[RFC v1 3/9] drm/xe/hw_engine_group: Register hw engine group's exec queues

[Matthew Brost]

On Mon, Jul 22, 2024 at 10:31:49AM +0200, Francois Dugast wrote:
> On Wed, Jul 17, 2024 at 11:19:13PM +0000, Matthew Brost wrote:
> > On Wed, Jul 17, 2024 at 03:07:24PM +0200, Francois Dugast wrote:
> > > Add helpers to safely add and delete the exec queues attached to a hw
> > > engine group, and make use them at the time of creation and destruction
> > > of the exec queues. Keeping track of them is required to control the
> > > execution mode of the hw engine group.
> > > 
> > 
> > Ugh, missed one thing.
> > 
> > > Signed-off-by: Francois Dugast <francois.dugast at intel.com>
> > > ---
> > >  drivers/gpu/drm/xe/xe_exec_queue.c |  7 +++++
> > >  drivers/gpu/drm/xe/xe_hw_engine.c  | 45 ++++++++++++++++++++++++++++++
> > >  drivers/gpu/drm/xe/xe_hw_engine.h  |  4 +++
> > >  3 files changed, 56 insertions(+)
> > > 
> > > diff --git a/drivers/gpu/drm/xe/xe_exec_queue.c b/drivers/gpu/drm/xe/xe_exec_queue.c
> > > index 0ba37835849b..645423a63434 100644
> > > --- a/drivers/gpu/drm/xe/xe_exec_queue.c
> > > +++ b/drivers/gpu/drm/xe/xe_exec_queue.c
> > > @@ -192,6 +192,9 @@ void xe_exec_queue_destroy(struct kref *ref)
> > >  			xe_exec_queue_put(eq);
> > >  	}
> > >  
> > > +	if (q->vm)
> > > +		xe_hw_engine_group_del_exec_queue(q->hwe->hw_engine_group, q);
> > > +
> > >  	q->ops->fini(q);
> > >  }
> > >  
> > > @@ -640,6 +643,10 @@ int xe_exec_queue_create_ioctl(struct drm_device *dev, void *data,
> > >  			if (XE_IOCTL_DBG(xe, err))
> > >  				goto put_exec_queue;
> > >  		}
> > > +
> > > +		err = xe_hw_engine_group_add_exec_queue(q->hwe->hw_engine_group, q);
> > > +		if (err)
> > > +			goto put_exec_queue;
> > >  	}
> > >  
> > >  	mutex_lock(&xef->exec_queue.lock);
> > > diff --git a/drivers/gpu/drm/xe/xe_hw_engine.c b/drivers/gpu/drm/xe/xe_hw_engine.c
> > > index f8df85d25617..4dcc885a55c8 100644
> > > --- a/drivers/gpu/drm/xe/xe_hw_engine.c
> > > +++ b/drivers/gpu/drm/xe/xe_hw_engine.c
> > > @@ -1178,3 +1178,48 @@ u64 xe_hw_engine_read_timestamp(struct xe_hw_engine *hwe)
> > >  {
> > >  	return xe_mmio_read64_2x32(hwe->gt, RING_TIMESTAMP(hwe->mmio_base));
> > >  }
> > > +
> > > +/**
> > > + * xe_hw_engine_group_add_exec_queue() - Add an exec queue to a hw engine group
> > > + * @group: The hw engine group
> > > + * @q: The exec_queue
> > > + *
> > > + * Return: 0 on success,
> > > + *	    -EINTR if the lock could not be acquired
> > > + */
> > > +int xe_hw_engine_group_add_exec_queue(struct xe_hw_engine_group *group, struct xe_exec_queue *q)
> > > +{
> > > +	int err;
> > > +
> > > +	err = down_write_killable(&group->mode_sem);
> > > +	if (err)
> > > +		return err;
> > > +
> > 
> > You need to ensure exec queue in fault mode as in the suspend state if
> > the group is in dma_fence mode:
> > 
> > if (xe_vm_in_fault_mode(q->vm) && group->cur_mode == EXEC_MODE_DMA_FENCE) {
> > 	q->suspend();
> > 	q->suspend_wait();
> > 	queue_work(resume);
> > }
> 
> This happens later in the series when switching execution mode to EXEC_MODE_DMA_FENCE
> at the time of job submission. But this function here is called at the time of exec
> queue creation, am I missing a path where the exec queue would be executing before
> a job submission happens?
> 

Oh, yea I guess if always switch modes in the exec IOCTL is this
probably wouldn't be needed. It would be needed resume worker is just
kicked in the exec IOCTL.

Conceptually I think it probably best to make all the queues in the
group in the same state though. Having mixed state just seems like a
good way to get into trouble.

Matt

> Francois
> 
> > 
> > Matt
> > 
> > > +	list_add(&q->hw_engine_group_link, &group->exec_queue_list);
> > > +	up_write(&group->mode_sem);
> > > +
> > > +	return 0;
> > > +}
> > > +
> > > +/**
> > > + * xe_hw_engine_group_del_exec_queue() - Delete an exec queue from a hw engine group
> > > + * @group: The hw engine group
> > > + * @q: The exec_queue
> > > + *
> > > + * Return: 0 on success,
> > > + *	    -EINTR if the lock could not be acquired
> > > + */
> > > +int xe_hw_engine_group_del_exec_queue(struct xe_hw_engine_group *group, struct xe_exec_queue *q)
> > > +{
> > > +	int err;
> > > +
> > > +	err = down_write_killable(&group->mode_sem);
> > > +	if (err)
> > > +		return err;
> > > +
> > > +	if (!list_empty(&group->exec_queue_list))
> > > +		list_del(&q->hw_engine_group_link);
> > > +	up_write(&group->mode_sem);
> > > +
> > > +	return 0;
> > > +}
> > > diff --git a/drivers/gpu/drm/xe/xe_hw_engine.h b/drivers/gpu/drm/xe/xe_hw_engine.h
> > > index 7f2d27c0ba1a..ce59d83a75ad 100644
> > > --- a/drivers/gpu/drm/xe/xe_hw_engine.h
> > > +++ b/drivers/gpu/drm/xe/xe_hw_engine.h
> > > @@ -9,6 +9,7 @@
> > >  #include "xe_hw_engine_types.h"
> > >  
> > >  struct drm_printer;
> > > +struct xe_exec_queue;
> > >  
> > >  #ifdef CONFIG_DRM_XE_JOB_TIMEOUT_MIN
> > >  #define XE_HW_ENGINE_JOB_TIMEOUT_MIN CONFIG_DRM_XE_JOB_TIMEOUT_MIN
> > > @@ -70,4 +71,7 @@ static inline bool xe_hw_engine_is_valid(struct xe_hw_engine *hwe)
> > >  const char *xe_hw_engine_class_to_str(enum xe_engine_class class);
> > >  u64 xe_hw_engine_read_timestamp(struct xe_hw_engine *hwe);
> > >  
> > > +int xe_hw_engine_group_add_exec_queue(struct xe_hw_engine_group *group, struct xe_exec_queue *q);
> > > +int xe_hw_engine_group_del_exec_queue(struct xe_hw_engine_group *group, struct xe_exec_queue *q);
> > > +
> > >  #endif
> > > -- 
> > > 2.43.0
> > >