net/mlx4: add external allocator for Verbs object

To support secondary process, the memory allocated by library such as
completion rings (CQ) and buffer rings (WQ) must be manageable by EAL,
in order to share it with secondary processes. With new changes in
rdma-core and kernel driver, it is possible to provide an external
allocator to the library layer for this purpose. All such resources
will now be allocated within DPDK framework.

Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
Acked-by: Shahaf Shuler <shahafs@mellanox.com>

drivers/net/mlx4/Makefile

@@ -87,6 +87,11 @@ mlx4_autoconf.h.new: FORCE
 mlx4_autoconf.h.new: $(RTE_SDK)/buildtools/auto-config-h.sh
 	$Q $(RM) -f -- '$@'
 	$Q : > '$@'
+	$Q sh -- '$<' '$@' \
+		HAVE_IBV_MLX4_BUF_ALLOCATORS \
+		infiniband/mlx4dv.h \
+		enum MLX4DV_SET_CTX_ATTR_BUF_ALLOCATORS \
+		$(AUTOCONF_OUTPUT)
 	$Q sh -- '$<' '$@' \
 		HAVE_IBV_MLX4_WQE_LSO_SEG \
 		infiniband/mlx4dv.h \

drivers/net/mlx4/meson.build

@@ -70,7 +70,17 @@ if build
 		[ 'HAVE_IBV_MLX4_WQE_LSO_SEG', 'infiniband/mlx4dv.h',
 		'struct mlx4_wqe_lso_seg', 'mss_hdr_size' ],
 	]
+	# input array for meson symbol search:
+	# [ "MACRO to define if found", "header for the search",
+	#   "symbol to search" ]
+	has_sym_args = [
+		[ 'HAVE_IBV_MLX4_BUF_ALLOCATORS', 'infiniband/mlx4dv.h',
+		'MLX4DV_SET_CTX_ATTR_BUF_ALLOCATORS' ],
+	]
 	config = configuration_data()
+	foreach arg:has_sym_args
+		config.set(arg[0], cc.has_header_symbol(arg[1], arg[2]))
+	endforeach
 	foreach arg:has_member_args
 		file_prefix = '#include<' + arg[1] + '>'
 		config.set(arg[0], cc.has_member(arg[2], arg[3],

drivers/net/mlx4/mlx4.c

@@ -69,6 +69,62 @@ const char *pmd_mlx4_init_params[] = {
 
 static void mlx4_dev_stop(struct rte_eth_dev *dev);
 
+#ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS
+/**
+ * Verbs callback to allocate a memory. This function should allocate the space
+ * according to the size provided residing inside a huge page.
+ * Please note that all allocation must respect the alignment from libmlx4
+ * (i.e. currently sysconf(_SC_PAGESIZE)).
+ *
+ * @param[in] size
+ *   The size in bytes of the memory to allocate.
+ * @param[in] data
+ *   A pointer to the callback data.
+ *
+ * @return
+ *   Allocated buffer, NULL otherwise and rte_errno is set.
+ */
+static void *
+mlx4_alloc_verbs_buf(size_t size, void *data)
+{
+	struct mlx4_priv *priv = data;
+	void *ret;
+	size_t alignment = sysconf(_SC_PAGESIZE);
+	unsigned int socket = SOCKET_ID_ANY;
+
+	if (priv->verbs_alloc_ctx.type == MLX4_VERBS_ALLOC_TYPE_TX_QUEUE) {
+		const struct txq *txq = priv->verbs_alloc_ctx.obj;
+
+		socket = txq->socket;
+	} else if (priv->verbs_alloc_ctx.type ==
+		   MLX4_VERBS_ALLOC_TYPE_RX_QUEUE) {
+		const struct rxq *rxq = priv->verbs_alloc_ctx.obj;
+
+		socket = rxq->socket;
+	}
+	assert(data != NULL);
+	ret = rte_malloc_socket(__func__, size, alignment, socket);
+	if (!ret && size)
+		rte_errno = ENOMEM;
+	return ret;
+}
+
+/**
+ * Verbs callback to free a memory.
+ *
+ * @param[in] ptr
+ *   A pointer to the memory to free.
+ * @param[in] data
+ *   A pointer to the callback data.
+ */
+static void
+mlx4_free_verbs_buf(void *ptr, void *data __rte_unused)
+{
+	assert(data != NULL);
+	rte_free(ptr);
+}
+#endif
+
 /**
  * DPDK callback for Ethernet device configuration.
  *
@@ -755,6 +811,17 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
 		eth_dev->intr_handle = &priv->intr_handle;
 		priv->dev_data = eth_dev->data;
 		eth_dev->dev_ops = &mlx4_dev_ops;
+#ifdef HAVE_IBV_MLX4_BUF_ALLOCATORS
+		/* Hint libmlx4 to use PMD allocator for data plane resources */
+		struct mlx4dv_ctx_allocators alctr = {
+			.alloc = &mlx4_alloc_verbs_buf,
+			.free = &mlx4_free_verbs_buf,
+			.data = priv,
+		};
+		mlx4_glue->dv_set_context_attr
+			(ctx, MLX4DV_SET_CTX_ATTR_BUF_ALLOCATORS,
+			 (void *)((uintptr_t)&alctr));
+#endif
 		/* Bring Ethernet device up. */
 		DEBUG("forcing Ethernet interface up");
 		mlx4_dev_set_link_up(eth_dev);

drivers/net/mlx4/mlx4.h

@@ -72,6 +72,24 @@ struct rxq;
 struct txq;
 struct rte_flow;
 
+/**
+ * Type of objet being allocated.
+ */
+enum mlx4_verbs_alloc_type {
+	MLX4_VERBS_ALLOC_TYPE_NONE,
+	MLX4_VERBS_ALLOC_TYPE_TX_QUEUE,
+	MLX4_VERBS_ALLOC_TYPE_RX_QUEUE,
+};
+
+/**
+ * Verbs allocator needs a context to know in the callback which kind of
+ * resources it is allocating.
+ */
+struct mlx4_verbs_alloc_ctx {
+	enum mlx4_verbs_alloc_type type; /* Kind of object being allocated. */
+	const void *obj; /* Pointer to the DPDK object. */
+};
+
 LIST_HEAD(mlx4_dev_list, mlx4_priv);
 LIST_HEAD(mlx4_mr_list, mlx4_mr);
 
@@ -111,6 +129,8 @@ struct mlx4_priv {
 	LIST_HEAD(, rte_flow) flows; /**< Configured flow rule handles. */
 	struct ether_addr mac[MLX4_MAX_MAC_ADDRESSES];
 	/**< Configured MAC addresses. Unused entries are zeroed. */
+	struct mlx4_verbs_alloc_ctx verbs_alloc_ctx;
+	/**< Context for Verbs allocator. */
 };
 
 #define PORT_ID(priv) ((priv)->dev_data->port_id)

drivers/net/mlx4/mlx4_rxq.c

@@ -513,6 +513,8 @@ mlx4_rxq_attach(struct rxq *rxq)
 	int ret;
 
 	assert(rte_is_power_of_2(elts_n));
+	priv->verbs_alloc_ctx.type = MLX4_VERBS_ALLOC_TYPE_RX_QUEUE;
+	priv->verbs_alloc_ctx.obj = rxq;
 	cq = mlx4_glue->create_cq(priv->ctx, elts_n / sges_n, NULL,
 				  rxq->channel, 0);
 	if (!cq) {
@@ -620,6 +622,7 @@ mlx4_rxq_attach(struct rxq *rxq)
 	rxq->rq_ci = elts_n / sges_n;
 	rte_wmb();
 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
+	priv->verbs_alloc_ctx.type = MLX4_VERBS_ALLOC_TYPE_NONE;
 	return 0;
 error:
 	if (wq)
@@ -630,6 +633,7 @@ mlx4_rxq_attach(struct rxq *rxq)
 	rte_errno = ret;
 	ERROR("error while attaching Rx queue %p: %s: %s",
 	      (void *)rxq, msg, strerror(ret));
+	priv->verbs_alloc_ctx.type = MLX4_VERBS_ALLOC_TYPE_NONE;
 	return -ret;
 }
 

drivers/net/mlx4/mlx4_txq.c

@@ -177,10 +177,8 @@ mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 	uint64_t offloads;
 
 	offloads = conf->offloads | dev->data->dev_conf.txmode.offloads;
-
 	DEBUG("%p: configuring queue %u for %u descriptors",
 	      (void *)dev, idx, desc);
-
 	if (idx >= dev->data->nb_tx_queues) {
 		rte_errno = EOVERFLOW;
 		ERROR("%p: queue index out of range (%u >= %u)",
@@ -241,6 +239,8 @@ mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 		.lb = !!priv->vf,
 		.bounce_buf = bounce_buf,
 	};
+	priv->verbs_alloc_ctx.type = MLX4_VERBS_ALLOC_TYPE_TX_QUEUE;
+	priv->verbs_alloc_ctx.obj = txq;
 	txq->cq = mlx4_glue->create_cq(priv->ctx, desc, NULL, NULL, 0);
 	if (!txq->cq) {
 		rte_errno = ENOMEM;
@@ -331,6 +331,7 @@ mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 	txq->mr_ctrl.dev_gen_ptr = &priv->mr.dev_gen;
 	DEBUG("%p: adding Tx queue %p to list", (void *)dev, (void *)txq);
 	dev->data->tx_queues[idx] = txq;
+	priv->verbs_alloc_ctx.type = MLX4_VERBS_ALLOC_TYPE_NONE;
 	return 0;
 error:
 	dev->data->tx_queues[idx] = NULL;
@@ -338,6 +339,7 @@ mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 	mlx4_tx_queue_release(txq);
 	rte_errno = ret;
 	assert(rte_errno > 0);
+	priv->verbs_alloc_ctx.type = MLX4_VERBS_ALLOC_TYPE_NONE;
 	return -rte_errno;
 }