Skip to contentSkip to conversation

perf: Change PMCR write to read-modify-write

PATCH In review Prasad Sodagudi 4 participants 6 messages Mar 21, 2019 – Apr 2, 2019

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf: Change PMCR write to read-modify-write

Preserves the bitfields of PMCR_EL0(AArch64) during PMU reset. Reset routine should write a 1 to PMCR.C and PMCR.P fields only to reset the counters. Other fields should not be changed as they could be set before PMU initialization and their value must be preserved even after reset.

---
 arch/arm64/kernel/perf_event.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
arch/arm64/kernel/perf_event.c +2 -2
index 4addb38..0c1afdd 100644
@@ -868,8 +868,8 @@ static void armv8pmu_reset(void *info)
 	 * Initialize & Reset PMNC. Request overflow interrupt for
 	 * 64 bit cycle counter but cheat in armv8pmu_write_counter().
 	 */
-	armv8pmu_pmcr_write(ARMV8_PMU_PMCR_P | ARMV8_PMU_PMCR_C |
-			    ARMV8_PMU_PMCR_LC);
+	armv8pmu_pmcr_write(armv8pmu_pmcr_read() | ARMV8_PMU_PMCR_P |
+			ARMV8_PMU_PMCR_C | ARMV8_PMU_PMCR_LC);
 }
 
 static int __armv8_pmuv3_map_event(struct perf_event *event,
Signature · 1 line
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,\na Linux Foundation Collaborative Project
#1 lore raw

Hi Prasad,

Quoted 6 lines · Prasad Sodagudi

Preserves the bitfields of PMCR_EL0(AArch64) during PMU reset. Reset routine should write a 1 to PMCR.C and PMCR.P fields only to reset the counters. Other fields should not be changed as they could be set before PMU initialization and their value must be preserved even after reset.

Are there any particular bit you are concerned about? Apart from the RO ones and the Res0 ones (to which we are already writing 0), I see:

DP -> irrelevant for non-secure
X -> This one is the only potentially interesting, however it resets to an architecturally unknown value, so unless we know for a fact it was set before hand, we probably want to clear it
D -> ignored when we have LC set (and we do)
E -> Since this is the function we use to reset the PMU on the current CPU, we probably want to set this bit to 0 regardless of its previous value

So, is there any issue this patch is solving?

Thanks,

Quoted 21 lines
---
 arch/arm64/kernel/perf_event.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
arch/arm64/kernel/perf_event.c +2 -2
index 4addb38..0c1afdd 100644
@@ -868,8 +868,8 @@ static void armv8pmu_reset(void *info)
 	 * Initialize & Reset PMNC. Request overflow interrupt for
 	 * 64 bit cycle counter but cheat in armv8pmu_write_counter().
 	 */
-	armv8pmu_pmcr_write(ARMV8_PMU_PMCR_P | ARMV8_PMU_PMCR_C |
-			    ARMV8_PMU_PMCR_LC);
+	armv8pmu_pmcr_write(armv8pmu_pmcr_read() | ARMV8_PMU_PMCR_P |
+			ARMV8_PMU_PMCR_C | ARMV8_PMU_PMCR_LC);
 }
 
 static int __armv8_pmuv3_map_event(struct perf_event *event,
Signature · 1 line
Julien Thierry
#2 lore raw
Quoted 21 lines · Julien Thierry

Hi Prasad,

Quoted 6 lines · Prasad Sodagudi

Preserves the bitfields of PMCR_EL0(AArch64) during PMU reset. Reset routine should write a 1 to PMCR.C and PMCR.P fields only to reset the counters. Other fields should not be changed as they could be set before PMU initialization and their value must be preserved even after reset.

Are there any particular bit you are concerned about? Apart from the RO ones and the Res0 ones (to which we are already writing 0), I see:

DP -> irrelevant for non-secure
X -> This one is the only potentially interesting, however it resets to an architecturally unknown value, so unless we know for a fact it was set before hand, we probably want to clear it
D -> ignored when we have LC set (and we do)
E -> Since this is the function we use to reset the PMU on the current CPU, we probably want to set this bit to 0 regardless of its previous value

Quoted 1 line

So, is there any issue this patch is solving?

Hi Julien,

Thanks for taking a look into this patch. Yes. On our Qualcomm platforms, observed that X bit is getting cleared by kernel. This bit is getting set by firmware for Qualcomm use cases and non-secure world is resetting without this patch. I think, changing that register this register modifications to read-modify-write style make sense.

-Thanks, Prasad

Quoted 24 lines

Thanks,

Quoted 21 lines
---
 arch/arm64/kernel/perf_event.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
arch/arm64/kernel/perf_event.c +2 -2
index 4addb38..0c1afdd 100644
@@ -868,8 +868,8 @@ static void armv8pmu_reset(void *info)
 	 * Initialize & Reset PMNC. Request overflow interrupt for
 	 * 64 bit cycle counter but cheat in armv8pmu_write_counter().
 	 */
-	armv8pmu_pmcr_write(ARMV8_PMU_PMCR_P | ARMV8_PMU_PMCR_C |
-			    ARMV8_PMU_PMCR_LC);
+	armv8pmu_pmcr_write(armv8pmu_pmcr_read() | ARMV8_PMU_PMCR_P |
+			ARMV8_PMU_PMCR_C | ARMV8_PMU_PMCR_LC);
 }
 
 static int __armv8_pmuv3_map_event(struct perf_event *event,
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
Linux Foundation Collaborative Project
#3 lore raw
Quoted 25 lines · Sodagudi Prasad
Quoted 22 lines · Julien Thierry

Hi Prasad,

Quoted 6 lines · Prasad Sodagudi

Preserves the bitfields of PMCR_EL0(AArch64) during PMU reset. Reset routine should write a 1 to PMCR.C and PMCR.P fields only to reset the counters. Other fields should not be changed as they could be set before PMU initialization and their value must be preserved even after reset.

Are there any particular bit you are concerned about? Apart from the RO ones and the Res0 ones (to which we are already writing 0), I see:

DP -> irrelevant for non-secure
X -> This one is the only potentially interesting, however it resets to an architecturally unknown value, so unless we know for a fact it was set before hand, we probably want to clear it
D -> ignored when we have LC set (and we do)
E -> Since this is the function we use to reset the PMU on the current CPU, we probably want to set this bit to 0 regardless of its previous value

Quoted 1 line

So, is there any issue this patch is solving?

Quoted 9 lines

Hi Julien,

Thanks for taking a look into this patch. Yes. On our Qualcomm platforms, observed that X bit is getting cleared by kernel. This bit is getting set by firmware for Qualcomm use cases and non-secure world is resetting without this patch.
I think, changing that register this register modifications to read-modify-write style make sense.

Maybe for the X bit, but for the E bit this seems like the wrong thing to do. We want to set the E bit to 0 here.

And for platforms that don't have a firmware touching the X bit (or rather the pmcr as a whole), I'd like to understand whether it would be valid to leave this bit set to an architecturally unknown value and preserving that value.

Thanks,

Quoted 23 lines
Quoted 22 lines
Quoted 22 lines

---
 arch/arm64/kernel/perf_event.c | 4 ++--

 1 file changed, 2 insertions(+), 2 deletions(-)
arch/arm64/kernel/perf_event.c +2 -2
index 4addb38..0c1afdd 100644
@@ -868,8 +868,8 @@ static void armv8pmu_reset(void *info)
       * Initialize & Reset PMNC. Request overflow interrupt for
       * 64 bit cycle counter but cheat in armv8pmu_write_counter().
       */
-    armv8pmu_pmcr_write(ARMV8_PMU_PMCR_P | ARMV8_PMU_PMCR_C |
-                ARMV8_PMU_PMCR_LC);
+    armv8pmu_pmcr_write(armv8pmu_pmcr_read() | ARMV8_PMU_PMCR_P |
+            ARMV8_PMU_PMCR_C | ARMV8_PMU_PMCR_LC);
  }
 
  static int __armv8_pmuv3_map_event(struct perf_event *event,
Signature · 1 line
Julien Thierry
maintainer
#4 lore raw
In reply to Sodagudi Prasad
Quoted 22 lines · Sodagudi Prasad
Quoted 21 lines · Julien Thierry

Hi Prasad,

Quoted 5 lines · Prasad Sodagudi

Preserves the bitfields of PMCR_EL0(AArch64) during PMU reset. Reset routine should write a 1 to PMCR.C and PMCR.P fields only to reset the counters. Other fields should not be changed as they could be set before PMU initialization and their value must be preserved even after reset.

Are there any particular bit you are concerned about? Apart from the RO ones and the Res0 ones (to which we are already writing 0), I see:

DP -> irrelevant for non-secure
X -> This one is the only potentially interesting, however it resets to an architecturally unknown value, so unless we know for a fact it was set before hand, we probably want to clear it
D -> ignored when we have LC set (and we do)
E -> Since this is the function we use to reset the PMU on the current CPU, we probably want to set this bit to 0 regardless of its previous value

Quoted 8 lines
Quoted 1 line

So, is there any issue this patch is solving?

Hi Julien,

Thanks for taking a look into this patch. Yes. On our Qualcomm platforms, observed that X bit is getting cleared by kernel.
This bit is getting set by firmware for Qualcomm use cases and non-secure world is resetting without this patch.

Given the kernel reconfigures the PMU dynamically at runtime (e.g. to multiplex events over counters), I don't follow how this is useful. Could you elaborate on how specifically you intend to use this?

Quoted 2 lines

I think, changing that register this register modifications to read-modify-write style make sense.

I disagree.

In general, RESx bits may be allocated new meanings in future (and will reset to UNKNOWN values), so it is not safe to simply preserve the reset value. Given that, any kernel _must_ explicitly reset registers in order to ensure the expected behaviour.

Thanks
Mark.

#5 lore raw
In reply to Prasad Sodagudi

[+MarkR]

Quoted 16 lines · Prasad Sodagudi

Preserves the bitfields of PMCR_EL0(AArch64) during PMU reset. Reset routine should write a 1 to PMCR.C and PMCR.P fields only to reset the counters. Other fields should not be changed as they could be set before PMU initialization and their value must be preserved even after reset.

---
 arch/arm64/kernel/perf_event.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
arch/arm64/kernel/perf_event.c +0 -0
index 4addb38..0c1afdd 100644

arch/arm64/kernel/perf_event.c static void armv8pmu_reset(void *info) · line 872

Quoted 7 lines
arch/arm64/kernel/perf_event.c +2 -2
@@ -868,8 +868,8 @@ static void armv8pmu_reset(void *info)
 	 * Initialize & Reset PMNC. Request overflow interrupt for
 	 * 64 bit cycle counter but cheat in armv8pmu_write_counter().
 	 */
-	armv8pmu_pmcr_write(ARMV8_PMU_PMCR_P | ARMV8_PMU_PMCR_C |
-			    ARMV8_PMU_PMCR_LC);
+	armv8pmu_pmcr_write(armv8pmu_pmcr_read() | ARMV8_PMU_PMCR_P |
+			ARMV8_PMU_PMCR_C | ARMV8_PMU_PMCR_LC);

No, I don't think this is right at all. Many of these bits are UNKNOWN out of reset, so you're throwing away our means of getting the PMU into a deterministic state.

Who is setting which bits that you want to preserve and why?

Will