Comment 0 for bug 1764956

Revision history for this message
Vincent Bernat (vbernat) wrote :

Hello!

As of Linux 4.4.0-119, when a KVM guest is using IBRS, this incurs a very large performance penalty on the hosts and other guests.

From my understanding, the patch f676aa34b4027d1a7a4bbcc58b81b20c68c7ce0c is incomplete. If host doesn't handle IBRS itself (which is now the case by default since 4.4.0-116: it relies on retpoline instead) but the guest does (eg running an earlier kernel), the guest will set IBRS for the CPU it is running on from time to time but if it gets preempted at some point, the IBRS bit will stay, incurring a major performance penalty for all other users of the CPU (host userland, host kernel and other guests not caring about IBRS). The equivalent patch in mainline (d28b387fb74da95d69d2615732f50cceb38e9a4d) ensure the appropriate MSR is correctly restored when switching from one guest to another or from one guest to host.

The issue is easy to reproduce: host running 4.4.0-119, exposing "spec_ctrl" to a guest running CentOS 7.4 with its January kernel. Wait a few minutes and the host will become pretty slow. A simple shell loop will take 10 more times to execute. Executing "sysctl -w kernel.ibrs_dump=1" will show that most real cores have now their IBRS bit set to 1.

A workaround is to reeanble IBRS on the host (sysctl -w kernel.ibrs_enabled=1). This way, IBRS will be correctly disabled when changing context.

A long term solution would be to properly backport the patch from mainline. It is not part of the 4.4 stable branch and it seems not trivial to port.

A mid term solution could be to remove the faulty patch (not exposing IBRS), since most VM don't need it anymore. This also salvage the ability to use IBPB (which doesn't seem to alter performance that much) but it isn't believed to be essential.