[PATCH v3] drm/xe/pmu: Add GT frequency events
Belgaumkar, Vinay
vinay.belgaumkar at intel.com
Thu Mar 6 01:47:02 UTC 2025
On 2/25/2025 9:47 AM, Lucas De Marchi wrote:
> On Wed, Feb 12, 2025 at 04:02:20PM -0800, Vinay Belgaumkar wrote:
>> Define PMU events for GT frequency (actual and requested). This is
>> a port from the i915 driver implementation, where an internal timer
>> is used to aggregate GT frequencies over certain fixed interval.
>> Following PMU events are being added-
>>
>> xe_0000_00_02.0/gt-actual-frequency/ [Kernel PMU event]
>> xe_0000_00_02.0/gt-requested-frequency/ [Kernel PMU event]
>>
>> Standard perf commands can be used to monitor GT frequency-
>> $ perf stat -e xe_0000_00_02.0/gt-requested-frequency,gt=0/ -I1000
>>
>> 1.001175175 700 M xe/gt-requested-frequency,gt=0/
>> 2.005891881 703 M xe/gt-requested-frequency,gt=0/
>> 3.007318169 700 M xe/gt-requested-frequency,gt=0/
>>
>> Actual frequencies will be reported as 0 when GT is in C6.
>>
>> v2: Use locks while storing/reading samples, keep track of multiple
>> clients (Lucas) and other general cleanup.
>>
>> Cc: Riana Tauro <riana.tauro at intel.com>
>> Cc: Lucas De Marchi <lucas.demarchi at intel.com>
>> Cc: Rodrigo Vivi <rodrigo.vivi at intel.com>
>> Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar at intel.com>
>> ---
>> drivers/gpu/drm/xe/xe_pmu.c | 211 ++++++++++++++++++++++++++++--
>> drivers/gpu/drm/xe/xe_pmu_types.h | 29 ++++
>> 2 files changed, 230 insertions(+), 10 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/xe/xe_pmu.c b/drivers/gpu/drm/xe/xe_pmu.c
>> index 3910a82328ee..ed76756403ea 100644
>> --- a/drivers/gpu/drm/xe/xe_pmu.c
>> +++ b/drivers/gpu/drm/xe/xe_pmu.c
>> @@ -8,6 +8,8 @@
>>
>> #include "xe_device.h"
>> #include "xe_gt_idle.h"
>> +#include "xe_guc_pc.h"
>> +#include "xe_macros.h"
>> #include "xe_pm.h"
>> #include "xe_pmu.h"
>>
>> @@ -37,6 +39,17 @@
>>
>> #define XE_PMU_EVENT_GT_MASK GENMASK_ULL(63, 60)
>> #define XE_PMU_EVENT_ID_MASK GENMASK_ULL(11, 0)
>> +#define XE_HRTIMER_INTERVAL_NS 5000000
>> +
>> +static struct xe_pmu *event_to_pmu(struct perf_event *event)
>> +{
>> + return container_of(event->pmu, typeof(struct xe_pmu), base);
>> +}
>> +
>> +static struct xe_device *event_to_xe(struct perf_event *event)
>> +{
>> + return container_of(event->pmu, typeof(struct xe_device),
>> pmu.base);
>> +}
>
> we shouldn't add refactors like this in the same patch you are adding a
> new feature, it creates noise in the patch and makes it harder to see
> what's going on. Please drop both event_to_pmu() and event_to_xe().
>
>>
>> static unsigned int config_to_event_id(u64 config)
>> {
>> @@ -49,10 +62,12 @@ static unsigned int config_to_gt_id(u64 config)
>> }
>>
>> #define XE_PMU_EVENT_GT_C6_RESIDENCY 0x01
>> +#define XE_PMU_EVENT_GT_ACTUAL_FREQUENCY 0x02
>> +#define XE_PMU_EVENT_GT_REQUESTED_FREQUENCY 0x03
>
> this will need a rebase
>
>>
>> static struct xe_gt *event_to_gt(struct perf_event *event)
>> {
>> - struct xe_device *xe = container_of(event->pmu, typeof(*xe),
>> pmu.base);
>> + struct xe_device *xe = event_to_xe(event);
>> u64 gt = config_to_gt_id(event->attr.config);
>>
>> return xe_device_get_gt(xe, gt);
>> @@ -70,7 +85,7 @@ static bool event_supported(struct xe_pmu *pmu,
>> unsigned int gt,
>>
>> static void xe_pmu_event_destroy(struct perf_event *event)
>> {
>> - struct xe_device *xe = container_of(event->pmu, typeof(*xe),
>> pmu.base);
>> + struct xe_device *xe = event_to_xe(event);
>>
>> drm_WARN_ON(&xe->drm, event->parent);
>> xe_pm_runtime_put(xe);
>> @@ -79,7 +94,7 @@ static void xe_pmu_event_destroy(struct perf_event
>> *event)
>>
>> static int xe_pmu_event_init(struct perf_event *event)
>> {
>> - struct xe_device *xe = container_of(event->pmu, typeof(*xe),
>> pmu.base);
>> + struct xe_device *xe = event_to_xe(event);
>> struct xe_pmu *pmu = &xe->pmu;
>> unsigned int id, gt;
>>
>> @@ -104,6 +119,8 @@ static int xe_pmu_event_init(struct perf_event
>> *event)
>> if (has_branch_stack(event))
>> return -EOPNOTSUPP;
>>
>> + XE_WARN_ON(id >= XE_PMU_MAX_EVENT_TYPES);
>
> I was going to say that we don't need the other ones. But then I noticed
> XE_PMU_MAX_EVENT_TYPES == 64, which is a lot more than we currently
> support.
>
> I suggested it before, but what's the issue of keeping a linked list of
> the events like rapl_pmu::active_list?
Ok, while testing out the implementation for this, I noticed a
limitation of using lists- if we have 2 perf requests for the same
event, we end up processing that event twice. Since I am adding up
values every time, that inflates the frequency values by a factor of 2.
Seems like we need to iterate per GT here instead of per event. So,
using a bitmask per GT might be better.
Thanks,
Vinay.
>
>> +
>> if (!event->parent) {
>> drm_dev_get(&xe->drm);
>> xe_pm_runtime_get(xe);
>> @@ -113,16 +130,50 @@ static int xe_pmu_event_init(struct perf_event
>> *event)
>> return 0;
>> }
>>
>> +static u64
>> +xe_get_pmu_sample(struct xe_pmu *pmu, unsigned int gt_id, uint32_t id)
>> +{
>> + unsigned long flags;
>> + u64 val;
>> +
>> + raw_spin_lock_irqsave(&pmu->lock, flags);
>> + val = pmu->event_sample[gt_id][id];
>> + raw_spin_unlock_irqrestore(&pmu->lock, flags);
>> +
>> + return val;
>> +}
>> +
>> +static void
>> +xe_store_pmu_sample(struct xe_pmu *pmu, uint32_t gt_id, uint32_t id,
>> + uint32_t val, uint32_t period)
>> +{
>> + unsigned long flags;
>> +
>> + raw_spin_lock_irqsave(&pmu->lock, flags);
>> + pmu->event_sample[gt_id][id] += mul_u32_u32(val, period);
>> + raw_spin_unlock_irqrestore(&pmu->lock, flags);
>> +}
>> +
>> static u64 __xe_pmu_event_read(struct perf_event *event)
>> {
>> struct xe_gt *gt = event_to_gt(event);
>> + struct xe_device *xe = event_to_xe(event);
>
> gt_to_xe() avoids the additional helper and you don't even need it.
> You are intested in the pmu, not xe... so you could just have used
>
> pmu = container_of(event->pmu, typeof(struct xe_pmu), base);
>
> ... like is done in other places.
>
>> + struct xe_pmu *pmu = &xe->pmu;
>> + unsigned long id;
>>
>> if (!gt)
>> return 0;
>>
>> - switch (config_to_event_id(event->attr.config)) {
>> + id = config_to_event_id(event->attr.config);
>> +
>> + switch (id) {
>> case XE_PMU_EVENT_GT_C6_RESIDENCY:
>> return xe_gt_idle_residency_msec(>->gtidle);
>> + case XE_PMU_EVENT_GT_ACTUAL_FREQUENCY:
>> + case XE_PMU_EVENT_GT_REQUESTED_FREQUENCY:
>> + return div_u64(xe_get_pmu_sample(pmu, gt->info.id,
>> + id),
>
> keep in the line above, it doesn't help readability.
>
>> + USEC_PER_SEC /* to MHz */);
>
> why is the unit nsec -> usec when storing, couldn't we just use
> NSEC_PER_SEC here?
>
>> }
>>
>> return 0;
>> @@ -143,7 +194,7 @@ static void xe_pmu_event_update(struct perf_event
>> *event)
>>
>> static void xe_pmu_event_read(struct perf_event *event)
>> {
>> - struct xe_device *xe = container_of(event->pmu, typeof(*xe),
>> pmu.base);
>> + struct xe_device *xe = event_to_xe(event);
>> struct xe_pmu *pmu = &xe->pmu;
>>
>> if (!pmu->registered) {
>> @@ -154,8 +205,105 @@ static void xe_pmu_event_read(struct perf_event
>> *event)
>> xe_pmu_event_update(event);
>> }
>>
>> +static bool gt_pmu_event_active(struct xe_pmu *pmu, uint32_t id,
>> uint32_t gt_id)
>> +{
>> + return pmu->active_mask[gt_id] & BIT_ULL(id);
>> +}
>> +
>> +static bool pmu_needs_timer(struct xe_pmu *pmu)
>> +{
>> + struct xe_device *xe = container_of(pmu, typeof(*xe), pmu);
>> + struct xe_gt *gt;
>> + int i;
>> +
>> + for_each_gt(gt, xe, i)
>> + if (gt_pmu_event_active(pmu,
>> XE_PMU_EVENT_GT_ACTUAL_FREQUENCY, i) ||
>> + gt_pmu_event_active(pmu,
>> XE_PMU_EVENT_GT_REQUESTED_FREQUENCY, i))
>> + return true;
>> +
>> + return false;
>> +}
>> +
>> +static void xe_pmu_start_timer(struct xe_pmu *pmu, uint32_t gt_id)
>> +{
>> + /* Timer is per device */
>> + 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(XE_HRTIMER_INTERVAL_NS), 0,
>> + HRTIMER_MODE_REL_PINNED);
>> + }
>> +}
>> +
>> +static void
>> +xe_sample_gt_frequency(struct xe_gt *gt, uint32_t period_ns)
>> +{
>> + struct xe_device *xe = gt_to_xe(gt);
>> + struct xe_pmu *pmu = &xe->pmu;
>> + uint32_t act_freq, cur_freq;
>> + int ret;
>> +
>> + if (gt_pmu_event_active(pmu, XE_PMU_EVENT_GT_ACTUAL_FREQUENCY,
>> gt->info.id)) {
>> +
>
> drop this line
>
>> + /* Actual freq will be 0 when GT is in C6 */
>> + act_freq = xe_guc_pc_get_act_freq(>->uc.guc.pc);
>> + xe_store_pmu_sample(pmu, gt->info.id,
>> XE_PMU_EVENT_GT_ACTUAL_FREQUENCY,
>> + act_freq, period_ns / NSEC_PER_USEC);
>> + }
>> +
>> + if (gt_pmu_event_active(pmu,
>> XE_PMU_EVENT_GT_REQUESTED_FREQUENCY, gt->info.id)) {
>> + ret = xe_guc_pc_get_cur_freq(>->uc.guc.pc, &cur_freq);
>> + if (!ret)
>> + xe_store_pmu_sample(pmu, gt->info.id,
>> XE_PMU_EVENT_GT_REQUESTED_FREQUENCY,
>> + cur_freq, period_ns / NSEC_PER_USEC);
>> + }
>> +}
>> +
>> +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 time_diff_ns;
>> + struct xe_gt *gt;
>> + ktime_t now;
>> + int i;
>> +
>> + if (!READ_ONCE(pmu->timer_enabled))
>> + return HRTIMER_NORESTART;
>> +
>> + now = ktime_get();
>> + time_diff_ns = ktime_to_ns(ktime_sub(now, pmu->timer_last));
>> + pmu->timer_last = now;
>> +
>> + /*
>> + * The passed in period includes the time needed for grabbing
>> the forcewake.
>
> I don't think this is true anymore, is it?
>
> xe_sample()
> 1. calculate time_diff at the beginning of the function
> 2. for each gt, readout freq samples, using the time_diff, may include
> delays on forcewake, but that is not included in the time_diff
> 3. htimer is adjusted to timer_last + period
>
>> + */
>> + for_each_gt(gt, xe, i)
>> + xe_sample_gt_frequency(gt, time_diff_ns);
>> +
>> + hrtimer_forward(hrtimer, now, ns_to_ktime(XE_HRTIMER_INTERVAL_NS));
>> +
>> + return HRTIMER_RESTART;
>> +}
>> +
>> static void xe_pmu_enable(struct perf_event *event)
>> {
>> + struct xe_pmu *pmu = event_to_pmu(event);
>> + struct xe_gt *gt = event_to_gt(event);
>> + uint32_t id = config_to_event_id(event->attr.config);
>> + unsigned long flags;
>> +
>> + raw_spin_lock_irqsave(&pmu->lock, flags);
>> +
>> + /* Update event mask */
>> + pmu->active_mask[gt->info.id] |= BIT_ULL(id);
>> + pmu->n_active[gt->info.id][id]++;
>> +
>> + /* Start a timer, if needed, to collect samples */
>> + xe_pmu_start_timer(pmu, gt->info.id);
>> +
>> + raw_spin_unlock_irqrestore(&pmu->lock, flags);
>> /*
>> * Store the current counter value so we can report the correct
>> delta
>> * for all listeners. Even when the event was already enabled and
>> has
>> @@ -166,7 +314,7 @@ static void xe_pmu_enable(struct perf_event *event)
>>
>> static void xe_pmu_event_start(struct perf_event *event, int flags)
>> {
>> - struct xe_device *xe = container_of(event->pmu, typeof(*xe),
>> pmu.base);
>> + struct xe_device *xe = event_to_xe(event);
>> struct xe_pmu *pmu = &xe->pmu;
>>
>> if (!pmu->registered)
>> @@ -176,26 +324,53 @@ static void xe_pmu_event_start(struct
>> perf_event *event, int flags)
>> event->hw.state = 0;
>> }
>>
>> +static void xe_pmu_disable(struct perf_event *event)
>> +{
>> + struct xe_device *xe = event_to_xe(event);
>> + struct xe_pmu *pmu = &xe->pmu;
>> + unsigned long flags;
>> + uint32_t id = config_to_event_id(event->attr.config);
>> + uint32_t gt_id = config_to_gt_id(event->attr.config);
>> +
>> + raw_spin_lock_irqsave(&pmu->lock, flags);
>> +
>> + if (--pmu->n_active[gt_id][id] == 0) {
>> + pmu->active_mask[gt_id] &= ~BIT_ULL(id);
>> + pmu->timer_enabled &= pmu_needs_timer(pmu);
>> + }
>> +
>> + raw_spin_unlock_irqrestore(&pmu->lock, flags);
>> +}
>> +
>> static void xe_pmu_event_stop(struct perf_event *event, int flags)
>> {
>> - struct xe_device *xe = container_of(event->pmu, typeof(*xe),
>> pmu.base);
>> + struct xe_device *xe = event_to_xe(event);
>> struct xe_pmu *pmu = &xe->pmu;
>> + uint32_t id = config_to_event_id(event->attr.config);
>> +
>> + XE_WARN_ON(id >= XE_PMU_MAX_EVENT_TYPES);
>
> if not using a linked list as suggested, please drop this. Instead block
> the event init from succeeding.
>
>>
>> - if (pmu->registered)
>> + if (pmu->registered) {
>> if (flags & PERF_EF_UPDATE)
>> xe_pmu_event_update(event);
>>
>> + xe_pmu_disable(event);
>> + }
>> +
>> event->hw.state = PERF_HES_STOPPED;
>> }
>>
>> static int xe_pmu_event_add(struct perf_event *event, int flags)
>> {
>> - struct xe_device *xe = container_of(event->pmu, typeof(*xe),
>> pmu.base);
>> + struct xe_device *xe = event_to_xe(event);
>> struct xe_pmu *pmu = &xe->pmu;
>> + uint32_t id = config_to_event_id(event->attr.config);
>>
>> if (!pmu->registered)
>> return -ENODEV;
>>
>> + XE_WARN_ON(id >= XE_PMU_MAX_EVENT_TYPES);
>
> ditto.
>
> Lucas De Marchi
>
>> +
>> if (flags & PERF_EF_START)
>> xe_pmu_event_start(event, flags);
>>
>> @@ -270,6 +445,10 @@ static ssize_t event_attr_show(struct device *dev,
>> XE_EVENT_ATTR_GROUP(v_, id_, &pmu_event_ ##v_.attr.attr)
>>
>> XE_EVENT_ATTR_SIMPLE(gt-c6-residency, gt_c6_residency,
>> XE_PMU_EVENT_GT_C6_RESIDENCY, "ms");
>> +XE_EVENT_ATTR_SIMPLE(gt-actual-frequency, gt_actual_frequency,
>> + XE_PMU_EVENT_GT_ACTUAL_FREQUENCY, "Mhz");
>> +XE_EVENT_ATTR_SIMPLE(gt-requested-frequency, gt_requested_frequency,
>> + XE_PMU_EVENT_GT_REQUESTED_FREQUENCY, "Mhz");
>>
>> static struct attribute *pmu_empty_event_attrs[] = {
>> /* Empty - all events are added as groups with .attr_update() */
>> @@ -283,6 +462,8 @@ static const struct attribute_group
>> pmu_events_attr_group = {
>>
>> static const struct attribute_group *pmu_events_attr_update[] = {
>> &pmu_group_gt_c6_residency,
>> + &pmu_group_gt_actual_frequency,
>> + &pmu_group_gt_requested_frequency,
>> NULL,
>> };
>>
>> @@ -290,8 +471,11 @@ static void set_supported_events(struct xe_pmu
>> *pmu)
>> {
>> struct xe_device *xe = container_of(pmu, typeof(*xe), pmu);
>>
>> - if (!xe->info.skip_guc_pc)
>> + if (!xe->info.skip_guc_pc) {
>> pmu->supported_events |= BIT_ULL(XE_PMU_EVENT_GT_C6_RESIDENCY);
>> + pmu->supported_events |=
>> BIT_ULL(XE_PMU_EVENT_GT_ACTUAL_FREQUENCY);
>> + pmu->supported_events |=
>> BIT_ULL(XE_PMU_EVENT_GT_REQUESTED_FREQUENCY);
>> + }
>> }
>>
>> /**
>> @@ -306,6 +490,8 @@ static void xe_pmu_unregister(void *arg)
>> if (!pmu->registered)
>> return;
>>
>> + hrtimer_cancel(&pmu->timer);
>> +
>> pmu->registered = false;
>>
>> perf_pmu_unregister(&pmu->base);
>> @@ -334,6 +520,11 @@ int xe_pmu_register(struct xe_pmu *pmu)
>> if (IS_SRIOV_VF(xe))
>> return 0;
>>
>> + raw_spin_lock_init(&pmu->lock);
>> +
>> + hrtimer_init(&pmu->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
>> + pmu->timer.function = xe_sample;
>> +
>> name = kasprintf(GFP_KERNEL, "xe_%s",
>> dev_name(xe->drm.dev));
>> if (!name)
>> diff --git a/drivers/gpu/drm/xe/xe_pmu_types.h
>> b/drivers/gpu/drm/xe/xe_pmu_types.h
>> index f5ba4d56622c..cd68cba58693 100644
>> --- a/drivers/gpu/drm/xe/xe_pmu_types.h
>> +++ b/drivers/gpu/drm/xe/xe_pmu_types.h
>> @@ -10,6 +10,7 @@
>> #include <linux/spinlock_types.h>
>>
>> #define XE_PMU_MAX_GT 2
>> +#define XE_PMU_MAX_EVENT_TYPES 64
>>
>> /**
>> * struct xe_pmu - PMU related data per Xe device
>> @@ -34,6 +35,34 @@ struct xe_pmu {
>> * @supported_events: Bitmap of supported events, indexed by
>> event id
>> */
>> u64 supported_events;
>> + /**
>> + * @lock: Lock protecting enable mask and sampling
>> + */
>> + raw_spinlock_t lock;
>> + /**
>> + * @active_mask: Bit representation of active events
>> + */
>> + u64 active_mask[XE_PMU_MAX_GT];
>> + /**
>> + * @n_active: Counts for the enabled events
>> + */
>> + unsigned int n_active[XE_PMU_MAX_GT][XE_PMU_MAX_EVENT_TYPES];
>> + /**
>> + * @timer: Timer for sampling GT frequency.
>> + */
>> + 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;
>> + /**
>> + * @event_sample: Store freq related counters.
>> + */
>> + u64 event_sample[XE_PMU_MAX_GT][XE_PMU_MAX_EVENT_TYPES];
>> };
>>
>> #endif
>> --
>> 2.38.1
>>
More information about the Intel-xe
mailing list