| Message ID | 20190307072424.18820-1-jin.xiao@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [RFC] spi: pxa2xx: Do cs if restart the SSP during pxa2xx_spi_transfer_one() | expand |
Hi
Is this also related to the regression with d5898e19c0d7 ("spi: pxa2xx:
Use core message processing loop") you have found or another issue?
Comments below.
On 3/7/19 9:24 AM, xiao jin wrote:
> The spi-pxa2xx can't read and write data correctly on our board.
> The pxa_ssp_type is LPSS_BXT_SSP in our case.
>
> With more debug we find that it's related to restart the SPP
> during pxa2xx_spi_transfer_one().
>
> In the normal case the spi_transfer_one_message() calls spi-pxa2xx
> cs_assert before transferring one message. After completing the
> transfer it calls spi-pxa2xx cs_deassert. The spi-pxa2xx works
> well.
>
> But in some other case pxa2xx_spi_unprepare_transfer() is called
> that clears SSCR0_SSE bit before the next transfer. In the next
> transfer the spi-pxa2xx driver will restart the SSP as the SSE
> bit is cleared. The cs_assert before the SSP restart can't ensure
> spi-pxa2xx work well.
>
> The patch is to do cs again if spi-pxa2xx restar the SSP during
> pxa2xx_spi_transfer_one()
>
Hmm.. please correct me if I'm wrong but pxa2xx_spi_unprepare_transfer()
is called always when there is no more messages pending and the spi core
should have deasserted the CS already?
More below.
> Signed-off-by: xiao jin <jin.xiao@intel.com>
> Signed-off-by: he, bo <bo.he@intel.com>
> ---
> drivers/spi/spi-pxa2xx.c | 8 ++++++++
> 1 file changed, 8 insertions(+)
>
> diff --git a/drivers/spi/spi-pxa2xx.c b/drivers/spi/spi-pxa2xx.c
> index 14f4ea59caff..1a2ea46858d9 100644
> --- a/drivers/spi/spi-pxa2xx.c
> +++ b/drivers/spi/spi-pxa2xx.c
> @@ -928,6 +928,7 @@ static int pxa2xx_spi_transfer_one(struct spi_controller *master,
> u32 cr1;
> int err;
> int dma_mapped;
> + bool need_cs_change = false;
>
> /* Check if we can DMA this transfer */
> if (transfer->len > MAX_DMA_LEN && chip->enable_dma) {
> @@ -1056,6 +1057,11 @@ static int pxa2xx_spi_transfer_one(struct spi_controller *master,
> if ((pxa2xx_spi_read(drv_data, SSCR0) != cr0)
> || (pxa2xx_spi_read(drv_data, SSCR1) & change_mask)
> != (cr1 & change_mask)) {
> + /* It needs to deassert the chip selection
> + * firstly before restart the SPP */
> + need_cs_change = true;
> + cs_deassert(spi);
> +
I think code comes here at the beginning of each transfer so will be hit
multiple times before pxa2xx_spi_unprepare_transfer() if SPI message
consists of multiple transfers.
This makes me wondering if the device driver setting up the "struct
spi_transfer" is maybe missing the cs_change flag set for transfers
before last one in case HW needs CS toggling between transfers? For
instance what following drivers are doing with the cs_change flag:
drivers/char/tpm/tpm_tis_spi.c: tpm_tis_spi_transfer()
drivers/input/touchscreen/ad7877.c: ad7877_read(), ad7877_read_adc()
On Thu, Mar 07, 2019 at 05:26:53PM +0200, Jarkko Nikula wrote: > On 3/7/19 9:24 AM, xiao jin wrote: > > The patch is to do cs again if spi-pxa2xx restar the SSP during > > pxa2xx_spi_transfer_one() > Hmm.. please correct me if I'm wrong but pxa2xx_spi_unprepare_transfer() is > called always when there is no more messages pending and the spi core should > have deasserted the CS already? Yes. > > @@ -1056,6 +1057,11 @@ static int pxa2xx_spi_transfer_one(struct spi_controller *master, > > if ((pxa2xx_spi_read(drv_data, SSCR0) != cr0) > > || (pxa2xx_spi_read(drv_data, SSCR1) & change_mask) > > != (cr1 & change_mask)) { > > + /* It needs to deassert the chip selection > > + * firstly before restart the SPP */ > > + need_cs_change = true; > > + cs_deassert(spi); > I think code comes here at the beginning of each transfer so will be hit > multiple times before pxa2xx_spi_unprepare_transfer() if SPI message > consists of multiple transfers. > This makes me wondering if the device driver setting up the "struct > spi_transfer" is maybe missing the cs_change flag set for transfers before > last one in case HW needs CS toggling between transfers? For instance what > following drivers are doing with the cs_change flag: > drivers/char/tpm/tpm_tis_spi.c: tpm_tis_spi_transfer() > drivers/input/touchscreen/ad7877.c: ad7877_read(), ad7877_read_adc() Right, this really feels like it's fixing the wrong thing.
Hi On 3/8/2019 12:09 AM, Mark Brown wrote: > On Thu, Mar 07, 2019 at 05:26:53PM +0200, Jarkko Nikula wrote: >> On 3/7/19 9:24 AM, xiao jin wrote: >>> The patch is to do cs again if spi-pxa2xx restar the SSP during >>> pxa2xx_spi_transfer_one() >> Hmm.. please correct me if I'm wrong but pxa2xx_spi_unprepare_transfer() is >> called always when there is no more messages pending and the spi core should >> have deasserted the CS already? > Yes. I try to describe the situation in more detail. cpu0 cpu1 => spi_transfer_one_message => __spi_pump_messages => __spi_sync => spi_sync => spi_sync_transfer.constprop.2 => spi_write => fm1388_spi_burst_write => fm1388_fw_loaded => request_firmware_work_func => process_one_work => pxa2xx_spi_unprepare_transfer => spi_pump_messages => kthread_worker_fn Yes, the spi core has de-asserted the CS before the pxa2xx_spi_unprepare_transfer(). The problem on my side is that the new transfer will restart the SSP in pxa2xx_spi_transfer_one(). The spi core has asserted the CS again before restart the SSP. In my test the CS assert before the restart SSP can't ensure the spi transfer working correctly. >>> @@ -1056,6 +1057,11 @@ static int pxa2xx_spi_transfer_one(struct spi_controller *master, >>> if ((pxa2xx_spi_read(drv_data, SSCR0) != cr0) >>> || (pxa2xx_spi_read(drv_data, SSCR1) & change_mask) >>> != (cr1 & change_mask)) { >>> + /* It needs to deassert the chip selection >>> + * firstly before restart the SPP */ >>> + need_cs_change = true; >>> + cs_deassert(spi); >> I think code comes here at the beginning of each transfer so will be hit >> multiple times before pxa2xx_spi_unprepare_transfer() if SPI message >> consists of multiple transfers. >> This makes me wondering if the device driver setting up the "struct >> spi_transfer" is maybe missing the cs_change flag set for transfers before >> last one in case HW needs CS toggling between transfers? For instance what >> following drivers are doing with the cs_change flag: >> drivers/char/tpm/tpm_tis_spi.c: tpm_tis_spi_transfer() >> drivers/input/touchscreen/ad7877.c: ad7877_read(), ad7877_read_adc() > Right, this really feels like it's fixing the wrong thing. I check the cs_change flag in the spi_transfer_one_message(). The cs_change flag is used in two cases. If the transfer is the last one it is used to keep the CS assert. If the transfer is not the last one it's used to generate the 10us pulse and then de-assert the CS. In my case the transfer is the last one and it can work correctly if I re-assert the CS after restart the SSP in the pxa2xx_spi_transfer_one() which is called from spi_transfer_one_message(). From my experiments both cs_change flag cases in the spi_transfer_one_message() can't help the problem. I am not sure if I fully understand your point.
Hi Jin On 3/8/19 9:28 AM, Xiao, Jin wrote: > > Yes, the spi core has de-asserted the CS before the > pxa2xx_spi_unprepare_transfer(). The problem on my side is that the new > transfer will restart the SSP in pxa2xx_spi_transfer_one(). The spi core > has asserted the CS again before restart the SSP. In my test the CS > assert before the restart SSP can't ensure the spi transfer working > correctly. > I wanted to ask have you found any other fix or reason than the patches you have sent? I've been trying to debug why the commit d5898e19c0d7 ("spi: pxa2xx: Use core message processing loop") is causing the regression you are seeing. I've been testing this by sending two or more messages each consisting of two 4-byte transfers through the spidev. SPI mode 3 and clock is 1 MHz. I see only slight timing difference between a commit before and at d5898e19c0d7 when the chip select is active. 8ae55af38817: CS act to CLK running ~40 us. CLK idle to CS deact ~26 us d5898e19c0d7: CS act to CLK running ~34 us. CLK idle to CS deact ~18 us There is more difference/variability during the times when the CS is not active between messages. Which is understandable since there happens message passing from userspace, queing, etc. At 8ae55af38817 total time over 2 messages varies from ~600 us to 1200 ~us and at d5898e19c0d7 from ~500 us to 1100 us. Then at v4.19 it goes a bit slover and varies between ~700 to 1600 us. Most probably due some other change than in SPI subsystem as there is not commit either in SPI core or spi-pxa2xx explaining it. Or just some random config change in my kernel configuration between 4.17 and 4.19. What I don't understand why additional CS toggling makes it work for you in case SSP was stopped in pxa2xx_spi_unprepare_transfer() and restarted again in pxa2xx_spi_transfer_one(). This SSP stopping and restarting is not visible in the SPI signals since clock is already stopped when FIFO gets empty. I guess the reason why you are seeing pxa2xx_spi_unprepare_transfer() only called sometimes is that in those cases where it is not called a new SPI message got queued while processing the current one and in other cases queue became empty before new message was queued. In both cases there shouldn't be difference in CS handling. Which makes me scratching my head why additional CS toggling is fixing the issue in case there is SSP restart. And which was due more time elapsed between messages and queue went empty.
diff --git a/drivers/spi/spi-pxa2xx.c b/drivers/spi/spi-pxa2xx.c index 14f4ea59caff..1a2ea46858d9 100644 --- a/drivers/spi/spi-pxa2xx.c +++ b/drivers/spi/spi-pxa2xx.c @@ -928,6 +928,7 @@ static int pxa2xx_spi_transfer_one(struct spi_controller *master, u32 cr1; int err; int dma_mapped; + bool need_cs_change = false; /* Check if we can DMA this transfer */ if (transfer->len > MAX_DMA_LEN && chip->enable_dma) { @@ -1056,6 +1057,11 @@ static int pxa2xx_spi_transfer_one(struct spi_controller *master, if ((pxa2xx_spi_read(drv_data, SSCR0) != cr0) || (pxa2xx_spi_read(drv_data, SSCR1) & change_mask) != (cr1 & change_mask)) { + /* It needs to deassert the chip selection + * firstly before restart the SPP */ + need_cs_change = true; + cs_deassert(spi); + /* stop the SSP, and update the other bits */ pxa2xx_spi_write(drv_data, SSCR0, cr0 & ~SSCR0_SSE); if (!pxa25x_ssp_comp(drv_data)) @@ -1070,6 +1076,8 @@ static int pxa2xx_spi_transfer_one(struct spi_controller *master, pxa2xx_spi_write(drv_data, SSTO, chip->timeout); } + if (need_cs_change) + cs_assert(spi); /* * Release the data by enabling service requests and interrupts, * without changing any mode bits