[Intel-gfx] [PATCH] drm/i915/pmu: Use GT parked for estimating RC6 while asleep
Tvrtko Ursulin
tvrtko.ursulin at linux.intel.com
Thu Sep 12 09:20:39 UTC 2019
On 11/09/2019 17:38, Chris Wilson wrote:
> As we track when we put the GT device to sleep upon idling, we can use
> that callback to sample the current rc6 counters and record the
> timestamp for estimating samples after that point while asleep.
>
> v2: Stick to using ktime_t
> v3: Track user_wakerefs that interfere with the new
> intel_gt_pm_wait_for_idle
>
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105010
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> ---
> drivers/gpu/drm/i915/gem/i915_gem_pm.c | 19 ++++
> drivers/gpu/drm/i915/gt/intel_gt_types.h | 1 +
> drivers/gpu/drm/i915/i915_debugfs.c | 22 ++---
> drivers/gpu/drm/i915/i915_pmu.c | 120 +++++++++++------------
> drivers/gpu/drm/i915/i915_pmu.h | 4 +-
> 5 files changed, 90 insertions(+), 76 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_pm.c b/drivers/gpu/drm/i915/gem/i915_gem_pm.c
> index 3bd764104d41..45a72cb698db 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_pm.c
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_pm.c
> @@ -141,6 +141,21 @@ bool i915_gem_load_power_context(struct drm_i915_private *i915)
> return switch_to_kernel_context_sync(&i915->gt);
> }
>
> +static void user_forcewake(struct intel_gt *gt, bool suspend)
> +{
> + int count = atomic_read(>->user_wakeref);
> +
> + if (likely(!count))
> + return;
> +
> + intel_gt_pm_get(gt);
> + if (suspend)
> + atomic_sub(count, >->wakeref.count);
> + else
> + atomic_add(count, >->wakeref.count);
> + intel_gt_pm_put(gt);
> +}
> +
> void i915_gem_suspend(struct drm_i915_private *i915)
> {
> GEM_TRACE("\n");
> @@ -148,6 +163,8 @@ void i915_gem_suspend(struct drm_i915_private *i915)
> intel_wakeref_auto(&i915->ggtt.userfault_wakeref, 0);
> flush_workqueue(i915->wq);
>
> + user_forcewake(&i915->gt, true);
This complication is needed only because you changed user forcewake
handling to use intel_gt_pm_get/put instead of intel_runtime_pm_get?
Which in turn is because of the CONFIG_PM ifdef removal below? Wouldn't
it be simpler to keep both as were? Maybe I am missing something...
> +
> mutex_lock(&i915->drm.struct_mutex);
>
> /*
> @@ -259,6 +276,8 @@ void i915_gem_resume(struct drm_i915_private *i915)
> if (!i915_gem_load_power_context(i915))
> goto err_wedged;
>
> + user_forcewake(&i915->gt, false);
> +
> out_unlock:
> intel_uncore_forcewake_put(&i915->uncore, FORCEWAKE_ALL);
> mutex_unlock(&i915->drm.struct_mutex);
> diff --git a/drivers/gpu/drm/i915/gt/intel_gt_types.h b/drivers/gpu/drm/i915/gt/intel_gt_types.h
> index dc295c196d11..3039cef64b11 100644
> --- a/drivers/gpu/drm/i915/gt/intel_gt_types.h
> +++ b/drivers/gpu/drm/i915/gt/intel_gt_types.h
> @@ -50,6 +50,7 @@ struct intel_gt {
> } timelines;
>
> struct intel_wakeref wakeref;
> + atomic_t user_wakeref;
>
> struct list_head closed_vma;
> spinlock_t closed_lock; /* guards the list of closed_vma */
> diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c
> index 708855e051b5..f00c0dc4f57e 100644
> --- a/drivers/gpu/drm/i915/i915_debugfs.c
> +++ b/drivers/gpu/drm/i915/i915_debugfs.c
> @@ -3988,13 +3988,12 @@ static int i915_sseu_status(struct seq_file *m, void *unused)
> static int i915_forcewake_open(struct inode *inode, struct file *file)
> {
> struct drm_i915_private *i915 = inode->i_private;
> + struct intel_gt *gt = &i915->gt;
>
> - if (INTEL_GEN(i915) < 6)
> - return 0;
> -
> - file->private_data =
> - (void *)(uintptr_t)intel_runtime_pm_get(&i915->runtime_pm);
> - intel_uncore_forcewake_user_get(&i915->uncore);
> + atomic_inc(>->user_wakeref);
> + intel_gt_pm_get(gt);
> + if (INTEL_GEN(i915) >= 6)
> + intel_uncore_forcewake_user_get(gt->uncore);
>
> return 0;
> }
> @@ -4002,13 +4001,12 @@ static int i915_forcewake_open(struct inode *inode, struct file *file)
> static int i915_forcewake_release(struct inode *inode, struct file *file)
> {
> struct drm_i915_private *i915 = inode->i_private;
> + struct intel_gt *gt = &i915->gt;
>
> - if (INTEL_GEN(i915) < 6)
> - return 0;
> -
> - intel_uncore_forcewake_user_put(&i915->uncore);
> - intel_runtime_pm_put(&i915->runtime_pm,
> - (intel_wakeref_t)(uintptr_t)file->private_data);
> + if (INTEL_GEN(i915) >= 6)
> + intel_uncore_forcewake_user_put(&i915->uncore);
> + intel_gt_pm_put(gt);
> + atomic_dec(>->user_wakeref);
>
> return 0;
> }
> diff --git a/drivers/gpu/drm/i915/i915_pmu.c b/drivers/gpu/drm/i915/i915_pmu.c
> index 8e251e719390..a3d3d3e39c15 100644
> --- a/drivers/gpu/drm/i915/i915_pmu.c
> +++ b/drivers/gpu/drm/i915/i915_pmu.c
> @@ -116,19 +116,51 @@ static bool pmu_needs_timer(struct i915_pmu *pmu, bool gpu_active)
> return enable;
> }
>
> +static u64 __get_rc6(struct intel_gt *gt)
> +{
> + struct drm_i915_private *i915 = gt->i915;
> + u64 val;
> +
> + val = intel_rc6_residency_ns(i915,
> + IS_VALLEYVIEW(i915) ?
> + VLV_GT_RENDER_RC6 :
> + GEN6_GT_GFX_RC6);
> +
> + if (HAS_RC6p(i915))
> + val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6p);
> +
> + if (HAS_RC6pp(i915))
> + val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6pp);
> +
> + return val;
> +}
> +
> void i915_pmu_gt_parked(struct drm_i915_private *i915)
> {
> struct i915_pmu *pmu = &i915->pmu;
> + u64 val;
>
> if (!pmu->base.event_init)
> return;
>
> spin_lock_irq(&pmu->lock);
> +
> + val = 0;
> + if (pmu->sample[__I915_SAMPLE_RC6].cur)
> + val = __get_rc6(&i915->gt);
> +
> + if (val >= pmu->sample[__I915_SAMPLE_RC6_ESTIMATED].cur) {
> + pmu->sample[__I915_SAMPLE_RC6_ESTIMATED].cur = 0;
> + pmu->sample[__I915_SAMPLE_RC6].cur = val;
> + }
> + pmu->sleep_last = ktime_get();
> +
> /*
> * Signal sampling timer to stop if only engine events are enabled and
> * GPU went idle.
> */
> pmu->timer_enabled = pmu_needs_timer(pmu, false);
> +
> spin_unlock_irq(&pmu->lock);
> }
>
> @@ -143,6 +175,11 @@ static void __i915_pmu_maybe_start_timer(struct i915_pmu *pmu)
> }
> }
>
> +static inline s64 ktime_since(const ktime_t kt)
> +{
> + return ktime_to_ns(ktime_sub(ktime_get(), kt));
> +}
> +
> void i915_pmu_gt_unparked(struct drm_i915_private *i915)
> {
> struct i915_pmu *pmu = &i915->pmu;
> @@ -151,10 +188,22 @@ void i915_pmu_gt_unparked(struct drm_i915_private *i915)
> return;
>
> spin_lock_irq(&pmu->lock);
> +
> /*
> * Re-enable sampling timer when GPU goes active.
> */
> __i915_pmu_maybe_start_timer(pmu);
> +
> + /* Estimate how long we slept and accumulate that into rc6 counters */
> + if (pmu->sample[__I915_SAMPLE_RC6].cur) {
> + u64 val;
> +
> + val = ktime_since(pmu->sleep_last);
> + val += pmu->sample[__I915_SAMPLE_RC6].cur;
> +
> + pmu->sample[__I915_SAMPLE_RC6_ESTIMATED].cur = val;
> + }
> +
> spin_unlock_irq(&pmu->lock);
> }
>
> @@ -426,39 +475,18 @@ static int i915_pmu_event_init(struct perf_event *event)
> return 0;
> }
>
> -static u64 __get_rc6(struct intel_gt *gt)
> -{
> - struct drm_i915_private *i915 = gt->i915;
> - u64 val;
> -
> - val = intel_rc6_residency_ns(i915,
> - IS_VALLEYVIEW(i915) ?
> - VLV_GT_RENDER_RC6 :
> - GEN6_GT_GFX_RC6);
> -
> - if (HAS_RC6p(i915))
> - val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6p);
> -
> - if (HAS_RC6pp(i915))
> - val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6pp);
> -
> - return val;
> -}
> -
> static u64 get_rc6(struct intel_gt *gt)
> {
> -#if IS_ENABLED(CONFIG_PM)
> struct drm_i915_private *i915 = gt->i915;
> - struct intel_runtime_pm *rpm = &i915->runtime_pm;
> struct i915_pmu *pmu = &i915->pmu;
> - intel_wakeref_t wakeref;
> unsigned long flags;
> u64 val;
>
> - wakeref = intel_runtime_pm_get_if_in_use(rpm);
> - if (wakeref) {
> + spin_lock_irqsave(&pmu->lock, flags);
> +
> + if (intel_gt_pm_get_if_awake(gt)) {
> val = __get_rc6(gt);
> - intel_runtime_pm_put(rpm, wakeref);
> + intel_gt_pm_put(gt);
>
> /*
> * If we are coming back from being runtime suspended we must
> @@ -466,19 +494,13 @@ static u64 get_rc6(struct intel_gt *gt)
> * previously.
> */
>
> - spin_lock_irqsave(&pmu->lock, flags);
> -
> if (val >= pmu->sample[__I915_SAMPLE_RC6_ESTIMATED].cur) {
> pmu->sample[__I915_SAMPLE_RC6_ESTIMATED].cur = 0;
> pmu->sample[__I915_SAMPLE_RC6].cur = val;
> } else {
> val = pmu->sample[__I915_SAMPLE_RC6_ESTIMATED].cur;
> }
> -
> - spin_unlock_irqrestore(&pmu->lock, flags);
For this branch pmu->lock is only needed over the estimation block, not
over __get_rc6() and intel_gt_pm_get_if_awake(). But I agree it's more
efficient to do it like this to avoid multiple irq-off-on transitions
via intel_rc6_residency_ns. I wanted to suggest local_irq_disable and
separate spin_(un)lock's on both if branches for more self-documenting ,
less confusion, but then single call also has it's benefits.
> } else {
> - struct device *kdev = rpm->kdev;
> -
> /*
> * We are runtime suspended.
> *
> @@ -486,42 +508,16 @@ static u64 get_rc6(struct intel_gt *gt)
> * on top of the last known real value, as the approximated RC6
> * counter value.
> */
> - spin_lock_irqsave(&pmu->lock, flags);
>
> - /*
> - * After the above branch intel_runtime_pm_get_if_in_use failed
> - * to get the runtime PM reference we cannot assume we are in
> - * runtime suspend since we can either: a) race with coming out
> - * of it before we took the power.lock, or b) there are other
> - * states than suspended which can bring us here.
> - *
> - * We need to double-check that we are indeed currently runtime
> - * suspended and if not we cannot do better than report the last
> - * known RC6 value.
> - */
> - if (pm_runtime_status_suspended(kdev)) {
> - val = pm_runtime_suspended_time(kdev);
> + val = ktime_since(pmu->sleep_last);
> + val += pmu->sample[__I915_SAMPLE_RC6].cur;
>
> - if (!pmu->sample[__I915_SAMPLE_RC6_ESTIMATED].cur)
> - pmu->suspended_time_last = val;
> -
> - val -= pmu->suspended_time_last;
> - val += pmu->sample[__I915_SAMPLE_RC6].cur;
> -
> - pmu->sample[__I915_SAMPLE_RC6_ESTIMATED].cur = val;
> - } else if (pmu->sample[__I915_SAMPLE_RC6_ESTIMATED].cur) {
> - val = pmu->sample[__I915_SAMPLE_RC6_ESTIMATED].cur;
> - } else {
> - val = pmu->sample[__I915_SAMPLE_RC6].cur;
> - }
> -
> - spin_unlock_irqrestore(&pmu->lock, flags);
> + pmu->sample[__I915_SAMPLE_RC6_ESTIMATED].cur = val;
> }
>
> + spin_unlock_irqrestore(&pmu->lock, flags);
> +
> return val;
> -#else
> - return __get_rc6(gt);
> -#endif
Don't we end up doing the irqsave spinlock needlessly when !CONFIG_PM?
> }
>
> static u64 __i915_pmu_event_read(struct perf_event *event)
> diff --git a/drivers/gpu/drm/i915/i915_pmu.h b/drivers/gpu/drm/i915/i915_pmu.h
> index 4fc4f2478301..067dbbf3bdff 100644
> --- a/drivers/gpu/drm/i915/i915_pmu.h
> +++ b/drivers/gpu/drm/i915/i915_pmu.h
> @@ -97,9 +97,9 @@ struct i915_pmu {
> */
> struct i915_pmu_sample sample[__I915_NUM_PMU_SAMPLERS];
> /**
> - * @suspended_time_last: Cached suspend time from PM core.
> + * @sleep_last: Last time GT parked for RC6 estimation.
> */
> - u64 suspended_time_last;
> + ktime_t sleep_last;
> /**
> * @i915_attr: Memory block holding device attributes.
> */
>
Regards,
Tvrtko
More information about the Intel-gfx
mailing list