[RFC PATCH] drm/xe: Docouple job seqno and lrc seqno
Thomas Hellström
thomas.hellstrom at linux.intel.com
Thu May 23 11:51:03 UTC 2024
On Wed, 2024-05-22 at 13:52 -0700, Matthew Brost wrote:
> Tightly coupling these seqno presents problems if alternative fences
> for
> jobs are used. Best to decouple these for correctness.
>
>
Nit: Imperative language.
> Cc: Thomas Hellström <thomas.hellstrom at linux.intel.com>
> Signed-off-by: Matthew Brost <mathhew.brost at intel.com>
> ---
> drivers/gpu/drm/xe/xe_exec_queue.c | 2 +-
> drivers/gpu/drm/xe/xe_guc_submit.c | 5 +++--
> drivers/gpu/drm/xe/xe_ring_ops.c | 10 +++++-----
> drivers/gpu/drm/xe/xe_sched_job.c | 12 ++++++------
> drivers/gpu/drm/xe/xe_sched_job.h | 5 +++++
> drivers/gpu/drm/xe/xe_sched_job_types.h | 2 ++
> drivers/gpu/drm/xe/xe_trace.h | 7 +++++--
> 7 files changed, 27 insertions(+), 16 deletions(-)
>
> diff --git a/drivers/gpu/drm/xe/xe_exec_queue.c
> b/drivers/gpu/drm/xe/xe_exec_queue.c
> index 395de93579fa..e4607f0e3456 100644
> --- a/drivers/gpu/drm/xe/xe_exec_queue.c
> +++ b/drivers/gpu/drm/xe/xe_exec_queue.c
> @@ -98,7 +98,7 @@ static struct xe_exec_queue
> *__xe_exec_queue_alloc(struct xe_device *xe,
>
> if (xe_exec_queue_is_parallel(q)) {
> q->parallel.composite_fence_ctx =
> dma_fence_context_alloc(1);
> - q->parallel.composite_fence_seqno =
> XE_FENCE_INITIAL_SEQNO;
> + q->parallel.composite_fence_seqno = 0;
> }
>
> return q;
> diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c
> b/drivers/gpu/drm/xe/xe_guc_submit.c
> index 4efb88e3e056..0a6a9471f1ce 100644
> --- a/drivers/gpu/drm/xe/xe_guc_submit.c
> +++ b/drivers/gpu/drm/xe/xe_guc_submit.c
> @@ -973,8 +973,9 @@ guc_exec_queue_timedout_job(struct drm_sched_job
> *drm_job)
> return DRM_GPU_SCHED_STAT_NOMINAL;
> }
>
> - drm_notice(&xe->drm, "Timedout job: seqno=%u, guc_id=%d,
> flags=0x%lx",
> - xe_sched_job_seqno(job), q->guc->id, q->flags);
> + drm_notice(&xe->drm, "Timedout job: seqno=%u, lrc_seqno=%d,
> guc_id=%d, flags=0x%lx",
> + xe_sched_job_seqno(job),
> xe_sched_job_lrc_seqno(job),
> + q->guc->id, q->flags);
> xe_gt_WARN(q->gt, q->flags & EXEC_QUEUE_FLAG_KERNEL,
> "Kernel-submitted job timed out\n");
> xe_gt_WARN(q->gt, q->flags & EXEC_QUEUE_FLAG_VM &&
> !exec_queue_killed(q),
> diff --git a/drivers/gpu/drm/xe/xe_ring_ops.c
> b/drivers/gpu/drm/xe/xe_ring_ops.c
> index a3ca718456f6..78bec455b714 100644
> --- a/drivers/gpu/drm/xe/xe_ring_ops.c
> +++ b/drivers/gpu/drm/xe/xe_ring_ops.c
> @@ -398,7 +398,7 @@ static void emit_job_gen12_gsc(struct
> xe_sched_job *job)
>
> __emit_job_gen12_simple(job, job->q->lrc,
> job->batch_addr[0],
> - xe_sched_job_seqno(job));
> + xe_sched_job_lrc_seqno(job));
> }
>
> static void emit_job_gen12_copy(struct xe_sched_job *job)
> @@ -407,14 +407,14 @@ static void emit_job_gen12_copy(struct
> xe_sched_job *job)
>
> if (xe_sched_job_is_migration(job->q)) {
> emit_migration_job_gen12(job, job->q->lrc,
> - xe_sched_job_seqno(job));
> +
> xe_sched_job_lrc_seqno(job));
> return;
> }
>
> for (i = 0; i < job->q->width; ++i)
> __emit_job_gen12_simple(job, job->q->lrc + i,
> job->batch_addr[i],
> - xe_sched_job_seqno(job));
> +
> xe_sched_job_lrc_seqno(job));
> }
>
> static void emit_job_gen12_video(struct xe_sched_job *job)
> @@ -425,7 +425,7 @@ static void emit_job_gen12_video(struct
> xe_sched_job *job)
> for (i = 0; i < job->q->width; ++i)
> __emit_job_gen12_video(job, job->q->lrc + i,
> job->batch_addr[i],
> - xe_sched_job_seqno(job));
> + xe_sched_job_lrc_seqno(job));
> }
>
> static void emit_job_gen12_render_compute(struct xe_sched_job *job)
> @@ -435,7 +435,7 @@ static void emit_job_gen12_render_compute(struct
> xe_sched_job *job)
> for (i = 0; i < job->q->width; ++i)
> __emit_job_gen12_render_compute(job, job->q->lrc +
> i,
> job->batch_addr[i],
> -
> xe_sched_job_seqno(job));
> + xe_sched_job_lrc_seq
> no(job));
> }
>
> static const struct xe_ring_ops ring_ops_gen12_gsc = {
> diff --git a/drivers/gpu/drm/xe/xe_sched_job.c
> b/drivers/gpu/drm/xe/xe_sched_job.c
> index cd8a2fba5438..b7b1669d0bea 100644
> --- a/drivers/gpu/drm/xe/xe_sched_job.c
> +++ b/drivers/gpu/drm/xe/xe_sched_job.c
> @@ -117,6 +117,7 @@ struct xe_sched_job *xe_sched_job_create(struct
> xe_exec_queue *q,
> err = PTR_ERR(job->fence);
> goto err_sched_job;
> }
> + job->lrc_seqno = job->fence->seqno;
> } else {
> struct dma_fence_array *cf;
>
> @@ -132,6 +133,8 @@ struct xe_sched_job *xe_sched_job_create(struct
> xe_exec_queue *q,
> err = PTR_ERR(fences[j]);
> goto err_fences;
> }
> + if (!j)
> + job->lrc_seqno = fences[0]->seqno;
> }
>
> cf = dma_fence_array_create(q->width, fences,
> @@ -144,10 +147,6 @@ struct xe_sched_job *xe_sched_job_create(struct
> xe_exec_queue *q,
> goto err_fences;
> }
>
> - /* Sanity check */
> - for (j = 0; j < q->width; ++j)
> - xe_assert(job_to_xe(job), cf->base.seqno ==
> fences[j]->seqno);
> -
> job->fence = &cf->base;
> }
>
> @@ -233,7 +232,7 @@ bool xe_sched_job_started(struct xe_sched_job
> *job)
> {
> struct xe_lrc *lrc = job->q->lrc;
>
> - return !__dma_fence_is_later(xe_sched_job_seqno(job),
> + return !__dma_fence_is_later(xe_sched_job_lrc_seqno(job),
> xe_lrc_start_seqno(lrc),
> job->fence->ops);
Fence ops must be the LRC ops here.
> }
> @@ -247,7 +246,8 @@ bool xe_sched_job_completed(struct xe_sched_job
> *job)
> * parallel handshake is done.
> */
>
> - return !__dma_fence_is_later(xe_sched_job_seqno(job),
> xe_lrc_seqno(lrc),
> + return !__dma_fence_is_later(xe_sched_job_lrc_seqno(job),
> + xe_lrc_seqno(lrc),
> job->fence->ops);
Same here.
Otherwise LGTM.
/Thomas
> }
>
> diff --git a/drivers/gpu/drm/xe/xe_sched_job.h
> b/drivers/gpu/drm/xe/xe_sched_job.h
> index c75018f4660d..002c3b5c0a5c 100644
> --- a/drivers/gpu/drm/xe/xe_sched_job.h
> +++ b/drivers/gpu/drm/xe/xe_sched_job.h
> @@ -73,6 +73,11 @@ static inline u32 xe_sched_job_seqno(struct
> xe_sched_job *job)
> return job->fence->seqno;
> }
>
> +static inline u32 xe_sched_job_lrc_seqno(struct xe_sched_job *job)
> +{
> + return job->lrc_seqno;
> +}
> +
> static inline void
> xe_sched_job_add_migrate_flush(struct xe_sched_job *job, u32 flags)
> {
> diff --git a/drivers/gpu/drm/xe/xe_sched_job_types.h
> b/drivers/gpu/drm/xe/xe_sched_job_types.h
> index 5e12724219fd..990ddac55ed6 100644
> --- a/drivers/gpu/drm/xe/xe_sched_job_types.h
> +++ b/drivers/gpu/drm/xe/xe_sched_job_types.h
> @@ -37,6 +37,8 @@ struct xe_sched_job {
> /** @user_fence.value: write back value */
> u64 value;
> } user_fence;
> + /** @lrc_seqno: LRC seqno */
> + u32 lrc_seqno;
> /** @migrate_flush_flags: Additional flush flags for
> migration jobs */
> u32 migrate_flush_flags;
> /** @ring_ops_flush_tlb: The ring ops need to flush TLB
> before payload. */
> diff --git a/drivers/gpu/drm/xe/xe_trace.h
> b/drivers/gpu/drm/xe/xe_trace.h
> index f18ae1c63031..657e2f1f4c31 100644
> --- a/drivers/gpu/drm/xe/xe_trace.h
> +++ b/drivers/gpu/drm/xe/xe_trace.h
> @@ -254,6 +254,7 @@ DECLARE_EVENT_CLASS(xe_sched_job,
>
> TP_STRUCT__entry(
> __field(u32, seqno)
> + __field(u32, lrc_seqno)
> __field(u16, guc_id)
> __field(u32, guc_state)
> __field(u32, flags)
> @@ -264,6 +265,7 @@ DECLARE_EVENT_CLASS(xe_sched_job,
>
> TP_fast_assign(
> __entry->seqno = xe_sched_job_seqno(job);
> + __entry->lrc_seqno =
> xe_sched_job_lrc_seqno(job);
> __entry->guc_id = job->q->guc->id;
> __entry->guc_state =
> atomic_read(&job->q->guc->state);
> @@ -273,8 +275,9 @@ DECLARE_EVENT_CLASS(xe_sched_job,
> __entry->batch_addr = (u64)job-
> >batch_addr[0];
> ),
>
> - TP_printk("fence=%p, seqno=%u, guc_id=%d,
> batch_addr=0x%012llx, guc_state=0x%x, flags=0x%x, error=%d",
> - __entry->fence, __entry->seqno,
> __entry->guc_id,
> + TP_printk("fence=%p, seqno=%u, lrc_seqno=%u,
> guc_id=%d, batch_addr=0x%012llx, guc_state=0x%x, flags=0x%x,
> error=%d",
> + __entry->fence, __entry->seqno,
> + __entry->lrc_seqno, __entry->guc_id,
> __entry->batch_addr, __entry-
> >guc_state,
> __entry->flags, __entry->error)
> );
More information about the Intel-xe
mailing list