As reported on the mailing list:
http://thread.gmane.org/gmane.linux.drivers.madwifi.devel/4744
In trunk after 0.9.3, at some point, scanning has become much more system intensive, to the degree where it noticably interrupts our realtime 3D graphics rendering.
I suspect this is because as of post-0.9.3, the device is generating an order of magnitude more interrupts when scanning. I'm measuring the interrupt count from /proc/interrupts, sampled every second for a 5 minute period (and then averaged to per-second values):
With 0.9.2 and 0.9.3, I find that when trying to make the device scan a lot, I get up to 100 interrupts per second. Not unreasonable for all the channel changing going on.
With post-0.9.3 (tested r2376 and r2518), I generally get more than 2000 interrupts per second performing the same test. The highest value I saw was 4600 interrupts per second -- rather excessive.