[patch] Memcached UDP

[Steven Grimm]

Part of this patch is wrong and will introduce failures. It changes  
the flow of control from

clean up connection data structure
if connection input buffer size is over a threshold, free the conn  
structure
else add it to the freelist

to

clean up connection data structure
add connection to the freelist
if connection input buffer size is over a threshold, free the conn  
structure

The problem is that if the buffer size is large, the freelist now  
contains a pointer to freed memory which will doubtless be allocated  
to something else at some point.

The other part of the patch (allocating a msghdr if needed) seems  
harmless to me. I haven't seen the failure mode it's apparently  
intended to address -- the UDP protocol is running fine on our  
network, processing billions of requests a day with no crashes or  
bogus responses -- but it doesn't seem like it can hurt anything and  
is probably a smart sanity check.

Any chance you (or the original author) can supply a test case to  
demonstrate the UDP failures? I saw Tomash Brechko's translation  
(thanks!) but the problem description is too vague.

-Steve