On Sun, Sep 25, 2022 at 11:05:18AM -0700, Yury Norov wrote: > On Fri, Sep 23, 2022 at 04:55:40PM +0100, Valentin Schneider wrote: > > 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@xxxxxxxxxx > > Signed-off-by: Valentin Schneider <vschneid@xxxxxxxxxx> > > --- > > include/linux/topology.h | 12 ++++++++++++ > > kernel/sched/topology.c | 31 +++++++++++++++++++++++++++++++ > > 2 files changed, 43 insertions(+) > > > > diff --git a/include/linux/topology.h b/include/linux/topology.h > > index 4564faafd0e1..3e91ae6d0ad5 100644 > > --- a/include/linux/topology.h > > +++ b/include/linux/topology.h > > @@ -245,5 +245,17 @@ 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) > > +{ > > + if (node == NUMA_NO_NODE && !hops) > > + return cpu_online_mask; > > + > > + 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..ee77706603c0 100644 > > --- a/kernel/sched/topology.c > > +++ b/kernel/sched/topology.c > > @@ -2067,6 +2067,37 @@ 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 distance; 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 == NUMA_NO_NODE && !hops) > > + return cpu_online_mask; > > + > > + if (node >= nr_node_ids || hops >= sched_domains_numa_levels) > > + return ERR_PTR(-EINVAL); > > This looks like a sanity check. If so, it should go before the snippet > above, so that client code would behave consistently. > > > + > > + if (!masks) > > + return NULL; > > In (node == NUMA_NO_NODE && !hops) case you return online cpus. Here > you return NULL just to convert it to cpu_online_mask in the caller. > This looks inconsistent. So, together with the above comment, this > makes me feel that you'd do it like this: > > const struct cpumask *sched_numa_hop_mask(int node, int hops) > { > struct cpumask ***masks; > > if (node >= nr_node_ids || hops >= sched_domains_numa_levels) > { > #ifdef CONFIG_SCHED_DEBUG > pr_err(...); > #endif > return ERR_PTR(-EINVAL); > } It's an exported function, and any lame driver may crash the system by dereferencing a random pointer. You need to check the node for -2, -3, etc, because only -1 is a valid negative value. For hops, it should be an unsigned int. Right? > > if (node == NUMA_NO_NODE && !hops) > return cpu_online_mask; /* or NULL */ > > masks = rcu_dereference(sched_domains_numa_masks); > if (!masks) > return cpu_online_mask; /* or NULL */ > > return masks[hops][node]; > }
