IB/hfi1: Add rx functions for dummy netdev
authorGrzegorz Andrejczuk <grzegorz.andrejczuk@intel.com>
Mon, 11 May 2020 16:06:43 +0000 (12:06 -0400)
committerJason Gunthorpe <jgg@mellanox.com>
Thu, 21 May 2020 14:23:56 +0000 (11:23 -0300)
This patch adds the rx functions for the dummy netdev:
- Functions to allocate/free the dummy netdev.
- Functions to allocate/free receiving contexts for the netdev.
- Functions to initialize/de-initialize the receive queue.
- Functions to enable/disable the receive queue.

Link: https://lore.kernel.org/r/20200511160643.173205.75087.stgit@awfm-01.aw.intel.com
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Sadanand Warrier <sadanand.warrier@intel.com>
Signed-off-by: Grzegorz Andrejczuk <grzegorz.andrejczuk@intel.com>
Signed-off-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
drivers/infiniband/hw/hfi1/ipoib.h
drivers/infiniband/hw/hfi1/ipoib_main.c
drivers/infiniband/hw/hfi1/ipoib_rx.c
drivers/infiniband/hw/hfi1/netdev.h
drivers/infiniband/hw/hfi1/netdev_rx.c

index ca00f6c..185c9b0 100644 (file)
@@ -154,6 +154,9 @@ int hfi1_ipoib_send_dma(struct net_device *dev,
 int hfi1_ipoib_txreq_init(struct hfi1_ipoib_dev_priv *priv);
 void hfi1_ipoib_txreq_deinit(struct hfi1_ipoib_dev_priv *priv);
 
+int hfi1_ipoib_rxq_init(struct net_device *dev);
+void hfi1_ipoib_rxq_deinit(struct net_device *dev);
+
 void hfi1_ipoib_napi_tx_enable(struct net_device *dev);
 void hfi1_ipoib_napi_tx_disable(struct net_device *dev);
 
index 304a5ac..014351e 100644 (file)
@@ -19,16 +19,31 @@ static u32 qpn_from_mac(u8 *mac_arr)
 static int hfi1_ipoib_dev_init(struct net_device *dev)
 {
        struct hfi1_ipoib_dev_priv *priv = hfi1_ipoib_priv(dev);
+       int ret;
 
        priv->netstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
 
-       return priv->netdev_ops->ndo_init(dev);
+       ret = priv->netdev_ops->ndo_init(dev);
+       if (ret)
+               return ret;
+
+       ret = hfi1_netdev_add_data(priv->dd,
+                                  qpn_from_mac(priv->netdev->dev_addr),
+                                  dev);
+       if (ret < 0) {
+               priv->netdev_ops->ndo_uninit(dev);
+               return ret;
+       }
+
+       return 0;
 }
 
 static void hfi1_ipoib_dev_uninit(struct net_device *dev)
 {
        struct hfi1_ipoib_dev_priv *priv = hfi1_ipoib_priv(dev);
 
+       hfi1_netdev_remove_data(priv->dd, qpn_from_mac(priv->netdev->dev_addr));
+
        priv->netdev_ops->ndo_uninit(dev);
 }
 
@@ -55,6 +70,7 @@ static int hfi1_ipoib_dev_open(struct net_device *dev)
                priv->qp = qp;
                rcu_read_unlock();
 
+               hfi1_netdev_enable_queues(priv->dd);
                hfi1_ipoib_napi_tx_enable(dev);
        }
 
@@ -69,6 +85,7 @@ static int hfi1_ipoib_dev_stop(struct net_device *dev)
                return 0;
 
        hfi1_ipoib_napi_tx_disable(dev);
+       hfi1_netdev_disable_queues(priv->dd);
 
        rvt_put_qp(priv->qp);
        priv->qp = NULL;
@@ -195,6 +212,7 @@ static void hfi1_ipoib_netdev_dtor(struct net_device *dev)
        struct hfi1_ipoib_dev_priv *priv = hfi1_ipoib_priv(dev);
 
        hfi1_ipoib_txreq_deinit(priv);
+       hfi1_ipoib_rxq_deinit(priv->netdev);
 
        free_percpu(priv->netstats);
 }
@@ -252,6 +270,13 @@ static int hfi1_ipoib_setup_rn(struct ib_device *device,
                return rc;
        }
 
+       rc = hfi1_ipoib_rxq_init(netdev);
+       if (rc) {
+               dd_dev_err(dd, "IPoIB netdev RX init - failed(%d)\n", rc);
+               hfi1_ipoib_free_rdma_netdev(netdev);
+               return rc;
+       }
+
        netdev->priv_destructor = hfi1_ipoib_netdev_dtor;
        netdev->needs_free_netdev = true;
 
@@ -268,7 +293,7 @@ int hfi1_ipoib_rn_get_params(struct ib_device *device,
        if (type != RDMA_NETDEV_IPOIB)
                return -EOPNOTSUPP;
 
-       if (!HFI1_CAP_IS_KSET(AIP))
+       if (!HFI1_CAP_IS_KSET(AIP) || !dd->num_netdev_contexts)
                return -EOPNOTSUPP;
 
        if (!port_num || port_num > dd->num_pports)
@@ -276,6 +301,7 @@ int hfi1_ipoib_rn_get_params(struct ib_device *device,
 
        params->sizeof_priv = sizeof(struct hfi1_ipoib_rdma_netdev);
        params->txqs = dd->num_sdma;
+       params->rxqs = dd->num_netdev_contexts;
        params->param = NULL;
        params->initialize_rdma_netdev = hfi1_ipoib_setup_rn;
 
index 2485663..606ac69 100644 (file)
@@ -69,3 +69,19 @@ struct sk_buff *hfi1_ipoib_prepare_skb(struct hfi1_netdev_rxq *rxq,
 
        return skb;
 }
+
+int hfi1_ipoib_rxq_init(struct net_device *netdev)
+{
+       struct hfi1_ipoib_dev_priv *ipoib_priv = hfi1_ipoib_priv(netdev);
+       struct hfi1_devdata *dd = ipoib_priv->dd;
+
+       return hfi1_netdev_rx_init(dd);
+}
+
+void hfi1_ipoib_rxq_deinit(struct net_device *netdev)
+{
+       struct hfi1_ipoib_dev_priv *ipoib_priv = hfi1_ipoib_priv(netdev);
+       struct hfi1_devdata *dd = ipoib_priv->dd;
+
+       hfi1_netdev_rx_destroy(dd);
+}
index 6740ec3..edb936f 100644 (file)
@@ -82,6 +82,12 @@ struct hfi1_ctxtdata *hfi1_netdev_get_ctxt(struct hfi1_devdata *dd, int ctxt)
        return priv->rxq[ctxt].rcd;
 }
 
+void hfi1_netdev_enable_queues(struct hfi1_devdata *dd);
+void hfi1_netdev_disable_queues(struct hfi1_devdata *dd);
+int hfi1_netdev_rx_init(struct hfi1_devdata *dd);
+int hfi1_netdev_rx_destroy(struct hfi1_devdata *dd);
+int hfi1_netdev_alloc(struct hfi1_devdata *dd);
+void hfi1_netdev_free(struct hfi1_devdata *dd);
 int hfi1_netdev_add_data(struct hfi1_devdata *dd, int id, void *data);
 void *hfi1_netdev_remove_data(struct hfi1_devdata *dd, int id);
 void *hfi1_netdev_get_data(struct hfi1_devdata *dd, int id);
index 3e286cb..124e4e8 100644 (file)
 #include <linux/etherdevice.h>
 #include <rdma/ib_verbs.h>
 
+static int hfi1_netdev_setup_ctxt(struct hfi1_netdev_priv *priv,
+                                 struct hfi1_ctxtdata *uctxt)
+{
+       unsigned int rcvctrl_ops;
+       struct hfi1_devdata *dd = priv->dd;
+       int ret;
+
+       uctxt->rhf_rcv_function_map = netdev_rhf_rcv_functions;
+       uctxt->do_interrupt = &handle_receive_interrupt_napi_sp;
+
+       /* Now allocate the RcvHdr queue and eager buffers. */
+       ret = hfi1_create_rcvhdrq(dd, uctxt);
+       if (ret)
+               goto done;
+
+       ret = hfi1_setup_eagerbufs(uctxt);
+       if (ret)
+               goto done;
+
+       clear_rcvhdrtail(uctxt);
+
+       rcvctrl_ops = HFI1_RCVCTRL_CTXT_DIS;
+       rcvctrl_ops |= HFI1_RCVCTRL_INTRAVAIL_DIS;
+
+       if (!HFI1_CAP_KGET_MASK(uctxt->flags, MULTI_PKT_EGR))
+               rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
+       if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_EGR_FULL))
+               rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
+       if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
+               rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
+       if (HFI1_CAP_KGET_MASK(uctxt->flags, DMA_RTAIL))
+               rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
+
+       hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt);
+done:
+       return ret;
+}
+
+static int hfi1_netdev_allocate_ctxt(struct hfi1_devdata *dd,
+                                    struct hfi1_ctxtdata **ctxt)
+{
+       struct hfi1_ctxtdata *uctxt;
+       int ret;
+
+       if (dd->flags & HFI1_FROZEN)
+               return -EIO;
+
+       ret = hfi1_create_ctxtdata(dd->pport, dd->node, &uctxt);
+       if (ret < 0) {
+               dd_dev_err(dd, "Unable to create ctxtdata, failing open\n");
+               return -ENOMEM;
+       }
+
+       uctxt->flags = HFI1_CAP_KGET(MULTI_PKT_EGR) |
+               HFI1_CAP_KGET(NODROP_RHQ_FULL) |
+               HFI1_CAP_KGET(NODROP_EGR_FULL) |
+               HFI1_CAP_KGET(DMA_RTAIL);
+       /* Netdev contexts are always NO_RDMA_RTAIL */
+       uctxt->fast_handler = handle_receive_interrupt_napi_fp;
+       uctxt->slow_handler = handle_receive_interrupt_napi_sp;
+       hfi1_set_seq_cnt(uctxt, 1);
+       uctxt->is_vnic = true;
+
+       hfi1_stats.sps_ctxts++;
+
+       dd_dev_info(dd, "created netdev context %d\n", uctxt->ctxt);
+       *ctxt = uctxt;
+
+       return 0;
+}
+
+static void hfi1_netdev_deallocate_ctxt(struct hfi1_devdata *dd,
+                                       struct hfi1_ctxtdata *uctxt)
+{
+       flush_wc();
+
+       /*
+        * Disable receive context and interrupt available, reset all
+        * RcvCtxtCtrl bits to default values.
+        */
+       hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
+                    HFI1_RCVCTRL_TIDFLOW_DIS |
+                    HFI1_RCVCTRL_INTRAVAIL_DIS |
+                    HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
+                    HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
+                    HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt);
+
+       if (uctxt->msix_intr != CCE_NUM_MSIX_VECTORS)
+               msix_free_irq(dd, uctxt->msix_intr);
+
+       uctxt->msix_intr = CCE_NUM_MSIX_VECTORS;
+       uctxt->event_flags = 0;
+
+       hfi1_clear_tids(uctxt);
+       hfi1_clear_ctxt_pkey(dd, uctxt);
+
+       hfi1_stats.sps_ctxts--;
+
+       hfi1_free_ctxt(uctxt);
+}
+
+static int hfi1_netdev_allot_ctxt(struct hfi1_netdev_priv *priv,
+                                 struct hfi1_ctxtdata **ctxt)
+{
+       int rc;
+       struct hfi1_devdata *dd = priv->dd;
+
+       rc = hfi1_netdev_allocate_ctxt(dd, ctxt);
+       if (rc) {
+               dd_dev_err(dd, "netdev ctxt alloc failed %d\n", rc);
+               return rc;
+       }
+
+       rc = hfi1_netdev_setup_ctxt(priv, *ctxt);
+       if (rc) {
+               dd_dev_err(dd, "netdev ctxt setup failed %d\n", rc);
+               hfi1_netdev_deallocate_ctxt(dd, *ctxt);
+               *ctxt = NULL;
+       }
+
+       return rc;
+}
+
+static int hfi1_netdev_rxq_init(struct net_device *dev)
+{
+       int i;
+       int rc;
+       struct hfi1_netdev_priv *priv = hfi1_netdev_priv(dev);
+       struct hfi1_devdata *dd = priv->dd;
+
+       priv->num_rx_q = dd->num_netdev_contexts;
+       priv->rxq = kcalloc_node(priv->num_rx_q, sizeof(struct hfi1_netdev_rxq),
+                                GFP_KERNEL, dd->node);
+
+       if (!priv->rxq) {
+               dd_dev_err(dd, "Unable to allocate netdev queue data\n");
+               return (-ENOMEM);
+       }
+
+       for (i = 0; i < priv->num_rx_q; i++) {
+               struct hfi1_netdev_rxq *rxq = &priv->rxq[i];
+
+               rc = hfi1_netdev_allot_ctxt(priv, &rxq->rcd);
+               if (rc)
+                       goto bail_context_irq_failure;
+
+               hfi1_rcd_get(rxq->rcd);
+               rxq->priv = priv;
+               rxq->rcd->napi = &rxq->napi;
+               dd_dev_info(dd, "Setting rcv queue %d napi to context %d\n",
+                           i, rxq->rcd->ctxt);
+               /*
+                * Disable BUSY_POLL on this NAPI as this is not supported
+                * right now.
+                */
+               set_bit(NAPI_STATE_NO_BUSY_POLL, &rxq->napi.state);
+               netif_napi_add(dev, &rxq->napi, hfi1_netdev_rx_napi, 64);
+               rc = msix_netdev_request_rcd_irq(rxq->rcd);
+               if (rc)
+                       goto bail_context_irq_failure;
+       }
+
+       return 0;
+
+bail_context_irq_failure:
+       dd_dev_err(dd, "Unable to allot receive context\n");
+       for (; i >= 0; i--) {
+               struct hfi1_netdev_rxq *rxq = &priv->rxq[i];
+
+               if (rxq->rcd) {
+                       hfi1_netdev_deallocate_ctxt(dd, rxq->rcd);
+                       hfi1_rcd_put(rxq->rcd);
+                       rxq->rcd = NULL;
+               }
+       }
+       kfree(priv->rxq);
+       priv->rxq = NULL;
+
+       return rc;
+}
+
+static void hfi1_netdev_rxq_deinit(struct net_device *dev)
+{
+       int i;
+       struct hfi1_netdev_priv *priv = hfi1_netdev_priv(dev);
+       struct hfi1_devdata *dd = priv->dd;
+
+       for (i = 0; i < priv->num_rx_q; i++) {
+               struct hfi1_netdev_rxq *rxq = &priv->rxq[i];
+
+               netif_napi_del(&rxq->napi);
+               hfi1_netdev_deallocate_ctxt(dd, rxq->rcd);
+               hfi1_rcd_put(rxq->rcd);
+               rxq->rcd = NULL;
+       }
+
+       kfree(priv->rxq);
+       priv->rxq = NULL;
+       priv->num_rx_q = 0;
+}
+
+static void enable_queues(struct hfi1_netdev_priv *priv)
+{
+       int i;
+
+       for (i = 0; i < priv->num_rx_q; i++) {
+               struct hfi1_netdev_rxq *rxq = &priv->rxq[i];
+
+               dd_dev_info(priv->dd, "enabling queue %d on context %d\n", i,
+                           rxq->rcd->ctxt);
+               napi_enable(&rxq->napi);
+               hfi1_rcvctrl(priv->dd,
+                            HFI1_RCVCTRL_CTXT_ENB | HFI1_RCVCTRL_INTRAVAIL_ENB,
+                            rxq->rcd);
+       }
+}
+
+static void disable_queues(struct hfi1_netdev_priv *priv)
+{
+       int i;
+
+       msix_vnic_synchronize_irq(priv->dd);
+
+       for (i = 0; i < priv->num_rx_q; i++) {
+               struct hfi1_netdev_rxq *rxq = &priv->rxq[i];
+
+               dd_dev_info(priv->dd, "disabling queue %d on context %d\n", i,
+                           rxq->rcd->ctxt);
+
+               /* wait for napi if it was scheduled */
+               hfi1_rcvctrl(priv->dd,
+                            HFI1_RCVCTRL_CTXT_DIS | HFI1_RCVCTRL_INTRAVAIL_DIS,
+                            rxq->rcd);
+               napi_synchronize(&rxq->napi);
+               napi_disable(&rxq->napi);
+       }
+}
+
+/**
+ * hfi1_netdev_rx_init - Incrememnts netdevs counter. When called first time,
+ * it allocates receive queue data and calls netif_napi_add
+ * for each queue.
+ *
+ * @dd: hfi1 dev data
+ */
+int hfi1_netdev_rx_init(struct hfi1_devdata *dd)
+{
+       struct hfi1_netdev_priv *priv = hfi1_netdev_priv(dd->dummy_netdev);
+       int res;
+
+       if (atomic_fetch_inc(&priv->netdevs))
+               return 0;
+
+       mutex_lock(&hfi1_mutex);
+       init_dummy_netdev(dd->dummy_netdev);
+       res = hfi1_netdev_rxq_init(dd->dummy_netdev);
+       mutex_unlock(&hfi1_mutex);
+       return res;
+}
+
+/**
+ * hfi1_netdev_rx_destroy - Decrements netdevs counter, when it reaches 0
+ * napi is deleted and receive queses memory is freed.
+ *
+ * @dd: hfi1 dev data
+ */
+int hfi1_netdev_rx_destroy(struct hfi1_devdata *dd)
+{
+       struct hfi1_netdev_priv *priv = hfi1_netdev_priv(dd->dummy_netdev);
+
+       /* destroy the RX queues only if it is the last netdev going away */
+       if (atomic_fetch_add_unless(&priv->netdevs, -1, 0) == 1) {
+               mutex_lock(&hfi1_mutex);
+               hfi1_netdev_rxq_deinit(dd->dummy_netdev);
+               mutex_unlock(&hfi1_mutex);
+       }
+
+       return 0;
+}
+
+/**
+ * hfi1_netdev_alloc - Allocates netdev and private data. It is required
+ * because RMT index and MSI-X interrupt can be set only
+ * during driver initialization.
+ *
+ * @dd: hfi1 dev data
+ */
+int hfi1_netdev_alloc(struct hfi1_devdata *dd)
+{
+       struct hfi1_netdev_priv *priv;
+       const int netdev_size = sizeof(*dd->dummy_netdev) +
+               sizeof(struct hfi1_netdev_priv);
+
+       dd_dev_info(dd, "allocating netdev size %d\n", netdev_size);
+       dd->dummy_netdev = kcalloc_node(1, netdev_size, GFP_KERNEL, dd->node);
+
+       if (!dd->dummy_netdev)
+               return -ENOMEM;
+
+       priv = hfi1_netdev_priv(dd->dummy_netdev);
+       priv->dd = dd;
+       xa_init(&priv->dev_tbl);
+       atomic_set(&priv->enabled, 0);
+       atomic_set(&priv->netdevs, 0);
+
+       return 0;
+}
+
+void hfi1_netdev_free(struct hfi1_devdata *dd)
+{
+       struct hfi1_netdev_priv *priv;
+
+       if (dd->dummy_netdev) {
+               priv = hfi1_netdev_priv(dd->dummy_netdev);
+               dd_dev_info(dd, "hfi1 netdev freed\n");
+               kfree(dd->dummy_netdev);
+               dd->dummy_netdev = NULL;
+       }
+}
+
+/**
+ * hfi1_netdev_enable_queues - This is napi enable function.
+ * It enables napi objects associated with queues.
+ * When at least one device has called it it increments atomic counter.
+ * Disable function decrements counter and when it is 0,
+ * calls napi_disable for every queue.
+ *
+ * @dd: hfi1 dev data
+ */
+void hfi1_netdev_enable_queues(struct hfi1_devdata *dd)
+{
+       struct hfi1_netdev_priv *priv;
+
+       if (!dd->dummy_netdev)
+               return;
+
+       priv = hfi1_netdev_priv(dd->dummy_netdev);
+       if (atomic_fetch_inc(&priv->enabled))
+               return;
+
+       mutex_lock(&hfi1_mutex);
+       enable_queues(priv);
+       mutex_unlock(&hfi1_mutex);
+}
+
+void hfi1_netdev_disable_queues(struct hfi1_devdata *dd)
+{
+       struct hfi1_netdev_priv *priv;
+
+       if (!dd->dummy_netdev)
+               return;
+
+       priv = hfi1_netdev_priv(dd->dummy_netdev);
+       if (atomic_dec_if_positive(&priv->enabled))
+               return;
+
+       mutex_lock(&hfi1_mutex);
+       disable_queues(priv);
+       mutex_unlock(&hfi1_mutex);
+}
+
 /**
  * hfi1_netdev_add_data - Registers data with unique identifier
  * to be requested later this is needed for VNIC and IPoIB VLANs