Skip to contentSkip to conversation

arm64: perf: Ensure EL0 access is disabled at reset

PATCH In review Rob Herring 2 participants 3 messages Apr 27, 2021

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

PATCH 1/1 arm64: perf: Ensure EL0 access is disabled at reset

The ER, SW, and EN bits in the PMUSERENR_EL0 register are UNKNOWN at reset and the register is never initialized, so EL0 access could be enabled by default on some implementations. Let's initialize PMUSERENR_EL0 to a known state with EL0 access disabled.

---
 arch/arm64/kernel/perf_event.c | 8 ++++++++
 1 file changed, 8 insertions(+)
arch/arm64/kernel/perf_event.c +8 -0
index 4658fcf88c2b..c32778ae5117 100644
@@ -450,6 +450,11 @@ static inline void armv8pmu_pmcr_write(u32 val)
 	write_sysreg(val, pmcr_el0);
 }
 
+static void armv8pmu_clear_pmuserenr(void)
+{
+	write_sysreg(0, pmuserenr_el0);
+}
+
 static inline int armv8pmu_has_overflowed(u32 pmovsr)
 {
 	return pmovsr & ARMV8_PMU_OVERFLOWED_MASK;
@@ -933,6 +938,9 @@ static void armv8pmu_reset(void *info)
 	armv8pmu_disable_counter(U32_MAX);
 	armv8pmu_disable_intens(U32_MAX);
 
+	/* User access is unknown at reset. */
+	armv8pmu_clear_pmuserenr();
+
 	/* Clear the counters we flip at guest entry/exit */
 	kvm_clr_pmu_events(U32_MAX);
 
Signature · 1 line
2.27.0
maintainer
#1 lore raw
Quoted 4 lines · Rob Herring

The ER, SW, and EN bits in the PMUSERENR_EL0 register are UNKNOWN at reset and the register is never initialized, so EL0 access could be enabled by default on some implementations. Let's initialize PMUSERENR_EL0 to a known state with EL0 access disabled.

We reset PMUSERENR_EL0 via the reset_pmuserenr_el0 macro, called from __cpu_setup when a CPU is onlined and from cpu_do_resume() when a CPU returns from a context-destructive idle state. We do it there so that it's handled even if a kernel isn't built with perf support.

AFAICT, that *should* do the right thing -- are you seeing UNKNOWN values, or was this found by inspection?

Thanks,
Mark.

Quoted 35 lines
---
 arch/arm64/kernel/perf_event.c | 8 ++++++++
 1 file changed, 8 insertions(+)
arch/arm64/kernel/perf_event.c +8 -0
index 4658fcf88c2b..c32778ae5117 100644
@@ -450,6 +450,11 @@ static inline void armv8pmu_pmcr_write(u32 val)
 	write_sysreg(val, pmcr_el0);
 }
 
+static void armv8pmu_clear_pmuserenr(void)
+{
+	write_sysreg(0, pmuserenr_el0);
+}
+
 static inline int armv8pmu_has_overflowed(u32 pmovsr)
 {
 	return pmovsr & ARMV8_PMU_OVERFLOWED_MASK;
@@ -933,6 +938,9 @@ static void armv8pmu_reset(void *info)
 	armv8pmu_disable_counter(U32_MAX);
 	armv8pmu_disable_intens(U32_MAX);
 
+	/* User access is unknown at reset. */
+	armv8pmu_clear_pmuserenr();
+
 	/* Clear the counters we flip at guest entry/exit */
 	kvm_clr_pmu_events(U32_MAX);
 
Signature · 1 line
2.27.0
#2 lore raw
Quoted 11 lines · Mark Rutland
Quoted 4 lines · Rob Herring

The ER, SW, and EN bits in the PMUSERENR_EL0 register are UNKNOWN at reset and the register is never initialized, so EL0 access could be enabled by default on some implementations. Let's initialize PMUSERENR_EL0 to a known state with EL0 access disabled.

We reset PMUSERENR_EL0 via the reset_pmuserenr_el0 macro, called from __cpu_setup when a CPU is onlined and from cpu_do_resume() when a CPU returns from a context-destructive idle state. We do it there so that it's handled even if a kernel isn't built with perf support.

Indeed.

Quoted 2 lines

AFAICT, that *should* do the right thing -- are you seeing UNKNOWN values, or was this found by inspection?

Inspection. Sorry for the noise.

Rob