Le vendredi 28 octobre 2011 à 05:40 -0700, Linus Torvalds a écrit :

On Fri, Oct 28, 2011 at 5:19 AM, Linus Torvalds
[off-list ref] wrote:

quoted

"Sane interfaces" are important. Insane interfaces lead to bugs.

Qutie frankly, if I do "atomic_read()", I do expect to get a single
value. If I don't get a single value, but some mixture of two values,
I'd personally go

  wtf, what does that "atomic" mean in "atomic_read()"?

and I think that's a reasonable wtf to ask.

That said, as mentioned, I don't know of any way to tell gcc "at most once".

Hmm.

Except perhaps using inline asm. Something like this might work:

  static inline int atomic_read(const atomic_t *v)
  {
       int val;
       asm("":"=r" (val):"0" (v->value));
       return val;
  }

(totally untested, but you get the idea: use a non-volatile asm to
make sure that gcc doesn't think it can re-load the value).

That's the trick we use in asmlinkage_protect() and a couple of other
places. It *should* make gcc able to optimize the value away entirely
if it isn't used, but will stop gcc from doing the reload magic.

Does that work for the test-case with VM_BUG_ON()?

On x86 it seems to work :

c050f80f: 0f 84 ea 00 00 00       je     c050f8ff <tcp_sendmsg+0xa1f>
c050f815: 8b 55 bc                mov    -0x44(%ebp),%edx
c050f818: f0 ff 42 10             lock incl 0x10(%edx)      atomic_inc(&page->count)
c050f81c: 8b 02                   mov    (%edx),%eax        page->flags
c050f81e: 25 00 40 02 00          and    $0x24000,%eax
c050f823: 3d 00 40 02 00          cmp    $0x24000,%eax
c050f828: 0f 84 9f 01 00 00       je     c050f9cd <tcp_sendmsg+0xaed>

On x86_64 (gcc-4.6.1 Debian-4.6.1-4) we still have a page->flags useless load,
but the atomic_read() itself is removed.


This looks like a CONFIG_PAGEFLAGS_EXTENDED difference.

In this case, PageTail() is :

#define TESTPAGEFLAG(uname, lname)                  \
static inline int Page##uname(const struct page *page)          \
	{ return test_bit(PG_##lname, &page->flags); } 


So we need a similar idea to remove the volatile from :

static __always_inline int constant_test_bit(unsigned int nr, const volatile unsigned long *addr)
{
return ((1UL << (nr % BITS_PER_LONG)) &
	 (addr[nr / BITS_PER_LONG])) != 0;
}