Avi Kivity wrote: > On 05/06/2010 07:23 AM, Cui, Dexuan wrote: >> >>>> + goto err; >>>> + vcpu->arch.guest_xstate_mask = new_bv; >>>> + xsetbv(XCR_XFEATURE_ENABLED_MASK, vcpu->arch.guest_xstate_mask); >>>> >>>> >>> Can't we run with the host xcr0? isn't it guaranteed to be a >>> superset of the guest xcr0? >>> >> I agree host xcr0 is a superset of guest xcr0. >> In the case guest xcr0 has less bits set than host xcr0, I suppose >> running with guest xcr0 would be a bit faster. > > However, switching xcr0 may be slow. That's our experience with msrs. > Can you measure its latency? We can measure that. However, I think the changing xcr0 to guest xcr0 in handle_xsetbv() is necessary -- or else, inside guest xgetbv() would return host xcr0 rather than guest xcr0 -- this is obviously incorrect. Once handle_xsetbv() changes the xcr0 to guest's value, the xsetbv() in kvm_fx_restore_host() is also necessary, and the xsetbv() in kvm_fx_restore_guest() is also necessary. So looks guest can't run with the host xcr0.
> Can also be avoided if guest and host xcr0 match. > >> If you think simplying the code by removing the field >> guest_xstate_mask is more important, we can do it. >> > > Well we need guest_xstate_mask, it's the guest's xcr0, no? I misread it. Yes, we need geust_xsave_mask. > > btw, it needs save/restore for live migration, as well as save/restore > for the new fpu state. Yes. This part is missing. Sheng and I is also doing this -- it may be a bittle troublesome as the number of XSTATEs can grown as time goes on. We'll have to handle the compatibility issue. > >>>> + skip_emulated_instruction(vcpu); >>>> + return 1; >>>> +err: >>>> + kvm_inject_gp(vcpu, 0); >>>> >>>> >>> Need to #UD in some circumstances. >>> >> I don't think so. When the CPU doesn't suport XSAVE, or guest >> doesn't set guest CR4.OSXSAVE, guest's attempt of exectuing xsetbv >> would get a #UD immediately >> and no VMexit would occur. >> > > Ok. > >>>> @@ -62,6 +64,7 @@ >>>> (~(unsigned long)(X86_CR4_VME | X86_CR4_PVI | X86_CR4_TSD | >>>> X86_CR4_DE\ | X86_CR4_PSE | X86_CR4_PAE | >>>> X86_CR4_MCE \ >>>> | X86_CR4_PGE | X86_CR4_PCE | X86_CR4_OSFXSR >>>> \ >>>> + | (cpu_has_xsave ? X86_CR4_OSXSAVE : 0) \ >>>> | X86_CR4_OSXMMEXCPT | X86_CR4_VMXE)) >>>> >>>> >>> It also depends on the guest cpuid value. Please add it outside the >>> >> If cpu_has_xsave is true, we always present xsave to guest via >> cpuid, so I >> think the code is correct here. >> > > We don't pass all host cpuid values to the guest. We expose them to > userspace via KVM_GET_SUPPORTED_CPUID, and then userspace decides what > cpuid to present to the guest via KVM_SET_CPUID2. So the guest may > run with fewer features than the host. Sorry, I didn't notice this. Will look into this. > >> I saw the 2 patches you sent. They (like the new APIs >> fpu_alloc/fpu_free) will simplify the implementation of kvm xsave >> support. Thanks a lot! >> > > Thanks. To simplify things, please separate host xsave support > (switching to the fpu api) and guest xsave support (cpuid, xsetbv, new > ioctls) into separate patches. In fact, guest xsave support is > probably best split into patches as well. Ok. Will try to do these. Thanks, -- Dexuan -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
