Comment 0 for bug 1417580

Revision history for this message
Rafael David Tinoco (rafaeldtinoco) wrote : HP Proliant Servers should use proper cmdline to avoid kernel panics

This bug will try to consolidate all HP Proliant Bugs related to kernel panics. Please do not use this to attach cores and/or files. Just to provide feedback on the cmdline and its explanations.

We had several talks (in the last 2 weeks) with HP ROM Engineering Team regarding NMIs (non maskable interrupts) being generated in some situations:

- NMIs caused during MWAIT instruction (caused by intel_idle module):

HP strongly uses ACPI for its power management features. HP is one of the most active members in the ACPI specification group and several features for their servers, available through their firmware, are heavily ACPI dependent. In the process of solving this and other bugs we have discovered that intel_idle module did not use ACPI tables (a way of firmware to say to OS what are the p-state/c-state values available) but queried processor directly for the available c-states. This is, probably, leading the OS to set a c-state (or sub-state) when the firmware is not "prepared" to handle. We have provided the following cmdline to be used: " intel_idle.max_cstate=0 ". This will tell OS to deactivate intel_idle and activate acpi_idle module, which gets c-state values to be used from the ACPI tables, given by firmware. HP is trying to figure out what is generating the NMIs with intel_idle but it might be the case to recommend all HP servers to deactivate intel_idle module (in a near future).

- Recently discovered NMIs caused by a BUG in Intel microcode

I've discovered that there is a recent microcode problem in some Intel Ivy Bridge microcode regarding a specific BIT not being cleared from the PMU (performance counter) register. This can lead to a NMI being wrongly handled (like if the PMU register was overflowed, without being) and a kernel panic. We have backported the fix to Ubuntu-3.13.0-35.61. So it is strongly advised that all Ubuntu Trusty Servers, running Xeon® Processor E7 v2, to be upgraded "at least" to kernel 3.13.0-35". The following cmdline: " nmi_watchdog=0 " can be used to disable regular x86 watchdog and use HP one but they don't recommend this to be the "default" cmdline. HP was advised by Canonical regarding Intel Errata # and that recommended workaround is a fix in firmware. Canonical has provided a kernel patch to "workaround" the issue in non-patched firmware (yet to be released by HP probably).

- X2APIC support for HP Proliant Servers

During this investigation we had to clarify with HP ROM Engineering Team whether this servers support X2APIC or not. It was told to us, by HP, that all Gen8 (and more recent generations) do support X2APIC but they still "ask" the OS to opt-out from X2APIC (not to use X2APIC). Running the PIC (programmable interrupt controller) in XAPIC mode might not be compatible with firmware if the CPU supports X2APIC because of one of the only features that differs XAPIC from X2APIC: IRQ remapping (for virtualization, basically). So it is recommended that on all HP Proliant Servers Gen8, or newer, to use the following cmdline: " intremap=no_x2apic_optout ".

Anyone affected, please provide proper feedback in this bug regarding the use of those cmdlines (and kernel version) and tell me if new kernel panics (regarding NMIs and/or APIC) happened on this Server Family. We are getting feedback from community that these options are being enough to avoid the Proliant Server Family to have kernel panics and they might be released as a "public recommendation" for HP HW compatibility soon.

Thank you

Rafael Tinoco