On Mon, Nov 04, 2024 at 10:22:37PM +0000, Catalin Marinas wrote:
> On Mon, Nov 04, 2024 at 07:16:20PM +0100, Vlastimil Babka wrote:
> > On 11/4/24 19:00, Matthew Wilcox wrote:
> > > On Tue, Nov 05, 2024 at 12:08:37AM +0900, Koichiro Den wrote:
> > >> Commit b035f5a6d852 ("mm: slab: reduce the kmalloc() minimum alignment
> > >> if DMA bouncing possible") reduced ARCH_KMALLOC_MINALIGN to 8 on arm64.
> > >> However, with KASAN_HW_TAGS enabled, arch_slab_minalign() becomes 16.
> > >> This causes kmalloc_caches[*][8] to be aliased to kmalloc_caches[*][16],
> > >> resulting in kmem_buckets_create() attempting to create a kmem_cache for
> > >> size 16 twice. This duplication triggers warnings on boot:
> > >
> > > Wouldn't this be easier?
> >
> > They wanted it to depend on actual HW capability / kernel parameter, see
> > d949a8155d13 ("mm: make minimum slab alignment a runtime property")
> >
> > Also Catalin's commit referenced above was part of the series that made the
> > alignment more dynamic for other cases IIRC. So I doubt we can simply reduce
> > it back to a build-time constant.
>
> I principle, I wouldn't reduce it back to constant though the 8 vs 16
> difference is not significant. It matter if one enables KASAN_HW_TAGS
> and wants to run it on hardware without MTE, getting the *-8 caches
> back.
>
> That said, I haven't managed to trigger this warning yet. Do I need
> other config options than KASAN_HW_TAGS and DEBUG_VM?
Ah, I was missing SLAB_BUCKETS. I'll have a look tomorrow.
--
Catalin
