Message ID | 20190118135907.2336-1-stefan@agner.ch (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | [1/2] ARM: dts: imx6q: add pmu interrupt-affinity | expand |
Am Freitag, den 18.01.2019, 14:59 +0100 schrieb Stefan Agner: > Explicitly specify interrupt affinity to avoid HW perfevents > need to guess. This avoids the following error upon boot: > hw perfevents: no interrupt-affinity property for /pmu, guessing. > But then it isn't correct either AFAICS. On i.MX6 all the PMU IRQs are ORed together into a single SPI, instead of each core dealing with its own PPI. So pretending that there are more IRQs with affinity to each core isn't the right thing to do, no? Regards, Lucas > Specifying all four CPUs shows no aversive effects on i.MX 6Dual > SoCs. > > > Signed-off-by: Stefan Agner <stefan@agner.ch> > --- > arch/arm/boot/dts/imx6q.dtsi | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/arch/arm/boot/dts/imx6q.dtsi b/arch/arm/boot/dts/imx6q.dtsi > index 8381d24eff7d..d2c1977c8b16 100644 > --- a/arch/arm/boot/dts/imx6q.dtsi > +++ b/arch/arm/boot/dts/imx6q.dtsi > @@ -537,6 +537,13 @@ > > <0x28 0x0000000c>; /* DCIC2_MUX_CTL */ > }; > > +&pmu { > > > + interrupt-affinity = <&{/cpus/cpu@0}>, > > > + <&{/cpus/cpu@1}>, > > > + <&{/cpus/cpu@2}>, > > > + <&{/cpus/cpu@3}>; > +}; > + > &vpu { > > compatible = "fsl,imx6q-vpu", "cnm,coda960"; > };
On 18.01.2019 15:12, Lucas Stach wrote: > Am Freitag, den 18.01.2019, 14:59 +0100 schrieb Stefan Agner: >> Explicitly specify interrupt affinity to avoid HW perfevents >> need to guess. This avoids the following error upon boot: >> hw perfevents: no interrupt-affinity property for /pmu, guessing. >> > But then it isn't correct either AFAICS. On i.MX6 all the PMU IRQs are > ORed together into a single SPI, instead of each core dealing with its > own PPI. So pretending that there are more IRQs with affinity to each > core isn't the right thing to do, no? > Oh I see, we only have a single interrupt in the i.MX 6 case. I agree, this patches are wrong. Hm, but why does hw perf then think it needs to guess? Doesn't seem hard to guess right if there is only one choice... We probably need to do something like this? --- a/drivers/perf/arm_pmu_platform.c +++ b/drivers/perf/arm_pmu_platform.c @@ -122,7 +122,8 @@ static int pmu_parse_irqs(struct arm_pmu *pmu) return pmu_parse_percpu_irq(pmu, irq); } - if (nr_cpu_ids != 1 && !pmu_has_irq_affinity(pdev->dev.of_node)) { + if ((nr_cpu_ids != 1 || num_irqs != 1) && + !pmu_has_irq_affinity(pdev->dev.of_node)) { pr_warn("no interrupt-affinity property for %pOF, guessing.\n", pdev->dev.of_node); } -- Stefan > Regards, > Lucas > >> Specifying all four CPUs shows no aversive effects on i.MX 6Dual >> SoCs. >> >> > Signed-off-by: Stefan Agner <stefan@agner.ch> >> --- >> arch/arm/boot/dts/imx6q.dtsi | 7 +++++++ >> 1 file changed, 7 insertions(+) >> >> diff --git a/arch/arm/boot/dts/imx6q.dtsi b/arch/arm/boot/dts/imx6q.dtsi >> index 8381d24eff7d..d2c1977c8b16 100644 >> --- a/arch/arm/boot/dts/imx6q.dtsi >> +++ b/arch/arm/boot/dts/imx6q.dtsi >> @@ -537,6 +537,13 @@ >> > <0x28 0x0000000c>; /* DCIC2_MUX_CTL */ >> }; >> >> +&pmu { >> > > + interrupt-affinity = <&{/cpus/cpu@0}>, >> > > + <&{/cpus/cpu@1}>, >> > > + <&{/cpus/cpu@2}>, >> > > + <&{/cpus/cpu@3}>; >> +}; >> + >> &vpu { >> > compatible = "fsl,imx6q-vpu", "cnm,coda960"; >> };
[adding Will/Mark] On 18.01.2019 16:41, Stefan Agner wrote: > On 18.01.2019 15:12, Lucas Stach wrote: >> Am Freitag, den 18.01.2019, 14:59 +0100 schrieb Stefan Agner: >>> Explicitly specify interrupt affinity to avoid HW perfevents >>> need to guess. This avoids the following error upon boot: >>> hw perfevents: no interrupt-affinity property for /pmu, guessing. >>> >> But then it isn't correct either AFAICS. On i.MX6 all the PMU IRQs are >> ORed together into a single SPI, instead of each core dealing with its >> own PPI. So pretending that there are more IRQs with affinity to each >> core isn't the right thing to do, no? >> > > Oh I see, we only have a single interrupt in the i.MX 6 case. > > I agree, this patches are wrong. > > Hm, but why does hw perf then think it needs to guess? Doesn't seem hard > to guess right if there is only one choice... > > We probably need to do something like this? > > --- a/drivers/perf/arm_pmu_platform.c > +++ b/drivers/perf/arm_pmu_platform.c > @@ -122,7 +122,8 @@ static int pmu_parse_irqs(struct arm_pmu *pmu) > return pmu_parse_percpu_irq(pmu, irq); > } > > - if (nr_cpu_ids != 1 && !pmu_has_irq_affinity(pdev->dev.of_node)) > { > + if ((nr_cpu_ids != 1 || num_irqs != 1) && > + !pmu_has_irq_affinity(pdev->dev.of_node)) { > pr_warn("no interrupt-affinity property for %pOF, > guessing.\n", > pdev->dev.of_node); > } Yeah, I see, it was on Friday... if anything, it should be: - if (nr_cpu_ids != 1 && !pmu_has_irq_affinity(pdev->dev.of_node)) { + if (nr_cpu_ids != 1 && num_irqs != 1 && + !pmu_has_irq_affinity(pdev->dev.of_node)) { However, I realized that arm_pmu_platform actually currently only assigns the one IRQ to the first CPU. This leads to perf not working if a process is scheduled on another CPU then the first. I could reproduce this. PID 763 is a long running task. root@colibri-imx6-05051054:~# taskset -p 0x1 763 pid 763's current affinity mask: 3 pid 763's new affinity mask: 1 root@colibri-imx6-05051054:~# perf stat -e cpu-cycles -p 763 ^C Performance counter stats for process id '763': 7581021 cpu-cycles 1.222248490 seconds time elapsed root@colibri-imx6-05051054:~# taskset -p 0x2 763 pid 763's current affinity mask: 1 pid 763's new affinity mask: 2 root@colibri-imx6-05051054:~# perf stat -e cpu-cycles -p 763 ^C Performance counter stats for process id '763': <not counted> cpu-cycles (0.00%) 1.050253575 seconds time elapsed I guess we need to tell the PMU driver that a single IRQ should be used for all CPUs? -- Stefan > > >> Regards, >> Lucas >> >>> Specifying all four CPUs shows no aversive effects on i.MX 6Dual >>> SoCs. >>> >>> > Signed-off-by: Stefan Agner <stefan@agner.ch> >>> --- >>> arch/arm/boot/dts/imx6q.dtsi | 7 +++++++ >>> 1 file changed, 7 insertions(+) >>> >>> diff --git a/arch/arm/boot/dts/imx6q.dtsi b/arch/arm/boot/dts/imx6q.dtsi >>> index 8381d24eff7d..d2c1977c8b16 100644 >>> --- a/arch/arm/boot/dts/imx6q.dtsi >>> +++ b/arch/arm/boot/dts/imx6q.dtsi >>> @@ -537,6 +537,13 @@ >>> > <0x28 0x0000000c>; /* DCIC2_MUX_CTL */ >>> }; >>> >>> +&pmu { >>> > > + interrupt-affinity = <&{/cpus/cpu@0}>, >>> > > + <&{/cpus/cpu@1}>, >>> > > + <&{/cpus/cpu@2}>, >>> > > + <&{/cpus/cpu@3}>; >>> +}; >>> + >>> &vpu { >>> > compatible = "fsl,imx6q-vpu", "cnm,coda960"; >>> };
diff --git a/arch/arm/boot/dts/imx6q.dtsi b/arch/arm/boot/dts/imx6q.dtsi index 8381d24eff7d..d2c1977c8b16 100644 --- a/arch/arm/boot/dts/imx6q.dtsi +++ b/arch/arm/boot/dts/imx6q.dtsi @@ -537,6 +537,13 @@ <0x28 0x0000000c>; /* DCIC2_MUX_CTL */ }; +&pmu { + interrupt-affinity = <&{/cpus/cpu@0}>, + <&{/cpus/cpu@1}>, + <&{/cpus/cpu@2}>, + <&{/cpus/cpu@3}>; +}; + &vpu { compatible = "fsl,imx6q-vpu", "cnm,coda960"; };
Explicitly specify interrupt affinity to avoid HW perfevents need to guess. This avoids the following error upon boot: hw perfevents: no interrupt-affinity property for /pmu, guessing. Specifying all four CPUs shows no aversive effects on i.MX 6Dual SoCs. Signed-off-by: Stefan Agner <stefan@agner.ch> --- arch/arm/boot/dts/imx6q.dtsi | 7 +++++++ 1 file changed, 7 insertions(+)