On Sat 12 Sep 16:08 CDT 2020, Douglas Anderson wrote: > When setting up a bidirectional transfer we need to program both the > TX and RX lengths. We don't need a memory barrier between those two > writes. Factor out the __iowmb() and use writel_relaxed(). This > saves a fraction of a microsecond of setup overhead on bidirectional > transfers. > > Signed-off-by: Douglas Anderson <dianders@xxxxxxxxxxxx> > --- > > drivers/spi/spi-geni-qcom.c | 13 ++++++++++--- > 1 file changed, 10 insertions(+), 3 deletions(-) > > diff --git a/drivers/spi/spi-geni-qcom.c b/drivers/spi/spi-geni-qcom.c > index 92d88bf85a90..6c7e12b68bf0 100644 > --- a/drivers/spi/spi-geni-qcom.c > +++ b/drivers/spi/spi-geni-qcom.c > @@ -376,15 +376,22 @@ static void setup_fifo_xfer(struct spi_transfer *xfer, > len &= TRANS_LEN_MSK; > > mas->cur_xfer = xfer; > + > + /* > + * Factor out the __iowmb() so that we can use writel_relaxed() for > + * both writes below and thus only incur the overhead once even if > + * we execute both of them. > + */ How many passes through this function do we have to take before saving the amount of time it took me to read this comment? Reviewed-by: Bjorn Andersson <bjorn.andersson@xxxxxxxxxx> Regards, Bjorn > + __iowmb(); > + > if (xfer->tx_buf) { > m_cmd |= SPI_TX_ONLY; > mas->tx_rem_bytes = xfer->len; > - writel(len, se->base + SE_SPI_TX_TRANS_LEN); > + writel_relaxed(len, se->base + SE_SPI_TX_TRANS_LEN); > } > - > if (xfer->rx_buf) { > m_cmd |= SPI_RX_ONLY; > - writel(len, se->base + SE_SPI_RX_TRANS_LEN); > + writel_relaxed(len, se->base + SE_SPI_RX_TRANS_LEN); > mas->rx_rem_bytes = xfer->len; > } > > -- > 2.28.0.618.gf4bc123cb7-goog >