net: dsa: felix: re-enable TAS guard band mode

Commit 316bcffe4479 ("net: dsa: felix: disable always guard band bit for
TAS config") disabled the guard band and broke 802.3Qbv compliance.

There are two issues here:
(1) Without the guard band the end of the scheduling window could be
overrun by a frame in transit.
(2) Frames that don't fit into a configured window will still be sent.

The reason for both issues is that the switch will schedule the _start_
of a frame transmission inside the predefined window without taking the
length of the frame into account. Thus, we'll need the guard band which
will close the gate early, so that a complete frame can still be sent.
Revert the commit and add a note.

For a lengthy discussion see [1].

[1] https://lore.kernel.org/netdev/[e-mail]/

Fixes: 316bcffe4479 ("net: dsa: felix: disable always guard band bit for TAS config")
Signed-off-by: Michael Walle <[e-mail]>
Reviewed-by: Vladimir Oltean <[e-mail]>
Signed-off-by: David S. Miller <[e-mail]>

drivers/net/dsa/ocelot/felix_vsc9959.c

···
 	if (taprio->num_entries > VSC9959_TAS_GCL_ENTRY_MAX)
 		return -ERANGE;
 
-	/* Set port num and disable ALWAYS_GUARD_BAND_SCH_Q, which means set
-	 * guard band to be implemented for nonschedule queues to schedule
-	 * queues transition.
+	/* Enable guard band. The switch will schedule frames without taking
+	 * their length into account. Thus we'll always need to enable the
+	 * guard band which reserves the time of a maximum sized frame at the
+	 * end of the time window.
+	 *
+	 * Although the ALWAYS_GUARD_BAND_SCH_Q bit is global for all ports, we
+	 * need to set PORT_NUM, because subsequent writes to PARAM_CFG_REG_n
+	 * operate on the port number.
 	 */
-	ocelot_rmw(ocelot,
-		   QSYS_TAS_PARAM_CFG_CTRL_PORT_NUM(port),
+	ocelot_rmw(ocelot, QSYS_TAS_PARAM_CFG_CTRL_PORT_NUM(port) |
+		   QSYS_TAS_PARAM_CFG_CTRL_ALWAYS_GUARD_BAND_SCH_Q,
 		   QSYS_TAS_PARAM_CFG_CTRL_PORT_NUM_M |
 		   QSYS_TAS_PARAM_CFG_CTRL_ALWAYS_GUARD_BAND_SCH_Q,
 		   QSYS_TAS_PARAM_CFG_CTRL);