diff options
Diffstat (limited to 'results/classifier/gemma3:12b/hypervisor/1921082')
| -rw-r--r-- | results/classifier/gemma3:12b/hypervisor/1921082 | 17 |
1 files changed, 17 insertions, 0 deletions
diff --git a/results/classifier/gemma3:12b/hypervisor/1921082 b/results/classifier/gemma3:12b/hypervisor/1921082 new file mode 100644 index 00000000..dc135037 --- /dev/null +++ b/results/classifier/gemma3:12b/hypervisor/1921082 @@ -0,0 +1,17 @@ + +VM crash when process broadcast MCE + +When i do memory SRAR test for VM, I meet the following issue: + +My VM has 16 vCPU, I will inject one UE error to memory which is accessed by VM, Then host MCE is raised and SIGBUS is send to VM, and qemu take control. +Qemu will check the broadcast attribute by following cpu_x86_support_mca_broadcast(); + +Then Qemu may inject MCE to all vCPU, as vCPU is just one process for HOST, we can't guarantee all the vCPUs will enter MCE hander in 1S sync time, and the VM may panic. + +This issue will be easily fixed by expand monarch_timeout configuration, but the exact monarch_timeout can't be easily got, as it will depand on the num of vCPUs and current system schedule status. + +I am wondering why VM need broadcast attribute for MCE, When qeme process MCE event form host, it will always be signaled for one vCPU? If so, why does qemu need boradcast the MCE event to all vCPUs? + +Can weu just deliver LMCE to one specifc vCPU and make this behavior default? + +If anything wrong, Please point out. \ No newline at end of file |