[ofw] memory 'leak' running ndconn

[Sean Hefty]

>number created increases faster than the number destroyed.
>
>created 2766,  destroyed 1762
>created 4934,  destroyed 2927
>created 6295,  destroyed 3292
>created 7604,  destroyed 3598
>created 8903,  destroyed 3902
>created 10169, destroyed 4168
>created 11460, destroyed 4457
>
>The result is a pathetic connection rate, with a whole lot of memory
>being consumed.  This may be a scheduling issue, but I'm guessing that
>somehow specific connections are not disconnecting properly.

I added more debug information to the test.  I gave every endpoint an
'order' number for its creation, along with the 'order' of its
destruction.  I compared the two order numbers to see if numerous
connections were stuck in the connect/disconnect somehow.

(The thought being that if the 1000th endpoint created was the 500th one
destroyed, then there was a good chance that 500 connections were
stuck.)

The result was that even though there were over 4000 more endpoints
created than destroyed at one point, the create/destroy order was rarely
off by over 100.

Based on this, I don't think we're dealing with a memory leak.  Winverbs
limits the number of work items that can be in progress at once, so
maybe one of the work items is starving a system thread, leading to a
slow down in work item processing.