[PATCH v2] firmware: xilinx: Clear IOCTL_SET_SD_TAPDELAY using PM_MMIO_WRITE

[Michal Simek]

On 10/20/22 22:02, Marek Vasut wrote:
> In case the tap delay required by Arasan SDHCI is set to 0, the current
> embeddedsw firmware unconditionally writes IOU_SLCR SD_ITAPDLY to 0x100
> (SD0_ITAPDLYENA=1, SD0_ITAPDLYSEL=0). Previous behavior was to keep the
> IOU_SLCR SD_ITAPDLY set to 0x0. There is some sort of difference in the
> behavior between SD0_ITAPDLYENA=1/0 with the same SD0_ITAPDLYSEL=0, even
> though the behavior should be identical -- zero delay added to rxclk_in
> line. The former breaks HS200 training in low temperature conditions.
> 
> Write IOU_SLCR SD_ITAPDLY register to 0 using PM_MMIO_WRITE which seem
> to allow unrestricted WRITE access (and PM_MMIO_READ which allows read
> access) to the entire address space. This way, it is possible to work
> around the defect in IOCTL_SET_SD_TAPDELAY design which does not permit
> clearing SDx_ITAPDLYENA bit.
> 
> Note that the embeddedsw firmware does not permit clearing the SD_ITAPDLY
> SD0_ITAPDLYENA bit, this bit can only ever be set by the firmware and it
> is often impossible to update the possibly broken firmware.
> 
> Signed-off-by: Marek Vasut <marex at denx.de>
> ---
> Cc: Abhyuday Godhasara <abhyuday.godhasara at xilinx.com>
> Cc: Harsha <harsha.harsha at xilinx.com>
> Cc: Michal Simek <michal.simek at amd.com>
> Cc: Rajan Vaja <rajan.vaja at xilinx.com>
> Cc: Ronak Jain <ronak.jain at xilinx.com>
> Cc: Tanmay Shah <tanmay.shah at xilinx.com>
> To: linux-arm-kernel at lists.infradead.org
> ---
> V2: - Use PM_MMIO_WRITE to clear SD_xTAPDLYx bitfields and work around
>        the IOCTL_SET_SD_TAPDELAY design defect.
>      - Update commit message accordingly.
> ---
>   drivers/firmware/xilinx/zynqmp.c | 29 +++++++++++++++++++++++++++--
>   1 file changed, 27 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/firmware/xilinx/zynqmp.c b/drivers/firmware/xilinx/zynqmp.c
> index 6bc6b6c84241c..c0ff3efd321fc 100644
> --- a/drivers/firmware/xilinx/zynqmp.c
> +++ b/drivers/firmware/xilinx/zynqmp.c
> @@ -738,8 +738,33 @@ EXPORT_SYMBOL_GPL(zynqmp_pm_get_pll_frac_data);
>    */
>   int zynqmp_pm_set_sd_tapdelay(u32 node_id, u32 type, u32 value)
>   {
> -	return zynqmp_pm_invoke_fn(PM_IOCTL, node_id, IOCTL_SET_SD_TAPDELAY,
> -				   type, value, NULL);
> +#define SD_ITAPDLY	0xFF180314
> +#define SD_OTAPDLYSEL	0xFF180318

These two macros should go to include/linux/firmware/xlnx-zynqmp.h


> +	u32 reg = (type == PM_TAPDELAY_INPUT) ? SD_ITAPDLY : SD_OTAPDLYSEL;
> +	u32 mask = (node_id == NODE_SD_0) ? GENMASK(15, 0) : GENMASK(31, 16);
> +
> +	if (value) {
> +		return zynqmp_pm_invoke_fn(PM_IOCTL, node_id,
> +					   IOCTL_SET_SD_TAPDELAY,
> +					   type, value, NULL);
> +	}
> +
> +	/*
> +	 * Work around completely misdesigned firmware API on Xilinx ZynqMP.
> +	 * The IOCTL_SET_SD_TAPDELAY firmware call allows the caller to only
> +	 * ever set IOU_SLCR SD_ITAPDLY Register SD0_ITAPDLYENA/SD1_ITAPDLYENA
> +	 * bits, but there is no matching call to clear those bits. If those
> +	 * bits are not cleared, SDMMC tuning may fail.
> +	 *
> +	 * Luckily, there are PM_MMIO_READ/PM_MMIO_WRITE calls which seem to
> +	 * allow complete unrestricted access to all address space, including
> +	 * IOU_SLCR SD_ITAPDLY Register and all the other registers, access
> +	 * to which was supposed to be protected by the current firmware API.
> +	 *
> +	 * Use PM_MMIO_READ/PM_MMIO_WRITE to re-implement the missing counter
> +	 * part of IOCTL_SET_SD_TAPDELAY which clears SDx_ITAPDLYENA bits.
> +	 */
> +	return zynqmp_pm_invoke_fn(PM_MMIO_WRITE, reg, mask, 0, 0, NULL);
>   }
>   EXPORT_SYMBOL_GPL(zynqmp_pm_set_sd_tapdelay);
>   

I am fine with this workaround. Just for the record the way forward is IMHO to 
deprecate the whole IOCTL_SET_SD_TAPDELAY and create new one 
IOCTL_SET_SD_TAPDELAY2 and use feature check interface to find it if new 
firmware provides it. If provides use it and don't get to PM_MMIO_WRITE 
interface. It is our plan of record that MMIO_READ/WRITE interface should be 
deprecated.

Once macro position is fixed I will queue it.

Thanks,
Michal