Re: Possible memory ordering bug in page reclaim?

["David S. Miller" <davem@davemloft.net>]

From: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Date: Sun, 16 Oct 2005 23:36:00 +0400

> My opinion is that we don't need a barrier even _after_ ll/sc sequences
> on Alpha - it's redundant; perhaps it's just a historical baggage.
> I strongly believe that ll/sc _does_ imply an SMP memory barrier, as can
> be seen from instructions timing: a standalone mb takes tens of cycles,
> the mb before/after ll/sc takes 0 to 1 cycle on ev6 (a bit more on ev56)
> depending on instruction slotting.
> Note that superfluous mb's around atomic stuff still can hurt -
> Alpha mb instruction also flushes IO write buffers, so it can
> be _extremely_ expensive under heavy IO...

Even a quick google tells me that on Alpha, LL/SC have implicit
barriers only with respect to the locations being acted upon
by the LL/SC, but don't have more general barrier properties.

This is in line with how PPC is defined as well.

I truly believe that you would be removing those Alpha memory barriers
only at your own peril. :-)  And since the PPC and Alpha defined
semantics aparently match, I think it would be wise to add the missing
memory barriers from the front of the LL/SC sequences which need them,
just as PPC does.

I just spent 6 months tracking down a bug that turned out to be a
couple of missing memory barriers on sparc64.  So, this has taught me
to be extra careful in this area :-)  I really think this is an area
in which it's better to be safe than sorry.