Re: [PATCH v3 2/8] drm/panfrost: Enable cycle counter register upon job submission

[Boris Brezillon]

On Tue,  5 Sep 2023 19:45:18 +0100
Adrián Larumbe <adrian.larumbe@xxxxxxxxxxxxx> wrote:

> In a future development, we will want to keep track of the number of GPU
> cycles spent on a given job. That means we should enable it only when the
> GPU has work to do, and switch it off whenever it is idle to avoid power
> waste.
> 
> To avoid race conditions during enablement/disabling, a reference counting
> mechanism was introduced, and a job flag that tells us whether a given job
> increased the refcount. This is necessary, because a future development
> will let user space toggle cycle counting through a debugfs file, and a
> given job might have been in flight by the time cycle counting was
> disabled.
> 
> Toggling of GPU cycle counting has to be done through a module parameter.
> 
> Signed-off-by: Adrián Larumbe <adrian.larumbe@xxxxxxxxxxxxx>
> ---
>  drivers/gpu/drm/panfrost/panfrost_device.c |  5 +++
>  drivers/gpu/drm/panfrost/panfrost_device.h |  6 +++
>  drivers/gpu/drm/panfrost/panfrost_gpu.c    | 43 ++++++++++++++++++++++
>  drivers/gpu/drm/panfrost/panfrost_gpu.h    |  6 +++
>  drivers/gpu/drm/panfrost/panfrost_job.c    | 10 +++++
>  drivers/gpu/drm/panfrost/panfrost_job.h    |  1 +
>  6 files changed, 71 insertions(+)
> 
> diff --git a/drivers/gpu/drm/panfrost/panfrost_device.c b/drivers/gpu/drm/panfrost/panfrost_device.c
> index fa1a086a862b..1ea2ac3804f0 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_device.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_device.c
> @@ -18,6 +18,9 @@
>  #include "panfrost_mmu.h"
>  #include "panfrost_perfcnt.h"
>  
> +static bool profile;
> +module_param(profile, bool, 0600);

Not sure if we should make that a module parameter. Might be better
exposed as a debugfs knob attached to the device (even if having
multiple Mali devices is rather unlikely, I think I'd prefer to make
this toggle per-device).

> +
>  static int panfrost_reset_init(struct panfrost_device *pfdev)
>  {
>  	pfdev->rstc = devm_reset_control_array_get_optional_exclusive(pfdev->dev);
> @@ -207,6 +210,8 @@ int panfrost_device_init(struct panfrost_device *pfdev)
>  
>  	spin_lock_init(&pfdev->as_lock);
>  
> +	atomic_set(&pfdev->profile_mode, profile);

So, profile_mode can only be set at probe time, meaning any changes to
the profile module param is not taken into account after that point.

> +
>  	err = panfrost_clk_init(pfdev);
>  	if (err) {
>  		dev_err(pfdev->dev, "clk init failed %d\n", err);
> diff --git a/drivers/gpu/drm/panfrost/panfrost_device.h b/drivers/gpu/drm/panfrost/panfrost_device.h
> index b0126b9fbadc..5c09c9f3ae08 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_device.h
> +++ b/drivers/gpu/drm/panfrost/panfrost_device.h
> @@ -107,6 +107,7 @@ struct panfrost_device {
>  	struct list_head scheduled_jobs;
>  
>  	struct panfrost_perfcnt *perfcnt;
> +	atomic_t profile_mode;
>  
>  	struct mutex sched_lock;
>  
> @@ -121,6 +122,11 @@ struct panfrost_device {
>  	struct shrinker shrinker;
>  
>  	struct panfrost_devfreq pfdevfreq;
> +
> +	struct {
> +		atomic_t use_count;
> +		spinlock_t lock;
> +	} cycle_counter;
>  };
>  
>  struct panfrost_mmu {
> diff --git a/drivers/gpu/drm/panfrost/panfrost_gpu.c b/drivers/gpu/drm/panfrost/panfrost_gpu.c
> index 2faa344d89ee..fddbc72bf093 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_gpu.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_gpu.c
> @@ -73,6 +73,8 @@ int panfrost_gpu_soft_reset(struct panfrost_device *pfdev)
>  	gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL);
>  	gpu_write(pfdev, GPU_INT_MASK, GPU_IRQ_MASK_ALL);
>  
> +	atomic_set(&pfdev->cycle_counter.use_count, 0);

I think I'd prefer if the jobs that were in-flight at the time a GPU
hang occurred explicitly release their reference on use_count. So maybe
something like

	/* When we reset the GPU we should have no cycle-counter users
	 * left.
	 */
	if (drm_WARN_ON(cycle_counter.use_count != 0))
		atomic_set(&pfdev->cycle_counter.use_count, 0);

to catch unbalanced get/put situations.

> +
>  	return 0;
>  }
>  
> @@ -321,6 +323,46 @@ static void panfrost_gpu_init_features(struct panfrost_device *pfdev)
>  		 pfdev->features.shader_present, pfdev->features.l2_present);
>  }
>  
> +void panfrost_cycle_counter_get(struct panfrost_device *pfdev)
> +{
> +	if (atomic_inc_not_zero(&pfdev->cycle_counter.use_count))
> +		return;
> +
> +	spin_lock(&pfdev->cycle_counter.lock);
> +	if (atomic_inc_return(&pfdev->cycle_counter.use_count) == 1)
> +		gpu_write(pfdev, GPU_CMD, GPU_CMD_CYCLE_COUNT_START);
> +	spin_unlock(&pfdev->cycle_counter.lock);
> +}
> +
> +void panfrost_cycle_counter_put(struct panfrost_device *pfdev)
> +{
> +	if (atomic_add_unless(&pfdev->cycle_counter.use_count, -1, 1))
> +		return;
> +
> +	spin_lock(&pfdev->cycle_counter.lock);
> +	if (atomic_dec_return(&pfdev->cycle_counter.use_count) == 0)
> +		gpu_write(pfdev, GPU_CMD, GPU_CMD_CYCLE_COUNT_STOP);
> +	spin_unlock(&pfdev->cycle_counter.lock);
> +}
> +
> +void panfrost_cycle_counter_stop(struct panfrost_device *pfdev)
> +{
> +	atomic_set(&pfdev->profile_mode, 0);
> +	gpu_write(pfdev, GPU_CMD, GPU_CMD_CYCLE_COUNT_STOP);

Why do we need to issue a STOP here. Setting profile_mode to false
should be enough to prevent future jobs from enabling the
cycle-counter, and the counter will be naturally disabled when all
in-flight jobs that had profiling enabled are done.

Actually I'm not even sure I understand why this function exists.

> +}
> +
> +unsigned long long panfrost_cycle_counter_read(struct panfrost_device *pfdev)
> +{
> +	u32 hi, lo;
> +
> +	do {
> +		hi = gpu_read(pfdev, GPU_CYCLE_COUNT_HI);
> +		lo = gpu_read(pfdev, GPU_CYCLE_COUNT_LO);
> +	} while (hi != gpu_read(pfdev, GPU_CYCLE_COUNT_HI));
> +
> +	return ((u64)hi << 32) | lo;
> +}
> +
>  void panfrost_gpu_power_on(struct panfrost_device *pfdev)
>  {
>  	int ret;
> @@ -367,6 +409,7 @@ void panfrost_gpu_power_on(struct panfrost_device *pfdev)
>  
>  void panfrost_gpu_power_off(struct panfrost_device *pfdev)
>  {
> +	panfrost_cycle_counter_stop(pfdev);

So, you're setting profile_mode = 0 in the suspend path, but AFAICT,
it's not set back to the module param profile value on resume, which
means it's disabled on suspend and never re-enabled after that.

Besides, I don't really see a reason to change the pfdev->profile_mode
value in this path. If we suspend the device, that means we have no
jobs running, and the use_count refcount should have dropped to zero,
thus disabling cycle counting.

>  	gpu_write(pfdev, TILER_PWROFF_LO, 0);
>  	gpu_write(pfdev, SHADER_PWROFF_LO, 0);
>  	gpu_write(pfdev, L2_PWROFF_LO, 0);
> diff --git a/drivers/gpu/drm/panfrost/panfrost_gpu.h b/drivers/gpu/drm/panfrost/panfrost_gpu.h
> index 468c51e7e46d..4d62e8901c79 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_gpu.h
> +++ b/drivers/gpu/drm/panfrost/panfrost_gpu.h
> @@ -16,6 +16,12 @@ int panfrost_gpu_soft_reset(struct panfrost_device *pfdev);
>  void panfrost_gpu_power_on(struct panfrost_device *pfdev);
>  void panfrost_gpu_power_off(struct panfrost_device *pfdev);
>  
> +void panfrost_stop_cycle_counter(struct panfrost_device *pfdev);
> +void panfrost_cycle_counter_get(struct panfrost_device *pfdev);
> +void panfrost_cycle_counter_stop(struct panfrost_device *pfdev);
> +void panfrost_cycle_counter_put(struct panfrost_device *pfdev);
> +unsigned long long panfrost_cycle_counter_read(struct panfrost_device *pfdev);
> +
>  void panfrost_gpu_amlogic_quirk(struct panfrost_device *pfdev);
>  
>  #endif
> diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c b/drivers/gpu/drm/panfrost/panfrost_job.c
> index 033f5e684707..8b1bf6ac48f8 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_job.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_job.c
> @@ -297,6 +297,11 @@ int panfrost_job_push(struct panfrost_job *job)
>  
>  	kref_get(&job->refcount); /* put by scheduler job completion */
>  
> +	if (atomic_read(&pfdev->profile_mode)) {
> +		panfrost_cycle_counter_get(pfdev);
> +		job->is_profiled = true;
> +	}
> +
>  	drm_sched_entity_push_job(&job->base);
>  
>  	mutex_unlock(&pfdev->sched_lock);
> @@ -351,6 +356,9 @@ static void panfrost_job_free(struct drm_sched_job *sched_job)
>  
>  	drm_sched_job_cleanup(sched_job);
>  
> +	if (job->is_profiled)
> +		panfrost_cycle_counter_put(job->pfdev);
> +
>  	panfrost_job_put(job);
>  }
>  
> @@ -842,6 +850,8 @@ int panfrost_job_init(struct panfrost_device *pfdev)
>  		}
>  	}
>  
> +	spin_lock_init(&pfdev->cycle_counter.lock);
> +
>  	panfrost_job_enable_interrupts(pfdev);
>  
>  	return 0;
> diff --git a/drivers/gpu/drm/panfrost/panfrost_job.h b/drivers/gpu/drm/panfrost/panfrost_job.h
> index 8becc1ba0eb9..2aa0add35459 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_job.h
> +++ b/drivers/gpu/drm/panfrost/panfrost_job.h
> @@ -32,6 +32,7 @@ struct panfrost_job {
>  
>  	/* Fence to be signaled by drm-sched once its done with the job */
>  	struct dma_fence *render_done_fence;
> +	bool is_profiled;
>  };
>  
>  int panfrost_job_init(struct panfrost_device *pfdev);