Arjan van de Ven wrote: > On Tue, 2006-07-18 at 00:00 -0700, Chris Wright wrote: > >> plain text document attachment (i386-cpuid) >> Allow subarchitectures to modify the CPUID instruction. This allows >> the subarch to provide a limited set of CPUID feature flags during CPU >> identification. Add a subarch implementation for Xen that traps to the >> hypervisor where unsupported feature flags can be hidden from guests. >> > > Hi, > > I'm wondering if this is entirely the wrong level of abstraction; to me > it feels the subarch shouldn't override the actual cpuid, but the cpu > feature flags that linux uses. That's a lot less messy: cpuid has many > many pieces of information which are near impossible to filter in > practice, however filtering the USAGE of it is trivial; linux basically > flattens the cpuid namespace into a simple bitmap of "what the kernel > can use". That is really what the subarch should filter/fixup, just like > we do for cpu quirks etc etc. > You really need a CPUID hook. The instruction is non-virtualizable, and anything claiming to be a hypervisor really has to support masking and flattening the cpuid namespace for the instruction itself. It is used in assembler code and very early in boot. The alternative is injecting a bunch of Xen-specific code to filter feature bits into the i386 layer, which is both bad for Linux and bad for Xen - and was quite ugly in the last set of Xen patches. Zach
