Add DMA engine support for transmit to the PL011 driver.  Based on a
patch from Linus Walliej, with the following changes:

- remove RX DMA support.  As PL011 doesn't give us receive timeout
  interrupts, we only get notified of received data when the RX DMA
  has completed.  This rather sucks for interactive use of the TTY.

- remove abuse of completions.  Completions are supposed to be for
  events, not to tell what condition buffers are in.  Replace it with
  a simple 'queued' bool.

- fix locking - it is only safe to access the circular buffer with the
  port lock held.

- only map the DMA buffer when required - if we're ever behind an IOMMU
  this helps keep IOMMU usage down, and also ensures that we're legal
  when we change the scatterlist entry length.

- fix XON/XOFF sending - we must send XON/XOFF characters out as soon
  as possible - waiting for up to 4095 characters in the DMA buffer
  to be sent first is not acceptable.

- fix XON/XOFF receive handling - we need to stop DMA when instructed
  to by the TTY layer, and restart it again when instructed to.  There
  is a subtle problem here: we must not completely empty the circular
  buffer with DMA, otherwise we will not be notified of XON.

- change the 'enable_dma' flag into a 'using DMA' flag, and track
  whether we can use TX DMA by whether the channel pointer is non-NULL.
  This gives us more control over whether we use DMA in the driver.

- we don't need to have the TX DMA buffer continually allocated for
  each port - instead, allocate it when the port starts up, and free
  it when it's shut down.  Update the 'using DMA' flag if we get
  the buffer, and adjust the TTY FIFO size appropriately.

- if we're going to use PIO to send characters, use the existing IRQ
  based functionality rather than reimplementing it.  This also ensures
  we call uart_write_wakeup() at the appropriate time, otherwise we'll
  stall.

- use DMA engine helper functions for type safety.

- fix init when built as a module - we can't have to initcall functions,
  so we must settle on one.  This means we can eliminate the deferred
  DMA initialization.

- there is no need to terminate transfers on a failed prep_slave_sg()
  call - nothing has been setup, so nothing needs to be terminated.
  This avoids a potential deadlock in the DMA engine code
  (tasklet->callback->failed prepare->terminate->tasklet_disable
   which then ends up waiting for the tasklet to finish running.)

- Dan says that the submission callback should not return an error:
  | dma_submit_error() is something I should have removed after commit
  | a0587bcf "ioat1: move descriptor allocation from submit to prep" all
  | errors should be notified by prep failing to return a descriptor
  | handle.  Negative dma_cookie_t values are only returned by the
  | dma_async_memcpy* calls which translate a prep failure into -ENOMEM.
  So remove the error handling at that point.  This also solves the
  potential deadlock mentioned in the previous comment.
Acked-by: Linus Walleij <linus.walleij@stericsson.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>