Re: [PATCH 2/6] drm/xe: Add generic dependecy jobs / scheduler

["Thomas Hellström" <thomas.hellstrom@linux.intel.com>]

On Tue, 2025-06-17 at 08:37 -0700, Matthew Brost wrote:
> Add generic dependecy jobs / scheduler which serves as wrapper for
> DRM
> scheduler. Useful when we want delay a generic operation until a
> dma-fence signals.
> 
> Existing use cases could be destroying of resources based fences /
> dma-resv, the preempt rebind worker, and pipelined GT TLB
> invalidations.
> 
> Written in such a way it could be moved to DRM subsystem if needed.
> 
> Signed-off-by: Matthew Brost <matthew.brost@intel.com>
> ---
>  drivers/gpu/drm/xe/Makefile           |   1 +
>  drivers/gpu/drm/xe/xe_dep_job_types.h |  29 ++++++
>  drivers/gpu/drm/xe/xe_dep_scheduler.c | 145
> ++++++++++++++++++++++++++
>  drivers/gpu/drm/xe/xe_dep_scheduler.h |  21 ++++
>  4 files changed, 196 insertions(+)
>  create mode 100644 drivers/gpu/drm/xe/xe_dep_job_types.h
>  create mode 100644 drivers/gpu/drm/xe/xe_dep_scheduler.c
>  create mode 100644 drivers/gpu/drm/xe/xe_dep_scheduler.h
> 
> diff --git a/drivers/gpu/drm/xe/Makefile
> b/drivers/gpu/drm/xe/Makefile
> index f5f5775acdc0..3523bd093b7a 100644
> --- a/drivers/gpu/drm/xe/Makefile
> +++ b/drivers/gpu/drm/xe/Makefile
> @@ -28,6 +28,7 @@ $(obj)/generated/%_wa_oob.c
> $(obj)/generated/%_wa_oob.h: $(obj)/xe_gen_wa_oob \
>  xe-y += xe_bb.o \
>  	xe_bo.o \
>  	xe_bo_evict.o \
> +	xe_dep_scheduler.o \
>  	xe_devcoredump.o \
>  	xe_device.o \
>  	xe_device_sysfs.o \
> diff --git a/drivers/gpu/drm/xe/xe_dep_job_types.h
> b/drivers/gpu/drm/xe/xe_dep_job_types.h
> new file mode 100644
> index 000000000000..c6a484f24c8c
> --- /dev/null
> +++ b/drivers/gpu/drm/xe/xe_dep_job_types.h
> @@ -0,0 +1,29 @@
> +/* SPDX-License-Identifier: MIT */
> +/*
> + * Copyright © 2025 Intel Corporation
> + */
> +
> +#ifndef _XE_DEP_JOB_TYPES_H_
> +#define _XE_DEP_JOB_TYPES_H_
> +
> +#include <drm/gpu_scheduler.h>
> +
> +struct xe_dep_job;
> +
> +/** struct xe_dep_job_ops - Generic Xe dependency job operations */
> +struct xe_dep_job_ops {
> +	/** @run_job: Run generic Xe dependency job */
> +	struct dma_fence *(*run_job)(struct xe_dep_job *job);
> +	/** @free_job: Free generic Xe dependency job */
> +	void (*free_job)(struct xe_dep_job *job);
> +};
> +
> +/** struct xe_dep_job - Generic dependency Xe job */
> +struct xe_dep_job {
> +	/** @drm: base DRM scheduler job */
> +	struct drm_sched_job drm;
> +	/** @ops: dependency job operations */
> +	const struct xe_dep_job_ops *ops;
> +};
> +
> +#endif
> diff --git a/drivers/gpu/drm/xe/xe_dep_scheduler.c
> b/drivers/gpu/drm/xe/xe_dep_scheduler.c
> new file mode 100644
> index 000000000000..fbd55577d787
> --- /dev/null
> +++ b/drivers/gpu/drm/xe/xe_dep_scheduler.c
> @@ -0,0 +1,145 @@
> +// SPDX-License-Identifier: MIT
> +/*
> + * Copyright © 2025 Intel Corporation
> + */
> +
> +#include <linux/slab.h>
> +
> +#include <drm/gpu_scheduler.h>
> +
> +#include "xe_dep_job_types.h"
> +#include "xe_dep_scheduler.h"
> +#include "xe_device_types.h"
> +
> +/**
> + * DOC: Xe Dependency Scheduler
> + *
> + * The Xe dependency scheduler is a simple wrapper built around the
> DRM
> + * scheduler to execute jobs once their dependencies are resolved
> (i.e., all
> + * input fences specified as dependencies are signaled). The jobs
> that are
> + * executed contain virtual functions to run (execute) and free the
> job,
> + * allowing a single dependency scheduler to handle jobs performing
> different
> + * operations.
> + *
> + * Example use cases include deferred resource freeing, TLB
> invalidations after
> + * bind jobs, etc.
> + */
> +
> +/** struct xe_dep_scheduler - Generic Xe dependency scheduler */
> +struct xe_dep_scheduler {
> +	/** @sched: DRM GPU scheduler */
> +	struct drm_gpu_scheduler sched;

Would it make sense to support sharing a single drm_gpu_scheduler for
many xe_dep_schedulers? Thinking of the first use-case where we'd
probably want just a single drm_gpu_scheduler per GuC since
invalidations are ordered anyway?

/Thomas


> +	/** @entity: DRM scheduler entity  */
> +	struct drm_sched_entity entity;
> +	/** @rcu: For safe freeing of exported dma fences */
> +	struct rcu_head rcu;
> +};
> +
> +static struct dma_fence *xe_dep_scheduler_run_job(struct
> drm_sched_job *drm_job)
> +{
> +	struct xe_dep_job *dep_job =
> +		container_of(drm_job, typeof(*dep_job), drm);
> +
> +	return dep_job->ops->run_job(dep_job);
> +}
> +
> +static void xe_dep_scheduler_free_job(struct drm_sched_job *drm_job)
> +{
> +	struct xe_dep_job *dep_job =
> +		container_of(drm_job, typeof(*dep_job), drm);
> +
> +	dep_job->ops->free_job(dep_job);
> +}
> +
> +static const struct drm_sched_backend_ops sched_ops = {
> +	.run_job = xe_dep_scheduler_run_job,
> +	.free_job = xe_dep_scheduler_free_job,
> +};
> +
> +/**
> + * xe_dep_scheduler_create() - Generic Xe dependency scheduler
> create
> + * @xe: Xe device
> + * @submit_wq: Submit workqueue struct (can be NULL)
> + * @name: Name of dependency scheduler
> + * @job_limit: Max dependency jobs that can be scheduled
> + *
> + * Create a generic Xe dependency scheduler and initialize internal
> DRM
> + * scheduler objects.
> + *
> + * Return: Generic Xe dependency scheduler object or ERR_PTR
> + */
> +struct xe_dep_scheduler *
> +xe_dep_scheduler_create(struct xe_device *xe,
> +			struct workqueue_struct *submit_wq,
> +			const char *name, u32 job_limit)
> +{
> +	struct xe_dep_scheduler *dep_scheduler;
> +	struct drm_gpu_scheduler *sched;
> +	const struct drm_sched_init_args args = {
> +		.ops = &sched_ops,
> +		.submit_wq = submit_wq,
> +		.num_rqs = 1,
> +		.credit_limit = job_limit,
> +		.timeout = MAX_SCHEDULE_TIMEOUT,
> +		.name = name,
> +		.dev = xe->drm.dev,
> +	};
> +	int err;
> +
> +	dep_scheduler = kzalloc(sizeof(*dep_scheduler), GFP_KERNEL);
> +	if (!dep_scheduler)
> +		return ERR_PTR(-ENOMEM);
> +
> +	err = drm_sched_init(&dep_scheduler->sched, &args);
> +	if (err)
> +		goto err_free;
> +
> +	sched = &dep_scheduler->sched;
> +	err = drm_sched_entity_init(&dep_scheduler->entity, 0,
> +				    (struct drm_gpu_scheduler
> **)&sched, 1,
> +				    NULL);
> +	if (err)
> +		goto err_sched;
> +
> +	init_rcu_head(&dep_scheduler->rcu);
> +
> +	return dep_scheduler;
> +
> +err_sched:
> +	drm_sched_fini(&dep_scheduler->sched);
> +err_free:
> +	kfree(dep_scheduler);
> +
> +	return ERR_PTR(err);
> +}
> +
> +/**
> + * xe_dep_scheduler_fini() - Generic Xe dependency scheduler
> finalize
> + * @dep_scheduler: Generic Xe dependency scheduler object
> + *
> + * Finalize internal DRM scheduler objects and free generic Xe
> dependency
> + * scheduler object
> + */
> +void xe_dep_scheduler_fini(struct xe_dep_scheduler *dep_scheduler)
> +{
> +	drm_sched_entity_fini(&dep_scheduler->entity);
> +	drm_sched_fini(&dep_scheduler->sched);
> +	/*
> +	 * RCU free due sched being exported via DRM scheduler
> fences
> +	 * (timeline name).
> +	 */
> +	kfree_rcu(dep_scheduler, rcu);
> +}
> +
> +/**
> + * xe_dep_scheduler_entity() - Retrieve a generic Xe dependency
> scheduler
> + *                             DRM scheduler entity
> + * @dep_scheduler: Generic Xe dependency scheduler object
> + *
> + * Return: The generic Xe dependency scheduler's DRM scheduler
> entity
> + */
> +struct drm_sched_entity *
> +xe_dep_scheduler_entity(struct xe_dep_scheduler *dep_scheduler)
> +{
> +	return &dep_scheduler->entity;
> +}
> diff --git a/drivers/gpu/drm/xe/xe_dep_scheduler.h
> b/drivers/gpu/drm/xe/xe_dep_scheduler.h
> new file mode 100644
> index 000000000000..853961eec64b
> --- /dev/null
> +++ b/drivers/gpu/drm/xe/xe_dep_scheduler.h
> @@ -0,0 +1,21 @@
> +/* SPDX-License-Identifier: MIT */
> +/*
> + * Copyright © 2025 Intel Corporation
> + */
> +
> +#include <linux/types.h>
> +
> +struct drm_sched_entity;
> +struct workqueue_struct;
> +struct xe_dep_scheduler;
> +struct xe_device;
> +
> +struct xe_dep_scheduler *
> +xe_dep_scheduler_create(struct xe_device *xe,
> +			struct workqueue_struct *submit_wq,
> +			const char *name, u32 job_limit);
> +
> +void xe_dep_scheduler_fini(struct xe_dep_scheduler *dep_scheduler);
> +
> +struct drm_sched_entity *
> +xe_dep_scheduler_entity(struct xe_dep_scheduler *dep_scheduler);