The kmalloc was taking up about 1.5% of the CPU on an ioctl-heavy workload
(x11perf -aa10text on 965).  Initial results look like they have a
corresponding improvement in performance for aa10text, but more numbers might
not hurt.

Thanks to ajax for pointing out this performance regression I'd introduced
back in 2007.

[airlied: well I introduced it sneakily inside Eric's patch]
Signed-off-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Dave Airlie <airlied@redhat.com>