1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /* Copyright (c) 2019 Mellanox Technologies. */
4 #include <linux/interrupt.h>
5 #include <linux/notifier.h>
6 #include <linux/module.h>
7 #include <linux/mlx5/driver.h>
9 #ifdef CONFIG_RFS_ACCEL
10 #include <linux/cpu_rmap.h>
13 #define MLX5_MAX_IRQ_NAME (32)
16 struct atomic_notifier_head nh;
18 char name[MLX5_MAX_IRQ_NAME];
21 struct mlx5_irq_table {
24 #ifdef CONFIG_RFS_ACCEL
25 struct cpu_rmap *rmap;
29 int mlx5_irq_table_init(struct mlx5_core_dev *dev)
31 struct mlx5_irq_table *irq_table;
33 irq_table = kvzalloc(sizeof(*irq_table), GFP_KERNEL);
37 dev->priv.irq_table = irq_table;
41 void mlx5_irq_table_cleanup(struct mlx5_core_dev *dev)
43 kvfree(dev->priv.irq_table);
46 int mlx5_irq_get_num_comp(struct mlx5_irq_table *table)
48 return table->nvec - MLX5_IRQ_VEC_COMP_BASE;
51 static struct mlx5_irq *mlx5_irq_get(struct mlx5_core_dev *dev, int vecidx)
53 struct mlx5_irq_table *irq_table = dev->priv.irq_table;
55 return &irq_table->irq[vecidx];
58 int mlx5_irq_attach_nb(struct mlx5_irq_table *irq_table, int vecidx,
59 struct notifier_block *nb)
63 irq = &irq_table->irq[vecidx];
64 return atomic_notifier_chain_register(&irq->nh, nb);
67 int mlx5_irq_detach_nb(struct mlx5_irq_table *irq_table, int vecidx,
68 struct notifier_block *nb)
72 irq = &irq_table->irq[vecidx];
73 return atomic_notifier_chain_unregister(&irq->nh, nb);
76 static irqreturn_t mlx5_irq_int_handler(int irq, void *nh)
78 atomic_notifier_call_chain(nh, 0, NULL);
82 static void irq_set_name(char *name, int vecidx)
85 snprintf(name, MLX5_MAX_IRQ_NAME, "mlx5_async");
89 snprintf(name, MLX5_MAX_IRQ_NAME, "mlx5_comp%d",
90 vecidx - MLX5_IRQ_VEC_COMP_BASE);
94 static int request_irqs(struct mlx5_core_dev *dev, int nvec)
96 char name[MLX5_MAX_IRQ_NAME];
100 for (i = 0; i < nvec; i++) {
101 struct mlx5_irq *irq = mlx5_irq_get(dev, i);
102 int irqn = pci_irq_vector(dev->pdev, i);
104 irq_set_name(name, i);
105 ATOMIC_INIT_NOTIFIER_HEAD(&irq->nh);
106 snprintf(irq->name, MLX5_MAX_IRQ_NAME,
107 "%s@pci:%s", name, pci_name(dev->pdev));
108 err = request_irq(irqn, mlx5_irq_int_handler, 0, irq->name,
111 mlx5_core_err(dev, "Failed to request irq\n");
112 goto err_request_irq;
118 for (; i >= 0; i--) {
119 struct mlx5_irq *irq = mlx5_irq_get(dev, i);
120 int irqn = pci_irq_vector(dev->pdev, i);
122 free_irq(irqn, &irq->nh);
127 static void irq_clear_rmap(struct mlx5_core_dev *dev)
129 #ifdef CONFIG_RFS_ACCEL
130 struct mlx5_irq_table *irq_table = dev->priv.irq_table;
132 free_irq_cpu_rmap(irq_table->rmap);
136 static int irq_set_rmap(struct mlx5_core_dev *mdev)
139 #ifdef CONFIG_RFS_ACCEL
140 struct mlx5_irq_table *irq_table = mdev->priv.irq_table;
141 int num_affinity_vec;
144 num_affinity_vec = mlx5_irq_get_num_comp(irq_table);
145 irq_table->rmap = alloc_irq_cpu_rmap(num_affinity_vec);
146 if (!irq_table->rmap) {
148 mlx5_core_err(mdev, "Failed to allocate cpu_rmap. err %d", err);
152 vecidx = MLX5_IRQ_VEC_COMP_BASE;
153 for (; vecidx < irq_table->nvec; vecidx++) {
154 err = irq_cpu_rmap_add(irq_table->rmap,
155 pci_irq_vector(mdev->pdev, vecidx));
157 mlx5_core_err(mdev, "irq_cpu_rmap_add failed. err %d",
159 goto err_irq_cpu_rmap_add;
164 err_irq_cpu_rmap_add:
165 irq_clear_rmap(mdev);
171 /* Completion IRQ vectors */
173 static int set_comp_irq_affinity_hint(struct mlx5_core_dev *mdev, int i)
175 int vecidx = MLX5_IRQ_VEC_COMP_BASE + i;
176 struct mlx5_irq *irq;
179 irq = mlx5_irq_get(mdev, vecidx);
180 irqn = pci_irq_vector(mdev->pdev, vecidx);
181 if (!zalloc_cpumask_var(&irq->mask, GFP_KERNEL)) {
182 mlx5_core_warn(mdev, "zalloc_cpumask_var failed");
186 cpumask_set_cpu(cpumask_local_spread(i, mdev->priv.numa_node),
188 if (IS_ENABLED(CONFIG_SMP) &&
189 irq_set_affinity_hint(irqn, irq->mask))
190 mlx5_core_warn(mdev, "irq_set_affinity_hint failed, irq 0x%.4x",
196 static void clear_comp_irq_affinity_hint(struct mlx5_core_dev *mdev, int i)
198 int vecidx = MLX5_IRQ_VEC_COMP_BASE + i;
199 struct mlx5_irq *irq;
202 irq = mlx5_irq_get(mdev, vecidx);
203 irqn = pci_irq_vector(mdev->pdev, vecidx);
204 irq_set_affinity_hint(irqn, NULL);
205 free_cpumask_var(irq->mask);
208 static int set_comp_irq_affinity_hints(struct mlx5_core_dev *mdev)
210 int nvec = mlx5_irq_get_num_comp(mdev->priv.irq_table);
214 for (i = 0; i < nvec; i++) {
215 err = set_comp_irq_affinity_hint(mdev, i);
223 for (i--; i >= 0; i--)
224 clear_comp_irq_affinity_hint(mdev, i);
229 static void clear_comp_irqs_affinity_hints(struct mlx5_core_dev *mdev)
231 int nvec = mlx5_irq_get_num_comp(mdev->priv.irq_table);
234 for (i = 0; i < nvec; i++)
235 clear_comp_irq_affinity_hint(mdev, i);
239 mlx5_irq_get_affinity_mask(struct mlx5_irq_table *irq_table, int vecidx)
241 return irq_table->irq[vecidx].mask;
244 #ifdef CONFIG_RFS_ACCEL
245 struct cpu_rmap *mlx5_irq_get_rmap(struct mlx5_irq_table *irq_table)
247 return irq_table->rmap;
251 static void unrequest_irqs(struct mlx5_core_dev *dev)
253 struct mlx5_irq_table *table = dev->priv.irq_table;
256 for (i = 0; i < table->nvec; i++)
257 free_irq(pci_irq_vector(dev->pdev, i),
258 &mlx5_irq_get(dev, i)->nh);
261 int mlx5_irq_table_create(struct mlx5_core_dev *dev)
263 struct mlx5_priv *priv = &dev->priv;
264 struct mlx5_irq_table *table = priv->irq_table;
265 int num_eqs = MLX5_CAP_GEN(dev, max_num_eqs) ?
266 MLX5_CAP_GEN(dev, max_num_eqs) :
267 1 << MLX5_CAP_GEN(dev, log_max_eq);
271 nvec = MLX5_CAP_GEN(dev, num_ports) * num_online_cpus() +
272 MLX5_IRQ_VEC_COMP_BASE;
273 nvec = min_t(int, nvec, num_eqs);
274 if (nvec <= MLX5_IRQ_VEC_COMP_BASE)
277 table->irq = kcalloc(nvec, sizeof(*table->irq), GFP_KERNEL);
281 nvec = pci_alloc_irq_vectors(dev->pdev, MLX5_IRQ_VEC_COMP_BASE + 1,
290 err = irq_set_rmap(dev);
294 err = request_irqs(dev, nvec);
296 goto err_request_irqs;
298 err = set_comp_irq_affinity_hints(dev);
300 mlx5_core_err(dev, "Failed to alloc affinity hint cpumask\n");
301 goto err_set_affinity;
311 pci_free_irq_vectors(dev->pdev);
317 void mlx5_irq_table_destroy(struct mlx5_core_dev *dev)
319 struct mlx5_irq_table *table = dev->priv.irq_table;
322 /* free_irq requires that affinity and rmap will be cleared
323 * before calling it. This is why there is asymmetry with set_rmap
324 * which should be called after alloc_irq but before request_irq.
327 clear_comp_irqs_affinity_hints(dev);
328 for (i = 0; i < table->nvec; i++)
329 free_irq(pci_irq_vector(dev->pdev, i),
330 &mlx5_irq_get(dev, i)->nh);
331 pci_free_irq_vectors(dev->pdev);