Re: [Xen-devel] [PATCH v3 3/5] xen/x86: introduce more cache maintenance operations

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 10.10.14 at 11:23, <stefano.stabellini@xxxxxxxxxxxxx> wrote:
> On Fri, 10 Oct 2014, Jan Beulich wrote:
>> >>> On 08.10.14 at 15:00, <stefano.stabellini@xxxxxxxxxxxxx> wrote:
>> > --- a/xen/include/asm-x86/page.h
>> > +++ b/xen/include/asm-x86/page.h
>> > @@ -346,6 +346,10 @@ static inline uint32_t 
>> > cacheattr_to_pte_flags(uint32_t 
> cacheattr)
>> >  
>> >  /* No cache maintenance required on x86 architecture. */
>> >  static inline void flush_page_to_ram(unsigned long mfn) {}
>> > +static inline void clean_xen_dcache_va_range(const void *p, unsigned long 
> size) {}
>> > +static inline void invalidate_xen_dcache_va_range(const void *p, unsigned 
> long size) {}
>> > +static inline void clean_and_invalidate_xen_dcache_va_range
>> > +    (const void *p, unsigned long size) {}
>> 
>> Mind explaining what purpose the "_xen" in these names serves?
> 
> It means a Xen virtual address (as opposed to a guest p2m address or a
> guest virtual address), but I guess it is pretty obvious so I should
> remove it?

Yes please.

>> Also, is simply stubbing these out correct? I.e. there are caches on
>> x86, and those can both be invalidated and cleaned, so why would
>> you not do so here?
> 
> Good point. The problem is that to be sure that I am doing the right
> thing I would need to know what kind of flushes are required with a
> non-coherent device on x86, but I don't have any examples at hand.
> 
> From the amd64 manual, I would implement invalidate with CLFLUSH, clean
> and invalidate with WBINVD (even though it operates on the entire cache
> so it is very heavyweight) and I would have no way to implement just
> "clean". Do you agree?

Except that "clean-without-invalidate" can fall back to "clean-and-
invalidate" just fine. The only thing you truly can't implement is
invalidate-without-clean, since INVD is global, yet you would want
range-restricted invalidation only. And no, CLFLUSH and WBINVD
do the same thing (address restricted vs globally).

Also x86 already has all of the logic in place, you just need to
call it (see xen/*/*x86/flushtlb.*). 

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel