drivers/perf: Add support for ARMv8.3-SPE
Subsystems (2)
PATCH v3 1/1 drivers/perf: Add support for ARMv8.3-SPE
Armv8.3 extends the SPE by adding:
- Alignment field in the Events packet, and filtering on this event using PMSEVFR_EL1.
- Support for the Scalable Vector Extension (SVE).
The main additions for SVE are:
- Recording the vector length for SVE operations in the Operation Type packet. It is not possible to filter on vector length.
- Incomplete predicate and empty predicate fields in the Events packet, and filtering on these events using PMSEVFR_EL1.
Update the check of pmsevfr for empty/partial predicated SVE and alignment event in SPE driver. For adaption by the version of SPE, expose 'pmsver' as cap attribute to userspace.
---
v2 -> v3:
- Make the definition of 'pmsevfr_res0' progressive and easy to check. (Suggested by Will.)
---
v1 -> v2:
- Rename 'pmuver' to 'pmsver', change it's type to 'u16' from 'int'. (Suggested by Will and Leo.)
- Expose 'pmsver' as cap attribute through sysfs, instead of printing. (Suggested by Will.)
--- arch/arm64/include/asm/sysreg.h | 9 ++++++++- drivers/perf/arm_spe_pmu.c | 21 +++++++++++++++++++-- 2 files changed, 27 insertions(+), 3 deletions(-)
arch/arm64/include/asm/sysreg.h +8 -1
@@ -287,7 +287,11 @@#define SYS_PMSFCR_EL1_ST_SHIFT 18#define SYS_PMSEVFR_EL1 sys_reg(3, 0, 9, 9, 5) -#define SYS_PMSEVFR_EL1_RES0 0x0000ffff00ff0f55UL +#define SYS_PMSEVFR_EL1_RES0_8_2 \ + (GENMASK_ULL(47, 32) | GENMASK_ULL(23, 16) | GENMASK_ULL(11, 8) |\ + BIT_ULL(6) | BIT_ULL(4) | BIT_ULL(2) | BIT_ULL(0)) +#define SYS_PMSEVFR_EL1_RES0_8_3 \ + (SYS_PMSEVFR_EL1_RES0_8_2 & ~(BIT_ULL(18) | BIT_ULL(17) | BIT_ULL(11)))#define SYS_PMSLATFR_EL1 sys_reg(3, 0, 9, 9, 6)#define SYS_PMSLATFR_EL1_MINLAT_SHIFT 0
@@ -829,6 +833,9 @@#define ID_AA64DFR0_PMUVER_8_5 0x6#define ID_AA64DFR0_PMUVER_IMP_DEF 0xf +#define ID_AA64DFR0_PMSVER_8_2 0x1 +#define ID_AA64DFR0_PMSVER_8_3 0x2 +#define ID_DFR0_PERFMON_SHIFT 24#define ID_DFR0_PERFMON_8_1 0x4
drivers/perf/arm_spe_pmu.c +19 -2
@@ -54,7 +54,7 @@ struct arm_spe_pmu {struct hlist_node hotplug_node;int irq;/* PPI */ - +u16 pmsver;u16 min_period;u16 counter_sz;
@@ -93,6 +93,7 @@ enum arm_spe_pmu_capabilities {SPE_PMU_CAP_FEAT_MAX ,SPE_PMU_CAP_CNT_SZ =SPE_PMU_CAP_FEAT_MAX ,SPE_PMU_CAP_MIN_IVAL , + SPE_PMU_CAP_PMSVER, }; static int arm_spe_pmu_feat_caps[SPE_PMU_CAP_FEAT_MAX] = {
@@ -110,6 +111,8 @@ static u32 arm_spe_pmu_cap_get(struct arm_spe_pmu *spe_pmu, int cap)return spe_pmu->counter_sz;case SPE_PMU_CAP_MIN_IVAL :return spe_pmu->min_period; +case SPE_PMU_CAP_PMSVER : +return spe_pmu->pmsver;default :WARN (1 ,"unknown cap %d\n" , cap); }
@@ -143,6 +146,7 @@ static struct attribute *arm_spe_pmu_cap_attr[] = {SPE_CAP_EXT_ATTR_ENTRY (ernd,SPE_PMU_CAP_ERND ),SPE_CAP_EXT_ATTR_ENTRY (count_size,SPE_PMU_CAP_CNT_SZ ),SPE_CAP_EXT_ATTR_ENTRY (min_interval,SPE_PMU_CAP_MIN_IVAL ), +SPE_CAP_EXT_ATTR_ENTRY (pmsver,SPE_PMU_CAP_PMSVER ), NULL, };
@@ -655,6 +659,18 @@ static irqreturn_t arm_spe_pmu_irq_handler(int irq, void *dev)return IRQ_HANDLED ; } +static u64 arm_spe_pmsevfr_res0 (u16 pmsver) +{ +switch (pmsver) { +case ID_AA64DFR0_PMSVER_8_2 : +return SYS_PMSEVFR_EL1_RES0_8_2 ; +case ID_AA64DFR0_PMSVER_8_3 : +return SYS_PMSEVFR_EL1_RES0_8_3 ; +default : +return -1 ; + } +} +/* Perf callbacks */ static int arm_spe_pmu_event_init (struct perf_event *event) {
@@ -670,7 +686,7 @@ static int arm_spe_pmu_event_init(struct perf_event *event) !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))return -ENOENT ; -if (arm_spe_event_to_pmsevfr (event) &SYS_PMSEVFR_EL1_RES0 ) +if (arm_spe_event_to_pmsevfr (event) &arm_spe_pmsevfr_res0 (spe_pmu->pmsver))return -EOPNOTSUPP ;if (attr->exclude_idle)
@@ -937,6 +953,7 @@ static void __arm_spe_pmu_dev_probe(void *info) fld, smp_processor_id());return ; } + spe_pmu->pmsver = (u16 )fld;/* Read PMBIDR first to determine whether or not we have access */ reg =read_sysreg_s (SYS_PMBIDR_EL1 );
Signature · 1 line
2.17.1
Quoted 11 lines · Wei Li
Armv8.3 extends the SPE by adding:
- Alignment field in the Events packet, and filtering on this event usingPMSEVFR_EL1.
- Support for the Scalable Vector Extension (SVE).The main additions for SVE are:
- Recording the vector length for SVE operations in the Operation Type packet. It is not possible to filter on vector length.
- Incomplete predicate and empty predicate fields in the Events packet, and filtering on these events usingPMSEVFR_EL1.
Quoted 5 lines
Update the check of pmsevfr for empty/partial predicated SVE and alignment event in SPE driver. For adaption by the version of SPE, expose 'pmsver' as cap attribute to userspace.
Signed-off-by Wei Li <[email protected]>
nit: Do we need to update the Kconfig help text too ? Right now that says :
--
tristate "Enable support for the ARMv8.2 Statistical Profiling Extension""
....
help
Enable perf support for the ARMv8.2 Statistical Profiling Extension, which provides periodic sampling of operations in
--
Suzuki
Will Deacon
maintainerQuoted 23 lines · Suzuki K Poulose
Quoted 11 lines · Wei Li
Armv8.3 extends the SPE by adding:
- Alignment field in the Events packet, and filtering on this event usingPMSEVFR_EL1.
- Support for the Scalable Vector Extension (SVE).The main additions for SVE are:
- Recording the vector length for SVE operations in the Operation Type packet. It is not possible to filter on vector length.
- Incomplete predicate and empty predicate fields in the Events packet, and filtering on these events usingPMSEVFR_EL1.Quoted 5 lines
Update the check of pmsevfr for empty/partial predicated SVE and alignment event in SPE driver. For adaption by the version of SPE, expose 'pmsver' as cap attribute to userspace.
Signed-off-by Wei Li <[email protected]>nit: Do we need to update the
Kconfighelp text too ? Right now that says :--
Quoted 6 lines
tristate "Enable support for the ARMv8.2 Statistical Profiling Extension""
....help
Enable perf support for the ARMv8.2 Statistical Profiling Extension, which provides periodic sampling of operations in
I think it's ok as-is, to be honest. It identifies the version of the architecture when the feature was introduced, and I think it's only reasonable to assume that subsequent versions of the architecture might add extensions.
Will
Will Deacon
maintainerQuoted 100 lines · Wei Li
Armv8.3 extends the SPE by adding:
- Alignment field in the Events packet, and filtering on this event usingPMSEVFR_EL1.
- Support for the Scalable Vector Extension (SVE).The main additions for SVE are:
- Recording the vector length for SVE operations in the Operation Type packet. It is not possible to filter on vector length.
- Incomplete predicate and empty predicate fields in the Events packet, and filtering on these events usingPMSEVFR_EL1.Update the check of pmsevfr for empty/partial predicated SVE and alignment event in SPE driver. For adaption by the version of SPE, expose 'pmsver' as cap attribute to userspace.
Signed-off-by Wei Li <[email protected]>---
v2 -> v3:
- Make the definition of 'pmsevfr_res0' progressive and easy to check. (Suggested by Will.)
---
v1 -> v2:
- Rename 'pmuver' to 'pmsver', change it's type to 'u16' from 'int'. (Suggested by Will and Leo.)
- Expose 'pmsver' as cap attribute through sysfs, instead of printing. (Suggested by Will.)--- arch/arm64/include/asm/sysreg.h | 9 ++++++++- drivers/perf/arm_spe_pmu.c | 21 +++++++++++++++++++-- 2 files changed, 27 insertions(+), 3 deletions(-)arch/arm64/include/asm/sysreg.h +8
-1@@ -287,7 +287,11 @@#define SYS_PMSFCR_EL1_ST_SHIFT 18#define SYS_PMSEVFR_EL1 sys_reg(3, 0, 9, 9, 5) -#define SYS_PMSEVFR_EL1_RES0 0x0000ffff00ff0f55UL +#define SYS_PMSEVFR_EL1_RES0_8_2 \ + (GENMASK_ULL(47, 32) | GENMASK_ULL(23, 16) | GENMASK_ULL(11, 8) |\ + BIT_ULL(6) | BIT_ULL(4) | BIT_ULL(2) | BIT_ULL(0)) +#define SYS_PMSEVFR_EL1_RES0_8_3 \ + (SYS_PMSEVFR_EL1_RES0_8_2 & ~(BIT_ULL(18) | BIT_ULL(17) | BIT_ULL(11)))#define SYS_PMSLATFR_EL1 sys_reg(3, 0, 9, 9, 6)#define SYS_PMSLATFR_EL1_MINLAT_SHIFT 0@@ -829,6 +833,9 @@#define ID_AA64DFR0_PMUVER_8_5 0x6#define ID_AA64DFR0_PMUVER_IMP_DEF 0xf +#define ID_AA64DFR0_PMSVER_8_2 0x1 +#define ID_AA64DFR0_PMSVER_8_3 0x2 +#define ID_DFR0_PERFMON_SHIFT 24#define ID_DFR0_PERFMON_8_1 0x4drivers/perf/arm_spe_pmu.c +5
-1@@ -54,7 +54,7 @@ struct arm_spe_pmu {struct hlist_node hotplug_node;int irq;/* PPI */ - +u16 pmsver;u16 min_period;u16 counter_sz;@@ -93,6 +93,7 @@ enum arm_spe_pmu_capabilities {SPE_PMU_CAP_FEAT_MAX ,SPE_PMU_CAP_CNT_SZ =SPE_PMU_CAP_FEAT_MAX ,SPE_PMU_CAP_MIN_IVAL , + SPE_PMU_CAP_PMSVER, }; static int arm_spe_pmu_feat_caps[SPE_PMU_CAP_FEAT_MAX] = {@@ -110,6 +111,8 @@ static u32 arm_spe_pmu_cap_get(struct arm_spe_pmu *spe_pmu, int cap)return spe_pmu->counter_sz;case SPE_PMU_CAP_MIN_IVAL :return spe_pmu->min_period; +case SPE_PMU_CAP_PMSVER : +return spe_pmu->pmsver;default :WARN (1 ,"unknown cap %d\n" , cap); }@@ -143,6 +146,7 @@ static struct attribute *arm_spe_pmu_cap_attr[] = {SPE_CAP_EXT_ATTR_ENTRY (ernd,SPE_PMU_CAP_ERND ),SPE_CAP_EXT_ATTR_ENTRY (count_size,SPE_PMU_CAP_CNT_SZ ),SPE_CAP_EXT_ATTR_ENTRY (min_interval,SPE_PMU_CAP_MIN_IVAL ), +SPE_CAP_EXT_ATTR_ENTRY (pmsver,SPE_PMU_CAP_PMSVER ), NULL, };@@ -655,6 +659,18 @@ static irqreturn_t arm_spe_pmu_irq_handler(int irq, void *dev)return IRQ_HANDLED ; }
drivers/perf/arm_spe_pmu.c static irqreturn_t arm_spe_pmu_irq_handler(int irq, void *dev) · line 671
Quoted 10 lines
drivers/perf/arm_spe_pmu.c +10
-0+static u64 arm_spe_pmsevfr_res0 (u16 pmsver) +{ +switch (pmsver) { +case ID_AA64DFR0_PMSVER_8_2 : +return SYS_PMSEVFR_EL1_RES0_8_2 ; +case ID_AA64DFR0_PMSVER_8_3 : +return SYS_PMSEVFR_EL1_RES0_8_3 ; +default : +return -1 ; + }
I don't think -1 is the right choice here if we don't recognise the version, as it means that old kernels won't work on new hardware. I think it would be better to return the highest architecture version we know about in that case...
Quoted 15 lines
drivers/perf/arm_spe_pmu.c +3
-1+} +/* Perf callbacks */ static int arm_spe_pmu_event_init (struct perf_event *event) {@@ -670,7 +686,7 @@ static int arm_spe_pmu_event_init(struct perf_event *event) !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))return -ENOENT ; -if (arm_spe_event_to_pmsevfr (event) &SYS_PMSEVFR_EL1_RES0 ) +if (arm_spe_event_to_pmsevfr (event) &arm_spe_pmsevfr_res0 (spe_pmu->pmsver))return -EOPNOTSUPP ;if (attr->exclude_idle)
drivers/perf/arm_spe_pmu.c static void __arm_spe_pmu_dev_probe(void *info) · line 956
Quoted 4 lines
drivers/perf/arm_spe_pmu.c +1
-0@@ -937,6 +953,7 @@ static void __arm_spe_pmu_dev_probe(void *info) fld, smp_processor_id());return ; } + spe_pmu->pmsver = (u16 )fld;
... which also means we should clamp this value now that we expose it to userspace. Otherwise, userspace can't rely on this field for anything.
That said -- please can you tell me what userspace intends to do with this version number?
Will
Hi Will,
Quoted 101 lines · Will Deacon
Quoted 100 lines · Wei Li
Armv8.3 extends the SPE by adding:
- Alignment field in the Events packet, and filtering on this event usingPMSEVFR_EL1.
- Support for the Scalable Vector Extension (SVE).The main additions for SVE are:
- Recording the vector length for SVE operations in the Operation Type packet. It is not possible to filter on vector length.
- Incomplete predicate and empty predicate fields in the Events packet, and filtering on these events usingPMSEVFR_EL1.Update the check of pmsevfr for empty/partial predicated SVE and alignment event in SPE driver. For adaption by the version of SPE, expose 'pmsver' as cap attribute to userspace.
Signed-off-by Wei Li <[email protected]>---
v2 -> v3:
- Make the definition of 'pmsevfr_res0' progressive and easy to check. (Suggested by Will.)
---
v1 -> v2:
- Rename 'pmuver' to 'pmsver', change it's type to 'u16' from 'int'. (Suggested by Will and Leo.)
- Expose 'pmsver' as cap attribute through sysfs, instead of printing. (Suggested by Will.)--- arch/arm64/include/asm/sysreg.h | 9 ++++++++- drivers/perf/arm_spe_pmu.c | 21 +++++++++++++++++++-- 2 files changed, 27 insertions(+), 3 deletions(-)arch/arm64/include/asm/sysreg.h +8
-1@@ -287,7 +287,11 @@#define SYS_PMSFCR_EL1_ST_SHIFT 18#define SYS_PMSEVFR_EL1 sys_reg(3, 0, 9, 9, 5) -#define SYS_PMSEVFR_EL1_RES0 0x0000ffff00ff0f55UL +#define SYS_PMSEVFR_EL1_RES0_8_2 \ + (GENMASK_ULL(47, 32) | GENMASK_ULL(23, 16) | GENMASK_ULL(11, 8) |\ + BIT_ULL(6) | BIT_ULL(4) | BIT_ULL(2) | BIT_ULL(0)) +#define SYS_PMSEVFR_EL1_RES0_8_3 \ + (SYS_PMSEVFR_EL1_RES0_8_2 & ~(BIT_ULL(18) | BIT_ULL(17) | BIT_ULL(11)))#define SYS_PMSLATFR_EL1 sys_reg(3, 0, 9, 9, 6)#define SYS_PMSLATFR_EL1_MINLAT_SHIFT 0@@ -829,6 +833,9 @@#define ID_AA64DFR0_PMUVER_8_5 0x6#define ID_AA64DFR0_PMUVER_IMP_DEF 0xf +#define ID_AA64DFR0_PMSVER_8_2 0x1 +#define ID_AA64DFR0_PMSVER_8_3 0x2 +#define ID_DFR0_PERFMON_SHIFT 24#define ID_DFR0_PERFMON_8_1 0x4drivers/perf/arm_spe_pmu.c +5
-1@@ -54,7 +54,7 @@ struct arm_spe_pmu {struct hlist_node hotplug_node;int irq;/* PPI */ - +u16 pmsver;u16 min_period;u16 counter_sz;@@ -93,6 +93,7 @@ enum arm_spe_pmu_capabilities {SPE_PMU_CAP_FEAT_MAX ,SPE_PMU_CAP_CNT_SZ =SPE_PMU_CAP_FEAT_MAX ,SPE_PMU_CAP_MIN_IVAL , + SPE_PMU_CAP_PMSVER, }; static int arm_spe_pmu_feat_caps[SPE_PMU_CAP_FEAT_MAX] = {@@ -110,6 +111,8 @@ static u32 arm_spe_pmu_cap_get(struct arm_spe_pmu *spe_pmu, int cap)return spe_pmu->counter_sz;case SPE_PMU_CAP_MIN_IVAL :return spe_pmu->min_period; +case SPE_PMU_CAP_PMSVER : +return spe_pmu->pmsver;default :WARN (1 ,"unknown cap %d\n" , cap); }@@ -143,6 +146,7 @@ static struct attribute *arm_spe_pmu_cap_attr[] = {SPE_CAP_EXT_ATTR_ENTRY (ernd,SPE_PMU_CAP_ERND ),SPE_CAP_EXT_ATTR_ENTRY (count_size,SPE_PMU_CAP_CNT_SZ ),SPE_CAP_EXT_ATTR_ENTRY (min_interval,SPE_PMU_CAP_MIN_IVAL ), +SPE_CAP_EXT_ATTR_ENTRY (pmsver,SPE_PMU_CAP_PMSVER ), NULL, };@@ -655,6 +659,18 @@ static irqreturn_t arm_spe_pmu_irq_handler(int irq, void *dev)return IRQ_HANDLED ; }
Quoted 15 lines
drivers/perf/arm_spe_pmu.c static irqreturn_t arm_spe_pmu_irq_handler(int irq, void *dev) · line 671
Quoted 10 lines
drivers/perf/arm_spe_pmu.c +10
-0+static u64 arm_spe_pmsevfr_res0 (u16 pmsver) +{ +switch (pmsver) { +case ID_AA64DFR0_PMSVER_8_2 : +return SYS_PMSEVFR_EL1_RES0_8_2 ; +case ID_AA64DFR0_PMSVER_8_3 : +return SYS_PMSEVFR_EL1_RES0_8_3 ; +default : +return -1 ; + }I don't think -1 is the right choice here if we don't recognise the version, as it means that old kernels won't work on new hardware. I think it would be better to return the highest architecture version we know about in that case...
ACK.
Quoted 23 lines
Quoted 15 lines
drivers/perf/arm_spe_pmu.c +3
-1+} +/* Perf callbacks */ static int arm_spe_pmu_event_init (struct perf_event *event) {@@ -670,7 +686,7 @@ static int arm_spe_pmu_event_init(struct perf_event *event) !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))return -ENOENT ; -if (arm_spe_event_to_pmsevfr (event) &SYS_PMSEVFR_EL1_RES0 ) +if (arm_spe_event_to_pmsevfr (event) &arm_spe_pmsevfr_res0 (spe_pmu->pmsver))return -EOPNOTSUPP ;if (attr->exclude_idle)drivers/perf/arm_spe_pmu.c static void __arm_spe_pmu_dev_probe(void *info) · line 956
Quoted 4 lines
drivers/perf/arm_spe_pmu.c +1
-0@@ -937,6 +953,7 @@ static void __arm_spe_pmu_dev_probe(void *info) fld, smp_processor_id());return ; } + spe_pmu->pmsver = (u16 )fld;... which also means we should clamp this value now that we expose it to userspace. Otherwise, userspace can't rely on this field for anything.
Quoted 3 lines
That said -- please can you tell me what userspace intends to do with this version number?
In fact, it is only used in our testcase for now, which needs to know the real version the chip has implemented, as these is no other way to get the info now. So i added it in the probe message in v1 at first.
Especially we use the of_device_id "arm,statistical-profiling-extension-v1" and the platform_device_id "arm,spe-v1". It's a little weird to ARMv8.3-SPE.
Further more i am wondering if we need to add the 'spe-v2' device_id. If not, i think it's broken on a big.LITTLE system [1] as we can register only one device and the driver will work incorrectly if the big and LITTLE cores have different version of SPE.
If add, we may really support two instance. Even two different versions of SPE pmus which work on different range of cores with the same PPI. Their functional scopes are the same as the PPI partitions. Exporting the version is not needed as we can get it from the device_id.
Thanks,
Wei
[1] https://www.spinics.net/lists/arm-kernel/msg827183.html
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Will Deacon
maintainerHi,
Quoted 23 lines · liwei (GF)
Quoted 22 lines · Will Deacon
Quoted 13 lines · Wei Li
/* Perf callbacks */
static intarm_spe_pmu_event_init(struct perf_event *event){drivers/perf/arm_spe_pmu.c +1
-1@@ -670,7 +686,7 @@ static int arm_spe_pmu_event_init(struct perf_event *event) !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))return -ENOENT ; -if (arm_spe_event_to_pmsevfr (event) &SYS_PMSEVFR_EL1_RES0 ) +if (arm_spe_event_to_pmsevfr (event) &arm_spe_pmsevfr_res0 (spe_pmu->pmsver))return -EOPNOTSUPP ;if (attr->exclude_idle)drivers/perf/arm_spe_pmu.c static void __arm_spe_pmu_dev_probe(void *info) · line 956
Quoted 4 lines
drivers/perf/arm_spe_pmu.c +1
-0@@ -937,6 +953,7 @@ static void __arm_spe_pmu_dev_probe(void *info) fld, smp_processor_id());return ; } + spe_pmu->pmsver = (u16 )fld;... which also means we should clamp this value now that we expose it to userspace. Otherwise, userspace can't rely on this field for anything.
Quoted 7 lines
Quoted 3 lines
That said -- please can you tell me what userspace intends to do with this version number?
In fact, it is only used in our testcase for now, which needs to know the real version the chip has implemented, as these is no other way to get the info now. So i added it in the probe message in v1 at first.
If it's not needed, then let's not expose it yet. ABIs aren't cheap!
Quoted 3 lines
Especially we use the
of_device_id"arm,statistical-profiling-extension-v1" and theplatform_device_id"arm,spe-v1". It's a little weird to ARMv8.3-SPE.
What exactly breaks? It looks like it should work fine to me, but I don't have any SPE systems to test on. The "v1" in the compatible string doesn't really do anything at the moment because we can probe everything other than the interrupt topology at runtime, so I wouldn't get hung up on that.
Quoted 4 lines
Further more i am wondering if we need to add the 'spe-v2'
device_id. If not, i think it's broken on a big.LITTLE system [1] as we can register only one device and the driver will work incorrectly if the big and LITTLE cores have different version of SPE.
This is only an issue for ACPI, right? I don't think that was ever intended to work with big.LITTLE. See:
pr_warn("ACPI: SPE must be homogeneous\n");DT should handle it though.
Anyway, please can you post a new version of this patch without the ABI changes? In other words, just including the RES0 part.
Thanks,
Will