[dpdk-dev] [PATCH 22/38] net/sfc: implement representor Tx routine

[Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>]

From: Igor Romanov <igor.romanov@oktetlabs.ru>

Forward traffic that is transmitted from a port representor to the
corresponding virtual function using the dedicated TxQ.

Signed-off-by: Igor Romanov <igor.romanov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
Reviewed-by: Andy Moreton <amoreton@xilinx.com>
Reviewed-by: Ivan Malov <ivan.malov@oktetlabs.ru>
---
 drivers/net/sfc/sfc_repr.c       | 45 ++++++++++++++++
 drivers/net/sfc/sfc_repr_proxy.c | 88 +++++++++++++++++++++++++++++++-
 drivers/net/sfc/sfc_repr_proxy.h |  8 +++
 3 files changed, 140 insertions(+), 1 deletion(-)

diff --git a/drivers/net/sfc/sfc_repr.c b/drivers/net/sfc/sfc_repr.c
index 7a34a0a904..e7386fb480 100644
--- a/drivers/net/sfc/sfc_repr.c
+++ b/drivers/net/sfc/sfc_repr.c
@@ -168,6 +168,49 @@ sfc_repr_tx_queue_stop(void *queue)
 	rte_ring_reset(txq->ring);
 }
 
+static uint16_t
+sfc_repr_tx_burst(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+{
+	struct sfc_repr_txq *txq = tx_queue;
+	unsigned int n_tx;
+	void **objs;
+	uint16_t i;
+
+	/*
+	 * mbuf is likely cache-hot. Set flag and egress m-port here instead of
+	 * doing that in representors proxy. Also, it should help to avoid
+	 * cache bounce. Moreover, potentially, it allows to use one
+	 * multi-producer single-consumer ring for all representors.
+	 *
+	 * The only potential problem is doing so many times if enqueue
+	 * fails and sender retries.
+	 */
+	for (i = 0; i < nb_pkts; ++i) {
+		struct rte_mbuf *m = tx_pkts[i];
+
+		m->ol_flags |= sfc_dp_mport_override;
+		*RTE_MBUF_DYNFIELD(m, sfc_dp_mport_offset,
+				   efx_mport_id_t *) = txq->egress_mport;
+	}
+
+	objs = (void *)&tx_pkts[0];
+	n_tx = rte_ring_sp_enqueue_burst(txq->ring, objs, nb_pkts, NULL);
+
+	/*
+	 * Remove m-port override flag from packets that were not enqueued
+	 * Setting the flag only for enqueued packets after the burst is
+	 * not possible since the ownership of enqueued packets is
+	 * transferred to representor proxy.
+	 */
+	for (i = n_tx; i < nb_pkts; ++i) {
+		struct rte_mbuf *m = tx_pkts[i];
+
+		m->ol_flags &= ~sfc_dp_mport_override;
+	}
+
+	return n_tx;
+}
+
 static int
 sfc_repr_start(struct rte_eth_dev *dev)
 {
@@ -782,6 +825,7 @@ sfc_repr_dev_close(struct rte_eth_dev *dev)
 
 	(void)sfc_repr_proxy_del_port(srs->pf_port_id, srs->repr_id);
 
+	dev->tx_pkt_burst = NULL;
 	dev->dev_ops = NULL;
 
 	sfc_repr_unlock(sr);
@@ -902,6 +946,7 @@ sfc_repr_eth_dev_init(struct rte_eth_dev *dev, void *init_params)
 		goto fail_mac_addrs;
 	}
 
+	dev->tx_pkt_burst = sfc_repr_tx_burst;
 	dev->dev_ops = &sfc_repr_dev_ops;
 
 	sr->state = SFC_ETHDEV_INITIALIZED;
diff --git a/drivers/net/sfc/sfc_repr_proxy.c b/drivers/net/sfc/sfc_repr_proxy.c
index ea03d5afdd..d8934bab65 100644
--- a/drivers/net/sfc/sfc_repr_proxy.c
+++ b/drivers/net/sfc/sfc_repr_proxy.c
@@ -25,6 +25,12 @@
  */
 #define SFC_REPR_PROXY_MBOX_POLL_TIMEOUT_MS	1000
 
+/**
+ * Amount of time to wait for the representor proxy routine (which is
+ * running on a service core) to terminate after service core is stopped.
+ */
+#define SFC_REPR_PROXY_ROUTINE_TERMINATE_TIMEOUT_MS	10000
+
 static struct sfc_repr_proxy *
 sfc_repr_proxy_by_adapter(struct sfc_adapter *sa)
 {
@@ -148,16 +154,71 @@ sfc_repr_proxy_mbox_handle(struct sfc_repr_proxy *rp)
 	__atomic_store_n(&mbox->ack, true, __ATOMIC_RELEASE);
 }
 
+static void
+sfc_repr_proxy_handle_tx(struct sfc_repr_proxy_dp_txq *rp_txq,
+			 struct sfc_repr_proxy_txq *repr_txq)
+{
+	/*
+	 * With multiple representor proxy queues configured it is
+	 * possible that not all of the corresponding representor
+	 * queues were created. Skip the queues that do not exist.
+	 */
+	if (repr_txq->ring == NULL)
+		return;
+
+	if (rp_txq->available < RTE_DIM(rp_txq->tx_pkts)) {
+		rp_txq->available +=
+			rte_ring_sc_dequeue_burst(repr_txq->ring,
+				(void **)(&rp_txq->tx_pkts[rp_txq->available]),
+				RTE_DIM(rp_txq->tx_pkts) - rp_txq->available,
+				NULL);
+
+		if (rp_txq->available == rp_txq->transmitted)
+			return;
+	}
+
+	rp_txq->transmitted += rp_txq->pkt_burst(rp_txq->dp,
+				&rp_txq->tx_pkts[rp_txq->transmitted],
+				rp_txq->available - rp_txq->transmitted);
+
+	if (rp_txq->available == rp_txq->transmitted) {
+		rp_txq->available = 0;
+		rp_txq->transmitted = 0;
+	}
+}
+
 static int32_t
 sfc_repr_proxy_routine(void *arg)
 {
+	struct sfc_repr_proxy_port *port;
 	struct sfc_repr_proxy *rp = arg;
+	unsigned int i;
 
 	sfc_repr_proxy_mbox_handle(rp);
 
+	TAILQ_FOREACH(port, &rp->ports, entries) {
+		if (!port->started)
+			continue;
+
+		for (i = 0; i < rp->nb_txq; i++)
+			sfc_repr_proxy_handle_tx(&rp->dp_txq[i], &port->txq[i]);
+	}
+
 	return 0;
 }
 
+static struct sfc_txq_info *
+sfc_repr_proxy_txq_info_get(struct sfc_adapter *sa, unsigned int repr_queue_id)
+{
+	struct sfc_adapter_shared *sas = sfc_sa2shared(sa);
+	struct sfc_repr_proxy_dp_txq *dp_txq;
+
+	SFC_ASSERT(repr_queue_id < sfc_repr_nb_txq(sas));
+	dp_txq = &sa->repr_proxy.dp_txq[repr_queue_id];
+
+	return &sas->txq_info[dp_txq->sw_index];
+}
+
 static int
 sfc_repr_proxy_txq_attach(struct sfc_adapter *sa)
 {
@@ -289,11 +350,20 @@ sfc_repr_proxy_txq_fini(struct sfc_adapter *sa)
 static int
 sfc_repr_proxy_txq_start(struct sfc_adapter *sa)
 {
+	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
 	struct sfc_repr_proxy *rp = &sa->repr_proxy;
+	unsigned int i;
 
 	sfc_log_init(sa, "entry");
 
-	RTE_SET_USED(rp);
+	for (i = 0; i < sfc_repr_nb_txq(sas); i++) {
+		struct sfc_repr_proxy_dp_txq *txq = &rp->dp_txq[i];
+
+		txq->dp = sfc_repr_proxy_txq_info_get(sa, i)->dp;
+		txq->pkt_burst = sa->eth_dev->tx_pkt_burst;
+		txq->available = 0;
+		txq->transmitted = 0;
+	}
 
 	sfc_log_init(sa, "done");
 
@@ -922,6 +992,8 @@ sfc_repr_proxy_start(struct sfc_adapter *sa)
 	if (rc != 0)
 		goto fail_txq_start;
 
+	rp->nb_txq = sfc_repr_nb_txq(sas);
+
 	/* Service core may be in "stopped" state, start it */
 	rc = rte_service_lcore_start(rp->service_core_id);
 	if (rc != 0 && rc != -EALREADY) {
@@ -1007,6 +1079,9 @@ sfc_repr_proxy_stop(struct sfc_adapter *sa)
 	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
 	struct sfc_repr_proxy *rp = &sa->repr_proxy;
 	struct sfc_repr_proxy_port *port;
+	const unsigned int wait_ms_total =
+		SFC_REPR_PROXY_ROUTINE_TERMINATE_TIMEOUT_MS;
+	unsigned int i;
 	int rc;
 
 	sfc_log_init(sa, "entry");
@@ -1050,6 +1125,17 @@ sfc_repr_proxy_stop(struct sfc_adapter *sa)
 
 	/* Service lcore may be shared and we never stop it */
 
+	/*
+	 * Wait for the representor proxy routine to finish the last iteration.
+	 * Give up on timeout.
+	 */
+	for (i = 0; i < wait_ms_total; i++) {
+		if (rte_service_may_be_active(rp->service_id) == 0)
+			break;
+
+		rte_delay_ms(1);
+	}
+
 	sfc_repr_proxy_rxq_stop(sa);
 	sfc_repr_proxy_txq_stop(sa);
 
diff --git a/drivers/net/sfc/sfc_repr_proxy.h b/drivers/net/sfc/sfc_repr_proxy.h
index c350713a55..d47e0a431a 100644
--- a/drivers/net/sfc/sfc_repr_proxy.h
+++ b/drivers/net/sfc/sfc_repr_proxy.h
@@ -79,6 +79,13 @@ struct sfc_repr_proxy_dp_rxq {
 };
 
 struct sfc_repr_proxy_dp_txq {
+	eth_tx_burst_t			pkt_burst;
+	struct sfc_dp_txq		*dp;
+
+	unsigned int			available;
+	unsigned int			transmitted;
+	struct rte_mbuf			*tx_pkts[SFC_REPR_PROXY_TX_BURST];
+
 	sfc_sw_index_t			sw_index;
 };
 
@@ -110,6 +117,7 @@ struct sfc_repr_proxy {
 	struct sfc_repr_proxy_filter	mport_filter;
 
 	struct sfc_repr_proxy_mbox	mbox;
+	unsigned int			nb_txq;
 };
 
 struct sfc_adapter;
-- 
2.30.2