net/virtio: fix Rx and Tx handler selection for ARM32
[dpdk.git] / drivers / net / mlx5 / mlx5.c
index ca44a0a..5a959f7 100644 (file)
@@ -39,6 +39,7 @@
 #include <stdlib.h>
 #include <errno.h>
 #include <net/if.h>
+#include <sys/mman.h>
 
 /* Verbs header. */
 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
 #endif
 
 #include <rte_malloc.h>
-#include <rte_ethdev.h>
+#include <rte_ethdev_driver.h>
 #include <rte_ethdev_pci.h>
 #include <rte_pci.h>
 #include <rte_bus_pci.h>
 #include <rte_common.h>
+#include <rte_eal_memconfig.h>
 #include <rte_kvargs.h>
 
 #include "mlx5.h"
@@ -85,9 +87,6 @@
 /* Device parameter to limit the size of inlining packet. */
 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
 
-/* Device parameter to enable hardware TSO offload. */
-#define MLX5_TSO "tso"
-
 /* Device parameter to enable hardware Tx vector. */
 #define MLX5_TX_VEC_EN "tx_vec_en"
 
@@ -142,10 +141,20 @@ mlx5_alloc_verbs_buf(size_t size, void *data)
        struct priv *priv = data;
        void *ret;
        size_t alignment = sysconf(_SC_PAGESIZE);
+       unsigned int socket = SOCKET_ID_ANY;
+
+       if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
+               const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 
+               socket = ctrl->socket;
+       } else if (priv->verbs_alloc_ctx.type ==
+                  MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
+               const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
+
+               socket = ctrl->socket;
+       }
        assert(data != NULL);
-       ret = rte_malloc_socket(__func__, size, alignment,
-                               priv->dev->device->numa_node);
+       ret = rte_malloc_socket(__func__, size, alignment, socket);
        DEBUG("Extern alloc size: %lu, align: %lu: %p", size, alignment, ret);
        return ret;
 }
@@ -287,6 +296,7 @@ const struct eth_dev_ops mlx5_dev_ops = {
        .tx_descriptor_status = mlx5_tx_descriptor_status,
        .rx_queue_intr_enable = mlx5_rx_intr_enable,
        .rx_queue_intr_disable = mlx5_rx_intr_disable,
+       .is_removed = mlx5_is_removed,
 };
 
 static const struct eth_dev_ops mlx5_dev_sec_ops = {
@@ -334,6 +344,7 @@ const struct eth_dev_ops mlx5_dev_ops_isolate = {
        .tx_descriptor_status = mlx5_tx_descriptor_status,
        .rx_queue_intr_enable = mlx5_rx_intr_enable,
        .rx_queue_intr_disable = mlx5_rx_intr_disable,
+       .is_removed = mlx5_is_removed,
 };
 
 static struct {
@@ -406,8 +417,6 @@ mlx5_args_check(const char *key, const char *val, void *opaque)
                config->mpw_hdr_dseg = !!tmp;
        } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
                config->inline_max_packet_sz = tmp;
-       } else if (strcmp(MLX5_TSO, key) == 0) {
-               config->tso = !!tmp;
        } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
                config->tx_vec_en = !!tmp;
        } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
@@ -440,7 +449,6 @@ mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
                MLX5_TXQ_MPW_EN,
                MLX5_TXQ_MPW_HDR_DSEG_EN,
                MLX5_TXQ_MAX_INLINE_LEN,
-               MLX5_TSO,
                MLX5_TX_VEC_EN,
                MLX5_RX_VEC_EN,
                NULL,
@@ -472,6 +480,106 @@ mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
 
 static struct rte_pci_driver mlx5_driver;
 
+/*
+ * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
+ * local resource used by both primary and secondary to avoid duplicate
+ * reservation.
+ * The space has to be available on both primary and secondary process,
+ * TXQ UAR maps to this area using fixed mmap w/o double check.
+ */
+static void *uar_base;
+
+/**
+ * Reserve UAR address space for primary process.
+ *
+ * @param[in] priv
+ *   Pointer to private structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+priv_uar_init_primary(struct priv *priv)
+{
+       void *addr = (void *)0;
+       int i;
+       const struct rte_mem_config *mcfg;
+       int ret;
+
+       if (uar_base) { /* UAR address space mapped. */
+               priv->uar_base = uar_base;
+               return 0;
+       }
+       /* find out lower bound of hugepage segments */
+       mcfg = rte_eal_get_configuration()->mem_config;
+       for (i = 0; i < RTE_MAX_MEMSEG && mcfg->memseg[i].addr; i++) {
+               if (addr)
+                       addr = RTE_MIN(addr, mcfg->memseg[i].addr);
+               else
+                       addr = mcfg->memseg[i].addr;
+       }
+       /* keep distance to hugepages to minimize potential conflicts. */
+       addr = RTE_PTR_SUB(addr, MLX5_UAR_OFFSET + MLX5_UAR_SIZE);
+       /* anonymous mmap, no real memory consumption. */
+       addr = mmap(addr, MLX5_UAR_SIZE,
+                   PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+       if (addr == MAP_FAILED) {
+               ERROR("Failed to reserve UAR address space, please adjust "
+                     "MLX5_UAR_SIZE or try --base-virtaddr");
+               ret = ENOMEM;
+               return ret;
+       }
+       /* Accept either same addr or a new addr returned from mmap if target
+        * range occupied.
+        */
+       INFO("Reserved UAR address space: %p", addr);
+       priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
+       uar_base = addr; /* process local, don't reserve again. */
+       return 0;
+}
+
+/**
+ * Reserve UAR address space for secondary process, align with
+ * primary process.
+ *
+ * @param[in] priv
+ *   Pointer to private structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+priv_uar_init_secondary(struct priv *priv)
+{
+       void *addr;
+       int ret;
+
+       assert(priv->uar_base);
+       if (uar_base) { /* already reserved. */
+               assert(uar_base == priv->uar_base);
+               return 0;
+       }
+       /* anonymous mmap, no real memory consumption. */
+       addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
+                   PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+       if (addr == MAP_FAILED) {
+               ERROR("UAR mmap failed: %p size: %llu",
+                     priv->uar_base, MLX5_UAR_SIZE);
+               ret = ENXIO;
+               return ret;
+       }
+       if (priv->uar_base != addr) {
+               ERROR("UAR address %p size %llu occupied, please adjust "
+                     "MLX5_UAR_OFFSET or try EAL parameter --base-virtaddr",
+                     priv->uar_base, MLX5_UAR_SIZE);
+               ret = ENXIO;
+               return ret;
+       }
+       uar_base = addr; /* process local, don't reserve again */
+       INFO("Reserved UAR address space: %p", addr);
+       return 0;
+}
+
 /**
  * DPDK callback to register a PCI device.
  *
@@ -614,6 +722,8 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
        INFO("%u port(s) detected", device_attr.orig_attr.phys_port_cnt);
 
        for (i = 0; i < device_attr.orig_attr.phys_port_cnt; i++) {
+               char name[RTE_ETH_NAME_MAX_LEN];
+               int len;
                uint32_t port = i + 1; /* ports are indexed from one */
                uint32_t test = (1 << i);
                struct ibv_context *ctx = NULL;
@@ -629,7 +739,6 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                        .cqe_comp = cqe_comp,
                        .mps = mps,
                        .tunnel_en = tunnel_en,
-                       .tso = 0,
                        .tx_vec_en = 1,
                        .rx_vec_en = 1,
                        .mpw_hdr_dseg = 0,
@@ -638,14 +747,15 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                        .inline_max_packet_sz = MLX5_ARG_UNSET,
                };
 
+               len = snprintf(name, sizeof(name), PCI_PRI_FMT,
+                        pci_dev->addr.domain, pci_dev->addr.bus,
+                        pci_dev->addr.devid, pci_dev->addr.function);
+               if (device_attr.orig_attr.phys_port_cnt > 1)
+                       snprintf(name + len, sizeof(name), " port %u", i);
+
                mlx5_dev[idx].ports |= test;
 
                if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
-                       /* from rte_ethdev.c */
-                       char name[RTE_ETH_NAME_MAX_LEN];
-
-                       snprintf(name, sizeof(name), "%s port %u",
-                                ibv_get_device_name(ibv_dev), port);
                        eth_dev = rte_eth_dev_attach_secondary(name);
                        if (eth_dev == NULL) {
                                ERROR("can not attach rte ethdev");
@@ -655,6 +765,11 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                        eth_dev->device = &pci_dev->device;
                        eth_dev->dev_ops = &mlx5_dev_sec_ops;
                        priv = eth_dev->data->dev_private;
+                       err = priv_uar_init_secondary(priv);
+                       if (err < 0) {
+                               err = -err;
+                               goto error;
+                       }
                        /* Receive command fd from primary process */
                        err = priv_socket_connect(priv);
                        if (err < 0) {
@@ -663,10 +778,8 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                        }
                        /* Remap UAR for Tx queues. */
                        err = priv_tx_uar_remap(priv, err);
-                       if (err < 0) {
-                               err = -err;
+                       if (err)
                                goto error;
-                       }
                        /*
                         * Ethdev pointer is still required as input since
                         * the primary device is not accessible from the
@@ -793,10 +906,9 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
 
                priv_get_num_vfs(priv, &num_vfs);
                config.sriov = (num_vfs || sriov);
-               if (config.tso)
-                       config.tso = ((device_attr_ex.tso_caps.max_tso > 0) &&
-                                     (device_attr_ex.tso_caps.supported_qpts &
-                                     (1 << IBV_QPT_RAW_PACKET)));
+               config.tso = ((device_attr_ex.tso_caps.max_tso > 0) &&
+                             (device_attr_ex.tso_caps.supported_qpts &
+                             (1 << IBV_QPT_RAW_PACKET)));
                if (config.tso)
                        config.tso_max_payload_sz =
                                        device_attr_ex.tso_caps.max_tso;
@@ -805,10 +917,6 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                              " (" MLX5_TXQ_MPW_EN ")");
                        err = ENOTSUP;
                        goto port_error;
-               } else if (config.mps && config.tso) {
-                       WARN("multi-packet send not supported in conjunction "
-                             "with TSO. MPS disabled");
-                       config.mps = 0;
                }
                INFO("%sMPS is %s",
                     config.mps == MLX5_MPW_ENHANCED ? "Enhanced " : "",
@@ -817,6 +925,9 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                        WARN("Rx CQE compression isn't supported");
                        config.cqe_comp = 0;
                }
+               err = priv_uar_init_primary(priv);
+               if (err)
+                       goto port_error;
                /* Configure the first MAC address by default. */
                if (priv_get_mac(priv, &mac.addr_bytes)) {
                        ERROR("cannot get MAC address, is mlx5_en loaded?"
@@ -844,14 +955,7 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                priv_get_mtu(priv, &priv->mtu);
                DEBUG("port %u MTU is %u", priv->port, priv->mtu);
 
-               /* from rte_ethdev.c */
-               {
-                       char name[RTE_ETH_NAME_MAX_LEN];
-
-                       snprintf(name, sizeof(name), "%s port %u",
-                                ibv_get_device_name(ibv_dev), port);
-                       eth_dev = rte_eth_dev_allocate(name);
-               }
+               eth_dev = rte_eth_dev_allocate(name);
                if (eth_dev == NULL) {
                        ERROR("can not allocate rte ethdev");
                        err = ENOMEM;
@@ -862,6 +966,11 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                eth_dev->device = &pci_dev->device;
                rte_eth_copy_pci_info(eth_dev, pci_dev);
                eth_dev->device->driver = &mlx5_driver.driver;
+               /*
+                * Initialize burst functions to prevent crashes before link-up.
+                */
+               eth_dev->rx_pkt_burst = removed_rx_burst;
+               eth_dev->tx_pkt_burst = removed_tx_burst;
                priv->dev = eth_dev;
                eth_dev->dev_ops = &mlx5_dev_ops;
                /* Register MAC address. */
@@ -881,7 +990,6 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                /* Bring Ethernet device up. */
                DEBUG("forcing Ethernet interface up");
                priv_set_flags(priv, ~IFF_UP, IFF_UP);
-               mlx5_link_update(priv->dev, 1);
                /* Store device configuration on private structure. */
                priv->config = config;
                continue;