Attached are 6 patches that try to improve the performance of sys_poll 
and sys_select:

- avoid dynamic memory allocations, stack storage is sufficient for most 
callers and faster.
- use the wakeup callbacks and use that info to speed up the 2nd scan 
for new events.

What do you think? Are there any apps/tests/benchmarks that stress 
sys_poll or sys_select?

The first 3 patches replace dynamic memory allocations with stack storage.
The 4th and 5th patch use wait queue callbacks for a more efficient 2nd 
scan.
The 6th patch merges common code.

The patch is against 2.5.48 - they do boot on my laptop, but that's all 
I can guarantee.

--
    Manfred