MFC r276480, r276485, r276498, r277225, r277226, r277227, r277230,

MFC r276480, r276485, r276498, r277225, r277226, r277227, r277230,
r277637, and r283149 (by emaste@).

r276485 is the real change here, the rest deal with the fallout of
mp_ring's reliance on 64b atomics.

Use the incorrectly spelled 'eigth' from struct pkthdr in this branch
instead of MFC'ing r261733, which would have renamed the field of a
public structure in a -STABLE branch.
---

r276480:
Temporarily unplug cxgbe(4) from !amd64 builds.

r276485:
cxgbe(4): major tx rework.

a) Front load as much work as possible in if_transmit, before any driver
lock or software queue has to get involved.

b) Replace buf_ring with a brand new mp_ring (multiproducer ring).  This
is specifically for the tx multiqueue model where one of the if_transmit
producer threads becomes the consumer and other producers carry on as
usual.  mp_ring is implemented as standalone code and it should be
possible to use it in any driver with tx multiqueue.  It also has:
- the ability to enqueue/dequeue multiple items.  This might become
  significant if packet batching is ever implemented.
- an abdication mechanism to allow a thread to give up writing tx
  descriptors and have another if_transmit thread take over.  A thread
  that's writing tx descriptors can end up doing so for an unbounded
  time period if a) there are other if_transmit threads continuously
  feeding the sofware queue, and b) the chip keeps up with whatever the
  thread is throwing at it.
- accurate statistics about interesting events even when the stats come
  at the expense of additional branches/conditional code.

The NIC txq lock is uncontested on the fast path at this point.  I've
left it there for synchronization with the control events (interface
up/down, modload/unload).

c) Add support for "type 1" coalescing work request in the normal NIC tx
path.  This work request is optimized for frames with a single item in
the DMA gather list.  These are very common when forwarding packets.
Note that netmap tx in cxgbe already uses these "type 1" work requests.

d) Do not request automatic cidx updates every 32 descriptors.  Instead,
request updates via bits in individual work requests (still every 32
descriptors approximately).  Also, request an automatic final update
when the queue idles after activity.  This means NIC tx reclaim is still
performed lazily but it will catch up quickly as soon as the queue
idles.  This seems to be the best middle ground and I'll probably do
something similar for netmap tx as well.

e) Implement a faster tx path for WRQs (used by TOE tx and control
queues, _not_ by the normal NIC tx).  Allow work requests to be written
directly to the hardware descriptor ring if room is available.  I will
convert t4_tom and iw_cxgbe modules to this faster style gradually.

r276498:
cxgbe(4): remove buf_ring specific restriction on the txq size.

r277225:
Make cxgbe(4) buildable with the gcc in base.

r277226:
Allow cxgbe(4) to be built on i386.  Driver attach will succeed only on
a subset of i386 systems.

r277227:
Plug cxgbe(4) back into !powerpc && !arm builds, instead of building it
on amd64 only.

r277230:
Build cxgbe(4) on powerpc64 too.

r277637:
Make sure the compiler flag to get cxgbe(4) to compile with gcc is used
only when gcc is being used.  This is what r277225 should have been.

git-svn-id: svn://svn.freebsd.org/base/stable/10@284052 ccf9f872-aa2e-dd11-9fc8-001c23d0bc1f