Re: [PATCH v3] skbuff: fix a data race in skb_queue_len()

["Jason A. Donenfeld" <Jason@zx2c4.com>]

On Thu, Feb 06, 2020 at 10:22:02AM -0800, Eric Dumazet wrote:
> On 2/6/20 10:12 AM, Jason A. Donenfeld wrote:
> > On Thu, Feb 6, 2020 at 6:10 PM Eric Dumazet <eric.dumazet@gmail.com> wrote:
> >> Unfortunately we do not have ADD_ONCE() or something like that.
> > 
> > I guess normally this is called "atomic_add", unless you're thinking
> > instead about something like this, which generates the same
> > inefficient code as WRITE_ONCE:
> > 
> > #define ADD_ONCE(d, s) *(volatile typeof(d) *)&(d) += (s)
> > 
> 
> Dmitry Vyukov had a nice suggestion few months back how to implement this.
> 
> https://lkml.org/lkml/2019/10/5/6

That trick appears to work well in clang but not gcc:

#define ADD_ONCE(d, i) ({ \
       typeof(d) *__p = &(d); \
       __atomic_store_n(__p, (i) + __atomic_load_n(__p, __ATOMIC_RELAXED), __ATOMIC_RELAXED); \
})

gcc 9.2 gives:

  0:   8b 47 10                mov    0x10(%rdi),%eax
  3:   83 e8 01                sub    $0x1,%eax
  6:   89 47 10                mov    %eax,0x10(%rdi)

clang 9.0.1 gives:

   0:   81 47 10 ff ff ff ff    addl   $0xffffffff,0x10(%rdi)

But actually, clang does equally as well with:

#define ADD_ONCE(d, i) *(volatile typeof(d) *)&(d) += (i)

And testing further back, it generates the same code with your original
WRITE_ONCE.

If clang's optimization here is technically correct, maybe we should go
talk to the gcc people about catching this case?