Re: [PATCH 08/10] mm/slab: Allow dynamic kmalloc() minimum alignment

[Catalin Marinas]

On Thu, Apr 07, 2022 at 06:18:16PM +0900, Hyeonggon Yoo wrote:
> On Thu, Apr 07, 2022 at 09:50:23AM +0100, Catalin Marinas wrote:
> > On Thu, Apr 07, 2022 at 03:46:37AM +0000, Hyeonggon Yoo wrote:
> > > On Tue, Apr 05, 2022 at 02:57:56PM +0100, Catalin Marinas wrote:
> > > > --- a/mm/slab_common.c
> > > > +++ b/mm/slab_common.c
> > > > @@ -838,9 +838,18 @@ void __init setup_kmalloc_cache_index_table(void)
> > > >  	}
> > > >  }
> > > >  
> > > > -static void __init
> > > > +unsigned int __weak arch_kmalloc_minalign(void)
> > > > +{
> > > > +	return ARCH_KMALLOC_MINALIGN;
> > > > +}
> > > > +
> > > 
> > > As ARCH_KMALLOC_ALIGN and arch_kmalloc_minalign() may not be same after
> > > patch 10, I think s/ARCH_KMALLOC_ALIGN/arch_kmalloc_minalign/g
> > > for every user of it would be more correct?
> > 
> > Not if the code currently using ARCH_KMALLOC_MINALIGN needs a constant.
> > Yes, there probably are a few places where the code can cope with a
> > dynamic arch_kmalloc_minalign() but there are two other cases where a
> > constant is needed:
> > 
> > 1. As a BUILD_BUG check because the code is storing some flags in the
> >    bottom bits of a pointer. A smaller ARCH_KMALLOC_MINALIGN works just
> >    fine here.
> > 
> > 2. As a static alignment for DMA requirements. That's where the newly
> >    exposed ARCH_DMA_MINALIGN should be used.
> > 
> > Note that this series doesn't make the situation any worse than before
> > since ARCH_DMA_MINALIGN stays at 128 bytes for arm64. Current users can
> > evolve to use a dynamic alignment in future patches. My main aim with
> > this series is to be able to create kmalloc-64 caches on arm64.
> 
> AFAIK there are bunch of drivers that directly calls kmalloc().

Well, lots of drivers call kmalloc() ;).

> It becomes tricky when e.g.) a driver allocates just 32 bytes,
> but architecture requires it to be 128-byte aligned.

That's the current behaviour, a 32 byte allocation would return an
object from kmalloc-128. I want to reduce this to at least kmalloc-64
(or smaller) if the CPU/SoC allows it.

> That's why everything allocated from kmalloc() need to be aligned in
> ARCH_DMA_MINALIGN.

I don't get your conclusion here. Would you mind explaining?

> So I'm yet skeptical on decoupling ARCH_DMA/KMALLOC_MINALIGN. Instead
> of decoupling it, I'm more into dynamically decreasing it.

The reason for decoupling is mostly that there are some static uses of
ARCH_KMALLOC_MINALIGN as per point 1 above. The other is the
__assume_kmalloc_alignment attribute. We shouldn't have such assumed
alignment larger than what a dynamic kmalloc() would return. To me it
makes a lot more sense for ARCH_KMALLOC_MINALIGN to be the minimum
guaranteed in a kernel build but kmalloc() returning a larger alignment
at run-time than the other way around.

Thanks.

-- 
Catalin