On Wed, Jul 18, 2018 at 12:25:32PM +0100, Sudeep Holla wrote:
> Commit 7f9545aa1a91 ("arm64: smp: remove cpu and numa topology
> information when hotplugging out CPU") updates the cpu topology when
> the CPU is hotplugged out. However the PSCI checker code uses the
> topology_core_cpumask pointers for some of the cpu hotplug testing.
> Since the pointer to the core_cpumask of the first CPU in the group
> is used, which when that CPU itself is hotpugged out is just set to
> itself, the testing terminates after that particular CPU is tested out.
> But the intention of this tests is to cover all the CPU in the group.
>
> In order to support that, we need to stash the topology_core_cpumask
> before the start of the test and use that value instead of pointer to
> a cpumask which will be updated on CPU hotplug.
>
> Fixes: 7f9545aa1a91a9a4 ("arm64: smp: remove cpu and numa topology
> information when hotplugging out CPU")
> Reported-by: Geert Uytterhoeven <geert@xxxxxxxxxxxxxx>
> Tested-by: Geert Uytterhoeven <geert+renesas@xxxxxxxxx>
> Cc: Mark Rutland <mark.rutland@xxxxxxx>
> Cc: Lorenzo Pieralisi <lorenzo.pieralisi@xxxxxxx>
> Signed-off-by: Sudeep Holla <sudeep.holla@xxxxxxx>
> ---
> drivers/firmware/psci_checker.c | 30 ++++++++++++++++++++----------
> 1 file changed, 20 insertions(+), 10 deletions(-)
>
> diff --git a/drivers/firmware/psci_checker.c b/drivers/firmware/psci_checker.c
> index bb1c068bff19..e330c6cb45f5 100644
> --- a/drivers/firmware/psci_checker.c
> +++ b/drivers/firmware/psci_checker.c
> @@ -77,21 +77,23 @@ static int psci_ops_check(void)
> return 0;
> }
>
> -static int find_cpu_groups(const struct cpumask *cpus,
> - const struct cpumask **cpu_groups)
> +static int find_cpu_groups(cpumask_var_t *cpu_groups)
> {
> unsigned int nb = 0;
> cpumask_var_t tmp;
>
> if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
> return -ENOMEM;
> - cpumask_copy(tmp, cpus);
> + cpumask_copy(tmp, cpu_online_mask);
>
> while (!cpumask_empty(tmp)) {
> const struct cpumask *cpu_group =
> topology_core_cpumask(cpumask_any(tmp));
>
> - cpu_groups[nb++] = cpu_group;
> + if (cpu_groups && cpu_groups[nb])
> + cpumask_copy(cpu_groups[nb], cpu_group);
> +
> + nb++;
> cpumask_andnot(tmp, tmp, cpu_group);
> }
>
> @@ -169,25 +171,31 @@ static unsigned int down_and_up_cpus(const struct cpumask *cpus,
> static int hotplug_tests(void)
> {
> int err;
> - cpumask_var_t offlined_cpus;
> + cpumask_var_t offlined_cpus, *cpu_groups;
> int i, nb_cpu_group;
> - const struct cpumask **cpu_groups;
> char *page_buf;
>
> + /* first run to just get the number of cpu groups */
> + nb_cpu_group = find_cpu_groups(NULL);
> +
> err = -ENOMEM;
> if (!alloc_cpumask_var(&offlined_cpus, GFP_KERNEL))
> return err;
> - /* We may have up to nb_available_cpus cpu_groups. */
> - cpu_groups = kmalloc_array(nb_available_cpus, sizeof(*cpu_groups),
> - GFP_KERNEL);
> + cpu_groups = kcalloc(nb_cpu_group, sizeof(cpu_groups), GFP_KERNEL);
> if (!cpu_groups)
> goto out_free_cpus;
> +
> + for (i = 0; i < nb_cpu_group; ++i)
> + if (!alloc_cpumask_var(&cpu_groups[i], GFP_KERNEL))
> + goto out_free_cpu_groups;
> +
> page_buf = (char *)__get_free_page(GFP_KERNEL);
> if (!page_buf)
> goto out_free_cpu_groups;
>
> err = 0;
> - nb_cpu_group = find_cpu_groups(cpu_online_mask, cpu_groups);
> + /* second run to populate/copy the cpumask */
> + nb_cpu_group = find_cpu_groups(cpu_groups);
>
> /*
> * Of course the last CPU cannot be powered down and cpu_down() should
> @@ -212,6 +220,8 @@ static int hotplug_tests(void)
>
> free_page((unsigned long)page_buf);
> out_free_cpu_groups:
> + for (i = 0; i < nb_cpu_group; ++i)
> + free_cpumask_var(cpu_groups[i]);
> kfree(cpu_groups);
> out_free_cpus:
> free_cpumask_var(offlined_cpus);
Hi Sudeep,
thanks for the patch. I reckon that adding two functions, say,
alloc_cpu_groups() and free_cpu_groups() would make the code
more readable instead of relying on find_cpu_groups() to first
count then copy; it is for readability rather than correctness.
Thanks,
Lorenzo
