Re: [PATCH v3 08/11] slub: Replace cmpxchg_double()

* Re: [PATCH v3 08/11] slub: Replace cmpxchg_double()
       [not found]   ` <20230524093246.GP83892@hirez.programming.kicks-ass.net>
@ 2023-05-30 14:22     ` Peter Zijlstra
  2023-05-30 19:32       ` Peter Zijlstra
  0 siblings, 1 reply; 2+ messages in thread
From: Peter Zijlstra @ 2023-05-30 14:22 UTC (permalink / raw)
  To: torvalds
  Cc: corbet, will, boqun.feng, mark.rutland, catalin.marinas, dennis,
	tj, cl, hca, gor, agordeev, borntraeger, svens, tglx, mingo, bp,
	dave.hansen, x86, hpa, joro, suravee.suthikulpanit, robin.murphy,
	dwmw2, baolu.lu, Arnd Bergmann, Herbert Xu, davem, penberg,
	rientjes, iamjoonsoo.kim, Andrew Morton, vbabka, roman.gushchin,
	42.hyeyoo, linux-doc, linux-kernel, linux-mm, linux-s390, iommu,
	linux-arch, linux-crypto, sfr, mpe, James.Bottomley, deller,
	linux-parisc

On Wed, May 24, 2023 at 11:32:47AM +0200, Peter Zijlstra wrote:
> On Mon, May 15, 2023 at 09:57:07AM +0200, Peter Zijlstra wrote:
> 
> > @@ -3008,6 +3029,22 @@ static inline bool pfmemalloc_match(stru
> >  }
> >  
> >  #ifndef CONFIG_SLUB_TINY
> > +static inline bool
> > +__update_cpu_freelist_fast(struct kmem_cache *s,
> > +			   void *freelist_old, void *freelist_new,
> > +			   unsigned long tid)
> > +{
> > +#ifdef system_has_freelist_aba
> > +	freelist_aba_t old = { .freelist = freelist_old, .counter = tid };
> > +	freelist_aba_t new = { .freelist = freelist_new, .counter = next_tid(tid) };
> > +
> > +	return this_cpu_cmpxchg_freelist(s->cpu_slab->freelist_tid.full,
> > +					 old.full, new.full) == old.full;
> > +#else
> > +	return false;
> > +#endif
> > +}
> > +
> >  /*
> >   * Check the slab->freelist and either transfer the freelist to the
> >   * per cpu freelist or deactivate the slab.
> > @@ -3359,11 +3396,7 @@ static __always_inline void *__slab_allo
> >  		 * against code executing on this cpu *not* from access by
> >  		 * other cpus.
> >  		 */
> > -		if (unlikely(!this_cpu_cmpxchg_double(
> > -				s->cpu_slab->freelist, s->cpu_slab->tid,
> > -				object, tid,
> > -				next_object, next_tid(tid)))) {
> > -
> > +		if (unlikely(!__update_cpu_freelist_fast(s, object, next_object, tid))) {
> >  			note_cmpxchg_failure("slab_alloc", s, tid);
> >  			goto redo;
> >  		}
> > @@ -3736,11 +3769,7 @@ static __always_inline void do_slab_free
> >  
> >  		set_freepointer(s, tail_obj, freelist);
> >  
> > -		if (unlikely(!this_cpu_cmpxchg_double(
> > -				s->cpu_slab->freelist, s->cpu_slab->tid,
> > -				freelist, tid,
> > -				head, next_tid(tid)))) {
> > -
> > +		if (unlikely(!__update_cpu_freelist_fast(s, freelist, head, tid))) {
> >  			note_cmpxchg_failure("slab_free", s, tid);
> >  			goto redo;
> >  		}
> 
> This isn't right; the this_cpu_cmpxchg_double() was unconditional and
> relied on the local_irq_save() fallback when no native cmpxchg128 is
> present.

This means this_cpu_cmpxchg128 is expected to be present on all 64bit
archs, except Mark just found out that HPPA doens't support __int128
until gcc-11.

(I've been building using gcc-12.2)

And because the cmpxchg128 fallback relies on '==' we can't trivally
fudge that with a struct type either :/ Now, afaict it all magically
works if I use:

#ifdef __SIZEOF_INT128__
typedef __s128 s128
typedef __u128 u128
#else
#if defined(CONFIG_PARISC) && defined(CONFIG_64BIT)
typedef long double u128;
#endif
#endif

but that is *super* gross.

The alternative is raising the minimum GCC for PARISC to gcc-11..

Yet another alternative is using a struct type and an equality function,
just for this.

Anybody?

^ permalink raw reply	[flat|nested] 2+ messages in thread