Rusty Russell wrote:
> Gruntwork, basically. You'll see why it's paravirt.h and no_paravirt.h
> in the next patch.
>
> Name: Make Paravirtualization-requiring Instructions into Macros
> Status: Booted on 2.6.16-rc2-git7
> Depends: ring_assumptions_cleanup.patch.gz
> Signed-off-by: Rusty Russell <rusty@xxxxxxxxxxxxxxx>
>
My only suggestion here is that sync_core be kept inline as a raw cpuid
instruction rather than being moved into paravirt.h. The purpose is not
to extract CPUID information, but to serialize execution, and it can
cause no trouble for a hypervisor, as cpuid doesn't trap anyway.
Further, it is only used in the microcode update module, which isn't
something that should be done in a hypervisor to begin with.
>
> @@ -286,13 +269,6 @@ static inline void clear_in_cr4 (unsigne
> outb((data), 0x23); \
> } while (0)
>
> -/* Stop speculative execution */
> -static inline void sync_core(void)
> -{
> - int tmp;
> - asm volatile("cpuid" : "=a" (tmp) : "0" (1) : "ebx","ecx","edx","memory");
> -}
That said, there is an open "bug" on i386 in module loading. We really
should insert a call to sync_core during module loading after the
alternative instructions have been patched. In theory, the processor
could prefetch the module code into the icache before any calls are made
into the module, and since the module initialization calls follow
shortly after the alternate patching, there is theoretically a window
where the old code (albeit more conservative) could be executed.
Zach
