Re: [PATCH] i386 spinlocks should use the full 32 bits, not only 8 bits

From: Andi Kleen
Date: Tue Oct 11 2005 - 12:59:35 EST


Linus Torvalds <torvalds@xxxxxxxx> writes:

> On Tue, 11 Oct 2005, Eric Dumazet wrote:
> >
> > As NR_CPUS might be > 128, and every spining CPU decrements the lock, we need
> > to use more than 8 bits for a spinlock. The current (i386/x86_64)
> > implementations have a (theorical) bug in this area.

I already queued a patch for x86-64 for this for post 2.6.14
following earlier complaints from Eric.

iirc on 32bit there are other issues though like limited
number of reader lock users.

But then I'm not sure why anybody would want to run a 32bit
kernel on such a big machine ...

> I don't think there are any x86 machines with > 128 CPU's right now.

Someone at Unisys just needs to put dual core hyper threaded CPUs
into their 64 socket systems, then you'll have a 256 CPU x86 machine
(which is currently the maximum the APICs can take anyways...)
With Multicores being on everybody's roadmap we'll probably see
such big systems soon.

-Andi
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/