We don't consistently advance the singlestep state machine when emulating instructions. We attempt to bodge around this when the host is stepping a guest, and fake a debug exception, but we don't always get this right. Additionally, we don't try to fix this up at all when a guest is stepping itself, so guests cannot single-step emulated instructions reliably. In both cases we're usually reliant on the HW singlestep state machine, so let's have our instruction emulation consistently advance that. Thus, when we return to a guest after emulating an instruction, the HW will generate the step exception for us, routed to host or guest appropriately. So far I have only compile-tested these patches. YMMV. These patches do not ensure that guest-stepping is reliable in the presence of host-stepping. We might need to say that it's userspace's responsibility to virtualize the guest singlestep state machine when stepping the guest. Otherwise, it's not clear to me if we can shadow this correctly within the kernel. Thanks, Mark. Mark Rutland (2): kvm/arm: skip MMIO insn after emulation kvm/arm: consistently advance singlestep when emulating instructions arch/arm/include/asm/kvm_host.h | 5 ---- arch/arm64/include/asm/kvm_emulate.h | 35 ++++++++++++++++++++------ arch/arm64/include/asm/kvm_host.h | 1 - arch/arm64/kvm/debug.c | 21 ---------------- arch/arm64/kvm/handle_exit.c | 14 +---------- arch/arm64/kvm/hyp/switch.c | 43 +++----------------------------- arch/arm64/kvm/hyp/vgic-v2-cpuif-proxy.c | 12 ++++++--- virt/kvm/arm/arm.c | 2 -- virt/kvm/arm/hyp/vgic-v3-sr.c | 6 ++++- virt/kvm/arm/mmio.c | 11 ++++---- 10 files changed, 52 insertions(+), 98 deletions(-) -- 2.11.0 _______________________________________________ kvmarm mailing list kvmarm@xxxxxxxxxxxxxxxxxxxxx https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
