I found some atomic_add/dec are replaced with atomic_add/dec_return,

I am going to replace -return variants with -fetch variants, potentially -fetch

those helpers with return value imply a full memory barrier around it, but
others without return value do not. Do you have any numbers to show
the impact? Maybe atomic_add/dec_return_relaxed can help this.

The generic variant uses  arch_cmpxchg() for all atomic variants
without any extra barriers. Therefore, on platforms that use generic
implementations there won't be performance differences except for an
extra branch that checks results when VM_BUG_ON is enabled.

On x86 the difference between the two is the following

atomic_add:
   lock add %eax,(%rsi)

atomic_fetch_add:
   lock xadd %eax,(%rsi)

atomic_fetch_add_relaxed:
   lock xadd %eax,(%rsi)

No differences between relaxed and non relaxed variants. However, we
used lock xadd instead of lock add. I am not sure if the performance
difference is going to be different.

Pasha