On 05/05/2010 09:19 AM, Glauber Costa wrote:
Currently, in the linux kernel, we reset kvmclock if we are rebooting into a crash kernel through kexec. The rationale, is that a new kernel won't follow the same memory addresses, and the memory where kvmclock is located in the first kernel, will be something else in the second one. We don't do it in normal reboots, because the second kernel ends up registering kvmclock again, which has the effect of turning off the first instance. This is, however, totally wrong. This assumes we're booting into a kernel that also has kvmclock enabled. If by some reason we reboot into something that doesn't do kvmclock including but not limited to: * rebooting into an older kernel without kvmclock support, * rebooting with no-kvmclock, * rebootint into another O.S, we'll simply have the hypervisor writting into a random memory position into the guest. Neat, uh? Moreover, I believe the fix belongs in qemu, since it is the entity more prepared to detect all kinds of reboots (by means of a cpu_reset), not to mention the presence of misbehaving guests, that can forget to turn kvmclock off. It is also necessary to reset other msrs, so this patch resets everything that kvm exports through its MSR list.
Does that include the TSC? Zach -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html