[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CAPLW+4k4qh4ZYBufZoGbUZN0yxSE2X8bOdkEQVw1Zg9YUVpbug@mail.gmail.com>
Date: Thu, 28 Mar 2024 12:58:07 -0500
From: Sam Protsenko <semen.protsenko@...aro.org>
To: Jaewon Kim <jaewon02.kim@...sung.com>
Cc: Andi Shyti <andi.shyti@...nel.org>, Mark Brown <broonie@...nel.org>,
Krzysztof Kozlowski <krzysztof.kozlowski@...aro.org>, Alim Akhtar <alim.akhtar@...sung.com>,
linux-spi@...r.kernel.org, linux-samsung-soc@...r.kernel.org,
linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] spi: s3c64xx: Use DMA mode from fifo size
On Tue, Mar 26, 2024 at 10:35 PM Jaewon Kim <jaewon02.kim@...sung.com> wrote:
>
> The SPI data size is smaller than FIFO, it operates in PIO mode,
Spelling: "The" -> "If the"
> and if it is larger than FIFO mode, DMA mode is selected.
>
> If the data size is the same as the FIFO size, it operates in PIO mode
> and data is separated into two transfer. In order to prevent,
Nit: "transfer" -> "transfers", "prevent" -> "prevent it"
> DMA mode must be used from the case of FIFO and data size.
>
You probably mean this code (it occurs two times in the driver):
xfer->len = fifo_len - 1;
Can you please elaborate on why it's done this way? Why can't we just
do "xfer->len = fifo_len" and use the whole FIFO for the transfer
instead? I don't understand the necessity to split the transfer into
two chunks if its size is of FIFO length -- wouldn't it fit into FIFO
in that case? (I'm pretty sure this change is correct, just want to
understand how exactly it works).
> Fixes: 1ee806718d5e ("spi: s3c64xx: support interrupt based pio mode")
Just wonder if that fixes some throughput regression, or something
worse (like failed transfers when the transfer size is the same as
FIFO size)?
> Signed-off-by: Jaewon Kim <jaewon02.kim@...sung.com>
> ---
> drivers/spi/spi-s3c64xx.c | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/spi/spi-s3c64xx.c b/drivers/spi/spi-s3c64xx.c
> index 9fcbe040cb2f..81ed5fddf83e 100644
> --- a/drivers/spi/spi-s3c64xx.c
> +++ b/drivers/spi/spi-s3c64xx.c
> @@ -430,7 +430,7 @@ static bool s3c64xx_spi_can_dma(struct spi_controller *host,
> struct s3c64xx_spi_driver_data *sdd = spi_controller_get_devdata(host);
>
> if (sdd->rx_dma.ch && sdd->tx_dma.ch)
> - return xfer->len > sdd->fifo_depth;
> + return xfer->len >= sdd->fifo_depth;
>
> return false;
> }
> @@ -826,11 +826,11 @@ static int s3c64xx_spi_transfer_one(struct spi_controller *host,
> return status;
> }
>
> - if (!is_polling(sdd) && (xfer->len > fifo_len) &&
> + if (!is_polling(sdd) && xfer->len >= fifo_len &&
> sdd->rx_dma.ch && sdd->tx_dma.ch) {
> use_dma = 1;
>
Would be nice to remove this empty line, while at it.
> - } else if (xfer->len >= fifo_len) {
> + } else if (xfer->len > fifo_len) {
Below in the same function I can see similar code:
if (target_len >= fifo_len)
xfer->len = fifo_len - 1;
Shouldn't that 'if' condition be fixed too? Or it's ok as it is? (Just
noticed it by searching, not sure myself, hence asking).
> tx_buf = xfer->tx_buf;
> rx_buf = xfer->rx_buf;
> origin_len = xfer->len;
> --
> 2.43.2
>
>
Powered by blists - more mailing lists