Re: [PATCH v3 10/25] drm/xe: Update GuC submission backend to run PT jobs

["Summers, Stuart" <stuart.summers@intel.com>]

On Tue, 2026-03-03 at 16:26 -0800, Matthew Brost wrote:
> On Tue, Mar 03, 2026 at 04:28:50PM -0700, Summers, Stuart wrote:
> > On Fri, 2026-02-27 at 17:34 -0800, Matthew Brost wrote:
> > > PT jobs bypass GPU execution for the final step of a bind job,
> > > using
> > > the
> > > CPU to program the required page tables. Teach the GuC submission
> > > backend
> > > how to execute these jobs.
> > > 
> > > PT job submission is implemented in the GuC backend for
> > > simplicity. A
> > > follow-up patch could introduce a dedicated backend for PT jobs.
> > 
> > Still looking through the whole series, but standing alone, this
> > patch
> > doesn't feel right to me. I don't see why we'd want to hook
> > together
> > the PT update flow with the GuC backend...
> > 
> 
> I don't think it is either, which is why I called out a follow-up to
> implement the PT backend. It’s likely a bigger refactor than one
> would
> expect, though...
> 
> - Build the backend on top of xe_dep_scheduler
> - Introduce xe_pt_job that inherits from xe_dep_job
> - Ripple these changes through CPU bind, the PT layer, etc.
> 
> A lot of that is just shuffling code around across those three steps,
> but it’s not too bad and will likely give us some nice layering
> cleanups
> along the way.
> 
> The real tricky part is handling all the flows that stop/start the
> backends while various global events occur (e.g., PM enter/exit, GT
> resets, VF migration, FLR (WIP), etc.). All of those flows are
> currently
> GT → UC → GuC layered (or, for VF migration, direct to GuC). So we’d
> need a refactor there as well. It’s doable, but it will end up
> touching
> quite a few files. Again, once we do this, I suspect we’ll get
> additional layering cleanups along the way.
> 
> So for now, given the already large size of the series, I’d like to
> get
> the functionality in first and then tackle the layering refactors.

I get what you're saying here. Other than complexity, is there a reason
we can't do that work first though? Is there some critical reason we
need to get the CPU binding work in first, basically?

-Stuart

> 
> Matt
> 
> > Thanks,
> > Stuart
> > 
> > > 
> > > Signed-off-by: Matthew Brost <matthew.brost@intel.com>
> > > ---
> > >  drivers/gpu/drm/xe/xe_guc_submit.c | 37
> > > ++++++++++++++++++++++++++--
> > > --
> > >  drivers/gpu/drm/xe/xe_migrate.c    | 13 ++++++++++-
> > >  drivers/gpu/drm/xe/xe_migrate.h    |  8 +++++++
> > >  3 files changed, 52 insertions(+), 6 deletions(-)
> > > 
> > > diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c
> > > b/drivers/gpu/drm/xe/xe_guc_submit.c
> > > index 453af51fe87b..1d6ac7a6563b 100644
> > > --- a/drivers/gpu/drm/xe/xe_guc_submit.c
> > > +++ b/drivers/gpu/drm/xe/xe_guc_submit.c
> > > @@ -36,8 +36,10 @@
> > >  #include "xe_lrc.h"
> > >  #include "xe_macros.h"
> > >  #include "xe_map.h"
> > > +#include "xe_migrate.h"
> > >  #include "xe_mocs.h"
> > >  #include "xe_pm.h"
> > > +#include "xe_pt.h"
> > >  #include "xe_ring_ops_types.h"
> > >  #include "xe_sched_job.h"
> > >  #include "xe_sleep.h"
> > > @@ -1183,6 +1185,20 @@ static void submit_exec_queue(struct
> > > xe_exec_queue *q, struct xe_sched_job *job)
> > >         }
> > >  }
> > >  
> > > +static bool is_pt_job(struct xe_sched_job *job)
> > > +{
> > > +       return job->is_pt_job;
> > > +}
> > > +
> > > +static void run_pt_job(struct xe_sched_job *job)
> > > +{
> > > +       xe_migrate_update_pgtables_cpu_execute(job-
> > > >pt_update[0].vm,
> > > +                                              job-
> > > > pt_update[0].tile,
> > > +                                              job-
> > > >pt_update[0].ops,
> > > +                                              job-
> > > > pt_update[0].pt_job_ops->ops,
> > > +                                              job-
> > > > pt_update[0].pt_job_ops->current_op);
> > > +}
> > > +
> > >  static struct dma_fence *
> > >  guc_exec_queue_run_job(struct drm_sched_job *drm_job)
> > >  {
> > > @@ -1210,14 +1226,25 @@ guc_exec_queue_run_job(struct
> > > drm_sched_job
> > > *drm_job)
> > >                                 register_exec_queue(primary,
> > > GUC_CONTEXT_NORMAL);
> > >                 }
> > >  
> > > -               if (!exec_queue_registered(q))
> > > -                       register_exec_queue(q,
> > > GUC_CONTEXT_NORMAL);
> > > -               if (!job->restore_replay)
> > > -                       q->ring_ops->emit_job(job);
> > > -               submit_exec_queue(q, job);
> > > +               if (is_pt_job(job)) {
> > > +                       xe_gt_assert(guc_to_gt(guc),
> > > !exec_queue_registered(q));
> > > +                       run_pt_job(job);
> > > +               } else {
> > > +                       if (!exec_queue_registered(q))
> > > +                               register_exec_queue(q,
> > > GUC_CONTEXT_NORMAL);
> > > +                       if (!job->restore_replay)
> > > +                               q->ring_ops->emit_job(job);
> > > +                       submit_exec_queue(q, job);
> > > +               }
> > >                 job->restore_replay = false;
> > >         }
> > >  
> > > +       if (is_pt_job(job)) {
> > > +               xe_pt_job_ops_put(job->pt_update[0].pt_job_ops);
> > > +               dma_fence_put(job->fence);      /* Drop ref from
> > > xe_sched_job_arm */
> > > +               return NULL;
> > > +       }
> > > +
> > >  run_job_out:
> > >  
> > >         return job->fence;
> > > diff --git a/drivers/gpu/drm/xe/xe_migrate.c
> > > b/drivers/gpu/drm/xe/xe_migrate.c
> > > index cd6802642ef3..e9b9dfe19e48 100644
> > > --- a/drivers/gpu/drm/xe/xe_migrate.c
> > > +++ b/drivers/gpu/drm/xe/xe_migrate.c
> > > @@ -1715,7 +1715,18 @@ struct migrate_test_params {
> > >         container_of(_priv, struct migrate_test_params, base)
> > >  #endif
> > >  
> > > -static void
> > > +/**
> > > + * xe_migrate_update_pgtables_cpu_execute() - Update a VM's PTEs
> > > via
> > > the CPU
> > > + * @vm: The VM being updated
> > > + * @tile: The tile being updated
> > > + * @ops: The migrate PT update ops
> > > + * @pt_ops: The VM PT update ops
> > > + * @num_ops: The number of The VM PT update ops
> > > + *
> > > + * Execute the VM PT update ops array which results in a VM's
> > > PTEs
> > > being updated
> > > + * via the CPU.
> > > + */
> > > +void
> > >  xe_migrate_update_pgtables_cpu_execute(struct xe_vm *vm, struct
> > > xe_tile *tile,
> > >                                        const struct
> > > xe_migrate_pt_update_ops *ops,
> > >                                        struct
> > > xe_vm_pgtable_update_op
> > > *pt_op,
> > > diff --git a/drivers/gpu/drm/xe/xe_migrate.h
> > > b/drivers/gpu/drm/xe/xe_migrate.h
> > > index c3c0740f908d..30c9c990a8b1 100644
> > > --- a/drivers/gpu/drm/xe/xe_migrate.h
> > > +++ b/drivers/gpu/drm/xe/xe_migrate.h
> > > @@ -24,6 +24,7 @@ struct xe_pt;
> > >  struct xe_tile;
> > >  struct xe_vm;
> > >  struct xe_vm_pgtable_update;
> > > +struct xe_vm_pgtable_update_op;
> > >  struct xe_vma;
> > >  
> > >  enum xe_sriov_vf_ccs_rw_ctxs;
> > > @@ -157,6 +158,13 @@ struct dma_fence *xe_migrate_clear(struct
> > > xe_migrate *m,
> > >  
> > >  struct xe_vm *xe_migrate_get_vm(struct xe_migrate *m);
> > >  
> > > +
> > > +void
> > > +xe_migrate_update_pgtables_cpu_execute(struct xe_vm *vm, struct
> > > xe_tile *tile,
> > > +                                      const struct
> > > xe_migrate_pt_update_ops *ops,
> > > +                                      struct
> > > xe_vm_pgtable_update_op
> > > *pt_op,
> > > +                                      int num_ops);
> > > +
> > >  struct dma_fence *
> > >  xe_migrate_update_pgtables(struct xe_migrate *m,
> > >                            struct xe_migrate_pt_update
> > > *pt_update);
> >