Re: [Bug #11516] severe performance degradation on x86_64 going from 2.6.26-rc9 -> 2.6.27-rc5

[Thomas Gleixner <tglx@linutronix.de>]

On Sun, 28 Sep 2008, Jason Vas Dias wrote:
> The problem is not "slow booting" - indeed, as noted in the bug, boot time up to the point
> where user space should start is very similar to the non-problematic 2.6.26 kernels.
> 
> The USB problem is that @ 10 messages per second are emitted to the log:
> Sep 28 00:41:51 localhost kernel: [   11.985257] hub 3-0:1.0: state 7 ports 2 chg 0000 evt 0002
> Sep 28 00:41:51 localhost kernel: [   12.789240] hub 1-0:1.0: state 7 ports 2 chg 0000 evt 0004
> Sep 28 00:41:51 localhost kernel: [   12.937169] hub 4-0:1.0: state 7 ports 2 chg 0000 evt 0000
> Sep 28 00:41:51 localhost kernel: [   12.937188] hub 3-0:1.0: state 7 ports 2 chg 0000 evt 0002
> Sep 28 00:41:51 localhost kernel: [   13.041167] hub 5-0:1.0: state 7 ports 2 chg 0000 evt 0000
> Sep 28 00:41:51 localhost kernel: [   13.041186] hub 6-0:1.0: state 7 ports 10 chg 0000 evt 0000
> Sep 28 00:41:51 localhost kernel: [   13.453734] hub 6-0:1.0: state 7 ports 10 chg 0000 evt 0020
> Sep 28 00:41:51 localhost kernel: [   13.453748] hub 1-0:1.0: state 7 ports 2 chg 0000 evt 0004
> Sep 28 00:41:51 localhost kernel: [   13.846512] hub 3-0:1.0: state 7 ports 2 chg 0000 evt 0002
> Sep 28 00:41:51 localhost kernel: [   14.220994] hub 3-0:1.0: state 7 ports 2 chg 0000 evt 0002
> 
> Why should USB be emitting state change messages when no state has changed ?

That's a separate question.

> I think this is because the core PM-timer (clocksource is acpi-pm) is kaput ,
> and every driver that depends on high-resolution timers gets confused .

No, if the pmtimer would be defect your machine would not even reach
user space.

I analysed the proc/acpi data of .25 and .27 and the machine is set to
throttling state T7 (12%) which would explain that behaviour halfways.

Can you please verify if that state is always T7 on your machine with
.27 ?

Also please add "acpi.debug_level=0x11" to the kernel command line so
we can get some more information out of the ACPI code.

> Displaying text on the terminal is @ 140 times slower under 2.6.27
> kernels than under 2.6.26 kernels .  Downloads over the network take
> @ 1000 times longer under 2.6.27 than under 2.6.26.  Disk I/O is @
> 100 times slower.

That's indeed insane.
 
> There are numerous log messages about "interrupt during system call"
> (EINTR) from user-space processes.

Separate problem as well.

Please let us concentrate on the throttling aspect and not mix
USB/EINTR stuff into it for now.

Thanks,

	tglx