[PATCH 3/3] drm/xe/pmu: Add GT frequency events

Riana Tauro riana.tauro at intel.com
Mon Sep 30 09:42:13 UTC 2024


Hi Vinay

On 9/27/2024 5:53 AM, Vinay Belgaumkar wrote:
> Define PMU events for GT frequency (actual and requested). This is
> a port from the i915 driver implementation.
> 
> Following PMU events are being added-
> 
>    xe_0000_00_02.0/actual-frequency-gt0/              [Kernel PMU event]
>    xe_0000_00_02.0/actual-frequency-gt1/              [Kernel PMU event]
>    xe_0000_00_02.0/requested-frequency-gt0/           [Kernel PMU event]
>    xe_0000_00_02.0/requested-frequency-gt1/           [Kernel PMU event]
> 
> v2: Checkpatch fix, moved timer code to this patch
> v3: Fix kunit issue
> 
> Cc: Rodrigo Vivi <rodrigo.vivi at intel.com>
> Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar at intel.com>
> ---
>   drivers/gpu/drm/xe/xe_gt.c        |   2 +
>   drivers/gpu/drm/xe/xe_pmu.c       | 218 ++++++++++++++++++++++++++++++
>   drivers/gpu/drm/xe/xe_pmu.h       |   2 +
>   drivers/gpu/drm/xe/xe_pmu_types.h |  22 +++
>   include/uapi/drm/xe_drm.h         |   5 +
>   5 files changed, 249 insertions(+)
> 
> diff --git a/drivers/gpu/drm/xe/xe_gt.c b/drivers/gpu/drm/xe/xe_gt.c
> index 7a190c49c573..796b330ffd62 100644
> --- a/drivers/gpu/drm/xe/xe_gt.c
> +++ b/drivers/gpu/drm/xe/xe_gt.c
> @@ -929,6 +929,8 @@ int xe_gt_resume(struct xe_gt *gt)
>   
>   	xe_gt_idle_enable_pg(gt);
>   
> +	xe_pmu_resume(gt);
> +
>   	XE_WARN_ON(xe_force_wake_put(gt_to_fw(gt), XE_FORCEWAKE_ALL));
>   	xe_gt_dbg(gt, "resumed\n");
>   
> diff --git a/drivers/gpu/drm/xe/xe_pmu.c b/drivers/gpu/drm/xe/xe_pmu.c
> index b1b38d245e00..39bae25a92ed 100644
> --- a/drivers/gpu/drm/xe/xe_pmu.c
> +++ b/drivers/gpu/drm/xe/xe_pmu.c
> @@ -63,6 +63,12 @@ static unsigned int other_bit(const u64 config)
>   	case XE_PMU_RC6_RESIDENCY:
>   		val = __XE_PMU_RC6_RESIDENCY_ENABLED;
>   		break;
> +	case XE_PMU_ACTUAL_FREQUENCY:
> +		val =  __XE_PMU_ACTUAL_FREQUENCY_ENABLED;
> +		break;
> +	case XE_PMU_REQUESTED_FREQUENCY:
> +		val = __XE_PMU_REQUESTED_FREQUENCY_ENABLED;
> +		break;
>   	default:
>   		/*
>   		 * Events that do not require sampling, or tracking state
> @@ -79,6 +85,22 @@ static unsigned int config_bit(const u64 config)
>   	return other_bit(config);
>   }
>   
> +static u32 config_mask(const u64 config)
> +{
> +	unsigned int bit = config_bit(config);
> +
> +	if (__builtin_constant_p(config))
> +		BUILD_BUG_ON(bit >
> +			     BITS_PER_TYPE(typeof_member(struct xe_pmu,
> +					   enable)) - 1);
> +	else
> +		WARN_ON_ONCE(bit >
> +			     BITS_PER_TYPE(typeof_member(struct xe_pmu,
> +							 enable)) - 1);
> +
> +	return BIT(config_bit(config));
> +}
> +
>   static unsigned int event_bit(struct perf_event *event)
>   {
>   	return config_bit(event->attr.config);
> @@ -107,6 +129,10 @@ config_status(struct xe_device *xe, u64 config)
>   		if (xe->info.skip_guc_pc)
>   			return -ENODEV;
>   		break;
> +	case XE_PMU_ACTUAL_FREQUENCY:
> +		fallthrough;
> +	case XE_PMU_REQUESTED_FREQUENCY:
> +		break;
>   	default:
>   		return -ENOENT;
>   	}
> @@ -169,6 +195,12 @@ store_sample(struct xe_pmu *pmu, unsigned int gt_id, int sample, u64 val)
>   	pmu->event_sample[gt_id][sample].cur = val;
>   }
>   
> +static void
> +add_sample_mult(struct xe_pmu *pmu, unsigned int gt_id, int sample, u32 val, u32 mul)
> +{
> +	pmu->event_sample[gt_id][sample].cur += mul_u32_u32(val, mul);
> +}
> +
>   static u64 get_rc6(struct xe_gt *gt)
>   {
>   	struct xe_device *xe = gt_to_xe(gt);
> @@ -214,6 +246,7 @@ static u64 __xe_pmu_event_read(struct perf_event *event)
>   {
>   	struct xe_device *xe =
>   		container_of(event->pmu, typeof(*xe), pmu.base);
> +	struct xe_pmu *pmu = &xe->pmu;
>   	const unsigned int gt_id = config_gt_id(event->attr.config);
>   	const u64 config = event->attr.config;
>   	struct xe_gt *gt = xe_device_get_gt(xe, gt_id);
> @@ -223,6 +256,18 @@ static u64 __xe_pmu_event_read(struct perf_event *event)
>   	case XE_PMU_RC6_RESIDENCY:
>   		val = get_rc6(gt);
>   		break;
> +	case XE_PMU_ACTUAL_FREQUENCY:
> +		val =
> +		   div_u64(read_sample(pmu, gt_id,
> +				       __XE_SAMPLE_FREQ_ACT),
> +			   USEC_PER_SEC /* to MHz */);
> +		break;
> +	case XE_PMU_REQUESTED_FREQUENCY:
> +		val =
> +		   div_u64(read_sample(pmu, gt_id,
> +				       __XE_SAMPLE_FREQ_REQ),
> +			   USEC_PER_SEC /* to MHz */);
> +		break;
>   	default:
>   		drm_warn(&gt->tile->xe->drm, "unknown pmu event\n");
>   	}
> @@ -252,6 +297,144 @@ static void xe_pmu_event_read(struct perf_event *event)
>   	local64_add(new - prev, &event->count);
>   }
>   
> +static u32 frequency_enabled_mask(void)
> +{
> +	unsigned int i;
> +	u32 mask = 0;
> +
> +	for (i = 0; i < XE_PMU_MAX_GT; i++)
> +		mask |= config_mask(__XE_PMU_ACTUAL_FREQUENCY(i)) |
> +		config_mask(__XE_PMU_REQUESTED_FREQUENCY(i));
> +
> +	return mask;
> +}
> +
> +static bool
> +frequency_sampling_enabled(struct xe_pmu *pmu, unsigned int gt)
> +{
> +	return pmu->enable &
> +	       (config_mask(__XE_PMU_ACTUAL_FREQUENCY(gt)) |
> +		config_mask(__XE_PMU_REQUESTED_FREQUENCY(gt)));
> +}
> +
> +static void
> +frequency_sample(struct xe_gt *gt, unsigned int period_ns)
> +{
> +	struct xe_device *xe = gt_to_xe(gt);
> +	const unsigned int gt_id = gt->info.id;
> +	struct xe_pmu *pmu = &xe->pmu;
> +	bool device_awake;
> +	int ret;
> +	u32 cur_freq;
> +
> +	if (!frequency_sampling_enabled(pmu, gt_id))
> +		return;
> +
> +	/* Report 0/0 (actual/requested) frequency while GT is suspended. */
> +	device_awake = xe_pm_runtime_get_if_active(xe);
> +	if (!device_awake)
> +		return;
> +
> +	if (pmu->enable & config_mask(__XE_PMU_ACTUAL_FREQUENCY(gt_id))) {
> +		u32 val;
> +
> +		/*
> +		 * We take a quick peek here without using forcewake
> +		 * so that we don't perturb the system under observation
> +		 * (forcewake => !rc6 => increased power use). We expect
> +		 * that if the read fails because it is outside of the
> +		 * mmio power well, then it will return 0 -- in which
> +		 * case we assume the system is running at the intended
> +		 * frequency. Fortunately, the read should rarely fail!
> +		 */
> +		val = xe_guc_pc_get_act_freq(&gt->uc.guc.pc);
> +
> +		add_sample_mult(pmu, gt_id, __XE_SAMPLE_FREQ_ACT,
> +				val, period_ns / 1000);
> +	}
> +
> +	if (pmu->enable & config_mask(__XE_PMU_REQUESTED_FREQUENCY(gt_id))) {
> +		ret = xe_guc_pc_get_cur_freq(&gt->uc.guc.pc, &cur_freq);
> +		if (!ret)
> +			add_sample_mult(pmu, gt_id, __XE_SAMPLE_FREQ_REQ,
> +					cur_freq,
> +					period_ns / 1000);
> +	}
> +
> +	xe_pm_runtime_put(xe);
> +}
> +
> +static enum hrtimer_restart xe_sample(struct hrtimer *hrtimer)
> +{
> +	struct xe_pmu *pmu = container_of(hrtimer, struct xe_pmu, timer);
> +	struct xe_device *xe = container_of(pmu, typeof(*xe), pmu);
> +	u64 period = max_t(u64, 10000, NSEC_PER_SEC / FREQUENCY);
> +	unsigned int period_ns;
> +	struct xe_gt *gt;
> +	unsigned int i;
> +	ktime_t now;
> +
> +	if (!READ_ONCE(pmu->timer_enabled))
> +		return HRTIMER_NORESTART;
> +
> +	now = ktime_get();
> +	period_ns = ktime_to_ns(ktime_sub(now, pmu->timer_last));
> +	pmu->timer_last = now;
> +
> +	/*
> +	 * Strictly speaking the passed in period may not be 100% accurate for
> +	 * all internal calculation, since some amount of time can be spent on
> +	 * grabbing the forcewake. However the potential error from timer call-
> +	 * back delay greatly dominates this so we keep it simple.
> +	 */
> +
> +	for_each_gt(gt, xe, i) {
> +		if (!(pmu->active_gts & BIT(i)))
> +			continue;
> +		frequency_sample(gt, period_ns);
> +	}
> +
> +	hrtimer_forward(hrtimer, now, ns_to_ktime(period));
> +
> +	return HRTIMER_RESTART;
> +}
> +
> +static bool pmu_needs_timer(struct xe_pmu *pmu)
> +{
> +	u32 enable;
> +
> +	/*
> +	 * Only some counters need the sampling timer.
> +	 *
> +	 * We start with a bitmask of all currently enabled events.
> +	 */
> +	enable = pmu->enable;
> +
> +	/*
> +	 * Mask out all the ones which do not need the timer, or in
> +	 * other words keep all the ones that could need the timer.
> +	 */
> +	enable &= frequency_enabled_mask();
> +
> +	/*
> +	 * If some bits remain it means we need the sampling timer running.
> +	 */
> +	return enable;
> +}
> +
> +static void __xe_pmu_maybe_start_timer(struct xe_pmu *pmu)
> +{
> +	u64 period = max_t(u64, 10000, NSEC_PER_SEC / FREQUENCY);
> +
> +	if (!pmu->timer_enabled && pmu_needs_timer(pmu)) {
> +		pmu->timer_enabled = true;
> +		pmu->timer_last = ktime_get();
> +		hrtimer_start_range_ns(&pmu->timer,
> +				       ns_to_ktime(period), 0,
> +				       HRTIMER_MODE_REL_PINNED);
> +	}
> +}
> +
>   static void xe_pmu_enable(struct perf_event *event)
>   {
>   	struct xe_pmu *pmu = event_to_pmu(event);
> @@ -263,6 +446,11 @@ static void xe_pmu_enable(struct perf_event *event)
>   
>   	spin_lock_irqsave(&pmu->lock, flags);
>   
> +	/*
> +	 * Start the sampling timer if needed and not already enabled.
> +	 */
> +	__xe_pmu_maybe_start_timer(pmu);
> +
>   	/*
>   	 * Update the bitmask of enabled events and increment
>   	 * the event reference counter.
> @@ -403,6 +591,8 @@ create_event_attributes(struct xe_pmu *pmu)
>   		const char *unit;
>   	} events[] = {
>   		__event(0, "rc6-residency", "ms"),
> +		__event(1, "actual-frequency", "M"),
> +		__event(2, "requested-frequency", "M"),
>   	};
>   
>   	struct perf_pmu_events_attr *pmu_attr = NULL, *pmu_iter;
> @@ -617,6 +807,30 @@ void xe_pmu_suspend(struct xe_gt *gt)
>   	spin_unlock_irq(&pmu->lock);
>   }
>   
> +/**
> + * xe_pmu_resume() - Restart the timer if needed
> + */missing kernel-doc format

Thanks,
Riana
> +void xe_pmu_resume(struct xe_gt *gt)
> +{
> +	struct xe_device *xe = gt_to_xe(gt);
> +	struct xe_pmu *pmu = &xe->pmu;
> +
> +	if (!pmu->base.event_init)
> +		return;
> +
> +	spin_lock_irq(&pmu->lock);
> +
> +	/*
> +	 * Re-enable sampling timer when GPU goes active.
> +	 */
> +	if (pmu->active_gts == 0)
> +		__xe_pmu_maybe_start_timer(pmu);
> +
> +	pmu->active_gts |= BIT(gt->info.id);
> +
> +	spin_unlock_irq(&pmu->lock);
> +}
> +
>   /**
>    * xe_pmu_unregister() - Remove/cleanup PMU registration
>    */
> @@ -635,6 +849,8 @@ void xe_pmu_unregister(void *arg)
>   	pmu->closed = true;
>   	synchronize_rcu();
>   
> +	hrtimer_cancel(&pmu->timer);
> +
>   	xe_pmu_unregister_cpuhp_state(pmu);
>   
>   	perf_pmu_unregister(&pmu->base);
> @@ -678,6 +894,8 @@ void xe_pmu_register(struct xe_pmu *pmu)
>   	int ret = -ENOMEM;
>   
>   	spin_lock_init(&pmu->lock);
> +	hrtimer_init(&pmu->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
> +	pmu->timer.function = xe_sample;
>   	pmu->cpuhp.cpu = -1;
>   
>   	pmu->name = kasprintf(GFP_KERNEL,
> diff --git a/drivers/gpu/drm/xe/xe_pmu.h b/drivers/gpu/drm/xe/xe_pmu.h
> index 17f5a8d7d45c..fb6e3819d7bc 100644
> --- a/drivers/gpu/drm/xe/xe_pmu.h
> +++ b/drivers/gpu/drm/xe/xe_pmu.h
> @@ -16,12 +16,14 @@ void xe_pmu_exit(void);
>   void xe_pmu_register(struct xe_pmu *pmu);
>   void xe_pmu_unregister(void *arg);
>   void xe_pmu_suspend(struct xe_gt *gt);
> +void xe_pmu_resume(struct xe_gt *gt);
>   #else
>   static inline int xe_pmu_init(void) { return 0; }
>   static inline void xe_pmu_exit(void) {}
>   static inline void xe_pmu_register(struct xe_pmu *pmu) {}
>   static inline void xe_pmu_unregister(void *arg) {}
>   static inline void xe_pmu_suspend(struct xe_gt *gt) {}
> +static inline void xe_pmu_resume(struct xe_gt *gt) {}
>   #endif
>   
>   #endif
> diff --git a/drivers/gpu/drm/xe/xe_pmu_types.h b/drivers/gpu/drm/xe/xe_pmu_types.h
> index 1213d2a73492..f4cf85d72aac 100644
> --- a/drivers/gpu/drm/xe/xe_pmu_types.h
> +++ b/drivers/gpu/drm/xe/xe_pmu_types.h
> @@ -13,6 +13,8 @@
>   enum {
>   	__XE_SAMPLE_RC6,
>   	__XE_SAMPLE_RC6_LAST_REPORTED,
> +	__XE_SAMPLE_FREQ_ACT,
> +	__XE_SAMPLE_FREQ_REQ,
>   	__XE_NUM_PMU_SAMPLERS
>   };
>   
> @@ -23,6 +25,8 @@ enum {
>    * current state.
>    */
>   enum xe_pmu_tracked_events {
> +	__XE_PMU_ACTUAL_FREQUENCY_ENABLED,
> +	__XE_PMU_REQUESTED_FREQUENCY_ENABLED,
>   	__XE_PMU_RC6_RESIDENCY_ENABLED,
>   	__XE_PMU_TRACKED_EVENT_COUNT, /* count marker */
>   };
> @@ -116,6 +120,24 @@ struct xe_pmu {
>   	 * @sleep_last: Last time GT parked for RC6 estimation.
>   	 */
>   	ktime_t sleep_last[XE_PMU_MAX_GT];
> +	/**
> +	 * @timer: Timer for internal Xe PMU sampling.
> +	 */
> +	struct hrtimer timer;
> +	/**
> +	 * @timer_last:
> +	 *
> +	 * Timestmap of the previous timer invocation.
> +	 */
> +	ktime_t timer_last;
> +	/**
> +	 * @timer_enabled: Should the internal sampling timer be running.
> +	 */
> +	bool timer_enabled;
> +	/**
> +	 * @active_gts: GT active mask.
> +	 */
> +	unsigned int active_gts;
>   };
>   
>   #endif
> diff --git a/include/uapi/drm/xe_drm.h b/include/uapi/drm/xe_drm.h
> index 3ef3926551dd..1f0ddbd77838 100644
> --- a/include/uapi/drm/xe_drm.h
> +++ b/include/uapi/drm/xe_drm.h
> @@ -1399,7 +1399,12 @@ struct drm_xe_wait_user_fence {
>   #define __XE_PMU_OTHER(x) ___XE_PMU_OTHER(0, x)
>   
>   #define XE_PMU_RC6_RESIDENCY                    __XE_PMU_OTHER(0)
> +#define XE_PMU_ACTUAL_FREQUENCY			__XE_PMU_OTHER(1)
> +#define XE_PMU_REQUESTED_FREQUENCY		__XE_PMU_OTHER(2)
> +
>   #define __XE_PMU_RC6_RESIDENCY(gt)              ___XE_PMU_OTHER(gt, 0)
> +#define __XE_PMU_ACTUAL_FREQUENCY(gt)		___XE_PMU_OTHER(gt, 1)
> +#define __XE_PMU_REQUESTED_FREQUENCY(gt)	___XE_PMU_OTHER(gt, 2)
>   
>   /**
>    * enum drm_xe_observation_type - Observation stream types


More information about the Intel-xe mailing list