Re: [RESEND PATCH V4] perf/x86/intel: Support RDPMC metrics clear mode
From: Liang, Kan
Date: Wed Oct 09 2024 - 15:31:48 EST
Hi Peter,
Gentle ping. Please let me know if you have any comments.
Thanks,
Kan
On 2024-09-26 2:45 p.m., kan.liang@xxxxxxxxxxxxxxx wrote:
> From: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
>
> The new RDPMC enhancement, metrics clear mode, is to clear the
> PERF_METRICS-related resources as well as the fixed-function performance
> monitoring counter 3 after the read is performed. It is available for
> ring 3. The feature is enumerated by the
> IA32_PERF_CAPABILITIES.RDPMC_CLEAR_METRICS[bit 19]. To enable the
> feature, the IA32_FIXED_CTR_CTRL.METRICS_CLEAR_EN[bit 14] must be set.
>
> Two ways were considered to enable the feature.
> - Expose a knob in the sysfs globally. One user may affect the
> measurement of other users when changing the knob. The solution is
> dropped.
> - Introduce a new event format, metrics_clear, for the slots event to
> disable/enable the feature only for the current process. Users can
> utilize the feature as needed.
> The latter solution is implemented in the patch.
>
> The current KVM doesn't support the perf metrics yet. For
> virtualization, the feature can be enabled later separately.
>
> Update the document of perf metrics.
>
> Suggested-by: Andi Kleen <ak@xxxxxxxxxxxxxxx>
> Reviewed-by: Andi Kleen <ak@xxxxxxxxxxxxxxx>
> Reviewed-by: Ian Rogers <irogers@xxxxxxxxxx>
> Signed-off-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
> ---
>
> The original V4 can be found at
> https://lore.kernel.org/lkml/20240731143835.771618-6-kan.liang@xxxxxxxxxxxxxxx/
>
> The patch was one of the PMU features in the LNL enabling patch set.
> The other feature is now blocked. Send the patch separately.
>
> arch/x86/events/intel/core.c | 20 +++++++++++++++++++-
> arch/x86/events/perf_event.h | 1 +
> arch/x86/include/asm/perf_event.h | 4 ++++
> tools/perf/Documentation/topdown.txt | 9 +++++++--
> 4 files changed, 31 insertions(+), 3 deletions(-)
>
> diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
> index 5182075e111b..342f8b1a2f93 100644
> --- a/arch/x86/events/intel/core.c
> +++ b/arch/x86/events/intel/core.c
> @@ -2816,6 +2816,9 @@ static void intel_pmu_enable_fixed(struct perf_event *event)
> return;
>
> idx = INTEL_PMC_IDX_FIXED_SLOTS;
> +
> + if (event->attr.config1 & INTEL_TD_CFG_METRIC_CLEAR)
> + bits |= INTEL_FIXED_3_METRICS_CLEAR;
> }
>
> intel_set_masks(event, idx);
> @@ -4067,7 +4070,12 @@ static int intel_pmu_hw_config(struct perf_event *event)
> * is used in a metrics group, it too cannot support sampling.
> */
> if (intel_pmu_has_cap(event, PERF_CAP_METRICS_IDX) && is_topdown_event(event)) {
> - if (event->attr.config1 || event->attr.config2)
> + /* The metrics_clear can only be set for the slots event */
> + if (event->attr.config1 &&
> + (!is_slots_event(event) || (event->attr.config1 & ~INTEL_TD_CFG_METRIC_CLEAR)))
> + return -EINVAL;
> +
> + if (event->attr.config2)
> return -EINVAL;
>
> /*
> @@ -4676,6 +4684,8 @@ PMU_FORMAT_ATTR(in_tx, "config:32" );
> PMU_FORMAT_ATTR(in_tx_cp, "config:33" );
> PMU_FORMAT_ATTR(eq, "config:36" ); /* v6 + */
>
> +PMU_FORMAT_ATTR(metrics_clear, "config1:0"); /* PERF_CAPABILITIES.RDPMC_METRICS_CLEAR */
> +
> static ssize_t umask2_show(struct device *dev,
> struct device_attribute *attr,
> char *page)
> @@ -4695,6 +4705,7 @@ static struct device_attribute format_attr_umask2 =
> static struct attribute *format_evtsel_ext_attrs[] = {
> &format_attr_umask2.attr,
> &format_attr_eq.attr,
> + &format_attr_metrics_clear.attr,
> NULL
> };
>
> @@ -4719,6 +4730,13 @@ evtsel_ext_is_visible(struct kobject *kobj, struct attribute *attr, int i)
> if (i == 1)
> return (mask & ARCH_PERFMON_EVENTSEL_EQ) ? attr->mode : 0;
>
> + /* PERF_CAPABILITIES.RDPMC_METRICS_CLEAR */
> + if (i == 2) {
> + union perf_capabilities intel_cap = hybrid(dev_get_drvdata(dev), intel_cap);
> +
> + return intel_cap.rdpmc_metrics_clear ? attr->mode : 0;
> + }
> +
> return 0;
> }
>
> diff --git a/arch/x86/events/perf_event.h b/arch/x86/events/perf_event.h
> index 82c6f45ce975..31c2771545a6 100644
> --- a/arch/x86/events/perf_event.h
> +++ b/arch/x86/events/perf_event.h
> @@ -624,6 +624,7 @@ union perf_capabilities {
> u64 pebs_output_pt_available:1;
> u64 pebs_timing_info:1;
> u64 anythread_deprecated:1;
> + u64 rdpmc_metrics_clear:1;
> };
> u64 capabilities;
> };
> diff --git a/arch/x86/include/asm/perf_event.h b/arch/x86/include/asm/perf_event.h
> index 91b73571412f..e3b5e8e96fb3 100644
> --- a/arch/x86/include/asm/perf_event.h
> +++ b/arch/x86/include/asm/perf_event.h
> @@ -41,6 +41,7 @@
> #define INTEL_FIXED_0_USER (1ULL << 1)
> #define INTEL_FIXED_0_ANYTHREAD (1ULL << 2)
> #define INTEL_FIXED_0_ENABLE_PMI (1ULL << 3)
> +#define INTEL_FIXED_3_METRICS_CLEAR (1ULL << 2)
>
> #define HSW_IN_TX (1ULL << 32)
> #define HSW_IN_TX_CHECKPOINTED (1ULL << 33)
> @@ -372,6 +373,9 @@ static inline bool use_fixed_pseudo_encoding(u64 code)
> #define INTEL_TD_METRIC_MAX INTEL_TD_METRIC_MEM_BOUND
> #define INTEL_TD_METRIC_NUM 8
>
> +#define INTEL_TD_CFG_METRIC_CLEAR_BIT 0
> +#define INTEL_TD_CFG_METRIC_CLEAR BIT_ULL(INTEL_TD_CFG_METRIC_CLEAR_BIT)
> +
> static inline bool is_metric_idx(int idx)
> {
> return (unsigned)(idx - INTEL_PMC_IDX_METRIC_BASE) < INTEL_TD_METRIC_NUM;
> diff --git a/tools/perf/Documentation/topdown.txt b/tools/perf/Documentation/topdown.txt
> index ae0aee86844f..f36c8ca1dc53 100644
> --- a/tools/perf/Documentation/topdown.txt
> +++ b/tools/perf/Documentation/topdown.txt
> @@ -280,8 +280,13 @@ with no longer interval than a few seconds
>
> perf stat -I 1000 --topdown ...
>
> -For user programs using RDPMC directly the counter can
> -be reset explicitly using ioctl:
> +Starting from the Lunar Lake p-core, a RDPMC metrics clear mode is
> +introduced. The metrics and the fixed counter 3 are automatically
> +cleared after the read is performed. It is recommended to always enable
> +the mode. To enable the mode, the config1 of slots event is set to 1.
> +
> +On the previous platforms, for user programs using RDPMC directly, the
> +counter has to be reset explicitly using ioctl:
>
> ioctl(perf_fd, PERF_EVENT_IOC_RESET, 0);
>