serial: sh-sci: Fix TX DMA buffer flushing and workqueue races

When uart_flush_buffer() is called, the .flush_buffer() callback zeroes
the tx_dma_len field.  This may race with the work queue function
handling transmit DMA requests:

  1. If the buffer is flushed before the first DMA API call,
     dmaengine_prep_slave_single() may be called with a zero length,
     causing the DMA request to never complete, leading to messages
     like:

        rcar-dmac e7300000.dma-controller: Channel Address Error happen

     and, with debug enabled:

	sh-sci e6e88000.serial: sci_dma_tx_work_fn: ffff800639b55000: 0...0, cookie 126

     and DMA timeouts.

  2. If the buffer is flushed after the first DMA API call, but before
     the second, dma_sync_single_for_device() may be called with a zero
     length, causing the transmit data not to be flushed to RAM, and
     leading to stale data being output.

Fix this by:
  1. Letting sci_dma_tx_work_fn() return immediately if the transmit
     buffer is empty,
  2. Extending the critical section to cover all DMA preparational work,
     so tx_dma_len stays consistent for all of it,
  3. Using local copies of circ_buf.head and circ_buf.tail, to make sure
     they match the actual operation above.

Reported-by: Eugeniu Rosca <erosca@de.adit-jv.com>
Suggested-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Reviewed-by: Eugeniu Rosca <erosca@de.adit-jv.com>
Tested-by: Eugeniu Rosca <erosca@de.adit-jv.com>
Link: https://lore.kernel.org/r/20190624123540.20629-2-geert+renesas@glider.be


Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>