[PATCH v2 3/5] drm/xe: Add exec_queue.sched_props.priority
Matthew Brost
matthew.brost at intel.com
Thu Jan 4 19:14:12 UTC 2024
On Thu, Jan 04, 2024 at 11:03:44AM -0800, Welty, Brian wrote:
>
>
> On 1/4/2024 8:46 AM, Matthew Brost wrote:
> > On Wed, Jan 03, 2024 at 10:44:06AM -0800, Brian Welty wrote:
> > > The purpose here is to allow to optimize exec_queue_set_priority()
> > > in follow-on patch. Currently it does q->ops->set_priority(...).
> > > But we'd like to apply exec_queue_user_extensions much earlier and
> > > q->ops cannot be called before __xe_exec_queue_init().
> > >
> > > It will be much more efficient to instead only have to set
> > > q->sched_props.priority when applying user extensions. That value will
> > > then already be set to the user requested value. So the setting of
> > > default value is moved from q->ops->init() to __xe_exec_queue_alloc.
> > >
> > > v2: fix existing bug such that q->sched_props.priority is now set
> > > before guc_exec_queue_add_msg() (Matt)
> > > fix existing bug in that xe_migrate_init() should use exec_queue's
> > > vfunc for updating priority.
> > >
> > > Signed-off-by: Brian Welty <brian.welty at intel.com>
> > > ---
> > > drivers/gpu/drm/xe/xe_exec_queue.c | 1 +
> > > drivers/gpu/drm/xe/xe_exec_queue_types.h | 4 ++--
> > > drivers/gpu/drm/xe/xe_guc_submit.c | 7 +++----
> > > drivers/gpu/drm/xe/xe_migrate.c | 2 +-
> > > 4 files changed, 7 insertions(+), 7 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/xe/xe_exec_queue.c b/drivers/gpu/drm/xe/xe_exec_queue.c
> > > index e78b13845417..9891cddba71c 100644
> > > --- a/drivers/gpu/drm/xe/xe_exec_queue.c
> > > +++ b/drivers/gpu/drm/xe/xe_exec_queue.c
> > > @@ -67,6 +67,7 @@ static struct xe_exec_queue *__xe_exec_queue_alloc(struct xe_device *xe,
> > > hwe->eclass->sched_props.preempt_timeout_us;
> > > q->sched_props.job_timeout_ms =
> > > hwe->eclass->sched_props.job_timeout_ms;
> > > + q->sched_props.priority = XE_EXEC_QUEUE_PRIORITY_NORMAL;
> > > if (xe_exec_queue_is_parallel(q)) {
> > > q->parallel.composite_fence_ctx = dma_fence_context_alloc(1);
> > > diff --git a/drivers/gpu/drm/xe/xe_exec_queue_types.h b/drivers/gpu/drm/xe/xe_exec_queue_types.h
> > > index 882eb5373980..6ae4f4e2ddca 100644
> > > --- a/drivers/gpu/drm/xe/xe_exec_queue_types.h
> > > +++ b/drivers/gpu/drm/xe/xe_exec_queue_types.h
> > > @@ -52,8 +52,6 @@ struct xe_exec_queue {
> > > struct xe_vm *vm;
> > > /** @class: class of this exec queue */
> > > enum xe_engine_class class;
> > > - /** @priority: priority of this exec queue */
> > > - enum xe_exec_queue_priority priority;
> > > /**
> > > * @logical_mask: logical mask of where job submitted to exec queue can run
> > > */
> > > @@ -144,6 +142,8 @@ struct xe_exec_queue {
> > > u32 preempt_timeout_us;
> > > /** @job_timeout_ms: job timeout in milliseconds */
> > > u32 job_timeout_ms;
> > > + /** @priority: priority of this exec queue */
> > > + enum xe_exec_queue_priority priority;
> > > } sched_props;
> > > /** @compute: compute exec queue state */
> > > diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c b/drivers/gpu/drm/xe/xe_guc_submit.c
> > > index 56c0a7bf554f..392cbde62957 100644
> > > --- a/drivers/gpu/drm/xe/xe_guc_submit.c
> > > +++ b/drivers/gpu/drm/xe/xe_guc_submit.c
> > > @@ -421,7 +421,7 @@ static void init_policies(struct xe_guc *guc, struct xe_exec_queue *q)
> > > {
> > > struct exec_queue_policy policy;
> > > struct xe_device *xe = guc_to_xe(guc);
> > > - enum xe_exec_queue_priority prio = q->priority;
> > > + enum xe_exec_queue_priority prio = q->sched_props.priority;
> > > u32 timeslice_us = q->sched_props.timeslice_us;
> > > u32 preempt_timeout_us = q->sched_props.preempt_timeout_us;
> > > @@ -1231,7 +1231,6 @@ static int guc_exec_queue_init(struct xe_exec_queue *q)
> > > err = xe_sched_entity_init(&ge->entity, sched);
> > > if (err)
> > > goto err_sched;
> > > - q->priority = XE_EXEC_QUEUE_PRIORITY_NORMAL;
> > > if (xe_exec_queue_is_lr(q))
> > > INIT_WORK(&q->guc->lr_tdr, xe_guc_exec_queue_lr_cleanup);
> > > @@ -1301,15 +1300,15 @@ static int guc_exec_queue_set_priority(struct xe_exec_queue *q,
> > > {
> > > struct xe_sched_msg *msg;
> > > - if (q->priority == priority || exec_queue_killed_or_banned(q))
> > > + if (q->sched_props.priority == priority || exec_queue_killed_or_banned(q))
> > > return 0;
> > > msg = kmalloc(sizeof(*msg), GFP_KERNEL);
> > > if (!msg)
> > > return -ENOMEM;
> > > + q->sched_props.priority = priority;
> > > guc_exec_queue_add_msg(q, msg, SET_SCHED_PROPS);
> > > - q->priority = priority;
> >
> > This probably should be its own patch and maybe with a Fixes tag too.
> >
> > > return 0;
> > > }
> > > diff --git a/drivers/gpu/drm/xe/xe_migrate.c b/drivers/gpu/drm/xe/xe_migrate.c
> > > index adf1dab5eba2..f967fa69769e 100644
> > > --- a/drivers/gpu/drm/xe/xe_migrate.c
> > > +++ b/drivers/gpu/drm/xe/xe_migrate.c
> > > @@ -356,7 +356,7 @@ struct xe_migrate *xe_migrate_init(struct xe_tile *tile)
> > > return ERR_CAST(m->q);
> > > }
> > > if (xe->info.has_usm)
> > > - m->q->priority = XE_EXEC_QUEUE_PRIORITY_KERNEL;
> > > + m->q->ops->set_priority(m->q, XE_EXEC_QUEUE_PRIORITY_KERNEL);
> >
> > Same here.
>
> Okay, will pursue.... but then I guess I'd rather pass a new flag into
> xe_exec_queue_create() and just set the priority upfront rather than use
> the vfunc to modify it here later. Will code it up and see how it looks.
> Then don't have to deal with the vfunc possibly failing.
>
This vfunc shouldn't fail in practice and if it does it actually does
not matter. We might want to audit the xe_exec_queue_ops return values
and change some of these to void.
> Would it be wrong to simply always use XE_EXEC_QUEUE_PRIORITY_KERNEL
> for exec_queues created with flags:
> EXEC_QUEUE_FLAG_KERNEL | EXEC_QUEUE_FLAG_PERMANENT
> ??
>
It seems like this works for now but need to careful going forward to
make sure assumption is correct.
Fine with it either way.
Matt
>
> >
> > Otherwise LGTM.
> >
> > Matt
> >
> > > mutex_init(&m->job_mutex);
> > > --
> > > 2.43.0
> > >
More information about the Intel-xe
mailing list