net/af_packet: release port upon close
[dpdk.git] / drivers / net / ena / ena_ethdev.c
index 1a7cc68..8c0dd30 100644 (file)
@@ -27,8 +27,8 @@
 #include <ena_eth_io_defs.h>
 
 #define DRV_MODULE_VER_MAJOR   2
-#define DRV_MODULE_VER_MINOR   0
-#define DRV_MODULE_VER_SUBMINOR        3
+#define DRV_MODULE_VER_MINOR   1
+#define DRV_MODULE_VER_SUBMINOR        0
 
 #define ENA_IO_TXQ_IDX(q)      (2 * (q))
 #define ENA_IO_RXQ_IDX(q)      (2 * (q) + 1)
@@ -72,6 +72,9 @@ struct ena_stats {
 #define ENA_STAT_TX_ENTRY(stat) \
        ENA_STAT_ENTRY(stat, tx)
 
+#define ENA_STAT_ENI_ENTRY(stat) \
+       ENA_STAT_ENTRY(stat, eni)
+
 #define ENA_STAT_GLOBAL_ENTRY(stat) \
        ENA_STAT_ENTRY(stat, dev)
 
@@ -91,6 +94,14 @@ static const struct ena_stats ena_stats_global_strings[] = {
        ENA_STAT_GLOBAL_ENTRY(tx_drops),
 };
 
+static const struct ena_stats ena_stats_eni_strings[] = {
+       ENA_STAT_ENI_ENTRY(bw_in_allowance_exceeded),
+       ENA_STAT_ENI_ENTRY(bw_out_allowance_exceeded),
+       ENA_STAT_ENI_ENTRY(pps_allowance_exceeded),
+       ENA_STAT_ENI_ENTRY(conntrack_allowance_exceeded),
+       ENA_STAT_ENI_ENTRY(linklocal_allowance_exceeded),
+};
+
 static const struct ena_stats ena_stats_tx_strings[] = {
        ENA_STAT_TX_ENTRY(cnt),
        ENA_STAT_TX_ENTRY(bytes),
@@ -114,6 +125,7 @@ static const struct ena_stats ena_stats_rx_strings[] = {
 };
 
 #define ENA_STATS_ARRAY_GLOBAL ARRAY_SIZE(ena_stats_global_strings)
+#define ENA_STATS_ARRAY_ENI    ARRAY_SIZE(ena_stats_eni_strings)
 #define ENA_STATS_ARRAY_TX     ARRAY_SIZE(ena_stats_tx_strings)
 #define ENA_STATS_ARRAY_RX     ARRAY_SIZE(ena_stats_rx_strings)
 
@@ -141,22 +153,6 @@ static const struct ena_stats ena_stats_rx_strings[] = {
 #define        ENA_TX_OFFLOAD_NOTSUP_MASK      \
        (PKT_TX_OFFLOAD_MASK ^ ENA_TX_OFFLOAD_MASK)
 
-int ena_logtype_init;
-int ena_logtype_driver;
-
-#ifdef RTE_LIBRTE_ENA_DEBUG_RX
-int ena_logtype_rx;
-#endif
-#ifdef RTE_LIBRTE_ENA_DEBUG_TX
-int ena_logtype_tx;
-#endif
-#ifdef RTE_LIBRTE_ENA_DEBUG_TX_FREE
-int ena_logtype_tx_free;
-#endif
-#ifdef RTE_LIBRTE_ENA_COM_DEBUG
-int ena_logtype_com;
-#endif
-
 static const struct rte_pci_id pci_id_ena_map[] = {
        { RTE_PCI_DEVICE(PCI_VENDOR_ID_AMAZON, PCI_DEVICE_ID_ENA_VF) },
        { RTE_PCI_DEVICE(PCI_VENDOR_ID_AMAZON, PCI_DEVICE_ID_ENA_LLQ_VF) },
@@ -195,13 +191,15 @@ static struct rte_mbuf *ena_rx_mbuf(struct ena_ring *rx_ring,
                                    uint8_t offset);
 static uint16_t eth_ena_recv_pkts(void *rx_queue,
                                  struct rte_mbuf **rx_pkts, uint16_t nb_pkts);
+static int ena_add_single_rx_desc(struct ena_com_io_sq *io_sq,
+                                 struct rte_mbuf *mbuf, uint16_t id);
 static int ena_populate_rx_queue(struct ena_ring *rxq, unsigned int count);
 static void ena_init_rings(struct ena_adapter *adapter,
                           bool disable_meta_caching);
 static int ena_mtu_set(struct rte_eth_dev *dev, uint16_t mtu);
 static int ena_start(struct rte_eth_dev *dev);
 static void ena_stop(struct rte_eth_dev *dev);
-static void ena_close(struct rte_eth_dev *dev);
+static int ena_close(struct rte_eth_dev *dev);
 static int ena_dev_reset(struct rte_eth_dev *dev);
 static int ena_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats);
 static void ena_rx_queue_release_all(struct rte_eth_dev *dev);
@@ -247,6 +245,7 @@ static int ena_process_bool_devarg(const char *key,
                                   void *opaque);
 static int ena_parse_devargs(struct ena_adapter *adapter,
                             struct rte_devargs *devargs);
+static int ena_copy_eni_stats(struct ena_adapter *adapter);
 
 static const struct eth_dev_ops ena_dev_ops = {
        .dev_configure        = ena_dev_configure,
@@ -465,7 +464,7 @@ err:
 /* This function calculates the number of xstats based on the current config */
 static unsigned int ena_xstats_calc_num(struct rte_eth_dev *dev)
 {
-       return ENA_STATS_ARRAY_GLOBAL +
+       return ENA_STATS_ARRAY_GLOBAL + ENA_STATS_ARRAY_ENI +
                (dev->data->nb_tx_queues * ENA_STATS_ARRAY_TX) +
                (dev->data->nb_rx_queues * ENA_STATS_ARRAY_RX);
 }
@@ -501,7 +500,7 @@ err:
        ena_com_delete_debug_area(&adapter->ena_dev);
 }
 
-static void ena_close(struct rte_eth_dev *dev)
+static int ena_close(struct rte_eth_dev *dev)
 {
        struct rte_pci_device *pci_dev = RTE_ETH_DEV_TO_PCI(dev);
        struct rte_intr_handle *intr_handle = &pci_dev->intr_handle;
@@ -527,6 +526,8 @@ static void ena_close(struct rte_eth_dev *dev)
         * release of the resource in the rte_eth_dev_release_port().
         */
        dev->data->mac_addrs = NULL;
+
+       return 0;
 }
 
 static int
@@ -584,7 +585,9 @@ static int ena_rss_reta_update(struct rte_eth_dev *dev,
                }
        }
 
+       rte_spinlock_lock(&adapter->admin_lock);
        rc = ena_com_indirect_table_set(ena_dev);
+       rte_spinlock_unlock(&adapter->admin_lock);
        if (unlikely(rc && rc != ENA_COM_UNSUPPORTED)) {
                PMD_DRV_LOG(ERR, "Cannot flush the indirect table\n");
                return rc;
@@ -613,7 +616,9 @@ static int ena_rss_reta_query(struct rte_eth_dev *dev,
            (reta_size > RTE_RETA_GROUP_SIZE && ((reta_conf + 1) == NULL)))
                return -EINVAL;
 
+       rte_spinlock_lock(&adapter->admin_lock);
        rc = ena_com_indirect_table_get(ena_dev, indirect_table);
+       rte_spinlock_unlock(&adapter->admin_lock);
        if (unlikely(rc && rc != ENA_COM_UNSUPPORTED)) {
                PMD_DRV_LOG(ERR, "cannot get indirect table\n");
                return -ENOTSUP;
@@ -968,7 +973,10 @@ static int ena_stats_get(struct rte_eth_dev *dev,
                return -ENOTSUP;
 
        memset(&ena_stats, 0, sizeof(ena_stats));
+
+       rte_spinlock_lock(&adapter->admin_lock);
        rc = ena_com_get_dev_basic_stats(ena_dev, &ena_stats);
+       rte_spinlock_unlock(&adapter->admin_lock);
        if (unlikely(rc)) {
                PMD_DRV_LOG(ERR, "Could not retrieve statistics from ENA\n");
                return rc;
@@ -1414,6 +1422,24 @@ static int ena_rx_queue_setup(struct rte_eth_dev *dev,
        return 0;
 }
 
+static int ena_add_single_rx_desc(struct ena_com_io_sq *io_sq,
+                                 struct rte_mbuf *mbuf, uint16_t id)
+{
+       struct ena_com_buf ebuf;
+       int rc;
+
+       /* prepare physical address for DMA transaction */
+       ebuf.paddr = mbuf->buf_iova + RTE_PKTMBUF_HEADROOM;
+       ebuf.len = mbuf->buf_len - RTE_PKTMBUF_HEADROOM;
+
+       /* pass resource to device */
+       rc = ena_com_add_single_rx_desc(io_sq, &ebuf, id);
+       if (unlikely(rc != 0))
+               PMD_DRV_LOG(WARNING, "failed adding rx desc\n");
+
+       return rc;
+}
+
 static int ena_populate_rx_queue(struct ena_ring *rxq, unsigned int count)
 {
        unsigned int i;
@@ -1441,7 +1467,6 @@ static int ena_populate_rx_queue(struct ena_ring *rxq, unsigned int count)
 
        for (i = 0; i < count; i++) {
                struct rte_mbuf *mbuf = mbufs[i];
-               struct ena_com_buf ebuf;
                struct ena_rx_buffer *rx_info;
 
                if (likely((i + 4) < count))
@@ -1454,16 +1479,10 @@ static int ena_populate_rx_queue(struct ena_ring *rxq, unsigned int count)
 
                rx_info = &rxq->rx_buffer_info[req_id];
 
-               /* prepare physical address for DMA transaction */
-               ebuf.paddr = mbuf->buf_iova + RTE_PKTMBUF_HEADROOM;
-               ebuf.len = mbuf->buf_len - RTE_PKTMBUF_HEADROOM;
-               /* pass resource to device */
-               rc = ena_com_add_single_rx_desc(rxq->ena_com_io_sq,
-                                               &ebuf, req_id);
-               if (unlikely(rc)) {
-                       PMD_DRV_LOG(WARNING, "failed adding rx desc\n");
+               rc = ena_add_single_rx_desc(rxq->ena_com_io_sq, mbuf, req_id);
+               if (unlikely(rc != 0))
                        break;
-               }
+
                rx_info->mbuf = mbuf;
                next_to_use = ENA_IDX_NEXT_MASKED(next_to_use, rxq->size_mask);
        }
@@ -1624,7 +1643,7 @@ static void ena_timer_wd_callback(__rte_unused struct rte_timer *timer,
 
        if (unlikely(adapter->trigger_reset)) {
                PMD_DRV_LOG(ERR, "Trigger reset is on\n");
-               _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RESET,
+               rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RESET,
                        NULL);
        }
 }
@@ -1743,7 +1762,7 @@ static int eth_ena_dev_init(struct rte_eth_dev *eth_dev)
        int rc;
        static int adapters_found;
        bool disable_meta_caching;
-       bool wd_state;
+       bool wd_state = false;
 
        eth_dev->dev_ops = &ena_dev_ops;
        eth_dev->rx_pkt_burst = &eth_ena_recv_pkts;
@@ -1879,6 +1898,8 @@ static int eth_ena_dev_init(struct rte_eth_dev *eth_dev)
                goto err_delete_debug_area;
        }
 
+       rte_spinlock_init(&adapter->admin_lock);
+
        rte_intr_callback_register(intr_handle,
                                   ena_interrupt_handler_rte,
                                   adapter);
@@ -2079,6 +2100,7 @@ static struct rte_mbuf *ena_rx_mbuf(struct ena_ring *rx_ring,
        struct rte_mbuf *mbuf;
        struct rte_mbuf *mbuf_head;
        struct ena_rx_buffer *rx_info;
+       int rc;
        uint16_t ntc, len, req_id, buf = 0;
 
        if (unlikely(descs == 0))
@@ -2121,13 +2143,44 @@ static struct rte_mbuf *ena_rx_mbuf(struct ena_ring *rx_ring,
                rx_info = &rx_ring->rx_buffer_info[req_id];
                RTE_ASSERT(rx_info->mbuf != NULL);
 
-               /* Create an mbuf chain. */
-               mbuf->next = rx_info->mbuf;
-               mbuf = mbuf->next;
+               if (unlikely(len == 0)) {
+                       /*
+                        * Some devices can pass descriptor with the length 0.
+                        * To avoid confusion, the PMD is simply putting the
+                        * descriptor back, as it was never used. We'll avoid
+                        * mbuf allocation that way.
+                        */
+                       rc = ena_add_single_rx_desc(rx_ring->ena_com_io_sq,
+                               rx_info->mbuf, req_id);
+                       if (unlikely(rc != 0)) {
+                               /* Free the mbuf in case of an error. */
+                               rte_mbuf_raw_free(rx_info->mbuf);
+                       } else {
+                               /*
+                                * If there was no error, just exit the loop as
+                                * 0 length descriptor is always the last one.
+                                */
+                               break;
+                       }
+               } else {
+                       /* Create an mbuf chain. */
+                       mbuf->next = rx_info->mbuf;
+                       mbuf = mbuf->next;
 
-               ena_init_rx_mbuf(mbuf, len);
-               mbuf_head->pkt_len += len;
+                       ena_init_rx_mbuf(mbuf, len);
+                       mbuf_head->pkt_len += len;
+               }
 
+               /*
+                * Mark the descriptor as depleted and perform necessary
+                * cleanup.
+                * This code will execute in two cases:
+                *  1. Descriptor len was greater than 0 - normal situation.
+                *  2. Descriptor len was 0 and we failed to add the descriptor
+                *     to the device. In that situation, we should try to add
+                *     the mbuf again in the populate routine and mark the
+                *     descriptor as used up by the device.
+                */
                rx_info->mbuf = NULL;
                rx_ring->empty_rx_reqs[ntc] = req_id;
                ntc = ENA_IDX_NEXT_MASKED(ntc, rx_ring->size_mask);
@@ -2570,6 +2623,31 @@ static uint16_t eth_ena_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
        return sent_idx;
 }
 
+int ena_copy_eni_stats(struct ena_adapter *adapter)
+{
+       struct ena_admin_eni_stats admin_eni_stats;
+       int rc;
+
+       rte_spinlock_lock(&adapter->admin_lock);
+       rc = ena_com_get_eni_stats(&adapter->ena_dev, &admin_eni_stats);
+       rte_spinlock_unlock(&adapter->admin_lock);
+       if (rc != 0) {
+               if (rc == ENA_COM_UNSUPPORTED) {
+                       PMD_DRV_LOG(DEBUG,
+                               "Retrieving ENI metrics is not supported.\n");
+               } else {
+                       PMD_DRV_LOG(WARNING,
+                               "Failed to get ENI metrics: %d\n", rc);
+               }
+               return rc;
+       }
+
+       rte_memcpy(&adapter->eni_stats, &admin_eni_stats,
+               sizeof(struct ena_stats_eni));
+
+       return 0;
+}
+
 /**
  * DPDK callback to retrieve names of extended device statistics
  *
@@ -2597,6 +2675,10 @@ static int ena_xstats_get_names(struct rte_eth_dev *dev,
                strcpy(xstats_names[count].name,
                        ena_stats_global_strings[stat].name);
 
+       for (stat = 0; stat < ENA_STATS_ARRAY_ENI; stat++, count++)
+               strcpy(xstats_names[count].name,
+                       ena_stats_eni_strings[stat].name);
+
        for (stat = 0; stat < ENA_STATS_ARRAY_RX; stat++)
                for (i = 0; i < dev->data->nb_rx_queues; i++, count++)
                        snprintf(xstats_names[count].name,
@@ -2652,6 +2734,19 @@ static int ena_xstats_get(struct rte_eth_dev *dev,
                        ((char *)stats_begin + stat_offset));
        }
 
+       /* Even if the function below fails, we should copy previous (or initial
+        * values) to keep structure of rte_eth_xstat consistent.
+        */
+       ena_copy_eni_stats(adapter);
+       for (stat = 0; stat < ENA_STATS_ARRAY_ENI; stat++, count++) {
+               stat_offset = ena_stats_eni_strings[stat].stat_offset;
+               stats_begin = &adapter->eni_stats;
+
+               xstats[count].id = count;
+               xstats[count].value = *((uint64_t *)
+                   ((char *)stats_begin + stat_offset));
+       }
+
        for (stat = 0; stat < ENA_STATS_ARRAY_RX; stat++) {
                for (i = 0; i < dev->data->nb_rx_queues; i++, count++) {
                        stat_offset = ena_stats_rx_strings[stat].stat_offset;
@@ -2688,6 +2783,8 @@ static int ena_xstats_get_by_id(struct rte_eth_dev *dev,
        unsigned int i;
        int qid;
        int valid = 0;
+       bool was_eni_copied = false;
+
        for (i = 0; i < n; ++i) {
                id = ids[i];
                /* Check if id belongs to global statistics */
@@ -2697,8 +2794,24 @@ static int ena_xstats_get_by_id(struct rte_eth_dev *dev,
                        continue;
                }
 
-               /* Check if id belongs to rx queue statistics */
+               /* Check if id belongs to ENI statistics */
                id -= ENA_STATS_ARRAY_GLOBAL;
+               if (id < ENA_STATS_ARRAY_ENI) {
+                       /* Avoid reading ENI stats multiple times in a single
+                        * function call, as it requires communication with the
+                        * admin queue.
+                        */
+                       if (!was_eni_copied) {
+                               was_eni_copied = true;
+                               ena_copy_eni_stats(adapter);
+                       }
+                       values[i] = *((uint64_t *)&adapter->eni_stats + id);
+                       ++valid;
+                       continue;
+               }
+
+               /* Check if id belongs to rx queue statistics */
+               id -= ENA_STATS_ARRAY_ENI;
                rx_entries = ENA_STATS_ARRAY_RX * dev->data->nb_rx_queues;
                if (id < rx_entries) {
                        qid = id % dev->data->nb_rx_queues;
@@ -2804,40 +2917,20 @@ RTE_PMD_REGISTER_PCI(net_ena, rte_ena_pmd);
 RTE_PMD_REGISTER_PCI_TABLE(net_ena, pci_id_ena_map);
 RTE_PMD_REGISTER_KMOD_DEP(net_ena, "* igb_uio | uio_pci_generic | vfio-pci");
 RTE_PMD_REGISTER_PARAM_STRING(net_ena, ENA_DEVARG_LARGE_LLQ_HDR "=<0|1>");
-
-RTE_INIT(ena_init_log)
-{
-       ena_logtype_init = rte_log_register("pmd.net.ena.init");
-       if (ena_logtype_init >= 0)
-               rte_log_set_level(ena_logtype_init, RTE_LOG_NOTICE);
-       ena_logtype_driver = rte_log_register("pmd.net.ena.driver");
-       if (ena_logtype_driver >= 0)
-               rte_log_set_level(ena_logtype_driver, RTE_LOG_NOTICE);
-
+RTE_LOG_REGISTER(ena_logtype_init, pmd.net.ena.init, NOTICE);
+RTE_LOG_REGISTER(ena_logtype_driver, pmd.net.ena.driver, NOTICE);
 #ifdef RTE_LIBRTE_ENA_DEBUG_RX
-       ena_logtype_rx = rte_log_register("pmd.net.ena.rx");
-       if (ena_logtype_rx >= 0)
-               rte_log_set_level(ena_logtype_rx, RTE_LOG_NOTICE);
+RTE_LOG_REGISTER(ena_logtype_rx, pmd.net.ena.rx, NOTICE);
 #endif
-
 #ifdef RTE_LIBRTE_ENA_DEBUG_TX
-       ena_logtype_tx = rte_log_register("pmd.net.ena.tx");
-       if (ena_logtype_tx >= 0)
-               rte_log_set_level(ena_logtype_tx, RTE_LOG_NOTICE);
+RTE_LOG_REGISTER(ena_logtype_tx, pmd.net.ena.tx, NOTICE);
 #endif
-
 #ifdef RTE_LIBRTE_ENA_DEBUG_TX_FREE
-       ena_logtype_tx_free = rte_log_register("pmd.net.ena.tx_free");
-       if (ena_logtype_tx_free >= 0)
-               rte_log_set_level(ena_logtype_tx_free, RTE_LOG_NOTICE);
+RTE_LOG_REGISTER(ena_logtype_tx_free, pmd.net.ena.tx_free, NOTICE);
 #endif
-
 #ifdef RTE_LIBRTE_ENA_COM_DEBUG
-       ena_logtype_com = rte_log_register("pmd.net.ena.com");
-       if (ena_logtype_com >= 0)
-               rte_log_set_level(ena_logtype_com, RTE_LOG_NOTICE);
+RTE_LOG_REGISTER(ena_logtype_com, pmd.net.ena.com, NOTICE);
 #endif
-}
 
 /******************************************************************************
  ******************************** AENQ Handlers *******************************
@@ -2858,7 +2951,7 @@ static void ena_update_on_link_change(void *adapter_data,
        adapter->link_status = status;
 
        ena_link_update(eth_dev, 0);
-       _rte_eth_dev_callback_process(eth_dev, RTE_ETH_EVENT_INTR_LSC, NULL);
+       rte_eth_dev_callback_process(eth_dev, RTE_ETH_EVENT_INTR_LSC, NULL);
 }
 
 static void ena_notification(void *data,