On Tue, 12 Jan 2021 21:43:29 +0100
Andrew Lunn [off-list ref] wrote:

quoted

+/* RollBall SFPs do not access internal PHY via I2C address 0x56, but
+ * instead via address 0x51, when SFP page is set to 0x03 and password to
+ * 0xffffffff.
+ * Since current SFP code does not modify SFP_PAGE, we set it to 0x03 only at
+ * bus creation time, and expect it to remain set to 0x03 throughout the
+ * lifetime of the module plugged into the system. If the SFP code starts
+ * modifying SFP_PAGE in the future, this code will need to change.  

...

quoted

+/* In order to not interfere with other SFP code (which possibly may manipulate
+ * SFP_PAGE), for every transfer we do this:
+ *   1. lock the bus
+ *   2. save content of SFP_PAGE
+ *   3. set SFP_PAGE to 3
+ *   4. do the transfer
+ *   5. restore original SFP_PAGE
+ *   6. unlock the bus  

These two comments seem to contradict each other?

I forgot about the first comment. I will rewrite it and send again.

quoted

+static int i2c_rollball_mii_poll(struct mii_bus *bus, int bus_addr, u8 *buf,
+				 size_t len)
+{
+	struct i2c_adapter *i2c = bus->priv;
+	struct i2c_msg msgs[2];
+	u8 cmd_addr, tmp, *res;
+	int i, ret;
+
+	cmd_addr = ROLLBALL_CMD_ADDR;
+
+	res = buf ? buf : &tmp;
+	len = buf ? len : 1;
+
+	msgs[0].addr = bus_addr;
+	msgs[0].flags = 0;
+	msgs[0].len = 1;
+	msgs[0].buf = &cmd_addr;
+
+	msgs[1].addr = bus_addr;
+	msgs[1].flags = I2C_M_RD;
+	msgs[1].len = len;
+	msgs[1].buf = res;
+
+	/* By experiment it takes up to 70 ms to access a register for these
+	 * SFPs. Sleep 20ms between iteratios and try 10 times.
+	 */  

iterations

THX.

quoted

+static int i2c_mii_read_rollball(struct mii_bus *bus, int phy_id, int reg)
+{
+	u8 buf[4], res[6];
+	int bus_addr, ret;
+	u16 val;
+
+	if (!(reg & MII_ADDR_C45))
+		return -EOPNOTSUPP;
+
+	bus_addr = i2c_mii_phy_addr(phy_id);
+	if (bus_addr != ROLLBALL_PHY_I2C_ADDR)
+		return 0xffff;
+
+	buf[0] = ROLLBALL_DATA_ADDR;
+	buf[1] = (reg >> 16) & 0x1f;
+	buf[2] = (reg >> 8) & 0xff;
+	buf[3] = reg & 0xff;  

This looks odd. There are only 32 registers for C22 transactions, so
it fits in one byte. You can set buf[1] and buf[2] to zero.

C22 is not supported by this protocol. A few line above there is
  if (!(reg & MII_ADDR_C45))
    return -EOPNOTSUPP;

C45 transactions allow for 65536 registers, and has the devtype field,
so would need 3 bytes. Which suggests that C45 might actually be
supported? At least by the protocol, if not the device.

quoted

+	dev_dbg(&bus->dev, "read reg %02x:%04x = %04x\n", (reg >> 16) & 0x1f,
+		reg & 0xffff, val);  

There is a tracepoint that allows access to this information, so you
can probably remove this.

    Andrew

OK then, I will remove this.

Thanks, Andrew.