tainted Linux kernel in default SMP QEMU/KVM guests
Hi,
since the default CPU model for QEMU (qemu64) is an AMD K7, the Linux kernel complains and taints the kernel when it detects multiple processors. The reason for this is a check for SMP safe CPUs in arch/x86/kernel/cpu/amd.c:amd_k7_smp_check(). In recent kernels (since about 2.6.29) this in only executed in 32bit kernels, since the check has been #ifdef'ed CONFIG_X86_32. The failing check triggers a nasty dump along with a call trace and taints the kernel, this confuses users, so I want to get rid of it.
One can work around it (using -cpu kvm64 or -cpu host), but I want to avoid it in the default case, too.
I see these possible solutions:
1) Change the default CPUID bits from 6/2/3 to 6/6/1, this passes the Linux kernel check. But I am not sure if that would introduce regressions, since some OSes apply quirks if they detect certain models (like we had with the sysenter issue in the past)
2) Only change the CPUID bits to 6/6/1 if we use SMP. Still has the above drawback, but would be limited to SMP guests only.
3) Set kvm64/kvm32 as the default CPU model if KVM is enabled. This would limit the report and taint to TCG, where SMP is rarely used. Additionally less people (if any) use it for production systems.
4) Make the Linux' kernel quirk dependent on the missing hypervisor bit. I don't think this will be accepted easily upstream (and I don't want to support Ingo's recent ideas ;-), also this would not fix older kernels.
I can easily provide patches for all solutions, but I'd like to get advice from people on which one to pursue.
Regards,
Andre.
--
Andre Przywara
AMD-Operating System Research Center (OSRC), Dresden, Germany
Tel: +49 351 448-3567-12
Paolo Bonzini
maintainerQuoted 13 lines
1) Change the default CPUID bits from 6/2/3 to 6/6/1, this passes the Linux kernel check. But I am not sure if that would introduce regressions, since some OSes apply quirks if they detect certain models (like we had with the sysenter issue in the past)
2) Only change the CPUID bits to 6/6/1 if we use SMP. Still has the above drawback, but would be limited to SMP guests only.
3) Set kvm64/kvm32 as the default CPU model if KVM is enabled. This would limit the report and taint to TCG, where SMP is rarely used. Additionally less people (if any) use it for production systems.
Quoted 6 lines
4) Make the Linux' kernel quirk dependent on the missing hypervisor bit. I don't think this will be accepted easily upstream (and I don't want to support Ingo's recent ideas ;-), also this would not fix older kernels.
I can easily provide patches for all solutions, but I'd like to get advice from people on which one to pursue.
Doing (3) seems the most sensible thing to do, and it does not prevent doing (1) later on for TCG only.
Paolo
Quoted 13 lines · Paolo Bonzini
Quoted 13 lines
1) Change the default CPUID bits from 6/2/3 to 6/6/1, this passes the Linux kernel check. But I am not sure if that would introduce regressions, since some OSes apply quirks if they detect certain models (like we had with the sysenter issue in the past)
2) Only change the CPUID bits to 6/6/1 if we use SMP. Still has the above drawback, but would be limited to SMP guests only.
3) Set kvm64/kvm32 as the default CPU model if KVM is enabled. This would limit the report and taint to TCG, where SMP is rarely used. Additionally less people (if any) use it for production systems.
Quoted 8 lines
Quoted 6 lines
4) Make the Linux' kernel quirk dependent on the missing hypervisor bit. I don't think this will be accepted easily upstream (and I don't want to support Ingo's recent ideas ;-), also this would not fix older kernels.
I can easily provide patches for all solutions, but I'd like to get advice from people on which one to pursue.
Doing (3) seems the most sensible thing to do, and it does not prevent doing (1) later on for TCG only.
Let's switch to -cpu host for kvm.
Regards,
Anthony Liguori
Quoted 5 lines
Paolo
Signature · 3 lines
To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
Quoted 18 lines · Anthony Liguori
Quoted 17 lines · Paolo Bonzini
Quoted 17 lines
1) Change the default CPUID bits from 6/2/3 to 6/6/1, this passes the Linux kernel check. But I am not sure if that would introduce regressions, since some OSes apply quirks if they detect certain models (like we had with the sysenter issue in the past)
2) Only change the CPUID bits to 6/6/1 if we use SMP. Still has the above drawback, but would be limited to SMP guests only.
3) Set kvm64/kvm32 as the default CPU model if KVM is enabled. This would limit the report and taint to TCG, where SMP is rarely used. Additionally less people (if any) use it for production systems.
4) Make the Linux' kernel quirk dependent on the missing hypervisor bit. I don't think this will be accepted easily upstream (and I don't want to support Ingo's recent ideas ;-), also this would not fix older kernels.
Quoted 6 lines
Quoted 4 lines
Quoted 2 lines
I can easily provide patches for all solutions, but I'd like to get advice from people on which one to pursue.
Doing (3) seems the most sensible thing to do, and it does not prevent doing (1) later on for TCG only.
Let's switch to -cpu host for kvm.
Except for -M old.
This has the nice advantage of exposing new features as they are rolled out.
Signature · 1 line
error compiling committee.c: too many arguments to function