| Message ID | 20240327033041.83625-1-jaewon02.kim@samsung.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | spi: s3c64xx: Use DMA mode from fifo size | expand |
On Tue, Mar 26, 2024 at 10:35 PM Jaewon Kim <jaewon02.kim@samsung.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@samsung.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 > >
Hi Sam, Thanks for your review. On 3/29/24 02:58, Sam Protsenko wrote: > On Tue, Mar 26, 2024 at 10:35 PM Jaewon Kim<jaewon02.kim@samsung.com> wrote: >> The SPI data size is smaller than FIFO, it operates in PIO mode, > Spelling: "The" -> "If the" Thanks, I will fix it v2. >> 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" Thanks, I will fix it v2. >> 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). In IRQ mode(S3C64XX_SPI_MODE_RX_RDY_LVL enable), TxOverrun/RxUnderrun irq occurs when FIFO is full. To avoid FIFO full, it is transmitted in a smaller size than fifo_len.(fifo-len - 1) However, in case of "fifo_len == data size" "fifo_len - 1" byte + "1" byte were transmitted separately. This problem can be solved by starting DMA transmission start size from fifo_len. >> 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)? It is not a critical issue, but When I look at the actual waveform, it seems strange that only the last 1-byte is transmitted separately. I thought it was "Fixes", but if not, I will remove it. >> Signed-off-by: Jaewon Kim<jaewon02.kim@samsung.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. Good, I will remove it also. >> - } 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). You are correct. This 'if' condition should not have been modified. >> tx_buf = xfer->tx_buf; >> rx_buf = xfer->rx_buf; >> origin_len = xfer->len; >> -- >> 2.43.2 >> >> Thanks Jaewon Kim
On Fri, Mar 29, 2024 at 12:53 AM Jaewon Kim <jaewon02.kim@samsung.com> wrote: > > Hi Sam, > > Thanks for your review. > > > On 3/29/24 02:58, Sam Protsenko wrote: > > On Tue, Mar 26, 2024 at 10:35 PM Jaewon Kim<jaewon02.kim@samsung.com> wrote: > >> The SPI data size is smaller than FIFO, it operates in PIO mode, > > Spelling: "The" -> "If the" > > Thanks, I will fix it v2. > > >> 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" > > Thanks, I will fix it v2. > > >> 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). > > In IRQ mode(S3C64XX_SPI_MODE_RX_RDY_LVL enable), TxOverrun/RxUnderrun > irq occurs when FIFO is full. > > To avoid FIFO full, it is transmitted in a smaller size than > fifo_len.(fifo-len - 1) > > However, in case of "fifo_len == data size" "fifo_len - 1" byte + "1" > byte were transmitted separately. > > This problem can be solved by starting DMA transmission start size from > fifo_len. > Thanks for the explanation! Please feel free to add: Reviewed-by: Sam Protsenko <semen.protsenko@linaro.org> > >> 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)? > > It is not a critical issue, but When I look at the actual waveform, it > seems strange that only the last 1-byte is transmitted separately. > > I thought it was "Fixes", but if not, I will remove it. > No no, I was just curious. "Fixes" is fine with me. > >> Signed-off-by: Jaewon Kim<jaewon02.kim@samsung.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. > Good, I will remove it also. > >> - } 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). > > You are correct. This 'if' condition should not have been modified. > > >> tx_buf = xfer->tx_buf; > >> rx_buf = xfer->rx_buf; > >> origin_len = xfer->len; > >> -- > >> 2.43.2 > >> > >> > > Thanks > > Jaewon Kim >
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; - } else if (xfer->len >= fifo_len) { + } else if (xfer->len > fifo_len) { tx_buf = xfer->tx_buf; rx_buf = xfer->rx_buf; origin_len = xfer->len;
The SPI data size is smaller than FIFO, it operates in PIO mode, 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, DMA mode must be used from the case of FIFO and data size. Fixes: 1ee806718d5e ("spi: s3c64xx: support interrupt based pio mode") Signed-off-by: Jaewon Kim <jaewon02.kim@samsung.com> --- drivers/spi/spi-s3c64xx.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)