delay loop calibration failed for 2.6.32 guest kernels on Intel Atom CPU 330

[Ralph Schmieder]

Going from CentOS5 to CentOS6 and the kernel change from 2.6.18 to 2.6.32 breaks the timer calibration and makes the system almost totally unusable (also on other 2.6.32 or newer based distros).

There's no VT-x on the Atom so HPET is turned off. However, I can observe the effect also on a Mac Mini w/ VirtualBox 4. The impact is not as dramatic as the Core2Duo on the Mac provides more performance to begin with. But the delay loop calibration on the Mac is also off (and not showing the value 2.6.18 shows).

I'm running VirtualBox 4.1.4 on a 64-Bit Ubuntu 10.04 w/ ~4G of DRAM and an Intel Atom 330 Dual Core CPU w/ HT

Please see the additional log files with cpuinfo etc on the virtual and the physical systems.

On 2.6.32 for the guest:

Fast TSC calibration failed
TSC: Unable to calibrate against PIT
TSC: HPET/PMTIMER calibration failed.
Marking TSC unstable due to could not calculate TSC khz
Calibrating delay loop... 293.88 BogoMIPS (lpj=146944)
Initializing cgroup subsys cpuacctmce: CPU supports 0 MCE banks
Performance Events: unsupported p6 CPU model 28 no PMU driver, software events only.
weird, boot CPU (#0) not listed by the BIOS.
APIC calibration not consistent with PM-Timer: 247ms instead of 100ms
APIC delta adjusted to PM-Timer: 6254153 (15508853)
Total of 1 processors activated (293.88 BogoMIPS).
Switching to clocksource jiffies
Switching to clocksource acpi_pm
hrtimer: interrupt took 3406704 ns

On 2.6.18 for the guest:

Calibrating delay loop (skipped), value calculated using timer frequency.. 3200.05 BogoMIPS (lpj=1600025)
Intel machine check reporting enabled on CPU#0.
CPU0: Intel(R) Atom(TM) CPU  330   @ 1.60GHz stepping 02
Using local APIC timer interrupts.
WARNING calibrate_APIC_clock: the APIC timer calibration may be wrong.
Brought up 1 CPUs
ACPI: (supports S0 S1 S4 S5<6>Time: acpi_pm clocksource has been installed.

Thanks for looking into this! -ralph

[Ralph Schmieder]

VBox.log, configuration of VM (2.6.32), log files of virtual hosts and physical hosts

[Ralph Schmieder]

Running Linux on VMware (Fusion, different Machine in this case) produces the following log messages in dmesg during CPU initialization on kernel start:

TSC freq read from hypervisor : 2653.482 MHz
Calibrating delay loop (skipped) preset value.. 5306.96 BogoMIPS (lpj=10613928)

The kernel obviously uses some mechanism to realize it's running on virtualized h/w and pulls some additional info thru some hypervisor API.

​Here's a link to the relevant post on LKML, it's quite date and VirtualBox could use the same / similar mechanism as the original poster mentioned: "This patch also adds a hypervisor_get_tsc_freq function, instead of calibrating the frequency which can be error prone in virtualized environment, we ask the hypervisor for it. We get the frequency from the hypervisor by accessing the backdoor port if we are running on VMware. Other hypervisors too can add code to get frequency on their platform to this routine. "

[Ralph Schmieder]

Upgraded everything to latest and greatest

• Host OS Linux asrock 2.6.32-36-server #79-Ubuntu SMP Tue Nov 8 22:44:38 UTC 2011 x86_64 GNU/Linux

• Virtualbox 4.1.6r74727

• Guest OS Linux centos6 2.6.32-131.21.1.el6.i686 #1 SMP Tue Nov 22 18:21:07 GMT 2011 i686 i686 i386 GNU/Linux

and it's still the same phenomena. It can be also observed on a Mac Mini w/ an Core2Dua which gives a lot more horsepower compared to the Atom 330 (and thus it's not that obvious) but still can be seen. Was wondering if this is going to be addressed eventually?

[Ralph Schmieder]

w/ current VBox version (4.1.20)

[Ralph Schmieder]

w/ current VBox version (4.1.20) on different hardware (Mac Mini w/ Intel C2D)

[Ralph Schmieder]

Upgraded host, guest install ISO and VBox to latest available versions. With a vanilla CentOS 6.3 install disk / ISO and given the config in info.tar the following boot error can be seen:

------------[ cut here ]------------
WARNING: at arch/x86/kernel/apic/apic.c:1304 setup_local_APIC+0x236/0x355() (Not
 tainted)
Hardware name: VirtualBox
Modules linked in:
Pid: 1, comm: swapper Not tainted 2.6.32-279.el6.i686 #1
Call Trace:
 [<c0455c11>] ? warn_slowpath_common+0x81/0xc0
 [<c0838bae>] ? setup_local_APIC+0x236/0x355
 [<c0838bae>] ? setup_local_APIC+0x236/0x355
 [<c0455c6b>] ? warn_slowpath_null+0x1b/0x20
 [<c0838bae>] ? setup_local_APIC+0x236/0x355
 [<c0ab9c40>] ? APIC_init_uniprocessor+0x9b/0x103
 [<c0ab75d9>] ? disable_smp+0x63/0x83
 [<c0ab776e>] ? native_smp_prepare_cpus+0x175/0x346
 [<c0aa72b9>] ? kernel_init+0x69/0x23a
 [<c0aa7250>] ? kernel_init+0x0/0x23a
 [<c0409fff>] ? kernel_thread_helper+0x7/0x10
---[ end trace a7919e7f17c0a725 ]---
APIC calibration not consistent with PM-Timer: 256ms instead of 100ms
APIC delta adjusted to PM-Timer: 6251117 (16060050)

Not seen in the same configuration on a different host (VBox 4.1.20 on a Mac Mini with Intel Core2 Duo CPU). But the Timer calibration seems to be way off on that platform as well even though this box has HardwareVirtEx (and the Atom box does not).

I would really like to see at least some confirmation here since incorrect timer calculation potentially affects a lot more users...

Thanks, -ralph

[Ralph Schmieder]

Please also note my edit on comment #1 above... with a link to existing code for timer calibration w/ Linux for the VMware hypervisor.

[Frank Mehnert]

You are right, this patch adds some hypervisor-specific code to the Linux kernel to make the detection of the TSC frequency more reliable and actually this is the only way to fix this problem. On the other hand, the code you mention is specific to VMware and emulating this functionality would mean to make VirtualBox appear a bit like VMware, ie. implement at least parts of the VMware backdoor. This can have negative consequences to other software which now could think that it runs straight on VMware and could use certain features of the backdoor which are not implemented.

I think we have to go this way sooner or later but this needs a lot of tests.