Re: [RESEND PATCH] mm: slab: Do not create kmalloc caches smaller than arch_slab_minalign()

[Catalin Marinas]

On Mon, Oct 09, 2023 at 10:22:52AM +0200, Vlastimil Babka wrote:
> On 10/6/23 18:39, Catalin Marinas wrote:
> > Commit b035f5a6d852 ("mm: slab: reduce the kmalloc() minimum alignment
> > if DMA bouncing possible") allows architectures with non-coherent DMA to
> > define a small ARCH_KMALLOC_MINALIGN (e.g. sizeof(unsigned long long))
> > and this has been enabled on arm64. With KASAN_HW_TAGS enabled, however,
> > ARCH_SLAB_MINALIGN becomes 16 on arm64 (arch_slab_minalign() dynamically
> > selects it since commit d949a8155d13 ("mm: make minimum slab alignment a
> > runtime property")). This can lead to a situation where kmalloc-8 caches
> > are attempted to be created with a kmem_caches.size aligned to 16. When
> > the cache is mergeable, it can lead to kernel warnings like:
> > 
> > sysfs: cannot create duplicate filename '/kernel/slab/:d-0000016'
> > CPU: 0 PID: 1 Comm: swapper/0 Not tainted 6.6.0-rc1-00001-gda98843cd306-dirty #5
> > Hardware name: QEMU QEMU Virtual Machine, BIOS 0.0.0 02/06/2015
> > Call trace:
> >  dump_backtrace+0x90/0xe8
> >  show_stack+0x18/0x24
> >  dump_stack_lvl+0x48/0x60
> >  dump_stack+0x18/0x24
> >  sysfs_warn_dup+0x64/0x80
> >  sysfs_create_dir_ns+0xe8/0x108
> >  kobject_add_internal+0x98/0x264
> >  kobject_init_and_add+0x8c/0xd8
> >  sysfs_slab_add+0x12c/0x248
> >  slab_sysfs_init+0x98/0x14c
> >  do_one_initcall+0x6c/0x1b0
> >  kernel_init_freeable+0x1c0/0x288
> >  kernel_init+0x24/0x1e0
> >  ret_from_fork+0x10/0x20
> > kobject: kobject_add_internal failed for :d-0000016 with -EEXIST, don't try to register things with the same name in the same directory.
> > SLUB: Unable to add boot slab dma-kmalloc-8 to sysfs
> > 
> > Limit the __kmalloc_minalign() return value (used to create the
> > kmalloc-* caches) to arch_slab_minalign() so that kmalloc-8 caches are
> > skipped when KASAN_HW_TAGS is enabled (both config and runtime).
> > 
> > Fixes: b035f5a6d852 ("mm: slab: reduce the kmalloc() minimum alignment if DMA bouncing possible")
> > Signed-off-by: Catalin Marinas <catalin.marinas@xxxxxxx>
> > Reported-by: Mark Rutland <mark.rutland@xxxxxxx>
> > Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> > Cc: Vlastimil Babka <vbabka@xxxxxxx>
> > Cc: Peter Collingbourne <pcc@xxxxxxxxxx>
> > Cc: <stable@xxxxxxxxxxxxxxx> # 6.5.x
> > ---
> > 
> > The previous post was messed up by my git send-email configuration, so
> > sending it again. Also cc'ing Vlastimil since he reviewed the previous
> > slab changes for ARCH_KMALLOC_MINALIGN. Thanks.
> 
> It also touches only slab files so I can take it via slab tree.

Thanks.

> >  mm/slab_common.c | 7 +++++--
> >  1 file changed, 5 insertions(+), 2 deletions(-)
> > 
> > diff --git a/mm/slab_common.c b/mm/slab_common.c
> > index cd71f9581e67..8b45922ed295 100644
> > --- a/mm/slab_common.c
> > +++ b/mm/slab_common.c
> > @@ -895,10 +895,13 @@ void __init setup_kmalloc_cache_index_table(void)
> >  
> >  static unsigned int __kmalloc_minalign(void)
> >  {
> > +	unsigned int minalign = dma_get_cache_alignment();
> > +
> >  	if (IS_ENABLED(CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC) &&
> >  	    is_swiotlb_allocated())
> > -		return ARCH_KMALLOC_MINALIGN;
> > -	return dma_get_cache_alignment();
> > +		minalign = ARCH_KMALLOC_MINALIGN;
> > +
> > +	return ALIGN(minalign, arch_slab_minalign());
> 
> Could it be max() instead of ALIGN()? It would be more obvious, at least to
> me :)

Yeah, max() would do since they are all a power of two. Do you want me
to repost?

-- 
Catalin