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, 8 insertions, 1 deletions

diff --git a/include/soc/mscc/ocelot.h b/include/soc/mscc/ocelot.h
index 89d17629efe5..33f2e8c9e88b 100644
--- a/include/soc/mscc/ocelot.h
+++ b/include/soc/mscc/ocelot.h
@@ -561,6 +561,7 @@ struct ocelot_ops {
 	int (*psfp_filter_del)(struct ocelot *ocelot, struct flow_cls_offload *f);
 	int (*psfp_stats_get)(struct ocelot *ocelot, struct flow_cls_offload *f,
 			      struct flow_stats *stats);
+	void (*cut_through_fwd)(struct ocelot *ocelot);
 };
 
 struct ocelot_vcap_policer {
@@ -655,6 +656,8 @@ struct ocelot_port {
 
 	struct net_device		*bridge;
 	u8				stp_state;
+
+	int				speed;
 };
 
 struct ocelot {
@@ -712,6 +715,8 @@ struct ocelot {
 
 	/* Lock for serializing access to the MAC table */
 	struct mutex			mact_lock;
+	/* Lock for serializing forwarding domain changes */
+	struct mutex			fwd_domain_lock;
 
 	struct workqueue_struct		*owq;
 
@@ -811,7 +816,9 @@ void ocelot_set_ageing_time(struct ocelot *ocelot, unsigned int msecs);
 int ocelot_port_vlan_filtering(struct ocelot *ocelot, int port, bool enabled,
 			       struct netlink_ext_ack *extack);
 void ocelot_bridge_stp_state_set(struct ocelot *ocelot, int port, u8 state);
-void ocelot_apply_bridge_fwd_mask(struct ocelot *ocelot);
+u32 ocelot_get_dsa_8021q_cpu_mask(struct ocelot *ocelot);
+u32 ocelot_get_bridge_fwd_mask(struct ocelot *ocelot, int src_port);
+void ocelot_apply_bridge_fwd_mask(struct ocelot *ocelot, bool joining);
 int ocelot_port_pre_bridge_flags(struct ocelot *ocelot, int port,
 				 struct switchdev_brport_flags val);
 void ocelot_port_bridge_flags(struct ocelot *ocelot, int port,