This isn't anything new. The same thing happened with ADX instructions on Broadwell, MOVBE on Haswell, F16C instructions on Ivebridge, etc. On Wed, Dec 20, 2017 at 5:08 PM, Haozhong Zhang <haozhong.zhang@xxxxxxxxx> wrote: > On 12/20/17 19:26 -0200, Eduardo Habkost wrote: >> On Tue, Dec 19, 2017 at 10:54:16AM +0800, Haozhong Zhang wrote: >> > On 12/18/17 16:36 -0200, Eduardo Habkost wrote: >> > > On Tue, Dec 12, 2017 at 04:35:24PM +0800, Haozhong Zhang wrote: >> > > > Intel VMX cannot intercept guest clwb and clflushopt. When clwb and >> > > > clflushopt are not exposed in guest cpuid, clwb and clflushopt >> > > > instructions in this test case can still succeed without #UD on the >> > > > host CPU which has clwb and clflushopt support, though failures with >> > > > UD are expected. >> > > > >> > > > In order to avoid false alarms in such cases, introduce the following >> > > > two arguments "has_clwb" and "has_clflushopt" to allow users to >> > > > specify whether clwb and clflushopt are supported on the host CPU. >> > > > >> > > > Signed-off-by: Haozhong Zhang <haozhong.zhang@xxxxxxxxx> >> > > > --- >> > > > x86/memory.c | 29 +++++++++++++++++++++++++++-- >> > > > 1 file changed, 27 insertions(+), 2 deletions(-) >> > > > >> > > > diff --git a/x86/memory.c b/x86/memory.c >> > > > index cd1eb46..03ff7d3 100644 >> > > > --- a/x86/memory.c >> > > > +++ b/x86/memory.c >> > > > @@ -23,10 +23,29 @@ static void handle_ud(struct ex_regs *regs) >> > > > regs->rip += isize; >> > > > } >> > > > >> > > > +/* >> > > > + * Intel VMX cannot intercept guest clwb and clflushopt. When clwb and >> > > > + * clflushopt are not exposed in guest cpuid, clwb and clflushopt >> > > > + * instructions in this test case can still succeed without #UD on >> > > > + * the host CPU which has clwb and clflushopt support. In order to avoid >> > > > + * false alarms in such cases, introduce the following two arguments >> > > > + * to allow users to specify whether clwb and clflushopt are supported on >> > > > + * the host CPU: >> > > > + * - has_clwb: indicates clwb is supported on the host CPU >> > > > + * - has_clflushopt: indicates clflushopt is supported on the host CPU >> > > > + */ >> > > >> > > Why not simply use "-cpu host" to make sure the guest CPUID flags >> > > match host CPUID? >> > > >> > >> > Can I understand that testing these two cases with host/guest CPUID >> > mismatch (specially clwb and clflushopt flags) is invalid? If yes, >> > please ignore this patch. >> >> I wouldn't say it's invalid to test what happens when the host >> and guest CPUID don't match. The question is: is it useful to do >> so? Are we testing different code paths when we do that? >> > > No, VMX cannot intercept guest clwb and clflushopt, so no KVM code > path is involved when guest executes those two instructions. > >> The inability to trigger #UD if the host CPUID includes the flag >> sounds like a bug/limitation we would like to get rid of as soon >> as hardware allow us to, and not a feature we need to test for. >> > > It's more the inability of VMX, which cannot intercept clwb and > clflushopt. > >> What's the right way to ensure memory.flat is always tested using >> "-cpu host"? >> > > I think so, at least '-cpu host' can mitigate the hardware inability. > > Haozhong
