dspi->words_in_flight is a variable populated in the *_write functions
and used in the dspi_fifo_read function. It is also used in
dspi_fifo_write, immediately after transmission, to update the
message->actual_length variable used by higher layers such as spi-mem
for integrity checking.

But it may happen that the IRQ which calls dspi_fifo_read to be
triggered before the updating of message->actual_length takes place. In
that case, dspi_fifo_read will decrement dspi->words_in_flight to -1,
and that will cause an invalid modification of message->actual_length.

For that, we make the simplest fix possible: to not decrement the actual
shared variable in dspi->words_in_flight from dspi_fifo_read, but
actually a copy of it which is on stack.

But even if dspi_fifo_read from the next IRQ does not interfere with the
dspi_fifo_write of the current chunk, the *next* dspi_fifo_write still
can. So we must assume that everything after the last write to the TX
FIFO can be preempted by the "TX complete" IRQ, and the dspi_fifo_write
function must be safe against that. This means refactoring the 2
flavours of FIFO writes (for EOQ and XSPI) such that the calculation of
the number of words to be written is common and happens a priori. This
way, the code for updating the message->actual_length variable works
with a copy and not with the volatile dspi->words_in_flight.

After some interior debate, the dspi->progress variable used for
software timestamping was *not* backed up against preemption in a copy
on stack. Because if preemption does occur between
spi_take_timestamp_pre and spi_take_timestamp_post, there's really no
point in trying to save anything. The first-in-time
spi_take_timestamp_post call with a dspi->progress higher than the
requested xfer->ptp_sts_word_post will trigger xfer->timestamped = true
anyway and will close the deal.

To understand the above a bit better, consider a transfer with
xfer->ptp_sts_word_pre = xfer->ptp_sts_word_post = 3, and
xfer->bits_per_words = 8 (so byte 3 needs to be timestamped). The DSPI
controller timestamps in chunks of 4 bytes at a time, and preemption
occurs in the middle of timestamping the first chunk:

  spi_take_timestamp_pre(0)
    .
    . (preemption)
    .
    . spi_take_timestamp_pre(4)
    .
    . spi_take_timestamp_post(7)
    .
  spi_take_timestamp_post(3)

So the reason I'm not bothering to back up dspi->progress for that
spi_take_timestamp_post(3) is that spi_take_timestamp_post(7) is going
to (a) be more honest, (b) provide better accuracy and (c) already
render the spi_take_timestamp_post(3) into a noop by setting
xfer->timestamped = true anyway.

Fixes: d59c90a2 ("spi: spi-fsl-dspi: Convert TCFQ users to XSPI FIFO mode")
Reported-by: Michael Walle <michael@walle.cc>
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Tested-by: Michael Walle <michael@walle.cc>
Link: https://lore.kernel.org/r/20200318001603.9650-6-olteanv@gmail.comSigned-off-by: Mark Brown <broonie@kernel.org>