Re: [PATCH v7 1/4] spinlock: A new lockref structure for lockless update of refcount

[Linus Torvalds <torvalds@linux-foundation.org>]

On Tue, Sep 3, 2013 at 2:13 PM, Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> This is the one that actually compiles.  Whether it *works* is still a
> total mystery.

It generates ok code, and it booted, so it seems to work at least for my config.

However, it seems to make no performance-difference what-so-ever, and
lg_local_lock is still using about 7% cpu per the profiles.

The code generation is slightly better, but the profile looks the same:

       │    ffffffff81078e70 <lg_local_lock>:
  0.62 │      push   %rbp
  0.28 │      mov    %rsp,%rbp
  0.22 │      add    %gs:0xcd48,%rdi
  0.27 │      mov    $0x100,%eax
 97.22 │      lock   xadd   %ax,(%rdi)
  0.01 │      movzbl %ah,%edx
       │      cmp    %al,%dl
  0.56 │    ↓ je     29
       │      xchg   %ax,%ax
  0.00 │20:   pause
  0.00 │      movzbl (%rdi),%eax
       │      cmp    %dl,%al
       │    ↑ jne    20
       │29:   pop    %rbp
  0.81 │    ← retq

but it still obviously doesn't do the "lock xadd %ax,%gs:(%rdi)"
(without the preceding 'add') that would be the optimal code.

I'll try to hack that up too, but it's looking like it really is just
the "lock xadd", not the memory dependency chain..

            Linus