net/sfc: avoid Tx queue setup failure if thresholds are set
[dpdk.git] / drivers / net / sfc / sfc_tx.c
index 3e64c0f..d1320f4 100644 (file)
@@ -1,5 +1,7 @@
 /*-
- * Copyright (c) 2016 Solarflare Communications Inc.
+ *   BSD LICENSE
+ *
+ * Copyright (c) 2016-2017 Solarflare Communications Inc.
  * All rights reserved.
  *
  * This software was jointly developed between OKTET Labs (under contract
@@ -33,6 +35,7 @@
 #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
@@ -76,17 +79,44 @@ sfc_tx_qcheck_conf(struct sfc_adapter *sa, uint16_t nb_tx_desc,
        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");
+       }
+
+       if (((flags & ETH_TXQ_FLAGS_NOMULTSEGS) == 0) &&
+           (~sa->dp_tx->features & SFC_DP_TX_FEAT_MULTI_SEG)) {
+               sfc_err(sa, "Multi-segment is not supported by %s datapath",
+                       sa->dp_tx->dp.name);
                rc = EINVAL;
        }
 
-       if (!encp->enc_hw_tx_insert_vlan_enabled &&
-           (flags & ETH_TXQ_FLAGS_NOVLANOFFL) == 0) {
-               sfc_err(sa, "VLAN offload is not supported");
+       if (((flags & ETH_TXQ_FLAGS_NOMULTMEMP) == 0) &&
+           (~sa->dp_tx->features & SFC_DP_TX_FEAT_MULTI_POOL)) {
+               sfc_err(sa, "multi-mempool is not supported by %s datapath",
+                       sa->dp_tx->dp.name);
+               rc = EINVAL;
+       }
+
+       if (((flags & ETH_TXQ_FLAGS_NOREFCOUNT) == 0) &&
+           (~sa->dp_tx->features & SFC_DP_TX_FEAT_REFCNT)) {
+               sfc_err(sa,
+                       "mbuf reference counters are neglected by %s datapath",
+                       sa->dp_tx->dp.name);
                rc = EINVAL;
        }
 
+       if ((flags & ETH_TXQ_FLAGS_NOVLANOFFL) == 0) {
+               if (!encp->enc_hw_tx_insert_vlan_enabled) {
+                       sfc_err(sa, "VLAN offload is not supported");
+                       rc = EINVAL;
+               } else if (~sa->dp_tx->features & SFC_DP_TX_FEAT_VLAN_INSERT) {
+                       sfc_err(sa,
+                               "VLAN offload is not supported by %s datapath",
+                               sa->dp_tx->dp.name);
+                       rc = EINVAL;
+               }
+       }
+
        if ((flags & ETH_TXQ_FLAGS_NOXSUMSCTP) == 0) {
                sfc_err(sa, "SCTP offload is not supported");
                rc = EINVAL;
@@ -109,39 +139,17 @@ sfc_tx_qflush_done(struct sfc_txq *txq)
        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;
-}
-
 int
 sfc_tx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
             uint16_t nb_tx_desc, unsigned int socket_id,
             const struct rte_eth_txconf *tx_conf)
 {
+       const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
        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;
 
        sfc_log_init(sa, "TxQ = %u", sw_index);
 
@@ -155,71 +163,64 @@ sfc_tx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
        SFC_ASSERT(nb_tx_desc <= sa->txq_max_entries);
        txq_info->entries = nb_tx_desc;
 
-       rc = sfc_ev_qinit(sa, evq_index, txq_info->entries, socket_id);
+       rc = sfc_ev_qinit(sa, SFC_EVQ_TYPE_TX, sw_index,
+                         txq_info->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_info->txq = txq;
+
+       txq->hw_index = sw_index;
+       txq->evq = evq;
+       txq->free_thresh =
+               (tx_conf->tx_free_thresh) ? tx_conf->tx_free_thresh :
+               SFC_TX_DEFAULT_FREE_THRESH;
+       txq->flags = tx_conf->txq_flags;
+
        rc = sfc_dma_alloc(sa, "txq", sw_index, EFX_TXQ_SIZE(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.free_thresh = txq->free_thresh;
+       info.flags = tx_conf->txq_flags;
+       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 = txq_info->entries;
+       info.evq_hw_ring = evq->mem.esm_base;
+       info.hw_index = txq->hw_index;
+       info.mem_bar = sa->mem_bar.esb_base;
+
+       rc = sa->dp_tx->qcreate(sa->eth_dev->data->port_id, sw_index,
+                               &RTE_ETH_DEV_TO_PCI(sa->eth_dev)->addr,
+                               socket_id, &info, &txq->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->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->hw_index = sw_index;
-       txq->flags = tx_conf->txq_flags;
-       txq->evq = evq;
 
-       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:
+       txq_info->txq = NULL;
        rte_free(txq);
 
 fail_txq_alloc:
-       sfc_ev_qfini(sa, evq_index);
+       sfc_ev_qfini(evq);
 
 fail_ev_qinit:
        txq_info->entries = 0;
@@ -244,14 +245,17 @@ sfc_tx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
        SFC_ASSERT(txq != NULL);
        SFC_ASSERT(txq->state == SFC_TXQ_INITIALIZED);
 
-       sfc_tso_free_tsoh_objs(txq->sw_ring, txq_info->entries);
+       sa->dp_tx->qdestroy(txq->dp);
+       txq->dp = NULL;
 
        txq_info->txq = NULL;
        txq_info->entries = 0;
 
-       rte_free(txq->sw_ring);
-       rte_free(txq->pend_desc);
        sfc_dma_free(sa, &txq->mem);
+
+       sfc_ev_qfini(txq->evq);
+       txq->evq = NULL;
+
        rte_free(txq);
 }
 
@@ -299,64 +303,108 @@ 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)
+{
+       int sw_index;
+
+       SFC_ASSERT(nb_tx_queues <= sa->txq_count);
+
+       sw_index = sa->txq_count;
+       while (--sw_index >= (int)nb_tx_queues) {
+               if (sa->txq_info[sw_index].txq != NULL)
+                       sfc_tx_qfini(sa, sw_index);
+       }
+
+       sa->txq_count = nb_tx_queues;
+}
+
 int
-sfc_tx_init(struct sfc_adapter *sa)
+sfc_tx_configure(struct sfc_adapter *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, sa->txq_count);
+
+       /*
+        * The datapath implementation assumes absence of boundary
+        * limits on Tx DMA descriptors. Addition of these checks on
+        * datapath would simply make the datapath slower.
+        */
+       if (encp->enc_tx_dma_desc_boundary != 0) {
+               rc = ENOTSUP;
+               goto fail_tx_dma_desc_boundary;
+       }
+
        rc = sfc_tx_check_mode(sa, &dev_conf->txmode);
        if (rc != 0)
                goto fail_check_mode;
 
-       sa->txq_count = sa->eth_dev->data->nb_tx_queues;
-
-       if (sa->tso)
-               sa->txq_count = MIN(sa->txq_count,
-                  efx_nic_cfg_get(sa->nic)->enc_fw_assisted_tso_v2_n_contexts /
-                  efx_nic_cfg_get(sa->nic)->enc_hw_pf_count);
+       if (nb_tx_queues == sa->txq_count)
+               goto done;
 
-       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;
+       if (sa->txq_info == NULL) {
+               sa->txq_info = rte_calloc_socket("sfc-txqs", nb_tx_queues,
+                                                sizeof(sa->txq_info[0]), 0,
+                                                sa->socket_id);
+               if (sa->txq_info == NULL)
+                       goto fail_txqs_alloc;
+       } else {
+               struct sfc_txq_info *new_txq_info;
+
+               if (nb_tx_queues < sa->txq_count)
+                       sfc_tx_fini_queues(sa, nb_tx_queues);
+
+               new_txq_info =
+                       rte_realloc(sa->txq_info,
+                                   nb_tx_queues * sizeof(sa->txq_info[0]), 0);
+               if (new_txq_info == NULL && nb_tx_queues > 0)
+                       goto fail_txqs_realloc;
+
+               sa->txq_info = new_txq_info;
+               if (nb_tx_queues > sa->txq_count)
+                       memset(&sa->txq_info[sa->txq_count], 0,
+                              (nb_tx_queues - sa->txq_count) *
+                              sizeof(sa->txq_info[0]));
+       }
 
-       for (sw_index = 0; sw_index < sa->txq_count; ++sw_index) {
-               rc = sfc_tx_qinit_info(sa, sw_index);
+       while (sa->txq_count < nb_tx_queues) {
+               rc = sfc_tx_qinit_info(sa, sa->txq_count);
                if (rc != 0)
                        goto fail_tx_qinit_info;
+
+               sa->txq_count++;
        }
 
+done:
        return 0;
 
 fail_tx_qinit_info:
-       rte_free(sa->txq_info);
-       sa->txq_info = NULL;
-
+fail_txqs_realloc:
 fail_txqs_alloc:
-       sa->txq_count = 0;
+       sfc_tx_close(sa);
 
 fail_check_mode:
+fail_tx_dma_desc_boundary:
        sfc_log_init(sa, "failed (rc = %d)", rc);
        return rc;
 }
 
 void
-sfc_tx_fini(struct sfc_adapter *sa)
+sfc_tx_close(struct sfc_adapter *sa)
 {
-       int sw_index;
-
-       sw_index = sa->txq_count;
-       while (--sw_index >= 0) {
-               if (sa->txq_info[sw_index].txq != NULL)
-                       sfc_tx_qfini(sa, sw_index);
-       }
+       sfc_tx_fini_queues(sa, 0);
 
        rte_free(sa->txq_info);
        sa->txq_info = NULL;
-       sa->txq_count = 0;
 }
 
 int
@@ -381,7 +429,7 @@ sfc_tx_qstart(struct sfc_adapter *sa, unsigned int 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;
 
@@ -410,12 +458,13 @@ 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->state |= SFC_TXQ_STARTED;
+
+       rc = sa->dp_tx->qstart(txq->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')
@@ -425,8 +474,12 @@ sfc_tx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 
        return 0;
 
+fail_dp_qstart:
+       txq->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;
@@ -440,7 +493,7 @@ sfc_tx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
        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);
 
@@ -454,7 +507,7 @@ sfc_tx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
 
        SFC_ASSERT(txq->state & SFC_TXQ_STARTED);
 
-       txq->state &= ~SFC_TXQ_RUNNING;
+       sa->dp_tx->qstop(txq->dp, &txq->evq->read_ptr);
 
        /*
         * Retry TX queue flushing in case of flush failed or
@@ -464,8 +517,10 @@ sfc_tx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
             ((txq->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->state |= (rc == EALREADY) ?
+                               SFC_TXQ_FLUSHED : SFC_TXQ_FLUSH_FAILED;
                        break;
                }
 
@@ -489,20 +544,13 @@ sfc_tx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
                        sfc_info(sa, "TxQ %u flushed", sw_index);
        }
 
-       sfc_tx_reap(txq);
-
-       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;
-               }
-       }
+       sa->dp_tx->qreap(txq->dp);
 
        txq->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')
@@ -568,6 +616,28 @@ sfc_tx_stop(struct sfc_adapter *sa)
        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
@@ -576,8 +646,8 @@ 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);
@@ -599,10 +669,11 @@ 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_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;
@@ -614,7 +685,7 @@ sfc_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
        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;
 
        /*
@@ -625,7 +696,7 @@ 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
@@ -648,16 +719,15 @@ sfc_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
                 * DEV_TX_VLAN_OFFLOAD 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);
 
-#ifdef RTE_LIBRTE_SFC_EFX_TSO
                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) {
+                       if (sfc_efx_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:
                                 *
@@ -688,14 +758,13 @@ sfc_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
                         * as for the usual non-TSO path
                         */
                }
-#endif /* RTE_LIBRTE_SFC_EFX_TSO */
 
                for (; m_seg != NULL; m_seg = m_seg->next) {
                        efsys_dma_addr_t        next_frag;
                        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,
@@ -711,9 +780,13 @@ sfc_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
                                efsys_dma_addr_t        frag_addr = next_frag;
                                size_t                  frag_len;
 
-                               next_frag = RTE_ALIGN(frag_addr + 1,
-                                                     SFC_TX_SEG_BOUNDARY);
-                               frag_len = MIN(next_frag - frag_addr, seg_len);
+                               /*
+                                * It is assumed here that there is no
+                                * limitation on address boundary
+                                * crossing by DMA descriptor.
+                                */
+                               frag_len = MIN(seg_len, txq->dma_desc_size_max);
+                               next_frag += frag_len;
                                seg_len -= frag_len;
                                pkt_len -= frag_len;
 
@@ -736,7 +809,7 @@ 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) {
@@ -765,9 +838,213 @@ 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;
 }
+
+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;
+       struct sfc_txq *txq;
+
+       SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
+       eth_dev = &rte_eth_devices[dpq->port_id];
+
+       sa = eth_dev->data->dev_private;
+
+       SFC_ASSERT(dpq->queue_id < sa->txq_count);
+       txq = sa->txq_info[dpq->queue_id].txq;
+
+       SFC_ASSERT(txq != NULL);
+       return txq;
+}
+
+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->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 >= EFX_TXQ_LIMIT(txq->ptr_mask + 1)))
+               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,
+       .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_burst              = sfc_efx_xmit_pkts,
+};