On Fri, Jul 02, 2004 at 01:00:19PM -0700, Chris Caputo wrote:
> On Fri, 25 Jun 2004, Marcelo Tosatti wrote:
> > On Wed, Jun 23, 2004 at 06:50:48PM -0700, Chris Caputo wrote:
> > > Is it safe to assume that the x86 version of atomic_dec_and_lock(), which
> > > iput() uses, is well trusted?  I figure it's got to be, but doesn't hurt
> > > to ask.
> > 
> > Pretty sure it is, used all over. You can try to use non-optimize version 
> > at lib/dec_and_lock.c for a test.
> 
> My current theory is that occasionally when irqbalance changes CPU
> affinities that the resulting set_ioapic_affinity() calls somehow cause
> either inter-CPU locking or cache coherency or ??? to fail.

or.... some spinlock is just incorrect and having the irqbalance irqlayout
unhides that.. irqbalance only balances very very rarely so I doubt it's the
cause of anything...