On Thu, Oct 17, 2013 at 11:37:48AM +0100, Jan Beulich wrote: > >>> On 17.10.13 at 12:23, Gleb Natapov <gleb@xxxxxxxxxx> wrote: > > On Thu, Oct 17, 2013 at 10:51:52AM +0100, Jan Beulich wrote: > >> >>> On 17.10.13 at 11:41, Gleb Natapov <gleb@xxxxxxxxxx> wrote: > >> > KVM obviously knows the complete state of virtual CPU. It can figure the > >> > situation above by looking at CS descriptor, not need to check > >> > is_long_mode() at all. Here is how emulator does it: > >> > >> And again - no: The last floating point operation may have > >> happened in 32-bit user mode context, while the state saving > >> may happen when the guest is already back in 64-bit kernel > >> mode. > >> > > Hmm, OK so the scenarios you are talking about is: > > 1. Guest's 32bit process uses FPU > > 2. Guest switch to 64bit kernel. > > 3. Before guest's kernel saves 32bit process's FPU state VMEXIT happens > > 4. KVM need to save FPU but it does not know what mode it is in > > Correct? > > Yes. > > > KVM gives FPU to a guest lazily, meaning that on a first FPU use #NM > > (intercepted by KVM) happens at which point FPU is granted to a guest. > > KVM can check what mode CPU was in at this point and use this info > > while saving FPU. But there is additional optimization that will prevent > > this from working for all cases: when FPU is granted to a guest KVM > > disabled CR0.TS/#NM intercepts, so guest is free to switch FPU from > > 32bit to 64bit mode without KVM knowing. Disabling this optimization > > will make FP intensive workload slow in a guest. > > Not sure what you're trying to tell me with this explanation. > Trying to think aloud how it can be fixed. -- Gleb. -- 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
