[dpdk-dev] [PATCH v2 10/11] net/sfc: add xstats for Rx/Tx doorbells

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

From: Ivan Ilchenko <ivan.ilchenko@oktetlabs.ru>

Rx/Tx doorbells statistics are collected in software and
available per queue. These stats are useful for performance
investigation.

Signed-off-by: Ivan Ilchenko <ivan.ilchenko@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
Reviewed-by: Andy Moreton <amoreton@xilinx.com>
---
 drivers/net/sfc/meson.build        |   1 +
 drivers/net/sfc/sfc.c              |  16 +
 drivers/net/sfc/sfc.h              |   9 +
 drivers/net/sfc/sfc_dp.h           |  10 +
 drivers/net/sfc/sfc_ef10.h         |   3 +-
 drivers/net/sfc/sfc_ef100_rx.c     |   1 +
 drivers/net/sfc/sfc_ef100_tx.c     |   1 +
 drivers/net/sfc/sfc_ef10_essb_rx.c |   3 +-
 drivers/net/sfc/sfc_ef10_rx.c      |   3 +-
 drivers/net/sfc/sfc_ef10_tx.c      |   1 +
 drivers/net/sfc/sfc_ethdev.c       | 124 +++++--
 drivers/net/sfc/sfc_port.c         |  10 +-
 drivers/net/sfc/sfc_rx.c           |   1 +
 drivers/net/sfc/sfc_sw_stats.c     | 572 +++++++++++++++++++++++++++++
 drivers/net/sfc/sfc_sw_stats.h     |  49 +++
 drivers/net/sfc/sfc_tx.c           |   4 +-
 16 files changed, 772 insertions(+), 36 deletions(-)
 create mode 100644 drivers/net/sfc/sfc_sw_stats.c
 create mode 100644 drivers/net/sfc/sfc_sw_stats.h

diff --git a/drivers/net/sfc/meson.build b/drivers/net/sfc/meson.build
index 4625859077..a912cdccfa 100644
--- a/drivers/net/sfc/meson.build
+++ b/drivers/net/sfc/meson.build
@@ -70,6 +70,7 @@ sources = files(
         'sfc.c',
         'sfc_mcdi.c',
         'sfc_sriov.c',
+        'sfc_sw_stats.c',
         'sfc_intr.c',
         'sfc_ev.c',
         'sfc_port.c',
diff --git a/drivers/net/sfc/sfc.c b/drivers/net/sfc/sfc.c
index 4097cf39de..274a98e228 100644
--- a/drivers/net/sfc/sfc.c
+++ b/drivers/net/sfc/sfc.c
@@ -24,6 +24,7 @@
 #include "sfc_tx.h"
 #include "sfc_kvargs.h"
 #include "sfc_tweak.h"
+#include "sfc_sw_stats.h"
 
 
 int
@@ -636,10 +637,17 @@ sfc_configure(struct sfc_adapter *sa)
 	if (rc != 0)
 		goto fail_tx_configure;
 
+	rc = sfc_sw_xstats_configure(sa);
+	if (rc != 0)
+		goto fail_sw_xstats_configure;
+
 	sa->state = SFC_ADAPTER_CONFIGURED;
 	sfc_log_init(sa, "done");
 	return 0;
 
+fail_sw_xstats_configure:
+	sfc_tx_close(sa);
+
 fail_tx_configure:
 	sfc_rx_close(sa);
 
@@ -666,6 +674,7 @@ sfc_close(struct sfc_adapter *sa)
 	SFC_ASSERT(sa->state == SFC_ADAPTER_CONFIGURED);
 	sa->state = SFC_ADAPTER_CLOSING;
 
+	sfc_sw_xstats_close(sa);
 	sfc_tx_close(sa);
 	sfc_rx_close(sa);
 	sfc_port_close(sa);
@@ -891,6 +900,10 @@ sfc_attach(struct sfc_adapter *sa)
 
 	sfc_flow_init(sa);
 
+	rc = sfc_sw_xstats_init(sa);
+	if (rc != 0)
+		goto fail_sw_xstats_init;
+
 	/*
 	 * Create vSwitch to be able to use VFs when PF is not started yet
 	 * as DPDK port. VFs should be able to talk to each other even
@@ -906,6 +919,9 @@ sfc_attach(struct sfc_adapter *sa)
 	return 0;
 
 fail_sriov_vswitch_create:
+	sfc_sw_xstats_close(sa);
+
+fail_sw_xstats_init:
 	sfc_flow_fini(sa);
 	sfc_mae_detach(sa);
 
diff --git a/drivers/net/sfc/sfc.h b/drivers/net/sfc/sfc.h
index 58b8c2c2ad..331e06bac6 100644
--- a/drivers/net/sfc/sfc.h
+++ b/drivers/net/sfc/sfc.h
@@ -217,6 +217,14 @@ struct sfc_counter_rxq {
 	struct rte_mempool		*mp;
 };
 
+struct sfc_sw_xstats {
+	uint64_t			*reset_vals;
+
+	rte_spinlock_t			queues_bitmap_lock;
+	void				*queues_bitmap_mem;
+	struct rte_bitmap		*queues_bitmap;
+};
+
 /* Adapter private data */
 struct sfc_adapter {
 	/*
@@ -249,6 +257,7 @@ struct sfc_adapter {
 	struct sfc_sriov		sriov;
 	struct sfc_intr			intr;
 	struct sfc_port			port;
+	struct sfc_sw_xstats		sw_xstats;
 	struct sfc_filter		filter;
 	struct sfc_mae			mae;
 
diff --git a/drivers/net/sfc/sfc_dp.h b/drivers/net/sfc/sfc_dp.h
index 61c1a3fbac..7fd8f34b0f 100644
--- a/drivers/net/sfc/sfc_dp.h
+++ b/drivers/net/sfc/sfc_dp.h
@@ -42,6 +42,16 @@ enum sfc_dp_type {
 
 /** Datapath queue run-time information */
 struct sfc_dp_queue {
+	/*
+	 * Typically the structure is located at the end of Rx/Tx queue
+	 * data structure and not used on datapath. So, it is not a
+	 * problem to have extra fields even if not used. However,
+	 * put stats at top of the structure to be closer to fields
+	 * used on datapath or reap to have more chances to be cache-hot.
+	 */
+	uint32_t			rx_dbells;
+	uint32_t			tx_dbells;
+
 	uint16_t			port_id;
 	uint16_t			queue_id;
 	struct rte_pci_addr		pci_addr;
diff --git a/drivers/net/sfc/sfc_ef10.h b/drivers/net/sfc/sfc_ef10.h
index ad4c1fdbef..e9bb72e28b 100644
--- a/drivers/net/sfc/sfc_ef10.h
+++ b/drivers/net/sfc/sfc_ef10.h
@@ -99,7 +99,7 @@ sfc_ef10_ev_present(const efx_qword_t ev)
 
 static inline void
 sfc_ef10_rx_qpush(volatile void *doorbell, unsigned int added,
-		  unsigned int ptr_mask)
+		  unsigned int ptr_mask, uint32_t *dbell_counter)
 {
 	efx_dword_t dword;
 
@@ -118,6 +118,7 @@ sfc_ef10_rx_qpush(volatile void *doorbell, unsigned int added,
 	 * operations that follow it (i.e. doorbell write).
 	 */
 	rte_write32(dword.ed_u32[0], doorbell);
+	(*dbell_counter)++;
 }
 
 static inline void
diff --git a/drivers/net/sfc/sfc_ef100_rx.c b/drivers/net/sfc/sfc_ef100_rx.c
index 10c74aa118..d4cb96881c 100644
--- a/drivers/net/sfc/sfc_ef100_rx.c
+++ b/drivers/net/sfc/sfc_ef100_rx.c
@@ -119,6 +119,7 @@ sfc_ef100_rx_qpush(struct sfc_ef100_rxq *rxq, unsigned int added)
 	 * operations that follow it (i.e. doorbell write).
 	 */
 	rte_write32(dword.ed_u32[0], rxq->doorbell);
+	rxq->dp.dpq.rx_dbells++;
 
 	sfc_ef100_rx_debug(rxq, "RxQ pushed doorbell at pidx %u (added=%u)",
 			   EFX_DWORD_FIELD(dword, ERF_GZ_RX_RING_PIDX),
diff --git a/drivers/net/sfc/sfc_ef100_tx.c b/drivers/net/sfc/sfc_ef100_tx.c
index f9ad6f7b73..522e9a0d34 100644
--- a/drivers/net/sfc/sfc_ef100_tx.c
+++ b/drivers/net/sfc/sfc_ef100_tx.c
@@ -489,6 +489,7 @@ sfc_ef100_tx_qpush(struct sfc_ef100_txq *txq, unsigned int added)
 	 * operations that follow it (i.e. doorbell write).
 	 */
 	rte_write32(dword.ed_u32[0], txq->doorbell);
+	txq->dp.dpq.tx_dbells++;
 
 	sfc_ef100_tx_debug(txq, "TxQ pushed doorbell at pidx %u (added=%u)",
 			   EFX_DWORD_FIELD(dword, ERF_GZ_TX_RING_PIDX),
diff --git a/drivers/net/sfc/sfc_ef10_essb_rx.c b/drivers/net/sfc/sfc_ef10_essb_rx.c
index 3c246eb149..991329e86f 100644
--- a/drivers/net/sfc/sfc_ef10_essb_rx.c
+++ b/drivers/net/sfc/sfc_ef10_essb_rx.c
@@ -220,7 +220,8 @@ sfc_ef10_essb_rx_qrefill(struct sfc_ef10_essb_rxq *rxq)
 
 	SFC_ASSERT(rxq->added != added);
 	rxq->added = added;
-	sfc_ef10_rx_qpush(rxq->doorbell, added, rxq_ptr_mask);
+	sfc_ef10_rx_qpush(rxq->doorbell, added, rxq_ptr_mask,
+			  &rxq->dp.dpq.rx_dbells);
 }
 
 static bool
diff --git a/drivers/net/sfc/sfc_ef10_rx.c b/drivers/net/sfc/sfc_ef10_rx.c
index 2b4393d232..49a7d4fb42 100644
--- a/drivers/net/sfc/sfc_ef10_rx.c
+++ b/drivers/net/sfc/sfc_ef10_rx.c
@@ -171,7 +171,8 @@ sfc_ef10_rx_qrefill(struct sfc_ef10_rxq *rxq)
 
 	SFC_ASSERT(rxq->added != added);
 	rxq->added = added;
-	sfc_ef10_rx_qpush(rxq->doorbell, added, ptr_mask);
+	sfc_ef10_rx_qpush(rxq->doorbell, added, ptr_mask,
+			  &rxq->dp.dpq.rx_dbells);
 }
 
 static void
diff --git a/drivers/net/sfc/sfc_ef10_tx.c b/drivers/net/sfc/sfc_ef10_tx.c
index a8d34ead33..ed43adb4ca 100644
--- a/drivers/net/sfc/sfc_ef10_tx.c
+++ b/drivers/net/sfc/sfc_ef10_tx.c
@@ -248,6 +248,7 @@ sfc_ef10_tx_qpush(struct sfc_ef10_txq *txq, unsigned int added,
 	rte_io_wmb();
 
 	*(volatile efsys_uint128_t *)txq->doorbell = oword.eo_u128[0];
+	txq->dp.dpq.tx_dbells++;
 }
 
 static unsigned int
diff --git a/drivers/net/sfc/sfc_ethdev.c b/drivers/net/sfc/sfc_ethdev.c
index dd7e5c253a..2db0d000c3 100644
--- a/drivers/net/sfc/sfc_ethdev.c
+++ b/drivers/net/sfc/sfc_ethdev.c
@@ -28,6 +28,10 @@
 #include "sfc_flow.h"
 #include "sfc_dp.h"
 #include "sfc_dp_rx.h"
+#include "sfc_sw_stats.h"
+
+#define SFC_XSTAT_ID_INVALID_VAL  UINT64_MAX
+#define SFC_XSTAT_ID_INVALID_NAME '\0'
 
 uint32_t sfc_logtype_driver;
 
@@ -714,29 +718,49 @@ sfc_stats_reset(struct rte_eth_dev *dev)
 	if (rc != 0)
 		sfc_err(sa, "failed to reset statistics (rc = %d)", rc);
 
+	sfc_sw_xstats_reset(sa);
+
 	sfc_adapter_unlock(sa);
 
 	SFC_ASSERT(rc >= 0);
 	return -rc;
 }
 
+static unsigned int
+sfc_xstats_get_nb_supported(struct sfc_adapter *sa)
+{
+	struct sfc_port *port = &sa->port;
+	unsigned int nb_supported;
+
+	sfc_adapter_lock(sa);
+	nb_supported = port->mac_stats_nb_supported +
+		       sfc_sw_xstats_get_nb_supported(sa);
+	sfc_adapter_unlock(sa);
+
+	return nb_supported;
+}
+
 static int
 sfc_xstats_get(struct rte_eth_dev *dev, struct rte_eth_xstat *xstats,
 	       unsigned int xstats_count)
 {
 	struct sfc_adapter *sa = sfc_adapter_by_eth_dev(dev);
-	struct sfc_port *port = &sa->port;
 	unsigned int nb_written = 0;
-	unsigned int nb_supp;
+	unsigned int nb_supported = 0;
+	int rc;
 
-	if (unlikely(xstats == NULL)) {
-		sfc_adapter_lock(sa);
-		nb_supp = port->mac_stats_nb_supported;
-		sfc_adapter_unlock(sa);
-		return nb_supp;
-	}
+	if (unlikely(xstats == NULL))
+		return sfc_xstats_get_nb_supported(sa);
+
+	rc = sfc_port_get_mac_stats(sa, xstats, xstats_count, &nb_written);
+	if (rc < 0)
+		return rc;
 
-	return sfc_port_get_mac_stats(sa, xstats, xstats_count, &nb_written);
+	nb_supported = rc;
+	sfc_sw_xstats_get_vals(sa, xstats, xstats_count, &nb_written,
+			       &nb_supported);
+
+	return nb_supported;
 }
 
 static int
@@ -748,24 +772,31 @@ sfc_xstats_get_names(struct rte_eth_dev *dev,
 	struct sfc_port *port = &sa->port;
 	unsigned int i;
 	unsigned int nstats = 0;
+	unsigned int nb_written = 0;
+	int ret;
 
-	if (unlikely(xstats_names == NULL)) {
-		sfc_adapter_lock(sa);
-		nstats = port->mac_stats_nb_supported;
-		sfc_adapter_unlock(sa);
-		return nstats;
-	}
+	if (unlikely(xstats_names == NULL))
+		return sfc_xstats_get_nb_supported(sa);
 
 	for (i = 0; i < EFX_MAC_NSTATS; ++i) {
 		if (EFX_MAC_STAT_SUPPORTED(port->mac_stats_mask, i)) {
-			if (nstats < xstats_count)
+			if (nstats < xstats_count) {
 				strlcpy(xstats_names[nstats].name,
 					efx_mac_stat_name(sa->nic, i),
 					sizeof(xstats_names[0].name));
+				nb_written++;
+			}
 			nstats++;
 		}
 	}
 
+	ret = sfc_sw_xstats_get_names(sa, xstats_names, xstats_count,
+				      &nb_written, &nstats);
+	if (ret != 0) {
+		SFC_ASSERT(ret < 0);
+		return ret;
+	}
+
 	return nstats;
 }
 
@@ -774,11 +805,35 @@ sfc_xstats_get_by_id(struct rte_eth_dev *dev, const uint64_t *ids,
 		     uint64_t *values, unsigned int n)
 {
 	struct sfc_adapter *sa = sfc_adapter_by_eth_dev(dev);
+	struct sfc_port *port = &sa->port;
+	unsigned int nb_supported;
+	unsigned int i;
+	int rc;
 
 	if (unlikely(ids == NULL || values == NULL))
 		return -EINVAL;
 
-	return sfc_port_get_mac_stats_by_id(sa, ids, values, n);
+	/*
+	 * Values array could be filled in nonsequential order. Fill values with
+	 * constant indicating invalid ID first.
+	 */
+	for (i = 0; i < n; i++)
+		values[i] = SFC_XSTAT_ID_INVALID_VAL;
+
+	rc = sfc_port_get_mac_stats_by_id(sa, ids, values, n);
+	if (rc != 0)
+		return rc;
+
+	nb_supported = port->mac_stats_nb_supported;
+	sfc_sw_xstats_get_vals_by_id(sa, ids, values, n, &nb_supported);
+
+	/* Return number of written stats before invalid ID is encountered. */
+	for (i = 0; i < n; i++) {
+		if (values[i] == SFC_XSTAT_ID_INVALID_VAL)
+			return i;
+	}
+
+	return n;
 }
 
 static int
@@ -790,18 +845,23 @@ sfc_xstats_get_names_by_id(struct rte_eth_dev *dev,
 	struct sfc_port *port = &sa->port;
 	unsigned int nb_supported;
 	unsigned int i;
+	int ret;
 
 	if (unlikely(xstats_names == NULL && ids != NULL) ||
 	    unlikely(xstats_names != NULL && ids == NULL))
 		return -EINVAL;
 
-	sfc_adapter_lock(sa);
+	if (unlikely(xstats_names == NULL && ids == NULL))
+		return sfc_xstats_get_nb_supported(sa);
 
-	if (unlikely(xstats_names == NULL && ids == NULL)) {
-		nb_supported = port->mac_stats_nb_supported;
-		sfc_adapter_unlock(sa);
-		return nb_supported;
-	}
+	/*
+	 * Names array could be filled in nonsequential order. Fill names with
+	 * string indicating invalid ID first.
+	 */
+	for (i = 0; i < size; i++)
+		xstats_names[i].name[0] = SFC_XSTAT_ID_INVALID_NAME;
+
+	sfc_adapter_lock(sa);
 
 	SFC_ASSERT(port->mac_stats_nb_supported <=
 		   RTE_DIM(port->mac_stats_by_id));
@@ -812,14 +872,26 @@ sfc_xstats_get_names_by_id(struct rte_eth_dev *dev,
 				efx_mac_stat_name(sa->nic,
 						 port->mac_stats_by_id[ids[i]]),
 				sizeof(xstats_names[0].name));
-		} else {
-			sfc_adapter_unlock(sa);
-			return i;
 		}
 	}
 
+	nb_supported = port->mac_stats_nb_supported;
+
 	sfc_adapter_unlock(sa);
 
+	ret = sfc_sw_xstats_get_names_by_id(sa, ids, xstats_names, size,
+					    &nb_supported);
+	if (ret != 0) {
+		SFC_ASSERT(ret < 0);
+		return ret;
+	}
+
+	/* Return number of written names before invalid ID is encountered. */
+	for (i = 0; i < size; i++) {
+		if (xstats_names[i].name[0] == SFC_XSTAT_ID_INVALID_NAME)
+			return i;
+	}
+
 	return size;
 }
 
diff --git a/drivers/net/sfc/sfc_port.c b/drivers/net/sfc/sfc_port.c
index f6689a17c0..adb2b2cb81 100644
--- a/drivers/net/sfc/sfc_port.c
+++ b/drivers/net/sfc/sfc_port.c
@@ -7,6 +7,8 @@
  * for Solarflare) and Solarflare Communications, Inc.
  */
 
+#include <rte_bitmap.h>
+
 #include "efx.h"
 
 #include "sfc.h"
@@ -701,15 +703,11 @@ sfc_port_get_mac_stats_by_id(struct sfc_adapter *sa, const uint64_t *ids,
 		   RTE_DIM(port->mac_stats_by_id));
 
 	for (i = 0; i < n; i++) {
-		if (ids[i] < port->mac_stats_nb_supported) {
+		if (ids[i] < port->mac_stats_nb_supported)
 			values[i] = mac_stats[port->mac_stats_by_id[ids[i]]];
-		} else {
-			ret = i;
-			goto unlock;
-		}
 	}
 
-	ret = n;
+	ret = 0;
 
 unlock:
 	sfc_adapter_unlock(sa);
diff --git a/drivers/net/sfc/sfc_rx.c b/drivers/net/sfc/sfc_rx.c
index f6a8ac68e8..280e8a61f9 100644
--- a/drivers/net/sfc/sfc_rx.c
+++ b/drivers/net/sfc/sfc_rx.c
@@ -138,6 +138,7 @@ sfc_efx_rx_qrefill(struct sfc_efx_rxq *rxq)
 	SFC_ASSERT(added != rxq->added);
 	rxq->added = added;
 	efx_rx_qpush(rxq->common, added, &rxq->pushed);
+	rxq->dp.dpq.rx_dbells++;
 }
 
 static uint64_t
diff --git a/drivers/net/sfc/sfc_sw_stats.c b/drivers/net/sfc/sfc_sw_stats.c
new file mode 100644
index 0000000000..8489b603f5
--- /dev/null
+++ b/drivers/net/sfc/sfc_sw_stats.c
@@ -0,0 +1,572 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ *
+ * Copyright(c) 2021 Xilinx, Inc.
+ */
+#include <rte_dev.h>
+#include <rte_bitmap.h>
+
+#include "sfc.h"
+#include "sfc_rx.h"
+#include "sfc_tx.h"
+#include "sfc_sw_stats.h"
+
+enum sfc_sw_stats_type {
+	SFC_SW_STATS_RX,
+	SFC_SW_STATS_TX,
+};
+
+typedef uint64_t sfc_get_sw_xstat_val_t(struct sfc_adapter *sa, uint16_t qid);
+
+struct sfc_sw_xstat_descr {
+	const char *name;
+	enum sfc_sw_stats_type type;
+	sfc_get_sw_xstat_val_t *get_val;
+};
+
+static sfc_get_sw_xstat_val_t sfc_get_sw_xstat_val_rx_dbells;
+static uint64_t
+sfc_get_sw_xstat_val_rx_dbells(struct sfc_adapter *sa, uint16_t qid)
+{
+	struct sfc_adapter_shared *sas = sfc_sa2shared(sa);
+	struct sfc_rxq_info *rxq_info;
+
+	rxq_info = sfc_rxq_info_by_ethdev_qid(sas, qid);
+	if (rxq_info->state & SFC_RXQ_INITIALIZED)
+		return rxq_info->dp->dpq.rx_dbells;
+	return 0;
+}
+
+static sfc_get_sw_xstat_val_t sfc_get_sw_xstat_val_tx_dbells;
+static uint64_t
+sfc_get_sw_xstat_val_tx_dbells(struct sfc_adapter *sa, uint16_t qid)
+{
+	struct sfc_adapter_shared *sas = sfc_sa2shared(sa);
+	struct sfc_txq_info *txq_info;
+
+	txq_info = sfc_txq_info_by_ethdev_qid(sas, qid);
+	if (txq_info->state & SFC_TXQ_INITIALIZED)
+		return txq_info->dp->dpq.tx_dbells;
+	return 0;
+}
+
+struct sfc_sw_xstat_descr sfc_sw_xstats[] = {
+	{
+		.name = "dbells",
+		.type = SFC_SW_STATS_RX,
+		.get_val  = sfc_get_sw_xstat_val_rx_dbells,
+	},
+	{
+		.name = "dbells",
+		.type = SFC_SW_STATS_TX,
+		.get_val  = sfc_get_sw_xstat_val_tx_dbells,
+	}
+};
+
+static int
+sfc_sw_stat_get_name(struct sfc_adapter *sa,
+		     const struct sfc_sw_xstat_descr *sw_xstat, char *name,
+		     size_t name_size, unsigned int id_off)
+{
+	const char *prefix;
+	int ret;
+
+	switch (sw_xstat->type) {
+	case SFC_SW_STATS_RX:
+		prefix = "rx";
+		break;
+	case SFC_SW_STATS_TX:
+		prefix = "tx";
+		break;
+	default:
+		sfc_err(sa, "%s: unknown software statistics type %d",
+			__func__, sw_xstat->type);
+		return -EINVAL;
+	}
+
+	if (id_off == 0) {
+		ret = snprintf(name, name_size, "%s_%s", prefix,
+							 sw_xstat->name);
+		if (ret < 0 || ret >= (int)name_size) {
+			sfc_err(sa, "%s: failed to fill xstat name %s_%s, err %d",
+				__func__, prefix, sw_xstat->name, ret);
+			return ret > 0 ? -EINVAL : ret;
+		}
+	} else {
+		uint16_t qid = id_off - 1;
+		ret = snprintf(name, name_size, "%s_q%u_%s", prefix, qid,
+							sw_xstat->name);
+		if (ret < 0 || ret >= (int)name_size) {
+			sfc_err(sa, "%s: failed to fill xstat name %s_q%u_%s, err %d",
+				__func__, prefix, qid, sw_xstat->name, ret);
+			return ret > 0 ? -EINVAL : ret;
+		}
+	}
+
+	return 0;
+}
+
+static unsigned int
+sfc_sw_stat_get_queue_count(struct sfc_adapter *sa,
+			    const struct sfc_sw_xstat_descr *sw_xstat)
+{
+	struct sfc_adapter_shared *sas = sfc_sa2shared(sa);
+
+	switch (sw_xstat->type) {
+	case SFC_SW_STATS_RX:
+		return sas->ethdev_rxq_count;
+	case SFC_SW_STATS_TX:
+		return sas->ethdev_txq_count;
+	default:
+		sfc_err(sa, "%s: unknown software statistics type %d",
+			__func__, sw_xstat->type);
+		return 0;
+	}
+}
+
+static unsigned int
+sfc_sw_xstat_per_queue_get_count(unsigned int nb_queues)
+{
+	/* Take into account the accumulative xstat of all queues */
+	return nb_queues > 0 ? 1 + nb_queues : 0;
+}
+
+static unsigned int
+sfc_sw_xstat_get_nb_supported(struct sfc_adapter *sa,
+			      const struct sfc_sw_xstat_descr *sw_xstat)
+{
+	unsigned int nb_queues;
+
+	nb_queues = sfc_sw_stat_get_queue_count(sa, sw_xstat);
+	return sfc_sw_xstat_per_queue_get_count(nb_queues);
+}
+
+static int
+sfc_sw_stat_get_names(struct sfc_adapter *sa,
+		      const struct sfc_sw_xstat_descr *sw_xstat,
+		      struct rte_eth_xstat_name *xstats_names,
+		      unsigned int xstats_names_sz,
+		      unsigned int *nb_written,
+		      unsigned int *nb_supported)
+{
+	const size_t name_size = sizeof(xstats_names[0].name);
+	unsigned int id_base = *nb_supported;
+	unsigned int nb_queues;
+	unsigned int qid;
+	int rc;
+
+	nb_queues = sfc_sw_stat_get_queue_count(sa, sw_xstat);
+	if (nb_queues == 0)
+		return 0;
+	*nb_supported += sfc_sw_xstat_per_queue_get_count(nb_queues);
+
+	/*
+	 * The order of each software xstat type is the accumulative xstat
+	 * followed by per-queue xstats.
+	 */
+	if (*nb_written < xstats_names_sz) {
+		rc = sfc_sw_stat_get_name(sa, sw_xstat,
+					  xstats_names[*nb_written].name,
+					  name_size, *nb_written - id_base);
+		if (rc != 0)
+			return rc;
+		(*nb_written)++;
+	}
+
+	for (qid = 0; qid < nb_queues; ++qid) {
+		if (*nb_written < xstats_names_sz) {
+			rc = sfc_sw_stat_get_name(sa, sw_xstat,
+					      xstats_names[*nb_written].name,
+					      name_size, *nb_written - id_base);
+			if (rc != 0)
+				return rc;
+			(*nb_written)++;
+		}
+	}
+
+	return 0;
+}
+
+static int
+sfc_sw_xstat_get_names_by_id(struct sfc_adapter *sa,
+			     const struct sfc_sw_xstat_descr *sw_xstat,
+			     const uint64_t *ids,
+			     struct rte_eth_xstat_name *xstats_names,
+			     unsigned int size,
+			     unsigned int *nb_supported)
+{
+	const size_t name_size = sizeof(xstats_names[0].name);
+	unsigned int id_base = *nb_supported;
+	unsigned int nb_queues;
+	unsigned int i;
+	int rc;
+
+	nb_queues = sfc_sw_stat_get_queue_count(sa, sw_xstat);
+	if (nb_queues == 0)
+		return 0;
+	*nb_supported += sfc_sw_xstat_per_queue_get_count(nb_queues);
+
+	/*
+	 * The order of each software xstat type is the accumulative xstat
+	 * followed by per-queue xstats.
+	 */
+	for (i = 0; i < size; i++) {
+		if (id_base <= ids[i] && ids[i] <= id_base + nb_queues) {
+			rc = sfc_sw_stat_get_name(sa, sw_xstat,
+						  xstats_names[i].name,
+						  name_size, ids[i] - id_base);
+			if (rc != 0)
+				return rc;
+		}
+	}
+
+	return 0;
+}
+
+static void
+sfc_sw_xstat_get_values(struct sfc_adapter *sa,
+			const struct sfc_sw_xstat_descr *sw_xstat,
+			struct rte_eth_xstat *xstats,
+			unsigned int xstats_size,
+			unsigned int *nb_written,
+			unsigned int *nb_supported)
+{
+	unsigned int qid;
+	uint64_t value;
+	struct rte_eth_xstat *accum_xstat;
+	bool count_accum_value = false;
+	unsigned int nb_queues;
+
+	nb_queues = sfc_sw_stat_get_queue_count(sa, sw_xstat);
+	if (nb_queues == 0)
+		return;
+	*nb_supported += sfc_sw_xstat_per_queue_get_count(nb_queues);
+
+	/*
+	 * The order of each software xstat type is the accumulative xstat
+	 * followed by per-queue xstats.
+	 */
+	if (*nb_written < xstats_size) {
+		count_accum_value = true;
+		accum_xstat = &xstats[*nb_written];
+		xstats[*nb_written].id = *nb_written;
+		xstats[*nb_written].value = 0;
+		(*nb_written)++;
+	}
+
+	for (qid = 0; qid < nb_queues; ++qid) {
+		value = sw_xstat->get_val(sa, qid);
+
+		if (*nb_written < xstats_size) {
+			xstats[*nb_written].id = *nb_written;
+			xstats[*nb_written].value = value;
+			(*nb_written)++;
+		}
+
+		if (count_accum_value)
+			accum_xstat->value += value;
+	}
+}
+
+static void
+sfc_sw_xstat_get_values_by_id(struct sfc_adapter *sa,
+			      const struct sfc_sw_xstat_descr *sw_xstat,
+			      const uint64_t *ids,
+			      uint64_t *values,
+			      unsigned int ids_size,
+			      unsigned int *nb_supported)
+{
+	rte_spinlock_t *bmp_lock = &sa->sw_xstats.queues_bitmap_lock;
+	struct rte_bitmap *bmp = sa->sw_xstats.queues_bitmap;
+	unsigned int id_base = *nb_supported;
+	bool count_accum_value = false;
+	unsigned int accum_value_idx;
+	uint64_t accum_value = 0;
+	unsigned int i, qid;
+	unsigned int nb_queues;
+
+
+	rte_spinlock_lock(bmp_lock);
+	rte_bitmap_reset(bmp);
+
+	nb_queues = sfc_sw_stat_get_queue_count(sa, sw_xstat);
+	if (nb_queues == 0)
+		goto unlock;
+	*nb_supported += sfc_sw_xstat_per_queue_get_count(nb_queues);
+
+	/*
+	 * The order of each software xstat type is the accumulative xstat
+	 * followed by per-queue xstats.
+	 */
+	for (i = 0; i < ids_size; i++) {
+		if (id_base <= ids[i] && ids[i] <= (id_base + nb_queues)) {
+			if (ids[i] == id_base) { /* Accumulative value */
+				count_accum_value = true;
+				accum_value_idx = i;
+				continue;
+			}
+			qid = ids[i] - id_base - 1;
+			values[i] = sw_xstat->get_val(sa, qid);
+			accum_value += values[i];
+
+			rte_bitmap_set(bmp, qid);
+		}
+	}
+
+	if (count_accum_value) {
+		for (qid = 0; qid < nb_queues; ++qid) {
+			if (rte_bitmap_get(bmp, qid) != 0)
+				continue;
+			values[accum_value_idx] += sw_xstat->get_val(sa, qid);
+		}
+		values[accum_value_idx] += accum_value;
+	}
+
+unlock:
+	rte_spinlock_unlock(bmp_lock);
+}
+
+unsigned int
+sfc_sw_xstats_get_nb_supported(struct sfc_adapter *sa)
+{
+	unsigned int nb_supported = 0;
+	unsigned int i;
+
+	SFC_ASSERT(sfc_adapter_is_locked(sa));
+
+	for (i = 0; i < RTE_DIM(sfc_sw_xstats); i++) {
+		nb_supported += sfc_sw_xstat_get_nb_supported(sa,
+							     &sfc_sw_xstats[i]);
+	}
+
+	return nb_supported;
+}
+
+void
+sfc_sw_xstats_get_vals(struct sfc_adapter *sa,
+		       struct rte_eth_xstat *xstats,
+		       unsigned int xstats_count,
+		       unsigned int *nb_written,
+		       unsigned int *nb_supported)
+{
+	uint64_t *reset_vals = sa->sw_xstats.reset_vals;
+	unsigned int sw_xstats_offset;
+	unsigned int i;
+
+	sfc_adapter_lock(sa);
+
+	sw_xstats_offset = *nb_supported;
+
+	for (i = 0; i < RTE_DIM(sfc_sw_xstats); i++) {
+		sfc_sw_xstat_get_values(sa, &sfc_sw_xstats[i], xstats,
+					xstats_count, nb_written, nb_supported);
+	}
+
+	for (i = sw_xstats_offset; i < *nb_written; i++)
+		xstats[i].value -= reset_vals[i - sw_xstats_offset];
+
+	sfc_adapter_unlock(sa);
+}
+
+int
+sfc_sw_xstats_get_names(struct sfc_adapter *sa,
+			struct rte_eth_xstat_name *xstats_names,
+			unsigned int xstats_count,
+			unsigned int *nb_written,
+			unsigned int *nb_supported)
+{
+	unsigned int i;
+	int ret;
+
+	sfc_adapter_lock(sa);
+
+	for (i = 0; i < RTE_DIM(sfc_sw_xstats); i++) {
+		ret = sfc_sw_stat_get_names(sa, &sfc_sw_xstats[i],
+					    xstats_names, xstats_count,
+					    nb_written, nb_supported);
+		if (ret != 0) {
+			sfc_adapter_unlock(sa);
+			return ret;
+		}
+	}
+
+	sfc_adapter_unlock(sa);
+
+	return 0;
+}
+
+void
+sfc_sw_xstats_get_vals_by_id(struct sfc_adapter *sa,
+			     const uint64_t *ids,
+			     uint64_t *values,
+			     unsigned int n,
+			     unsigned int *nb_supported)
+{
+	uint64_t *reset_vals = sa->sw_xstats.reset_vals;
+	unsigned int sw_xstats_offset;
+	unsigned int i;
+
+	sfc_adapter_lock(sa);
+
+	sw_xstats_offset = *nb_supported;
+
+	for (i = 0; i < RTE_DIM(sfc_sw_xstats); i++) {
+		sfc_sw_xstat_get_values_by_id(sa, &sfc_sw_xstats[i], ids,
+					      values, n, nb_supported);
+	}
+
+	for (i = 0; i < n; i++) {
+		if (sw_xstats_offset <= ids[i] && ids[i] < *nb_supported)
+			values[i] -= reset_vals[ids[i] - sw_xstats_offset];
+	}
+
+	sfc_adapter_unlock(sa);
+}
+
+int
+sfc_sw_xstats_get_names_by_id(struct sfc_adapter *sa,
+			      const uint64_t *ids,
+			      struct rte_eth_xstat_name *xstats_names,
+			      unsigned int size,
+			      unsigned int *nb_supported)
+{
+	unsigned int i;
+	int ret;
+
+	sfc_adapter_lock(sa);
+
+	for (i = 0; i < RTE_DIM(sfc_sw_xstats); i++) {
+		ret = sfc_sw_xstat_get_names_by_id(sa, &sfc_sw_xstats[i], ids,
+						   xstats_names, size,
+						   nb_supported);
+		if (ret != 0) {
+			sfc_adapter_unlock(sa);
+			SFC_ASSERT(ret < 0);
+			return ret;
+		}
+	}
+
+	sfc_adapter_unlock(sa);
+
+	return 0;
+}
+
+static void
+sfc_sw_xstat_reset(struct sfc_adapter *sa, struct sfc_sw_xstat_descr *sw_xstat,
+		   uint64_t *reset_vals)
+{
+	unsigned int nb_queues;
+	unsigned int qid;
+	uint64_t *accum_xstat_reset;
+
+	SFC_ASSERT(sfc_adapter_is_locked(sa));
+
+	nb_queues = sfc_sw_stat_get_queue_count(sa, sw_xstat);
+	if (nb_queues == 0)
+		return;
+
+	/*
+	 * The order of each software xstat type is the accumulative xstat
+	 * followed by per-queue xstats.
+	 */
+	accum_xstat_reset = reset_vals;
+	*accum_xstat_reset = 0;
+	reset_vals++;
+
+	for (qid = 0; qid < nb_queues; ++qid) {
+		reset_vals[qid] = sw_xstat->get_val(sa, qid);
+		*accum_xstat_reset += reset_vals[qid];
+	}
+}
+
+void
+sfc_sw_xstats_reset(struct sfc_adapter *sa)
+{
+	uint64_t *reset_vals = sa->sw_xstats.reset_vals;
+	struct sfc_sw_xstat_descr *sw_xstat;
+	unsigned int i;
+
+	SFC_ASSERT(sfc_adapter_is_locked(sa));
+
+	for (i = 0; i < RTE_DIM(sfc_sw_xstats); i++) {
+		sw_xstat = &sfc_sw_xstats[i];
+		sfc_sw_xstat_reset(sa, sw_xstat, reset_vals);
+		reset_vals += sfc_sw_xstat_get_nb_supported(sa, sw_xstat);
+	}
+}
+
+int
+sfc_sw_xstats_configure(struct sfc_adapter *sa)
+{
+	uint64_t **reset_vals = &sa->sw_xstats.reset_vals;
+	size_t nb_supported = 0;
+	unsigned int i;
+
+	for (i = 0; i < RTE_DIM(sfc_sw_xstats); i++)
+		nb_supported += sfc_sw_xstat_get_nb_supported(sa,
+							&sfc_sw_xstats[i]);
+
+	*reset_vals = rte_realloc(*reset_vals,
+				  nb_supported * sizeof(**reset_vals), 0);
+	if (*reset_vals == NULL)
+		return ENOMEM;
+
+	memset(*reset_vals, 0, nb_supported * sizeof(**reset_vals));
+
+	return 0;
+}
+
+static void
+sfc_sw_xstats_free_queues_bitmap(struct sfc_adapter *sa)
+{
+	rte_bitmap_free(sa->sw_xstats.queues_bitmap);
+	rte_free(sa->sw_xstats.queues_bitmap_mem);
+}
+
+static int
+sfc_sw_xstats_alloc_queues_bitmap(struct sfc_adapter *sa)
+{
+	struct rte_bitmap **queues_bitmap = &sa->sw_xstats.queues_bitmap;
+	void **queues_bitmap_mem = &sa->sw_xstats.queues_bitmap_mem;
+	uint32_t bmp_size;
+	int rc;
+
+	bmp_size = rte_bitmap_get_memory_footprint(RTE_MAX_QUEUES_PER_PORT);
+	*queues_bitmap_mem = NULL;
+	*queues_bitmap = NULL;
+
+	*queues_bitmap_mem = rte_calloc_socket("bitmap_mem", bmp_size, 1, 0,
+					       sa->socket_id);
+	if (*queues_bitmap_mem == NULL)
+		return ENOMEM;
+
+	*queues_bitmap = rte_bitmap_init(RTE_MAX_QUEUES_PER_PORT,
+					 *queues_bitmap_mem, bmp_size);
+	if (*queues_bitmap == NULL) {
+		rc = EINVAL;
+		goto fail;
+	}
+
+	rte_spinlock_init(&sa->sw_xstats.queues_bitmap_lock);
+	return 0;
+
+fail:
+	sfc_sw_xstats_free_queues_bitmap(sa);
+	return rc;
+}
+
+int
+sfc_sw_xstats_init(struct sfc_adapter *sa)
+{
+	sa->sw_xstats.reset_vals = NULL;
+
+	return sfc_sw_xstats_alloc_queues_bitmap(sa);
+}
+
+void
+sfc_sw_xstats_close(struct sfc_adapter *sa)
+{
+	rte_free(sa->sw_xstats.reset_vals);
+	sa->sw_xstats.reset_vals = NULL;
+
+	sfc_sw_xstats_free_queues_bitmap(sa);
+}
diff --git a/drivers/net/sfc/sfc_sw_stats.h b/drivers/net/sfc/sfc_sw_stats.h
new file mode 100644
index 0000000000..1abded8018
--- /dev/null
+++ b/drivers/net/sfc/sfc_sw_stats.h
@@ -0,0 +1,49 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ *
+ * Copyright(c) 2021 Xilinx, Inc.
+ */
+#ifndef _SFC_SW_STATS_H
+#define _SFC_SW_STATS_H
+
+#include <rte_dev.h>
+
+#include "sfc.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void sfc_sw_xstats_get_vals(struct sfc_adapter *sa,
+			    struct rte_eth_xstat *xstats,
+			    unsigned int xstats_count, unsigned int *nb_written,
+			    unsigned int *nb_supported);
+
+int sfc_sw_xstats_get_names(struct sfc_adapter *sa,
+			    struct rte_eth_xstat_name *xstats_names,
+			    unsigned int xstats_count, unsigned int *nb_written,
+			    unsigned int *nb_supported);
+
+void sfc_sw_xstats_get_vals_by_id(struct sfc_adapter *sa, const uint64_t *ids,
+				  uint64_t *values, unsigned int n,
+				  unsigned int *nb_supported);
+
+int sfc_sw_xstats_get_names_by_id(struct sfc_adapter *sa, const uint64_t *ids,
+				  struct rte_eth_xstat_name *xstats_names,
+				  unsigned int size,
+				  unsigned int *nb_supported);
+
+unsigned int sfc_sw_xstats_get_nb_supported(struct sfc_adapter *sa);
+
+int sfc_sw_xstats_configure(struct sfc_adapter *sa);
+
+void sfc_sw_xstats_reset(struct sfc_adapter *sa);
+
+int sfc_sw_xstats_init(struct sfc_adapter *sa);
+
+void sfc_sw_xstats_close(struct sfc_adapter *sa);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  /* _SFC_SW_STATS_H */
diff --git a/drivers/net/sfc/sfc_tx.c b/drivers/net/sfc/sfc_tx.c
index ce2a9a6a4f..49b239f4d2 100644
--- a/drivers/net/sfc/sfc_tx.c
+++ b/drivers/net/sfc/sfc_tx.c
@@ -980,8 +980,10 @@ sfc_efx_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 				       txq->completed, &txq->added);
 		SFC_ASSERT(rc == 0);
 
-		if (likely(pushed != txq->added))
+		if (likely(pushed != txq->added)) {
 			efx_tx_qpush(txq->common, txq->added, pushed);
+			txq->dp.dpq.tx_dbells++;
+		}
 	}
 
 #if SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE
-- 
2.30.2