summary refs log tree commit diff stats
path: root/results/classifier/gemma3:12b/kvm/1837851
diff options
context:
space:
mode:
Diffstat (limited to 'results/classifier/gemma3:12b/kvm/1837851')
-rw-r--r--results/classifier/gemma3:12b/kvm/183785112
1 files changed, 12 insertions, 0 deletions
diff --git a/results/classifier/gemma3:12b/kvm/1837851 b/results/classifier/gemma3:12b/kvm/1837851
new file mode 100644
index 000000000..376f55a35
--- /dev/null
+++ b/results/classifier/gemma3:12b/kvm/1837851
@@ -0,0 +1,12 @@
+
+hv-tlbflush malfunctions on Intel host CPUs with neither EPT nor VPID (qemu-kvm)
+
+Enabling hv-tlbflush on older hosts using Intel CPUs supporting VT-x but neither EPT nor VPID will lead to bluescreens on the guest.
+
+It seems KVM only checks if EPT is available, and if it isn't it forcibly uses VPID. If that's *also* not available, it defaults to basically a no-op hypercall, though windows is expecting the TLB to be flushed.
+
+hv-tlbflush is pretty useless on machines not supporting these extensions anyway (only reasonably fix I can see would be to flush the *entire* TLB on tlbflush hypercall in KVM (i.e. a kernel fix), but that would remove any performance benefits), so I would suggest some kind of preliminary check and warning/error if hv-tlbflush is specified on such a host.
+
+All CPUs mentioned in this thread[0] are confirmed to be affected by the bug, and I have successfully reproduced it on an Intel Core2Duo E8500.
+
+[0] https://forum.proxmox.com/threads/windows-guest-bluescreen-with-proxmox-6.56053/
\ No newline at end of file