Once the GuC context has its scheduling disabled by TDR or kill work item, the queue is taken off the hardware and can no longer touch memory without risking corruption. Invoke drm_dep_queue_kill, which bypasses any remaining job dependencies in the queue and calls run_job immediately for each remaining job. In run_job, if the queue returns drm_dep_queue_is_killed, immediately signal the hardware fence, as the queue can no longer access any memory associated with the fence being signaled.
Signed-off-by: Matthew Brost <[email protected]> --- drivers/gpu/drm/xe/xe_guc_exec_queue_types.h | 2 + drivers/gpu/drm/xe/xe_guc_submit.c | 139 +++++++++++++------ 2 files changed, 97 insertions(+), 44 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_guc_exec_queue_types.h b/drivers/gpu/drm/xe/xe_guc_exec_queue_types.h index cb15e86823d2..72de6d0a754a 100644 --- a/drivers/gpu/drm/xe/xe_guc_exec_queue_types.h +++ b/drivers/gpu/drm/xe/xe_guc_exec_queue_types.h @@ -29,6 +29,8 @@ struct xe_guc_exec_queue { */ #define MAX_STATIC_MSG_TYPE 3 struct xe_sched_msg static_msgs[MAX_STATIC_MSG_TYPE]; + /** @kill_work: Kill work item */ + struct delayed_work kill_work; /** @resume_time: time of last resume */ u64 resume_time; /** @state: GuC specific state for this xe_exec_queue */ diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c b/drivers/gpu/drm/xe/xe_guc_submit.c index 064cf15166b9..58569969b4c7 100644 --- a/drivers/gpu/drm/xe/xe_guc_submit.c +++ b/drivers/gpu/drm/xe/xe_guc_submit.c @@ -1216,6 +1216,8 @@ guc_exec_queue_run_job(struct drm_dep_job *drm_job) trace_xe_sched_job_run(job); if (!killed_or_banned_or_wedged && !xe_sched_job_is_error(job)) { + xe_gt_assert(guc_to_gt(guc), !drm_dep_queue_is_killed(&q->dep_q)); + if (xe_exec_queue_is_multi_queue_secondary(q)) { struct xe_exec_queue *primary = xe_exec_queue_multi_queue_primary(q); @@ -1234,6 +1236,15 @@ guc_exec_queue_run_job(struct drm_dep_job *drm_job) q->ring_ops->emit_job(job); submit_exec_queue(q, job); job->restore_replay = false; + } else if (drm_dep_queue_is_killed(&q->dep_q)) { + xe_sched_job_set_error(job, -ECANCELED); /* fence signal */ + dma_fence_put(job->fence); /* drop the DRM dep reference */ + + /* + * Our queue is off hardware, so fences can be signalled + * immediately with an error. + */ + return ERR_PTR(-ECANCELED); } run_job_out: @@ -1479,29 +1490,21 @@ static void disable_scheduling(struct xe_exec_queue *q, bool immediate) } static enum drm_dep_timedout_stat -guc_exec_queue_timedout_job(struct drm_dep_job *drm_job) +__guc_exec_queue_timedout_job(struct xe_guc *guc, struct xe_exec_queue *q, + struct xe_sched_job *job) { - struct xe_sched_job *job = to_xe_sched_job(drm_job); struct drm_dep_job *tmp_job; - struct xe_exec_queue *q = job->q, *primary; + struct xe_exec_queue *primary = xe_exec_queue_multi_queue_primary(q); struct xe_gpu_scheduler *sched = &q->guc->sched; - struct xe_guc *guc = exec_queue_to_guc(q); const char *process_name = "no process"; struct xe_device *xe = guc_to_xe(guc); int err = -ETIME; pid_t pid = -1; bool wedged = false, skip_timeout_check; - xe_gt_assert(guc_to_gt(guc), !exec_queue_destroyed(q)); - - if (drm_dep_job_is_finished(&job->drm)) - return DRM_DEP_TIMEDOUT_STAT_JOB_SIGNALED; - if (vf_recovery(guc)) return DRM_DEP_TIMEDOUT_STAT_REQUEUE_JOB; - primary = xe_exec_queue_multi_queue_primary(q); - /* Kill the run_job entry point */ if (xe_exec_queue_is_multi_queue(q)) xe_guc_exec_queue_group_stop(q); @@ -1509,7 +1512,7 @@ guc_exec_queue_timedout_job(struct drm_dep_job *drm_job) xe_sched_submission_stop(sched); /* Must check all state after stopping scheduler */ - skip_timeout_check = exec_queue_reset(q) || + skip_timeout_check = !job || exec_queue_reset(q) || exec_queue_killed_or_banned_or_wedged(q); /* Skip timeout check if multi-queue group is banned */ @@ -1603,43 +1606,45 @@ guc_exec_queue_timedout_job(struct drm_dep_job *drm_job) } } - if (q->vm && q->vm->xef) { - process_name = q->vm->xef->process_name; - pid = q->vm->xef->pid; - } + if (job) { + if (q->vm && q->vm->xef) { + process_name = q->vm->xef->process_name; + pid = q->vm->xef->pid; + } - if (!exec_queue_killed(q)) - xe_gt_notice(guc_to_gt(guc), - "Timedout job: seqno=%u, lrc_seqno=%u, guc_id=%d, flags=0x%lx in %s [%d]", - xe_sched_job_seqno(job), xe_sched_job_lrc_seqno(job), - q->guc->id, q->flags, process_name, pid); + if (!exec_queue_killed(q)) + xe_gt_notice(guc_to_gt(guc), + "Timedout job: seqno=%u, lrc_seqno=%u, guc_id=%d, flags=0x%lx in %s [%d]", + xe_sched_job_seqno(job), xe_sched_job_lrc_seqno(job), + q->guc->id, q->flags, process_name, pid); - trace_xe_sched_job_timedout(job); + trace_xe_sched_job_timedout(job); - if (!exec_queue_killed(q)) - xe_devcoredump(q, job, - "Timedout job - seqno=%u, lrc_seqno=%u, guc_id=%d, flags=0x%lx", - xe_sched_job_seqno(job), xe_sched_job_lrc_seqno(job), - q->guc->id, q->flags); + if (!exec_queue_killed(q)) + xe_devcoredump(q, job, + "Timedout job - seqno=%u, lrc_seqno=%u, guc_id=%d, flags=0x%lx", + xe_sched_job_seqno(job), xe_sched_job_lrc_seqno(job), + q->guc->id, q->flags); - /* - * Kernel jobs should never fail, nor should VM jobs if they do - * somethings has gone wrong and the GT needs a reset - */ - 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), - "VM job timed out on non-killed execqueue\n"); - if (!wedged && (q->flags & EXEC_QUEUE_FLAG_KERNEL || - (q->flags & EXEC_QUEUE_FLAG_VM && !exec_queue_killed(q)))) { - if (!xe_sched_invalidate_job(job, 2)) - xe_gt_reset_async(q->gt); - } + /* + * Kernel jobs should never fail, nor should VM jobs if they do + * somethings has gone wrong and the GT needs a reset + */ + 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), + "VM job timed out on non-killed execqueue\n"); + if (!wedged && (q->flags & EXEC_QUEUE_FLAG_KERNEL || + (q->flags & EXEC_QUEUE_FLAG_VM && !exec_queue_killed(q)))) { + if (!xe_sched_invalidate_job(job, 2)) + xe_gt_reset_async(q->gt); + } - /* Mark all outstanding jobs as bad, thus completing them */ - xe_sched_job_set_error(job, err); - drm_dep_queue_for_each_pending_job(tmp_job, sched->dep_q) - xe_sched_job_set_error(to_xe_sched_job(tmp_job), -ECANCELED); + /* Mark all outstanding jobs as bad, thus completing them */ + xe_sched_job_set_error(job, err); + drm_dep_queue_for_each_pending_job(tmp_job, sched->dep_q) + xe_sched_job_set_error(to_xe_sched_job(tmp_job), -ECANCELED); + } if (xe_exec_queue_is_multi_queue(q)) { xe_guc_exec_queue_group_start(q); @@ -1649,6 +1654,10 @@ guc_exec_queue_timedout_job(struct drm_dep_job *drm_job) xe_guc_exec_queue_trigger_cleanup(q); } + /* Queue is off hardware; start flushing jobs bypassing dependencies. */ + drm_dep_queue_kill(&q->dep_q); + cancel_delayed_work(&q->guc->kill_work); + return DRM_DEP_TIMEDOUT_STAT_REQUEUE_JOB; rearm: @@ -1665,6 +1674,43 @@ guc_exec_queue_timedout_job(struct drm_dep_job *drm_job) return DRM_DEP_TIMEDOUT_STAT_REQUEUE_JOB; } +static enum drm_dep_timedout_stat +guc_exec_queue_timedout_job(struct drm_dep_job *drm_job) +{ + struct xe_sched_job *job = to_xe_sched_job(drm_job); + struct xe_exec_queue *q = job->q; + struct xe_guc *guc = exec_queue_to_guc(q); + + xe_gt_assert(guc_to_gt(guc), !exec_queue_destroyed(q)); + + if (drm_dep_job_is_finished(&job->drm)) + return DRM_DEP_TIMEDOUT_STAT_JOB_SIGNALED; + + return __guc_exec_queue_timedout_job(guc, q, job); +} + +static void guc_exec_queue_kill_work_func(struct work_struct *w) +{ + struct xe_guc_exec_queue *ge = + container_of(w, typeof(*ge), kill_work.work); + struct xe_exec_queue *q = ge->q; + struct xe_exec_queue *primary = xe_exec_queue_multi_queue_primary(q); + struct xe_guc *guc = exec_queue_to_guc(q); + + xe_gt_assert(guc_to_gt(guc), exec_queue_killed(q)); + + if (drm_dep_queue_is_killed(&q->dep_q)) + return; + + if (!(exec_queue_enabled(primary) || + exec_queue_pending_disable(primary))) { + drm_dep_queue_kill(&q->dep_q); + return; + } + + __guc_exec_queue_timedout_job(guc, q, NULL); +} + static void guc_exec_queue_fini(struct xe_exec_queue *q) { struct xe_guc_exec_queue *ge = q->guc; @@ -1915,6 +1961,7 @@ static void guc_dep_queue_fini(struct drm_dep_queue *dep_q) { struct xe_exec_queue *q = container_of(dep_q, typeof(*q), dep_q); + cancel_delayed_work(&q->guc->kill_work); xe_exec_queue_destroy(q); } @@ -1949,6 +1996,7 @@ static int guc_exec_queue_init(struct xe_exec_queue *q) q->guc = ge; ge->q = q; init_waitqueue_head(&ge->suspend_wait); + INIT_DELAYED_WORK(&ge->kill_work, guc_exec_queue_kill_work_func); for (i = 0; i < MAX_STATIC_MSG_TYPE; ++i) INIT_LIST_HEAD(&ge->static_msgs[i].link); @@ -2028,6 +2076,9 @@ static void guc_exec_queue_kill(struct xe_exec_queue *q) set_exec_queue_killed(q); __suspend_fence_signal(q); xe_guc_exec_queue_trigger_cleanup(q); + + mod_delayed_work(drm_dep_queue_timeout_wq(&q->dep_q), + &q->guc->kill_work, 2); } static void guc_exec_queue_add_msg(struct xe_exec_queue *q, struct xe_sched_msg *msg, -- 2.34.1
