On Mon, Mar 20, 2023 at 2:22 PM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
>
> Hi Rob,
>
> On Sun, Mar 19, 2023 at 4:00 PM Rob Herring <robh@xxxxxxxxxx> wrote:
> > Replace open coded CPU nodes reading of "reg" and translation to logical
> > ID with of_cpu_node_to_id().
> >
> > Signed-off-by: Rob Herring <robh@xxxxxxxxxx>
>
> Thanks for your patch!
>
> > --- a/arch/arm/mach-shmobile/platsmp-apmu.c
> > +++ b/arch/arm/mach-shmobile/platsmp-apmu.c
> > @@ -10,6 +10,7 @@
> > #include <linux/init.h>
> > #include <linux/io.h>
> > #include <linux/ioport.h>
> > +#include <linux/of.h>
> > #include <linux/of_address.h>
> > #include <linux/smp.h>
> > #include <linux/suspend.h>
> > @@ -210,7 +211,6 @@ static void apmu_parse_dt(void (*fn)(struct resource *res, int cpu, int bit))
> > struct device_node *np_apmu, *np_cpu;
> > struct resource res;
> > int bit, index;
> > - u32 id;
> >
> > for_each_matching_node(np_apmu, apmu_ids) {
> > /* only enable the cluster that includes the boot CPU */
> > @@ -218,33 +218,27 @@ static void apmu_parse_dt(void (*fn)(struct resource *res, int cpu, int bit))
> >
> > for (bit = 0; bit < CONFIG_NR_CPUS; bit++) {
>
> This loops over all CPUs....
>
> > np_cpu = of_parse_phandle(np_apmu, "cpus", bit);
Have you looked at what this does? :)
> > - if (np_cpu) {
> > - if (!of_property_read_u32(np_cpu, "reg", &id)) {
> > - if (id == cpu_logical_map(0)) {
> > - is_allowed = true;
> > - of_node_put(np_cpu);
> > - break;
> > - }
> > -
> > - }
> > + if (np_cpu && of_cpu_node_to_id(np_cpu) == 0) {
>
> As of_cpu_node_to_id() uses for_each_possible_cpu(), you're
> converting an O(n) operation to O(n²). I'm sure this can be done
> more efficiently, using for_each_possible_cpu() as the outer loop?
>
> Meh, cpu_logical_map() also loops over all CPUs, so it was already
> O(n²)... Still, we should do better...
Can you measure the performance impact?
I would assume 'cpus' is less than NR_CPUS or possible cpus. We should
cap the outer loop based on 'cpus' length. The simplest fix is bail
when of_parse_phandle() fails. That will work unless you expect empty
entries in 'cpus':
cpus = <&cpu0>, <0>, <&cpu3>;
Otherwise, we'd need to use of_parse_phandle_with_fixed_args() instead
to distinguish empty entry vs. the end of the list. There is a
of_for_each_phandle() iterator which would avoid walking the 'cpus'
entry each time from the beginning, but it's a bit more complicated
since it handles arg cells.
This is the simple fix:
diff --git a/arch/arm/mach-shmobile/platsmp-apmu.c
b/arch/arm/mach-shmobile/platsmp-apmu.c
index 27cfe753c467..ec6f421c0f4d 100644
--- a/arch/arm/mach-shmobile/platsmp-apmu.c
+++ b/arch/arm/mach-shmobile/platsmp-apmu.c
@@ -218,7 +218,9 @@ static void apmu_parse_dt(void (*fn)(struct
resource *res, int cpu, int bit))
for (bit = 0; bit < CONFIG_NR_CPUS; bit++) {
np_cpu = of_parse_phandle(np_apmu, "cpus", bit);
- if (np_cpu && of_cpu_node_to_id(np_cpu) == 0) {
+ if (!np_cpu)
+ break;
+ if (of_cpu_node_to_id(np_cpu) == 0) {
is_allowed = true;
of_node_put(np_cpu);
break;
@@ -231,7 +233,7 @@ static void apmu_parse_dt(void (*fn)(struct
resource *res, int cpu, int bit))
for (bit = 0; bit < CONFIG_NR_CPUS; bit++) {
np_cpu = of_parse_phandle(np_apmu, "cpus", bit);
if (!np_cpu)
- continue;
+ break;
index = of_cpu_node_to_id(np_cpu);
if ((index >= 0) &&
