Use kmalloc() instead of slab to allocate the MD and MSG portal descriptors.
On x86_64 the MD and MSG structs got quite large, into the 2k range, though
pages are still 4k there.  This tricked slab into trying to back their slabs
with multi-page allocations.  These allocations failed under high load which
lead to LBUG()s and timeouts.

While we're in here, properly initialize the in_use counts and stop using
sleeping allocations while holding the state lock.