Re: [PATCH] kernel: fix data race in put_pid

[James Hogan]

Hi Peter,

On Fri, Sep 18, 2015 at 11:27:32AM +0200, Peter Zijlstra wrote:
> On Fri, Sep 18, 2015 at 10:57:32AM +0200, Peter Zijlstra wrote:
> > diff --git a/arch/metag/include/asm/atomic_lnkget.h b/arch/metag/include/asm/atomic_lnkget.h
> > index 21c4c268b86c..1bd21c933435 100644
> > --- a/arch/metag/include/asm/atomic_lnkget.h
> > +++ b/arch/metag/include/asm/atomic_lnkget.h
> > @@ -3,7 +3,8 @@
> >  
> >  #define ATOMIC_INIT(i)	{ (i) }
> >  
> > -#define atomic_set(v, i)		((v)->counter = (i))
> > +/* XXX: should be LNKSETD ? */
> > +#define atomic_set(v, i)		WRITE_ONCE((v)->counter, (i))
> >  
> >  #include <linux/compiler.h>
> >  
> 
> James?

It should be fine without a LNKSET. The only reason LNKGET is required
for atomic_read() is that on certain cores the LNKGET/LNKSET
instructions went around the cache, but those cores also have write
through caches, so it shouldn't result in incoherence.

I've just double checked, and the pipeline seems to handle the linked
read after write correctly, so e.g. after:

0220002c    MOV       D0FrT,#0x5
b6208042    SETD      [D1Ar3],D0FrT
ad1080cc    LNKGETD   D0Ar4,[D1Ar3]

D0Ar4 is correct (0x5).

Cheers
James

Attachment: signature.asc
Description: Digital signature