X-Git-Url: http://git.droids-corp.org/?a=blobdiff_plain;f=drivers%2Fnet%2Fsfc%2Fsfc_tx.c;h=4037802e6af7048fc7e3e97328d360d0dd099f31;hb=f7a66f9365fe442d743639d50638a8439e712b96;hp=6131a498c366707203433b573acbb0520dc199b9;hpb=244cfa79a41c70f8c58692dd2601687a65e928c3;p=dpdk.git

diff --git a/drivers/net/sfc/sfc_tx.c b/drivers/net/sfc/sfc_tx.c
index 6131a498c3..4037802e6a 100644
--- a/drivers/net/sfc/sfc_tx.c
+++ b/drivers/net/sfc/sfc_tx.c
@@ -1,32 +1,10 @@
-/*-
- *   BSD LICENSE
+/* SPDX-License-Identifier: BSD-3-Clause
  *
- * Copyright (c) 2016-2017 Solarflare Communications Inc.
+ * Copyright (c) 2016-2018 Solarflare Communications Inc.
  * All rights reserved.
  *
  * This software was jointly developed between OKTET Labs (under contract
  * for Solarflare) and Solarflare Communications, Inc.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *
- * 1. Redistributions of source code must retain the above copyright notice,
- *    this list of conditions and the following disclaimer.
- * 2. Redistributions in binary form must reproduce the above copyright notice,
- *    this list of conditions and the following disclaimer in the documentation
- *    and/or other materials provided with the distribution.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
- * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
- * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
- * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
- * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
- * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
- * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
- * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
- * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
- * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
- * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */
 
 #include "sfc.h"
@@ -35,6 +13,7 @@
 #include "sfc_ev.h"
 #include "sfc_tx.h"
 #include "sfc_tweak.h"
+#include "sfc_kvargs.h"
 
 /*
  * Maximum number of TX queue flush attempts in case of
@@ -55,12 +34,50 @@
  */
 #define SFC_TX_QFLUSH_POLL_ATTEMPTS	(2000)
 
-static int
-sfc_tx_qcheck_conf(struct sfc_adapter *sa, uint16_t nb_tx_desc,
-		   const struct rte_eth_txconf *tx_conf)
+uint64_t
+sfc_tx_get_dev_offload_caps(struct sfc_adapter *sa)
+{
+	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
+	uint64_t caps = 0;
+
+	if ((sa->priv.dp_tx->features & SFC_DP_TX_FEAT_VLAN_INSERT) &&
+	    encp->enc_hw_tx_insert_vlan_enabled)
+		caps |= DEV_TX_OFFLOAD_VLAN_INSERT;
+
+	if (sa->priv.dp_tx->features & SFC_DP_TX_FEAT_MULTI_SEG)
+		caps |= DEV_TX_OFFLOAD_MULTI_SEGS;
+
+	if ((~sa->priv.dp_tx->features & SFC_DP_TX_FEAT_MULTI_POOL) &&
+	    (~sa->priv.dp_tx->features & SFC_DP_TX_FEAT_REFCNT))
+		caps |= DEV_TX_OFFLOAD_MBUF_FAST_FREE;
+
+	return caps;
+}
+
+uint64_t
+sfc_tx_get_queue_offload_caps(struct sfc_adapter *sa)
 {
-	unsigned int flags = tx_conf->txq_flags;
 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
+	uint64_t caps = 0;
+
+	caps |= DEV_TX_OFFLOAD_IPV4_CKSUM;
+	caps |= DEV_TX_OFFLOAD_UDP_CKSUM;
+	caps |= DEV_TX_OFFLOAD_TCP_CKSUM;
+
+	if (encp->enc_tunnel_encapsulations_supported)
+		caps |= DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM;
+
+	if (sa->tso)
+		caps |= DEV_TX_OFFLOAD_TCP_TSO;
+
+	return caps;
+}
+
+static int
+sfc_tx_qcheck_conf(struct sfc_adapter *sa, unsigned int txq_max_fill_level,
+		   const struct rte_eth_txconf *tx_conf,
+		   uint64_t offloads)
+{
 	int rc = 0;
 
 	if (tx_conf->tx_rs_thresh != 0) {
@@ -68,35 +85,23 @@ sfc_tx_qcheck_conf(struct sfc_adapter *sa, uint16_t nb_tx_desc,
 		rc = EINVAL;
 	}
 
-	if (tx_conf->tx_free_thresh > EFX_TXQ_LIMIT(nb_tx_desc)) {
+	if (tx_conf->tx_free_thresh > txq_max_fill_level) {
 		sfc_err(sa,
 			"TxQ free threshold too large: %u vs maximum %u",
-			tx_conf->tx_free_thresh, EFX_TXQ_LIMIT(nb_tx_desc));
+			tx_conf->tx_free_thresh, txq_max_fill_level);
 		rc = EINVAL;
 	}
 
 	if (tx_conf->tx_thresh.pthresh != 0 ||
 	    tx_conf->tx_thresh.hthresh != 0 ||
 	    tx_conf->tx_thresh.wthresh != 0) {
-		sfc_err(sa,
+		sfc_warn(sa,
 			"prefetch/host/writeback thresholds are not supported");
-		rc = EINVAL;
-	}
-
-	if (!encp->enc_hw_tx_insert_vlan_enabled &&
-	    (flags & ETH_TXQ_FLAGS_NOVLANOFFL) == 0) {
-		sfc_err(sa, "VLAN offload is not supported");
-		rc = EINVAL;
-	}
-
-	if ((flags & ETH_TXQ_FLAGS_NOXSUMSCTP) == 0) {
-		sfc_err(sa, "SCTP offload is not supported");
-		rc = EINVAL;
 	}
 
 	/* We either perform both TCP and UDP offload, or no offload at all */
-	if (((flags & ETH_TXQ_FLAGS_NOXSUMTCP) == 0) !=
-	    ((flags & ETH_TXQ_FLAGS_NOXSUMUDP) == 0)) {
+	if (((offloads & DEV_TX_OFFLOAD_TCP_CKSUM) == 0) !=
+	    ((offloads & DEV_TX_OFFLOAD_UDP_CKSUM) == 0)) {
 		sfc_err(sa, "TCP and UDP offloads can't be set independently");
 		rc = EINVAL;
 	}
@@ -105,33 +110,10 @@ sfc_tx_qcheck_conf(struct sfc_adapter *sa, uint16_t nb_tx_desc,
 }
 
 void
-sfc_tx_qflush_done(struct sfc_txq *txq)
+sfc_tx_qflush_done(struct sfc_txq_info *txq_info)
 {
-	txq->state |= SFC_TXQ_FLUSHED;
-	txq->state &= ~SFC_TXQ_FLUSHING;
-}
-
-static void
-sfc_tx_reap(struct sfc_txq *txq)
-{
-	unsigned int    completed;
-
-
-	sfc_ev_qpoll(txq->evq);
-
-	for (completed = txq->completed;
-	     completed != txq->pending; completed++) {
-		struct sfc_tx_sw_desc *txd;
-
-		txd = &txq->sw_ring[completed & txq->ptr_mask];
-
-		if (txd->mbuf != NULL) {
-			rte_pktmbuf_free(txd->mbuf);
-			txd->mbuf = NULL;
-		}
-	}
-
-	txq->completed = completed;
+	txq_info->state |= SFC_TXQ_FLUSHED;
+	txq_info->state &= ~SFC_TXQ_FLUSHING;
 }
 
 int
@@ -140,95 +122,103 @@ sfc_tx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
 	     const struct rte_eth_txconf *tx_conf)
 {
 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
+	unsigned int txq_entries;
+	unsigned int evq_entries;
+	unsigned int txq_max_fill_level;
 	struct sfc_txq_info *txq_info;
 	struct sfc_evq *evq;
 	struct sfc_txq *txq;
-	unsigned int evq_index = sfc_evq_index_by_txq_sw_index(sa, sw_index);
 	int rc = 0;
+	struct sfc_dp_tx_qcreate_info info;
+	uint64_t offloads;
+	struct sfc_dp_tx_hw_limits hw_limits;
 
 	sfc_log_init(sa, "TxQ = %u", sw_index);
 
-	rc = sfc_tx_qcheck_conf(sa, nb_tx_desc, tx_conf);
+	memset(&hw_limits, 0, sizeof(hw_limits));
+	hw_limits.txq_max_entries = sa->txq_max_entries;
+	hw_limits.txq_min_entries = sa->txq_min_entries;
+
+	rc = sa->priv.dp_tx->qsize_up_rings(nb_tx_desc, &hw_limits,
+					    &txq_entries, &evq_entries,
+					    &txq_max_fill_level);
+	if (rc != 0)
+		goto fail_size_up_rings;
+	SFC_ASSERT(txq_entries >= sa->txq_min_entries);
+	SFC_ASSERT(txq_entries <= sa->txq_max_entries);
+	SFC_ASSERT(txq_entries >= nb_tx_desc);
+	SFC_ASSERT(txq_max_fill_level <= nb_tx_desc);
+
+	offloads = tx_conf->offloads |
+		sa->eth_dev->data->dev_conf.txmode.offloads;
+	rc = sfc_tx_qcheck_conf(sa, txq_max_fill_level, tx_conf, offloads);
 	if (rc != 0)
 		goto fail_bad_conf;
 
-	SFC_ASSERT(sw_index < sa->txq_count);
-	txq_info = &sa->txq_info[sw_index];
+	SFC_ASSERT(sw_index < sfc_sa2shared(sa)->txq_count);
+	txq_info = &sfc_sa2shared(sa)->txq_info[sw_index];
 
-	SFC_ASSERT(nb_tx_desc <= sa->txq_max_entries);
-	txq_info->entries = nb_tx_desc;
+	txq_info->entries = txq_entries;
 
-	rc = sfc_ev_qinit(sa, evq_index, txq_info->entries, socket_id);
+	rc = sfc_ev_qinit(sa, SFC_EVQ_TYPE_TX, sw_index,
+			  evq_entries, socket_id, &evq);
 	if (rc != 0)
 		goto fail_ev_qinit;
 
-	evq = sa->evq_info[evq_index].evq;
-
-	rc = ENOMEM;
-	txq = rte_zmalloc_socket("sfc-txq", sizeof(*txq), 0, socket_id);
-	if (txq == NULL)
-		goto fail_txq_alloc;
+	txq = &sa->txq_ctrl[sw_index];
+	txq->hw_index = sw_index;
+	txq->evq = evq;
+	txq_info->free_thresh =
+		(tx_conf->tx_free_thresh) ? tx_conf->tx_free_thresh :
+		SFC_TX_DEFAULT_FREE_THRESH;
+	txq_info->offloads = offloads;
 
-	rc = sfc_dma_alloc(sa, "txq", sw_index, EFX_TXQ_SIZE(txq_info->entries),
+	rc = sfc_dma_alloc(sa, "txq", sw_index,
+			   efx_txq_size(sa->nic, txq_info->entries),
 			   socket_id, &txq->mem);
 	if (rc != 0)
 		goto fail_dma_alloc;
 
-	rc = ENOMEM;
-	txq->pend_desc = rte_calloc_socket("sfc-txq-pend-desc",
-					   EFX_TXQ_LIMIT(txq_info->entries),
-					   sizeof(efx_desc_t), 0, socket_id);
-	if (txq->pend_desc == NULL)
-		goto fail_pend_desc_alloc;
-
-	rc = ENOMEM;
-	txq->sw_ring = rte_calloc_socket("sfc-txq-desc", txq_info->entries,
-					 sizeof(*txq->sw_ring), 0, socket_id);
-	if (txq->sw_ring == NULL)
-		goto fail_desc_alloc;
+	memset(&info, 0, sizeof(info));
+	info.max_fill_level = txq_max_fill_level;
+	info.free_thresh = txq_info->free_thresh;
+	info.offloads = offloads;
+	info.txq_entries = txq_info->entries;
+	info.dma_desc_size_max = encp->enc_tx_dma_desc_size_max;
+	info.txq_hw_ring = txq->mem.esm_base;
+	info.evq_entries = evq_entries;
+	info.evq_hw_ring = evq->mem.esm_base;
+	info.hw_index = txq->hw_index;
+	info.mem_bar = sa->mem_bar.esb_base;
+	info.vi_window_shift = encp->enc_vi_window_shift;
+	info.tso_tcp_header_offset_limit =
+		encp->enc_tx_tso_tcp_header_offset_limit;
+
+	rc = sa->priv.dp_tx->qcreate(sa->eth_dev->data->port_id, sw_index,
+				     &RTE_ETH_DEV_TO_PCI(sa->eth_dev)->addr,
+				     socket_id, &info, &txq_info->dp);
+	if (rc != 0)
+		goto fail_dp_tx_qinit;
 
-	if (sa->tso) {
-		rc = sfc_tso_alloc_tsoh_objs(txq->sw_ring, txq_info->entries,
-					     socket_id);
-		if (rc != 0)
-			goto fail_alloc_tsoh_objs;
-	}
+	evq->dp_txq = txq_info->dp;
 
-	txq->state = SFC_TXQ_INITIALIZED;
-	txq->ptr_mask = txq_info->entries - 1;
-	txq->free_thresh = (tx_conf->tx_free_thresh) ? tx_conf->tx_free_thresh :
-						     SFC_TX_DEFAULT_FREE_THRESH;
-	txq->dma_desc_size_max = encp->enc_tx_dma_desc_size_max;
-	txq->hw_index = sw_index;
-	txq->flags = tx_conf->txq_flags;
-	txq->evq = evq;
+	txq_info->state = SFC_TXQ_INITIALIZED;
 
-	evq->txq = txq;
-
-	txq_info->txq = txq;
 	txq_info->deferred_start = (tx_conf->tx_deferred_start != 0);
 
 	return 0;
 
-fail_alloc_tsoh_objs:
-	rte_free(txq->sw_ring);
-
-fail_desc_alloc:
-	rte_free(txq->pend_desc);
-
-fail_pend_desc_alloc:
+fail_dp_tx_qinit:
 	sfc_dma_free(sa, &txq->mem);
 
 fail_dma_alloc:
-	rte_free(txq);
-
-fail_txq_alloc:
-	sfc_ev_qfini(sa, evq_index);
+	sfc_ev_qfini(evq);
 
 fail_ev_qinit:
 	txq_info->entries = 0;
 
 fail_bad_conf:
+fail_size_up_rings:
 	sfc_log_init(sa, "failed (TxQ = %u, rc = %d)", sw_index, rc);
 	return rc;
 }
@@ -241,22 +231,25 @@ sfc_tx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
 
 	sfc_log_init(sa, "TxQ = %u", sw_index);
 
-	SFC_ASSERT(sw_index < sa->txq_count);
-	txq_info = &sa->txq_info[sw_index];
+	SFC_ASSERT(sw_index < sfc_sa2shared(sa)->txq_count);
+	sa->eth_dev->data->tx_queues[sw_index] = NULL;
+
+	txq_info = &sfc_sa2shared(sa)->txq_info[sw_index];
 
-	txq = txq_info->txq;
-	SFC_ASSERT(txq != NULL);
-	SFC_ASSERT(txq->state == SFC_TXQ_INITIALIZED);
+	SFC_ASSERT(txq_info->state == SFC_TXQ_INITIALIZED);
 
-	sfc_tso_free_tsoh_objs(txq->sw_ring, txq_info->entries);
+	sa->priv.dp_tx->qdestroy(txq_info->dp);
+	txq_info->dp = NULL;
 
-	txq_info->txq = NULL;
+	txq_info->state &= ~SFC_TXQ_INITIALIZED;
 	txq_info->entries = 0;
 
-	rte_free(txq->sw_ring);
-	rte_free(txq->pend_desc);
+	txq = &sa->txq_ctrl[sw_index];
+
 	sfc_dma_free(sa, &txq->mem);
-	rte_free(txq);
+
+	sfc_ev_qfini(txq->evq);
+	txq->evq = NULL;
 }
 
 static int
@@ -303,14 +296,39 @@ sfc_tx_check_mode(struct sfc_adapter *sa, const struct rte_eth_txmode *txmode)
 	return rc;
 }
 
+/**
+ * Destroy excess queues that are no longer needed after reconfiguration
+ * or complete close.
+ */
+static void
+sfc_tx_fini_queues(struct sfc_adapter *sa, unsigned int nb_tx_queues)
+{
+	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
+	int sw_index;
+
+	SFC_ASSERT(nb_tx_queues <= sas->txq_count);
+
+	sw_index = sas->txq_count;
+	while (--sw_index >= (int)nb_tx_queues) {
+		if (sas->txq_info[sw_index].state & SFC_TXQ_INITIALIZED)
+			sfc_tx_qfini(sa, sw_index);
+	}
+
+	sas->txq_count = nb_tx_queues;
+}
+
 int
-sfc_tx_init(struct sfc_adapter *sa)
+sfc_tx_configure(struct sfc_adapter *sa)
 {
+	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 	const struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
-	unsigned int sw_index;
+	const unsigned int nb_tx_queues = sa->eth_dev->data->nb_tx_queues;
 	int rc = 0;
 
+	sfc_log_init(sa, "nb_tx_queues=%u (old %u)",
+		     nb_tx_queues, sas->txq_count);
+
 	/*
 	 * The datapath implementation assumes absence of boundary
 	 * limits on Tx DMA descriptors. Addition of these checks on
@@ -325,28 +343,71 @@ sfc_tx_init(struct sfc_adapter *sa)
 	if (rc != 0)
 		goto fail_check_mode;
 
-	sa->txq_count = sa->eth_dev->data->nb_tx_queues;
+	if (nb_tx_queues == sas->txq_count)
+		goto done;
+
+	if (sas->txq_info == NULL) {
+		sas->txq_info = rte_calloc_socket("sfc-txqs", nb_tx_queues,
+						  sizeof(sas->txq_info[0]), 0,
+						  sa->socket_id);
+		if (sas->txq_info == NULL)
+			goto fail_txqs_alloc;
 
-	sa->txq_info = rte_calloc_socket("sfc-txqs", sa->txq_count,
-					 sizeof(sa->txq_info[0]), 0,
-					 sa->socket_id);
-	if (sa->txq_info == NULL)
-		goto fail_txqs_alloc;
+		/*
+		 * Allocate primary process only TxQ control from heap
+		 * since it should not be shared.
+		 */
+		rc = ENOMEM;
+		sa->txq_ctrl = calloc(nb_tx_queues, sizeof(sa->txq_ctrl[0]));
+		if (sa->txq_ctrl == NULL)
+			goto fail_txqs_ctrl_alloc;
+	} else {
+		struct sfc_txq_info *new_txq_info;
+		struct sfc_txq *new_txq_ctrl;
+
+		if (nb_tx_queues < sas->txq_count)
+			sfc_tx_fini_queues(sa, nb_tx_queues);
+
+		new_txq_info =
+			rte_realloc(sas->txq_info,
+				    nb_tx_queues * sizeof(sas->txq_info[0]), 0);
+		if (new_txq_info == NULL && nb_tx_queues > 0)
+			goto fail_txqs_realloc;
+
+		new_txq_ctrl = realloc(sa->txq_ctrl,
+				       nb_tx_queues * sizeof(sa->txq_ctrl[0]));
+		if (new_txq_ctrl == NULL && nb_tx_queues > 0)
+			goto fail_txqs_ctrl_realloc;
+
+		sas->txq_info = new_txq_info;
+		sa->txq_ctrl = new_txq_ctrl;
+		if (nb_tx_queues > sas->txq_count) {
+			memset(&sas->txq_info[sas->txq_count], 0,
+			       (nb_tx_queues - sas->txq_count) *
+			       sizeof(sas->txq_info[0]));
+			memset(&sa->txq_ctrl[sas->txq_count], 0,
+			       (nb_tx_queues - sas->txq_count) *
+			       sizeof(sa->txq_ctrl[0]));
+		}
+	}
 
-	for (sw_index = 0; sw_index < sa->txq_count; ++sw_index) {
-		rc = sfc_tx_qinit_info(sa, sw_index);
+	while (sas->txq_count < nb_tx_queues) {
+		rc = sfc_tx_qinit_info(sa, sas->txq_count);
 		if (rc != 0)
 			goto fail_tx_qinit_info;
+
+		sas->txq_count++;
 	}
 
+done:
 	return 0;
 
 fail_tx_qinit_info:
-	rte_free(sa->txq_info);
-	sa->txq_info = NULL;
-
+fail_txqs_ctrl_realloc:
+fail_txqs_realloc:
+fail_txqs_ctrl_alloc:
 fail_txqs_alloc:
-	sa->txq_count = 0;
+	sfc_tx_close(sa);
 
 fail_check_mode:
 fail_tx_dma_desc_boundary:
@@ -355,62 +416,63 @@ fail_tx_dma_desc_boundary:
 }
 
 void
-sfc_tx_fini(struct sfc_adapter *sa)
+sfc_tx_close(struct sfc_adapter *sa)
 {
-	int sw_index;
+	sfc_tx_fini_queues(sa, 0);
 
-	sw_index = sa->txq_count;
-	while (--sw_index >= 0) {
-		if (sa->txq_info[sw_index].txq != NULL)
-			sfc_tx_qfini(sa, sw_index);
-	}
+	free(sa->txq_ctrl);
+	sa->txq_ctrl = NULL;
 
-	rte_free(sa->txq_info);
-	sa->txq_info = NULL;
-	sa->txq_count = 0;
+	rte_free(sfc_sa2shared(sa)->txq_info);
+	sfc_sa2shared(sa)->txq_info = NULL;
 }
 
 int
 sfc_tx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 {
+	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
+	uint64_t offloads_supported = sfc_tx_get_dev_offload_caps(sa) |
+				      sfc_tx_get_queue_offload_caps(sa);
 	struct rte_eth_dev_data *dev_data;
 	struct sfc_txq_info *txq_info;
 	struct sfc_txq *txq;
 	struct sfc_evq *evq;
-	uint16_t flags;
+	uint16_t flags = 0;
 	unsigned int desc_index;
 	int rc = 0;
 
 	sfc_log_init(sa, "TxQ = %u", sw_index);
 
-	SFC_ASSERT(sw_index < sa->txq_count);
-	txq_info = &sa->txq_info[sw_index];
-
-	txq = txq_info->txq;
+	SFC_ASSERT(sw_index < sas->txq_count);
+	txq_info = &sas->txq_info[sw_index];
 
-	SFC_ASSERT(txq->state == SFC_TXQ_INITIALIZED);
+	SFC_ASSERT(txq_info->state == SFC_TXQ_INITIALIZED);
 
+	txq = &sa->txq_ctrl[sw_index];
 	evq = txq->evq;
 
-	rc = sfc_ev_qstart(sa, evq->evq_index);
+	rc = sfc_ev_qstart(evq, sfc_evq_index_by_txq_sw_index(sa, sw_index));
 	if (rc != 0)
 		goto fail_ev_qstart;
 
-	/*
-	 * It seems that DPDK has no controls regarding IPv4 offloads,
-	 * hence, we always enable it here
-	 */
-	if ((txq->flags & ETH_TXQ_FLAGS_NOXSUMTCP) ||
-	    (txq->flags & ETH_TXQ_FLAGS_NOXSUMUDP)) {
-		flags = EFX_TXQ_CKSUM_IPV4;
-	} else {
-		flags = EFX_TXQ_CKSUM_IPV4 | EFX_TXQ_CKSUM_TCPUDP;
+	if (txq_info->offloads & DEV_TX_OFFLOAD_IPV4_CKSUM)
+		flags |= EFX_TXQ_CKSUM_IPV4;
+
+	if (txq_info->offloads & DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)
+		flags |= EFX_TXQ_CKSUM_INNER_IPV4;
 
-		if (sa->tso)
-			flags |= EFX_TXQ_FATSOV2;
+	if ((txq_info->offloads & DEV_TX_OFFLOAD_TCP_CKSUM) ||
+	    (txq_info->offloads & DEV_TX_OFFLOAD_UDP_CKSUM)) {
+		flags |= EFX_TXQ_CKSUM_TCPUDP;
+
+		if (offloads_supported & DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)
+			flags |= EFX_TXQ_CKSUM_INNER_TCPUDP;
 	}
 
-	rc = efx_tx_qcreate(sa->nic, sw_index, 0, &txq->mem,
+	if (txq_info->offloads & DEV_TX_OFFLOAD_TCP_TSO)
+		flags |= EFX_TXQ_FATSOV2;
+
+	rc = efx_tx_qcreate(sa->nic, txq->hw_index, 0, &txq->mem,
 			    txq_info->entries, 0 /* not used on EF10 */,
 			    flags, evq->common,
 			    &txq->common, &desc_index);
@@ -421,12 +483,13 @@ sfc_tx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 		goto fail_tx_qcreate;
 	}
 
-	txq->added = txq->pending = txq->completed = desc_index;
-	txq->hw_vlan_tci = 0;
-
 	efx_tx_qenable(txq->common);
 
-	txq->state |= (SFC_TXQ_STARTED | SFC_TXQ_RUNNING);
+	txq_info->state |= SFC_TXQ_STARTED;
+
+	rc = sa->priv.dp_tx->qstart(txq_info->dp, evq->read_ptr, desc_index);
+	if (rc != 0)
+		goto fail_dp_qstart;
 
 	/*
 	 * It seems to be used by DPDK for debug purposes only ('rte_ether')
@@ -436,8 +499,12 @@ sfc_tx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 
 	return 0;
 
+fail_dp_qstart:
+	txq_info->state = SFC_TXQ_INITIALIZED;
+	efx_tx_qdestroy(txq->common);
+
 fail_tx_qcreate:
-	sfc_ev_qstop(sa, evq->evq_index);
+	sfc_ev_qstop(evq);
 
 fail_ev_qstart:
 	return rc;
@@ -446,37 +513,39 @@ fail_ev_qstart:
 void
 sfc_tx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
 {
+	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
 	struct rte_eth_dev_data *dev_data;
 	struct sfc_txq_info *txq_info;
 	struct sfc_txq *txq;
 	unsigned int retry_count;
 	unsigned int wait_count;
-	unsigned int txds;
+	int rc;
 
 	sfc_log_init(sa, "TxQ = %u", sw_index);
 
-	SFC_ASSERT(sw_index < sa->txq_count);
-	txq_info = &sa->txq_info[sw_index];
+	SFC_ASSERT(sw_index < sas->txq_count);
+	txq_info = &sas->txq_info[sw_index];
 
-	txq = txq_info->txq;
-
-	if (txq->state == SFC_TXQ_INITIALIZED)
+	if (txq_info->state == SFC_TXQ_INITIALIZED)
 		return;
 
-	SFC_ASSERT(txq->state & SFC_TXQ_STARTED);
+	SFC_ASSERT(txq_info->state & SFC_TXQ_STARTED);
 
-	txq->state &= ~SFC_TXQ_RUNNING;
+	txq = &sa->txq_ctrl[sw_index];
+	sa->priv.dp_tx->qstop(txq_info->dp, &txq->evq->read_ptr);
 
 	/*
 	 * Retry TX queue flushing in case of flush failed or
 	 * timeout; in the worst case it can delay for 6 seconds
 	 */
 	for (retry_count = 0;
-	     ((txq->state & SFC_TXQ_FLUSHED) == 0) &&
+	     ((txq_info->state & SFC_TXQ_FLUSHED) == 0) &&
 	     (retry_count < SFC_TX_QFLUSH_ATTEMPTS);
 	     ++retry_count) {
-		if (efx_tx_qflush(txq->common) != 0) {
-			txq->state |= SFC_TXQ_FLUSHING;
+		rc = efx_tx_qflush(txq->common);
+		if (rc != 0) {
+			txq_info->state |= (rc == EALREADY) ?
+				SFC_TXQ_FLUSHED : SFC_TXQ_FLUSH_FAILED;
 			break;
 		}
 
@@ -490,30 +559,23 @@ sfc_tx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
 		do {
 			rte_delay_ms(SFC_TX_QFLUSH_POLL_WAIT_MS);
 			sfc_ev_qpoll(txq->evq);
-		} while ((txq->state & SFC_TXQ_FLUSHING) &&
+		} while ((txq_info->state & SFC_TXQ_FLUSHING) &&
 			 wait_count++ < SFC_TX_QFLUSH_POLL_ATTEMPTS);
 
-		if (txq->state & SFC_TXQ_FLUSHING)
+		if (txq_info->state & SFC_TXQ_FLUSHING)
 			sfc_err(sa, "TxQ %u flush timed out", sw_index);
 
-		if (txq->state & SFC_TXQ_FLUSHED)
-			sfc_info(sa, "TxQ %u flushed", sw_index);
+		if (txq_info->state & SFC_TXQ_FLUSHED)
+			sfc_notice(sa, "TxQ %u flushed", sw_index);
 	}
 
-	sfc_tx_reap(txq);
+	sa->priv.dp_tx->qreap(txq_info->dp);
 
-	for (txds = 0; txds < txq_info->entries; txds++) {
-		if (txq->sw_ring[txds].mbuf != NULL) {
-			rte_pktmbuf_free(txq->sw_ring[txds].mbuf);
-			txq->sw_ring[txds].mbuf = NULL;
-		}
-	}
-
-	txq->state = SFC_TXQ_INITIALIZED;
+	txq_info->state = SFC_TXQ_INITIALIZED;
 
 	efx_tx_qdestroy(txq->common);
 
-	sfc_ev_qstop(sa, txq->evq->evq_index);
+	sfc_ev_qstop(txq->evq);
 
 	/*
 	 * It seems to be used by DPDK for debug purposes only ('rte_ether')
@@ -525,10 +587,11 @@ sfc_tx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
 int
 sfc_tx_start(struct sfc_adapter *sa)
 {
+	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
 	unsigned int sw_index;
 	int rc = 0;
 
-	sfc_log_init(sa, "txq_count = %u", sa->txq_count);
+	sfc_log_init(sa, "txq_count = %u", sas->txq_count);
 
 	if (sa->tso) {
 		if (!efx_nic_cfg_get(sa->nic)->enc_fw_assisted_tso_v2_enabled) {
@@ -541,9 +604,10 @@ sfc_tx_start(struct sfc_adapter *sa)
 	if (rc != 0)
 		goto fail_efx_tx_init;
 
-	for (sw_index = 0; sw_index < sa->txq_count; ++sw_index) {
-		if (!(sa->txq_info[sw_index].deferred_start) ||
-		    sa->txq_info[sw_index].deferred_started) {
+	for (sw_index = 0; sw_index < sas->txq_count; ++sw_index) {
+		if (sas->txq_info[sw_index].state == SFC_TXQ_INITIALIZED &&
+		    (!(sas->txq_info[sw_index].deferred_start) ||
+		     sas->txq_info[sw_index].deferred_started)) {
 			rc = sfc_tx_qstart(sa, sw_index);
 			if (rc != 0)
 				goto fail_tx_qstart;
@@ -566,19 +630,42 @@ fail_efx_tx_init:
 void
 sfc_tx_stop(struct sfc_adapter *sa)
 {
+	struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
 	unsigned int sw_index;
 
-	sfc_log_init(sa, "txq_count = %u", sa->txq_count);
+	sfc_log_init(sa, "txq_count = %u", sas->txq_count);
 
-	sw_index = sa->txq_count;
+	sw_index = sas->txq_count;
 	while (sw_index-- > 0) {
-		if (sa->txq_info[sw_index].txq != NULL)
+		if (sas->txq_info[sw_index].state & SFC_TXQ_STARTED)
 			sfc_tx_qstop(sa, sw_index);
 	}
 
 	efx_tx_fini(sa->nic);
 }
 
+static void
+sfc_efx_tx_reap(struct sfc_efx_txq *txq)
+{
+	unsigned int completed;
+
+	sfc_ev_qpoll(txq->evq);
+
+	for (completed = txq->completed;
+	     completed != txq->pending; completed++) {
+		struct sfc_efx_tx_sw_desc *txd;
+
+		txd = &txq->sw_ring[completed & txq->ptr_mask];
+
+		if (txd->mbuf != NULL) {
+			rte_pktmbuf_free(txd->mbuf);
+			txd->mbuf = NULL;
+		}
+	}
+
+	txq->completed = completed;
+}
+
 /*
  * The function is used to insert or update VLAN tag;
  * the firmware has state of the firmware tag to insert per TxQ
@@ -587,8 +674,8 @@ sfc_tx_stop(struct sfc_adapter *sa)
  * the function will update it
  */
 static unsigned int
-sfc_tx_maybe_insert_tag(struct sfc_txq *txq, struct rte_mbuf *m,
-			efx_desc_t **pend)
+sfc_efx_tx_maybe_insert_tag(struct sfc_efx_txq *txq, struct rte_mbuf *m,
+			    efx_desc_t **pend)
 {
 	uint16_t this_tag = ((m->ol_flags & PKT_TX_VLAN_PKT) ?
 			     m->vlan_tci : 0);
@@ -610,22 +697,53 @@ sfc_tx_maybe_insert_tag(struct sfc_txq *txq, struct rte_mbuf *m,
 	return 1;
 }
 
-uint16_t
-sfc_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+static uint16_t
+sfc_efx_prepare_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
+		     uint16_t nb_pkts)
+{
+	struct sfc_dp_txq *dp_txq = tx_queue;
+	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
+	const efx_nic_cfg_t *encp = efx_nic_cfg_get(txq->evq->sa->nic);
+	uint16_t i;
+
+	for (i = 0; i < nb_pkts; i++) {
+		int ret;
+
+		/*
+		 * EFX Tx datapath may require extra VLAN descriptor if VLAN
+		 * insertion offload is requested regardless the offload
+		 * requested/supported.
+		 */
+		ret = sfc_dp_tx_prepare_pkt(tx_pkts[i],
+				encp->enc_tx_tso_tcp_header_offset_limit,
+				txq->max_fill_level, EFX_TX_FATSOV2_OPT_NDESCS,
+				1);
+		if (unlikely(ret != 0)) {
+			rte_errno = ret;
+			break;
+		}
+	}
+
+	return i;
+}
+
+static uint16_t
+sfc_efx_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 {
-	struct sfc_txq *txq = (struct sfc_txq *)tx_queue;
+	struct sfc_dp_txq *dp_txq = (struct sfc_dp_txq *)tx_queue;
+	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
 	unsigned int added = txq->added;
 	unsigned int pushed = added;
 	unsigned int pkts_sent = 0;
 	efx_desc_t *pend = &txq->pend_desc[0];
-	const unsigned int hard_max_fill = EFX_TXQ_LIMIT(txq->ptr_mask + 1);
+	const unsigned int hard_max_fill = txq->max_fill_level;
 	const unsigned int soft_max_fill = hard_max_fill - txq->free_thresh;
 	unsigned int fill_level = added - txq->completed;
 	boolean_t reap_done;
 	int rc __rte_unused;
 	struct rte_mbuf **pktp;
 
-	if (unlikely((txq->state & SFC_TXQ_RUNNING) == 0))
+	if (unlikely((txq->flags & SFC_EFX_TXQ_FLAG_RUNNING) == 0))
 		goto done;
 
 	/*
@@ -636,7 +754,7 @@ sfc_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 	reap_done = (fill_level > soft_max_fill);
 
 	if (reap_done) {
-		sfc_tx_reap(txq);
+		sfc_efx_tx_reap(txq);
 		/*
 		 * Recalculate fill level since 'txq->completed'
 		 * might have changed on reap
@@ -647,6 +765,7 @@ sfc_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 	for (pkts_sent = 0, pktp = &tx_pkts[0];
 	     (pkts_sent < nb_pkts) && (fill_level <= soft_max_fill);
 	     pkts_sent++, pktp++) {
+		uint16_t		hw_vlan_tci_prev = txq->hw_vlan_tci;
 		struct rte_mbuf		*m_seg = *pktp;
 		size_t			pkt_len = m_seg->pkt_len;
 		unsigned int		pkt_descs = 0;
@@ -654,27 +773,24 @@ sfc_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 
 		/*
 		 * Here VLAN TCI is expected to be zero in case if no
-		 * DEV_TX_VLAN_OFFLOAD capability is advertised;
+		 * DEV_TX_OFFLOAD_VLAN_INSERT capability is advertised;
 		 * if the calling app ignores the absence of
-		 * DEV_TX_VLAN_OFFLOAD and pushes VLAN TCI, then
+		 * DEV_TX_OFFLOAD_VLAN_INSERT and pushes VLAN TCI, then
 		 * TX_ERROR will occur
 		 */
-		pkt_descs += sfc_tx_maybe_insert_tag(txq, m_seg, &pend);
+		pkt_descs += sfc_efx_tx_maybe_insert_tag(txq, m_seg, &pend);
 
 		if (m_seg->ol_flags & PKT_TX_TCP_SEG) {
 			/*
 			 * We expect correct 'pkt->l[2, 3, 4]_len' values
 			 * to be set correctly by the caller
 			 */
-			if (sfc_tso_do(txq, added, &m_seg, &in_off, &pend,
-				       &pkt_descs, &pkt_len) != 0) {
-				/* We may have reached this place for
-				 * one of the following reasons:
-				 *
-				 * 1) Packet header length is greater
-				 *    than SFC_TSOH_STD_LEN
-				 * 2) TCP header starts at more then
-				 *    208 bytes into the frame
+			if (sfc_efx_tso_do(txq, added, &m_seg, &in_off, &pend,
+					   &pkt_descs, &pkt_len) != 0) {
+				/* We may have reached this place if packet
+				 * header linearization is needed but the
+				 * header length is greater than
+				 * SFC_TSOH_STD_LEN
 				 *
 				 * We will deceive RTE saying that we have sent
 				 * the packet, but we will actually drop it.
@@ -685,6 +801,7 @@ sfc_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 				 * mbuf shouldn't be orphaned
 				 */
 				pend -= pkt_descs;
+				txq->hw_vlan_tci = hw_vlan_tci_prev;
 
 				rte_pktmbuf_free(*pktp);
 
@@ -704,7 +821,7 @@ sfc_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 			size_t			seg_len;
 
 			seg_len = m_seg->data_len;
-			next_frag = rte_mbuf_data_dma_addr(m_seg);
+			next_frag = rte_mbuf_data_iova(m_seg);
 
 			/*
 			 * If we've started TSO transaction few steps earlier,
@@ -749,15 +866,17 @@ sfc_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 			 * Try to reap (if we haven't yet).
 			 */
 			if (!reap_done) {
-				sfc_tx_reap(txq);
+				sfc_efx_tx_reap(txq);
 				reap_done = B_TRUE;
 				fill_level = added - txq->completed;
 				if (fill_level > hard_max_fill) {
 					pend -= pkt_descs;
+					txq->hw_vlan_tci = hw_vlan_tci_prev;
 					break;
 				}
 			} else {
 				pend -= pkt_descs;
+				txq->hw_vlan_tci = hw_vlan_tci_prev;
 				break;
 			}
 		}
@@ -778,9 +897,257 @@ sfc_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 
 #if SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE
 	if (!reap_done)
-		sfc_tx_reap(txq);
+		sfc_efx_tx_reap(txq);
 #endif
 
 done:
 	return pkts_sent;
 }
+
+const struct sfc_dp_tx *
+sfc_dp_tx_by_dp_txq(const struct sfc_dp_txq *dp_txq)
+{
+	const struct sfc_dp_queue *dpq = &dp_txq->dpq;
+	struct rte_eth_dev *eth_dev;
+	struct sfc_adapter_priv *sap;
+
+	SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
+	eth_dev = &rte_eth_devices[dpq->port_id];
+
+	sap = sfc_adapter_priv_by_eth_dev(eth_dev);
+
+	return sap->dp_tx;
+}
+
+struct sfc_txq_info *
+sfc_txq_info_by_dp_txq(const struct sfc_dp_txq *dp_txq)
+{
+	const struct sfc_dp_queue *dpq = &dp_txq->dpq;
+	struct rte_eth_dev *eth_dev;
+	struct sfc_adapter_shared *sas;
+
+	SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
+	eth_dev = &rte_eth_devices[dpq->port_id];
+
+	sas = sfc_adapter_shared_by_eth_dev(eth_dev);
+
+	SFC_ASSERT(dpq->queue_id < sas->txq_count);
+	return &sas->txq_info[dpq->queue_id];
+}
+
+struct sfc_txq *
+sfc_txq_by_dp_txq(const struct sfc_dp_txq *dp_txq)
+{
+	const struct sfc_dp_queue *dpq = &dp_txq->dpq;
+	struct rte_eth_dev *eth_dev;
+	struct sfc_adapter *sa;
+
+	SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
+	eth_dev = &rte_eth_devices[dpq->port_id];
+
+	sa = sfc_adapter_by_eth_dev(eth_dev);
+
+	SFC_ASSERT(dpq->queue_id < sfc_sa2shared(sa)->txq_count);
+	return &sa->txq_ctrl[dpq->queue_id];
+}
+
+static sfc_dp_tx_qsize_up_rings_t sfc_efx_tx_qsize_up_rings;
+static int
+sfc_efx_tx_qsize_up_rings(uint16_t nb_tx_desc,
+			  __rte_unused struct sfc_dp_tx_hw_limits *limits,
+			  unsigned int *txq_entries,
+			  unsigned int *evq_entries,
+			  unsigned int *txq_max_fill_level)
+{
+	*txq_entries = nb_tx_desc;
+	*evq_entries = nb_tx_desc;
+	*txq_max_fill_level = EFX_TXQ_LIMIT(*txq_entries);
+	return 0;
+}
+
+static sfc_dp_tx_qcreate_t sfc_efx_tx_qcreate;
+static int
+sfc_efx_tx_qcreate(uint16_t port_id, uint16_t queue_id,
+		   const struct rte_pci_addr *pci_addr,
+		   int socket_id,
+		   const struct sfc_dp_tx_qcreate_info *info,
+		   struct sfc_dp_txq **dp_txqp)
+{
+	struct sfc_efx_txq *txq;
+	struct sfc_txq *ctrl_txq;
+	int rc;
+
+	rc = ENOMEM;
+	txq = rte_zmalloc_socket("sfc-efx-txq", sizeof(*txq),
+				 RTE_CACHE_LINE_SIZE, socket_id);
+	if (txq == NULL)
+		goto fail_txq_alloc;
+
+	sfc_dp_queue_init(&txq->dp.dpq, port_id, queue_id, pci_addr);
+
+	rc = ENOMEM;
+	txq->pend_desc = rte_calloc_socket("sfc-efx-txq-pend-desc",
+					   EFX_TXQ_LIMIT(info->txq_entries),
+					   sizeof(*txq->pend_desc), 0,
+					   socket_id);
+	if (txq->pend_desc == NULL)
+		goto fail_pend_desc_alloc;
+
+	rc = ENOMEM;
+	txq->sw_ring = rte_calloc_socket("sfc-efx-txq-sw_ring",
+					 info->txq_entries,
+					 sizeof(*txq->sw_ring),
+					 RTE_CACHE_LINE_SIZE, socket_id);
+	if (txq->sw_ring == NULL)
+		goto fail_sw_ring_alloc;
+
+	ctrl_txq = sfc_txq_by_dp_txq(&txq->dp);
+	if (ctrl_txq->evq->sa->tso) {
+		rc = sfc_efx_tso_alloc_tsoh_objs(txq->sw_ring,
+						 info->txq_entries, socket_id);
+		if (rc != 0)
+			goto fail_alloc_tsoh_objs;
+	}
+
+	txq->evq = ctrl_txq->evq;
+	txq->ptr_mask = info->txq_entries - 1;
+	txq->max_fill_level = info->max_fill_level;
+	txq->free_thresh = info->free_thresh;
+	txq->dma_desc_size_max = info->dma_desc_size_max;
+
+	*dp_txqp = &txq->dp;
+	return 0;
+
+fail_alloc_tsoh_objs:
+	rte_free(txq->sw_ring);
+
+fail_sw_ring_alloc:
+	rte_free(txq->pend_desc);
+
+fail_pend_desc_alloc:
+	rte_free(txq);
+
+fail_txq_alloc:
+	return rc;
+}
+
+static sfc_dp_tx_qdestroy_t sfc_efx_tx_qdestroy;
+static void
+sfc_efx_tx_qdestroy(struct sfc_dp_txq *dp_txq)
+{
+	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
+
+	sfc_efx_tso_free_tsoh_objs(txq->sw_ring, txq->ptr_mask + 1);
+	rte_free(txq->sw_ring);
+	rte_free(txq->pend_desc);
+	rte_free(txq);
+}
+
+static sfc_dp_tx_qstart_t sfc_efx_tx_qstart;
+static int
+sfc_efx_tx_qstart(struct sfc_dp_txq *dp_txq,
+		  __rte_unused unsigned int evq_read_ptr,
+		  unsigned int txq_desc_index)
+{
+	/* libefx-based datapath is specific to libefx-based PMD */
+	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
+	struct sfc_txq *ctrl_txq = sfc_txq_by_dp_txq(dp_txq);
+
+	txq->common = ctrl_txq->common;
+
+	txq->pending = txq->completed = txq->added = txq_desc_index;
+	txq->hw_vlan_tci = 0;
+
+	txq->flags |= (SFC_EFX_TXQ_FLAG_STARTED | SFC_EFX_TXQ_FLAG_RUNNING);
+
+	return 0;
+}
+
+static sfc_dp_tx_qstop_t sfc_efx_tx_qstop;
+static void
+sfc_efx_tx_qstop(struct sfc_dp_txq *dp_txq,
+		 __rte_unused unsigned int *evq_read_ptr)
+{
+	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
+
+	txq->flags &= ~SFC_EFX_TXQ_FLAG_RUNNING;
+}
+
+static sfc_dp_tx_qreap_t sfc_efx_tx_qreap;
+static void
+sfc_efx_tx_qreap(struct sfc_dp_txq *dp_txq)
+{
+	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
+	unsigned int txds;
+
+	sfc_efx_tx_reap(txq);
+
+	for (txds = 0; txds <= txq->ptr_mask; txds++) {
+		if (txq->sw_ring[txds].mbuf != NULL) {
+			rte_pktmbuf_free(txq->sw_ring[txds].mbuf);
+			txq->sw_ring[txds].mbuf = NULL;
+		}
+	}
+
+	txq->flags &= ~SFC_EFX_TXQ_FLAG_STARTED;
+}
+
+static sfc_dp_tx_qdesc_status_t sfc_efx_tx_qdesc_status;
+static int
+sfc_efx_tx_qdesc_status(struct sfc_dp_txq *dp_txq, uint16_t offset)
+{
+	struct sfc_efx_txq *txq = sfc_efx_txq_by_dp_txq(dp_txq);
+
+	if (unlikely(offset > txq->ptr_mask))
+		return -EINVAL;
+
+	if (unlikely(offset >= txq->max_fill_level))
+		return RTE_ETH_TX_DESC_UNAVAIL;
+
+	/*
+	 * Poll EvQ to derive up-to-date 'txq->pending' figure;
+	 * it is required for the queue to be running, but the
+	 * check is omitted because API design assumes that it
+	 * is the duty of the caller to satisfy all conditions
+	 */
+	SFC_ASSERT((txq->flags & SFC_EFX_TXQ_FLAG_RUNNING) ==
+		   SFC_EFX_TXQ_FLAG_RUNNING);
+	sfc_ev_qpoll(txq->evq);
+
+	/*
+	 * Ring tail is 'txq->pending', and although descriptors
+	 * between 'txq->completed' and 'txq->pending' are still
+	 * in use by the driver, they should be reported as DONE
+	 */
+	if (unlikely(offset < (txq->added - txq->pending)))
+		return RTE_ETH_TX_DESC_FULL;
+
+	/*
+	 * There is no separate return value for unused descriptors;
+	 * the latter will be reported as DONE because genuine DONE
+	 * descriptors will be freed anyway in SW on the next burst
+	 */
+	return RTE_ETH_TX_DESC_DONE;
+}
+
+struct sfc_dp_tx sfc_efx_tx = {
+	.dp = {
+		.name		= SFC_KVARG_DATAPATH_EFX,
+		.type		= SFC_DP_TX,
+		.hw_fw_caps	= 0,
+	},
+	.features		= SFC_DP_TX_FEAT_VLAN_INSERT |
+				  SFC_DP_TX_FEAT_TSO |
+				  SFC_DP_TX_FEAT_MULTI_POOL |
+				  SFC_DP_TX_FEAT_REFCNT |
+				  SFC_DP_TX_FEAT_MULTI_SEG,
+	.qsize_up_rings		= sfc_efx_tx_qsize_up_rings,
+	.qcreate		= sfc_efx_tx_qcreate,
+	.qdestroy		= sfc_efx_tx_qdestroy,
+	.qstart			= sfc_efx_tx_qstart,
+	.qstop			= sfc_efx_tx_qstop,
+	.qreap			= sfc_efx_tx_qreap,
+	.qdesc_status		= sfc_efx_tx_qdesc_status,
+	.pkt_prepare		= sfc_efx_prepare_pkts,
+	.pkt_burst		= sfc_efx_xmit_pkts,
+};