igb: enable Rx queue interrupts for PF
authorCunming Liang <cunming.liang@intel.com>
Mon, 20 Jul 2015 03:02:28 +0000 (11:02 +0800)
committerThomas Monjalon <thomas.monjalon@6wind.com>
Mon, 27 Jul 2015 21:13:57 +0000 (23:13 +0200)
The patch does below for igb PF:
- Setup NIC to generate MSI-X interrupts
- Set the IVAR register to map interrupt causes to vectors
- Implement interrupt enable/disable functions

Signed-off-by: Danny Zhou <danny.zhou@intel.com>
Signed-off-by: Cunming Liang <cunming.liang@intel.com>
drivers/net/e1000/igb_ethdev.c

index ddc7186..56734a3 100644 (file)
@@ -105,6 +105,9 @@ static int  eth_igb_flow_ctrl_get(struct rte_eth_dev *dev,
 static int  eth_igb_flow_ctrl_set(struct rte_eth_dev *dev,
                                struct rte_eth_fc_conf *fc_conf);
 static int eth_igb_lsc_interrupt_setup(struct rte_eth_dev *dev);
+#ifdef RTE_NEXT_ABI
+static int eth_igb_rxq_interrupt_setup(struct rte_eth_dev *dev);
+#endif
 static int eth_igb_interrupt_get_status(struct rte_eth_dev *dev);
 static int eth_igb_interrupt_action(struct rte_eth_dev *dev);
 static void eth_igb_interrupt_handler(struct rte_intr_handle *handle,
@@ -218,7 +221,6 @@ static int eth_igb_get_eeprom(struct rte_eth_dev *dev,
                struct rte_dev_eeprom_info *eeprom);
 static int eth_igb_set_eeprom(struct rte_eth_dev *dev,
                struct rte_dev_eeprom_info *eeprom);
-
 static int eth_igb_set_mc_addr_list(struct rte_eth_dev *dev,
                                    struct ether_addr *mc_addr_set,
                                    uint32_t nb_mc_addr);
@@ -229,6 +231,17 @@ static int igb_timesync_read_rx_timestamp(struct rte_eth_dev *dev,
                                          uint32_t flags);
 static int igb_timesync_read_tx_timestamp(struct rte_eth_dev *dev,
                                          struct timespec *timestamp);
+#ifdef RTE_NEXT_ABI
+static int eth_igb_rx_queue_intr_enable(struct rte_eth_dev *dev,
+                                       uint16_t queue_id);
+static int eth_igb_rx_queue_intr_disable(struct rte_eth_dev *dev,
+                                        uint16_t queue_id);
+static void eth_igb_assign_msix_vector(struct e1000_hw *hw, int8_t direction,
+                                      uint8_t queue, uint8_t msix_vector);
+static void eth_igb_write_ivar(struct e1000_hw *hw, uint8_t msix_vector,
+                              uint8_t index, uint8_t offset);
+#endif
+static void eth_igb_configure_msix_intr(struct rte_eth_dev *dev);
 
 /*
  * Define VF Stats MACRO for Non "cleared on read" register
@@ -289,6 +302,10 @@ static const struct eth_dev_ops eth_igb_ops = {
        .vlan_tpid_set        = eth_igb_vlan_tpid_set,
        .vlan_offload_set     = eth_igb_vlan_offload_set,
        .rx_queue_setup       = eth_igb_rx_queue_setup,
+#ifdef RTE_NEXT_ABI
+       .rx_queue_intr_enable = eth_igb_rx_queue_intr_enable,
+       .rx_queue_intr_disable = eth_igb_rx_queue_intr_disable,
+#endif
        .rx_queue_release     = eth_igb_rx_queue_release,
        .rx_queue_count       = eth_igb_rx_queue_count,
        .rx_descriptor_done   = eth_igb_rx_descriptor_done,
@@ -639,12 +656,6 @@ eth_igb_dev_init(struct rte_eth_dev *eth_dev)
                     eth_dev->data->port_id, pci_dev->id.vendor_id,
                     pci_dev->id.device_id);
 
-       rte_intr_callback_register(&(pci_dev->intr_handle),
-               eth_igb_interrupt_handler, (void *)eth_dev);
-
-       /* enable uio intr after callback register */
-       rte_intr_enable(&(pci_dev->intr_handle));
-
        /* enable support intr */
        igb_intr_enable(eth_dev);
 
@@ -879,7 +890,11 @@ eth_igb_start(struct rte_eth_dev *dev)
                E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
        struct e1000_adapter *adapter =
                E1000_DEV_PRIVATE(dev->data->dev_private);
-       int ret, i, mask;
+       struct rte_intr_handle *intr_handle = &dev->pci_dev->intr_handle;
+       int ret, mask;
+#ifdef RTE_NEXT_ABI
+       uint32_t intr_vector = 0;
+#endif
        uint32_t ctrl_ext;
 
        PMD_INIT_FUNC_TRACE();
@@ -920,6 +935,29 @@ eth_igb_start(struct rte_eth_dev *dev)
        /* configure PF module if SRIOV enabled */
        igb_pf_host_configure(dev);
 
+#ifdef RTE_NEXT_ABI
+       /* check and configure queue intr-vector mapping */
+       if (dev->data->dev_conf.intr_conf.rxq != 0)
+               intr_vector = dev->data->nb_rx_queues;
+
+       if (rte_intr_efd_enable(intr_handle, intr_vector))
+               return -1;
+
+       if (rte_intr_dp_is_en(intr_handle)) {
+               intr_handle->intr_vec =
+                       rte_zmalloc("intr_vec",
+                                   dev->data->nb_rx_queues * sizeof(int), 0);
+               if (intr_handle->intr_vec == NULL) {
+                       PMD_INIT_LOG(ERR, "Failed to allocate %d rx_queues"
+                                    " intr_vec\n", dev->data->nb_rx_queues);
+                       return -ENOMEM;
+               }
+       }
+#endif
+
+       /* confiugre msix for rx interrupt */
+       eth_igb_configure_msix_intr(dev);
+
        /* Configure for OS presence */
        igb_init_manageability(hw);
 
@@ -947,33 +985,9 @@ eth_igb_start(struct rte_eth_dev *dev)
                igb_vmdq_vlan_hw_filter_enable(dev);
        }
 
-       /*
-        * Configure the Interrupt Moderation register (EITR) with the maximum
-        * possible value (0xFFFF) to minimize "System Partial Write" issued by
-        * spurious [DMA] memory updates of RX and TX ring descriptors.
-        *
-        * With a EITR granularity of 2 microseconds in the 82576, only 7/8
-        * spurious memory updates per second should be expected.
-        * ((65535 * 2) / 1000.1000 ~= 0.131 second).
-        *
-        * Because interrupts are not used at all, the MSI-X is not activated
-        * and interrupt moderation is controlled by EITR[0].
-        *
-        * Note that having [almost] disabled memory updates of RX and TX ring
-        * descriptors through the Interrupt Moderation mechanism, memory
-        * updates of ring descriptors are now moderated by the configurable
-        * value of Write-Back Threshold registers.
-        */
        if ((hw->mac.type == e1000_82576) || (hw->mac.type == e1000_82580) ||
                (hw->mac.type == e1000_i350) || (hw->mac.type == e1000_i210) ||
                (hw->mac.type == e1000_i211)) {
-               uint32_t ivar;
-
-               /* Enable all RX & TX queues in the IVAR registers */
-               ivar = (uint32_t) ((E1000_IVAR_VALID << 16) | E1000_IVAR_VALID);
-               for (i = 0; i < 8; i++)
-                       E1000_WRITE_REG_ARRAY(hw, E1000_IVAR0, i, ivar);
-
                /* Configure EITR with the maximum possible value (0xFFFF) */
                E1000_WRITE_REG(hw, E1000_EITR(0), 0xFFFF);
        }
@@ -1024,8 +1038,25 @@ eth_igb_start(struct rte_eth_dev *dev)
        e1000_setup_link(hw);
 
        /* check if lsc interrupt feature is enabled */
-       if (dev->data->dev_conf.intr_conf.lsc != 0)
-               ret = eth_igb_lsc_interrupt_setup(dev);
+       if (dev->data->dev_conf.intr_conf.lsc != 0) {
+               if (rte_intr_allow_others(intr_handle)) {
+                       rte_intr_callback_register(intr_handle,
+                                                  eth_igb_interrupt_handler,
+                                                  (void *)dev);
+                       eth_igb_lsc_interrupt_setup(dev);
+               } else
+                       PMD_INIT_LOG(INFO, "lsc won't enable because of"
+                                    " no intr multiplex\n");
+       }
+
+#ifdef RTE_NEXT_ABI
+       /* check if rxq interrupt is enabled */
+       if (dev->data->dev_conf.intr_conf.rxq != 0)
+               eth_igb_rxq_interrupt_setup(dev);
+#endif
+
+       /* enable uio/vfio intr/eventfd mapping */
+       rte_intr_enable(intr_handle);
 
        /* resume enabled intr since hw reset */
        igb_intr_enable(dev);
@@ -1058,8 +1089,13 @@ eth_igb_stop(struct rte_eth_dev *dev)
        struct e1000_flex_filter *p_flex;
        struct e1000_5tuple_filter *p_5tuple, *p_5tuple_next;
        struct e1000_2tuple_filter *p_2tuple, *p_2tuple_next;
+       struct rte_intr_handle *intr_handle = &dev->pci_dev->intr_handle;
 
        igb_intr_disable(hw);
+
+       /* disable intr eventfd mapping */
+       rte_intr_disable(intr_handle);
+
        igb_pf_reset_hw(hw);
        E1000_WRITE_REG(hw, E1000_WUC, 0);
 
@@ -1108,6 +1144,15 @@ eth_igb_stop(struct rte_eth_dev *dev)
                rte_free(p_2tuple);
        }
        filter_info->twotuple_mask = 0;
+
+#ifdef RTE_NEXT_ABI
+       /* Clean datapath event and queue/vec mapping */
+       rte_intr_efd_disable(intr_handle);
+       if (intr_handle->intr_vec != NULL) {
+               rte_free(intr_handle->intr_vec);
+               intr_handle->intr_vec = NULL;
+       }
+#endif
 }
 
 static void
@@ -1117,6 +1162,9 @@ eth_igb_close(struct rte_eth_dev *dev)
        struct e1000_adapter *adapter =
                E1000_DEV_PRIVATE(dev->data->dev_private);
        struct rte_eth_link link;
+#ifdef RTE_NEXT_ABI
+       struct rte_pci_device *pci_dev;
+#endif
 
        eth_igb_stop(dev);
        adapter->stopped = 1;
@@ -1136,6 +1184,14 @@ eth_igb_close(struct rte_eth_dev *dev)
 
        igb_dev_free_queues(dev);
 
+#ifdef RTE_NEXT_ABI
+       pci_dev = dev->pci_dev;
+       if (pci_dev->intr_handle.intr_vec) {
+               rte_free(pci_dev->intr_handle.intr_vec);
+               pci_dev->intr_handle.intr_vec = NULL;
+       }
+#endif
+
        memset(&link, 0, sizeof(link));
        rte_igb_dev_atomic_write_link_status(dev, &link);
 }
@@ -1960,6 +2016,35 @@ eth_igb_lsc_interrupt_setup(struct rte_eth_dev *dev)
        return 0;
 }
 
+#ifdef RTE_NEXT_ABI
+/* It clears the interrupt causes and enables the interrupt.
+ * It will be called once only during nic initialized.
+ *
+ * @param dev
+ *  Pointer to struct rte_eth_dev.
+ *
+ * @return
+ *  - On success, zero.
+ *  - On failure, a negative value.
+ */
+static int eth_igb_rxq_interrupt_setup(struct rte_eth_dev *dev)
+{
+       uint32_t mask, regval;
+       struct e1000_hw *hw =
+               E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
+       struct rte_eth_dev_info dev_info;
+
+       memset(&dev_info, 0, sizeof(dev_info));
+       eth_igb_infos_get(dev, &dev_info);
+
+       mask = 0xFFFFFFFF >> (32 - dev_info.max_rx_queues);
+       regval = E1000_READ_REG(hw, E1000_EIMS);
+       E1000_WRITE_REG(hw, E1000_EIMS, regval | mask);
+
+       return 0;
+}
+#endif
+
 /*
  * It reads ICR and gets interrupt causes, check it and set a bit flag
  * to update link status.
@@ -4051,5 +4136,163 @@ static struct rte_driver pmd_igbvf_drv = {
        .init = rte_igbvf_pmd_init,
 };
 
+#ifdef RTE_NEXT_ABI
+static int
+eth_igb_rx_queue_intr_disable(struct rte_eth_dev *dev, uint16_t queue_id)
+{
+       struct e1000_hw *hw =
+               E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
+       uint32_t mask = 1 << queue_id;
+
+       E1000_WRITE_REG(hw, E1000_EIMC, mask);
+       E1000_WRITE_FLUSH(hw);
+
+       return 0;
+}
+
+static int
+eth_igb_rx_queue_intr_enable(struct rte_eth_dev *dev, uint16_t queue_id)
+{
+       struct e1000_hw *hw =
+               E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
+       uint32_t mask = 1 << queue_id;
+       uint32_t regval;
+
+       regval = E1000_READ_REG(hw, E1000_EIMS);
+       E1000_WRITE_REG(hw, E1000_EIMS, regval | mask);
+       E1000_WRITE_FLUSH(hw);
+
+       rte_intr_enable(&dev->pci_dev->intr_handle);
+
+       return 0;
+}
+
+static void
+eth_igb_write_ivar(struct e1000_hw *hw, uint8_t  msix_vector,
+                  uint8_t index, uint8_t offset)
+{
+       uint32_t val = E1000_READ_REG_ARRAY(hw, E1000_IVAR0, index);
+
+       /* clear bits */
+       val &= ~((uint32_t)0xFF << offset);
+
+       /* write vector and valid bit */
+       val |= (msix_vector | E1000_IVAR_VALID) << offset;
+
+       E1000_WRITE_REG_ARRAY(hw, E1000_IVAR0, index, val);
+}
+
+static void
+eth_igb_assign_msix_vector(struct e1000_hw *hw, int8_t direction,
+                          uint8_t queue, uint8_t msix_vector)
+{
+       uint32_t tmp = 0;
+
+       if (hw->mac.type == e1000_82575) {
+               if (direction == 0)
+                       tmp = E1000_EICR_RX_QUEUE0 << queue;
+               else if (direction == 1)
+                       tmp = E1000_EICR_TX_QUEUE0 << queue;
+               E1000_WRITE_REG(hw, E1000_MSIXBM(msix_vector), tmp);
+       } else if (hw->mac.type == e1000_82576) {
+               if ((direction == 0) || (direction == 1))
+                       eth_igb_write_ivar(hw, msix_vector, queue & 0x7,
+                                          ((queue & 0x8) << 1) +
+                                          8 * direction);
+       } else if ((hw->mac.type == e1000_82580) ||
+                       (hw->mac.type == e1000_i350) ||
+                       (hw->mac.type == e1000_i354) ||
+                       (hw->mac.type == e1000_i210) ||
+                       (hw->mac.type == e1000_i211)) {
+               if ((direction == 0) || (direction == 1))
+                       eth_igb_write_ivar(hw, msix_vector,
+                                          queue >> 1,
+                                          ((queue & 0x1) << 4) +
+                                          8 * direction);
+       }
+}
+#endif
+
+/* Sets up the hardware to generate MSI-X interrupts properly
+ * @hw
+ *  board private structure
+ */
+static void
+eth_igb_configure_msix_intr(struct rte_eth_dev *dev)
+{
+#ifdef RTE_NEXT_ABI
+       int queue_id;
+       uint32_t tmpval, regval, intr_mask;
+       struct e1000_hw *hw =
+               E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
+       uint32_t vec = 0;
+#endif
+       struct rte_intr_handle *intr_handle = &dev->pci_dev->intr_handle;
+
+       /* won't configure msix register if no mapping is done
+        * between intr vector and event fd
+        */
+       if (!rte_intr_dp_is_en(intr_handle))
+               return;
+
+#ifdef RTE_NEXT_ABI
+       /* set interrupt vector for other causes */
+       if (hw->mac.type == e1000_82575) {
+               tmpval = E1000_READ_REG(hw, E1000_CTRL_EXT);
+               /* enable MSI-X PBA support */
+               tmpval |= E1000_CTRL_EXT_PBA_CLR;
+
+               /* Auto-Mask interrupts upon ICR read */
+               tmpval |= E1000_CTRL_EXT_EIAME;
+               tmpval |= E1000_CTRL_EXT_IRCA;
+
+               E1000_WRITE_REG(hw, E1000_CTRL_EXT, tmpval);
+
+               /* enable msix_other interrupt */
+               E1000_WRITE_REG_ARRAY(hw, E1000_MSIXBM(0), 0, E1000_EIMS_OTHER);
+               regval = E1000_READ_REG(hw, E1000_EIAC);
+               E1000_WRITE_REG(hw, E1000_EIAC, regval | E1000_EIMS_OTHER);
+               regval = E1000_READ_REG(hw, E1000_EIAM);
+               E1000_WRITE_REG(hw, E1000_EIMS, regval | E1000_EIMS_OTHER);
+       } else if ((hw->mac.type == e1000_82576) ||
+                       (hw->mac.type == e1000_82580) ||
+                       (hw->mac.type == e1000_i350) ||
+                       (hw->mac.type == e1000_i354) ||
+                       (hw->mac.type == e1000_i210) ||
+                       (hw->mac.type == e1000_i211)) {
+               /* turn on MSI-X capability first */
+               E1000_WRITE_REG(hw, E1000_GPIE, E1000_GPIE_MSIX_MODE |
+                                       E1000_GPIE_PBA | E1000_GPIE_EIAME |
+                                       E1000_GPIE_NSICR);
+
+               intr_mask = (1 << intr_handle->max_intr) - 1;
+               regval = E1000_READ_REG(hw, E1000_EIAC);
+               E1000_WRITE_REG(hw, E1000_EIAC, regval | intr_mask);
+
+               /* enable msix_other interrupt */
+               regval = E1000_READ_REG(hw, E1000_EIMS);
+               E1000_WRITE_REG(hw, E1000_EIMS, regval | intr_mask);
+               tmpval = (dev->data->nb_rx_queues | E1000_IVAR_VALID) << 8;
+               E1000_WRITE_REG(hw, E1000_IVAR_MISC, tmpval);
+       }
+
+       /* use EIAM to auto-mask when MSI-X interrupt
+        * is asserted, this saves a register write for every interrupt
+        */
+       intr_mask = (1 << intr_handle->nb_efd) - 1;
+       regval = E1000_READ_REG(hw, E1000_EIAM);
+       E1000_WRITE_REG(hw, E1000_EIAM, regval | intr_mask);
+
+       for (queue_id = 0; queue_id < dev->data->nb_rx_queues; queue_id++) {
+               eth_igb_assign_msix_vector(hw, 0, queue_id, vec);
+               intr_handle->intr_vec[queue_id] = vec;
+               if (vec < intr_handle->nb_efd - 1)
+                       vec++;
+       }
+
+       E1000_WRITE_FLUSH(hw);
+#endif
+}
+
 PMD_REGISTER_DRIVER(pmd_igb_drv);
 PMD_REGISTER_DRIVER(pmd_igbvf_drv);