[squid-users] Pegging CPU with epoll_wait

I'm running Squid 2.6STABLE6 (the RedHat-distributed version) on a stock
RHEL 5.1 64-bit server with kernel 2.6.18-53.1.13.el5. Some time after
restarting squid (maybe a few hours or a few days), it starts pegging
the CPU at 100%. Running strace on the squid processes scrolls:

    epoll_wait(3, {}, 256, 0) = 0

as fast as my screen will scroll. Restarting squid makes it settle down
again for a while. This server sees only a few hits an hour. What's
causing this, and how do I stop it?