[idd] transport: track asynchronous frame readiness

idd/LGIdd/capture/CFrameScheduler.cpp

@@ -701,13 +701,12 @@ bool CFrameScheduler::TryFrameSubmitted(const Schedule& schedule,
     publication.deadlineSerial = schedule.deadlineSerial;
     publication.frameSerial    = frameSerial;
     publication.deadline       = schedule.deadline;
-    publication.committed      = true;
     registered                 = true;
   }
   return registered;
 }
 
-void CFrameScheduler::FramePublished(const Schedule& schedule,
+void CFrameScheduler::FrameCommitted(const Schedule& schedule,
   uint32_t frameSerial, uint64_t now, bool periodic)
 {
   CSRWExclusiveLock lock(m_lock);
@@ -721,6 +720,25 @@ void CFrameScheduler::FramePublished(const Schedule& schedule,
     if (publication)
       publication->committed = true;
 
+    if (periodic)
+    {
+      AdvanceCurrentDeadline();
+      AdvanceDeadline(now);
+    }
+    Client * client = FindClient(m_schedule.clientID);
+    if (client)
+      client->nextDelivery = m_nextDeadline;
+  }
+}
+
+void CFrameScheduler::FramePublished(const Schedule& schedule,
+  uint32_t frameSerial)
+{
+  CSRWExclusiveLock lock(m_lock);
+  if (m_scheduling && schedule.clientID == m_schedule.clientID &&
+      schedule.generation == m_schedule.generation &&
+      schedule.epoch == m_schedule.epoch)
+  {
     if (schedule.forceTicket > m_forceAckTicket)
       m_forceAckTicket = schedule.forceTicket;
     if (schedule.republishTicket > m_republishAckTicket)
@@ -735,26 +753,16 @@ void CFrameScheduler::FramePublished(const Schedule& schedule,
       client->deliveredFrameValid      = true;
     }
     ++m_publishedFrames;
-    if (periodic)
-    {
-      AdvanceCurrentDeadline();
-      AdvanceDeadline(now);
-    }
-    if (client)
-      client->nextDelivery = m_nextDeadline;
   }
 }
 
-void CFrameScheduler::FrameRetained(const Schedule& schedule,
+void CFrameScheduler::FrameRetained(const Schedule& schedule)
-  uint64_t now, bool periodic)
 {
   bool wake = false;
   CSRWExclusiveLock lock(m_lock);
   if (m_scheduling && schedule.clientID == m_schedule.clientID &&
       schedule.generation == m_schedule.generation &&
-      schedule.epoch == m_schedule.epoch &&
+      schedule.epoch == m_schedule.epoch)
-      schedule.deadlineSerial == m_deadlineSerial &&
-      schedule.deadline == m_nextDeadline)
   {
     if (schedule.forceTicket > m_forceAckTicket)
       m_forceAckTicket = schedule.forceTicket;
@@ -766,16 +774,6 @@ void CFrameScheduler::FrameRetained(const Schedule& schedule,
       ++m_republishRequestTicket;
       wake = true;
     }
-
-    if (periodic)
-    {
-      AdvanceCurrentDeadline();
-      AdvanceDeadline(now);
-    }
-
-    Client * client = FindClient(m_schedule.clientID);
-    if (client)
-      client->nextDelivery = m_nextDeadline;
   }
   lock.Unlock();
 
@@ -783,6 +781,11 @@ void CFrameScheduler::FrameRetained(const Schedule& schedule,
     WakePublisher();
 }
 
+void CFrameScheduler::FrameFailed()
+{
+  ForceFrame();
+}
+
 bool CFrameScheduler::TryFrameCompleted(const Schedule& schedule,
   uint32_t frameSerial, uint64_t completedAt)
 {

idd/LGIdd/capture/CFrameScheduler.h

@@ -172,10 +172,11 @@ public:
   void FrameMissed(const Schedule& schedule, uint64_t now, bool periodic);
   void FrameSuperseded();
   bool TryFrameSubmitted(const Schedule& schedule, uint32_t frameSerial);
-  void FramePublished(const Schedule& schedule, uint32_t frameSerial,
+  void FrameCommitted(const Schedule& schedule, uint32_t frameSerial,
     uint64_t now, bool periodic);
-  void FrameRetained(const Schedule& schedule, uint64_t now,
+  void FramePublished(const Schedule& schedule, uint32_t frameSerial);
-    bool periodic);
+  void FrameRetained(const Schedule& schedule);
+  void FrameFailed();
   void FrameRepublished(const Schedule& schedule, uint32_t frameSerial);
   bool TryFrameCompleted(const Schedule& schedule, uint32_t frameSerial,
     uint64_t completedAt);

idd/LGIdd/transport/CFrameHub.cpp

@@ -27,6 +27,13 @@ static bool ContentAfter(uint64_t value, uint64_t current)
     static_cast<int64_t>(value - current) > 0);
 }
 
+static bool TokenMatches(const FrameToken& left, const FrameToken& right)
+{
+  return left.backend == right.backend &&
+    left.epoch == right.epoch && left.slot == right.slot &&
+    left.serial == right.serial;
+}
+
 CFrameHub::CFrameHub()
 {
   m_wakeEvent = CreateEvent(nullptr, FALSE, FALSE, nullptr);
@@ -89,8 +96,14 @@ bool CFrameHub::Bind(BackendId backend, uint32_t epoch, bool primary,
     selected->backend.store(backend, std::memory_order_release);
     selected->epoch.store(epoch, std::memory_order_release);
     selected->primary       = primary;
+    selected->outstanding.store(0, std::memory_order_release);
+    SetEvent(selected->drained);
+  }
+  {
+    CSRWExclusiveLock lock(selected->laneLock);
     selected->needsFullCopy = true;
     selected->lastContent   = 0;
+    selected->lastTerminalContent = 0;
     selected->blockedContent   = 0;
     selected->blockedFrameSize = 0;
     selected->blockedAllocation = false;
@@ -101,11 +114,10 @@ bool CFrameHub::Bind(BackendId backend, uint32_t epoch, bool primary,
     selected->frameType     = FRAME_TYPE_INVALID;
     for (Sink::ResourceLane& lane : selected->lanes)
       lane = {};
-    selected->outstanding.store(0, std::memory_order_release);
-    SetEvent(selected->drained);
-    target.SetFrameScheduleEvent(m_wakeEvent);
   }
 
+  target.SetFrameEvents(this);
+  target.SetFrameScheduleEvent(m_wakeEvent);
   {
     CSRWExclusiveLock lock(m_listLock);
     selected->active.store(true, std::memory_order_release);
@@ -135,21 +147,52 @@ void CFrameHub::Unbind(BackendId backend, uint32_t epoch)
   if (!selected)
     return;
 
+  IFrameSink * target;
   {
     CSRWExclusiveLock lock(selected->callLock);
-    if (selected->target)
+    target = selected->target;
-      selected->target->SetFrameScheduleEvent(nullptr);
+  }
+  if (target)
+    target->SetFrameScheduleEvent(nullptr);
+
+  FrameToken cancel[FRAME_SINK_BUFFERS] = {};
+  unsigned cancelCount = 0;
+  {
+    CSRWExclusiveLock lock(selected->laneLock);
+    for (Sink::ResourceLane& lane : selected->lanes)
+    {
+      if (lane.phase == Sink::ResourceLane::PENDING)
+        cancel[cancelCount++] = lane.token;
+      else if (lane.phase == Sink::ResourceLane::FILLING)
+        lane.cancelRequested = true;
+    }
+  }
+  for (unsigned i = 0; i < cancelCount; ++i)
+    for (unsigned lane = 0; lane < FRAME_SINK_BUFFERS; ++lane)
+    {
+      CSRWSharedLock lock(selected->laneLock);
+      if (!TokenMatches(selected->lanes[lane].token, cancel[i]))
+        continue;
+      lock.Unlock();
+      CancelLane(*selected, lane, cancel[i]);
+      break;
     }
 
   WaitForSingleObject(selected->drained, INFINITE);
+  if (target)
+    target->SetFrameEvents(nullptr);
   {
     CSRWExclusiveLock lock(selected->callLock);
     selected->target        = nullptr;
     selected->backend.store(0, std::memory_order_release);
     selected->epoch.store(0, std::memory_order_release);
     selected->primary       = false;
+  }
+  {
+    CSRWExclusiveLock lock(selected->laneLock);
     selected->needsFullCopy = true;
     selected->lastContent   = 0;
+    selected->lastTerminalContent = 0;
     selected->blockedContent   = 0;
     selected->blockedFrameSize = 0;
     selected->blockedAllocation = false;
@@ -186,7 +229,10 @@ void CFrameHub::ReleaseTarget(Batch& batch, BatchTarget& target)
     return;
   target.releasePending = true;
   if (target.resourceLane < FRAME_SINK_BUFFERS)
+  {
+    CSRWExclusiveLock lock(target.sink->laneLock);
     target.sink->lanes[target.resourceLane].busy = false;
+  }
   if (target.sink->outstanding.fetch_sub(
         1, std::memory_order_acq_rel) == 1)
     SetEvent(target.sink->drained);
@@ -199,37 +245,276 @@ void CFrameHub::ReleaseTarget(Batch& batch, BatchTarget& target)
   batch.active = false;
 }
 
-void CFrameHub::CompleteTarget(
+bool CFrameHub::DetachTarget(Batch& batch, BatchTarget& target,
-  Batch& batch, BatchTarget& target, bool succeeded)
+  FrameToken& token)
 {
-  CSRWExclusiveLock call(target.sink->callLock);
+  if (!target.active || target.releasePending ||
-  if (target.sink->target &&
+      target.resourceLane >= FRAME_SINK_BUFFERS)
-      target.sink->backend.load(std::memory_order_acquire) ==
+    return false;
-        target.backend &&
+
-      target.sink->epoch.load(std::memory_order_acquire) == target.epoch)
+  token.backend = target.backend;
+  token.epoch   = target.epoch;
+  token.slot    = target.localSlot;
+  token.serial  = batch.serial;
   {
-    target.sink->target->CompleteFrameBuffer(target.localSlot, succeeded);
+    CSRWExclusiveLock lock(target.sink->laneLock);
-    if (succeeded)
+    Sink::ResourceLane& lane =
-    {
+      target.sink->lanes[target.resourceLane];
-      if (ContentAfter(target.content, target.sink->lastContent))
+    if (!lane.busy || lane.phase != Sink::ResourceLane::IDLE)
-      {
+      return false;
-        target.sink->lastContent   = target.content;
+    lane.token           = token;
-        target.sink->pitch         = target.pitch;
+    lane.schedule        = target.schedule;
-        target.sink->width         = target.width;
+    lane.content         = target.content;
-        target.sink->height        = target.height;
+    lane.captureTime     = target.captureTime;
-        target.sink->format        = target.format;
+    lane.postProcessTime = target.postProcessTime;
-        target.sink->frameType     = target.frameType;
+    lane.copyTime        = target.copyTime;
-        target.sink->needsFullCopy = false;
+    lane.readyTime       = target.readyTime;
-        target.sink->blockedContent   = 0;
+    lane.holdTime        = target.holdTime;
-        target.sink->blockedFrameSize = 0;
+    lane.filledAt        = target.filledAt;
-        target.sink->blockedAllocation = false;
+    lane.workStart       = target.workStart;
+    lane.pitch           = target.pitch;
+    lane.width           = target.width;
+    lane.height          = target.height;
+    lane.format          = target.format;
+    lane.frameType       = target.frameType;
+    lane.phase           = Sink::ResourceLane::FILLING;
+    lane.timingValid     = target.timingValid;
+    lane.resultPending   = false;
+    lane.callActive      = false;
+    lane.cancelRequested = false;
   }
 
+  target.releasePending = true;
+  target.active         = false;
+  for (unsigned i = 0; i < batch.count; ++i)
+    if (batch.targets[i].active)
+      return true;
+  batch.active = false;
+  return true;
+}
+
+void CFrameHub::FillLane(Sink& sink, unsigned laneIndex,
+  const FrameToken& token)
+{
+  if (laneIndex >= FRAME_SINK_BUFFERS)
+    return;
+
+  {
+    CSRWExclusiveLock lock(sink.laneLock);
+    Sink::ResourceLane& lane = sink.lanes[laneIndex];
+    if (lane.phase != Sink::ResourceLane::FILLING ||
+        !TokenMatches(lane.token, token))
+      return;
+    lane.callActive = true;
   }
+
+  IFrameSink * target = sink.target;
+  const FrameFill fill = target ? target->FrameFilled(token) :
+    FrameFill::REJECTED;
+
+  bool      terminal = false;
+  bool      cancel   = false;
+  FrameDone result   = FrameDone::FAILED;
+  uint64_t  readyAt  = 0;
+  {
+    CSRWExclusiveLock lock(sink.laneLock);
+    Sink::ResourceLane& lane = sink.lanes[laneIndex];
+    if (lane.phase != Sink::ResourceLane::FILLING ||
+        !TokenMatches(lane.token, token))
+      return;
+    lane.callActive = false;
+    if (lane.resultPending)
+    {
+      terminal           = true;
+      result             = lane.pendingResult;
+      readyAt            = lane.pendingReadyAt;
+      lane.resultPending = false;
+    }
+    else if (fill == FrameFill::READY)
+    {
+      terminal = true;
+      result   = FrameDone::READY;
+    }
+    else if (fill == FrameFill::REJECTED)
+      terminal = true;
     else
-      target.sink->needsFullCopy = true;
+    {
+      lane.phase = Sink::ResourceLane::PENDING;
+      cancel = lane.cancelRequested ||
+        !sink.active.load(std::memory_order_acquire);
+    }
+  }
+
+  if (terminal)
+    CompleteLane(sink, laneIndex, token, result, readyAt);
+  else if (cancel)
+    CancelLane(sink, laneIndex, token);
+}
+
+void CFrameHub::CancelLane(Sink& sink, unsigned laneIndex,
+  const FrameToken& token)
+{
+  if (laneIndex >= FRAME_SINK_BUFFERS)
+    return;
+
+  {
+    CSRWExclusiveLock lock(sink.laneLock);
+    Sink::ResourceLane& lane = sink.lanes[laneIndex];
+    if (lane.phase != Sink::ResourceLane::PENDING ||
+        !TokenMatches(lane.token, token))
+      return;
+    lane.phase      = Sink::ResourceLane::CANCELING;
+    lane.callActive = true;
+  }
+
+  IFrameSink * target = sink.target;
+  if (target)
+    target->CancelFrame(token);
+
+  FrameDone result   = FrameDone::SUPERSEDED;
+  uint64_t  readyAt  = 0;
+  {
+    CSRWExclusiveLock lock(sink.laneLock);
+    Sink::ResourceLane& lane = sink.lanes[laneIndex];
+    if (lane.phase != Sink::ResourceLane::CANCELING ||
+        !TokenMatches(lane.token, token))
+      return;
+    lane.callActive = false;
+    if (lane.resultPending)
+    {
+      result             = lane.pendingResult;
+      readyAt            = lane.pendingReadyAt;
+      lane.resultPending = false;
+    }
+  }
+  CompleteLane(sink, laneIndex, token, result, readyAt);
+}
+
+void CFrameHub::CompleteLane(Sink& sink, unsigned laneIndex,
+  const FrameToken& token, FrameDone result, uint64_t readyAt)
+{
+  if (laneIndex >= FRAME_SINK_BUFFERS)
+    return;
+
+  Sink::ResourceLane completed;
+  {
+    CSRWExclusiveLock lock(sink.laneLock);
+    Sink::ResourceLane& lane = sink.lanes[laneIndex];
+    if (!lane.busy || lane.phase == Sink::ResourceLane::IDLE ||
+        lane.phase == Sink::ResourceLane::COMPLETING ||
+        !TokenMatches(lane.token, token))
+      return;
+    lane.phase = Sink::ResourceLane::COMPLETING;
+    completed  = lane;
+  }
+
+  if (!readyAt)
+    readyAt = CFrameScheduler::Nanotime();
+  IFrameSink * target = sink.target;
+  if (target)
+  {
+    if (result == FrameDone::READY && completed.timingValid)
+    {
+      uint64_t readyTime = completed.readyTime;
+      if (readyAt >= completed.filledAt)
+        readyTime += readyAt - completed.filledAt;
+      target->SetFrameTiming(completed.localSlot, completed.captureTime,
+        completed.postProcessTime, completed.copyTime, readyTime,
+        completed.holdTime, completed.schedule, readyAt);
+      if (completed.workStart && readyAt >= completed.workStart)
+        target->TryRecordFrameTiming(readyAt - completed.workStart);
+    }
+    target->CompleteFrameBuffer(completed.localSlot, result);
+  }
+
+  {
+    CSRWExclusiveLock lock(sink.laneLock);
+    const bool newestTerminal =
+      ContentAfter(completed.content, sink.lastTerminalContent);
+    if (newestTerminal)
+      sink.lastTerminalContent = completed.content;
+    if (result == FrameDone::READY &&
+        (newestTerminal ||
+         completed.content == sink.lastTerminalContent))
+    {
+      if (ContentAfter(completed.content, sink.lastContent))
+      {
+        sink.lastContent   = completed.content;
+        sink.pitch         = completed.pitch;
+        sink.width         = completed.width;
+        sink.height        = completed.height;
+        sink.format        = completed.format;
+        sink.frameType     = completed.frameType;
+        sink.needsFullCopy = false;
+        sink.blockedContent   = 0;
+        sink.blockedFrameSize = 0;
+        sink.blockedAllocation = false;
+      }
+    }
+    else if (newestTerminal)
+      sink.needsFullCopy = true;
+    Sink::ResourceLane& lane = sink.lanes[laneIndex];
+    if (lane.phase == Sink::ResourceLane::COMPLETING &&
+        TokenMatches(lane.token, token))
+    {
+      lane.token           = {};
+      lane.schedule        = {};
+      lane.content         = 0;
+      lane.phase           = Sink::ResourceLane::IDLE;
+      lane.timingValid     = false;
+      lane.resultPending   = false;
+      lane.callActive      = false;
+      lane.cancelRequested = false;
+      lane.busy            = false;
+    }
+  }
+  if (sink.outstanding.fetch_sub(1, std::memory_order_acq_rel) == 1)
+    SetEvent(sink.drained);
+  SetEvent(m_wakeEvent);
+}
+
+void CFrameHub::OnFrameDone(const FrameToken& token, FrameDone result,
+  uint64_t readyAt)
+{
+  if (!token.backend || !token.epoch || !token.serial)
+    return;
+
+  for (Sink& sink : m_sinks)
+  {
+    if (sink.backend.load(std::memory_order_acquire) != token.backend ||
+        sink.epoch.load(std::memory_order_acquire) != token.epoch)
+      continue;
+
+    unsigned laneIndex = FRAME_SINK_BUFFERS;
+    {
+      CSRWExclusiveLock lock(sink.laneLock);
+      for (unsigned i = 0; i < FRAME_SINK_BUFFERS; ++i)
+      {
+        Sink::ResourceLane& lane = sink.lanes[i];
+        if (!lane.busy || !TokenMatches(lane.token, token) ||
+            lane.phase == Sink::ResourceLane::IDLE ||
+            lane.phase == Sink::ResourceLane::COMPLETING)
+          continue;
+        if (lane.callActive)
+        {
+          if (!lane.resultPending)
+          {
+            lane.pendingResult  = result;
+            lane.pendingReadyAt = readyAt;
+            lane.resultPending  = true;
+          }
+          return;
+        }
+        if (lane.phase == Sink::ResourceLane::PENDING ||
+            lane.phase == Sink::ResourceLane::CANCELING)
+          laneIndex = i;
+        break;
+      }
+    }
+    if (laneIndex < FRAME_SINK_BUFFERS)
+      CompleteLane(sink, laneIndex, token, result, readyAt);
+    return;
   }
-  ReleaseTarget(batch, target);
 }
 
 size_t CFrameHub::GetMaxFrameSize() const
@@ -281,11 +566,12 @@ bool CFrameHub::NeedsFrame() const
     if (!refs[i].sink->active.load(std::memory_order_acquire) ||
         !refs[i].sink->target)
       continue;
+    const size_t maxFrameSize = refs[i].sink->target->GetMaxFrameSize();
+    CSRWSharedLock state(refs[i].sink->laneLock);
     if (refs[i].sink->blockedContent == newest &&
         refs[i].sink->blockedFrameSize &&
         (refs[i].sink->blockedAllocation ||
-         refs[i].sink->target->GetMaxFrameSize() <
+         maxFrameSize < refs[i].sink->blockedFrameSize))
-           refs[i].sink->blockedFrameSize))
       continue;
     if (refs[i].sink->needsFullCopy ||
         ContentAfter(newest, refs[i].sink->lastContent))
@@ -485,18 +771,24 @@ bool CFrameHub::PrepareFrameBatch(const FramePlan& plan,
     const size_t maxFrameSize = sink.target->GetMaxFrameSize();
     if (!maxFrameSize || frameSize > maxFrameSize)
     {
+      CSRWExclusiveLock lock(sink.laneLock);
       sink.blockedContent    = contentSerial;
       sink.blockedFrameSize  = frameSize;
       sink.blockedAllocation = false;
       continue;
     }
 
+    bool forceFull;
+    {
+      CSRWSharedLock lock(sink.laneLock);
       const bool layoutChanged = sink.pitch != pitch ||
-      sink.width != dstFormat.width || sink.height != dstFormat.height ||
+        sink.width != dstFormat.width ||
+        sink.height != dstFormat.height ||
         sink.format != dstFormat.desc.Format ||
         sink.frameType != dstFormat.format;
-    const bool forceFull = sink.needsFullCopy || layoutChanged ||
+      forceFull = sink.needsFullCopy || layoutChanged ||
         sink.lastContent + 1 != contentSerial;
+    }
     const RECT * damage = forceFull ? nullptr : dirtyRects;
     const unsigned damageCount = forceFull ? 0 : nbDirtyRects;
     SinkTarget result = sink.target->PrepareFrameBuffer(
@@ -506,14 +798,19 @@ bool CFrameHub::PrepareFrameBatch(const FramePlan& plan,
     {
       if (result.mem)
       {
+        {
+          CSRWExclusiveLock lock(sink.laneLock);
           sink.blockedContent    = contentSerial;
           sink.blockedFrameSize  = frameSize;
           sink.blockedAllocation = true;
+        }
         sink.target->AbortFrameBuffer(result.slot);
       }
       continue;
     }
 
+    {
+      CSRWExclusiveLock lock(sink.laneLock);
       unsigned resourceLane = FRAME_SINK_BUFFERS;
       for (unsigned lane = 0; lane < FRAME_SINK_BUFFERS; ++lane)
       {
@@ -544,6 +841,7 @@ bool CFrameHub::PrepareFrameBatch(const FramePlan& plan,
           }
       if (resourceLane == FRAME_SINK_BUFFERS)
       {
+        lock.Unlock();
         sink.target->AbortFrameBuffer(result.slot);
         continue;
       }
@@ -555,6 +853,7 @@ bool CFrameHub::PrepareFrameBatch(const FramePlan& plan,
       lane.direct     = request.primary;
       lane.valid      = true;
       lane.busy       = true;
+      lane.phase      = Sink::ResourceLane::IDLE;
 
       unsigned index = batch->count;
       if (sink.primary && index)
@@ -588,7 +887,7 @@ bool CFrameHub::PrepareFrameBatch(const FramePlan& plan,
         ResetEvent(sink.drained);
 
       PreparedFrameBuffer& output = prepared.targets[index];
-    output.token.sink   = request.backend;
+      output.token.backend = request.backend;
       output.token.epoch  = request.epoch;
       output.token.slot   = result.slot;
       output.token.serial = batch->serial;
@@ -600,6 +899,7 @@ bool CFrameHub::PrepareFrameBatch(const FramePlan& plan,
       output.direct       = request.primary;
       output.fullCopy     = forceFull || result.fullCopy;
     }
+  }
   prepared.count = batch->count;
   if (!batch->count)
   {
@@ -636,7 +936,10 @@ uint32_t CFrameHub::PublishFrameBatch(const FrameBatchToken& token)
     {
       if (valid)
         target.sink->target->AbortFrameBuffer(target.localSlot);
+      {
+        CSRWExclusiveLock state(target.sink->laneLock);
         target.sink->needsFullCopy = true;
+      }
       ReleaseTarget(batch, target);
       continue;
     }
@@ -658,6 +961,14 @@ void CFrameHub::CommitFrameBatch(const FrameBatchToken& token)
   if (token.slot >= FRAME_BATCHES)
     return;
   Batch& batch = m_batches[token.slot];
+  struct FillRequest
+  {
+    Sink       * sink;
+    unsigned     lane;
+    FrameToken  token;
+    bool        succeeded;
+  } fill[FRAME_MAX_SINKS] = {};
+  unsigned fillCount = 0;
   CSRWExclusiveLock lock(batch.lock);
   if (!BatchValid(batch, token))
     return;
@@ -678,8 +989,22 @@ void CFrameHub::CommitFrameBatch(const FrameBatchToken& token)
     }
     target.committed = true;
     if (target.completionPending)
-      CompleteTarget(batch, target, target.completionSucceeded);
+    {
+      FrameToken frameToken;
+      const unsigned lane = target.resourceLane;
+      Sink * sink = target.sink;
+      if (DetachTarget(batch, target, frameToken))
+        fill[fillCount++] = {
+          sink, lane, frameToken, target.completionSucceeded };
     }
+  }
+  lock.Unlock();
+  for (unsigned i = 0; i < fillCount; ++i)
+    if (fill[i].succeeded)
+      FillLane(*fill[i].sink, fill[i].lane, fill[i].token);
+    else
+      CompleteLane(*fill[i].sink, fill[i].lane, fill[i].token,
+        FrameDone::FAILED, CFrameScheduler::Nanotime());
 }
 
 void CFrameHub::AbortFrameBatch(const FrameBatchToken& token)
@@ -701,7 +1026,10 @@ void CFrameHub::AbortFrameBatch(const FrameBatchToken& token)
           target.backend &&
         target.sink->epoch.load(std::memory_order_acquire) == target.epoch)
       target.sink->target->AbortFrameBuffer(target.localSlot);
+    {
+      CSRWExclusiveLock state(target.sink->laneLock);
       target.sink->needsFullCopy = true;
+    }
     ReleaseTarget(batch, target);
   }
 }
@@ -730,7 +1058,10 @@ void CFrameHub::FailFrameBatch(const FrameBatchToken& token)
       else
         target.sink->target->AbortFrameBuffer(target.localSlot);
     }
+    {
+      CSRWExclusiveLock state(target.sink->laneLock);
       target.sink->needsFullCopy = true;
+    }
     ReleaseTarget(batch, target);
   }
 }
@@ -804,19 +1135,18 @@ void CFrameHub::SetFrameTargetTiming(const FrameBatchToken& token,
   if (token.slot >= FRAME_BATCHES)
     return;
   Batch& batch = m_batches[token.slot];
-  CSRWSharedLock lock(batch.lock);
+  CSRWExclusiveLock lock(batch.lock);
   if (!BatchValid(batch, token) || index >= batch.count ||
       !batch.targets[index].active)
     return;
   BatchTarget& target = batch.targets[index];
-  CSRWSharedLock call(target.sink->callLock);
+  target.captureTime     = captureTime;
-  if (target.sink->target &&
+  target.postProcessTime = postProcessTime;
-      target.sink->backend.load(std::memory_order_acquire) ==
+  target.copyTime        = copyTime;
-        target.backend &&
+  target.readyTime       = readyTime;
-      target.sink->epoch.load(std::memory_order_acquire) == target.epoch)
+  target.holdTime        = holdTime;
-    target.sink->target->SetFrameTiming(target.localSlot, captureTime,
+  target.filledAt        = completedAt;
-      postProcessTime, copyTime, readyTime, holdTime,
+  target.timingValid     = true;
-      target.schedule, completedAt);
 }
 
 void CFrameHub::TryRecordFrameTiming(const FrameBatchToken& token,
@@ -825,17 +1155,13 @@ void CFrameHub::TryRecordFrameTiming(const FrameBatchToken& token,
   if (token.slot >= FRAME_BATCHES)
     return;
   Batch& batch = m_batches[token.slot];
-  CSRWSharedLock lock(batch.lock);
+  CSRWExclusiveLock lock(batch.lock);
   if (!BatchValid(batch, token) || index >= batch.count ||
       !batch.targets[index].active)
     return;
   BatchTarget& target = batch.targets[index];
-  CSRWSharedLock call(target.sink->callLock);
+  if (target.filledAt >= duration)
-  if (target.sink->target &&
+    target.workStart = target.filledAt - duration;
-      target.sink->backend.load(std::memory_order_acquire) ==
-        target.backend &&
-      target.sink->epoch.load(std::memory_order_acquire) == target.epoch)
-    target.sink->target->TryRecordFrameTiming(duration);
 }
 
 void CFrameHub::CompleteFrameTarget(const FrameBatchToken& token,
@@ -855,7 +1181,17 @@ void CFrameHub::CompleteFrameTarget(const FrameBatchToken& token,
     target.completionSucceeded = succeeded;
     return;
   }
-  CompleteTarget(batch, target, succeeded);
+  FrameToken frameToken;
+  const unsigned lane = target.resourceLane;
+  Sink * sink = target.sink;
+  if (!DetachTarget(batch, target, frameToken))
+    return;
+  lock.Unlock();
+  if (succeeded)
+    FillLane(*sink, lane, frameToken);
+  else
+    CompleteLane(*sink, lane, frameToken, FrameDone::FAILED,
+      CFrameScheduler::Nanotime());
 }
 
 void CFrameHub::ObserveFrame(uint64_t now)

idd/LGIdd/transport/CFrameHub.h

@@ -27,22 +27,55 @@
 
 #include <atomic>
 
-class CFrameHub final : public IFrameTransport
+class CFrameHub final : public IFrameTransport, public IFrameEvents
 {
 private:
   struct Sink
   {
     struct ResourceLane
     {
+      enum Phase
+      {
+        IDLE,
+        FILLING,
+        PENDING,
+        CANCELING,
+        COMPLETING,
+      };
+
       uint8_t  * mem        = nullptr;
       uint64_t   heapOffset = 0;
       unsigned   localSlot  = 0;
       bool       direct     = false;
       bool       valid      = false;
       bool       busy       = false;
+
+      FrameToken                 token       = {};
+      CFrameScheduler::Schedule schedule    = {};
+      uint64_t                   content     = 0;
+      uint64_t                   captureTime = 0;
+      uint64_t                   postProcessTime = 0;
+      uint64_t                   copyTime    = 0;
+      uint64_t                   readyTime   = 0;
+      uint64_t                   holdTime    = 0;
+      uint64_t                   filledAt    = 0;
+      uint64_t                   workStart   = 0;
+      unsigned                   pitch       = 0;
+      unsigned                   width       = 0;
+      unsigned                   height      = 0;
+      DXGI_FORMAT                format      = DXGI_FORMAT_UNKNOWN;
+      FrameType                  frameType   = FRAME_TYPE_INVALID;
+      Phase                      phase       = IDLE;
+      FrameDone                  pendingResult = FrameDone::FAILED;
+      uint64_t                   pendingReadyAt = 0;
+      bool                       timingValid = false;
+      bool                       resultPending = false;
+      bool                       callActive  = false;
+      bool                       cancelRequested = false;
     };
 
     CSRWLock          callLock;
+    CSRWLock          laneLock;
     IFrameSink      * target        = nullptr;
     HANDLE            drained       = nullptr;
     std::atomic<unsigned> outstanding = 0;
@@ -53,6 +86,7 @@ private:
     bool               primary       = false;
     bool               needsFullCopy = true;
     uint64_t           lastContent   = 0;
+    uint64_t           lastTerminalContent = 0;
     uint64_t           blockedContent = 0;
     size_t             blockedFrameSize = 0;
     bool               blockedAllocation = false;
@@ -87,6 +121,14 @@ private:
     bool                      completionPending   = false;
     bool                      completionSucceeded = false;
     bool                      releasePending = false;
+    uint64_t                  captureTime = 0;
+    uint64_t                  postProcessTime = 0;
+    uint64_t                  copyTime    = 0;
+    uint64_t                  readyTime   = 0;
+    uint64_t                  holdTime    = 0;
+    uint64_t                  filledAt    = 0;
+    uint64_t                  workStart   = 0;
+    bool                      timingValid = false;
     bool                      active      = false;
   };
 
@@ -117,7 +159,14 @@ private:
   unsigned Snapshot(SinkRef refs[FRAME_MAX_SINKS]) const;
   bool BatchValid(const Batch& batch, const FrameBatchToken& token) const;
   void ReleaseTarget(Batch& batch, BatchTarget& target);
-  void CompleteTarget(Batch& batch, BatchTarget& target, bool succeeded);
+  bool DetachTarget(Batch& batch, BatchTarget& target,
+    FrameToken& token);
+  void FillLane(Sink& sink, unsigned laneIndex,
+    const FrameToken& token);
+  void CancelLane(Sink& sink, unsigned laneIndex,
+    const FrameToken& token);
+  void CompleteLane(Sink& sink, unsigned laneIndex,
+    const FrameToken& token, FrameDone result, uint64_t readyAt);
 
 public:
   CFrameHub();
@@ -130,6 +179,9 @@ public:
     IFrameSink& sink);
   void Unbind(BackendId backend, uint32_t epoch);
 
+  void OnFrameDone(const FrameToken& token, FrameDone result,
+    uint64_t readyAt) override;
+
   size_t GetMaxFrameSize() const override;
   uint64_t NextContentSerial() override;
   void FrameProductReady(uint64_t contentSerial) override;

idd/LGIdd/transport/IFrameSink.h

@@ -28,11 +28,22 @@
 #include <stddef.h>
 #include <stdint.h>
 
+class IFrameEvents
+{
+public:
+  virtual ~IFrameEvents() = default;
+
+  virtual void OnFrameDone(const FrameToken& token, FrameDone result,
+    uint64_t readyAt) = 0;
+};
+
 class IFrameSink
 {
 public:
   virtual ~IFrameSink() = default;
 
+  // Clearing events waits for callbacks through the previous pointer to end.
+  virtual void SetFrameEvents(IFrameEvents * events) = 0;
   virtual size_t GetMaxFrameSize() const = 0;
 
   virtual bool FrameBufferAvailable(
@@ -60,7 +71,12 @@ public:
     bool deliveredToOwner) = 0;
   virtual void AbortFrameBuffer(unsigned slot) = 0;
   virtual void FailFrameBuffer(unsigned slot) = 0;
-  virtual void CompleteFrameBuffer(unsigned slot, bool succeeded) = 0;
+  // A terminal callback raised before this returns takes precedence over the
+  // return value.
+  virtual FrameFill FrameFilled(const FrameToken& token) = 0;
+  // On return, the sink no longer accesses the frame identified by token.
+  virtual void CancelFrame(const FrameToken& token) = 0;
+  virtual void CompleteFrameBuffer(unsigned slot, FrameDone result) = 0;
   virtual void SetFrameTiming(unsigned slot, uint64_t captureTime,
     uint64_t postProcessTime, uint64_t copyTime, uint64_t readyTime,
     uint64_t holdTime, const CFrameScheduler::Schedule& schedule,

idd/LGIdd/transport/PreparedFrameBuffer.h

@@ -33,12 +33,26 @@ enum : unsigned
 
 struct FrameToken
 {
-  uint32_t sink   = 0;
+  uint32_t backend = 0;
   uint32_t epoch  = 0;
   uint32_t slot   = 0;
   uint64_t serial = 0;
 };
 
+enum class FrameFill
+{
+  READY,
+  PENDING,
+  REJECTED,
+};
+
+enum class FrameDone
+{
+  READY,
+  FAILED,
+  SUPERSEDED,
+};
+
 struct PreparedFrameBuffer
 {
   FrameToken token;

idd/LGIdd/transport/lgmp/CLGMPFrameTransport.cpp

@@ -202,6 +202,7 @@ bool CLGMPFrameTransport::Setup(size_t alignSize)
   m_readyFrameIndex.store(-1, std::memory_order_release);
   m_deferredOwnerFrameIndex = -1;
   m_framePublishSequence = 0;
+  m_frameReadySequence   = 0;
   memset(m_frameLastPublishSequence, 0,
     sizeof(m_frameLastPublishSequence));
   for (FrameDelivery& delivery : m_frameDelivery)
@@ -222,6 +223,7 @@ void CLGMPFrameTransport::DeInit()
     m_readyFrameIndex.store(-1, std::memory_order_release);
     m_deferredOwnerFrameIndex = -1;
     m_framePublishSequence = 0;
+    m_frameReadySequence   = 0;
     memset(m_frameLastPublishSequence, 0,
       sizeof(m_frameLastPublishSequence));
     memset(m_frameCompleted, 0, sizeof(m_frameCompleted));
@@ -845,6 +847,8 @@ SinkTarget CLGMPFrameTransport::PrepareFrameBuffer(
     flags |= FRAME_FLAG_TRUNCATED;
 
   fi->formatVer        = m_formatVer;
+  if (!m_frameSerial)
+    ++m_frameSerial;
   fi->frameSerial      = m_frameSerial++;
   fi->screenWidth      = srcFormat.width;
   fi->screenHeight     = srcFormat.height;
@@ -1008,6 +1012,8 @@ bool CLGMPFrameTransport::PublishFrameBuffer(unsigned frameIndex,
   if (published)
   {
     m_frameLastPublishSequence[frameIndex] = ++m_framePublishSequence;
+    m_frameSchedule[frameIndex]   = schedule;
+    m_frameDelivered[frameIndex]  = deliveredToOwner;
     m_deferredOwnerFrameIndex = schedule.clientID && !deliveredToOwner ?
       static_cast<LONG>(frameIndex) : -1;
     m_submittedFrameIndex.store(
@@ -1122,15 +1128,14 @@ void CLGMPFrameTransport::CommitFrameBuffer(unsigned frameIndex,
   const CFrameScheduler::Schedule& schedule, bool periodic,
   bool deliveredToOwner)
 {
+  UNREFERENCED_PARAMETER(deliveredToOwner);
+
   if (frameIndex >= LGMP_Q_FRAME_BUFFER_LEN)
     return;
 
   const uint64_t now = CFrameScheduler::Nanotime();
-  if (deliveredToOwner)
+  m_frameScheduler.FrameCommitted(
-    m_frameScheduler.FramePublished(
     schedule, m_frame[frameIndex]->frameSerial, now, periodic);
-  else
-    m_frameScheduler.FrameRetained(schedule, now, periodic);
 }
 
 bool CLGMPFrameTransport::TryFrameSubmitted(unsigned frameIndex,
@@ -1200,16 +1205,21 @@ void CLGMPFrameTransport::FailFrameBuffer(unsigned frameIndex)
   InterlockedExchange(
     (volatile LONG *)&m_frame[frameIndex]->timingValid, 0);
   FinalizeFrameBuffer(frameIndex);
-  CompleteFrameBuffer(frameIndex, false);
+  AbortFrameBuffer(frameIndex);
 }
 
 void CLGMPFrameTransport::CompleteFrameBuffer(
-  unsigned frameIndex, bool succeeded)
+  unsigned frameIndex, FrameDone result)
 {
   if (frameIndex >= LGMP_Q_FRAME_BUFFER_LEN)
     return;
 
+  const bool succeeded = result == FrameDone::READY;
   CSRWExclusiveLock lock(m_framePublishLock);
+  const CFrameScheduler::Schedule schedule = m_frameSchedule[frameIndex];
+  const uint32_t frameSerial = m_frame[frameIndex]->frameSerial;
+  const bool delivered = m_frameDelivered[frameIndex];
+  const uint64_t sequence = m_frameLastPublishSequence[frameIndex];
   m_frameCompleted[frameIndex] = succeeded;
   if (!succeeded &&
       m_deferredOwnerFrameIndex == static_cast<LONG>(frameIndex))
@@ -1218,7 +1228,6 @@ void CLGMPFrameTransport::CompleteFrameBuffer(
   {
     // Completion callbacks may run out of order. Never replace a newer ready
     // frame with an older submission.
-    const uint64_t sequence = m_frameLastPublishSequence[frameIndex];
     const LONG readyFrameIndex =
       m_readyFrameIndex.load(std::memory_order_acquire);
     if (sequence &&
@@ -1228,6 +1237,22 @@ void CLGMPFrameTransport::CompleteFrameBuffer(
         static_cast<LONG>(frameIndex), std::memory_order_release);
   }
   m_frameInFlight[frameIndex].store(false, std::memory_order_release);
+  const bool newerThanReady =
+    sequence && sequence > m_frameReadySequence;
+  if (result == FrameDone::READY && newerThanReady)
+    m_frameReadySequence = sequence;
+
+  if (result == FrameDone::READY && newerThanReady)
+  {
+    if (delivered)
+      m_frameScheduler.FramePublished(schedule, frameSerial);
+    else
+      m_frameScheduler.FrameRetained(schedule);
+  }
+  else if (result == FrameDone::FAILED && newerThanReady)
+    m_frameScheduler.FrameFailed();
+  else if (result == FrameDone::SUPERSEDED && newerThanReady)
+    m_frameScheduler.FrameSuperseded();
 }
 
 void CLGMPFrameTransport::SetFrameTiming(unsigned frameIndex,

idd/LGIdd/transport/lgmp/CLGMPFrameTransport.h

@@ -101,7 +101,11 @@ private:
   bool              m_frameCompleted[LGMP_Q_FRAME_BUFFER_LEN] = {};
   CSRWLock          m_framePublishLock;
   uint64_t          m_framePublishSequence = 0;
+  uint64_t          m_frameReadySequence   = 0;
   uint64_t          m_frameLastPublishSequence[LGMP_Q_FRAME_BUFFER_LEN] = {};
+  CFrameScheduler::Schedule
+                    m_frameSchedule[LGMP_Q_FRAME_BUFFER_LEN] = {};
+  bool              m_frameDelivered[LGMP_Q_FRAME_BUFFER_LEN] = {};
 
   FrameDelivery m_frameDelivery[LGMP_Q_FRAME_BUFFER_LEN] = {};
   OwnerDelivery m_ownerDelivery[LGMP_Q_FRAME_LEN]        = {};
@@ -160,6 +164,7 @@ public:
   CLGMPFrameTransport(const CLGMPFrameTransport&) = delete;
   CLGMPFrameTransport& operator=(const CLGMPFrameTransport&) = delete;
 
+  void SetFrameEvents(IFrameEvents *) override {}
   size_t GetMaxFrameSize() const override { return m_maxFrameSize; }
 
   bool FrameBufferAvailable(const CFrameScheduler::Schedule& schedule,
@@ -189,8 +194,12 @@ public:
     bool deliveredToOwner) override;
   void AbortFrameBuffer(unsigned frameIndex) override;
   void FailFrameBuffer(unsigned frameIndex) override;
-  void CompleteFrameBuffer(
+  FrameFill FrameFilled(const FrameToken&) override
-    unsigned frameIndex, bool succeeded) override;
+  {
+    return FrameFill::READY;
+  }
+  void CancelFrame(const FrameToken&) override {}
+  void CompleteFrameBuffer(unsigned frameIndex, FrameDone result) override;
   void SetFrameTiming(unsigned frameIndex, uint64_t captureTime,
     uint64_t postProcessTime, uint64_t copyTime, uint64_t readyTime,
     uint64_t holdTime, const CFrameScheduler::Schedule& schedule,