Re: [Intel-gfx] [PATCH 17/27] drm/i915/guc: Implement multi-lrc reset

On 8/20/2021 15:44, Matthew Brost wrote:

Update context and full GPU reset to work with multi-lrc. The idea is parent context tracks all the active requests inflight for itself and its' children. The parent context owns the reset replaying / canceling

its' -> its

requests as needed.

Signed-off-by: Matthew Brost matthew.brost@intel.com

drivers/gpu/drm/i915/gt/intel_context.c | 11 ++-- .../gpu/drm/i915/gt/uc/intel_guc_submission.c | 63 +++++++++++++------ 2 files changed, 51 insertions(+), 23 deletions(-)

diff --git a/drivers/gpu/drm/i915/gt/intel_context.c b/drivers/gpu/drm/i915/gt/intel_context.c index 00d1aee6d199..5615be32879c 100644 --- a/drivers/gpu/drm/i915/gt/intel_context.c +++ b/drivers/gpu/drm/i915/gt/intel_context.c @@ -528,20 +528,21 @@ struct i915_request *intel_context_create_request(struct intel_context *ce)

struct i915_request *intel_context_find_active_request(struct intel_context *ce) {

• struct intel_context *parent = intel_context_to_parent(ce); struct i915_request *rq, *active = NULL; unsigned long flags;

  GEM_BUG_ON(!intel_engine_uses_guc(ce->engine));

Should this not check the parent as well/instead?

And to be clear, this can be called on regular contexts (where ce == parent) and on both the parent or child contexts of multi-LRC contexts (where ce may or may not match parent)?

• spin_lock_irqsave(&ce->guc_state.lock, flags);
• list_for_each_entry_reverse(rq, &ce->guc_state.requests,

• spin_lock_irqsave(&parent->guc_state.lock, flags);
• list_for_each_entry_reverse(rq, &parent->guc_state.requests, sched.link) {

• if (i915_request_completed(rq))
  

• if (i915_request_completed(rq) && rq->context == ce)
  

'rq->context == ce' means:

1. single-LRC context, rq is owned by ce 2. multi-LRC context, ce is child, rq really belongs to ce but is being tracked by parent 3. multi-LRC context, ce is parent, rq really is owned by ce

So when 'rq->ce != ce', it means that the request is owned by a different child to 'ce' but within the same multi-LRC group. So we want to ignore that request and keep searching until we find one that is really owned by the target ce?

	break;

• active = rq;
  

• active = (rq->context == ce) ? rq : active;
  

Would be clearer to say 'if(rq->ce != ce) continue;' and leave 'active = rq;' alone?

And again, the intention is to ignore requests that are owned by other members of the same multi-LRC group?

Would be good to add some documentation to this function to explain the above (assuming my description is correct?).

}

• spin_unlock_irqrestore(&ce->guc_state.lock, flags);

• spin_unlock_irqrestore(&parent->guc_state.lock, flags);

  return active; }

diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c index f0b60fecf253..e34e0ea9136a 100644 --- a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c +++ b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c @@ -670,6 +670,11 @@ static int rq_prio(const struct i915_request *rq) return rq->sched.attr.priority; }

+static inline bool is_multi_lrc(struct intel_context *ce) +{

• return intel_context_is_parallel(ce);

+}

• static bool is_multi_lrc_rq(struct i915_request *rq) { return intel_context_is_parallel(rq->context);

@@ -1179,10 +1184,13 @@ __unwind_incomplete_requests(struct intel_context *ce)

static void __guc_reset_context(struct intel_context *ce, bool stalled) {

• bool local_stalled; struct i915_request *rq; unsigned long flags; u32 head;
• int i, number_children = ce->guc_number_children;

If this is a child context, does it not need to pull the child count from the parent? Likewise the list/link pointers below? Or does each child context have a full list of its siblings + parent?

bool skip = false;

• struct intel_context *parent = ce;

  intel_context_get(ce);

@@ -1209,25 +1217,34 @@ static void __guc_reset_context(struct intel_context *ce, bool stalled) if (unlikely(skip)) goto out_put;

• rq = intel_context_find_active_request(ce);
• if (!rq) {

• head = ce->ring->tail;
  

• stalled = false;
  

• goto out_replay;
  

• }

• for (i = 0; i < number_children + 1; ++i) {

• if (!intel_context_is_pinned(ce))
  

• 	goto next_context;
  

• local_stalled = false;
  

• rq = intel_context_find_active_request(ce);
  

• if (!rq) {
  

• 	head = ce->ring->tail;
  

• 	goto out_replay;
  

• }
  
  

• if (!i915_request_started(rq))

• stalled = false;
  

• GEM_BUG_ON(i915_active_is_idle(&ce->active));
  

• head = intel_ring_wrap(ce->ring, rq->head);
  
  

• GEM_BUG_ON(i915_active_is_idle(&ce->active));
• head = intel_ring_wrap(ce->ring, rq->head);
• __i915_request_reset(rq, stalled);

• if (i915_request_started(rq))
  

Why change the ordering of the started test versus the wrap/reset call? Is it significant? Why is it now important to be reversed?

• 	local_stalled = true;
  
  

• __i915_request_reset(rq, local_stalled && stalled);
  

  out_replay:

• guc_reset_state(ce, head, stalled);
• __unwind_incomplete_requests(ce);

• guc_reset_state(ce, head, local_stalled && stalled);
  

+next_context:

• if (i != number_children)
  

• 	ce = list_next_entry(ce, guc_child_link);
  

Can this not be put in to the step clause of the for statement?

• }
• __unwind_incomplete_requests(parent); out_put:

• intel_context_put(ce);

• intel_context_put(parent);

As above, I think this function would benefit from some comments to explain exactly what is being done and why.

John.

}

void intel_guc_submission_reset(struct intel_guc *guc, bool stalled) @@ -1248,7 +1265,8 @@ void intel_guc_submission_reset(struct intel_guc *guc, bool stalled)

xa_unlock(&guc->context_lookup);

• if (intel_context_is_pinned(ce))
  

• if (intel_context_is_pinned(ce) &&
  

•     !intel_context_is_child(ce))
  __guc_reset_context(ce, stalled);
  
  

  intel_context_put(ce);

@@ -1340,7 +1358,8 @@ void intel_guc_submission_cancel_requests(struct intel_guc *guc)

xa_unlock(&guc->context_lookup);

• if (intel_context_is_pinned(ce))
  

• if (intel_context_is_pinned(ce) &&
  

•     !intel_context_is_child(ce))
  guc_cancel_context_requests(ce);
  
  

  intel_context_put(ce);

@@ -2031,6 +2050,8 @@ static struct i915_sw_fence *guc_context_block(struct intel_context *ce) u16 guc_id; bool enabled;

• GEM_BUG_ON(intel_context_is_child(ce));

• spin_lock_irqsave(&ce->guc_state.lock, flags);

  incr_context_blocked(ce);

@@ -2068,6 +2089,7 @@ static void guc_context_unblock(struct intel_context *ce) bool enable;

GEM_BUG_ON(context_enabled(ce));

• GEM_BUG_ON(intel_context_is_child(ce));

  spin_lock_irqsave(&ce->guc_state.lock, flags);

@@ -2099,11 +2121,14 @@ static void guc_context_unblock(struct intel_context *ce) static void guc_context_cancel_request(struct intel_context *ce, struct i915_request *rq) {

• struct intel_context *block_context =

• request_to_scheduling_context(rq);
  

• if (i915_sw_fence_signaled(&rq->submit)) { struct i915_sw_fence *fence;

  intel_context_get(ce);

• fence = guc_context_block(ce);
  

• fence = guc_context_block(block_context);
  

  i915_sw_fence_wait(fence); if (!i915_request_completed(rq)) { __i915_request_skip(rq);

@@ -2117,7 +2142,7 @@ static void guc_context_cancel_request(struct intel_context *ce, */ flush_work(&ce_to_guc(ce)->ct.requests.worker);

• guc_context_unblock(ce);
  

• guc_context_unblock(block_context);
  

  intel_context_put(ce); } }

@@ -2143,6 +2168,8 @@ static void guc_context_ban(struct intel_context *ce, struct i915_request *rq) intel_wakeref_t wakeref; unsigned long flags;

• GEM_BUG_ON(intel_context_is_child(ce));

• guc_flush_submissions(guc);

  spin_lock_irqsave(&ce->guc_state.lock, flags);