Re: [RFC][PATCH] ia64: Fix atomic ops vs memory barriers

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Feb 04, 2014 at 08:29:36AM -0800, Linus Torvalds wrote:
> On Tue, Feb 4, 2014 at 4:22 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> >
> > The below patch assumes the SDM is right (TM), and fixes the atomic_t,
> > cmpxchg() and xchg() implementations by inserting a mf before the
> > cmpxchg.acq (or xchg).
> 
> You picked the wrong thing to be right. The SDM is wrong.

Figured, just my luck :-)

> Last time this came up, Tony explained it thus:
> 
> >> Worse still - early processor implementations actually just ignored
> >> the acquire/release and did a full fence all the time.  Unfortunately
> >> this meant a lot of badly written code that used .acq when they really
> >> wanted .rel became legacy out in the wild - so when we made a cpu
> >> that strictly did the .acq or .rel ... all that code started breaking - so
> >> we had to back-pedal and keep the "legacy" behavior of a full fence :-(

That would make a lovely comment near ia64_cmpxchg().

> and since ia64 is basically on life support as an architecture, we can
> pretty much agree that the SDM is dead, and the only thing that
> matters is implementation.
> 
> The above quote was strictly in the context of just cmpxchg, though,
> so it's possible that the "fetchadd" instruction acts differently. I
> would personally expect it to have the same issues, but let's see what
> Tony says.. Tony?

I would suspect it to be a full fence too, let me do the reverse patch.

---
--- a/arch/ia64/include/asm/bitops.h
+++ b/arch/ia64/include/asm/bitops.h
@@ -65,11 +65,8 @@ __set_bit (int nr, volatile void *addr)
 	*((__u32 *) addr + (nr >> 5)) |= (1 << (nr & 31));
 }
 
-/*
- * clear_bit() has "acquire" semantics.
- */
-#define smp_mb__before_clear_bit()	smp_mb()
-#define smp_mb__after_clear_bit()	do { /* skip */; } while (0)
+#define smp_mb__before_clear_bit()	barrier();
+#define smp_mb__after_clear_bit()	barrier();
 
 /**
  * clear_bit - Clears a bit in memory
--- a/arch/ia64/include/uapi/asm/cmpxchg.h
+++ b/arch/ia64/include/uapi/asm/cmpxchg.h
@@ -118,6 +118,15 @@ extern long ia64_cmpxchg_called_with_bad
 #define cmpxchg_rel(ptr, o, n)	\
 	ia64_cmpxchg(rel, (ptr), (o), (n), sizeof(*(ptr)))
 
+/*
+ * Worse still - early processor implementations actually just ignored
+ * the acquire/release and did a full fence all the time.  Unfortunately
+ * this meant a lot of badly written code that used .acq when they really
+ * wanted .rel became legacy out in the wild - so when we made a cpu
+ * that strictly did the .acq or .rel ... all that code started breaking - so
+ * we had to back-pedal and keep the "legacy" behavior of a full fence :-(
+ */
+
 /* for compatibility with other platforms: */
 #define cmpxchg(ptr, o, n)	cmpxchg_acq((ptr), (o), (n))
 #define cmpxchg64(ptr, o, n)	cmpxchg_acq((ptr), (o), (n))