Tariq has pointed out that drivers allocating IRQ vectors would benefit
from having smarter NUMA-awareness - cpumask_local_spread() only knows
about the local node and everything outside is in the same bucket.

sched_domains_numa_masks is pretty much what we want to hand out (a cpumask
of CPUs reachable within a given distance budget), introduce
sched_numa_hop_mask() to export those cpumasks.

Link: http://lore.kernel.org/r/20220728191203.4055-1-tariqt@nvidia.com
Signed-off-by: Valentin Schneider <vschneid@redhat.com>
---
 include/linux/topology.h |  9 +++++++++
 kernel/sched/topology.c  | 28 ++++++++++++++++++++++++++++
 2 files changed, 37 insertions(+)

diff --git a/include/linux/topology.h b/include/linux/topology.h
index 4564faafd0e1..13b82b83e547 100644
--- a/include/linux/topology.h
+++ b/include/linux/topology.h
@@ -245,5 +245,14 @@ static inline const struct cpumask *cpu_cpu_mask(int cpu)
 	return cpumask_of_node(cpu_to_node(cpu));
 }
 
+#ifdef CONFIG_NUMA
+extern const struct cpumask *sched_numa_hop_mask(int node, int hops);
+#else
+static inline const struct cpumask *sched_numa_hop_mask(int node, int hops)
+{
+	return ERR_PTR(-EOPNOTSUPP);
+}
+#endif	/* CONFIG_NUMA */
+
 
 #endif /* _LINUX_TOPOLOGY_H */
diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
index 8739c2a5a54e..f0236a0ae65c 100644
--- a/kernel/sched/topology.c
+++ b/kernel/sched/topology.c
@@ -2067,6 +2067,34 @@ int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
 	return found;
 }
 
+/**
+ * sched_numa_hop_mask() - Get the cpumask of CPUs at most @hops hops away.
+ * @node: The node to count hops from.
+ * @hops: Include CPUs up to that many hops away. 0 means local node.
+ *
+ * Requires rcu_lock to be held. Returned cpumask is only valid within that
+ * read-side section, copy it if required beyond that.
+ *
+ * Note that not all hops are equal in size; see sched_init_numa() for how
+ * distances and masks are handled.
+ *
+ * Also note that this is a reflection of sched_domains_numa_masks, which may change
+ * during the lifetime of the system (offline nodes are taken out of the masks).
+ */
+const struct cpumask *sched_numa_hop_mask(int node, int hops)
+{
+	struct cpumask ***masks = rcu_dereference(sched_domains_numa_masks);
+
+	if (node >= nr_node_ids || hops >= sched_domains_numa_levels)
+		return ERR_PTR(-EINVAL);
+
+	if (!masks)
+		return NULL;
+
+	return masks[hops][node];
+}
+EXPORT_SYMBOL_GPL(sched_numa_hop_mask);
+
 #endif /* CONFIG_NUMA */
 
 static int __sdt_alloc(const struct cpumask *cpu_map)
-- 
2.31.1

The recently introduced sched_numa_hop_mask() exposes cpumasks of CPUs
reachable within a given distance budget, but this means each successive
cpumask is a superset of the previous one.

Code wanting to allocate one item per CPU (e.g. IRQs) at increasing
distances would thus need to allocate a temporary cpumask to note which
CPUs have already been visited. This can be prevented by leveraging
for_each_cpu_andnot() - package all that logic into one ugl^D fancy macro.

Signed-off-by: Valentin Schneider <vschneid@redhat.com>
---
 include/linux/topology.h | 37 +++++++++++++++++++++++++++++++++++++
 1 file changed, 37 insertions(+)

diff --git a/include/linux/topology.h b/include/linux/topology.h
index 13b82b83e547..6c671dc3252c 100644
--- a/include/linux/topology.h
+++ b/include/linux/topology.h
@@ -254,5 +254,42 @@ static inline const struct cpumask *sched_numa_hop_mask(int node, int hops)
 }
 #endif	/* CONFIG_NUMA */
 
+/**
+ * for_each_numa_hop_cpu - iterate over CPUs by increasing NUMA distance,
+ *                         starting from a given node.
+ * @cpu: the iteration variable.
+ * @node: the NUMA node to start the search from.
+ *
+ * Requires rcu_lock to be held.
+ * Careful: this is a double loop, 'break' won't work as expected.
+ *
+ *
+ * Implementation notes:
+ *
+ * Providing it is valid, the mask returned by
+ *  sched_numa_hop_mask(node, hops+1)
+ * is a superset of the one returned by
+ *   sched_numa_hop_mask(node, hops)
+ * which may not be that useful for drivers that try to spread things out and
+ * want to visit a CPU not more than once.
+ *
+ * To accommodate for that, we use for_each_cpu_andnot() to iterate over the cpus
+ * of sched_numa_hop_mask(node, hops+1) with the CPUs of
+ * sched_numa_hop_mask(node, hops) removed, IOW we only iterate over CPUs
+ * a given distance away (rather than *up to* a given distance).
+ *
+ * hops=0 forces us to play silly games: we pass cpu_none_mask to
+ * for_each_cpu_andnot(), which turns it into for_each_cpu().
+ */
+#define for_each_numa_hop_cpu(cpu, node)				       \
+	for (struct { const struct cpumask *curr, *prev; int hops; } __v =     \
+		     { sched_numa_hop_mask(node, 0), NULL, 0 };		       \
+	     !IS_ERR_OR_NULL(__v.curr);					       \
+	     __v.hops++,                                                       \
+	     __v.prev = __v.curr,					       \
+	     __v.curr = sched_numa_hop_mask(node, __v.hops))                   \
+		for_each_cpu_andnot(cpu,				       \
+				    __v.curr,				       \
+				    __v.hops ? __v.prev : cpu_none_mask)
 
 #endif /* _LINUX_TOPOLOGY_H */
-- 
2.31.1

Not-signed-off-by: Valentin Schneider <vschneid@redhat.com>
---
 drivers/net/ethernet/mellanox/mlx5/core/eq.c | 12 ++++++++++--
 1 file changed, 10 insertions(+), 2 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx5/core/eq.c b/drivers/net/ethernet/mellanox/mlx5/core/eq.c
index 229728c80233..0a5432903edd 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/eq.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/eq.c
@@ -812,6 +812,7 @@ static int comp_irqs_request(struct mlx5_core_dev *dev)
 	int ncomp_eqs = table->num_comp_eqs;
 	u16 *cpus;
 	int ret;
+	int cpu;
 	int i;
 
 	ncomp_eqs = table->num_comp_eqs;
@@ -830,8 +831,15 @@ static int comp_irqs_request(struct mlx5_core_dev *dev)
 		ret = -ENOMEM;
 		goto free_irqs;
 	}
-	for (i = 0; i < ncomp_eqs; i++)
-		cpus[i] = cpumask_local_spread(i, dev->priv.numa_node);
+
+	rcu_read_lock();
+	for_each_numa_hop_cpus(cpu, dev->priv.numa_node) {
+		cpus[i] = cpu;
+		if (++i == ncomp_eqs)
+			goto spread_done;
+	}
+spread_done:
+	rcu_read_unlock();
 	ret = mlx5_irqs_request_vectors(dev, cpus, ncomp_eqs, table->comp_irqs);
 	kfree(cpus);
 	if (ret < 0)
-- 
2.31.1