Re: [kvm-unit-tests PATCH] x86: Assign a canonical address before execute invpcid

[Sean Christopherson <seanjc@google.com>]

On Thu, Dec 30, 2021, Zhenzhong Duan wrote:
> Occasionally we see pcid test fail as INVPCID_DESC[127:64] is
> uninitialized before execute invpcid.
> 
> According to Intel spec: "#GP If INVPCID_TYPE is 0 and the linear
> address in INVPCID_DESC[127:64] is not canonical."
> 
> Assign desc's address which is guaranteed to be a real memory
> address and canonical.
> 
> Fixes: b44d84dae10c ("Add PCID/INVPCID test")
> Signed-off-by: Zhenzhong Duan <zhenzhong.duan@intel.com>
> ---
>  x86/pcid.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/x86/pcid.c b/x86/pcid.c
> index 527a4a9..4828bbc 100644
> --- a/x86/pcid.c
> +++ b/x86/pcid.c
> @@ -75,6 +75,9 @@ static void test_invpcid_enabled(int pcid_enabled)
>      struct invpcid_desc desc;
>      desc.rsv = 0;
>  
> +    /* Initialize INVPCID_DESC[127:64] with a canonical address */
> +    desc.addr = (u64)&desc;

Casting to a u64 is arguably wrong since the address is an unsigned long.  It
doesn't cause problems because the test is 64-bit only, but it's a bit odd.

What about just replacing "desc.rsv = 0" with "memset(&desc, 0, sizeof(desc))"?

> +
>      /* try executing invpcid when CR4.PCIDE=0, desc.pcid=0 and type=0..3
>       * no exception expected
>       */
> -- 
> 2.25.1
>