gpu: GPU service scheduler.

gpu/command_buffer/service/scheduler.cc

+// Copyright (c) 2017 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "gpu/command_buffer/service/scheduler.h"
+
+#include <algorithm>
+
+#include "base/callback.h"
+#include "base/memory/ptr_util.h"
+#include "base/stl_util.h"
+#include "base/trace_event/trace_event.h"
+#include "base/trace_event/trace_event_argument.h"
+#include "gpu/command_buffer/service/sync_point_manager.h"
+
+namespace gpu {
+
+class Scheduler::Sequence {
+ public:
+  Sequence(SequenceId sequence_id,
+           GpuStreamPriority priority,
+           scoped_refptr<SyncPointOrderData> order_data);
+
+  ~Sequence();
+
+  void Destroy();
+
+  bool destroyed() const { return destroyed_; }
+
+  SequenceId sequence_id() const { return sequence_id_; }
+
+  const SchedulingState& scheduling_state() const { return scheduling_state_; }
+
+  bool enabled() const { return enabled_; }
+
+  bool scheduled() const { return running_state_ == SCHEDULED; }
+
+  bool running() const { return running_state_ == RUNNING; }
+
+  // The sequence is runnable if its enabled and has tasks which are not blocked
+  // by wait fences.
+  bool IsRunnable() const;
+
+  bool NeedsRescheduling() const;
+
+  void UpdateSchedulingState();
+
+  // If this sequence runs before the other sequence.
+  bool RunsBefore(const Sequence* other) const;
+
+  void SetEnabled(bool enabled);
+
+  // Sets running state to SCHEDULED.
+  void SetScheduled();
+
+  // Called before running the next task on the sequence. Returns the closure
+  // for the task. Sets running state to RUNNING.
+  base::OnceClosure BeginTask();
+
+  // Called after running the closure returned by BeginTask. Sets running state
+  // to IDLE.
+  void FinishTask();
+
+  // Enqueues a task in the sequence and returns the generated order number.
+  uint32_t ScheduleTask(base::OnceClosure closure);
+
+  // Continue running the current task with the given closure. Must be called in
+  // between |BeginTask| and |FinishTask|.
+  void ContinueTask(base::OnceClosure closure);
+
+  // Add a sync token fence that this sequence should wait on.
+  void AddWaitFence(const SyncToken& sync_token, uint32_t order_num);
+
+  // Remove a waiting sync token fence.
+  void RemoveWaitFence(const SyncToken& sync_token, uint32_t order_num);
+
+  // Add a sync token fence that this sequence is expected to release.
+  void AddReleaseFence(const SyncToken& sync_token, uint32_t order_num);
+
+  // Remove a release sync token fence.
+  void RemoveReleaseFence(const SyncToken& sync_token, uint32_t order_num);
+
+ private:
+  enum RunningState { IDLE, SCHEDULED, RUNNING };
+
+  struct Fence {
+    SyncToken sync_token;
+    uint32_t order_num;
+
+    bool operator==(const Fence& other) const {
+      return std::tie(sync_token, order_num) ==
+             std::tie(other.sync_token, other.order_num);
+    }
+  };
+
+  struct Task {
+    base::OnceClosure closure;
+    uint32_t order_num;
+  };
+
+  GpuStreamPriority GetSchedulingPriority() const;
+
+  bool destroyed_ = false;
+
+  // If the sequence is enabled. Sequences are disabled/enabled based on when
+  // the command buffer is descheduled/sc

[piman, 2017/05/10 00:38:50]

nit: "/sc" typo or missing end of comment?

[sunnyps, 2017/05/10 23:15:15]

Done.

+  bool enabled_ = true;
+
+  RunningState running_state_;
+
+  // Cached scheduling state used for comparison with other sequences using
+  // |RunsBefore|. Updated in |UpdateSchedulingState|.
+  SchedulingState scheduling_state_;
+
+  const SequenceId sequence_id_;
+
+  const GpuStreamPriority priority_;
+
+  scoped_refptr<SyncPointOrderData> order_data_;
+
+  // Deque of tasks. Tasks are inserted at the back with increasing order number
+  // generated from SyncPointOrderData. If a running task needs to be continued,
+  // it is inserted at the front with the same order number.
+  std::deque<Task> tasks_;
+
+  // List of fences that this sequence is waiting on. Fences are inserted in
+  // increasing order number but may be removed out of order. Tasks are blocked
+  // if there's a wait fence with order number less than or equal to the task's
+  // order number.
+  std::vector<Fence> wait_fences_;
+
+  // List of fences that this sequence is expected to release. If this list is
+  // non-empty, the priority of the sequence is raised.
+  std::vector<Fence> release_fences_;
+
+  DISALLOW_COPY_AND_ASSIGN(Sequence);
+};
+
+Scheduler::SchedulingState::SchedulingState() = default;
+Scheduler::SchedulingState::SchedulingState(const SchedulingState& other) =
+    default;
+Scheduler::SchedulingState::~SchedulingState() = default;
+
+std::unique_ptr<base::trace_event::ConvertableToTraceFormat>
+Scheduler::SchedulingState::AsValue() const {
+  std::unique_ptr<base::trace_event::TracedValue> state(
+      new base::trace_event::TracedValue());
+  state->SetInteger("sequence_id", sequence_id.GetUnsafeValue());
+  state->SetString("priority", GpuStreamPriorityToString(priority));
+  state->SetInteger("order_num", order_num);
+  return std::move(state);
+}
+
+Scheduler::Sequence::Sequence(SequenceId sequence_id,
+                              GpuStreamPriority priority,
+                              scoped_refptr<SyncPointOrderData> order_data)
+    : sequence_id_(sequence_id), priority_(priority), order_data_(order_data) {}
+
+Scheduler::Sequence::~Sequence() = default;
+
+void Scheduler::Sequence::Destroy() {
+  DCHECK(!destroyed_);
+  destroyed_ = true;
+}
+
+bool Scheduler::Sequence::NeedsRescheduling() const {
+  return running_state_ != IDLE &&
+         scheduling_state_.priority != GetSchedulingPriority();
+}
+
+bool Scheduler::Sequence::IsRunnable() const {
+  return enabled_ && !tasks_.empty() &&
+         (wait_fences_.empty() ||
+          wait_fences_.front().order_num > tasks_.front().order_num);
+}
+
+GpuStreamPriority Scheduler::Sequence::GetSchedulingPriority() const {
+  if (!release_fences_.empty())
+    return std::min(priority_, GpuStreamPriority::HIGH);
+  return priority_;
+}
+
+bool Scheduler::Sequence::RunsBefore(const Scheduler::Sequence* other) const {
+  return other->scheduling_state() < scheduling_state_;
+}
+
+void Scheduler::Sequence::SetEnabled(bool enabled) {
+  if (enabled_ == enabled)
+    return;
+  DCHECK_EQ(running_state_, enabled ? IDLE : RUNNING);
+  enabled_ = enabled;
+}
+
+void Scheduler::Sequence::SetScheduled() {
+  DCHECK_NE(running_state_, RUNNING);
+  running_state_ = SCHEDULED;
+  UpdateSchedulingState();
+}
+
+void Scheduler::Sequence::UpdateSchedulingState() {
+  scheduling_state_.sequence_id = sequence_id_;
+  scheduling_state_.priority = GetSchedulingPriority();
+
+  uint32_t order_num = UINT32_MAX;  // IDLE
+  if (running_state_ == SCHEDULED) {
+    DCHECK(!tasks_.empty());
+    order_num = tasks_.front().order_num;
+  } else if (running_state_ == RUNNING) {
+    order_num = order_data_->current_order_num();
+  }
+  scheduling_state_.order_num = order_num;
+}
+
+void Scheduler::Sequence::ContinueTask(base::OnceClosure closure) {
+  DCHECK_EQ(running_state_, RUNNING);
+  tasks_.push_front({std::move(closure), order_data_->current_order_num()});
+}
+
+uint32_t Scheduler::Sequence::ScheduleTask(base::OnceClosure closure) {
+  uint32_t order_num = order_data_->GenerateUnprocessedOrderNumber();
+  tasks_.push_back({std::move(closure), order_num});
+  return order_num;
+}
+
+base::OnceClosure Scheduler::Sequence::BeginTask() {
+  DCHECK(!tasks_.empty());
+
+  DCHECK_EQ(running_state_, SCHEDULED);
+  running_state_ = RUNNING;
+
+  base::OnceClosure closure = std::move(tasks_.front().closure);
+  uint32_t order_num = tasks_.front().order_num;
+  tasks_.pop_front();
+
+  order_data_->BeginProcessingOrderNumber(order_num);
+
+  UpdateSchedulingState();
+
+  return closure;
+}
+
+void Scheduler::Sequence::FinishTask() {
+  DCHECK_EQ(running_state_, RUNNING);
+  running_state_ = IDLE;

[piman, 2017/05/10 00:38:50]

nit: could be premature optimization, but in Scheduler::RunNextTask, we call
FinishTask, then if (IsRunnable()) SetScheduled()
Would it make sense to combine it here, and set the state to SCHEDULED directly
if IsRunnable() ? That way we avoid calling UpdateSchedulingState twice. Little
things, not sure if they matter.

[sunnyps, 2017/05/10 23:15:15]

I kept it this way because I wanted to be explicit that SCHEDULED means that the
sequence is in the scheduling queue.

+  uint32_t order_num = order_data_->current_order_num();
+  if (!tasks_.empty() && tasks_.front().order_num == order_num) {
+    order_data_->PauseProcessingOrderNumber(order_num);
+  } else {
+    order_data_->FinishProcessingOrderNumber(order_num);
+  }
+  UpdateSchedulingState();
+}
+
+void Scheduler::Sequence::AddWaitFence(const SyncToken& sync_token,
+                                       uint32_t order_num) {
+  wait_fences_.push_back({sync_token, order_num});
+}
+
+void Scheduler::Sequence::RemoveWaitFence(const SyncToken& sync_token,
+                                          uint32_t order_num) {
+  base::Erase(wait_fences_, Fence{sync_token, order_num});
+}
+
+void Scheduler::Sequence::AddReleaseFence(const SyncToken& sync_token,
+                                          uint32_t order_num) {
+  release_fences_.push_back({sync_token, order_num});
+}
+
+void Scheduler::Sequence::RemoveReleaseFence(const SyncToken& sync_token,
+                                             uint32_t order_num) {
+  base::Erase(release_fences_, Fence{sync_token, order_num});
+}
+
+Scheduler::Scheduler(scoped_refptr<base::SingleThreadTaskRunner> task_runner,
+                     SyncPointManager* sync_point_manager)
+    : task_runner_(std::move(task_runner)),
+      sync_point_manager_(sync_point_manager),
+      weak_factory_(this) {
+  DCHECK(thread_checker_.CalledOnValidThread());
+}
+
+Scheduler::~Scheduler() {
+  DCHECK(thread_checker_.CalledOnValidThread());
+}
+
+SequenceId Scheduler::CreateSequence(GpuStreamPriority priority) {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  base::AutoLock auto_lock(lock_);
+  scoped_refptr<SyncPointOrderData> order_data =
+      sync_point_manager_->CreateSyncPointOrderData();
+  SequenceId sequence_id = order_data->sequence_id();
+  std::unique_ptr<Sequence> sequence =

[piman, 2017/05/10 00:38:50]

nit: it's ok to use auto when using MakeUnique (that's more or less the point of
MakeUnique).

[sunnyps, 2017/05/10 23:15:15]

Done.

+      base::MakeUnique<Sequence>(sequence_id, priority, std::move(order_data));
+  sequences_.emplace(sequence_id, std::move(sequence));
+  return sequence_id;
+}
+
+void Scheduler::DestroySequence(SequenceId sequence_id) {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  base::AutoLock auto_lock(lock_);
+
+  Sequence* sequence = GetSequence(sequence_id);
+  DCHECK(sequence);
+  sequence->Destroy();

[vmiura, 2017/05/04 21:27:37]

Is it necessary to defer the Destroy?  Is it due to sequences_ having ownership
of the sequence lifetime?

[sunnyps, 2017/05/10 23:15:15]

This isn't necessary as long as destroying the stream won't have nasty
side-effects. I was going to run pending tasks on destruction but after
discussing with piman@ we decided that's not necessary if we're destroying
streams when we destroy the channel (no need to reply to sync messages). I've
removed the two-stage destruction for now.

+
+  if (sequence->running())
+    return;
+
+  if (sequence->scheduled())
+    rebuild_scheduling_queue_ = true;
+  sequences_.erase(sequence_id);
+}
+
+Scheduler::Sequence* Scheduler::GetSequence(SequenceId sequence_id) {
+  lock_.AssertAcquired();
+  auto it = sequences_.find(sequence_id);
+  if (it != sequences_.end())
+    return it->second.get();
+  return nullptr;
+}
+
+void Scheduler::EnableSequence(SequenceId sequence_id) {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  base::AutoLock auto_lock(lock_);
+  Sequence* sequence = GetSequence(sequence_id);
+  DCHECK(sequence);
+  sequence->SetEnabled(true);
+  TryScheduleSequence(sequence);
+}
+
+void Scheduler::DisableSequence(SequenceId sequence_id) {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  base::AutoLock auto_lock(lock_);
+  Sequence* sequence = GetSequence(sequence_id);
+  DCHECK(sequence);
+  sequence->SetEnabled(false);
+}
+
+void Scheduler::ScheduleTask(SequenceId sequence_id,
+                             base::OnceClosure closure,
+                             const std::vector<SyncToken>& sync_token_fences) {
+  base::AutoLock auto_lock(lock_);
+  Sequence* sequence = GetSequence(sequence_id);
+  DCHECK(sequence);
+
+  uint32_t order_num = sequence->ScheduleTask(std::move(closure));
+
+  for (const SyncToken& sync_token : sync_token_fences) {
+    SequenceId release_id =
+        sync_point_manager_->GetSyncTokenReleaseSequenceId(sync_token);
+    Sequence* release_sequence = GetSequence(release_id);
+    if (!release_sequence)
+      continue;
+    if (sync_point_manager_->Wait(
+            sync_token, order_num,
+            base::Bind(&Scheduler::SyncTokenFenceReleased,
+                       weak_factory_.GetWeakPtr(), sync_token, order_num,
+                       release_id, sequence_id))) {
+      sequence->AddWaitFence(sync_token, order_num);
+      release_sequence->AddReleaseFence(sync_token, order_num);
+      TryScheduleSequence(release_sequence);
+    }
+  }
+
+  TryScheduleSequence(sequence);
+}
+
+void Scheduler::ContinueTask(SequenceId sequence_id,
+                             base::OnceClosure closure) {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  base::AutoLock auto_lock(lock_);
+  Sequence* sequence = GetSequence(sequence_id);
+  DCHECK(sequence);
+  sequence->ContinueTask(std::move(closure));
+}
+
+bool Scheduler::ShouldYield(SequenceId sequence_id) {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  base::AutoLock auto_lock(lock_);
+
+  Sequence* sequence = GetSequence(sequence_id);
+  DCHECK(sequence);
+  DCHECK(sequence->running());
+
+  if (should_yield_)
+    return true;
+
+  RebuildSchedulingQueue();
+
+  sequence->UpdateSchedulingState();
+
+  if (!scheduling_queue_.empty()) {
+    Sequence* next_sequence = GetSequence(scheduling_queue_.top().sequence_id);
+    DCHECK(next_sequence);
+    if (next_sequence->RunsBefore(sequence))
+      should_yield_ = true;
+  }
+
+  return should_yield_;
+}
+
+void Scheduler::SyncTokenFenceReleased(const SyncToken& sync_token,
+                                       uint32_t order_num,
+                                       SequenceId release_sequence_id,
+                                       SequenceId waiting_sequence_id) {
+  base::AutoLock auto_lock(lock_);
+  Sequence* sequence = GetSequence(waiting_sequence_id);
+  if (sequence) {
+    sequence->RemoveWaitFence(sync_token, order_num);
+    TryScheduleSequence(sequence);
+  }
+  Sequence* release_sequence = GetSequence(release_sequence_id);
+  if (release_sequence) {
+    release_sequence->RemoveReleaseFence(sync_token, order_num);
+    TryScheduleSequence(release_sequence);
+  }
+}
+
+void Scheduler::TryScheduleSequence(Sequence* sequence) {
+  lock_.AssertAcquired();
+
+  if (sequence->running())
+    return;
+
+  if (sequence->NeedsRescheduling()) {
+    DCHECK(sequence->IsRunnable());
+    rebuild_scheduling_queue_ = true;
+  } else if (!sequence->scheduled() && sequence->IsRunnable()) {
+    sequence->SetScheduled();
+    scheduling_queue_.push(sequence->scheduling_state());
+  }
+
+  if (!running_) {
+    TRACE_EVENT_ASYNC_BEGIN0("gpu", "Scheduler::Running", this);
+    running_ = true;
+    task_runner_->PostTask(FROM_HERE, base::Bind(&Scheduler::RunNextTask,
+                                                 weak_factory_.GetWeakPtr()));
+  }
+}
+
+void Scheduler::RebuildSchedulingQueue() {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  lock_.AssertAcquired();
+
+  if (!rebuild_scheduling_queue_)
+    return;
+  rebuild_scheduling_queue_ = false;
+
+  std::vector<SchedulingState> states;

[piman, 2017/05/10 00:38:50]

nit: states.reserve(sequences_.count()); to avoid reallocations during the loop?

[sunnyps, 2017/05/10 23:15:15]

Not needed any more (see below).

+  for (const auto& kv : sequences_) {
+    Sequence* sequence = kv.second.get();
+    if (!sequence->IsRunnable() || sequence->running())
+      continue;
+    sequence->SetScheduled();
+    states.push_back(sequence->scheduling_state());
+  }
+
+  scheduling_queue_ = SchedulingQueue(states.begin(), states.end());

[vmiura, 2017/05/04 21:27:37]

For efficiency, it would be better to use the std::priority_queue move
constructor form.

[piman, 2017/05/10 00:38:50]

I'd even argue, if we access underlying container (scheduling_queue_.c), we can
save allocations, e.g. by reusing the underlying vector here (rather that create
a new one in |states|) and calling std::make_heap.
It's a bit more code, but could make things a bit faster.

[sunnyps, 2017/05/10 23:15:15]

Changed to use make_heap, etc.

+}
+
+void Scheduler::RunNextTask() {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  base::AutoLock auto_lock(lock_);
+
+  should_yield_ = false;
+
+  RebuildSchedulingQueue();
+
+  if (scheduling_queue_.empty()) {
+    TRACE_EVENT_ASYNC_END0("gpu", "Scheduler::Running", this);
+    running_ = false;
+    return;
+  }
+
+  SchedulingState state = scheduling_queue_.top();
+  scheduling_queue_.pop();
+
+  TRACE_EVENT1("gpu", "Scheduler::RunNextTask", "state", state.AsValue());
+
+  Sequence* sequence = GetSequence(state.sequence_id);
+  DCHECK(sequence);
+
+  base::OnceClosure closure = sequence->BeginTask();
+
+  {
+    base::AutoUnlock auto_unlock(lock_);
+    std::move(closure).Run();
+  }
+
+  sequence->FinishTask();
+
+  if (sequence->destroyed()) {
+    sequences_.erase(sequence->sequence_id());
+  } else if (sequence->IsRunnable()) {
+    sequence->SetScheduled();
+    scheduling_queue_.push(sequence->scheduling_state());
+  }
+
+  task_runner_->PostTask(FROM_HERE, base::Bind(&Scheduler::RunNextTask,
+                                               weak_factory_.GetWeakPtr()));
+}
+
+}  // namespace gpu