CentOS

Hello,

I have a X99 system with i7 5820K, which is used to run a emulator on X86; In the emulator the EPT is used to accelerate the address translating.
After switch to "3.10.0-693.11.6.el7.x86_64", the performance dropped 60-80% without any error from the EPT kernel module; If switch back to old kernel everything is back normal; And I tried to repalce the i7 5820K with i7 6800K, the problem is gone, there is no performance drop.

What could be the cause of the problem? Where should I look to solve it?

Thank you.

Try toggling the 3 switches that control the new code and see which one makes the most difference. Look at the files /sys/kernel/debug/x86/*enabled - should be 3 of them, ibpb, ibrs and pti_enabled. First run cat /sys/kernel/debug/x86/*enabled and see what the current values are. You can change them by using e.g. echo 0 > /sys/kernel/debug/x86/pti_enabled

If ibrs_enabled=1 on my work machines then it doubles cpu usage.

TrevorH wrote:Try toggling the 3 switches that control the new code and see which one makes the most difference. Look at the files /sys/kernel/debug/x86/*enabled - should be 3 of them, ibpb, ibrs and pti_enabled. First run cat /sys/kernel/debug/x86/*enabled and see what the current values are. You can change them by using e.g. echo 0 > /sys/kernel/debug/x86/pti_enabled

If ibrs_enabled=1 on my work machines then it doubles cpu usage.

Thanks a lot, finally find the real culprit. the performance drop is due to the ibrs_enable, disable it turns the performance back.
Tried 4790/4790K/5820K/6700K/6800K/E5 2690, only 5820K is affected, it is strange.

Check the output from dmesg on each of those processors and see which ones have microcode that support IBRS. I suspect that only the one that slows down is using the latest. Like this
That's on a VM and shows that there is no microcode that supports the IBRS mitigation. On a machine with fixed microcode it should say

[339488.642229] FEATURE SPEC_CTRL Present
[339488.642231] FEATURE IBPB_SUPPORT Present

TrevorH wrote: That's on a VM and shows that there is no microcode that supports the IBRS mitigation. On a machine with fixed microcode it should say

Yes, you are right, the 5820K is the only one with updated microcode, we are doomed. Updating the microcode will affect all systems.