On Fri, Oct 21, 2022, Vitaly Kuznetsov wrote: > @@ -48,6 +49,8 @@ static inline void rdmsr_gs_base(void) > > void l2_guest_code(void) > { > + u64 unused; > + > GUEST_SYNC(7); > > GUEST_SYNC(8); > @@ -64,15 +67,33 @@ void l2_guest_code(void) > vmcall(); > rdmsr_gs_base(); /* intercepted */ > > + /* L2 TLB flush tests */ > + hyperv_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE | HV_HYPERCALL_FAST_BIT, 0x0, > + HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES | HV_FLUSH_ALL_PROCESSORS); > + rdmsr_fs_base(); > + /* > + * Note: hypercall status (RAX) is not preserved correctly by L1 after > + * synthetic vmexit, use unchecked version. > + */ > + __hyperv_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE | HV_HYPERCALL_FAST_BIT, 0x0, > + HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES | HV_FLUSH_ALL_PROCESSORS, > + &unused); > + /* Make sure we're not issuing Hyper-V TLB flush call again */ > + __asm__ __volatile__ ("mov $0xdeadbeef, %rcx"); This needs a clobber. It won't cause problems in the current code, but it will make someone really sad if they add more code after this. Using %ecx instead of %rcx will also suffice, 32-bit accesses clear bits 63:32. Even better, if a "nop" isn't required to get the compiler to emit preamble, would be to load ECX through an input constraint, that way the compiler "knows" the value of ECX and can optimize for it (though it's extremely unlikely 0xdeadbeef will be a useful value). Actually, not setting RCX in vmcall() is a nasty bug waiting to happen, e.g. if RCX just happens to contain a value that gets routed to L0. Rather than handle this as a one-off, can you insert a prep patch to have the common vmcall() stuff RCX with a "safe" value? Related side topic, rdmsr_{f,g}s_base() should also use input constraints, and should use a proper #define for the MSRs. Also, why earth do those clobber all GPRs? Oooh, because they get routed to L1 and L1 doesn't preserve GPRs. Related side topic #2, KVM's kvm_xen_hypercall() is broken, it checks the wrong input register for 64-bit (checks RAX instead of RCX). Not sure that's even a fixable bug though.