When chasing down some of those 'bad entry' diagnostics, I once
got suspicious that the problem was slab corruption coming from
the way the td hashtable code worked.   So I put together this
patch, eliminating some kmallocation, and the next times I ran
that test, the oops went away and it worked like a charm.  Hmm.

This patch is good because it shrinks memory and code, and gets
rid of some could-fail allocations, so I figured I'd send it on
(low priority) even if I don't think it fixes the root problem.