Re: [RFC v2 2/3] atomic: Specify alignment for atomic_t and atomic64_t

[Arnd Bergmann]

On Tue, Sep 23, 2025, at 08:28, Finn Thain wrote:
> On Mon, 22 Sep 2025, Arnd Bergmann wrote:
>> On Mon, Sep 22, 2025, at 10:16, Finn Thain wrote:
> @@ -8,7 +8,7 @@ struct vfsmount;
>  struct path {
>         struct vfsmount *mnt;
>         struct dentry *dentry;
> -} __randomize_layout;
> +} __aligned(sizeof(long)) __randomize_layout;
>
> There's no need: struct path contains a struct dentry, which contains a 
> seqcount_spinlock_t, which contains a spinlock_t which contains an 
> atomic_t member, which is explicitly aligned.
>
> Despite that, there's still some kmem cache or other allocator somewhere 
> that has produced some misaligned path and dentry structures. So we get 
> misaligned atomics somewhere in the VFS and TTY layers. I was unable to 
> find those allocations.

Ok, I see. Those would certainly be good to find. I would assume that
all kmem caches have a sensible alignment on each architecture, but
I think the definition in linux/slab.h actually ends up setting the
minimum to 2 here:

#ifndef ARCH_KMALLOC_MINALIGN
#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
#elif ARCH_KMALLOC_MINALIGN > 8
#define KMALLOC_MIN_SIZE ARCH_KMALLOC_MINALIGN
#define KMALLOC_SHIFT_LOW ilog2(KMALLOC_MIN_SIZE)
#endif

#ifndef ARCH_SLAB_MINALIGN
#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
#endif

Maybe we should just change __alignof__(unsigned long long)
to a plain '8' here and make that the minimum alignment
everywhere, same as the atomic64_t alignment change.

Alternatively, we can keep the __alignof__ here in order
to reduce padding on architectures with a default 4-byte
alignment for __u64, but then override ARCH_SLAB_MINALIGN
and ARCH_KMALLOC_MINALIGN on m68k to be '4' instead of '2'.

     Arnd