Re: [PATCH] x86/apic: Use byte array apic_version[], not int array. Saves up to 96k

[Denys Vlasenko]

On 09/13/2016 05:33 PM, Thomas Gleixner wrote:
> On Sun, 11 Sep 2016, Borislav Petkov wrote:
> > On Fri, Sep 09, 2016 at 10:32:04AM +0200, Denys Vlasenko wrote:
> > > This array is [MAX_LOCAL_APIC], and MAX_LOCAL_APIC can easily be up to 32k.
> > >
> > > This patch changes apic_version[] array elements from int to u8 -
> > > APIC version values as of year 2016 are no larger than 0x1f on all known CPUs.
> > > Version field in the APIC register is 8 bit wide - not likely
> > > to overflow byte range in foreseeable future.
> > >
> > > The "ver" argument of generic_processor_info(id,ver), which goes into apic_version[id],
> > > is also changed from int to u8: make it obvious that assignment can't overflow.
> > >
> > > generic_processor_info() has four callsites, none of them can put an out-of-range value
> > > into this argument.
> >
> > Right, so I dug a bit into this and found:
> >
> > http://marc.info/?l=linux-kernel&m=123230551709711
> >
> > and
> >
> > b2b815d80a5c ("x86: put trigger in to detect mismatched apic versions")
> >
> > It is from 2009 and I don't know how relevant 16-bit APIC IDs are
> > anymore... I guess you probably want to run this by SGI folk first.
> >
> > Otherwise I was going to propose to kill that apic_version array
> > altogether and cache only the version of the previous CPU and compare it
> > to the current one to catch mismatches...
>
> Yeah, the idea was back then to eliminate the array, but we wanted to make
> sure that we don't have systems out in the wild which have different apic
> versions. I really doubt that we can deal with that proper, so having a
> single version entry and yelling loudly when we detect a mismatch is good
> enough.

Makes sense. I'll send a patch