+
+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
+ * (controlled by option descriptors), hence, if the tag of the
+ * packet to be sent is different from one remembered by the firmware,
+ * the function will update it
+ */
+static unsigned int
+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);
+
+ if (this_tag == txq->hw_vlan_tci)
+ return 0;
+
+ /*
+ * The expression inside SFC_ASSERT() is not desired to be checked in
+ * a non-debug build because it might be too expensive on the data path
+ */
+ SFC_ASSERT(efx_nic_cfg_get(txq->evq->sa->nic)->enc_hw_tx_insert_vlan_enabled);
+
+ efx_tx_qdesc_vlantci_create(txq->common, rte_cpu_to_be_16(this_tag),
+ *pend);
+ (*pend)++;
+ txq->hw_vlan_tci = this_tag;
+
+ return 1;
+}
+
+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], 0, SFC_TSOH_STD_LEN,
+ 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_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 = 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->flags & SFC_EFX_TXQ_FLAG_RUNNING) == 0))
+ goto done;
+
+ /*
+ * If insufficient space for a single packet is present,
+ * we should reap; otherwise, we shouldn't do that all the time
+ * to avoid latency increase
+ */
+ reap_done = (fill_level > soft_max_fill);
+
+ if (reap_done) {
+ sfc_efx_tx_reap(txq);
+ /*
+ * Recalculate fill level since 'txq->completed'
+ * might have changed on reap
+ */
+ fill_level = added - txq->completed;
+ }
+
+ 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;
+ size_t in_off = 0;
+
+ /*
+ * Here VLAN TCI is expected to be zero in case if no
+ * DEV_TX_OFFLOAD_VLAN_INSERT capability is advertised;
+ * if the calling app ignores the absence of
+ * DEV_TX_OFFLOAD_VLAN_INSERT and pushes VLAN TCI, then
+ * TX_ERROR will occur
+ */
+ 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_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.
+ * Hence, we should revert 'pend' to the
+ * previous state (in case we have added
+ * VLAN descriptor) and start processing
+ * another one packet. But the original
+ * mbuf shouldn't be orphaned
+ */
+ pend -= pkt_descs;
+ txq->hw_vlan_tci = hw_vlan_tci_prev;
+
+ rte_pktmbuf_free(*pktp);
+
+ continue;
+ }
+
+ /*
+ * We've only added 2 FATSOv2 option descriptors
+ * and 1 descriptor for the linearized packet header.
+ * The outstanding work will be done in the same manner
+ * as for the usual non-TSO path
+ */
+ }
+
+ 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_iova(m_seg);
+
+ /*
+ * If we've started TSO transaction few steps earlier,
+ * we'll skip packet header using an offset in the
+ * current segment (which has been set to the
+ * first one containing payload)
+ */
+ seg_len -= in_off;
+ next_frag += in_off;
+ in_off = 0;
+
+ do {
+ efsys_dma_addr_t frag_addr = next_frag;
+ size_t frag_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;
+
+ efx_tx_qdesc_dma_create(txq->common,
+ frag_addr, frag_len,
+ (pkt_len == 0),
+ pend++);
+
+ pkt_descs++;
+ } while (seg_len != 0);
+ }
+
+ added += pkt_descs;
+
+ fill_level += pkt_descs;
+ if (unlikely(fill_level > hard_max_fill)) {
+ /*
+ * Our estimation for maximum number of descriptors
+ * required to send a packet seems to be wrong.
+ * Try to reap (if we haven't yet).
+ */
+ if (!reap_done) {
+ 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;
+ }
+ }
+
+ /* Assign mbuf to the last used desc */
+ txq->sw_ring[(added - 1) & txq->ptr_mask].mbuf = *pktp;
+ }
+
+ if (likely(pkts_sent > 0)) {
+ rc = efx_tx_qdesc_post(txq->common, txq->pend_desc,
+ pend - &txq->pend_desc[0],
+ txq->completed, &txq->added);
+ SFC_ASSERT(rc == 0);
+
+ if (likely(pushed != txq->added))
+ efx_tx_qpush(txq->common, txq->added, pushed);
+ }
+
+#if SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE
+ if (!reap_done)
+ 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 = SFC_DP_HW_FW_CAP_TX_EFX,
+ },
+ .features = 0,
+ .dev_offload_capa = DEV_TX_OFFLOAD_VLAN_INSERT |
+ DEV_TX_OFFLOAD_MULTI_SEGS,
+ .queue_offload_capa = DEV_TX_OFFLOAD_IPV4_CKSUM |
+ DEV_TX_OFFLOAD_UDP_CKSUM |
+ DEV_TX_OFFLOAD_TCP_CKSUM |
+ DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM |
+ DEV_TX_OFFLOAD_TCP_TSO,
+ .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,
+};