callout scheduling, which should reduce the use of et_hw mutex. I fixed a couple recently but did not try to do an exhaustive search, just fixed what showed up in profiling.

In D35735#810622, @markj wrote:

Are there any popular platforms without a per-CPU eventtimer?

Last I remember were about Gen2-3 Cores. There were still unreliable LAPIC timers, so they were falling back to HPET, which often could not provide enough per-CPU vectors. These days I don't know any. So I'd care about it mostly if it is cheap, just to not hurt somebody's relics too much. Though the global eventtimers support in general we'll probably keep for a while.