Re: [HACKERS] Deadlock in XLogInsert at AIX

On 2018-01-16 16:12:11 +0900, Michael Paquier wrote:
> On Fri, Feb 03, 2017 at 12:26:50AM +0000, Noah Misch wrote:
> > On Wed, Feb 01, 2017 at 02:39:25PM +0200, Heikki Linnakangas wrote:
> >> @@ -73,11 +73,19 @@ pg_atomic_compare_exchange_u32_impl(volatile pg_atomic_uint32 *ptr,
> >> static inline uint32
> >> pg_atomic_fetch_add_u32_impl(volatile pg_atomic_uint32 *ptr, int32 add_)
> >> {
> >> + uint32 ret;
> >> +
> >> /*
> >> - * __fetch_and_add() emits a leading "sync" and trailing "isync", thereby
> >> - * providing sequential consistency. This is undocumented.
> >> + * Use __sync() before and __isync() after, like in compare-exchange
> >> + * above.
> >> */
> >> - return __fetch_and_add((volatile int *)&ptr->value, add_);
> >> + __sync();
> >> +
> >> + ret = __fetch_and_add((volatile int *)&ptr->value, add_);
> >> +
> >> + __isync();
> >> +
> >> + return ret;
> >> }
> >
> > Since this emits double syncs with older xlc, I recommend instead replacing
> > the whole thing with inline asm. As I opined in the last message of the
> > thread you linked above, the intrinsics provide little value as abstractions
> > if one checks the generated code to deduce how to use them. Now that the
> > generated code is xlc-version-dependent, the port is better off with
> > compiler-independent asm like we have for ppc in s_lock.h.
>
> Could it be cleaner to just use __xlc_ver__ to avoid double syncs on
> past versions? I think that it would make the code more understandable
> than just listing directly the instructions.

Given the quality of the intrinsics on AIX, see past commits and the
comment in the code quoted above, I think we're much better of doing
this via inline asm.

Greetings,

Andres Freund