directory change notification patch

[Mark Weaver]

> I've reviewed it *very* crefully, and wrote sume test code to reproduce
> the races. His fix (in sys_select) is correct, the fix in sys_select_intr
> is a little overkill (IMHO) so I've trimmed down the paranoia a little.
> 
This looks good to me.  You are correct in that my original changes to 
sys_select_intr were overkill: they went zeroing fd_sets that were then 
reinitialised on the next loop.  (I had basically copied all of the code 
out of sys_select minus the read pipe without thinking through the 
implications properly).

> Yes, he (and you) are correct in the EINTR must have a higher priority
> than incoming data, otherwise an active client could starve asynchronous
> processing.
> 
This makes sense -- I hadn't thought about that.

Thanks,

Mark