Bugzilla Bug 203826

A kernel built from sys-kernel/gentoo-sources-2.6.23-r3 or
sys-kernel/vanilla-sources-2.6.23.9 leads to an abysmal overall performance on
machines using HIL input devices (CONFIG_KEYBOARD_HIL=y in the kernel .config
file).
The HIL driver is known to put a lot of work on ksoftirqd, on this certain
machine ksoftirqd eats up about 60% of the avalable CPU time on an idle system.
This did not affect the performance with kernels built from the now masked
sys-kernel/hppa-sources, because the default nice level of ksoftirqd was set to
19, so under situations of high system load, ksoftirqd did only use 3-4% of the
CPU time.
Kernels built from vanilla-sources or gentoo-sources have a default nice level
of -5 for ksoftirqd which results in the situation that all other processes
have to share the remaining 40% of the total CPU power and the overall system
performance being cut down to less than the half of what it used to be.

Reproducible: Always

Steps to Reproduce:
1. Fetch a HPPA machine with HIL keyboard and/or mouse
2. Build a kernel from sys-kernel/hppa-sources (and HIL driver) to see it run
at nominal speed
3. Build a kernel from sys-kernel/gentoo-sources or sys-kernel/hppa-sources to
see it crawl




One can work around the performance loss by renice-ing the ksoftirqd process to
19 after the system has booted up, but the system startup process is _slow_
anyway.
In hppa-sources, the niceness of ksoftirqd was set in kernel/softirq.c via
set_user_nice(), but in recent versions of gentoo-sources and vanilla-sources
this has changed, and up to now, I did not find the corresponding snippet of
code. Possible locations are kernel/workqueue.c and kernel/kthread.c, but I am
unsure about this.

Created an attachment (id=139666) [details]
top showing high load and high priority of ksoftirqd

Created an attachment (id=139667) [details]
emerge --info

Created an attachment (id=139668) [details]
.config file of sys-kernel/vanilla-sources-2.6.23.9

In the meantime I worked out that the nice level of ksoftirqd is set in
kernel/kthread.c. If the nice level is changed, it is also applied to some
other threads, on this particular these are kthreadd, kseriod, kswapd0,
cifsoplockd, cifsdnotifyd and scsi_eh_0.

(In reply to comment #4)
Whoops, missed one word

> other threads, on this particular
machine
> these are kthreadd, kseriod, kswapd0,

Created an attachment (id=145054) [details]
Set nice level of 15 to ksoftirq when HP_MLC is set

Kernel team, please include this patch.

This set the nice level to 15 of ksoftirqd when CONFIG_HIL_MLC is set.

This must be done to avoid the HIL driver to take all the ressources. Known
problem^W feature of the driver.

Kernel team, please commit HIL-ksoftirqd-nice-level.patch into next
gentoo-sources patchset.

Have there been any efforts to submit this patch upstream?

(In reply to comment #8)
> Kernel team, please commit HIL-ksoftirqd-nice-level.patch into next
> gentoo-sources patchset.
> 

Are you the author of this patch?

Yep I'm the author of this patch.

I've submitted this upstream. Will update bug with the outcome.

Any status on your submission? Can you tell me where and how you submitted the
patch?

any news?

Please reopen when you've had a chance to follow up with upstream