net: dsa: felix: enable cut-through forwarding between ports by default

The VSC9959 switch embedded within NXP LS1028A (and that version of
Ocelot switches only) supports cut-through forwarding - meaning it can
start the process of looking up the destination ports for a packet, and
forward towards those ports, before the entire packet has been received
(as opposed to the store-and-forward mode).

The up side is having lower forwarding latency for large packets. The
down side is that frames with FCS errors are forwarded instead of being
dropped. However, erroneous frames do not result in incorrect updates of
the FDB or incorrect policer updates, since these processes are deferred
inside the switch to the end of frame. Since the switch starts the
cut-through forwarding process after all packet headers (including IP,
if any) have been processed, packets with large headers and small
payload do not see the benefit of lower forwarding latency.

There are two cases that need special attention.

The first is when a packet is multicast (or flooded) to multiple
destinations, one of which doesn't have cut-through forwarding enabled.
The switch deals with this automatically by disabling cut-through
forwarding for the frame towards all destination ports.

The second is when a packet is forwarded from a port of lower link speed
towards a port of higher link speed. This is not handled by the hardware
and needs software intervention.

Since we practically need to update the cut-through forwarding domain
from paths that aren't serialized by the rtnl_mutex (phylink
mac_link_down/mac_link_up ops), this means we need to serialize physical
link events with user space updates of bonding/bridging domains.

Enabling cut-through forwarding is done per {egress port, traffic class}.
I don't see any reason why this would be a configurable option as long
as it works without issues, and there doesn't appear to be any user
space configuration tool to toggle this on/off, so this patch enables
cut-through forwarding on all eligible ports and traffic classes.

Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Link: https://lore.kernel.org/r/20211125125808.2383984-2-vladimir.oltean@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

1 files changed, 69 insertions, 8 deletions

diff --git a/drivers/net/ethernet/mscc/ocelot.c b/drivers/net/ethernet/mscc/ocelot.c
index 26feb030d1a6..03c716ef39cf 100644
--- a/drivers/net/ethernet/mscc/ocelot.c
+++ b/drivers/net/ethernet/mscc/ocelot.c
@@ -663,9 +663,17 @@ void ocelot_phylink_mac_link_down(struct ocelot *ocelot, int port,
 	struct ocelot_port *ocelot_port = ocelot->ports[port];
 	int err;
 
+	ocelot_port->speed = SPEED_UNKNOWN;
+
 	ocelot_port_rmwl(ocelot_port, 0, DEV_MAC_ENA_CFG_RX_ENA,
 			 DEV_MAC_ENA_CFG);
 
+	if (ocelot->ops->cut_through_fwd) {
+		mutex_lock(&ocelot->fwd_domain_lock);
+		ocelot->ops->cut_through_fwd(ocelot);
+		mutex_unlock(&ocelot->fwd_domain_lock);
+	}
+
 	ocelot_fields_write(ocelot, port, QSYS_SWITCH_PORT_MODE_PORT_ENA, 0);
 
 	err = ocelot_port_flush(ocelot, port);
@@ -697,6 +705,8 @@ void ocelot_phylink_mac_link_up(struct ocelot *ocelot, int port,
 	int mac_speed, mode = 0;
 	u32 mac_fc_cfg;
 
+	ocelot_port->speed = speed;
+
 	/* The MAC might be integrated in systems where the MAC speed is fixed
 	 * and it's the PCS who is performing the rate adaptation, so we have
 	 * to write "1000Mbps" into the LINK_SPEED field of DEV_CLOCK_CFG
@@ -769,6 +779,15 @@ void ocelot_phylink_mac_link_up(struct ocelot *ocelot, int port,
 	ocelot_port_writel(ocelot_port, DEV_MAC_ENA_CFG_RX_ENA |
 			   DEV_MAC_ENA_CFG_TX_ENA, DEV_MAC_ENA_CFG);
 
+	/* If the port supports cut-through forwarding, update the masks before
+	 * enabling forwarding on the port.
+	 */
+	if (ocelot->ops->cut_through_fwd) {
+		mutex_lock(&ocelot->fwd_domain_lock);
+		ocelot->ops->cut_through_fwd(ocelot);
+		mutex_unlock(&ocelot->fwd_domain_lock);
+	}
+
 	/* Core: Enable port for frame transfer */
 	ocelot_fields_write(ocelot, port,
 			    QSYS_SWITCH_PORT_MODE_PORT_ENA, 1);
@@ -1542,7 +1561,7 @@ static u32 ocelot_get_bond_mask(struct ocelot *ocelot, struct net_device *bond,
 	return mask;
 }
 
-static u32 ocelot_get_bridge_fwd_mask(struct ocelot *ocelot, int src_port)
+u32 ocelot_get_bridge_fwd_mask(struct ocelot *ocelot, int src_port)
 {
 	struct ocelot_port *ocelot_port = ocelot->ports[src_port];
 	const struct net_device *bridge;
@@ -1569,8 +1588,9 @@ static u32 ocelot_get_bridge_fwd_mask(struct ocelot *ocelot, int src_port)
 
 	return mask;
 }
+EXPORT_SYMBOL_GPL(ocelot_get_bridge_fwd_mask);
 
-static u32 ocelot_get_dsa_8021q_cpu_mask(struct ocelot *ocelot)
+u32 ocelot_get_dsa_8021q_cpu_mask(struct ocelot *ocelot)
 {
 	u32 mask = 0;
 	int port;
@@ -1587,12 +1607,22 @@ static u32 ocelot_get_dsa_8021q_cpu_mask(struct ocelot *ocelot)
 
 	return mask;
 }
+EXPORT_SYMBOL_GPL(ocelot_get_dsa_8021q_cpu_mask);
 
-void ocelot_apply_bridge_fwd_mask(struct ocelot *ocelot)
+void ocelot_apply_bridge_fwd_mask(struct ocelot *ocelot, bool joining)
 {
 	unsigned long cpu_fwd_mask;
 	int port;
 
+	lockdep_assert_held(&ocelot->fwd_domain_lock);
+
+	/* If cut-through forwarding is supported, update the masks before a
+	 * port joins the forwarding domain, to avoid potential underruns if it
+	 * has the highest speed from the new domain.
+	 */
+	if (joining && ocelot->ops->cut_through_fwd)
+		ocelot->ops->cut_through_fwd(ocelot);
+
 	/* If a DSA tag_8021q CPU exists, it needs to be included in the
 	 * regular forwarding path of the front ports regardless of whether
 	 * those are bridged or standalone.
@@ -1639,6 +1669,16 @@ void ocelot_apply_bridge_fwd_mask(struct ocelot *ocelot)
 
 		ocelot_write_rix(ocelot, mask, ANA_PGID_PGID, PGID_SRC + port);
 	}
+
+	/* If cut-through forwarding is supported and a port is leaving, there
+	 * is a chance that cut-through was disabled on the other ports due to
+	 * the port which is leaving (it has a higher link speed). We need to
+	 * update the cut-through masks of the remaining ports no earlier than
+	 * after the port has left, to prevent underruns from happening between
+	 * the cut-through update and the forwarding domain update.
+	 */
+	if (!joining && ocelot->ops->cut_through_fwd)
+		ocelot->ops->cut_through_fwd(ocelot);
 }
 EXPORT_SYMBOL(ocelot_apply_bridge_fwd_mask);
 
@@ -1647,6 +1687,8 @@ void ocelot_bridge_stp_state_set(struct ocelot *ocelot, int port, u8 state)
 	struct ocelot_port *ocelot_port = ocelot->ports[port];
 	u32 learn_ena = 0;
 
+	mutex_lock(&ocelot->fwd_domain_lock);
+
 	ocelot_port->stp_state = state;
 
 	if ((state == BR_STATE_LEARNING || state == BR_STATE_FORWARDING) &&
@@ -1656,7 +1698,9 @@ void ocelot_bridge_stp_state_set(struct ocelot *ocelot, int port, u8 state)
 	ocelot_rmw_gix(ocelot, learn_ena, ANA_PORT_PORT_CFG_LEARN_ENA,
 		       ANA_PORT_PORT_CFG, port);
 
-	ocelot_apply_bridge_fwd_mask(ocelot);
+	ocelot_apply_bridge_fwd_mask(ocelot, state == BR_STATE_FORWARDING);
+
+	mutex_unlock(&ocelot->fwd_domain_lock);
 }
 EXPORT_SYMBOL(ocelot_bridge_stp_state_set);
 
@@ -1886,9 +1930,13 @@ void ocelot_port_bridge_join(struct ocelot *ocelot, int port,
 {
 	struct ocelot_port *ocelot_port = ocelot->ports[port];
 
+	mutex_lock(&ocelot->fwd_domain_lock);
+
 	ocelot_port->bridge = bridge;
 
-	ocelot_apply_bridge_fwd_mask(ocelot);
+	ocelot_apply_bridge_fwd_mask(ocelot, true);
+
+	mutex_unlock(&ocelot->fwd_domain_lock);
 }
 EXPORT_SYMBOL(ocelot_port_bridge_join);
 
@@ -1897,11 +1945,15 @@ void ocelot_port_bridge_leave(struct ocelot *ocelot, int port,
 {
 	struct ocelot_port *ocelot_port = ocelot->ports[port];
 
+	mutex_lock(&ocelot->fwd_domain_lock);
+
 	ocelot_port->bridge = NULL;
 
 	ocelot_port_set_pvid(ocelot, port, NULL);
 	ocelot_port_manage_port_tag(ocelot, port);
-	ocelot_apply_bridge_fwd_mask(ocelot);
+	ocelot_apply_bridge_fwd_mask(ocelot, false);
+
+	mutex_unlock(&ocelot->fwd_domain_lock);
 }
 EXPORT_SYMBOL(ocelot_port_bridge_leave);
 
@@ -2023,12 +2075,16 @@ int ocelot_port_lag_join(struct ocelot *ocelot, int port,
 	if (info->tx_type != NETDEV_LAG_TX_TYPE_HASH)
 		return -EOPNOTSUPP;
 
+	mutex_lock(&ocelot->fwd_domain_lock);
+
 	ocelot->ports[port]->bond = bond;
 
 	ocelot_setup_logical_port_ids(ocelot);
-	ocelot_apply_bridge_fwd_mask(ocelot);
+	ocelot_apply_bridge_fwd_mask(ocelot, true);
 	ocelot_set_aggr_pgids(ocelot);
 
+	mutex_unlock(&ocelot->fwd_domain_lock);
+
 	return 0;
 }
 EXPORT_SYMBOL(ocelot_port_lag_join);
@@ -2036,11 +2092,15 @@ EXPORT_SYMBOL(ocelot_port_lag_join);
 void ocelot_port_lag_leave(struct ocelot *ocelot, int port,
 			   struct net_device *bond)
 {
+	mutex_lock(&ocelot->fwd_domain_lock);
+
 	ocelot->ports[port]->bond = NULL;
 
 	ocelot_setup_logical_port_ids(ocelot);
-	ocelot_apply_bridge_fwd_mask(ocelot);
+	ocelot_apply_bridge_fwd_mask(ocelot, false);
 	ocelot_set_aggr_pgids(ocelot);
+
+	mutex_unlock(&ocelot->fwd_domain_lock);
 }
 EXPORT_SYMBOL(ocelot_port_lag_leave);
 
@@ -2331,6 +2391,7 @@ int ocelot_init(struct ocelot *ocelot)
 	mutex_init(&ocelot->stats_lock);
 	mutex_init(&ocelot->ptp_lock);
 	mutex_init(&ocelot->mact_lock);
+	mutex_init(&ocelot->fwd_domain_lock);
 	spin_lock_init(&ocelot->ptp_clock_lock);
 	spin_lock_init(&ocelot->ts_id_lock);
 	snprintf(queue_name, sizeof(queue_name), "%s-stats",