net/mlx5: SF, Use all available cpu for setting cpu affinity

Currently all SFs are using the same CPUs. Spreading SF over CPUs, in
round-robin manner, in order to achieve better distribution of the SFs
over available CPUs.

Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Moshe Shemesh <moshe@nvidia.com>
Reviewed-by: Parav Pandit <parav@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>

5 files changed, 201 insertions, 34 deletions

diff --git a/drivers/net/ethernet/mellanox/mlx5/core/eq.c b/drivers/net/ethernet/mellanox/mlx5/core/eq.c
index 14547b6f2894..48a45aa54a3c 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/eq.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/eq.c
@@ -798,7 +798,10 @@ static void comp_irqs_release(struct mlx5_core_dev *dev)
 {
 	struct mlx5_eq_table *table = dev->priv.eq_table;
 
-	mlx5_irqs_release_vectors(table->comp_irqs, table->num_comp_eqs);
+	if (mlx5_core_is_sf(dev))
+		mlx5_irq_affinity_irqs_release(dev, table->comp_irqs, table->num_comp_eqs);
+	else
+		mlx5_irqs_release_vectors(table->comp_irqs, table->num_comp_eqs);
 	kfree(table->comp_irqs);
 }
 
@@ -814,6 +817,12 @@ static int comp_irqs_request(struct mlx5_core_dev *dev)
 	table->comp_irqs = kcalloc(ncomp_eqs, sizeof(*table->comp_irqs), GFP_KERNEL);
 	if (!table->comp_irqs)
 		return -ENOMEM;
+	if (mlx5_core_is_sf(dev)) {
+		ret = mlx5_irq_affinity_irqs_request_auto(dev, ncomp_eqs, table->comp_irqs);
+		if (ret < 0)
+			goto free_irqs;
+		return ret;
+	}
 
 	cpus = kcalloc(ncomp_eqs, sizeof(*cpus), GFP_KERNEL);
 	if (!cpus) {
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/irq_affinity.c b/drivers/net/ethernet/mellanox/mlx5/core/irq_affinity.c
index 4ff0af0fc58a..380a208ab137 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/irq_affinity.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/irq_affinity.c
@@ -5,21 +5,81 @@
 #include "mlx5_irq.h"
 #include "pci_irq.h"
 
+static void cpu_put(struct mlx5_irq_pool *pool, int cpu)
+{
+	pool->irqs_per_cpu[cpu]--;
+}
+
+static void cpu_get(struct mlx5_irq_pool *pool, int cpu)
+{
+	pool->irqs_per_cpu[cpu]++;
+}
+
+/* Gets the least loaded CPU. e.g.: the CPU with least IRQs bound to it */
+static int cpu_get_least_loaded(struct mlx5_irq_pool *pool,
+				const struct cpumask *req_mask)
+{
+	int best_cpu = -1;
+	int cpu;
+
+	for_each_cpu_and(cpu, req_mask, cpu_online_mask) {
+		/* CPU has zero IRQs on it. No need to search any more CPUs. */
+		if (!pool->irqs_per_cpu[cpu]) {
+			best_cpu = cpu;
+			break;
+		}
+		if (best_cpu < 0)
+			best_cpu = cpu;
+		if (pool->irqs_per_cpu[cpu] < pool->irqs_per_cpu[best_cpu])
+			best_cpu = cpu;
+	}
+	if (best_cpu == -1) {
+		/* There isn't online CPUs in req_mask */
+		mlx5_core_err(pool->dev, "NO online CPUs in req_mask (%*pbl)\n",
+			      cpumask_pr_args(req_mask));
+		best_cpu = cpumask_first(cpu_online_mask);
+	}
+	pool->irqs_per_cpu[best_cpu]++;
+	return best_cpu;
+}
+
 /* Creating an IRQ from irq_pool */
 static struct mlx5_irq *
 irq_pool_request_irq(struct mlx5_irq_pool *pool, const struct cpumask *req_mask)
 {
+	cpumask_var_t auto_mask;
+	struct mlx5_irq *irq;
 	u32 irq_index;
 	int err;
 
-	err = xa_alloc(&pool->irqs, &irq_index, NULL, pool->xa_num_irqs,
-		       GFP_KERNEL);
+	if (!zalloc_cpumask_var(&auto_mask, GFP_KERNEL))
+		return ERR_PTR(-ENOMEM);
+	err = xa_alloc(&pool->irqs, &irq_index, NULL, pool->xa_num_irqs, GFP_KERNEL);
 	if (err)
 		return ERR_PTR(err);
-	return mlx5_irq_alloc(pool, irq_index, req_mask);
+	if (pool->irqs_per_cpu) {
+		if (cpumask_weight(req_mask) > 1)
+			/* if req_mask contain more then one CPU, set the least loadad CPU
+			 * of req_mask
+			 */
+			cpumask_set_cpu(cpu_get_least_loaded(pool, req_mask), auto_mask);
+		else
+			cpu_get(pool, cpumask_first(req_mask));
+	}
+	irq = mlx5_irq_alloc(pool, irq_index, cpumask_empty(auto_mask) ? req_mask : auto_mask);
+	free_cpumask_var(auto_mask);
+	return irq;
 }
 
-/* Looking for the IRQ with the smallest refcount and the same mask */
+/* Looking for the IRQ with the smallest refcount that fits req_mask.
+ * If pool is sf_comp_pool, then we are looking for an IRQ with any of the
+ * requested CPUs in req_mask.
+ * for example: req_mask = 0xf, irq0_mask = 0x10, irq1_mask = 0x1. irq0_mask
+ * isn't subset of req_mask, so we will skip it. irq1_mask is subset of req_mask,
+ * we don't skip it.
+ * If pool is sf_ctrl_pool, then all IRQs have the same mask, so any IRQ will
+ * fit. And since mask is subset of itself, we will pass the first if bellow.
+ */
 static struct mlx5_irq *
 irq_pool_find_least_loaded(struct mlx5_irq_pool *pool, const struct cpumask *req_mask)
 {
@@ -35,8 +95,8 @@ irq_pool_find_least_loaded(struct mlx5_irq_pool *pool, const struct cpumask *req
 		struct cpumask *iter_mask = mlx5_irq_get_affinity_mask(iter);
 		int iter_refcount = mlx5_irq_read_locked(iter);
 
-		if (!cpumask_equal(iter_mask, req_mask))
-			/* If a user request a mask, skip IRQs that's aren't a match */
+		if (!cpumask_subset(iter_mask, req_mask))
+			/* skip IRQs with a mask which is not subset of req_mask */
 			continue;
 		if (iter_refcount < pool->min_threshold)
 			/* If we found an IRQ with less than min_thres, return it */
@@ -97,3 +157,70 @@ unlock:
 	mutex_unlock(&pool->lock);
 	return least_loaded_irq;
 }
+
+void mlx5_irq_affinity_irqs_release(struct mlx5_core_dev *dev, struct mlx5_irq **irqs,
+				    int num_irqs)
+{
+	struct mlx5_irq_pool *pool = mlx5_irq_pool_get(dev);
+	int i;
+
+	for (i = 0; i < num_irqs; i++) {
+		int cpu = cpumask_first(mlx5_irq_get_affinity_mask(irqs[i]));
+
+		synchronize_irq(pci_irq_vector(pool->dev->pdev,
+					       mlx5_irq_get_index(irqs[i])));
+		if (mlx5_irq_put(irqs[i]))
+			if (pool->irqs_per_cpu)
+				cpu_put(pool, cpu);
+	}
+}
+
+/**
+ * mlx5_irq_affinity_irqs_request_auto - request one or more IRQs for mlx5 device.
+ * @dev: mlx5 device that is requesting the IRQs.
+ * @nirqs: number of IRQs to request.
+ * @irqs: an output array of IRQs pointers.
+ *
+ * Each IRQ is bounded to at most 1 CPU.
+ * This function is requesting IRQs according to the default assignment.
+ * The default assignment policy is:
+ * - in each iteration, request the least loaded IRQ which is not bound to any
+ *   CPU of the previous IRQs requested.
+ *
+ * This function returns the number of IRQs requested, (which might be smaller than
+ * @nirqs), if successful, or a negative error code in case of an error.
+ */
+int mlx5_irq_affinity_irqs_request_auto(struct mlx5_core_dev *dev, int nirqs,
+					struct mlx5_irq **irqs)
+{
+	struct mlx5_irq_pool *pool = mlx5_irq_pool_get(dev);
+	cpumask_var_t req_mask;
+	struct mlx5_irq *irq;
+	int i = 0;
+
+	if (!zalloc_cpumask_var(&req_mask, GFP_KERNEL))
+		return -ENOMEM;
+	cpumask_copy(req_mask, cpu_online_mask);
+	for (i = 0; i < nirqs; i++) {
+		if (mlx5_irq_pool_is_sf_pool(pool))
+			irq = mlx5_irq_affinity_request(pool, req_mask);
+		else
+			/* In case SF pool doesn't exists, fallback to the PF IRQs.
+			 * The PF IRQs are already allocated and binded to CPU
+			 * at this point. Hence, only an index is needed.
+			 */
+			irq = mlx5_irq_request(dev, i, NULL);
+		if (IS_ERR(irq))
+			break;
+		irqs[i] = irq;
+		cpumask_clear_cpu(cpumask_first(mlx5_irq_get_affinity_mask(irq)), req_mask);
+		mlx5_core_dbg(pool->dev, "IRQ %u mapped to cpu %*pbl, %u EQs on this irq\n",
+			      pci_irq_vector(dev->pdev, mlx5_irq_get_index(irq)),
+			      cpumask_pr_args(mlx5_irq_get_affinity_mask(irq)),
+			      mlx5_irq_read_locked(irq) / MLX5_EQ_REFS_PER_IRQ);
+	}
+	free_cpumask_var(req_mask);
+	if (!i)
+		return PTR_ERR(irq);
+	return i;
+}
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/mlx5_irq.h b/drivers/net/ethernet/mellanox/mlx5/core/mlx5_irq.h
index f2ad3c21802f..23cb63fa4588 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/mlx5_irq.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/mlx5_irq.h
@@ -36,13 +36,26 @@ int mlx5_irq_get_index(struct mlx5_irq *irq);
 
 struct mlx5_irq_pool;
 #ifdef CONFIG_MLX5_SF
+int mlx5_irq_affinity_irqs_request_auto(struct mlx5_core_dev *dev, int nirqs,
+					struct mlx5_irq **irqs);
 struct mlx5_irq *mlx5_irq_affinity_request(struct mlx5_irq_pool *pool,
 					   const struct cpumask *req_mask);
+void mlx5_irq_affinity_irqs_release(struct mlx5_core_dev *dev, struct mlx5_irq **irqs,
+				    int num_irqs);
 #else
+static inline int mlx5_irq_affinity_irqs_request_auto(struct mlx5_core_dev *dev, int nirqs,
+						      struct mlx5_irq **irqs)
+{
+	return -EOPNOTSUPP;
+}
+
 static inline struct mlx5_irq *
 mlx5_irq_affinity_request(struct mlx5_irq_pool *pool, const struct cpumask *req_mask)
 {
 	return ERR_PTR(-EOPNOTSUPP);
 }
+
+static inline void mlx5_irq_affinity_irqs_release(struct mlx5_core_dev *dev,
+						  struct mlx5_irq **irqs, int num_irqs) {}
 #endif
 #endif /* __MLX5_IRQ_H__ */
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/pci_irq.c b/drivers/net/ethernet/mellanox/mlx5/core/pci_irq.c
index 2e19c3c222fe..90fec0649ef5 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/pci_irq.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/pci_irq.c
@@ -139,15 +139,19 @@ static void irq_release(struct mlx5_irq *irq)
 	kfree(irq);
 }
 
-static void irq_put(struct mlx5_irq *irq)
+int mlx5_irq_put(struct mlx5_irq *irq)
 {
 	struct mlx5_irq_pool *pool = irq->pool;
+	int ret = 0;
 
 	mutex_lock(&pool->lock);
 	irq->refcount--;
-	if (!irq->refcount)
+	if (!irq->refcount) {
 		irq_release(irq);
+		ret = 1;
+	}
 	mutex_unlock(&pool->lock);
+	return ret;
 }
 
 int mlx5_irq_read_locked(struct mlx5_irq *irq)
@@ -202,11 +206,6 @@ static void irq_set_name(struct mlx5_irq_pool *pool, char *name, int vecidx)
 	snprintf(name, MLX5_MAX_IRQ_NAME, "mlx5_comp%d", vecidx);
 }
 
-static bool irq_pool_is_sf_pool(struct mlx5_irq_pool *pool)
-{
-	return !strncmp("mlx5_sf", pool->name, strlen("mlx5_sf"));
-}
-
 struct mlx5_irq *mlx5_irq_alloc(struct mlx5_irq_pool *pool, int i,
 				const struct cpumask *affinity)
 {
@@ -219,7 +218,7 @@ struct mlx5_irq *mlx5_irq_alloc(struct mlx5_irq_pool *pool, int i,
 	if (!irq)
 		return ERR_PTR(-ENOMEM);
 	irq->irqn = pci_irq_vector(dev->pdev, i);
-	if (!irq_pool_is_sf_pool(pool))
+	if (!mlx5_irq_pool_is_sf_pool(pool))
 		irq_set_name(pool, name, i);
 	else
 		irq_sf_set_name(pool, name, i);
@@ -273,7 +272,7 @@ int mlx5_irq_attach_nb(struct mlx5_irq *irq, struct notifier_block *nb)
 		return -ENOENT;
 	ret = atomic_notifier_chain_register(&irq->nh, nb);
 	if (ret)
-		irq_put(irq);
+		mlx5_irq_put(irq);
 	return ret;
 }
 
@@ -282,7 +281,7 @@ int mlx5_irq_detach_nb(struct mlx5_irq *irq, struct notifier_block *nb)
 	int err = 0;
 
 	err = atomic_notifier_chain_unregister(&irq->nh, nb);
-	irq_put(irq);
+	mlx5_irq_put(irq);
 	return err;
 }
 
@@ -327,6 +326,20 @@ static struct mlx5_irq_pool *sf_irq_pool_get(struct mlx5_irq_table *irq_table)
 	return irq_table->sf_comp_pool;
 }
 
+struct mlx5_irq_pool *mlx5_irq_pool_get(struct mlx5_core_dev *dev)
+{
+	struct mlx5_irq_table *irq_table = mlx5_irq_table_get(dev);
+	struct mlx5_irq_pool *pool = NULL;
+
+	if (mlx5_core_is_sf(dev))
+		pool = sf_irq_pool_get(irq_table);
+
+	/* In some configs, there won't be a pool of SFs IRQs. Hence, returning
+	 * the PF IRQs pool in case the SF pool doesn't exist.
+	 */
+	return pool ? pool : irq_table->pf_pool;
+}
+
 static struct mlx5_irq_pool *ctrl_irq_pool_get(struct mlx5_core_dev *dev)
 {
 	struct mlx5_irq_table *irq_table = mlx5_irq_table_get(dev);
@@ -352,7 +365,7 @@ static void mlx5_irqs_release(struct mlx5_irq **irqs, int nirqs)
 
 	for (i = 0; i < nirqs; i++) {
 		synchronize_irq(irqs[i]->irqn);
-		irq_put(irqs[i]);
+		mlx5_irq_put(irqs[i]);
 	}
 }
 
@@ -380,7 +393,7 @@ struct mlx5_irq *mlx5_ctrl_irq_request(struct mlx5_core_dev *dev)
 	if (!zalloc_cpumask_var(&req_mask, GFP_KERNEL))
 		return ERR_PTR(-ENOMEM);
 	cpumask_copy(req_mask, cpu_online_mask);
-	if (!irq_pool_is_sf_pool(pool)) {
+	if (!mlx5_irq_pool_is_sf_pool(pool)) {
 		/* In case we are allocating a control IRQ for PF/VF */
 		if (!pool->xa_num_irqs.max) {
 			cpumask_clear(req_mask);
@@ -398,7 +411,7 @@ struct mlx5_irq *mlx5_ctrl_irq_request(struct mlx5_core_dev *dev)
 }
 
 /**
- * mlx5_irq_request - request an IRQ for mlx5 device.
+ * mlx5_irq_request - request an IRQ for mlx5 PF/VF device.
  * @dev: mlx5 device that requesting the IRQ.
  * @vecidx: vector index of the IRQ. This argument is ignore if affinity is
  * provided.
@@ -413,22 +426,8 @@ struct mlx5_irq *mlx5_irq_request(struct mlx5_core_dev *dev, u16 vecidx,
 	struct mlx5_irq_pool *pool;
 	struct mlx5_irq *irq;
 
-	if (mlx5_core_is_sf(dev)) {
-		pool = sf_irq_pool_get(irq_table);
-		if (!pool)
-			/* we don't have IRQs for SFs, using the PF IRQs */
-			goto pf_irq;
-		if (cpumask_empty(affinity) && !strcmp(pool->name, "mlx5_sf_comp"))
-			/* In case an SF user request IRQ with vecidx */
-			irq = irq_pool_request_vector(pool, vecidx, NULL);
-		else
-			irq = mlx5_irq_affinity_request(pool, affinity);
-		goto out;
-	}
-pf_irq:
 	pool = irq_table->pf_pool;
 	irq = irq_pool_request_vector(pool, vecidx, affinity);
-out:
 	if (IS_ERR(irq))
 		return irq;
 	mlx5_core_dbg(dev, "irq %u mapped to cpu %*pbl, %u EQs on this irq\n",
@@ -518,6 +517,7 @@ static void irq_pool_free(struct mlx5_irq_pool *pool)
 		irq_release(irq);
 	xa_destroy(&pool->irqs);
 	mutex_destroy(&pool->lock);
+	kfree(pool->irqs_per_cpu);
 	kvfree(pool);
 }
 
@@ -565,7 +565,17 @@ static int irq_pools_init(struct mlx5_core_dev *dev, int sf_vec, int pf_vec)
 		err = PTR_ERR(table->sf_comp_pool);
 		goto err_sf_ctrl;
 	}
+
+	table->sf_comp_pool->irqs_per_cpu = kcalloc(nr_cpu_ids, sizeof(u16), GFP_KERNEL);
+	if (!table->sf_comp_pool->irqs_per_cpu) {
+		err = -ENOMEM;
+		goto err_irqs_per_cpu;
+	}
+
 	return 0;
+
+err_irqs_per_cpu:
+	irq_pool_free(table->sf_comp_pool);
 err_sf_ctrl:
 	irq_pool_free(table->sf_ctrl_pool);
 err_pf:
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/pci_irq.h b/drivers/net/ethernet/mellanox/mlx5/core/pci_irq.h
index 5fee4ce57d6c..5c7e68bee43a 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/pci_irq.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/pci_irq.h
@@ -20,12 +20,20 @@ struct mlx5_irq_pool {
 	struct xarray irqs;
 	u32 max_threshold;
 	u32 min_threshold;
+	u16 *irqs_per_cpu;
 	struct mlx5_core_dev *dev;
 };
 
+struct mlx5_irq_pool *mlx5_irq_pool_get(struct mlx5_core_dev *dev);
+static inline bool mlx5_irq_pool_is_sf_pool(struct mlx5_irq_pool *pool)
+{
+	return !strncmp("mlx5_sf", pool->name, strlen("mlx5_sf"));
+}
+
 struct mlx5_irq *mlx5_irq_alloc(struct mlx5_irq_pool *pool, int i,
 				const struct cpumask *affinity);
 int mlx5_irq_get_locked(struct mlx5_irq *irq);
 int mlx5_irq_read_locked(struct mlx5_irq *irq);
+int mlx5_irq_put(struct mlx5_irq *irq);
 
 #endif /* __PCI_IRQ_H__ */