Skip to contentSkip to conversation

arm64: dts: allwinner: a64: Drop PMU node

PATCH Applied Vasily Khoruzhick 9 participants 26 messages Aug 6, 2019 – Nov 11, 2019

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 arm64: dts: allwinner: a64: Drop PMU node

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Tested on Pine64-LTS.

Fixes 34a97fcc71c2 ("arm64: dts: allwinner: a64: Add PMU node")
Cc Harald Geyer <[email protected]>
Cc Jared D. McNeill <[email protected]>
---
 arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi | 9 ---------
 1 file changed, 9 deletions(-)
arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi +0 -9
index 9cc9bdde81ac..cd92f546c483 100644
@@ -142,15 +142,6 @@
 		clock-output-names = "ext-osc32k";
 	};
 
-	pmu {
-		compatible = "arm,cortex-a53-pmu";
-		interrupts = <GIC_SPI 152 IRQ_TYPE_LEVEL_HIGH>,
-			     <GIC_SPI 153 IRQ_TYPE_LEVEL_HIGH>,
-			     <GIC_SPI 154 IRQ_TYPE_LEVEL_HIGH>,
-			     <GIC_SPI 155 IRQ_TYPE_LEVEL_HIGH>;
-		interrupt-affinity = <&cpu0>, <&cpu1>, <&cpu2>, <&cpu3>;
-	};
-
 	psci {
 		compatible = "arm,psci-0.2";
 		method = "smc";
Signature · 1 line
2.22.0
#1 lore raw
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Quoted 15 lines

Tested on Pine64-LTS.

Fixes 34a97fcc71c2 ("arm64: dts: allwinner: a64: Add PMU node")
Cc Harald Geyer <[email protected]>
Cc Jared D. McNeill <[email protected]>
---
  arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi | 9 ---------
  1 file changed, 9 deletions(-)
arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi +0 -0
index 9cc9bdde81ac..cd92f546c483 100644

arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi line 150

Quoted 9 lines
arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi +0 -6
@@ -142,15 +142,6 @@
  		clock-output-names = "ext-osc32k";
  	};
 
-	pmu {
-		compatible = "arm,cortex-a53-pmu";
-		interrupts = <GIC_SPI 152 IRQ_TYPE_LEVEL_HIGH>,
-			     <GIC_SPI 153 IRQ_TYPE_LEVEL_HIGH>,
-			     <GIC_SPI 154 IRQ_TYPE_LEVEL_HIGH>,
-			     <GIC_SPI 155 IRQ_TYPE_LEVEL_HIGH>;

Cross-referencing between some random DTs in the H6 BSP I happen to have to hand and the A64 User Manual, it looks a lot like someone just forgot to subtract 32 from these to satisfy the awkward GIC binding - that wants the SPI index rather than the actual interrupt source number, which implies these should probably be 120-123 rather than 152-155.

Robin.

Quoted 7 lines
arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi +0 -3
-		interrupt-affinity = <&cpu0>, <&cpu1>, <&cpu2>, <&cpu3>;
-	};
-
  	psci {
  		compatible = "arm,psci-0.2";
  		method = "smc";
#2 lore raw
Quoted 7 lines · Robin Murphy
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

$ perf stat sleep 1

Performance counter stats for 'sleep 1':

4.08 msec task-clock:u # 0.004 CPUs
utilized

                0      context-switches:u        #    0.000 K/sec
                0      cpu-migrations:u          #    0.000 K/sec
               55      page-faults:u             #    0.013 M/sec
          527,711      cycles:u                  #    0.129 GHz
          197,262      instructions:u            #    0.37  insn per
cycle
           24,242      branches:u                #    5.947 M/sec
            5,083      branch-misses:u           #   20.97% of all

branches

      1.011928625 seconds time elapsed

      0.000000000 seconds user
      0.007196000 seconds sys
Quoted 18 lines
Quoted 18 lines

Tested on Pine64-LTS.

Fixes 34a97fcc71c2 ("arm64: dts: allwinner: a64: Add PMU node")
Cc Harald Geyer <[email protected]>
Cc Jared D. McNeill <[email protected]>
---
  arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi | 9 ---------
  1 file changed, 9 deletions(-)
arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi +0 -0
index 9cc9bdde81ac..cd92f546c483 100644
@@ -142,15 +142,6 @@
              clock-output-names = "ext-osc32k";
      };
 
Quoted 12 lines

arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi line 150

Quoted 6 lines
arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi +0 -6
-     pmu {
-             compatible = "arm,cortex-a53-pmu";
-             interrupts = <GIC_SPI 152 IRQ_TYPE_LEVEL_HIGH>,
-                          <GIC_SPI 153 IRQ_TYPE_LEVEL_HIGH>,
-                          <GIC_SPI 154 IRQ_TYPE_LEVEL_HIGH>,
-                          <GIC_SPI 155 IRQ_TYPE_LEVEL_HIGH>;

Cross-referencing between some random DTs in the H6 BSP I happen to have to hand and the A64 User Manual, it looks a lot like someone just forgot to subtract 32 from these to satisfy the awkward GIC binding - that wants the SPI index rather than the actual interrupt source number, which implies these should probably be 120-123 rather than 152-155.

Tried that, didn't work. 'grep pmu /proc/interrupts' shows zeroes, and 'perf top' is completely silent.

Quoted 9 lines

Robin.

Quoted 7 lines
arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi +0 -3
-             interrupt-affinity = <&cpu0>, <&cpu1>, <&cpu2>, <&cpu3>;
-     };
-
      psci {
              compatible = "arm,psci-0.2";
              method = "smc";
#3 lore raw
Quoted 10 lines · Vasily Khoruzhick
Quoted 7 lines · Robin Murphy
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

Harald

#4 lore raw
Quoted 15 lines · Harald Geyer
Quoted 10 lines · Vasily Khoruzhick
Quoted 7 lines · Robin Murphy
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Quoted 7 lines

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Quoted 2 lines

Harald

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#5 lore raw
Quoted 20 lines · Vasily Khoruzhick
Quoted 19 lines · Harald Geyer
Quoted 10 lines · Vasily Khoruzhick
Quoted 7 lines · Robin Murphy
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

Quoted 6 lines
Quoted 3 lines

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

Robin.

#6 lore raw
Quoted 26 lines · Robin Murphy
Quoted 24 lines · Vasily Khoruzhick
Quoted 22 lines · Harald Geyer
Quoted 10 lines · Vasily Khoruzhick
Quoted 7 lines · Robin Murphy
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

Quoted 8 lines
Quoted 2 lines

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

I tried to drop interrupts completely from the node but 'perf top' is still broken. Though now in different way: it complains "cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'"

Is there any way to make it working?

Quoted 2 lines

Robin.

#7 lore raw
Quoted 36 lines · Vasily Khoruzhick
Quoted 26 lines · Robin Murphy
Quoted 24 lines · Vasily Khoruzhick
Quoted 22 lines · Harald Geyer
Quoted 10 lines · Vasily Khoruzhick
Quoted 7 lines · Robin Murphy
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

Quoted 8 lines
Quoted 2 lines

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

Quoted 4 lines

I tried to drop interrupts completely from the node but 'perf top' is still broken. Though now in different way: it complains "cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'"

I have no idea if that's the culprit, but what is the state of the 0x09010000 register?

(in particular, are the bits 16-19 and 24 set or not?

Maxime

--
Maxime Ripard, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

#8 lore raw
Quoted 38 lines · Maxime Ripard
Quoted 36 lines · Vasily Khoruzhick
Quoted 34 lines · Robin Murphy
Quoted 20 lines · Vasily Khoruzhick
Quoted 19 lines · Harald Geyer
Quoted 10 lines · Vasily Khoruzhick
Quoted 7 lines · Robin Murphy
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

Quoted 6 lines
Quoted 3 lines

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

Quoted 7 lines
Quoted 4 lines

I tried to drop interrupts completely from the node but 'perf top' is still broken. Though now in different way: it complains "cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'"

I have no idea if that's the culprit, but what is the state of the 0x09010000 register?

What register is that and how do I check it?

Quoted 8 lines

(in particular, are the bits 16-19 and 24 set or not?

Maxime

--
Maxime Ripard, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

#9 lore raw
Quoted 44 lines · Vasily Khoruzhick
Quoted 43 lines · Maxime Ripard
Quoted 40 lines · Vasily Khoruzhick
Quoted 26 lines · Robin Murphy
Quoted 24 lines · Vasily Khoruzhick
Quoted 22 lines · Harald Geyer
Quoted 10 lines · Vasily Khoruzhick
Quoted 7 lines · Robin Murphy
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

Quoted 8 lines
Quoted 2 lines

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

I tried to drop interrupts completely from the node but 'perf top' is still broken. Though now in different way: it complains "cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'"

Quoted 4 lines
Quoted 2 lines

I have no idea if that's the culprit, but what is the state of the 0x09010000 register?

What register is that and how do I check it?

It's in the CPUX Configuration block, and the bits are labelled as CPU Debug Reset.

And if you have busybox, you can use devmem.

Maxime

--
Maxime Ripard, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#10 lore raw
Quoted 49 lines · Maxime Ripard
Quoted 47 lines · Vasily Khoruzhick
Quoted 45 lines · Maxime Ripard
Quoted 36 lines · Vasily Khoruzhick
Quoted 26 lines · Robin Murphy
Quoted 24 lines · Vasily Khoruzhick
Quoted 22 lines · Harald Geyer
Quoted 10 lines · Vasily Khoruzhick
Quoted 7 lines · Robin Murphy
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

Quoted 8 lines
Quoted 2 lines

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

Quoted 4 lines

I tried to drop interrupts completely from the node but 'perf top' is still broken. Though now in different way: it complains "cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'"

I have no idea if that's the culprit, but what is the state of the 0x09010000 register?

Quoted 6 lines
Quoted 1 line

What register is that and how do I check it?

It's in the CPUX Configuration block, and the bits are labelled as CPU Debug Reset.

And if you have busybox, you can use devmem.

CPUX configuration block is at 0x01700000 according to A64 user manual, and particular register you're interested in is at 0x01700080, its value is 0x1110110F.

Bits 16-19 are not defined in user manual and are not set.

Quoted 6 lines

Maxime

--
Maxime Ripard, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

#11 lore raw
Quoted 55 lines · Vasily Khoruzhick
Quoted 54 lines · Maxime Ripard
Quoted 44 lines · Vasily Khoruzhick
Quoted 43 lines · Maxime Ripard
Quoted 40 lines · Vasily Khoruzhick
Quoted 26 lines · Robin Murphy
Quoted 24 lines · Vasily Khoruzhick
Quoted 22 lines · Harald Geyer
Quoted 10 lines · Vasily Khoruzhick
Quoted 7 lines · Robin Murphy
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

Quoted 8 lines
Quoted 2 lines

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

I tried to drop interrupts completely from the node but 'perf top' is still broken. Though now in different way: it complains "cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'"

Quoted 4 lines
Quoted 2 lines

I have no idea if that's the culprit, but what is the state of the 0x09010000 register?

What register is that and how do I check it?

It's in the CPUX Configuration block, and the bits are labelled as CPU Debug Reset.

Quoted 7 lines
Quoted 1 line

And if you have busybox, you can use devmem.

CPUX configuration block is at 0x01700000 according to A64 user manual, and particular register you're interested in is at 0x01700080, its value is 0x1110110F.

Bits 16-19 are not defined in user manual and are not set.

Sorry, I somehow thought this was for the H6...

I don't have any idea then :/

Maxime

--
Maxime Ripard, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#12 lore raw
Quoted 63 lines · Maxime Ripard
Quoted 61 lines · Vasily Khoruzhick
Quoted 49 lines · Maxime Ripard
Quoted 47 lines · Vasily Khoruzhick
Quoted 45 lines · Maxime Ripard
Quoted 36 lines · Vasily Khoruzhick
Quoted 26 lines · Robin Murphy
Quoted 24 lines · Vasily Khoruzhick
Quoted 22 lines · Harald Geyer
Quoted 10 lines · Vasily Khoruzhick
Quoted 7 lines · Robin Murphy
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

Quoted 8 lines
Quoted 2 lines

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

Quoted 4 lines

I tried to drop interrupts completely from the node but 'perf top' is still broken. Though now in different way: it complains "cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'"

I have no idea if that's the culprit, but what is the state of the 0x09010000 register?

Quoted 6 lines
Quoted 1 line

What register is that and how do I check it?

It's in the CPUX Configuration block, and the bits are labelled as CPU Debug Reset.

And if you have busybox, you can use devmem.

CPUX configuration block is at 0x01700000 according to A64 user manual, and particular register you're interested in is at 0x01700080, its value is 0x1110110F.

Quoted 5 lines
Quoted 1 line

Bits 16-19 are not defined in user manual and are not set.

Sorry, I somehow thought this was for the H6...

I don't have any idea then :/

OK, so what should we do? 'perf top'/'perf record' work fine if PMU node is dropped, but they don't work if PMU node is present (even with interrupts dropped). I'd prefer to have 'perf top' and 'perf record' working instead of 'perf stat'

Quoted 6 lines

Maxime

--
Maxime Ripard, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

#13 lore raw
Quoted 3 lines · Vasily Khoruzhick

OK, so what should we do? 'perf top'/'perf record' work fine if PMU node is dropped, but they don't work if PMU node is present (even with interrupts dropped).

Really? Even if you tell it to only listen to software events? (Which is the only thing you get without a PMU anyway, I believe.)

Quoted 2 lines

I'd prefer to have 'perf top' and 'perf record'
working instead of 'perf stat'

I think, if a broken PMU confuses 'perf top' beyond usability, it should be fixed.

Harald

Signature · 9 lines
Nationalratswahl: Ich trete als unabhängiger Experte für Klimaschutz
und freie Software an! Ich will mit Vorzugsstimmen ins Parlament kommen,
weil wenn ich es nicht mache, kümmert sich womöglich niemand darum.
https://haraldgeyer.at

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
#14 lore raw
In reply to Vasily Khoruzhick
Quoted 66 lines · Vasily Khoruzhick
Quoted 65 lines · Maxime Ripard
Quoted 62 lines · Vasily Khoruzhick
Quoted 49 lines · Maxime Ripard
Quoted 47 lines · Vasily Khoruzhick
Quoted 45 lines · Maxime Ripard
Quoted 36 lines · Vasily Khoruzhick
Quoted 26 lines · Robin Murphy
Quoted 24 lines · Vasily Khoruzhick
Quoted 22 lines · Harald Geyer
Quoted 10 lines · Vasily Khoruzhick
Quoted 7 lines · Robin Murphy
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and
as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

Quoted 8 lines
Quoted 2 lines

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

Quoted 4 lines

I tried to drop interrupts completely from the node but 'perf top' is still broken. Though now in different way: it complains "cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'"

I have no idea if that's the culprit, but what is the state of the 0x09010000 register?

Quoted 6 lines
Quoted 1 line

What register is that and how do I check it?

It's in the CPUX Configuration block, and the bits are labelled as CPU Debug Reset.

And if you have busybox, you can use devmem.

CPUX configuration block is at 0x01700000 according to A64 user manual, and particular register you're interested in is at 0x01700080, its value is 0x1110110F.

Bits 16-19 are not defined in user manual and are not set.

Quoted 8 lines
Quoted 3 lines

Sorry, I somehow thought this was for the H6...

I don't have any idea then :/

OK, so what should we do? 'perf top'/'perf record' work fine if PMU node is dropped, but they don't work if PMU node is present (even with interrupts dropped). I'd prefer to have 'perf top' and 'perf record' working instead of 'perf stat'

Well, it doesn't work so we should just remove the node, and if someone wants it back, they should figure it out.

Maxime

--
Maxime Ripard, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

#15 lore raw
Quoted 72 lines · Maxime Ripard
Quoted 70 lines · Vasily Khoruzhick
Quoted 68 lines · Maxime Ripard
Quoted 55 lines · Vasily Khoruzhick
Quoted 54 lines · Maxime Ripard
Quoted 44 lines · Vasily Khoruzhick
Quoted 43 lines · Maxime Ripard
Quoted 40 lines · Vasily Khoruzhick
Quoted 26 lines · Robin Murphy
Quoted 24 lines · Vasily Khoruzhick
Quoted 22 lines · Harald Geyer
Quoted 10 lines · Vasily Khoruzhick
Quoted 7 lines · Robin Murphy

On 06/08/2019 15:01, Vasily Khoruzhick wrote:
Looks like PMU in A64 is broken, it generates no interrupts at all and
as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly?
It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

Quoted 8 lines
Quoted 2 lines

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

I tried to drop interrupts completely from the node but 'perf top' is still broken. Though now in different way: it complains "cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'"

Quoted 4 lines
Quoted 2 lines

I have no idea if that's the culprit, but what is the state of the 0x09010000 register?

What register is that and how do I check it?

It's in the CPUX Configuration block, and the bits are labelled as CPU Debug Reset.

Quoted 7 lines
Quoted 1 line

And if you have busybox, you can use devmem.

CPUX configuration block is at 0x01700000 according to A64 user manual, and particular register you're interested in is at 0x01700080, its value is 0x1110110F.

Bits 16-19 are not defined in user manual and are not set.

Sorry, I somehow thought this was for the H6...

I don't have any idea then :/

Quoted 7 lines
Quoted 4 lines

OK, so what should we do? 'perf top'/'perf record' work fine if PMU node is dropped, but they don't work if PMU node is present (even with interrupts dropped). I'd prefer to have 'perf top' and 'perf record' working instead of 'perf stat'

Well, it doesn't work so we should just remove the node, and if someone wants it back, they should figure it out.

Hey Maxime,

So can you merge this patch?

Quoted 6 lines

Maxime

--
Maxime Ripard, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

#16 lore raw
Quoted 80 lines · Vasily Khoruzhick
Quoted 77 lines · Maxime Ripard
Quoted 74 lines · Vasily Khoruzhick
Quoted 63 lines · Maxime Ripard
Quoted 61 lines · Vasily Khoruzhick
Quoted 49 lines · Maxime Ripard
Quoted 47 lines · Vasily Khoruzhick
Quoted 45 lines · Maxime Ripard
Quoted 36 lines · Vasily Khoruzhick
Quoted 26 lines · Robin Murphy
Quoted 24 lines · Vasily Khoruzhick
Quoted 22 lines · Harald Geyer
Quoted 10 lines · Vasily Khoruzhick

On Tue, Aug 6, 2019 at 7:35 AM Robin Murphy <[email protected]> wrote:

On 06/08/2019 15:01, Vasily Khoruzhick wrote:
Looks like PMU in A64 is broken, it generates no interrupts at all and
as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly?
It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

Quoted 8 lines
Quoted 2 lines

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

Quoted 4 lines

I tried to drop interrupts completely from the node but 'perf top' is still broken. Though now in different way: it complains "cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'"

I have no idea if that's the culprit, but what is the state of the 0x09010000 register?

Quoted 6 lines
Quoted 1 line

What register is that and how do I check it?

It's in the CPUX Configuration block, and the bits are labelled as CPU Debug Reset.

And if you have busybox, you can use devmem.

CPUX configuration block is at 0x01700000 according to A64 user manual, and particular register you're interested in is at 0x01700080, its value is 0x1110110F.

Quoted 5 lines
Quoted 1 line

Bits 16-19 are not defined in user manual and are not set.

Sorry, I somehow thought this was for the H6...

I don't have any idea then :/

OK, so what should we do? 'perf top'/'perf record' work fine if PMU node is dropped, but they don't work if PMU node is present (even with interrupts dropped). I'd prefer to have 'perf top' and 'perf record' working instead of 'perf stat'

Quoted 6 lines
Quoted 2 lines

Well, it doesn't work so we should just remove the node, and if someone wants it back, they should figure it out.

Hey Maxime,

So can you merge this patch?

Added new Maxime's email to CC

Quoted 6 lines
Quoted 6 lines

Maxime

--
Maxime Ripard, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#17 lore raw
Quoted 81 lines · Vasily Khoruzhick
Quoted 80 lines · Vasily Khoruzhick
Quoted 77 lines · Maxime Ripard
Quoted 74 lines · Vasily Khoruzhick
Quoted 63 lines · Maxime Ripard
Quoted 61 lines · Vasily Khoruzhick
Quoted 49 lines · Maxime Ripard
Quoted 47 lines · Vasily Khoruzhick
Quoted 45 lines · Maxime Ripard
Quoted 36 lines · Vasily Khoruzhick
Quoted 26 lines · Robin Murphy
Quoted 24 lines · Vasily Khoruzhick
Quoted 22 lines · Harald Geyer

Vasily Khoruzhick writes:
On Tue, Aug 6, 2019 at 7:35 AM Robin Murphy <[email protected]> wrote:

On 06/08/2019 15:01, Vasily Khoruzhick wrote:
Looks like PMU in A64 is broken, it generates no interrupts at all and
as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly?
It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for
the interrupts part of the node, when I added it. So I'm not too
surprised I got it wrong.

However, it would be unfortunate if the node got removed completely,
because perf stat would not work anymore. Maybe we can only remove
the interrupts or just fix them even if the HW doesn't work?

Quoted 8 lines
Quoted 2 lines

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

Quoted 4 lines

I tried to drop interrupts completely from the node but 'perf top' is still broken. Though now in different way: it complains "cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'"

I have no idea if that's the culprit, but what is the state of the 0x09010000 register?

Quoted 6 lines
Quoted 1 line

What register is that and how do I check it?

It's in the CPUX Configuration block, and the bits are labelled as CPU Debug Reset.

And if you have busybox, you can use devmem.

CPUX configuration block is at 0x01700000 according to A64 user manual, and particular register you're interested in is at 0x01700080, its value is 0x1110110F.

Quoted 5 lines
Quoted 1 line

Bits 16-19 are not defined in user manual and are not set.

Sorry, I somehow thought this was for the H6...

I don't have any idea then :/

OK, so what should we do? 'perf top'/'perf record' work fine if PMU node is dropped, but they don't work if PMU node is present (even with interrupts dropped). I'd prefer to have 'perf top' and 'perf record' working instead of 'perf stat'

Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

Well, it doesn't work so we should just remove the node, and if someone wants it back, they should figure it out.

Hey Maxime,

So can you merge this patch?

Added new Maxime's email to CC

Queued as a fix for 5.4, thanks!
Maxime

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#18 lore raw

Hi,

Just a remark here but the interrupt are from 152 to 155 SPI. But there is an offset of 32 no (remove SGI/PPI)? This should be from 120 to 123

Regards,
Clément

Quoted 98 lines · Maxime Ripard
Quoted 81 lines · Vasily Khoruzhick
Quoted 80 lines · Vasily Khoruzhick
Quoted 77 lines · Maxime Ripard
Quoted 74 lines · Vasily Khoruzhick
Quoted 63 lines · Maxime Ripard
Quoted 61 lines · Vasily Khoruzhick
Quoted 49 lines · Maxime Ripard
Quoted 47 lines · Vasily Khoruzhick
Quoted 45 lines · Maxime Ripard
Quoted 36 lines · Vasily Khoruzhick
Quoted 26 lines · Robin Murphy
Quoted 24 lines · Vasily Khoruzhick

On Tue, Aug 6, 2019 at 1:19 PM Harald Geyer <[email protected]> wrote:

Vasily Khoruzhick writes:
On Tue, Aug 6, 2019 at 7:35 AM Robin Murphy <[email protected]> wrote:

On 06/08/2019 15:01, Vasily Khoruzhick wrote:
Looks like PMU in A64 is broken, it generates no interrupts at all and
as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly?
It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for
the interrupts part of the node, when I added it. So I'm not too
surprised I got it wrong.

However, it would be unfortunate if the node got removed completely,
because perf stat would not work anymore. Maybe we can only remove
the interrupts or just fix them even if the HW doesn't work?

Quoted 8 lines
Quoted 2 lines

I'm not familiar with PMU driver. Is it possible to get it working
without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

Quoted 4 lines

I tried to drop interrupts completely from the node but 'perf top' is still broken. Though now in different way: it complains "cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'"

I have no idea if that's the culprit, but what is the state of the 0x09010000 register?

Quoted 6 lines
Quoted 1 line

What register is that and how do I check it?

It's in the CPUX Configuration block, and the bits are labelled as CPU Debug Reset.

And if you have busybox, you can use devmem.

CPUX configuration block is at 0x01700000 according to A64 user manual, and particular register you're interested in is at 0x01700080, its value is 0x1110110F.

Quoted 5 lines
Quoted 1 line

Bits 16-19 are not defined in user manual and are not set.

Sorry, I somehow thought this was for the H6...

I don't have any idea then :/

OK, so what should we do? 'perf top'/'perf record' work fine if PMU node is dropped, but they don't work if PMU node is present (even with interrupts dropped). I'd prefer to have 'perf top' and 'perf record' working instead of 'perf stat'

Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

Well, it doesn't work so we should just remove the node, and if someone wants it back, they should figure it out.

Hey Maxime,

So can you merge this patch?

Added new Maxime's email to CC

Queued as a fix for 5.4, thanks!
Maxime
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#19 lore raw
Quoted 6 lines · Clément Péron

Hi,

Just a remark here but the interrupt are from 152 to 155 SPI. But there is an offset of 32 no (remove SGI/PPI)? This should be from 120 to 123

I already tried it (and I believe someone already suggested it above), it doesn't fix PMU interrupts though.

Quoted 102 lines

Regards,
Clément

Quoted 98 lines · Maxime Ripard
Quoted 81 lines · Vasily Khoruzhick
Quoted 80 lines · Vasily Khoruzhick
Quoted 77 lines · Maxime Ripard
Quoted 74 lines · Vasily Khoruzhick
Quoted 63 lines · Maxime Ripard
Quoted 61 lines · Vasily Khoruzhick
Quoted 49 lines · Maxime Ripard
Quoted 47 lines · Vasily Khoruzhick
Quoted 45 lines · Maxime Ripard
Quoted 36 lines · Vasily Khoruzhick
Quoted 34 lines · Robin Murphy

On 2019-08-06 9:52 pm, Vasily Khoruzhick wrote:
On Tue, Aug 6, 2019 at 1:19 PM Harald Geyer <[email protected]> wrote:

Vasily Khoruzhick writes:
On Tue, Aug 6, 2019 at 7:35 AM Robin Murphy <[email protected]> wrote:

On 06/08/2019 15:01, Vasily Khoruzhick wrote:
Looks like PMU in A64 is broken, it generates no interrupts at all and
as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly?
It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for
the interrupts part of the node, when I added it. So I'm not too
surprised I got it wrong.

However, it would be unfortunate if the node got removed completely,
because perf stat would not work anymore. Maybe we can only remove
the interrupts or just fix them even if the HW doesn't work?

I'm not familiar with PMU driver. Is it possible to get it working
without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling
events (the ones which weren't working anyway), and if you measure
anything for long enough that a counter overflows you'll get wonky
results. But for counting hardware events over relatively short periods
it'll still do the job.

Quoted 4 lines

I tried to drop interrupts completely from the node but 'perf top' is still broken. Though now in different way: it complains "cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'"

I have no idea if that's the culprit, but what is the state of the 0x09010000 register?

Quoted 6 lines
Quoted 1 line

What register is that and how do I check it?

It's in the CPUX Configuration block, and the bits are labelled as CPU Debug Reset.

And if you have busybox, you can use devmem.

CPUX configuration block is at 0x01700000 according to A64 user manual, and particular register you're interested in is at 0x01700080, its value is 0x1110110F.

Quoted 5 lines
Quoted 1 line

Bits 16-19 are not defined in user manual and are not set.

Sorry, I somehow thought this was for the H6...

I don't have any idea then :/

OK, so what should we do? 'perf top'/'perf record' work fine if PMU node is dropped, but they don't work if PMU node is present (even with interrupts dropped). I'd prefer to have 'perf top' and 'perf record' working instead of 'perf stat'

Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

Well, it doesn't work so we should just remove the node, and if someone wants it back, they should figure it out.

Hey Maxime,

So can you merge this patch?

Added new Maxime's email to CC

Queued as a fix for 5.4, thanks!
Maxime
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#20 lore raw

Hi

Quoted 11 lines · Vasily Khoruzhick
Quoted 6 lines · Clément Péron

Hi,

Just a remark here but the interrupt are from 152 to 155 SPI. But there is an offset of 32 no (remove SGI/PPI)? This should be from 120 to 123

I already tried it (and I believe someone already suggested it above), it doesn't fix PMU interrupts though.

Ok thanks for the confirmation.

Made a research about the PMU for A64 and found that Andre Przywara made a patch to enable it:
https://gist.github.com/apritzel/d025abaa1425fcaf5991b5ffcf18a0a3

Maybe he can confirm or not the issue on A64 ?

Regards,
Clément

Quoted 103 lines
Quoted 103 lines

Regards,
Clément

Quoted 98 lines · Maxime Ripard
Quoted 81 lines · Vasily Khoruzhick
Quoted 80 lines · Vasily Khoruzhick
Quoted 77 lines · Maxime Ripard
Quoted 74 lines · Vasily Khoruzhick
Quoted 63 lines · Maxime Ripard
Quoted 61 lines · Vasily Khoruzhick
Quoted 49 lines · Maxime Ripard
Quoted 47 lines · Vasily Khoruzhick
Quoted 45 lines · Maxime Ripard
Quoted 40 lines · Vasily Khoruzhick

On Tue, Aug 6, 2019 at 2:14 PM Robin Murphy <[email protected]> wrote:

On 2019-08-06 9:52 pm, Vasily Khoruzhick wrote:
On Tue, Aug 6, 2019 at 1:19 PM Harald Geyer <[email protected]> wrote:

Vasily Khoruzhick writes:
On Tue, Aug 6, 2019 at 7:35 AM Robin Murphy <[email protected]> wrote:

On 06/08/2019 15:01, Vasily Khoruzhick wrote:
Looks like PMU in A64 is broken, it generates no interrupts at all and
as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly?
It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for
the interrupts part of the node, when I added it. So I'm not too
surprised I got it wrong.

However, it would be unfortunate if the node got removed completely,
because perf stat would not work anymore. Maybe we can only remove
the interrupts or just fix them even if the HW doesn't work?

I'm not familiar with PMU driver. Is it possible to get it working
without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling
events (the ones which weren't working anyway), and if you measure
anything for long enough that a counter overflows you'll get wonky
results. But for counting hardware events over relatively short periods
it'll still do the job.

I tried to drop interrupts completely from the node but 'perf top' is
still broken. Though now in different way: it complains "cycles: PMU
Hardware doesn't support sampling/overflow-interrupts. Try 'perf
stat'"

I have no idea if that's the culprit, but what is the state of the 0x09010000 register?

Quoted 6 lines
Quoted 1 line

What register is that and how do I check it?

It's in the CPUX Configuration block, and the bits are labelled as CPU Debug Reset.

And if you have busybox, you can use devmem.

CPUX configuration block is at 0x01700000 according to A64 user manual, and particular register you're interested in is at 0x01700080, its value is 0x1110110F.

Quoted 5 lines
Quoted 1 line

Bits 16-19 are not defined in user manual and are not set.

Sorry, I somehow thought this was for the H6...

I don't have any idea then :/

OK, so what should we do? 'perf top'/'perf record' work fine if PMU node is dropped, but they don't work if PMU node is present (even with interrupts dropped). I'd prefer to have 'perf top' and 'perf record' working instead of 'perf stat'

Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

Well, it doesn't work so we should just remove the node, and if someone wants it back, they should figure it out.

Hey Maxime,

So can you merge this patch?

Added new Maxime's email to CC

Queued as a fix for 5.4, thanks!
Maxime
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#21 lore raw

On 11/1/19 11:30 AM, Clément Péron wrote:

Hi,

Quoted 13 lines · Vasily Khoruzhick
Quoted 11 lines · Clément Péron
Quoted 6 lines

Hi,

Just a remark here but the interrupt are from 152 to 155 SPI. But there is an offset of 32 no (remove SGI/PPI)? This should be from 120 to 123

I already tried it (and I believe someone already suggested it above), it doesn't fix PMU interrupts though.

Quoted 7 lines

Ok thanks for the confirmation.

Made a research about the PMU for A64 and found that Andre Przywara made a patch to enable it:
https://gist.github.com/apritzel/d025abaa1425fcaf5991b5ffcf18a0a3

Maybe he can confirm or not the issue on A64 ?

Well, I tried it back then, but couldn't make the interrupts work (and yes, I tried +/- 32). That's the reason I didn't send it back then.

I can't say whether the IRQ lines are not wired or the manual just gives the wrong numbers. I don't have access to a board before Sunday, but if someone wants to beat me to it:
- Hack U-Boot to add a command to program one PMU counter to expire quickly, and enable overflow interrupts.
- Enable *all* SPIs on the GIC distributor level, and enable the distributor. Keep the GIC CPU interface disabled.
- Trigger the U-Boot command, and inspect the GICD_ISPENDR registers to see if any SPI fired.
- Also double check the PMU overflow status register to verify that the event triggered.

Cheers,
Andre.

Quoted 107 lines

Regards,
Clément

Quoted 103 lines
Quoted 103 lines

Regards,
Clément

Quoted 98 lines · Maxime Ripard
Quoted 81 lines · Vasily Khoruzhick
Quoted 80 lines · Vasily Khoruzhick
Quoted 77 lines · Maxime Ripard
Quoted 74 lines · Vasily Khoruzhick
Quoted 63 lines · Maxime Ripard
Quoted 61 lines · Vasily Khoruzhick
Quoted 49 lines · Maxime Ripard
Quoted 47 lines · Vasily Khoruzhick
Quoted 45 lines · Maxime Ripard

On Tue, Aug 06, 2019 at 07:39:26PM -0700, Vasily Khoruzhick wrote:
On Tue, Aug 6, 2019 at 2:14 PM Robin Murphy <[email protected]> wrote:

On 2019-08-06 9:52 pm, Vasily Khoruzhick wrote:
On Tue, Aug 6, 2019 at 1:19 PM Harald Geyer <[email protected]> wrote:

Vasily Khoruzhick writes:
On Tue, Aug 6, 2019 at 7:35 AM Robin Murphy <[email protected]> wrote:

On 06/08/2019 15:01, Vasily Khoruzhick wrote:
Looks like PMU in A64 is broken, it generates no interrupts at all and
as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly?
It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for
the interrupts part of the node, when I added it. So I'm not too
surprised I got it wrong.

However, it would be unfortunate if the node got removed completely,
because perf stat would not work anymore. Maybe we can only remove
the interrupts or just fix them even if the HW doesn't work?

I'm not familiar with PMU driver. Is it possible to get it working
without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling
events (the ones which weren't working anyway), and if you measure
anything for long enough that a counter overflows you'll get wonky
results. But for counting hardware events over relatively short periods
it'll still do the job.

I tried to drop interrupts completely from the node but 'perf top' is
still broken. Though now in different way: it complains "cycles: PMU
Hardware doesn't support sampling/overflow-interrupts. Try 'perf
stat'"

I have no idea if that's the culprit, but what is the state of the
0x09010000 register?

Quoted 6 lines
Quoted 1 line

What register is that and how do I check it?

It's in the CPUX Configuration block, and the bits are labelled as CPU Debug Reset.

And if you have busybox, you can use devmem.

CPUX configuration block is at 0x01700000 according to A64 user manual, and particular register you're interested in is at 0x01700080, its value is 0x1110110F.

Quoted 5 lines
Quoted 1 line

Bits 16-19 are not defined in user manual and are not set.

Sorry, I somehow thought this was for the H6...

I don't have any idea then :/

OK, so what should we do? 'perf top'/'perf record' work fine if PMU node is dropped, but they don't work if PMU node is present (even with interrupts dropped). I'd prefer to have 'perf top' and 'perf record' working instead of 'perf stat'

Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

Well, it doesn't work so we should just remove the node, and if someone wants it back, they should figure it out.

Hey Maxime,

So can you merge this patch?

Added new Maxime's email to CC

Queued as a fix for 5.4, thanks!
Maxime
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#22 lore raw
Quoted 1 line · Andre Przywara

On 11/1/19 11:30 AM, Clément Péron wrote:

Hi,

for the sake of finishing this thread, in case search engines point to it:

Quoted 17 lines
Quoted 17 lines
Quoted 4 lines · Vasily Khoruzhick
Quoted 1 line · Clément Péron
Quoted 8 lines
Quoted 5 lines

Hi,

Just a remark here but the interrupt are from 152 to 155 SPI. But there is an offset of 32 no (remove SGI/PPI)? This should be from 120 to 123

I already tried it (and I believe someone already suggested it above), it doesn't fix PMU interrupts though.

Ok thanks for the confirmation.

Quoted 8 lines
Quoted 5 lines

Made a research about the PMU for A64 and found that Andre Przywara made a patch to enable it:
https://gist.github.com/apritzel/d025abaa1425fcaf5991b5ffcf18a0a3

Maybe he can confirm or not the issue on A64 ?

Well, I tried it back then, but couldn't make the interrupts work (and yes, I tried +/- 32). That's the reason I didn't send it back then.

I did the experiment drafted below, and found that the interrupts are not 152-155, as the manual describes, but 148-151 instead. And yes, those are the GIC interrupt IDs used, but the GIC binding requires us to give the SPI numbers, so we need to subtract 32, which was not the case for the original patch.
Thanks to Maxime for merging the fix patch:
https://archive.armlinux.org.uk/lurker/message/20191105.110651.914455de.en.html

In case someone wants to confirm this or in general needs to find/confirm IRQ numbers for some peripherals:

Quoted 5 lines

I can't say whether the IRQ lines are not wired or the manual just gives the wrong numbers. I don't have access to a board before Sunday, but if someone wants to beat me to it:
- Hack U-Boot to add a command to program one PMU counter to expire quickly, and enable overflow interrupts.

Attaching a U-Boot patch to trigger a PMU cycle counter overflow IRQ.

Quoted 2 lines

- Enable *all* SPIs on the GIC distributor level, and enable the distributor. Keep the GIC CPU interface disabled.

The GIC distributor on most Allwinner SoCs is located at 0x01c81000. This is the first address shown in the GIC's DT node.
To enable all IRQs in U-Boot, on the U-Boot command prompt:
=> mw.l 0x01c81100 0xffffffff 0x20
This will set all possible 1024 bits in the GICD_ISENABLERn registers (offset 0x100) to 1. Reading this back will reveal which IRQs the GIC actually is configured for, typically we have much fewer SPIs supported. To enable the distributor (really: group 1 interrupts):
=> mw.l 0x01c81000 1
As suggested, we don't touch the GIC CPU interface, so no interrupts will actually reach the core.

Quoted 2 lines

- Trigger the U-Boot command, and inspect the GICD_ISPENDR registers to see if any SPI fired.

Trigger the IRQ, for instance using the command provided by the patch above. Keep the PMU enabled, because it will deassert the IRQ lines otherwise:

=> pmuc start
=> pmuc
(check that the counter overflows)
=> md.l 0x01c81200 0x20

This will print the pending bits (GICD_ISPENDRn at offset 0x200) for each IRQ. In U-Boot we expect no other IRQs to ever fire, so any bit set in here would be due to our experiment.
In this case the output looked like:
01c81200: 00000000 00000000 00000000 00000000
01c81210: 00100000 00000000 00000000 00000000
If you do the counting, this is bit 20 in word 4, so 4 * 32 + 20 = 148. Subtract 32 to get the number that goes into the DT node. U-Boot is UP, so this is probably the IRQ on core 0. A reasonable assumption is that the other cores just follow behind this, which you can confirm in Linux, by running some perf command on only a single core and watching the interrupt count increasing:

$ grep pmu /proc/interrupts
$ taskset -cp 1 $$
$ perf record sleep 3
$ grep pmu /proc/interrupts
Quoted 2 lines

- Also double check the PMU overflow status register to verify that the event triggered.

Just calling "pmuc" will print the PMOVSCLR register.

Hope that helps!

Cheers,
Andre

Quoted 124 lines
Quoted 124 lines
Quoted 124 lines
Quoted 124 lines · Maxime Ripard
Quoted 123 lines · Vasily Khoruzhick
Quoted 121 lines
Quoted 118 lines · Vasily Khoruzhick
Quoted 109 lines · Maxime Ripard
Quoted 99 lines · Vasily Khoruzhick
Quoted 91 lines · Maxime Ripard
Quoted 88 lines · Vasily Khoruzhick
Quoted 73 lines · Maxime Ripard
Quoted 70 lines · Vasily Khoruzhick

On Wed, Aug 7, 2019 at 4:56 AM Maxime Ripard
<[email protected]> wrote:

On Tue, Aug 06, 2019 at 07:39:26PM -0700, Vasily
Khoruzhick wrote:
On Tue, Aug 6, 2019 at 2:14 PM Robin Murphy
<[email protected]> wrote:

On 2019-08-06 9:52 pm, Vasily Khoruzhick wrote:
On Tue, Aug 6, 2019 at 1:19 PM Harald Geyer
<[email protected]> wrote:

Vasily Khoruzhick writes:
On Tue, Aug 6, 2019 at 7:35 AM Robin Murphy
<[email protected]> wrote:

On 06/08/2019 15:01, Vasily Khoruzhick wrote:
Looks like PMU in A64 is broken, it generates no
interrupts at all and
as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least
count cycles correctly?
It could well just be that the interrupt numbers are
wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and
it works fine.

Unfortunately I wasn't aware that perf stat is a poor
test for
the interrupts part of the node, when I added it. So
I'm not too
surprised I got it wrong.

However, it would be unfortunate if the node got
removed completely,
because perf stat would not work anymore. Maybe we can
only remove
the interrupts or just fix them even if the HW doesn't
work?

I'm not familiar with PMU driver. Is it possible to get
it working
without interrupts?

Yup - you get a grumpy message from the driver, it will
refuse sampling
events (the ones which weren't working anyway), and if
you measure
anything for long enough that a counter overflows you'll
get wonky
results. But for counting hardware events over
relatively short periods
it'll still do the job.

I tried to drop interrupts completely from the node but
'perf top' is
still broken. Though now in different way: it complains
"cycles: PMU
Hardware doesn't support sampling/overflow-interrupts.
Try 'perf
stat'"

I have no idea if that's the culprit, but what is the
state of the
0x09010000 register?

Quoted 7 lines
Quoted 1 line

What register is that and how do I check it?

It's in the CPUX Configuration block, and the bits are labelled as CPU
Debug Reset.

And if you have busybox, you can use devmem.

CPUX configuration block is at 0x01700000 according to A64 user manual, and particular register you're interested in is at 0x01700080,
its value is 0x1110110F.

Quoted 5 lines
Quoted 1 line

Bits 16-19 are not defined in user manual and are not set.

Sorry, I somehow thought this was for the H6...

I don't have any idea then :/

Quoted 7 lines

OK, so what should we do? 'perf top'/'perf record' work fine if PMU
node is dropped, but they don't work if PMU node is present (even with
interrupts dropped). I'd prefer to have 'perf top' and 'perf record'
working instead of 'perf stat'

Quoted 2 lines

Well, it doesn't work so we should just remove the node, and if someone wants it back, they should figure it out.

Hey Maxime,

So can you merge this patch?

Quoted 9 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 1 line

Added new Maxime's email to CC

Queued as a fix for 5.4, thanks!
Maxime
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

From 169f01c8545a68b5f36515c49d171b9e0555ec1b Mon Sep 17 00:00:00 2001
From: Andre Przywara <[email protected]>
Date: Mon, 4 Nov 2019 21:54:51 +0000
Subject: [PATCH] arm64: Add performance monitoring unit test command

Mostly useful for identifying the interrupts used.

---
 cmd/Makefile  |  1 +
 cmd/arm_pmu.c | 78 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 79 insertions(+)
 create mode 100644 cmd/arm_pmu.c
cmd/Makefile +1 -0
index ac843b4b16..bced333222 100644
@@ -182,6 +182,7 @@ endif # !CONFIG_SPL_BUILD
 
 # core command
 obj-y += nvedit.o
+obj-$(CONFIG_ARM64) += arm_pmu.o
 
 obj-$(CONFIG_TI_COMMON_CMD_OPTIONS) += ti/
 
cmd/arm_pmu.c new file +78 -0

78 lines not on this page · raw message

Signature · 1 line
2.14.5
#23 lore raw
In reply to Vasily Khoruzhick
Quoted 36 lines · Vasily Khoruzhick
Quoted 26 lines · Robin Murphy
Quoted 24 lines · Vasily Khoruzhick
Quoted 22 lines · Harald Geyer
Quoted 10 lines · Vasily Khoruzhick
Quoted 7 lines · Robin Murphy
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

Quoted 8 lines
Quoted 2 lines

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

Quoted 6 lines

I tried to drop interrupts completely from the node but 'perf top' is still broken. Though now in different way: it complains "cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'"

Is there any way to make it working?

As the message implies, 'perf top' can't work because it uses sampling events, which are based on periodic interrupts. If the IRQs aren't there, then too bad, as there's no alternative.

One other possibility is that the IRQs really are wired up, but the firmware is somehow leaving them configured as Secure group 0, such that Linux has no visibility of them.

Robin.

#24 lore raw
In reply to Robin Murphy
Quoted 25 lines · Robin Murphy
Quoted 24 lines · Vasily Khoruzhick
Quoted 22 lines · Harald Geyer
Quoted 10 lines · Vasily Khoruzhick
Quoted 7 lines · Robin Murphy
Quoted 2 lines · Vasily Khoruzhick

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Does something like 'perf stat sleep 1' at least count cycles correctly? It could well just be that the interrupt numbers are wrong...

Looks like it does, at least result looks plausible:

I'm using perf stat regularly (cache benchmarks) and it works fine.

Unfortunately I wasn't aware that perf stat is a poor test for the interrupts part of the node, when I added it. So I'm not too surprised I got it wrong.

However, it would be unfortunate if the node got removed completely, because perf stat would not work anymore. Maybe we can only remove the interrupts or just fix them even if the HW doesn't work?

Quoted 8 lines
Quoted 2 lines

I'm not familiar with PMU driver. Is it possible to get it working without interrupts?

Yup - you get a grumpy message from the driver, it will refuse sampling events (the ones which weren't working anyway), and if you measure anything for long enough that a counter overflows you'll get wonky results. But for counting hardware events over relatively short periods it'll still do the job.

Even that's stupidly dodgy; a CPU_CYCLES event could easily overflow several times between the kernel sampling it, so you can lose billions of events without any idea that happened.

For other PMUs we can fix that with a hrtimer, but I think for a CPU PMU it has to be at such a high frequency that it imposes a ridiculous overhead, even assuming we can choose a sufficient frequency. :/

Thanks,
Mark.

Reviewed-by on v1 1/1
#25 lore raw
In reply to Vasily Khoruzhick
Quoted 35 lines · On Tue, 6 Aug 2019 07:01:35 -0700 Vasily Khoruzhick <[email protected]> wrote:

Looks like PMU in A64 is broken, it generates no interrupts at all and as result 'perf top' shows no events.

Tested on Pine64-LTS.

Fixes 34a97fcc71c2 ("arm64: dts: allwinner: a64: Add PMU node")
Cc Harald Geyer <[email protected]>
Cc Jared D. McNeill <[email protected]>
---
 arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi | 9 ---------
 1 file changed, 9 deletions(-)
arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi +0 -9
index 9cc9bdde81ac..cd92f546c483 100644
@@ -142,15 +142,6 @@
 		clock-output-names = "ext-osc32k";
 	};
 
-	pmu {
-		compatible = "arm,cortex-a53-pmu";
-		interrupts = <GIC_SPI 152 IRQ_TYPE_LEVEL_HIGH>,
-			     <GIC_SPI 153 IRQ_TYPE_LEVEL_HIGH>,
-			     <GIC_SPI 154 IRQ_TYPE_LEVEL_HIGH>,
-			     <GIC_SPI 155 IRQ_TYPE_LEVEL_HIGH>;
-		interrupt-affinity = <&cpu0>, <&cpu1>, <&cpu2>, <&cpu3>;
-	};
-
 	psci {
 		compatible = "arm,psci-0.2";
 		method = "smc";
Signature · 1 line
2.22.0

It doesn't work for me too on FreeBSD, and yes the interrupts are
wrong but this is not the problem. Maybe there is a reset line but it's not documented in the documentation.

Signature · 1 line
Emmanuel Vadot <[email protected]> <[email protected]>