Re: Page Fault Scalability patch V19 [1/4]: pte_cmpxchg andCONFIG_ATOMIC_TABLE_OPS

[Christoph Lameter]

On Thu, 10 Mar 2005, Andi Kleen wrote:

> Christoph Lameter <clameter@xxxxxxx> writes:
> >
> > Atomic operations may be enabled in the kernel configuration on
> > i386, ia64 and x86_64 if a suitable CPU is configured in SMP mode.
> > Generic atomic definitions for ptep_xchg and ptep_cmpxchg
> > have been provided based on the existing xchg() and cmpxchg() functions
> > that already work atomically on many platforms. It is very
>
> I'm curious - do you have any micro benchmarks on i386 or x86-64 systems
> about the difference between spin_lock(ptl) access; spin_unlock(ptl);
> and cmpxchg ?

There is a benchmark for UP on
http://oss.sgi.com/projects/page_fault_performance.

> cmpxchg can be quite slow, with bad luck it could be slower than
> the spinlocks.

Spinlocks also require atomic operations like a lock decb on i386 in order
to acquire the locks. And the page_table_lock is acquired two
times in the page fault handler. In order for this to be faster

2*spinlock acquisition and release would have to be faster than a cmpxchg.

> A P4 would be good to benchmark this because it seems to be the worst
> case.

The numbers on the webpage are for an AMD64. But I can try
to get some testing done on a P4 too.
-
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/