MFC r212109,212124,212185:

MFC r212109,212124,212185:
r212109:
  bus_dma(9) cleanup.
   o Enforce TX/RX descriptor ring alignment. NS data sheet says the
     controller needs 4 bytes alignment but use 16 to cover both SiS
     and NS controllers. I don't have SiS data sheet so I'm not sure
     what is alignment restriction of SiS controller but 16 would be
     enough because it's larger than the size of a TX/RX descriptor.
     Previously sis(4) ignored the alignment restriction.
   o Enforce RX buffer alignment, 4.
     Previously sis(4) ignored RX buffer alignment restriction.
   o Limit number of TX DMA segment to be used to 16. It seems
     controller has no restriction on number of DMA segments but
     using more than 16 looks resource waste.
   o Collapse long mbuf chains with m_collapse(9) instead of calling
     expensive m_defrag(9).
   o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove
     unnecessary callbacks.
   o Initial endianness support.
   o Prefer local alignment fixup code to m_devget(9).
   o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying
     these maps in fast TX/RX path. On non-x86 architectures, this is
     very expensive operation and there is no need to do that.
   o Add missing bus_dmamap_sync(9) in TX/RX path.
   o watchdog is now unarmed only when there are no pending frames
     on controller. Previously sis(4) blindly unarmed watchdog
     without checking the number of queued frames.
   o For efficiency, loaded DMA map is reused for error frames.
   o DMA map loading failure is now gracefully handled. Previously
     sis(4) ignored any DMA map loading errors.
   o Nuke unused macros which are not appropriate for endianness
     operation.
   o Stop embedding driver maintained structures into descriptor
     rings. Because TX/RX descriptor structures are shared between
     host and controller, frequent bus_dmamap_sync(9) operations are
     required in fast path. Embedding driver structures will increase
     the size of DMA map which in turn will slow down performance.

r212124:
  Fix stupid error in r212109 which didn't swap DMA maps. This caused
  IOMMU panic on sparc64 under high TX load.

r212185:
  Fix another bug introduced in r212109. We should unload DMA maps
  only after sending the last fragment of a frame so the mbuf pointer
  also should be stored in the last descriptor index.

git-svn-id: svn://svn.freebsd.org/base/stable/8@213612 ccf9f872-aa2e-dd11-9fc8-001c23d0bc1f