net/sfc: avoid Rx queue setup failure if thresholds are set
[dpdk.git] / drivers / net / sfc / sfc_rx.c
index bdaf21b..7816393 100644 (file)
@@ -122,13 +122,13 @@ sfc_efx_rx_qrefill(struct sfc_efx_rxq *rxq)
                        rxd = &rxq->sw_desc[id];
                        rxd->mbuf = m;
 
-                       rte_mbuf_refcnt_set(m, 1);
+                       SFC_ASSERT(rte_mbuf_refcnt_read(m) == 1);
                        m->data_off = RTE_PKTMBUF_HEADROOM;
-                       m->next = NULL;
-                       m->nb_segs = 1;
+                       SFC_ASSERT(m->next == NULL);
+                       SFC_ASSERT(m->nb_segs == 1);
                        m->port = port_id;
 
-                       addr[i] = rte_pktmbuf_mtophys(m);
+                       addr[i] = rte_pktmbuf_iova(m);
                }
 
                efx_rx_qpost(rxq->common, addr, rxq->buf_size,
@@ -207,11 +207,11 @@ sfc_efx_supported_ptypes_get(void)
        return ptypes;
 }
 
+#if EFSYS_OPT_RX_SCALE
 static void
 sfc_efx_rx_set_rss_hash(struct sfc_efx_rxq *rxq, unsigned int flags,
                        struct rte_mbuf *m)
 {
-#if EFSYS_OPT_RX_SCALE
        uint8_t *mbuf_data;
 
 
@@ -227,8 +227,15 @@ sfc_efx_rx_set_rss_hash(struct sfc_efx_rxq *rxq, unsigned int flags,
 
                m->ol_flags |= PKT_RX_RSS_HASH;
        }
-#endif
 }
+#else
+static void
+sfc_efx_rx_set_rss_hash(__rte_unused struct sfc_efx_rxq *rxq,
+                       __rte_unused unsigned int flags,
+                       __rte_unused struct rte_mbuf *m)
+{
+}
+#endif
 
 static uint16_t
 sfc_efx_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
@@ -292,7 +299,7 @@ sfc_efx_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
                if (desc_flags & EFX_PKT_CONT) {
                        /* The packet is scattered, more fragments to come */
                        scatter_pkt = m;
-                       /* Futher fragments have no prefix */
+                       /* Further fragments have no prefix */
                        prefix_size = 0;
                        continue;
                }
@@ -349,6 +356,43 @@ sfc_efx_rx_qdesc_npending(struct sfc_dp_rxq *dp_rxq)
        return rxq->pending - rxq->completed;
 }
 
+static sfc_dp_rx_qdesc_status_t sfc_efx_rx_qdesc_status;
+static int
+sfc_efx_rx_qdesc_status(struct sfc_dp_rxq *dp_rxq, uint16_t offset)
+{
+       struct sfc_efx_rxq *rxq = sfc_efx_rxq_by_dp_rxq(dp_rxq);
+
+       if (unlikely(offset > rxq->ptr_mask))
+               return -EINVAL;
+
+       /*
+        * Poll EvQ to derive up-to-date 'rxq->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((rxq->flags & SFC_EFX_RXQ_FLAG_RUNNING) ==
+                  SFC_EFX_RXQ_FLAG_RUNNING);
+       sfc_ev_qpoll(rxq->evq);
+
+       /*
+        * There is a handful of reserved entries in the ring,
+        * but an explicit check whether the offset points to
+        * a reserved entry is neglected since the two checks
+        * below rely on the figures which take the HW limits
+        * into account and thus if an entry is reserved, the
+        * checks will fail and UNAVAIL code will be returned
+        */
+
+       if (offset < (rxq->pending - rxq->completed))
+               return RTE_ETH_RX_DESC_DONE;
+
+       if (offset < (rxq->added - rxq->completed))
+               return RTE_ETH_RX_DESC_AVAIL;
+
+       return RTE_ETH_RX_DESC_UNAVAIL;
+}
+
 struct sfc_rxq *
 sfc_rxq_by_dp_rxq(const struct sfc_dp_rxq *dp_rxq)
 {
@@ -498,6 +542,7 @@ struct sfc_dp_rx sfc_efx_rx = {
        .qpurge                 = sfc_efx_rx_qpurge,
        .supported_ptypes_get   = sfc_efx_supported_ptypes_get,
        .qdesc_npending         = sfc_efx_rx_qdesc_npending,
+       .qdesc_status           = sfc_efx_rx_qdesc_status,
        .pkt_burst              = sfc_efx_recv_pkts,
 };
 
@@ -529,6 +574,7 @@ sfc_rx_qflush(struct sfc_adapter *sa, unsigned int sw_index)
        struct sfc_rxq *rxq;
        unsigned int retry_count;
        unsigned int wait_count;
+       int rc;
 
        rxq = sa->rxq_info[sw_index].rxq;
        SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
@@ -541,8 +587,10 @@ sfc_rx_qflush(struct sfc_adapter *sa, unsigned int sw_index)
             ((rxq->state & SFC_RXQ_FLUSHED) == 0) &&
             (retry_count < SFC_RX_QFLUSH_ATTEMPTS);
             ++retry_count) {
-               if (efx_rx_qflush(rxq->common) != 0) {
-                       rxq->state |= SFC_RXQ_FLUSH_FAILED;
+               rc = efx_rx_qflush(rxq->common);
+               if (rc != 0) {
+                       rxq->state |= (rc == EALREADY) ?
+                               SFC_RXQ_FLUSHED : SFC_RXQ_FLUSH_FAILED;
                        break;
                }
                rxq->state &= ~SFC_RXQ_FLUSH_FAILED;
@@ -577,7 +625,7 @@ sfc_rx_qflush(struct sfc_adapter *sa, unsigned int sw_index)
 static int
 sfc_rx_default_rxq_set_filter(struct sfc_adapter *sa, struct sfc_rxq *rxq)
 {
-       boolean_t rss = (sa->rss_channels > 1) ? B_TRUE : B_FALSE;
+       boolean_t rss = (sa->rss_channels > 0) ? B_TRUE : B_FALSE;
        struct sfc_port *port = &sa->port;
        int rc;
 
@@ -627,6 +675,7 @@ retry:
 int
 sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 {
+       struct sfc_port *port = &sa->port;
        struct sfc_rxq_info *rxq_info;
        struct sfc_rxq *rxq;
        struct sfc_evq *evq;
@@ -642,7 +691,7 @@ sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 
        evq = rxq->evq;
 
-       rc = sfc_ev_qstart(sa, evq->evq_index);
+       rc = sfc_ev_qstart(evq, sfc_evq_index_by_rxq_sw_index(sa, sw_index));
        if (rc != 0)
                goto fail_ev_qstart;
 
@@ -661,7 +710,7 @@ sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 
        rxq->state |= SFC_RXQ_STARTED;
 
-       if (sw_index == 0) {
+       if ((sw_index == 0) && !port->isolated) {
                rc = sfc_rx_default_rxq_set_filter(sa, rxq);
                if (rc != 0)
                        goto fail_mac_filter_default_rxq_set;
@@ -680,7 +729,7 @@ fail_dp_qstart:
        sfc_rx_qflush(sa, sw_index);
 
 fail_rx_qcreate:
-       sfc_ev_qstop(sa, evq->evq_index);
+       sfc_ev_qstop(evq);
 
 fail_ev_qstart:
        return rc;
@@ -718,7 +767,7 @@ sfc_rx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
 
        efx_rx_qdestroy(rxq->common);
 
-       sfc_ev_qstop(sa, rxq->evq->evq_index);
+       sfc_ev_qstop(rxq->evq);
 }
 
 static int
@@ -731,9 +780,8 @@ sfc_rx_qcheck_conf(struct sfc_adapter *sa, uint16_t nb_rx_desc,
        if (rx_conf->rx_thresh.pthresh != 0 ||
            rx_conf->rx_thresh.hthresh != 0 ||
            rx_conf->rx_thresh.wthresh != 0) {
-               sfc_err(sa,
+               sfc_warn(sa,
                        "RxQ prefetch/host/writeback thresholds are not supported");
-               rc = EINVAL;
        }
 
        if (rx_conf->rx_free_thresh > rx_free_thresh_max) {
@@ -861,7 +909,6 @@ sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
        int rc;
        uint16_t buf_size;
        struct sfc_rxq_info *rxq_info;
-       unsigned int evq_index;
        struct sfc_evq *evq;
        struct sfc_rxq *rxq;
        struct sfc_dp_rx_qcreate_info info;
@@ -899,15 +946,11 @@ sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
                sa->eth_dev->data->dev_conf.rxmode.enable_scatter ?
                EFX_RXQ_TYPE_SCATTER : EFX_RXQ_TYPE_DEFAULT;
 
-       evq_index = sfc_evq_index_by_rxq_sw_index(sa, sw_index);
-
-       rc = sfc_ev_qinit(sa, evq_index, SFC_EVQ_TYPE_RX, sw_index,
-                         rxq_info->entries, socket_id);
+       rc = sfc_ev_qinit(sa, SFC_EVQ_TYPE_RX, sw_index,
+                         rxq_info->entries, socket_id, &evq);
        if (rc != 0)
                goto fail_ev_qinit;
 
-       evq = sa->evq_info[evq_index].evq;
-
        rc = ENOMEM;
        rxq = rte_zmalloc_socket("sfc-rxq", sizeof(*rxq), RTE_CACHE_LINE_SIZE,
                                 socket_id);
@@ -935,7 +978,7 @@ sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
        info.prefix_size = encp->enc_rx_prefix_size;
 
 #if EFSYS_OPT_RX_SCALE
-       if (sa->hash_support == EFX_RX_HASH_AVAILABLE)
+       if (sa->hash_support == EFX_RX_HASH_AVAILABLE && sa->rss_channels > 0)
                info.flags |= SFC_RXQ_FLAG_RSS_HASH;
 #endif
 
@@ -947,7 +990,7 @@ sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
        info.mem_bar = sa->mem_bar.esb_base;
 
        rc = sa->dp_rx->qcreate(sa->eth_dev->data->port_id, sw_index,
-                               &SFC_DEV_TO_PCI(sa->eth_dev)->addr,
+                               &RTE_ETH_DEV_TO_PCI(sa->eth_dev)->addr,
                                socket_id, &info, &rxq->dp);
        if (rc != 0)
                goto fail_dp_rx_qcreate;
@@ -968,7 +1011,7 @@ fail_dma_alloc:
        rte_free(rxq);
 
 fail_rxq_alloc:
-       sfc_ev_qfini(sa, evq_index);
+       sfc_ev_qfini(evq);
 
 fail_ev_qinit:
        rxq_info->entries = 0;
@@ -998,9 +1041,11 @@ sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
        rxq_info->entries = 0;
 
        sfc_dma_free(sa, &rxq->mem);
-       rte_free(rxq);
 
-       sfc_ev_qfini(sa, sfc_evq_index_by_rxq_sw_index(sa, sw_index));
+       sfc_ev_qfini(rxq->evq);
+       rxq->evq = NULL;
+
+       rte_free(rxq);
 }
 
 #if EFSYS_OPT_RX_SCALE
@@ -1049,31 +1094,39 @@ sfc_efx_to_rte_hash_type(efx_rx_hash_type_t efx_hash_types)
 }
 #endif
 
+#if EFSYS_OPT_RX_SCALE
 static int
 sfc_rx_rss_config(struct sfc_adapter *sa)
 {
        int rc = 0;
 
-#if EFSYS_OPT_RX_SCALE
-       if (sa->rss_channels > 1) {
-               rc = efx_rx_scale_mode_set(sa->nic, EFX_RX_HASHALG_TOEPLITZ,
+       if (sa->rss_channels > 0) {
+               rc = efx_rx_scale_mode_set(sa->nic, EFX_RSS_CONTEXT_DEFAULT,
+                                          EFX_RX_HASHALG_TOEPLITZ,
                                           sa->rss_hash_types, B_TRUE);
                if (rc != 0)
                        goto finish;
 
-               rc = efx_rx_scale_key_set(sa->nic, sa->rss_key,
+               rc = efx_rx_scale_key_set(sa->nic, EFX_RSS_CONTEXT_DEFAULT,
+                                         sa->rss_key,
                                          sizeof(sa->rss_key));
                if (rc != 0)
                        goto finish;
 
-               rc = efx_rx_scale_tbl_set(sa->nic, sa->rss_tbl,
-                                         sizeof(sa->rss_tbl));
+               rc = efx_rx_scale_tbl_set(sa->nic, EFX_RSS_CONTEXT_DEFAULT,
+                                         sa->rss_tbl, RTE_DIM(sa->rss_tbl));
        }
 
 finish:
-#endif
        return rc;
 }
+#else
+static int
+sfc_rx_rss_config(__rte_unused struct sfc_adapter *sa)
+{
+       return 0;
+}
+#endif
 
 int
 sfc_rx_start(struct sfc_adapter *sa)
@@ -1209,58 +1262,106 @@ sfc_rx_check_mode(struct sfc_adapter *sa, struct rte_eth_rxmode *rxmode)
        return rc;
 }
 
+/**
+ * Destroy excess queues that are no longer needed after reconfiguration
+ * or complete close.
+ */
+static void
+sfc_rx_fini_queues(struct sfc_adapter *sa, unsigned int nb_rx_queues)
+{
+       int sw_index;
+
+       SFC_ASSERT(nb_rx_queues <= sa->rxq_count);
+
+       sw_index = sa->rxq_count;
+       while (--sw_index >= (int)nb_rx_queues) {
+               if (sa->rxq_info[sw_index].rxq != NULL)
+                       sfc_rx_qfini(sa, sw_index);
+       }
+
+       sa->rxq_count = nb_rx_queues;
+}
+
 /**
  * Initialize Rx subsystem.
  *
- * Called at device configuration stage when number of receive queues is
+ * Called at device (re)configuration stage when number of receive queues is
  * specified together with other device level receive configuration.
  *
  * It should be used to allocate NUMA-unaware resources.
  */
 int
-sfc_rx_init(struct sfc_adapter *sa)
+sfc_rx_configure(struct sfc_adapter *sa)
 {
        struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
-       unsigned int sw_index;
+       const unsigned int nb_rx_queues = sa->eth_dev->data->nb_rx_queues;
        int rc;
 
+       sfc_log_init(sa, "nb_rx_queues=%u (old %u)",
+                    nb_rx_queues, sa->rxq_count);
+
        rc = sfc_rx_check_mode(sa, &dev_conf->rxmode);
        if (rc != 0)
                goto fail_check_mode;
 
-       sa->rxq_count = sa->eth_dev->data->nb_rx_queues;
+       if (nb_rx_queues == sa->rxq_count)
+               goto done;
 
-       rc = ENOMEM;
-       sa->rxq_info = rte_calloc_socket("sfc-rxqs", sa->rxq_count,
-                                        sizeof(struct sfc_rxq_info), 0,
-                                        sa->socket_id);
-       if (sa->rxq_info == NULL)
-               goto fail_rxqs_alloc;
+       if (sa->rxq_info == NULL) {
+               rc = ENOMEM;
+               sa->rxq_info = rte_calloc_socket("sfc-rxqs", nb_rx_queues,
+                                                sizeof(sa->rxq_info[0]), 0,
+                                                sa->socket_id);
+               if (sa->rxq_info == NULL)
+                       goto fail_rxqs_alloc;
+       } else {
+               struct sfc_rxq_info *new_rxq_info;
+
+               if (nb_rx_queues < sa->rxq_count)
+                       sfc_rx_fini_queues(sa, nb_rx_queues);
+
+               rc = ENOMEM;
+               new_rxq_info =
+                       rte_realloc(sa->rxq_info,
+                                   nb_rx_queues * sizeof(sa->rxq_info[0]), 0);
+               if (new_rxq_info == NULL && nb_rx_queues > 0)
+                       goto fail_rxqs_realloc;
+
+               sa->rxq_info = new_rxq_info;
+               if (nb_rx_queues > sa->rxq_count)
+                       memset(&sa->rxq_info[sa->rxq_count], 0,
+                              (nb_rx_queues - sa->rxq_count) *
+                              sizeof(sa->rxq_info[0]));
+       }
 
-       for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
-               rc = sfc_rx_qinit_info(sa, sw_index);
+       while (sa->rxq_count < nb_rx_queues) {
+               rc = sfc_rx_qinit_info(sa, sa->rxq_count);
                if (rc != 0)
                        goto fail_rx_qinit_info;
+
+               sa->rxq_count++;
        }
 
 #if EFSYS_OPT_RX_SCALE
        sa->rss_channels = (dev_conf->rxmode.mq_mode == ETH_MQ_RX_RSS) ?
-                          MIN(sa->rxq_count, EFX_MAXRSS) : 1;
+                          MIN(sa->rxq_count, EFX_MAXRSS) : 0;
+
+       if (sa->rss_channels > 0) {
+               unsigned int sw_index;
 
-       if (sa->rss_channels > 1) {
                for (sw_index = 0; sw_index < EFX_RSS_TBL_SIZE; ++sw_index)
                        sa->rss_tbl[sw_index] = sw_index % sa->rss_channels;
        }
 #endif
 
+done:
        return 0;
 
 fail_rx_qinit_info:
-       rte_free(sa->rxq_info);
-       sa->rxq_info = NULL;
-
+fail_rxqs_realloc:
 fail_rxqs_alloc:
-       sa->rxq_count = 0;
+       sfc_rx_close(sa);
+
 fail_check_mode:
        sfc_log_init(sa, "failed %d", rc);
        return rc;
@@ -1269,21 +1370,15 @@ fail_check_mode:
 /**
  * Shutdown Rx subsystem.
  *
- * Called at device close stage, for example, before device
- * reconfiguration or shutdown.
+ * Called at device close stage, for example, before device shutdown.
  */
 void
-sfc_rx_fini(struct sfc_adapter *sa)
+sfc_rx_close(struct sfc_adapter *sa)
 {
-       unsigned int sw_index;
+       sfc_rx_fini_queues(sa, 0);
 
-       sw_index = sa->rxq_count;
-       while (sw_index-- > 0) {
-               if (sa->rxq_info[sw_index].rxq != NULL)
-                       sfc_rx_qfini(sa, sw_index);
-       }
+       sa->rss_channels = 0;
 
        rte_free(sa->rxq_info);
        sa->rxq_info = NULL;
-       sa->rxq_count = 0;
 }