Re: Qemu problems in -next with 's390/spinlock: add niai spinlock hints'

[Martin Schwidefsky]

On Wed, 26 Jul 2017 08:05:27 +0200
Heiko Carstens <heiko.carstens@xxxxxxxxxx> wrote:

> On Wed, Jul 26, 2017 at 07:40:44AM +0200, Martin Schwidefsky wrote:
> > On Wed, 26 Jul 2017 07:00:33 +0200
> > Heiko Carstens <heiko.carstens@xxxxxxxxxx> wrote:
> >   
> > > On Tue, Jul 25, 2017 at 06:14:51PM -0700, Guenter Roeck wrote:  
> > > > Hi Martin,
> > > > 
> > > > my s390 qemu tests in linux-next stopped working a few days ago.
> > > > Bisect points to commit 's390/spinlock: add niai spinlock hints'.
> > > > 
> > > > Looking at the patch, this isn't really surprising; at least to me it looks
> > > > like the patch is making instructions mandatory which are only available in
> > > > Z14 CPUs. Does this mean that older s390 CPUs (such as the Z900 used in my
> > > > qemu tests) are no longer going to be supported in Linux ?    
> > > 
> > > No, that means that the patch has a bug. The NIAI instruction is only
> > > available if the execution-hint facility is installed. That facility came
> > > with zEC12. Luckily it uses the same facility indicator bit like the
> > > miscellaneous-instruction-extensions facility, which we already use anyway
> > > if the kernel gets compiled for zEC12. In that case we have early code
> > > which verifies if all required facilities to run the kernel are installed,
> > > and if not it will print a message to the console and stop the machine.
> > > 
> > > So the easiest fix would be to generate the NIAI instruction only if the
> > > kernel gets compiled for zEC12 or newer.  
> > 
> > Hmm, I though that NIAI is a NOP on older machines. A runtime check for
> > the facility bit is out of the question as the NIAI-7 gets inlined in
> > the spin_unlock code. So yes, the only available fix is to make the
> > NIAI hinting conditional on zEC12. Which is quite ugly as we would need
> > an architecture level set to zEC12 for the distribution kernel to make
> > use of NIAI.  
> 
> Alternatively you could generate a four-byte nop, and replace that at IPL
> time with the needed NIAI instruction, if the facility is available. Some
> sort of "alternative" code patching infrastructure that x86 already has.
> Not sure if it is worth it, however...

Patching all spin_unlock inlines? There are a lot of callers for this
function. We could think about an out-of-line spin_unlock and patch this
single function but then we'd loose the advantage of inlining.
I do not think it is worthwhile.

I pushed an updated patch to the features branch of s390/linux. Should
be in linux-next tomorroy. Thanks. 

-- 
blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.