> On Nov 24, 2020, at 10:51 AM, Len Brown <lenb@xxxxxxxxxx> wrote: > > On Fri, Nov 20, 2020 at 12:03 AM Andy Lutomirski <luto@xxxxxxxxxx> wrote: >> >>> On Thu, Nov 19, 2020 at 3:37 PM Chang S. Bae <chang.seok.bae@xxxxxxxxx> wrote: >>> "xstate.enable=0x60000" will enable AMX on a system that does NOT have AMX >>> compiled into XFEATURE_MASK_USER_ENABLED (assuming the kernel is new enough >>> to support this feature). >>> >> >> What's the purpose of xstate.enable? I can't really imagine it's >> useful for AMX. I suppose it could be useful for hypothetical >> post-AMX features, but that sounds extremely dangerous. Intel has >> changed its strategy so many times on XSTATE extensibility that I find >> it quite hard to believe that supporting unknown states is wise. > > Not hypothetical -- there are subsequent hardware features coming that > will use the same > exact XSTATE support that this series puts in place for AMX. > > We know that when those features ship in new hardware, there will be > a set of customers who want to exercise those features immediately, > but their kernel binary has not yet been re-compiled to see those > features by-default. > > The purpose of "xstate.enable" is to empower those users to be able to > explicitly enable support using their existing binary. > > You are right -- the feature isn't needed to enable AMX, unless somebody went to > the trouble of building a kernel with the AMX source update, but chose > to disable > AMX-specific recognition, by-default. > > We may want to taint the kernel if one of these flags is used because, frankly, Intel’s track record is poor. Suppose we get a new feature with PKRU-like semantics -- switching it blindly using XSAVE(C,S,OPT,whatever) would simply incorrect. And XFD itself has problems — supposedly it’s difficult or impossible to virtualize. It wouldn’t utterly shock me if Intel were to drop IA32_XFD_ERR and replace it with a new mechanism that’s less janky. So this whole thing makes me quite nervous. (I'm not a virtualization expert, but AIUI IA32_XFD_ERR has some issues. If it's too late to fix those issues, Intel could probably get away with completely dropping IA32_XFD_ERR from the spec -- OSes can handle AMX just fine without it. Then the next XFD-able feature could introduce a new improved way of reporting which feature triggered #NM.)
