[PATCH net v3] net: phy: meson-gxl: detect LPA corruption

[Jerome Brunet]

The purpose of this change is to fix the incorrect detection of the link
partner (LP) advertised capabilities which sometimes happens with this PHY
(roughly 1 time in a dozen)

This issue may cause the link to be negotiated at 10Mbps/Full or
10Mbps/Half when 100MBps/Full is actually possible. In some case, the link
is even completely broken and no communication is possible.

To detect the corruption, we must look for a magic undocumented bit in the
WOL bank (hint given by the SoC vendor kernel) but this is not enough to
cover all cases. We also have to look at the LPA ack. If the LP supports
Aneg but did not ack our base code when aneg is completed, we assume
something went wrong.

The detection of a corrupted LPA triggers a restart of the aneg process.
This solves the problem but may take up to 6 retries to complete.

Fixes: 7334b3e47aee ("net: phy: Add Meson GXL Internal PHY driver")
Signed-off-by: Jerome Brunet <jbrunet at baylibre.com>
---

I suppose this patch probably seems a bit hacky, especially the part
about the link partner acknowledge. I'm trying to figure out if the
value in MII_LPA makes sense but I don't have such a deep knowledge
of the ethernet spec.

To me, it does not makes sense for the LP to support ANEG (Bit 1 in
MII_EXPENSION), the aneg to have successfully complete and, at the
same time, LP does not ACK our base code word, which we should have
sent during this aneg.

If you think this may have unintended consequences or if you have
an idea to do this differently, feel free to let me know.

This fix has been tested on the libretech-cc and khadas VIM

The v2 [0] had dependencies on other changes which were not fixes.
This v3 is re-implementation of v2 without the dependencies to be
merged as a fix and easily backportable.

All the ugly phy_write() with hexadecimal values will go away when
then clean-up gets through.

[0]: https://lkml.kernel.org/r/20171207142715.32578-6-jbrunet@baylibre.com

 drivers/net/phy/meson-gxl.c | 74 ++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 73 insertions(+), 1 deletion(-)

diff --git a/drivers/net/phy/meson-gxl.c b/drivers/net/phy/meson-gxl.c
index 1ea69b7585d9..700007dd4be5 100644
--- a/drivers/net/phy/meson-gxl.c
+++ b/drivers/net/phy/meson-gxl.c
@@ -22,6 +22,7 @@
 #include <linux/ethtool.h>
 #include <linux/phy.h>
 #include <linux/netdevice.h>
+#include <linux/bitfield.h>
 
 static int meson_gxl_config_init(struct phy_device *phydev)
 {
@@ -50,6 +51,77 @@ static int meson_gxl_config_init(struct phy_device *phydev)
 	return 0;
 }
 
+/* This function is provided to cope with the possible failures of this phy
+ * during aneg process. When aneg fails, the PHY reports that aneg is done
+ * but the value found in MII_LPA is wrong:
+ *  - Early failures: MII_LPA is just 0x0001. if MII_EXPANSION reports that
+ *    the link partner (LP) supports aneg but the LP never acked our base
+ *    code word, it is likely that we never sent it to begin with.
+ *  - Late failures: MII_LPA is filled with a value which seems to make sense
+ *    but it actually is not what the LP is advertising. It seems that we
+ *    can detect this using a magic bit in the WOL bank (reg 12 - bit 12).
+ *    If this particular bit is not set when aneg is reported being done,
+ *    it means MII_LPA is likely to be wrong.
+ *
+ * In both case, forcing a restart of the aneg process solve the problem.
+ * When this failure happens, the first retry is usually successful but,
+ * in some cases, it may take up to 6 retries to get a decent result
+ */
+int meson_gxl_read_status(struct phy_device *phydev)
+{
+	int ret, wol, lpa, exp;
+
+	if (phydev->autoneg == AUTONEG_ENABLE) {
+		ret = genphy_aneg_done(phydev);
+		if (ret < 0)
+			return ret;
+		else if (!ret)
+			goto read_status_continue;
+
+		/* Need to access WOL bank, make sure the access is open */
+		ret = phy_write(phydev, 0x14, 0x0000);
+		if (ret)
+			return ret;
+		ret = phy_write(phydev, 0x14, 0x0400);
+		if (ret)
+			return ret;
+		ret = phy_write(phydev, 0x14, 0x0000);
+		if (ret)
+			return ret;
+		ret = phy_write(phydev, 0x14, 0x0400);
+		if (ret)
+			return ret;
+
+		/* Request LPI_STATUS WOL register */
+		ret = phy_write(phydev, 0x14, 0x8D80);
+		if (ret)
+			return ret;
+
+		/* Read LPI_STATUS value */
+		wol = phy_read(phydev, 0x15);
+		if (wol < 0)
+			return wol;
+
+		lpa = phy_read(phydev, MII_LPA);
+		if (lpa < 0)
+			return lpa;
+
+		exp = phy_read(phydev, MII_EXPANSION);
+		if (exp < 0)
+			return exp;
+
+		if (!(wol & BIT(12)) ||
+		    ((exp & EXPANSION_NWAY) && !(lpa & LPA_LPACK))) {
+			/* Looks like aneg failed after all */
+			phydev_dbg(phydev, "LPA corruption - aneg restart\n");
+			return genphy_restart_aneg(phydev);
+		}
+	}
+
+read_status_continue:
+	return genphy_read_status(phydev);
+}
+
 static struct phy_driver meson_gxl_phy[] = {
 	{
 		.phy_id		= 0x01814400,
@@ -60,7 +132,7 @@ static struct phy_driver meson_gxl_phy[] = {
 		.config_init	= meson_gxl_config_init,
 		.config_aneg	= genphy_config_aneg,
 		.aneg_done      = genphy_aneg_done,
-		.read_status	= genphy_read_status,
+		.read_status	= meson_gxl_read_status,
 		.suspend        = genphy_suspend,
 		.resume         = genphy_resume,
 	},
-- 
2.14.3