Re: [PATCH] drm/sched: Extend the documentation.

[Daniel Vetter]

On Thu, Apr 5, 2018 at 3:44 PM, Alex Deucher  wrote:
> On Thu, Apr 5, 2018 at 9:41 AM, Nayan Deshmukh
>  wrote:
>> On Thu, Apr 5, 2018 at 6:59 PM, Daniel Vetter  wrote:
>>> On Thu, Apr 5, 2018 at 3:27 PM, Alex Deucher  wrote:
 On Thu, Apr 5, 2018 at 2:16 AM, Daniel Vetter  wrote:
> On Thu, Apr 5, 2018 at 12:32 AM, Eric Anholt  wrote:
>> These comments answer all the questions I had for myself when
>> implementing a driver using the GPU scheduler.
>>
>> Signed-off-by: Eric Anholt 
>
> Pulling all these comments into the generated kerneldoc would be
> awesome, maybe as a new "GPU Scheduler" chapter at the end of
> drm-mm.rst? Would mean a bit of busywork to convert the existing raw
> comments into proper kerneldoc. Also has the benefit that 0day will
> complain when you forget to update the comment when editing the
> function prototype - kerneldoc which isn't included anywhere in .rst
> won't be checked automatically.

 I was actually planning to do this myself, but Nayan wanted to do this
 a prep work for his proposed GSoC project so I was going to see how
 far he got first.
>>
>> It is still on my TODO list. Just got a bit busy with my coursework. I
>> will try to look at it during the weekend.
>
> No worries.  Take your time.
>
>>>
>>> Awesome. I'm also happy to help out with any kerneldoc questions and
>>> best practices. Technically ofc no clue about the scheduler :-)
>>>
>> I was thinking of adding a different rst for scheduler altogther. Will
>> it be better to add it in drm-mm.rst itself?
>
> I had been planning to add a separate file too since it's a separate
> entity.  Do whatever you think works best.

My recommendation is that to put the gist of the docs all into
source-code comments. That way there's a much higher chance to spot
them. In the docs you'll then only have the chapter structure (and not
sure the scheduler needs more than 1 chapter).

drm-mm.rst is a bit misnamed, since it contains all the stuff for
handling rendering: MM, fences, dma-buf/prime, drm_syncobjs. I think
scheduler fits very well in that topic range. We can ofc rename it to
drm-rendering.rst or similar, if the -mm.rst is misleading (plus
adjust the title).
-Daniel

>
> Alex
>
>>
>>> Cheers, Daniel
 Alex

> -Daniel
>
>> ---
>>  include/drm/gpu_scheduler.h | 46 +
>>  1 file changed, 42 insertions(+), 4 deletions(-)
>>
>> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
>> index dfd54fb94e10..c053a32341bf 100644
>> --- a/include/drm/gpu_scheduler.h
>> +++ b/include/drm/gpu_scheduler.h
>> @@ -43,10 +43,12 @@ enum drm_sched_priority {
>>  };
>>
>>  /**
>> - * A scheduler entity is a wrapper around a job queue or a group
>> - * of other entities. Entities take turns emitting jobs from their
>> - * job queues to corresponding hardware ring based on scheduling
>> - * policy.
>> + * drm_sched_entity - A wrapper around a job queue (typically attached
>> + * to the DRM file_priv).
>> + *
>> + * Entities will emit jobs in order to their corresponding hardware
>> + * ring, and the scheduler will alternate between entities based on
>> + * scheduling policy.
>>  */
>>  struct drm_sched_entity {
>> struct list_headlist;
>> @@ -78,7 +80,18 @@ struct drm_sched_rq {
>>
>>  struct drm_sched_fence {
>> struct dma_fencescheduled;
>> +
>> +   /* This fence is what will be signaled by the scheduler when
>> +* the job is completed.
>> +*
>> +* When setting up an out fence for the job, you should use
>> +* this, since it's available immediately upon
>> +* drm_sched_job_init(), and the fence returned by the driver
>> +* from run_job() won't be created until the dependencies have
>> +* resolved.
>> +*/
>> struct dma_fencefinished;
>> +
>> struct dma_fence_cb cb;
>> struct dma_fence*parent;
>> struct drm_gpu_scheduler*sched;
>> @@ -88,6 +101,13 @@ struct drm_sched_fence {
>>
>>  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
>>
>> +/**
>> + * drm_sched_job - A job to be run by an entity.
>> + *
>> + * A job is created by the driver using drm_sched_job_init(), and
>> + * should call drm_sched_entity_push_job() once it wants the scheduler
>> + * to schedule the job.
>> + */
>>  struct drm_sched_job {
>> struct spsc_nodequeue_node;
>> struct drm_gpu_scheduler*sched;
>> @@ -112,10 +132,28 

Re: [PATCH] drm/sched: Extend the documentation.

[Daniel Vetter]

On Thu, Apr 5, 2018 at 3:44 PM, Alex Deucher  wrote:
> On Thu, Apr 5, 2018 at 9:41 AM, Nayan Deshmukh
>  wrote:
>> On Thu, Apr 5, 2018 at 6:59 PM, Daniel Vetter  wrote:
>>> On Thu, Apr 5, 2018 at 3:27 PM, Alex Deucher  wrote:
 On Thu, Apr 5, 2018 at 2:16 AM, Daniel Vetter  wrote:
> On Thu, Apr 5, 2018 at 12:32 AM, Eric Anholt  wrote:
>> These comments answer all the questions I had for myself when
>> implementing a driver using the GPU scheduler.
>>
>> Signed-off-by: Eric Anholt 
>
> Pulling all these comments into the generated kerneldoc would be
> awesome, maybe as a new "GPU Scheduler" chapter at the end of
> drm-mm.rst? Would mean a bit of busywork to convert the existing raw
> comments into proper kerneldoc. Also has the benefit that 0day will
> complain when you forget to update the comment when editing the
> function prototype - kerneldoc which isn't included anywhere in .rst
> won't be checked automatically.

 I was actually planning to do this myself, but Nayan wanted to do this
 a prep work for his proposed GSoC project so I was going to see how
 far he got first.
>>
>> It is still on my TODO list. Just got a bit busy with my coursework. I
>> will try to look at it during the weekend.
>
> No worries.  Take your time.
>
>>>
>>> Awesome. I'm also happy to help out with any kerneldoc questions and
>>> best practices. Technically ofc no clue about the scheduler :-)
>>>
>> I was thinking of adding a different rst for scheduler altogther. Will
>> it be better to add it in drm-mm.rst itself?
>
> I had been planning to add a separate file too since it's a separate
> entity.  Do whatever you think works best.

My recommendation is that to put the gist of the docs all into
source-code comments. That way there's a much higher chance to spot
them. In the docs you'll then only have the chapter structure (and not
sure the scheduler needs more than 1 chapter).

drm-mm.rst is a bit misnamed, since it contains all the stuff for
handling rendering: MM, fences, dma-buf/prime, drm_syncobjs. I think
scheduler fits very well in that topic range. We can ofc rename it to
drm-rendering.rst or similar, if the -mm.rst is misleading (plus
adjust the title).
-Daniel

>
> Alex
>
>>
>>> Cheers, Daniel
 Alex

> -Daniel
>
>> ---
>>  include/drm/gpu_scheduler.h | 46 +
>>  1 file changed, 42 insertions(+), 4 deletions(-)
>>
>> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
>> index dfd54fb94e10..c053a32341bf 100644
>> --- a/include/drm/gpu_scheduler.h
>> +++ b/include/drm/gpu_scheduler.h
>> @@ -43,10 +43,12 @@ enum drm_sched_priority {
>>  };
>>
>>  /**
>> - * A scheduler entity is a wrapper around a job queue or a group
>> - * of other entities. Entities take turns emitting jobs from their
>> - * job queues to corresponding hardware ring based on scheduling
>> - * policy.
>> + * drm_sched_entity - A wrapper around a job queue (typically attached
>> + * to the DRM file_priv).
>> + *
>> + * Entities will emit jobs in order to their corresponding hardware
>> + * ring, and the scheduler will alternate between entities based on
>> + * scheduling policy.
>>  */
>>  struct drm_sched_entity {
>> struct list_headlist;
>> @@ -78,7 +80,18 @@ struct drm_sched_rq {
>>
>>  struct drm_sched_fence {
>> struct dma_fencescheduled;
>> +
>> +   /* This fence is what will be signaled by the scheduler when
>> +* the job is completed.
>> +*
>> +* When setting up an out fence for the job, you should use
>> +* this, since it's available immediately upon
>> +* drm_sched_job_init(), and the fence returned by the driver
>> +* from run_job() won't be created until the dependencies have
>> +* resolved.
>> +*/
>> struct dma_fencefinished;
>> +
>> struct dma_fence_cb cb;
>> struct dma_fence*parent;
>> struct drm_gpu_scheduler*sched;
>> @@ -88,6 +101,13 @@ struct drm_sched_fence {
>>
>>  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
>>
>> +/**
>> + * drm_sched_job - A job to be run by an entity.
>> + *
>> + * A job is created by the driver using drm_sched_job_init(), and
>> + * should call drm_sched_entity_push_job() once it wants the scheduler
>> + * to schedule the job.
>> + */
>>  struct drm_sched_job {
>> struct spsc_nodequeue_node;
>> struct drm_gpu_scheduler*sched;
>> @@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
>> drm_sched_job *s_job,
>>   * these functions should be implemented in 

Re: [PATCH] drm/sched: Extend the documentation.

[Alex Deucher]

On Thu, Apr 5, 2018 at 9:41 AM, Nayan Deshmukh
 wrote:
> On Thu, Apr 5, 2018 at 6:59 PM, Daniel Vetter  wrote:
>> On Thu, Apr 5, 2018 at 3:27 PM, Alex Deucher  wrote:
>>> On Thu, Apr 5, 2018 at 2:16 AM, Daniel Vetter  wrote:
 On Thu, Apr 5, 2018 at 12:32 AM, Eric Anholt  wrote:
> These comments answer all the questions I had for myself when
> implementing a driver using the GPU scheduler.
>
> Signed-off-by: Eric Anholt 

 Pulling all these comments into the generated kerneldoc would be
 awesome, maybe as a new "GPU Scheduler" chapter at the end of
 drm-mm.rst? Would mean a bit of busywork to convert the existing raw
 comments into proper kerneldoc. Also has the benefit that 0day will
 complain when you forget to update the comment when editing the
 function prototype - kerneldoc which isn't included anywhere in .rst
 won't be checked automatically.
>>>
>>> I was actually planning to do this myself, but Nayan wanted to do this
>>> a prep work for his proposed GSoC project so I was going to see how
>>> far he got first.
>
> It is still on my TODO list. Just got a bit busy with my coursework. I
> will try to look at it during the weekend.

No worries.  Take your time.

>>
>> Awesome. I'm also happy to help out with any kerneldoc questions and
>> best practices. Technically ofc no clue about the scheduler :-)
>>
> I was thinking of adding a different rst for scheduler altogther. Will
> it be better to add it in drm-mm.rst itself?

I had been planning to add a separate file too since it's a separate
entity.  Do whatever you think works best.

Alex

>
>> Cheers, Daniel
>>> Alex
>>>
 -Daniel

> ---
>  include/drm/gpu_scheduler.h | 46 +
>  1 file changed, 42 insertions(+), 4 deletions(-)
>
> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
> index dfd54fb94e10..c053a32341bf 100644
> --- a/include/drm/gpu_scheduler.h
> +++ b/include/drm/gpu_scheduler.h
> @@ -43,10 +43,12 @@ enum drm_sched_priority {
>  };
>
>  /**
> - * A scheduler entity is a wrapper around a job queue or a group
> - * of other entities. Entities take turns emitting jobs from their
> - * job queues to corresponding hardware ring based on scheduling
> - * policy.
> + * drm_sched_entity - A wrapper around a job queue (typically attached
> + * to the DRM file_priv).
> + *
> + * Entities will emit jobs in order to their corresponding hardware
> + * ring, and the scheduler will alternate between entities based on
> + * scheduling policy.
>  */
>  struct drm_sched_entity {
> struct list_headlist;
> @@ -78,7 +80,18 @@ struct drm_sched_rq {
>
>  struct drm_sched_fence {
> struct dma_fencescheduled;
> +
> +   /* This fence is what will be signaled by the scheduler when
> +* the job is completed.
> +*
> +* When setting up an out fence for the job, you should use
> +* this, since it's available immediately upon
> +* drm_sched_job_init(), and the fence returned by the driver
> +* from run_job() won't be created until the dependencies have
> +* resolved.
> +*/
> struct dma_fencefinished;
> +
> struct dma_fence_cb cb;
> struct dma_fence*parent;
> struct drm_gpu_scheduler*sched;
> @@ -88,6 +101,13 @@ struct drm_sched_fence {
>
>  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
>
> +/**
> + * drm_sched_job - A job to be run by an entity.
> + *
> + * A job is created by the driver using drm_sched_job_init(), and
> + * should call drm_sched_entity_push_job() once it wants the scheduler
> + * to schedule the job.
> + */
>  struct drm_sched_job {
> struct spsc_nodequeue_node;
> struct drm_gpu_scheduler*sched;
> @@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
> drm_sched_job *s_job,
>   * these functions should be implemented in driver side
>  */
>  struct drm_sched_backend_ops {
> +   /* Called when the scheduler is considering scheduling this
> +* job next, to get another struct dma_fence for this job to
> +* block on.  Once it returns NULL, run_job() may be called.
> +*/
> struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
> struct drm_sched_entity 
> *s_entity);
> +
> +   /* Called to execute the job once all of the dependencies have
> +* been resolved.  This may be 

Re: [PATCH] drm/sched: Extend the documentation.

[Alex Deucher]

On Thu, Apr 5, 2018 at 9:41 AM, Nayan Deshmukh
 wrote:
> On Thu, Apr 5, 2018 at 6:59 PM, Daniel Vetter  wrote:
>> On Thu, Apr 5, 2018 at 3:27 PM, Alex Deucher  wrote:
>>> On Thu, Apr 5, 2018 at 2:16 AM, Daniel Vetter  wrote:
 On Thu, Apr 5, 2018 at 12:32 AM, Eric Anholt  wrote:
> These comments answer all the questions I had for myself when
> implementing a driver using the GPU scheduler.
>
> Signed-off-by: Eric Anholt 

 Pulling all these comments into the generated kerneldoc would be
 awesome, maybe as a new "GPU Scheduler" chapter at the end of
 drm-mm.rst? Would mean a bit of busywork to convert the existing raw
 comments into proper kerneldoc. Also has the benefit that 0day will
 complain when you forget to update the comment when editing the
 function prototype - kerneldoc which isn't included anywhere in .rst
 won't be checked automatically.
>>>
>>> I was actually planning to do this myself, but Nayan wanted to do this
>>> a prep work for his proposed GSoC project so I was going to see how
>>> far he got first.
>
> It is still on my TODO list. Just got a bit busy with my coursework. I
> will try to look at it during the weekend.

No worries.  Take your time.

>>
>> Awesome. I'm also happy to help out with any kerneldoc questions and
>> best practices. Technically ofc no clue about the scheduler :-)
>>
> I was thinking of adding a different rst for scheduler altogther. Will
> it be better to add it in drm-mm.rst itself?

I had been planning to add a separate file too since it's a separate
entity.  Do whatever you think works best.

Alex

>
>> Cheers, Daniel
>>> Alex
>>>
 -Daniel

> ---
>  include/drm/gpu_scheduler.h | 46 +
>  1 file changed, 42 insertions(+), 4 deletions(-)
>
> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
> index dfd54fb94e10..c053a32341bf 100644
> --- a/include/drm/gpu_scheduler.h
> +++ b/include/drm/gpu_scheduler.h
> @@ -43,10 +43,12 @@ enum drm_sched_priority {
>  };
>
>  /**
> - * A scheduler entity is a wrapper around a job queue or a group
> - * of other entities. Entities take turns emitting jobs from their
> - * job queues to corresponding hardware ring based on scheduling
> - * policy.
> + * drm_sched_entity - A wrapper around a job queue (typically attached
> + * to the DRM file_priv).
> + *
> + * Entities will emit jobs in order to their corresponding hardware
> + * ring, and the scheduler will alternate between entities based on
> + * scheduling policy.
>  */
>  struct drm_sched_entity {
> struct list_headlist;
> @@ -78,7 +80,18 @@ struct drm_sched_rq {
>
>  struct drm_sched_fence {
> struct dma_fencescheduled;
> +
> +   /* This fence is what will be signaled by the scheduler when
> +* the job is completed.
> +*
> +* When setting up an out fence for the job, you should use
> +* this, since it's available immediately upon
> +* drm_sched_job_init(), and the fence returned by the driver
> +* from run_job() won't be created until the dependencies have
> +* resolved.
> +*/
> struct dma_fencefinished;
> +
> struct dma_fence_cb cb;
> struct dma_fence*parent;
> struct drm_gpu_scheduler*sched;
> @@ -88,6 +101,13 @@ struct drm_sched_fence {
>
>  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
>
> +/**
> + * drm_sched_job - A job to be run by an entity.
> + *
> + * A job is created by the driver using drm_sched_job_init(), and
> + * should call drm_sched_entity_push_job() once it wants the scheduler
> + * to schedule the job.
> + */
>  struct drm_sched_job {
> struct spsc_nodequeue_node;
> struct drm_gpu_scheduler*sched;
> @@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
> drm_sched_job *s_job,
>   * these functions should be implemented in driver side
>  */
>  struct drm_sched_backend_ops {
> +   /* Called when the scheduler is considering scheduling this
> +* job next, to get another struct dma_fence for this job to
> +* block on.  Once it returns NULL, run_job() may be called.
> +*/
> struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
> struct drm_sched_entity 
> *s_entity);
> +
> +   /* Called to execute the job once all of the dependencies have
> +* been resolved.  This may be called multiple times, if
> +* timedout_job() has happened and drm_sched_job_recovery()
> +* 

Re: [PATCH] drm/sched: Extend the documentation.

[Nayan Deshmukh]

On Thu, Apr 5, 2018 at 6:59 PM, Daniel Vetter  wrote:
> On Thu, Apr 5, 2018 at 3:27 PM, Alex Deucher  wrote:
>> On Thu, Apr 5, 2018 at 2:16 AM, Daniel Vetter  wrote:
>>> On Thu, Apr 5, 2018 at 12:32 AM, Eric Anholt  wrote:
 These comments answer all the questions I had for myself when
 implementing a driver using the GPU scheduler.

 Signed-off-by: Eric Anholt 
>>>
>>> Pulling all these comments into the generated kerneldoc would be
>>> awesome, maybe as a new "GPU Scheduler" chapter at the end of
>>> drm-mm.rst? Would mean a bit of busywork to convert the existing raw
>>> comments into proper kerneldoc. Also has the benefit that 0day will
>>> complain when you forget to update the comment when editing the
>>> function prototype - kerneldoc which isn't included anywhere in .rst
>>> won't be checked automatically.
>>
>> I was actually planning to do this myself, but Nayan wanted to do this
>> a prep work for his proposed GSoC project so I was going to see how
>> far he got first.

It is still on my TODO list. Just got a bit busy with my coursework. I
will try to look at it during the weekend.
>
> Awesome. I'm also happy to help out with any kerneldoc questions and
> best practices. Technically ofc no clue about the scheduler :-)
>
I was thinking of adding a different rst for scheduler altogther. Will
it be better to add it in drm-mm.rst itself?

> Cheers, Daniel
>> Alex
>>
>>> -Daniel
>>>
 ---
  include/drm/gpu_scheduler.h | 46 +
  1 file changed, 42 insertions(+), 4 deletions(-)

 diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
 index dfd54fb94e10..c053a32341bf 100644
 --- a/include/drm/gpu_scheduler.h
 +++ b/include/drm/gpu_scheduler.h
 @@ -43,10 +43,12 @@ enum drm_sched_priority {
  };

  /**
 - * A scheduler entity is a wrapper around a job queue or a group
 - * of other entities. Entities take turns emitting jobs from their
 - * job queues to corresponding hardware ring based on scheduling
 - * policy.
 + * drm_sched_entity - A wrapper around a job queue (typically attached
 + * to the DRM file_priv).
 + *
 + * Entities will emit jobs in order to their corresponding hardware
 + * ring, and the scheduler will alternate between entities based on
 + * scheduling policy.
  */
  struct drm_sched_entity {
 struct list_headlist;
 @@ -78,7 +80,18 @@ struct drm_sched_rq {

  struct drm_sched_fence {
 struct dma_fencescheduled;
 +
 +   /* This fence is what will be signaled by the scheduler when
 +* the job is completed.
 +*
 +* When setting up an out fence for the job, you should use
 +* this, since it's available immediately upon
 +* drm_sched_job_init(), and the fence returned by the driver
 +* from run_job() won't be created until the dependencies have
 +* resolved.
 +*/
 struct dma_fencefinished;
 +
 struct dma_fence_cb cb;
 struct dma_fence*parent;
 struct drm_gpu_scheduler*sched;
 @@ -88,6 +101,13 @@ struct drm_sched_fence {

  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);

 +/**
 + * drm_sched_job - A job to be run by an entity.
 + *
 + * A job is created by the driver using drm_sched_job_init(), and
 + * should call drm_sched_entity_push_job() once it wants the scheduler
 + * to schedule the job.
 + */
  struct drm_sched_job {
 struct spsc_nodequeue_node;
 struct drm_gpu_scheduler*sched;
 @@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
 drm_sched_job *s_job,
   * these functions should be implemented in driver side
  */
  struct drm_sched_backend_ops {
 +   /* Called when the scheduler is considering scheduling this
 +* job next, to get another struct dma_fence for this job to
 +* block on.  Once it returns NULL, run_job() may be called.
 +*/
 struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
 struct drm_sched_entity *s_entity);
 +
 +   /* Called to execute the job once all of the dependencies have
 +* been resolved.  This may be called multiple times, if
 +* timedout_job() has happened and drm_sched_job_recovery()
 +* decides to try it again.
 +*/
 struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
 +
 +   /* Called when a job has taken too long to execute, to trigger
 +* GPU recovery.
 +  

Re: [PATCH] drm/sched: Extend the documentation.

[Nayan Deshmukh]

On Thu, Apr 5, 2018 at 6:59 PM, Daniel Vetter  wrote:
> On Thu, Apr 5, 2018 at 3:27 PM, Alex Deucher  wrote:
>> On Thu, Apr 5, 2018 at 2:16 AM, Daniel Vetter  wrote:
>>> On Thu, Apr 5, 2018 at 12:32 AM, Eric Anholt  wrote:
 These comments answer all the questions I had for myself when
 implementing a driver using the GPU scheduler.

 Signed-off-by: Eric Anholt 
>>>
>>> Pulling all these comments into the generated kerneldoc would be
>>> awesome, maybe as a new "GPU Scheduler" chapter at the end of
>>> drm-mm.rst? Would mean a bit of busywork to convert the existing raw
>>> comments into proper kerneldoc. Also has the benefit that 0day will
>>> complain when you forget to update the comment when editing the
>>> function prototype - kerneldoc which isn't included anywhere in .rst
>>> won't be checked automatically.
>>
>> I was actually planning to do this myself, but Nayan wanted to do this
>> a prep work for his proposed GSoC project so I was going to see how
>> far he got first.

It is still on my TODO list. Just got a bit busy with my coursework. I
will try to look at it during the weekend.
>
> Awesome. I'm also happy to help out with any kerneldoc questions and
> best practices. Technically ofc no clue about the scheduler :-)
>
I was thinking of adding a different rst for scheduler altogther. Will
it be better to add it in drm-mm.rst itself?

> Cheers, Daniel
>> Alex
>>
>>> -Daniel
>>>
 ---
  include/drm/gpu_scheduler.h | 46 +
  1 file changed, 42 insertions(+), 4 deletions(-)

 diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
 index dfd54fb94e10..c053a32341bf 100644
 --- a/include/drm/gpu_scheduler.h
 +++ b/include/drm/gpu_scheduler.h
 @@ -43,10 +43,12 @@ enum drm_sched_priority {
  };

  /**
 - * A scheduler entity is a wrapper around a job queue or a group
 - * of other entities. Entities take turns emitting jobs from their
 - * job queues to corresponding hardware ring based on scheduling
 - * policy.
 + * drm_sched_entity - A wrapper around a job queue (typically attached
 + * to the DRM file_priv).
 + *
 + * Entities will emit jobs in order to their corresponding hardware
 + * ring, and the scheduler will alternate between entities based on
 + * scheduling policy.
  */
  struct drm_sched_entity {
 struct list_headlist;
 @@ -78,7 +80,18 @@ struct drm_sched_rq {

  struct drm_sched_fence {
 struct dma_fencescheduled;
 +
 +   /* This fence is what will be signaled by the scheduler when
 +* the job is completed.
 +*
 +* When setting up an out fence for the job, you should use
 +* this, since it's available immediately upon
 +* drm_sched_job_init(), and the fence returned by the driver
 +* from run_job() won't be created until the dependencies have
 +* resolved.
 +*/
 struct dma_fencefinished;
 +
 struct dma_fence_cb cb;
 struct dma_fence*parent;
 struct drm_gpu_scheduler*sched;
 @@ -88,6 +101,13 @@ struct drm_sched_fence {

  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);

 +/**
 + * drm_sched_job - A job to be run by an entity.
 + *
 + * A job is created by the driver using drm_sched_job_init(), and
 + * should call drm_sched_entity_push_job() once it wants the scheduler
 + * to schedule the job.
 + */
  struct drm_sched_job {
 struct spsc_nodequeue_node;
 struct drm_gpu_scheduler*sched;
 @@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
 drm_sched_job *s_job,
   * these functions should be implemented in driver side
  */
  struct drm_sched_backend_ops {
 +   /* Called when the scheduler is considering scheduling this
 +* job next, to get another struct dma_fence for this job to
 +* block on.  Once it returns NULL, run_job() may be called.
 +*/
 struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
 struct drm_sched_entity *s_entity);
 +
 +   /* Called to execute the job once all of the dependencies have
 +* been resolved.  This may be called multiple times, if
 +* timedout_job() has happened and drm_sched_job_recovery()
 +* decides to try it again.
 +*/
 struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
 +
 +   /* Called when a job has taken too long to execute, to trigger
 +* GPU recovery.
 +*/
 void (*timedout_job)(struct drm_sched_job *sched_job);
 +
 + 

Re: [PATCH] drm/sched: Extend the documentation.

[Daniel Vetter]

On Thu, Apr 5, 2018 at 3:27 PM, Alex Deucher  wrote:
> On Thu, Apr 5, 2018 at 2:16 AM, Daniel Vetter  wrote:
>> On Thu, Apr 5, 2018 at 12:32 AM, Eric Anholt  wrote:
>>> These comments answer all the questions I had for myself when
>>> implementing a driver using the GPU scheduler.
>>>
>>> Signed-off-by: Eric Anholt 
>>
>> Pulling all these comments into the generated kerneldoc would be
>> awesome, maybe as a new "GPU Scheduler" chapter at the end of
>> drm-mm.rst? Would mean a bit of busywork to convert the existing raw
>> comments into proper kerneldoc. Also has the benefit that 0day will
>> complain when you forget to update the comment when editing the
>> function prototype - kerneldoc which isn't included anywhere in .rst
>> won't be checked automatically.
>
> I was actually planning to do this myself, but Nayan wanted to do this
> a prep work for his proposed GSoC project so I was going to see how
> far he got first.

Awesome. I'm also happy to help out with any kerneldoc questions and
best practices. Technically ofc no clue about the scheduler :-)

Cheers, Daniel
> Alex
>
>> -Daniel
>>
>>> ---
>>>  include/drm/gpu_scheduler.h | 46 +
>>>  1 file changed, 42 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
>>> index dfd54fb94e10..c053a32341bf 100644
>>> --- a/include/drm/gpu_scheduler.h
>>> +++ b/include/drm/gpu_scheduler.h
>>> @@ -43,10 +43,12 @@ enum drm_sched_priority {
>>>  };
>>>
>>>  /**
>>> - * A scheduler entity is a wrapper around a job queue or a group
>>> - * of other entities. Entities take turns emitting jobs from their
>>> - * job queues to corresponding hardware ring based on scheduling
>>> - * policy.
>>> + * drm_sched_entity - A wrapper around a job queue (typically attached
>>> + * to the DRM file_priv).
>>> + *
>>> + * Entities will emit jobs in order to their corresponding hardware
>>> + * ring, and the scheduler will alternate between entities based on
>>> + * scheduling policy.
>>>  */
>>>  struct drm_sched_entity {
>>> struct list_headlist;
>>> @@ -78,7 +80,18 @@ struct drm_sched_rq {
>>>
>>>  struct drm_sched_fence {
>>> struct dma_fencescheduled;
>>> +
>>> +   /* This fence is what will be signaled by the scheduler when
>>> +* the job is completed.
>>> +*
>>> +* When setting up an out fence for the job, you should use
>>> +* this, since it's available immediately upon
>>> +* drm_sched_job_init(), and the fence returned by the driver
>>> +* from run_job() won't be created until the dependencies have
>>> +* resolved.
>>> +*/
>>> struct dma_fencefinished;
>>> +
>>> struct dma_fence_cb cb;
>>> struct dma_fence*parent;
>>> struct drm_gpu_scheduler*sched;
>>> @@ -88,6 +101,13 @@ struct drm_sched_fence {
>>>
>>>  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
>>>
>>> +/**
>>> + * drm_sched_job - A job to be run by an entity.
>>> + *
>>> + * A job is created by the driver using drm_sched_job_init(), and
>>> + * should call drm_sched_entity_push_job() once it wants the scheduler
>>> + * to schedule the job.
>>> + */
>>>  struct drm_sched_job {
>>> struct spsc_nodequeue_node;
>>> struct drm_gpu_scheduler*sched;
>>> @@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
>>> drm_sched_job *s_job,
>>>   * these functions should be implemented in driver side
>>>  */
>>>  struct drm_sched_backend_ops {
>>> +   /* Called when the scheduler is considering scheduling this
>>> +* job next, to get another struct dma_fence for this job to
>>> +* block on.  Once it returns NULL, run_job() may be called.
>>> +*/
>>> struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
>>> struct drm_sched_entity *s_entity);
>>> +
>>> +   /* Called to execute the job once all of the dependencies have
>>> +* been resolved.  This may be called multiple times, if
>>> +* timedout_job() has happened and drm_sched_job_recovery()
>>> +* decides to try it again.
>>> +*/
>>> struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
>>> +
>>> +   /* Called when a job has taken too long to execute, to trigger
>>> +* GPU recovery.
>>> +*/
>>> void (*timedout_job)(struct drm_sched_job *sched_job);
>>> +
>>> +   /* Called once the job's finished fence has been signaled and
>>> +* it's time to clean it up.
>>> +*/
>>> void (*free_job)(struct drm_sched_job *sched_job);
>>>  };
>>>
>>> --
>>> 2.17.0
>>>
>>> ___
>>> dri-devel mailing list
>>> 

Re: [PATCH] drm/sched: Extend the documentation.

[Daniel Vetter]

On Thu, Apr 5, 2018 at 3:27 PM, Alex Deucher  wrote:
> On Thu, Apr 5, 2018 at 2:16 AM, Daniel Vetter  wrote:
>> On Thu, Apr 5, 2018 at 12:32 AM, Eric Anholt  wrote:
>>> These comments answer all the questions I had for myself when
>>> implementing a driver using the GPU scheduler.
>>>
>>> Signed-off-by: Eric Anholt 
>>
>> Pulling all these comments into the generated kerneldoc would be
>> awesome, maybe as a new "GPU Scheduler" chapter at the end of
>> drm-mm.rst? Would mean a bit of busywork to convert the existing raw
>> comments into proper kerneldoc. Also has the benefit that 0day will
>> complain when you forget to update the comment when editing the
>> function prototype - kerneldoc which isn't included anywhere in .rst
>> won't be checked automatically.
>
> I was actually planning to do this myself, but Nayan wanted to do this
> a prep work for his proposed GSoC project so I was going to see how
> far he got first.

Awesome. I'm also happy to help out with any kerneldoc questions and
best practices. Technically ofc no clue about the scheduler :-)

Cheers, Daniel
> Alex
>
>> -Daniel
>>
>>> ---
>>>  include/drm/gpu_scheduler.h | 46 +
>>>  1 file changed, 42 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
>>> index dfd54fb94e10..c053a32341bf 100644
>>> --- a/include/drm/gpu_scheduler.h
>>> +++ b/include/drm/gpu_scheduler.h
>>> @@ -43,10 +43,12 @@ enum drm_sched_priority {
>>>  };
>>>
>>>  /**
>>> - * A scheduler entity is a wrapper around a job queue or a group
>>> - * of other entities. Entities take turns emitting jobs from their
>>> - * job queues to corresponding hardware ring based on scheduling
>>> - * policy.
>>> + * drm_sched_entity - A wrapper around a job queue (typically attached
>>> + * to the DRM file_priv).
>>> + *
>>> + * Entities will emit jobs in order to their corresponding hardware
>>> + * ring, and the scheduler will alternate between entities based on
>>> + * scheduling policy.
>>>  */
>>>  struct drm_sched_entity {
>>> struct list_headlist;
>>> @@ -78,7 +80,18 @@ struct drm_sched_rq {
>>>
>>>  struct drm_sched_fence {
>>> struct dma_fencescheduled;
>>> +
>>> +   /* This fence is what will be signaled by the scheduler when
>>> +* the job is completed.
>>> +*
>>> +* When setting up an out fence for the job, you should use
>>> +* this, since it's available immediately upon
>>> +* drm_sched_job_init(), and the fence returned by the driver
>>> +* from run_job() won't be created until the dependencies have
>>> +* resolved.
>>> +*/
>>> struct dma_fencefinished;
>>> +
>>> struct dma_fence_cb cb;
>>> struct dma_fence*parent;
>>> struct drm_gpu_scheduler*sched;
>>> @@ -88,6 +101,13 @@ struct drm_sched_fence {
>>>
>>>  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
>>>
>>> +/**
>>> + * drm_sched_job - A job to be run by an entity.
>>> + *
>>> + * A job is created by the driver using drm_sched_job_init(), and
>>> + * should call drm_sched_entity_push_job() once it wants the scheduler
>>> + * to schedule the job.
>>> + */
>>>  struct drm_sched_job {
>>> struct spsc_nodequeue_node;
>>> struct drm_gpu_scheduler*sched;
>>> @@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
>>> drm_sched_job *s_job,
>>>   * these functions should be implemented in driver side
>>>  */
>>>  struct drm_sched_backend_ops {
>>> +   /* Called when the scheduler is considering scheduling this
>>> +* job next, to get another struct dma_fence for this job to
>>> +* block on.  Once it returns NULL, run_job() may be called.
>>> +*/
>>> struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
>>> struct drm_sched_entity *s_entity);
>>> +
>>> +   /* Called to execute the job once all of the dependencies have
>>> +* been resolved.  This may be called multiple times, if
>>> +* timedout_job() has happened and drm_sched_job_recovery()
>>> +* decides to try it again.
>>> +*/
>>> struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
>>> +
>>> +   /* Called when a job has taken too long to execute, to trigger
>>> +* GPU recovery.
>>> +*/
>>> void (*timedout_job)(struct drm_sched_job *sched_job);
>>> +
>>> +   /* Called once the job's finished fence has been signaled and
>>> +* it's time to clean it up.
>>> +*/
>>> void (*free_job)(struct drm_sched_job *sched_job);
>>>  };
>>>
>>> --
>>> 2.17.0
>>>
>>> ___
>>> dri-devel mailing list
>>> dri-de...@lists.freedesktop.org
>>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>>

Re: [PATCH] drm/sched: Extend the documentation.

[Alex Deucher]

On Thu, Apr 5, 2018 at 2:16 AM, Daniel Vetter  wrote:
> On Thu, Apr 5, 2018 at 12:32 AM, Eric Anholt  wrote:
>> These comments answer all the questions I had for myself when
>> implementing a driver using the GPU scheduler.
>>
>> Signed-off-by: Eric Anholt 
>
> Pulling all these comments into the generated kerneldoc would be
> awesome, maybe as a new "GPU Scheduler" chapter at the end of
> drm-mm.rst? Would mean a bit of busywork to convert the existing raw
> comments into proper kerneldoc. Also has the benefit that 0day will
> complain when you forget to update the comment when editing the
> function prototype - kerneldoc which isn't included anywhere in .rst
> won't be checked automatically.

I was actually planning to do this myself, but Nayan wanted to do this
a prep work for his proposed GSoC project so I was going to see how
far he got first.

Alex

> -Daniel
>
>> ---
>>  include/drm/gpu_scheduler.h | 46 +
>>  1 file changed, 42 insertions(+), 4 deletions(-)
>>
>> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
>> index dfd54fb94e10..c053a32341bf 100644
>> --- a/include/drm/gpu_scheduler.h
>> +++ b/include/drm/gpu_scheduler.h
>> @@ -43,10 +43,12 @@ enum drm_sched_priority {
>>  };
>>
>>  /**
>> - * A scheduler entity is a wrapper around a job queue or a group
>> - * of other entities. Entities take turns emitting jobs from their
>> - * job queues to corresponding hardware ring based on scheduling
>> - * policy.
>> + * drm_sched_entity - A wrapper around a job queue (typically attached
>> + * to the DRM file_priv).
>> + *
>> + * Entities will emit jobs in order to their corresponding hardware
>> + * ring, and the scheduler will alternate between entities based on
>> + * scheduling policy.
>>  */
>>  struct drm_sched_entity {
>> struct list_headlist;
>> @@ -78,7 +80,18 @@ struct drm_sched_rq {
>>
>>  struct drm_sched_fence {
>> struct dma_fencescheduled;
>> +
>> +   /* This fence is what will be signaled by the scheduler when
>> +* the job is completed.
>> +*
>> +* When setting up an out fence for the job, you should use
>> +* this, since it's available immediately upon
>> +* drm_sched_job_init(), and the fence returned by the driver
>> +* from run_job() won't be created until the dependencies have
>> +* resolved.
>> +*/
>> struct dma_fencefinished;
>> +
>> struct dma_fence_cb cb;
>> struct dma_fence*parent;
>> struct drm_gpu_scheduler*sched;
>> @@ -88,6 +101,13 @@ struct drm_sched_fence {
>>
>>  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
>>
>> +/**
>> + * drm_sched_job - A job to be run by an entity.
>> + *
>> + * A job is created by the driver using drm_sched_job_init(), and
>> + * should call drm_sched_entity_push_job() once it wants the scheduler
>> + * to schedule the job.
>> + */
>>  struct drm_sched_job {
>> struct spsc_nodequeue_node;
>> struct drm_gpu_scheduler*sched;
>> @@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
>> drm_sched_job *s_job,
>>   * these functions should be implemented in driver side
>>  */
>>  struct drm_sched_backend_ops {
>> +   /* Called when the scheduler is considering scheduling this
>> +* job next, to get another struct dma_fence for this job to
>> +* block on.  Once it returns NULL, run_job() may be called.
>> +*/
>> struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
>> struct drm_sched_entity *s_entity);
>> +
>> +   /* Called to execute the job once all of the dependencies have
>> +* been resolved.  This may be called multiple times, if
>> +* timedout_job() has happened and drm_sched_job_recovery()
>> +* decides to try it again.
>> +*/
>> struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
>> +
>> +   /* Called when a job has taken too long to execute, to trigger
>> +* GPU recovery.
>> +*/
>> void (*timedout_job)(struct drm_sched_job *sched_job);
>> +
>> +   /* Called once the job's finished fence has been signaled and
>> +* it's time to clean it up.
>> +*/
>> void (*free_job)(struct drm_sched_job *sched_job);
>>  };
>>
>> --
>> 2.17.0
>>
>> ___
>> dri-devel mailing list
>> dri-de...@lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>
>
>
> --
> Daniel Vetter
> Software Engineer, Intel Corporation
> +41 (0) 79 365 57 48 - http://blog.ffwll.ch
> ___
> dri-devel mailing list
> dri-de...@lists.freedesktop.org
> 

Re: [PATCH] drm/sched: Extend the documentation.

[Alex Deucher]

On Thu, Apr 5, 2018 at 2:16 AM, Daniel Vetter  wrote:
> On Thu, Apr 5, 2018 at 12:32 AM, Eric Anholt  wrote:
>> These comments answer all the questions I had for myself when
>> implementing a driver using the GPU scheduler.
>>
>> Signed-off-by: Eric Anholt 
>
> Pulling all these comments into the generated kerneldoc would be
> awesome, maybe as a new "GPU Scheduler" chapter at the end of
> drm-mm.rst? Would mean a bit of busywork to convert the existing raw
> comments into proper kerneldoc. Also has the benefit that 0day will
> complain when you forget to update the comment when editing the
> function prototype - kerneldoc which isn't included anywhere in .rst
> won't be checked automatically.

I was actually planning to do this myself, but Nayan wanted to do this
a prep work for his proposed GSoC project so I was going to see how
far he got first.

Alex

> -Daniel
>
>> ---
>>  include/drm/gpu_scheduler.h | 46 +
>>  1 file changed, 42 insertions(+), 4 deletions(-)
>>
>> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
>> index dfd54fb94e10..c053a32341bf 100644
>> --- a/include/drm/gpu_scheduler.h
>> +++ b/include/drm/gpu_scheduler.h
>> @@ -43,10 +43,12 @@ enum drm_sched_priority {
>>  };
>>
>>  /**
>> - * A scheduler entity is a wrapper around a job queue or a group
>> - * of other entities. Entities take turns emitting jobs from their
>> - * job queues to corresponding hardware ring based on scheduling
>> - * policy.
>> + * drm_sched_entity - A wrapper around a job queue (typically attached
>> + * to the DRM file_priv).
>> + *
>> + * Entities will emit jobs in order to their corresponding hardware
>> + * ring, and the scheduler will alternate between entities based on
>> + * scheduling policy.
>>  */
>>  struct drm_sched_entity {
>> struct list_headlist;
>> @@ -78,7 +80,18 @@ struct drm_sched_rq {
>>
>>  struct drm_sched_fence {
>> struct dma_fencescheduled;
>> +
>> +   /* This fence is what will be signaled by the scheduler when
>> +* the job is completed.
>> +*
>> +* When setting up an out fence for the job, you should use
>> +* this, since it's available immediately upon
>> +* drm_sched_job_init(), and the fence returned by the driver
>> +* from run_job() won't be created until the dependencies have
>> +* resolved.
>> +*/
>> struct dma_fencefinished;
>> +
>> struct dma_fence_cb cb;
>> struct dma_fence*parent;
>> struct drm_gpu_scheduler*sched;
>> @@ -88,6 +101,13 @@ struct drm_sched_fence {
>>
>>  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
>>
>> +/**
>> + * drm_sched_job - A job to be run by an entity.
>> + *
>> + * A job is created by the driver using drm_sched_job_init(), and
>> + * should call drm_sched_entity_push_job() once it wants the scheduler
>> + * to schedule the job.
>> + */
>>  struct drm_sched_job {
>> struct spsc_nodequeue_node;
>> struct drm_gpu_scheduler*sched;
>> @@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
>> drm_sched_job *s_job,
>>   * these functions should be implemented in driver side
>>  */
>>  struct drm_sched_backend_ops {
>> +   /* Called when the scheduler is considering scheduling this
>> +* job next, to get another struct dma_fence for this job to
>> +* block on.  Once it returns NULL, run_job() may be called.
>> +*/
>> struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
>> struct drm_sched_entity *s_entity);
>> +
>> +   /* Called to execute the job once all of the dependencies have
>> +* been resolved.  This may be called multiple times, if
>> +* timedout_job() has happened and drm_sched_job_recovery()
>> +* decides to try it again.
>> +*/
>> struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
>> +
>> +   /* Called when a job has taken too long to execute, to trigger
>> +* GPU recovery.
>> +*/
>> void (*timedout_job)(struct drm_sched_job *sched_job);
>> +
>> +   /* Called once the job's finished fence has been signaled and
>> +* it's time to clean it up.
>> +*/
>> void (*free_job)(struct drm_sched_job *sched_job);
>>  };
>>
>> --
>> 2.17.0
>>
>> ___
>> dri-devel mailing list
>> dri-de...@lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>
>
>
> --
> Daniel Vetter
> Software Engineer, Intel Corporation
> +41 (0) 79 365 57 48 - http://blog.ffwll.ch
> ___
> dri-devel mailing list
> dri-de...@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH] drm/sched: Extend the documentation.

[Christian König]

Am 05.04.2018 um 00:32 schrieb Eric Anholt:

These comments answer all the questions I had for myself when
implementing a driver using the GPU scheduler.

Signed-off-by: Eric Anholt 


Reviewed-by: Christian König 

Already pushed to amd-staging-drm-next as well.

Thanks,
Christian.


---
  include/drm/gpu_scheduler.h | 46 +
  1 file changed, 42 insertions(+), 4 deletions(-)

diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
index dfd54fb94e10..c053a32341bf 100644
--- a/include/drm/gpu_scheduler.h
+++ b/include/drm/gpu_scheduler.h
@@ -43,10 +43,12 @@ enum drm_sched_priority {
  };
  
  /**

- * A scheduler entity is a wrapper around a job queue or a group
- * of other entities. Entities take turns emitting jobs from their
- * job queues to corresponding hardware ring based on scheduling
- * policy.
+ * drm_sched_entity - A wrapper around a job queue (typically attached
+ * to the DRM file_priv).
+ *
+ * Entities will emit jobs in order to their corresponding hardware
+ * ring, and the scheduler will alternate between entities based on
+ * scheduling policy.
  */
  struct drm_sched_entity {
struct list_headlist;
@@ -78,7 +80,18 @@ struct drm_sched_rq {
  
  struct drm_sched_fence {

struct dma_fencescheduled;
+
+   /* This fence is what will be signaled by the scheduler when
+* the job is completed.
+*
+* When setting up an out fence for the job, you should use
+* this, since it's available immediately upon
+* drm_sched_job_init(), and the fence returned by the driver
+* from run_job() won't be created until the dependencies have
+* resolved.
+*/
struct dma_fencefinished;
+
struct dma_fence_cb cb;
struct dma_fence*parent;
struct drm_gpu_scheduler*sched;
@@ -88,6 +101,13 @@ struct drm_sched_fence {
  
  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
  
+/**

+ * drm_sched_job - A job to be run by an entity.
+ *
+ * A job is created by the driver using drm_sched_job_init(), and
+ * should call drm_sched_entity_push_job() once it wants the scheduler
+ * to schedule the job.
+ */
  struct drm_sched_job {
struct spsc_nodequeue_node;
struct drm_gpu_scheduler*sched;
@@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
drm_sched_job *s_job,
   * these functions should be implemented in driver side
  */
  struct drm_sched_backend_ops {
+   /* Called when the scheduler is considering scheduling this
+* job next, to get another struct dma_fence for this job to
+* block on.  Once it returns NULL, run_job() may be called.
+*/
struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
struct drm_sched_entity *s_entity);
+
+   /* Called to execute the job once all of the dependencies have
+* been resolved.  This may be called multiple times, if
+* timedout_job() has happened and drm_sched_job_recovery()
+* decides to try it again.
+*/
struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
+
+   /* Called when a job has taken too long to execute, to trigger
+* GPU recovery.
+*/
void (*timedout_job)(struct drm_sched_job *sched_job);
+
+   /* Called once the job's finished fence has been signaled and
+* it's time to clean it up.
+*/
void (*free_job)(struct drm_sched_job *sched_job);
  };
  

Re: [PATCH] drm/sched: Extend the documentation.

[Christian König]

Am 05.04.2018 um 00:32 schrieb Eric Anholt:

These comments answer all the questions I had for myself when
implementing a driver using the GPU scheduler.

Signed-off-by: Eric Anholt 


Reviewed-by: Christian König 

Already pushed to amd-staging-drm-next as well.

Thanks,
Christian.


---
  include/drm/gpu_scheduler.h | 46 +
  1 file changed, 42 insertions(+), 4 deletions(-)

diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
index dfd54fb94e10..c053a32341bf 100644
--- a/include/drm/gpu_scheduler.h
+++ b/include/drm/gpu_scheduler.h
@@ -43,10 +43,12 @@ enum drm_sched_priority {
  };
  
  /**

- * A scheduler entity is a wrapper around a job queue or a group
- * of other entities. Entities take turns emitting jobs from their
- * job queues to corresponding hardware ring based on scheduling
- * policy.
+ * drm_sched_entity - A wrapper around a job queue (typically attached
+ * to the DRM file_priv).
+ *
+ * Entities will emit jobs in order to their corresponding hardware
+ * ring, and the scheduler will alternate between entities based on
+ * scheduling policy.
  */
  struct drm_sched_entity {
struct list_headlist;
@@ -78,7 +80,18 @@ struct drm_sched_rq {
  
  struct drm_sched_fence {

struct dma_fencescheduled;
+
+   /* This fence is what will be signaled by the scheduler when
+* the job is completed.
+*
+* When setting up an out fence for the job, you should use
+* this, since it's available immediately upon
+* drm_sched_job_init(), and the fence returned by the driver
+* from run_job() won't be created until the dependencies have
+* resolved.
+*/
struct dma_fencefinished;
+
struct dma_fence_cb cb;
struct dma_fence*parent;
struct drm_gpu_scheduler*sched;
@@ -88,6 +101,13 @@ struct drm_sched_fence {
  
  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
  
+/**

+ * drm_sched_job - A job to be run by an entity.
+ *
+ * A job is created by the driver using drm_sched_job_init(), and
+ * should call drm_sched_entity_push_job() once it wants the scheduler
+ * to schedule the job.
+ */
  struct drm_sched_job {
struct spsc_nodequeue_node;
struct drm_gpu_scheduler*sched;
@@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
drm_sched_job *s_job,
   * these functions should be implemented in driver side
  */
  struct drm_sched_backend_ops {
+   /* Called when the scheduler is considering scheduling this
+* job next, to get another struct dma_fence for this job to
+* block on.  Once it returns NULL, run_job() may be called.
+*/
struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
struct drm_sched_entity *s_entity);
+
+   /* Called to execute the job once all of the dependencies have
+* been resolved.  This may be called multiple times, if
+* timedout_job() has happened and drm_sched_job_recovery()
+* decides to try it again.
+*/
struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
+
+   /* Called when a job has taken too long to execute, to trigger
+* GPU recovery.
+*/
void (*timedout_job)(struct drm_sched_job *sched_job);
+
+   /* Called once the job's finished fence has been signaled and
+* it's time to clean it up.
+*/
void (*free_job)(struct drm_sched_job *sched_job);
  };
  

Re: [PATCH] drm/sched: Extend the documentation.

[Daniel Vetter]

On Thu, Apr 5, 2018 at 12:32 AM, Eric Anholt  wrote:
> These comments answer all the questions I had for myself when
> implementing a driver using the GPU scheduler.
>
> Signed-off-by: Eric Anholt 

Pulling all these comments into the generated kerneldoc would be
awesome, maybe as a new "GPU Scheduler" chapter at the end of
drm-mm.rst? Would mean a bit of busywork to convert the existing raw
comments into proper kerneldoc. Also has the benefit that 0day will
complain when you forget to update the comment when editing the
function prototype - kerneldoc which isn't included anywhere in .rst
won't be checked automatically.
-Daniel

> ---
>  include/drm/gpu_scheduler.h | 46 +
>  1 file changed, 42 insertions(+), 4 deletions(-)
>
> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
> index dfd54fb94e10..c053a32341bf 100644
> --- a/include/drm/gpu_scheduler.h
> +++ b/include/drm/gpu_scheduler.h
> @@ -43,10 +43,12 @@ enum drm_sched_priority {
>  };
>
>  /**
> - * A scheduler entity is a wrapper around a job queue or a group
> - * of other entities. Entities take turns emitting jobs from their
> - * job queues to corresponding hardware ring based on scheduling
> - * policy.
> + * drm_sched_entity - A wrapper around a job queue (typically attached
> + * to the DRM file_priv).
> + *
> + * Entities will emit jobs in order to their corresponding hardware
> + * ring, and the scheduler will alternate between entities based on
> + * scheduling policy.
>  */
>  struct drm_sched_entity {
> struct list_headlist;
> @@ -78,7 +80,18 @@ struct drm_sched_rq {
>
>  struct drm_sched_fence {
> struct dma_fencescheduled;
> +
> +   /* This fence is what will be signaled by the scheduler when
> +* the job is completed.
> +*
> +* When setting up an out fence for the job, you should use
> +* this, since it's available immediately upon
> +* drm_sched_job_init(), and the fence returned by the driver
> +* from run_job() won't be created until the dependencies have
> +* resolved.
> +*/
> struct dma_fencefinished;
> +
> struct dma_fence_cb cb;
> struct dma_fence*parent;
> struct drm_gpu_scheduler*sched;
> @@ -88,6 +101,13 @@ struct drm_sched_fence {
>
>  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
>
> +/**
> + * drm_sched_job - A job to be run by an entity.
> + *
> + * A job is created by the driver using drm_sched_job_init(), and
> + * should call drm_sched_entity_push_job() once it wants the scheduler
> + * to schedule the job.
> + */
>  struct drm_sched_job {
> struct spsc_nodequeue_node;
> struct drm_gpu_scheduler*sched;
> @@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
> drm_sched_job *s_job,
>   * these functions should be implemented in driver side
>  */
>  struct drm_sched_backend_ops {
> +   /* Called when the scheduler is considering scheduling this
> +* job next, to get another struct dma_fence for this job to
> +* block on.  Once it returns NULL, run_job() may be called.
> +*/
> struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
> struct drm_sched_entity *s_entity);
> +
> +   /* Called to execute the job once all of the dependencies have
> +* been resolved.  This may be called multiple times, if
> +* timedout_job() has happened and drm_sched_job_recovery()
> +* decides to try it again.
> +*/
> struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
> +
> +   /* Called when a job has taken too long to execute, to trigger
> +* GPU recovery.
> +*/
> void (*timedout_job)(struct drm_sched_job *sched_job);
> +
> +   /* Called once the job's finished fence has been signaled and
> +* it's time to clean it up.
> +*/
> void (*free_job)(struct drm_sched_job *sched_job);
>  };
>
> --
> 2.17.0
>
> ___
> dri-devel mailing list
> dri-de...@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/dri-devel



-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

Re: [PATCH] drm/sched: Extend the documentation.

[Daniel Vetter]

On Thu, Apr 5, 2018 at 12:32 AM, Eric Anholt  wrote:
> These comments answer all the questions I had for myself when
> implementing a driver using the GPU scheduler.
>
> Signed-off-by: Eric Anholt 

Pulling all these comments into the generated kerneldoc would be
awesome, maybe as a new "GPU Scheduler" chapter at the end of
drm-mm.rst? Would mean a bit of busywork to convert the existing raw
comments into proper kerneldoc. Also has the benefit that 0day will
complain when you forget to update the comment when editing the
function prototype - kerneldoc which isn't included anywhere in .rst
won't be checked automatically.
-Daniel

> ---
>  include/drm/gpu_scheduler.h | 46 +
>  1 file changed, 42 insertions(+), 4 deletions(-)
>
> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
> index dfd54fb94e10..c053a32341bf 100644
> --- a/include/drm/gpu_scheduler.h
> +++ b/include/drm/gpu_scheduler.h
> @@ -43,10 +43,12 @@ enum drm_sched_priority {
>  };
>
>  /**
> - * A scheduler entity is a wrapper around a job queue or a group
> - * of other entities. Entities take turns emitting jobs from their
> - * job queues to corresponding hardware ring based on scheduling
> - * policy.
> + * drm_sched_entity - A wrapper around a job queue (typically attached
> + * to the DRM file_priv).
> + *
> + * Entities will emit jobs in order to their corresponding hardware
> + * ring, and the scheduler will alternate between entities based on
> + * scheduling policy.
>  */
>  struct drm_sched_entity {
> struct list_headlist;
> @@ -78,7 +80,18 @@ struct drm_sched_rq {
>
>  struct drm_sched_fence {
> struct dma_fencescheduled;
> +
> +   /* This fence is what will be signaled by the scheduler when
> +* the job is completed.
> +*
> +* When setting up an out fence for the job, you should use
> +* this, since it's available immediately upon
> +* drm_sched_job_init(), and the fence returned by the driver
> +* from run_job() won't be created until the dependencies have
> +* resolved.
> +*/
> struct dma_fencefinished;
> +
> struct dma_fence_cb cb;
> struct dma_fence*parent;
> struct drm_gpu_scheduler*sched;
> @@ -88,6 +101,13 @@ struct drm_sched_fence {
>
>  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
>
> +/**
> + * drm_sched_job - A job to be run by an entity.
> + *
> + * A job is created by the driver using drm_sched_job_init(), and
> + * should call drm_sched_entity_push_job() once it wants the scheduler
> + * to schedule the job.
> + */
>  struct drm_sched_job {
> struct spsc_nodequeue_node;
> struct drm_gpu_scheduler*sched;
> @@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
> drm_sched_job *s_job,
>   * these functions should be implemented in driver side
>  */
>  struct drm_sched_backend_ops {
> +   /* Called when the scheduler is considering scheduling this
> +* job next, to get another struct dma_fence for this job to
> +* block on.  Once it returns NULL, run_job() may be called.
> +*/
> struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
> struct drm_sched_entity *s_entity);
> +
> +   /* Called to execute the job once all of the dependencies have
> +* been resolved.  This may be called multiple times, if
> +* timedout_job() has happened and drm_sched_job_recovery()
> +* decides to try it again.
> +*/
> struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
> +
> +   /* Called when a job has taken too long to execute, to trigger
> +* GPU recovery.
> +*/
> void (*timedout_job)(struct drm_sched_job *sched_job);
> +
> +   /* Called once the job's finished fence has been signaled and
> +* it's time to clean it up.
> +*/
> void (*free_job)(struct drm_sched_job *sched_job);
>  };
>
> --
> 2.17.0
>
> ___
> dri-devel mailing list
> dri-de...@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/dri-devel



-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

[PATCH] drm/sched: Extend the documentation.

[Eric Anholt]

These comments answer all the questions I had for myself when
implementing a driver using the GPU scheduler.

Signed-off-by: Eric Anholt 
---
 include/drm/gpu_scheduler.h | 46 +
 1 file changed, 42 insertions(+), 4 deletions(-)

diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
index dfd54fb94e10..c053a32341bf 100644
--- a/include/drm/gpu_scheduler.h
+++ b/include/drm/gpu_scheduler.h
@@ -43,10 +43,12 @@ enum drm_sched_priority {
 };
 
 /**
- * A scheduler entity is a wrapper around a job queue or a group
- * of other entities. Entities take turns emitting jobs from their
- * job queues to corresponding hardware ring based on scheduling
- * policy.
+ * drm_sched_entity - A wrapper around a job queue (typically attached
+ * to the DRM file_priv).
+ *
+ * Entities will emit jobs in order to their corresponding hardware
+ * ring, and the scheduler will alternate between entities based on
+ * scheduling policy.
 */
 struct drm_sched_entity {
struct list_headlist;
@@ -78,7 +80,18 @@ struct drm_sched_rq {
 
 struct drm_sched_fence {
struct dma_fencescheduled;
+
+   /* This fence is what will be signaled by the scheduler when
+* the job is completed.
+*
+* When setting up an out fence for the job, you should use
+* this, since it's available immediately upon
+* drm_sched_job_init(), and the fence returned by the driver
+* from run_job() won't be created until the dependencies have
+* resolved.
+*/
struct dma_fencefinished;
+
struct dma_fence_cb cb;
struct dma_fence*parent;
struct drm_gpu_scheduler*sched;
@@ -88,6 +101,13 @@ struct drm_sched_fence {
 
 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
 
+/**
+ * drm_sched_job - A job to be run by an entity.
+ *
+ * A job is created by the driver using drm_sched_job_init(), and
+ * should call drm_sched_entity_push_job() once it wants the scheduler
+ * to schedule the job.
+ */
 struct drm_sched_job {
struct spsc_nodequeue_node;
struct drm_gpu_scheduler*sched;
@@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
drm_sched_job *s_job,
  * these functions should be implemented in driver side
 */
 struct drm_sched_backend_ops {
+   /* Called when the scheduler is considering scheduling this
+* job next, to get another struct dma_fence for this job to
+* block on.  Once it returns NULL, run_job() may be called.
+*/
struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
struct drm_sched_entity *s_entity);
+
+   /* Called to execute the job once all of the dependencies have
+* been resolved.  This may be called multiple times, if
+* timedout_job() has happened and drm_sched_job_recovery()
+* decides to try it again.
+*/
struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
+
+   /* Called when a job has taken too long to execute, to trigger
+* GPU recovery.
+*/
void (*timedout_job)(struct drm_sched_job *sched_job);
+
+   /* Called once the job's finished fence has been signaled and
+* it's time to clean it up.
+*/
void (*free_job)(struct drm_sched_job *sched_job);
 };
 
-- 
2.17.0

[PATCH] drm/sched: Extend the documentation.

[Eric Anholt]

These comments answer all the questions I had for myself when
implementing a driver using the GPU scheduler.

Signed-off-by: Eric Anholt 
---
 include/drm/gpu_scheduler.h | 46 +
 1 file changed, 42 insertions(+), 4 deletions(-)

diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
index dfd54fb94e10..c053a32341bf 100644
--- a/include/drm/gpu_scheduler.h
+++ b/include/drm/gpu_scheduler.h
@@ -43,10 +43,12 @@ enum drm_sched_priority {
 };
 
 /**
- * A scheduler entity is a wrapper around a job queue or a group
- * of other entities. Entities take turns emitting jobs from their
- * job queues to corresponding hardware ring based on scheduling
- * policy.
+ * drm_sched_entity - A wrapper around a job queue (typically attached
+ * to the DRM file_priv).
+ *
+ * Entities will emit jobs in order to their corresponding hardware
+ * ring, and the scheduler will alternate between entities based on
+ * scheduling policy.
 */
 struct drm_sched_entity {
struct list_headlist;
@@ -78,7 +80,18 @@ struct drm_sched_rq {
 
 struct drm_sched_fence {
struct dma_fencescheduled;
+
+   /* This fence is what will be signaled by the scheduler when
+* the job is completed.
+*
+* When setting up an out fence for the job, you should use
+* this, since it's available immediately upon
+* drm_sched_job_init(), and the fence returned by the driver
+* from run_job() won't be created until the dependencies have
+* resolved.
+*/
struct dma_fencefinished;
+
struct dma_fence_cb cb;
struct dma_fence*parent;
struct drm_gpu_scheduler*sched;
@@ -88,6 +101,13 @@ struct drm_sched_fence {
 
 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
 
+/**
+ * drm_sched_job - A job to be run by an entity.
+ *
+ * A job is created by the driver using drm_sched_job_init(), and
+ * should call drm_sched_entity_push_job() once it wants the scheduler
+ * to schedule the job.
+ */
 struct drm_sched_job {
struct spsc_nodequeue_node;
struct drm_gpu_scheduler*sched;
@@ -112,10 +132,28 @@ static inline bool drm_sched_invalidate_job(struct 
drm_sched_job *s_job,
  * these functions should be implemented in driver side
 */
 struct drm_sched_backend_ops {
+   /* Called when the scheduler is considering scheduling this
+* job next, to get another struct dma_fence for this job to
+* block on.  Once it returns NULL, run_job() may be called.
+*/
struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
struct drm_sched_entity *s_entity);
+
+   /* Called to execute the job once all of the dependencies have
+* been resolved.  This may be called multiple times, if
+* timedout_job() has happened and drm_sched_job_recovery()
+* decides to try it again.
+*/
struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
+
+   /* Called when a job has taken too long to execute, to trigger
+* GPU recovery.
+*/
void (*timedout_job)(struct drm_sched_job *sched_job);
+
+   /* Called once the job's finished fence has been signaled and
+* it's time to clean it up.
+*/
void (*free_job)(struct drm_sched_job *sched_job);
 };
 
-- 
2.17.0