Comment 0 for bug 1727331

bugproxy (bugproxy) wrote :

Environment:
Host OS: Ubuntu 16.04 (4.4.0-97-generic)
Guest OS: Ubuntu 16.04

Host is running NovaLink and GPFS. Guest gets filesystem from gpfs cluster.

Issue:

While booting the guest OS went into recovery. dmesg shows CPU soft lockup. This issue is easily recreatable when multiple VMs (about 10) are being started simultaneously.

.......................................................................................................................

 - Paul Mackerras <email address hidden> - 2017-10-24 00:42:25 ==

Looking at kernel sources:

The Ubuntu host kernel (4.4.0-97) does not include the patch "KVM: PPC: Book3S: Treat VTB as a per-subcore register, not per-thread", which is commit 88b02cf97bb7 in the upstream Linux kernel repository. The symptom of not having this patch is that guests running with threads > 1 (that is, in SMT2, SMT4 or SMT8 mode) can give spurious soft-lockup messages when they are not in fact locked up, if the guest kernel uses the VTB (virtual timebase) register in its softlockup detector code.

This is a backport of commit 88b02cf97bb7 from the upstream Linux kernel repository to the Ubuntu 4.4 kernel. It is the fix for the problem of seeing spurious soft lockup messages in guests running in an SMT mode greater than 1.

These tests were run on a system with 20 cores. Each VM had 1 core and 4 threads (SMT=4).
It is working very well with this patch