Hello,
We've been talking internally about Spectre-BHB (disclosed a couple of
weeks ago [0])
and thought it could be interesting to start some discussion on the mailing
list, especially around the software mitigation.
As a start, here is the overhead of the mitigation [1] for some guest operations
on a Neoverse N1 core [0]. The table shows the avg. latency of each operation,
measured using kvm-unit-tests/arm/micro-bench.
Micro-bench not-mitigated mitigated
nsecs nsecs overhead
=============================================================
hvc 292 317 8.56%
mmio_read_user 1609 1704 5.90%
mmio_read_vgic 419 444 5.97%
eoi 29 29 0.00%
ipi 1509 1548 2.58%
lpi 1643 1758 7.00%
timer_10ms 798 844 5.76%
Some details:
- Using 56e337f2cf13 with and without MITIGATE_SPECTRE_BRANCH_HISTORY.
- Each test (row) runs for 1 minute in a tight loop.
We are still evaluating the effect of this on real workloads.
Thanks!
Ricardo
[0] Spectre-BHB white paper:
https://developer.arm.com/support/arm-security-updates/speculative-processor-vulnerability/spectre-bhb
[1] 558c303c9734af5a813739cd284879227f7297d2 arm64: Mitigate spectre
style branch history side channels
[2] the specific vector is spectre_bhb_k24 (Arm Neoverse-N1), from the
Spectre-BHB white paper in [0].
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
