diff mbox series

[v3,8/8] perf: arm_spe: Add support for SPEv1.2 inverted event filtering

Message ID 20220825-arm-spe-v8-7-v3-8-87682f78caac@kernel.org (mailing list archive)
State New, archived
Headers show
Series perf: Arm SPEv1.2 support | expand

Commit Message

Rob Herring (Arm) Nov. 4, 2022, 3:55 p.m. UTC 
Arm SPEv1.2 (Arm v8.7/v9.2) adds a new feature called Inverted Event
Filter which excludes samples matching the event filter. The feature
mirrors the existing event filter in PMSEVFR_EL1 adding a new register,
PMSNEVFR_EL1, which has the same event bit assignments.

Tested-by: James Clark <james.clark@arm.com>
Signed-off-by: Rob Herring <robh@kernel.org>
---
v3:
 - No change
v2:
 - Update for auto generated register defines
 - Avoid accessing SYS_PMSNEVFR_EL1 on < v8.7
---
 drivers/perf/arm_spe_pmu.c | 45 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 45 insertions(+)

Comments

Namhyung Kim Nov. 7, 2022, 7:48 p.m. UTC | #1 
Hello,

On Fri, Nov 4, 2022 at 8:55 AM Rob Herring <robh@kernel.org> wrote:
>
> Arm SPEv1.2 (Arm v8.7/v9.2) adds a new feature called Inverted Event
> Filter which excludes samples matching the event filter. The feature
> mirrors the existing event filter in PMSEVFR_EL1 adding a new register,
> PMSNEVFR_EL1, which has the same event bit assignments.

Just out of curiosity, is it possible to have the event filter and
inverted filter at the same time?  What if they have the same?

Thanks,
Namhyung

>
> Tested-by: James Clark <james.clark@arm.com>
> Signed-off-by: Rob Herring <robh@kernel.org>
> ---
> v3:
>  - No change
> v2:
>  - Update for auto generated register defines
>  - Avoid accessing SYS_PMSNEVFR_EL1 on < v8.7
> ---
Rob Herring (Arm) Nov. 8, 2022, 1:12 p.m. UTC | #2 
On Mon, Nov 7, 2022 at 1:48 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> Hello,
>
> On Fri, Nov 4, 2022 at 8:55 AM Rob Herring <robh@kernel.org> wrote:
> >
> > Arm SPEv1.2 (Arm v8.7/v9.2) adds a new feature called Inverted Event
> > Filter which excludes samples matching the event filter. The feature
> > mirrors the existing event filter in PMSEVFR_EL1 adding a new register,
> > PMSNEVFR_EL1, which has the same event bit assignments.
>
> Just out of curiosity, is it possible to have the event filter and
> inverted filter at the same time?

Yes, that's why config3 is needed. Samples with all the events set in
PMSEVFR_EL1 and none of the events set in PMSNEVFR_EL1 are recorded.
For example if an event bit is 0 in PMSEVFR_EL1, then recorded samples
may or may not have that event. If the same event bit is then set in
PMSNEVFR_EL1, then record samples will not have that event.

> What if they have the same?

You'd get no samples.

Rob
diff mbox series

Patch

diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
index 82f67e941bc4..573db4211acd 100644
--- a/drivers/perf/arm_spe_pmu.c
+++ b/drivers/perf/arm_spe_pmu.c
@@ -85,6 +85,7 @@  struct arm_spe_pmu {
 #define SPE_PMU_FEAT_ARCH_INST			(1UL << 3)
 #define SPE_PMU_FEAT_LDS			(1UL << 4)
 #define SPE_PMU_FEAT_ERND			(1UL << 5)
+#define SPE_PMU_FEAT_INV_FILT_EVT		(1UL << 6)
 #define SPE_PMU_FEAT_DEV_PROBED			(1UL << 63)
 	u64					features;
 
@@ -202,6 +203,10 @@  static const struct attribute_group arm_spe_pmu_cap_group = {
 #define ATTR_CFG_FLD_min_latency_LO		0
 #define ATTR_CFG_FLD_min_latency_HI		11
 
+#define ATTR_CFG_FLD_inv_event_filter_CFG	config3	/* PMSNEVFR_EL1 */
+#define ATTR_CFG_FLD_inv_event_filter_LO	0
+#define ATTR_CFG_FLD_inv_event_filter_HI	63
+
 /* Why does everything I do descend into this? */
 #define __GEN_PMU_FORMAT_ATTR(cfg, lo, hi)				\
 	(lo) == (hi) ? #cfg ":" #lo "\n" : #cfg ":" #lo "-" #hi
@@ -232,6 +237,7 @@  GEN_PMU_FORMAT_ATTR(branch_filter);
 GEN_PMU_FORMAT_ATTR(load_filter);
 GEN_PMU_FORMAT_ATTR(store_filter);
 GEN_PMU_FORMAT_ATTR(event_filter);
+GEN_PMU_FORMAT_ATTR(inv_event_filter);
 GEN_PMU_FORMAT_ATTR(min_latency);
 
 static struct attribute *arm_spe_pmu_formats_attr[] = {
@@ -243,12 +249,27 @@  static struct attribute *arm_spe_pmu_formats_attr[] = {
 	&format_attr_load_filter.attr,
 	&format_attr_store_filter.attr,
 	&format_attr_event_filter.attr,
+	&format_attr_inv_event_filter.attr,
 	&format_attr_min_latency.attr,
 	NULL,
 };
 
+static umode_t arm_spe_pmu_format_attr_is_visible(struct kobject *kobj,
+						  struct attribute *attr,
+						  int unused)
+	{
+	struct device *dev = kobj_to_dev(kobj);
+	struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
+
+	if (attr == &format_attr_inv_event_filter.attr && !(spe_pmu->features & SPE_PMU_FEAT_INV_FILT_EVT))
+		return 0;
+
+	return attr->mode;
+}
+
 static const struct attribute_group arm_spe_pmu_format_group = {
 	.name	= "format",
+	.is_visible = arm_spe_pmu_format_attr_is_visible,
 	.attrs	= arm_spe_pmu_formats_attr,
 };
 
@@ -343,6 +364,9 @@  static u64 arm_spe_event_to_pmsfcr(struct perf_event *event)
 	if (ATTR_CFG_GET_FLD(attr, event_filter))
 		reg |= PMSFCR_EL1_FE;
 
+	if (ATTR_CFG_GET_FLD(attr, inv_event_filter))
+		reg |= PMSFCR_EL1_FnE;
+
 	if (ATTR_CFG_GET_FLD(attr, min_latency))
 		reg |= PMSFCR_EL1_FL;
 
@@ -355,6 +379,12 @@  static u64 arm_spe_event_to_pmsevfr(struct perf_event *event)
 	return ATTR_CFG_GET_FLD(attr, event_filter);
 }
 
+static u64 arm_spe_event_to_pmsnevfr(struct perf_event *event)
+{
+	struct perf_event_attr *attr = &event->attr;
+	return ATTR_CFG_GET_FLD(attr, inv_event_filter);
+}
+
 static u64 arm_spe_event_to_pmslatfr(struct perf_event *event)
 {
 	struct perf_event_attr *attr = &event->attr;
@@ -703,6 +733,9 @@  static int arm_spe_pmu_event_init(struct perf_event *event)
 	if (arm_spe_event_to_pmsevfr(event) & arm_spe_pmsevfr_res0(spe_pmu->pmsver))
 		return -EOPNOTSUPP;
 
+	if (arm_spe_event_to_pmsnevfr(event) & arm_spe_pmsevfr_res0(spe_pmu->pmsver))
+		return -EOPNOTSUPP;
+
 	if (attr->exclude_idle)
 		return -EOPNOTSUPP;
 
@@ -721,6 +754,10 @@  static int arm_spe_pmu_event_init(struct perf_event *event)
 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_EVT))
 		return -EOPNOTSUPP;
 
+	if ((FIELD_GET(PMSFCR_EL1_FnE, reg)) &&
+	    !(spe_pmu->features & SPE_PMU_FEAT_INV_FILT_EVT))
+		return -EOPNOTSUPP;
+
 	if ((FIELD_GET(PMSFCR_EL1_FT, reg)) &&
 	    !(spe_pmu->features & SPE_PMU_FEAT_FILT_TYP))
 		return -EOPNOTSUPP;
@@ -756,6 +793,11 @@  static void arm_spe_pmu_start(struct perf_event *event, int flags)
 	reg = arm_spe_event_to_pmsevfr(event);
 	write_sysreg_s(reg, SYS_PMSEVFR_EL1);
 
+	if (spe_pmu->features & SPE_PMU_FEAT_INV_FILT_EVT) {
+		reg = arm_spe_event_to_pmsnevfr(event);
+		write_sysreg_s(reg, SYS_PMSNEVFR_EL1);
+	}
+
 	reg = arm_spe_event_to_pmslatfr(event);
 	write_sysreg_s(reg, SYS_PMSLATFR_EL1);
 
@@ -990,6 +1032,9 @@  static void __arm_spe_pmu_dev_probe(void *info)
 	if (FIELD_GET(PMSIDR_EL1_FE, reg))
 		spe_pmu->features |= SPE_PMU_FEAT_FILT_EVT;
 
+	if (FIELD_GET(PMSIDR_EL1_FnE, reg))
+		spe_pmu->features |= SPE_PMU_FEAT_INV_FILT_EVT;
+
 	if (FIELD_GET(PMSIDR_EL1_FT, reg))
 		spe_pmu->features |= SPE_PMU_FEAT_FILT_TYP;