arm: pmu-overflow-interrupt: Fix failures on Amberwing
v2 cover arm: pmu-overflow-interrupt: Fix failures on Amberwing
On Qualcomm Amberwing, some pmu-overflow-interrupt failures can be observed. Although the even counter overflows, the interrupt is not seen as expected on guest side. This happens in the subtest after "promote to 64-b" comment.
After analysis, the PMU overflow interrupt actually hits, ie. kvm_pmu_perf_overflow() gets called and KVM_REQ_IRQ_PENDING is set, as expected. However the PMCR.E is reset by the handle_exit path, at kvm_pmu_handle_pmcr() before the next guest entry and kvm_pmu_flush_hwstate/kvm_pmu_update_state subsequent call. There, since the enable bit has been reset, kvm_pmu_update_state() does not inject the interrupt into the guest.
This does not seem to be a KVM bug but rather an unfortunate scenario where the test disables the PMCR.E too closely to the advent of the overflow interrupt.
Since it looks like a benign and inlikely case, let's resize the number of iterations to prevent the PMCR enable bit from being resetted immediately at the same time as the actual overflow event.
Also make pmu_stats volatile to prevent any optimizations.
Eric Auger (2):
arm: pmu: Declare pmu_stats as volatile
arm: pmu-overflow-interrupt: Increase count values
arm/pmu.c | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-)
Signature · 1 line
2.41.0
PATCH v2 1/2 arm: pmu: Declare pmu_stats as volatile
Declare pmu_stats as volatile in order to prevent the compiler from caching previously read values. This actually fixes pmu-overflow-interrupt failures on some HW.
--- arm/pmu.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arm/pmu.c +1 -1
@@ -328,7 +328,7 @@ asm volatile( : "x9", "x10", "cc"); } -static struct pmu_stats pmu_stats; +static volatile struct pmu_stats pmu_stats;static void irq_handler (struct pt_regs *regs) {
Signature · 1 line
2.41.0
Re: [kvm-unit-tests PATCH v2 1/2] arm: pmu: Declare pmu_stats as volatile
Quoted 6 lines · Eric Auger
Declare
pmu_statsas volatile in order to prevent the compiler from caching previously read values. This actually fixes pmu-overflow-interrupt failures on some HW.Reported-by Alexandru Elisei <[email protected]>Signed-off-by Eric Auger <[email protected]>
Quoted 17 lines
--- arm/pmu.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)arm/pmu.c +1
-1@@ -328,7 +328,7 @@ asm volatile( : "x9", "x10", "cc"); } -static struct pmu_stats pmu_stats; +static volatile struct pmu_stats pmu_stats;static void irq_handler (struct pt_regs *regs) {
Signature · 1 line
Shaoqin
Re: [kvm-unit-tests PATCH v2 1/2] arm: pmu: Declare pmu_stats as volatile
In reply to Eric AugerHi,
Quoted 3 lines · Eric Auger
Declare
pmu_statsas volatile in order to prevent the compiler from caching previously read values. This actually fixes pmu-overflow-interrupt failures on some HW.
Looks good to me:
Thanks,
Alex
Quoted 23 lines
Reported-by Alexandru Elisei <[email protected]>Signed-off-by Eric Auger <[email protected]>--- arm/pmu.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)arm/pmu.c +1
-1@@ -328,7 +328,7 @@ asm volatile( : "x9", "x10", "cc"); } -static struct pmu_stats pmu_stats; +static volatile struct pmu_stats pmu_stats;static void irq_handler (struct pt_regs *regs) {Signature · 1 line
2.41.0
PATCH v2 2/2 arm: pmu-overflow-interrupt: Increase count values
In reply to Eric AugerOn some hardware, some pmu-overflow-interrupt failures can be observed. Although the even counter overflows, the interrupt is not seen as expected. This happens in the subtest after "promote to 64-b" comment. After analysis, the PMU overflow interrupt actually hits, ie. kvm_pmu_perf_overflow() gets called and KVM_REQ_IRQ_PENDING is set, as expected. However the PMCR.E is reset by the handle_exit path, at kvm_pmu_handle_pmcr() before the next guest entry and kvm_pmu_flush_hwstate/kvm_pmu_update_state subsequent call. There, since the enable bit has been reset, kvm_pmu_update_state() does not inject the interrupt into the guest.
This does not seem to be a KVM bug but rather an unfortunate scenario where the test disables the PMCR.E too closely to the advent of the overflow interrupt.
Since it looks like a benign and inlikely case, let's resize the number of iterations to prevent the PMCR enable bit from being resetted immediately at the same time as the actual overflow event.
COUNT_INT is introduced, arbitrarily set to 1000 iterations and is used in this test.
An alternative would be to let the PMU enabled and wait for the interrupt but those extra executions might disturb the counting.
---
v1 -> v2:
- Only increase mem_access_loop iterations
--- arm/pmu.c | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-)
arm/pmu.c +6 -5
@@ -66,6 +66,7 @@#define PRE_OVERFLOW_64 0xFFFFFFFFFFFFFFF0ULL#define COUNT 250#define MARGIN 100 +#define COUNT_INT 1000/* * PRE_OVERFLOW2 is set so that 1st @COUNT iterations do not * produce 32b overflow and 2nd @COUNT iterations do. To accommodate
@@ -978,7 +979,7 @@ static void test_overflow_interrupt(bool overflow_at_64bits)/* interrupts are disabled (PMINTENSET_EL1 == 0) */ -mem_access_loop (addr,200 , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp); +mem_access_loop (addr,COUNT_INT , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp);report (expect_interrupts (0 ),"no overflow interrupt after preset" );set_pmcr (pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp);
@@ -1002,7 +1003,7 @@ static void test_overflow_interrupt(bool overflow_at_64bits)write_sysreg (ALL_SET_32 , pmintenset_el1);isb (); -mem_access_loop (addr,200 , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp); +mem_access_loop (addr,COUNT_INT , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp);set_pmcr (pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp);isb ();
@@ -1010,7 +1011,7 @@ static void test_overflow_interrupt(bool overflow_at_64bits)for (i =0 ; i <100 ; i++)write_sysreg (0x3 , pmswinc_el0); -mem_access_loop (addr,200 , pmu.pmcr_ro); +mem_access_loop (addr,COUNT_INT , pmu.pmcr_ro);report_info ("overflow=0x%lx" ,read_sysreg (pmovsclr_el0));report (expect_interrupts (0x3 ),"overflow interrupts expected on #0 and #1" );
@@ -1029,7 +1030,7 @@ static void test_overflow_interrupt(bool overflow_at_64bits)write_regn_el0 (pmevtyper,1 ,CHAIN |PMEVTYPER_EXCLUDE_EL0 );write_regn_el0 (pmevcntr,0 , pre_overflow);isb (); -mem_access_loop (addr,200 , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp); +mem_access_loop (addr,COUNT_INT , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp);report (expect_interrupts (0x1 ),"expect overflow interrupt" );/* overflow on odd counter */
@@ -1037,7 +1038,7 @@ static void test_overflow_interrupt(bool overflow_at_64bits)write_regn_el0 (pmevcntr,0 , pre_overflow);write_regn_el0 (pmevcntr,1 , all_set);isb (); -mem_access_loop (addr,400 , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp); +mem_access_loop (addr,COUNT_INT , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp);if (overflow_at_64bits) {report (expect_interrupts (0x1 ),"expect overflow interrupt on even counter" );
Signature · 1 line
2.41.0
Re: [kvm-unit-tests PATCH v2 2/2] arm: pmu-overflow-interrupt: Increase count values
Quoted 19 lines · Eric Auger
On some hardware, some pmu-overflow-interrupt failures can be observed. Although the even counter overflows, the interrupt is not seen as expected. This happens in the subtest after "promote to 64-b" comment. After analysis, the PMU overflow interrupt actually hits, ie.
kvm_pmu_perf_overflow()gets called andKVM_REQ_IRQ_PENDINGis set, as expected. However the PMCR.E is reset by thehandle_exitpath, atkvm_pmu_handle_pmcr()before the next guest entry and kvm_pmu_flush_hwstate/kvm_pmu_update_state subsequent call. There, since the enable bit has been reset,kvm_pmu_update_state()does not inject the interrupt into the guest.This does not seem to be a KVM bug but rather an unfortunate scenario where the test disables the PMCR.E too closely to the advent of the overflow interrupt.
Since it looks like a benign and inlikely case, let's resize the number of iterations to prevent the PMCR enable bit from being resetted immediately at the same time as the actual overflow event.
Quoted 8 lines
COUNT_INTis introduced, arbitrarily set to 1000 iterations and is used in this test.An alternative would be to let the PMU enabled and wait for the interrupt but those extra executions might disturb the counting.
Reported-by Jan Richter <[email protected]>Signed-off-by Eric Auger <[email protected]>
Quoted 66 lines
---
v1 -> v2:
- Only increasemem_access_loopiterations--- arm/pmu.c | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-)arm/pmu.c +6
-5@@ -66,6 +66,7 @@#define PRE_OVERFLOW_64 0xFFFFFFFFFFFFFFF0ULL#define COUNT 250#define MARGIN 100 +#define COUNT_INT 1000/* * PRE_OVERFLOW2 is set so that 1st @COUNT iterations do not * produce 32b overflow and 2nd @COUNT iterations do. To accommodate @@ -978,7 +979,7 @@ static void test_overflow_interrupt(bool overflow_at_64bits)/* interrupts are disabled (PMINTENSET_EL1 == 0) */ -mem_access_loop (addr,200 , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp); +mem_access_loop (addr,COUNT_INT , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp);report (expect_interrupts (0 ),"no overflow interrupt after preset" );set_pmcr (pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp);@@ -1002,7 +1003,7 @@ static void test_overflow_interrupt(bool overflow_at_64bits)write_sysreg (ALL_SET_32 , pmintenset_el1);isb (); -mem_access_loop (addr,200 , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp); +mem_access_loop (addr,COUNT_INT , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp);set_pmcr (pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp);isb ();@@ -1010,7 +1011,7 @@ static void test_overflow_interrupt(bool overflow_at_64bits)for (i =0 ; i <100 ; i++)write_sysreg (0x3 , pmswinc_el0); -mem_access_loop (addr,200 , pmu.pmcr_ro); +mem_access_loop (addr,COUNT_INT , pmu.pmcr_ro);report_info ("overflow=0x%lx" ,read_sysreg (pmovsclr_el0));report (expect_interrupts (0x3 ),"overflow interrupts expected on #0 and #1" );@@ -1029,7 +1030,7 @@ static void test_overflow_interrupt(bool overflow_at_64bits)write_regn_el0 (pmevtyper,1 ,CHAIN |PMEVTYPER_EXCLUDE_EL0 );write_regn_el0 (pmevcntr,0 , pre_overflow);isb (); -mem_access_loop (addr,200 , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp); +mem_access_loop (addr,COUNT_INT , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp);report (expect_interrupts (0x1 ),"expect overflow interrupt" );/* overflow on odd counter */ @@ -1037,7 +1038,7 @@ static void test_overflow_interrupt(bool overflow_at_64bits)write_regn_el0 (pmevcntr,0 , pre_overflow);write_regn_el0 (pmevcntr,1 , all_set);isb (); -mem_access_loop (addr,400 , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp); +mem_access_loop (addr,COUNT_INT , pmu.pmcr_ro |PMU_PMCR_E | pmcr_lp);if (overflow_at_64bits) {report (expect_interrupts (0x1 ),"expect overflow interrupt on even counter" );
Signature · 1 line
Shaoqin
Quoted 21 lines · Eric Auger
On Qualcomm Amberwing, some pmu-overflow-interrupt failures can be observed. Although the even counter overflows, the interrupt is not seen as expected on guest side. This happens in the subtest after "promote to 64-b" comment.
After analysis, the PMU overflow interrupt actually hits, ie.
kvm_pmu_perf_overflow()gets called andKVM_REQ_IRQ_PENDINGis set, as expected. However the PMCR.E is reset by thehandle_exitpath, atkvm_pmu_handle_pmcr()before the next guest entry and kvm_pmu_flush_hwstate/kvm_pmu_update_state subsequent call. There, since the enable bit has been reset,kvm_pmu_update_state()does not inject the interrupt into the guest.This does not seem to be a KVM bug but rather an unfortunate scenario where the test disables the PMCR.E too closely to the advent of the overflow interrupt.
Since it looks like a benign and inlikely case, let's resize the number of iterations to prevent the PMCR enable bit from being resetted immediately at the same time as the actual overflow event.
Quoted 12 lines
Also make
pmu_statsvolatile to prevent any optimizations.Eric Auger (2):
arm: pmu: Declarepmu_statsas volatile
arm: pmu-overflow-interrupt: Increase count valuesarm/pmu.c | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-)Signature · 1 line
2.41.0
Merged into https://gitlab.com/kvm-unit-tests/kvm-unit-tests master
Thanks,
drew