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

[Vlastimil Babka <vbabka@suse.cz>]

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@arm.com>
> Reported-by: Mark Rutland <mark.rutland@arm.com>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: Vlastimil Babka <vbabka@suse.cz>
> Cc: Peter Collingbourne <pcc@google.com>
> Cc: <stable@vger.kernel.org> # 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.

>  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 :)

>  }
>  
>  void __init