cachefs broken on ppc64

cachefs broken on ppc64

[Paul Mackerras]

David,

I just tried compiling 2.6.10-rc2-mm2, and just for fun, I turned on
cachefs, and found out that cachefs won't build on ppc64.  The problem
is that it is using xchg() on 16-bit quantities, which we don't
support on ppc (32 or 64).  Is there a good reason why
cachefs_super.ujnl_serial has to be 16 bits rather than 32?

It worries me a bit that you are using xchg() so much, actually.  It
feels like you are trying to be clever and do things without taking
any locks, but there are no memory barriers anywhere in fs/cachefs
that I could see.  So I suspect it would have problems on SMP ppc64 or
ia64 systems, which have weak memory consistency.  Or is there some
serialization at a higher level that saves you?  (If so, why do you
need to use xchg()?)

Paul.

Re: cachefs broken on ppc64

[David Howells]

> I just tried compiling 2.6.10-rc2-mm2, and just for fun, I turned on
> cachefs, and found out that cachefs won't build on ppc64.  The problem
> is that it is using xchg() on 16-bit quantities, which we don't
> support on ppc (32 or 64).  Is there a good reason why
> cachefs_super.ujnl_serial has to be 16 bits rather than 32?

Because the on-disc value is 16 bits, I suppose.

Actually, if you are looking at the one in journal.c, that probably doesn't
need the xchg() at all. There's a write-locked rwsem covering the region.

> It worries me a bit that you are using xchg() so much, actually.  It
> feels like you are trying to be clever and do things without taking
> any locks, but there are no memory barriers anywhere in fs/cachefs
> that I could see.

In many of the cases, there're spinlocks or semaphores covering the xchg(), so
those don't really need to be xchg calls, I suppose.

> So I suspect it would have problems on SMP ppc64 or ia64 systems, which have
> weak memory consistency.  Or is there some serialization at a higher level
> that saves you?  (If so, why do you need to use xchg()?)

There is serialisation in the form of spinlocks and semaphores covering many
of the exchanges, but not all. I'm not entirely sure where I need memory
barriers - i386 doesn't really have them.

David