common/dpaax: fallback to check separate memory node for VM
[dpdk.git] / drivers / net / sfc / sfc_tx.c
index 94baf45..7d0e18a 100644 (file)
  */
 #define SFC_TX_QFLUSH_POLL_ATTEMPTS    (2000)
 
-uint64_t
-sfc_tx_get_dev_offload_caps(struct sfc_adapter *sa)
+static uint64_t
+sfc_tx_get_offload_mask(struct sfc_adapter *sa)
 {
        const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
-       uint64_t caps = 0;
+       uint64_t no_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 (!encp->enc_hw_tx_insert_vlan_enabled)
+               no_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 (!encp->enc_tunnel_encapsulations_supported)
+               no_caps |= DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM;
 
-       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;
+       if (!sa->tso)
+               no_caps |= DEV_TX_OFFLOAD_TCP_TSO;
 
-       return caps;
+       if (!sa->tso_encap)
+               no_caps |= (DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
+                           DEV_TX_OFFLOAD_GENEVE_TNL_TSO);
+
+       return ~no_caps;
 }
 
 uint64_t
-sfc_tx_get_queue_offload_caps(struct sfc_adapter *sa)
+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;
-
-       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 sa->priv.dp_tx->dev_offload_capa & sfc_tx_get_offload_mask(sa);
+}
 
-       return caps;
+uint64_t
+sfc_tx_get_queue_offload_caps(struct sfc_adapter *sa)
+{
+       return sa->priv.dp_tx->queue_offload_capa & sfc_tx_get_offload_mask(sa);
 }
 
 static int
@@ -131,14 +126,20 @@ sfc_tx_qinit(struct sfc_adapter *sa, unsigned int 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 = sa->priv.dp_tx->qsize_up_rings(nb_tx_desc, &txq_entries,
-                                           &evq_entries, &txq_max_fill_level);
+       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 >= EFX_TXQ_MINNDESCS);
+       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);
@@ -149,8 +150,8 @@ sfc_tx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
        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];
 
        txq_info->entries = txq_entries;
 
@@ -167,7 +168,8 @@ sfc_tx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
                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;
@@ -224,10 +226,10 @@ 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);
+       SFC_ASSERT(sw_index < sfc_sa2shared(sa)->txq_count);
        sa->eth_dev->data->tx_queues[sw_index] = NULL;
 
-       txq_info = &sa->txq_info[sw_index];
+       txq_info = &sfc_sa2shared(sa)->txq_info[sw_index];
 
        SFC_ASSERT(txq_info->state == SFC_TXQ_INITIALIZED);
 
@@ -296,29 +298,31 @@ sfc_tx_check_mode(struct sfc_adapter *sa, const struct rte_eth_txmode *txmode)
 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 <= sa->txq_count);
+       SFC_ASSERT(nb_tx_queues <= sas->txq_count);
 
-       sw_index = sa->txq_count;
+       sw_index = sas->txq_count;
        while (--sw_index >= (int)nb_tx_queues) {
-               if (sa->txq_info[sw_index].state & SFC_TXQ_INITIALIZED)
+               if (sas->txq_info[sw_index].state & SFC_TXQ_INITIALIZED)
                        sfc_tx_qfini(sa, sw_index);
        }
 
-       sa->txq_count = nb_tx_queues;
+       sas->txq_count = nb_tx_queues;
 }
 
 int
 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;
        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);
+                    nb_tx_queues, sas->txq_count);
 
        /*
         * The datapath implementation assumes absence of boundary
@@ -334,14 +338,14 @@ sfc_tx_configure(struct sfc_adapter *sa)
        if (rc != 0)
                goto fail_check_mode;
 
-       if (nb_tx_queues == sa->txq_count)
+       if (nb_tx_queues == sas->txq_count)
                goto done;
 
-       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)
+       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;
 
                /*
@@ -356,12 +360,12 @@ sfc_tx_configure(struct sfc_adapter *sa)
                struct sfc_txq_info *new_txq_info;
                struct sfc_txq *new_txq_ctrl;
 
-               if (nb_tx_queues < sa->txq_count)
+               if (nb_tx_queues < sas->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);
+                       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;
 
@@ -370,24 +374,24 @@ sfc_tx_configure(struct sfc_adapter *sa)
                if (new_txq_ctrl == NULL && nb_tx_queues > 0)
                        goto fail_txqs_ctrl_realloc;
 
-               sa->txq_info = new_txq_info;
+               sas->txq_info = new_txq_info;
                sa->txq_ctrl = new_txq_ctrl;
-               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]));
-                       memset(&sa->txq_ctrl[sa->txq_count], 0,
-                              (nb_tx_queues - sa->txq_count) *
+               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]));
                }
        }
 
-       while (sa->txq_count < nb_tx_queues) {
-               rc = sfc_tx_qinit_info(sa, sa->txq_count);
+       while (sas->txq_count < nb_tx_queues) {
+               rc = sfc_tx_qinit_info(sa, sas->txq_count);
                if (rc != 0)
                        goto fail_tx_qinit_info;
 
-               sa->txq_count++;
+               sas->txq_count++;
        }
 
 done:
@@ -414,13 +418,14 @@ sfc_tx_close(struct sfc_adapter *sa)
        free(sa->txq_ctrl);
        sa->txq_ctrl = NULL;
 
-       rte_free(sa->txq_info);
-       sa->txq_info = NULL;
+       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;
@@ -433,8 +438,8 @@ sfc_tx_qstart(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 < sas->txq_count);
+       txq_info = &sas->txq_info[sw_index];
 
        SFC_ASSERT(txq_info->state == SFC_TXQ_INITIALIZED);
 
@@ -459,7 +464,9 @@ sfc_tx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
                        flags |= EFX_TXQ_CKSUM_INNER_TCPUDP;
        }
 
-       if (txq_info->offloads & DEV_TX_OFFLOAD_TCP_TSO)
+       if (txq_info->offloads & (DEV_TX_OFFLOAD_TCP_TSO |
+                                 DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
+                                 DEV_TX_OFFLOAD_GENEVE_TNL_TSO))
                flags |= EFX_TXQ_FATSOV2;
 
        rc = efx_tx_qcreate(sa->nic, txq->hw_index, 0, &txq->mem,
@@ -503,6 +510,7 @@ 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;
@@ -512,8 +520,8 @@ sfc_tx_qstop(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 < sas->txq_count);
+       txq_info = &sas->txq_info[sw_index];
 
        if (txq_info->state == SFC_TXQ_INITIALIZED)
                return;
@@ -576,26 +584,34 @@ 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);
+       const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
        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) {
+               if (!encp->enc_fw_assisted_tso_v2_enabled) {
                        sfc_warn(sa, "TSO support was unable to be restored");
                        sa->tso = B_FALSE;
+                       sa->tso_encap = B_FALSE;
                }
        }
 
+       if (sa->tso_encap && !encp->enc_fw_assisted_tso_v2_encap_enabled) {
+               sfc_warn(sa, "Encapsulated TSO support was unable to be restored");
+               sa->tso_encap = B_FALSE;
+       }
+
        rc = efx_tx_init(sa->nic);
        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].state == SFC_TXQ_INITIALIZED &&
-                   (!(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;
@@ -618,13 +634,14 @@ 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].state & SFC_TXQ_STARTED)
+               if (sas->txq_info[sw_index].state & SFC_TXQ_STARTED)
                        sfc_tx_qstop(sa, sw_index);
        }
 
@@ -684,6 +701,36 @@ sfc_efx_tx_maybe_insert_tag(struct sfc_efx_txq *txq, struct rte_mbuf *m,
        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],
+                               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)
 {
@@ -744,13 +791,10 @@ sfc_efx_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
                         */
                        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:
-                                *
-                                * 1) Packet header length is greater
-                                *    than SFC_TSOH_STD_LEN
-                                * 2) TCP header starts at more then
-                                *    208 bytes into the frame
+                               /* 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.
@@ -884,15 +928,15 @@ 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 *sa;
+       struct sfc_adapter_shared *sas;
 
        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;
+       sas = sfc_adapter_shared_by_eth_dev(eth_dev);
 
-       SFC_ASSERT(dpq->queue_id < sa->txq_count);
-       return &sa->txq_info[dpq->queue_id];
+       SFC_ASSERT(dpq->queue_id < sas->txq_count);
+       return &sas->txq_info[dpq->queue_id];
 }
 
 struct sfc_txq *
@@ -905,15 +949,16 @@ sfc_txq_by_dp_txq(const struct sfc_dp_txq *dp_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;
+       sa = sfc_adapter_by_eth_dev(eth_dev);
 
-       SFC_ASSERT(dpq->queue_id < sa->txq_count);
+       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)
@@ -1095,11 +1140,14 @@ struct sfc_dp_tx sfc_efx_tx = {
                .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,
+       .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,
@@ -1107,5 +1155,6 @@ struct sfc_dp_tx sfc_efx_tx = {
        .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,
 };