firmware: xilinx: Clear IOCTL_SET_SD_TAPDELAY using PM_MMIO_WRITE

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@denx.de>
Link: https://lore.kernel.org/r/20221215152023.8387-1-marex@denx.deSigned-off-by: Michal Simek <michal.simek@amd.com>

drivers/firmware/xilinx/zynqmp.c

@@ -738,8 +738,31 @@ 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);
+	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);
 

include/linux/firmware/xlnx-zynqmp.h

@@ -79,6 +79,10 @@
 #define EVENT_ERROR_PSM_ERR1	(0x28108000U)
 #define EVENT_ERROR_PSM_ERR2	(0x2810C000U)
 
+/* ZynqMP SD tap delay tuning */
+#define SD_ITAPDLY	0xFF180314
+#define SD_OTAPDLYSEL	0xFF180318
+
 enum pm_api_cb_id {
 	PM_INIT_SUSPEND_CB = 30,
 	PM_ACKNOWLEDGE_CB = 31,