On Wed, Aug 07, 2024 at 01:36:33PM GMT, Andrew Jones wrote: ... > I just tried 32-bit KVM and see that the DBCN write test fails the > 'write success' test. That may be a KVM bug. > We can blame both KVM and kvmtool. KVM sets sbiret.error to a0 and sbiret.value to a1 before exiting to userspace[1]. I think that comes from thinking about how a real ecall would set them. However, as this isn't an ecall, they should get set directly by userspace, not through registers. Also, we should initialize them to some known value before calling userspace, and zero is probably the best choice. kvmtool neglects to set sbiret.error to SBI_SUCCESS on a successful write. QEMU does set it, so this failure shouldn't happen with QEMU, but I haven't tried it. Patching both KVM and kvmtool is best, as it would allow the test to pass when running new KVM with old kvmtool and when running old KVM with new kvmtool. [1] arch/riscv/kvm/vcpu_sbi.c:130 Thanks, drew
