src/tracing/internal/tracing_muxer_impl.cc

/*
 * Copyright (C) 2019 The Android Open Source Project
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *      http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

#include "src/tracing/internal/tracing_muxer_impl.h"

#include <algorithm>
#include <atomic>
#include <mutex>
#include <vector>

#include "perfetto/base/build_config.h"
#include "perfetto/base/logging.h"
#include "perfetto/base/task_runner.h"
#include "perfetto/ext/base/hash.h"
#include "perfetto/ext/base/thread_checker.h"
#include "perfetto/ext/base/waitable_event.h"
#include "perfetto/ext/tracing/core/shared_memory_arbiter.h"
#include "perfetto/ext/tracing/core/trace_packet.h"
#include "perfetto/ext/tracing/core/trace_stats.h"
#include "perfetto/ext/tracing/core/trace_writer.h"
#include "perfetto/ext/tracing/core/tracing_service.h"
#include "perfetto/tracing/buffer_exhausted_policy.h"
#include "perfetto/tracing/core/data_source_config.h"
#include "perfetto/tracing/core/tracing_service_state.h"
#include "perfetto/tracing/data_source.h"
#include "perfetto/tracing/internal/data_source_internal.h"
#include "perfetto/tracing/internal/interceptor_trace_writer.h"
#include "perfetto/tracing/trace_writer_base.h"
#include "perfetto/tracing/tracing.h"
#include "perfetto/tracing/tracing_backend.h"

#include "protos/perfetto/config/interceptor_config.gen.h"

namespace perfetto {
namespace internal {

namespace {

class StopArgsImpl : public DataSourceBase::StopArgs {
 public:
  std::function<void()> HandleStopAsynchronously() const override {
    auto closure = std::move(async_stop_closure);
    async_stop_closure = std::function<void()>();
    return closure;
  }

  mutable std::function<void()> async_stop_closure;
};

uint64_t ComputeConfigHash(const DataSourceConfig& config) {
  base::Hash hasher;
  std::string config_bytes = config.SerializeAsString();
  hasher.Update(config_bytes.data(), config_bytes.size());
  return hasher.digest();
}

}  // namespace

// ----- Begin of TracingMuxerImpl::ProducerImpl
TracingMuxerImpl::ProducerImpl::ProducerImpl(
    TracingMuxerImpl* muxer,
    TracingBackendId backend_id,
    uint32_t shmem_batch_commits_duration_ms)
    : muxer_(muxer),
      backend_id_(backend_id),
      shmem_batch_commits_duration_ms_(shmem_batch_commits_duration_ms) {}

TracingMuxerImpl::ProducerImpl::~ProducerImpl() = default;

void TracingMuxerImpl::ProducerImpl::Initialize(
    std::unique_ptr<ProducerEndpoint> endpoint) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  PERFETTO_DCHECK(!connected_);
  connection_id_++;

  // Adopt the endpoint into a shared pointer so that we can safely share it
  // across threads that create trace writers. The custom deleter function
  // ensures that the endpoint is always destroyed on the muxer's thread. (Note
  // that |task_runner| is assumed to outlive tracing sessions on all threads.)
  auto* task_runner = muxer_->task_runner_.get();
  auto deleter = [task_runner](ProducerEndpoint* e) {
    task_runner->PostTask([e] { delete e; });
  };
  std::shared_ptr<ProducerEndpoint> service(endpoint.release(), deleter);
  // This atomic store is needed because another thread might be concurrently
  // creating a trace writer using the previous (disconnected) |service_|. See
  // CreateTraceWriter().
  std::atomic_store(&service_, std::move(service));
  // Don't try to use the service here since it may not have connected yet. See
  // OnConnect().
}

void TracingMuxerImpl::ProducerImpl::OnConnect() {
  PERFETTO_DLOG("Producer connected");
  PERFETTO_DCHECK_THREAD(thread_checker_);
  PERFETTO_DCHECK(!connected_);
  connected_ = true;
  muxer_->UpdateDataSourcesOnAllBackends();
}

void TracingMuxerImpl::ProducerImpl::OnDisconnect() {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  connected_ = false;
  // Active data sources for this producer will be stopped by
  // DestroyStoppedTraceWritersForCurrentThread() since the reconnected producer
  // will have a different connection id (even before it has finished
  // connecting).
  registered_data_sources_.reset();
  // Keep the old service around as a dead connection in case it has active
  // trace writers. We can't clear |service_| here because other threads may be
  // concurrently creating new trace writers. The reconnection below will
  // atomically swap the new service in place of the old one.
  dead_services_.push_back(service_);
  // Try reconnecting the producer.
  muxer_->OnProducerDisconnected(this);
}

void TracingMuxerImpl::ProducerImpl::OnTracingSetup() {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  service_->MaybeSharedMemoryArbiter()->SetBatchCommitsDuration(
      shmem_batch_commits_duration_ms_);
}

void TracingMuxerImpl::ProducerImpl::SetupDataSource(
    DataSourceInstanceID id,
    const DataSourceConfig& cfg) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  muxer_->SetupDataSource(backend_id_, connection_id_, id, cfg);
}

void TracingMuxerImpl::ProducerImpl::StartDataSource(DataSourceInstanceID id,
                                                     const DataSourceConfig&) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  muxer_->StartDataSource(backend_id_, id);
  service_->NotifyDataSourceStarted(id);
}

void TracingMuxerImpl::ProducerImpl::StopDataSource(DataSourceInstanceID id) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  muxer_->StopDataSource_AsyncBegin(backend_id_, id);
}

void TracingMuxerImpl::ProducerImpl::Flush(FlushRequestID flush_id,
                                           const DataSourceInstanceID*,
                                           size_t) {
  // Flush is not plumbed for now, we just ack straight away.
  PERFETTO_DCHECK_THREAD(thread_checker_);
  service_->NotifyFlushComplete(flush_id);
}

void TracingMuxerImpl::ProducerImpl::ClearIncrementalState(
    const DataSourceInstanceID*,
    size_t) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  // TODO(skyostil): Mark each affected data source's incremental state as
  // needing to be cleared.
}

void TracingMuxerImpl::ProducerImpl::SweepDeadServices() {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  auto is_unused = [](const std::shared_ptr<ProducerEndpoint>& endpoint) {
    auto* arbiter = endpoint->MaybeSharedMemoryArbiter();
    return !arbiter || arbiter->TryShutdown();
  };
  for (auto it = dead_services_.begin(); it != dead_services_.end();) {
    auto next_it = it;
    next_it++;
    if (is_unused(*it)) {
      dead_services_.erase(it);
    }
    it = next_it;
  }
}

// ----- End of TracingMuxerImpl::ProducerImpl methods.

// ----- Begin of TracingMuxerImpl::ConsumerImpl
TracingMuxerImpl::ConsumerImpl::ConsumerImpl(TracingMuxerImpl* muxer,
                                             BackendType backend_type,
                                             TracingBackendId backend_id,
                                             TracingSessionGlobalID session_id)
    : muxer_(muxer),
      backend_type_(backend_type),
      backend_id_(backend_id),
      session_id_(session_id) {}

TracingMuxerImpl::ConsumerImpl::~ConsumerImpl() = default;

void TracingMuxerImpl::ConsumerImpl::Initialize(
    std::unique_ptr<ConsumerEndpoint> endpoint) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  service_ = std::move(endpoint);
  // Don't try to use the service here since it may not have connected yet. See
  // OnConnect().
}

void TracingMuxerImpl::ConsumerImpl::OnConnect() {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  PERFETTO_DCHECK(!connected_);
  connected_ = true;

  // Observe data source instance events so we get notified when tracing starts.
  service_->ObserveEvents(ObservableEvents::TYPE_DATA_SOURCES_INSTANCES |
                          ObservableEvents::TYPE_ALL_DATA_SOURCES_STARTED);

  // If the API client configured and started tracing before we connected,
  // tell the backend about it now.
  if (trace_config_)
    muxer_->SetupTracingSession(session_id_, trace_config_);
  if (start_pending_)
    muxer_->StartTracingSession(session_id_);
  if (get_trace_stats_pending_) {
    auto callback = std::move(get_trace_stats_callback_);
    get_trace_stats_callback_ = nullptr;
    muxer_->GetTraceStats(session_id_, std::move(callback));
  }
  if (query_service_state_callback_) {
    auto callback = std::move(query_service_state_callback_);
    query_service_state_callback_ = nullptr;
    muxer_->QueryServiceState(session_id_, std::move(callback));
  }
  if (stop_pending_)
    muxer_->StopTracingSession(session_id_);
}

void TracingMuxerImpl::ConsumerImpl::OnDisconnect() {
  PERFETTO_DCHECK_THREAD(thread_checker_);
#if PERFETTO_BUILDFLAG(PERFETTO_OS_ANDROID)
  if (!connected_ && backend_type_ == kSystemBackend) {
    PERFETTO_ELOG(
        "Unable to connect to the system tracing service as a consumer. On "
        "Android, use the \"perfetto\" command line tool instead to start "
        "system-wide tracing sessions");
  }
#endif  // PERFETTO_BUILDFLAG(PERFETTO_OS_ANDROID)

  // Notify the client about disconnection.
  NotifyError(TracingError{TracingError::kDisconnected, "Peer disconnected"});

  // Make sure the client doesn't hang in a blocking start/stop because of the
  // disconnection.
  NotifyStartComplete();
  NotifyStopComplete();

  // It shouldn't be necessary to call StopTracingSession. If we get this call
  // it means that the service did shutdown before us, so there is no point
  // trying it to ask it to stop the session. We should just remember to cleanup
  // the consumer vector.
  connected_ = false;

  // Notify the muxer that it is safe to destroy |this|. This is needed because
  // the ConsumerEndpoint stored in |service_| requires that |this| be safe to
  // access until OnDisconnect() is called.
  muxer_->OnConsumerDisconnected(this);
}

void TracingMuxerImpl::ConsumerImpl::Disconnect() {
  // This is weird and deserves a comment.
  //
  // When we called the ConnectConsumer method on the service it returns
  // us a ConsumerEndpoint which we stored in |service_|, however this
  // ConsumerEndpoint holds a pointer to the ConsumerImpl pointed to by
  // |this|. Part of the API contract to TracingService::ConnectConsumer is that
  // the ConsumerImpl pointer has to be valid until the
  // ConsumerImpl::OnDisconnect method is called. Therefore we reset the
  // ConsumerEndpoint |service_|. Eventually this will call
  // ConsumerImpl::OnDisconnect and we will inform the muxer it is safe to
  // call the destructor of |this|.
  service_.reset();
}

void TracingMuxerImpl::ConsumerImpl::OnTracingDisabled(
    const std::string& error) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  PERFETTO_DCHECK(!stopped_);
  stopped_ = true;

  if (!error.empty())
    NotifyError(TracingError{TracingError::kTracingFailed, error});

  // If we're still waiting for the start event, fire it now. This may happen if
  // there are no active data sources in the session.
  NotifyStartComplete();
  NotifyStopComplete();
}

void TracingMuxerImpl::ConsumerImpl::NotifyStartComplete() {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  if (start_complete_callback_) {
    muxer_->task_runner_->PostTask(std::move(start_complete_callback_));
    start_complete_callback_ = nullptr;
  }
  if (blocking_start_complete_callback_) {
    muxer_->task_runner_->PostTask(
        std::move(blocking_start_complete_callback_));
    blocking_start_complete_callback_ = nullptr;
  }
}

void TracingMuxerImpl::ConsumerImpl::NotifyError(const TracingError& error) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  if (error_callback_) {
    muxer_->task_runner_->PostTask(
        std::bind(std::move(error_callback_), error));
  }
}

void TracingMuxerImpl::ConsumerImpl::NotifyStopComplete() {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  if (stop_complete_callback_) {
    muxer_->task_runner_->PostTask(std::move(stop_complete_callback_));
    stop_complete_callback_ = nullptr;
  }
  if (blocking_stop_complete_callback_) {
    muxer_->task_runner_->PostTask(std::move(blocking_stop_complete_callback_));
    blocking_stop_complete_callback_ = nullptr;
  }
}

void TracingMuxerImpl::ConsumerImpl::OnTraceData(
    std::vector<TracePacket> packets,
    bool has_more) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  if (!read_trace_callback_)
    return;

  size_t capacity = 0;
  for (const auto& packet : packets) {
    // 16 is an over-estimation of the proto preamble size
    capacity += packet.size() + 16;
  }

  // The shared_ptr is to avoid making a copy of the buffer when PostTask-ing.
  std::shared_ptr<std::vector<char>> buf(new std::vector<char>());
  buf->reserve(capacity);
  for (auto& packet : packets) {
    char* start;
    size_t size;
    std::tie(start, size) = packet.GetProtoPreamble();
    buf->insert(buf->end(), start, start + size);
    for (auto& slice : packet.slices()) {
      const auto* slice_data = reinterpret_cast<const char*>(slice.start);
      buf->insert(buf->end(), slice_data, slice_data + slice.size);
    }
  }

  auto callback = read_trace_callback_;
  muxer_->task_runner_->PostTask([callback, buf, has_more] {
    TracingSession::ReadTraceCallbackArgs callback_arg{};
    callback_arg.data = buf->size() ? &(*buf)[0] : nullptr;
    callback_arg.size = buf->size();
    callback_arg.has_more = has_more;
    callback(callback_arg);
  });

  if (!has_more)
    read_trace_callback_ = nullptr;
}

void TracingMuxerImpl::ConsumerImpl::OnObservableEvents(
    const ObservableEvents& events) {
  if (events.instance_state_changes_size()) {
    for (const auto& state_change : events.instance_state_changes()) {
      DataSourceHandle handle{state_change.producer_name(),
                              state_change.data_source_name()};
      data_source_states_[handle] =
          state_change.state() ==
          ObservableEvents::DATA_SOURCE_INSTANCE_STATE_STARTED;
    }
  }

  if (events.instance_state_changes_size() ||
      events.all_data_sources_started()) {
    // Data sources are first reported as being stopped before starting, so once
    // all the data sources we know about have started we can declare tracing
    // begun. In the case where there are no matching data sources for the
    // session, the service will report the all_data_sources_started() event
    // without adding any instances (only since Android S / Perfetto v10.0).
    if (start_complete_callback_ || blocking_start_complete_callback_) {
      bool all_data_sources_started = std::all_of(
          data_source_states_.cbegin(), data_source_states_.cend(),
          [](std::pair<DataSourceHandle, bool> state) { return state.second; });
      if (all_data_sources_started)
        NotifyStartComplete();
    }
  }
}

void TracingMuxerImpl::ConsumerImpl::OnTraceStats(
    bool success,
    const TraceStats& trace_stats) {
  if (!get_trace_stats_callback_)
    return;
  TracingSession::GetTraceStatsCallbackArgs callback_arg{};
  callback_arg.success = success;
  callback_arg.trace_stats_data = trace_stats.SerializeAsArray();
  muxer_->task_runner_->PostTask(
      std::bind(std::move(get_trace_stats_callback_), std::move(callback_arg)));
  get_trace_stats_callback_ = nullptr;
}

// The callbacks below are not used.
void TracingMuxerImpl::ConsumerImpl::OnDetach(bool) {}
void TracingMuxerImpl::ConsumerImpl::OnAttach(bool, const TraceConfig&) {}
// ----- End of TracingMuxerImpl::ConsumerImpl

// ----- Begin of TracingMuxerImpl::TracingSessionImpl

// TracingSessionImpl is the RAII object returned to API clients when they
// invoke Tracing::CreateTracingSession. They use it for starting/stopping
// tracing.

TracingMuxerImpl::TracingSessionImpl::TracingSessionImpl(
    TracingMuxerImpl* muxer,
    TracingSessionGlobalID session_id)
    : muxer_(muxer), session_id_(session_id) {}

// Can be destroyed from any thread.
TracingMuxerImpl::TracingSessionImpl::~TracingSessionImpl() {
  auto* muxer = muxer_;
  auto session_id = session_id_;
  muxer->task_runner_->PostTask(
      [muxer, session_id] { muxer->DestroyTracingSession(session_id); });
}

// Can be called from any thread.
void TracingMuxerImpl::TracingSessionImpl::Setup(const TraceConfig& cfg,
                                                 int fd) {
  auto* muxer = muxer_;
  auto session_id = session_id_;
  std::shared_ptr<TraceConfig> trace_config(new TraceConfig(cfg));
  if (fd >= 0) {
    trace_config->set_write_into_file(true);
    fd = dup(fd);
  }
  muxer->task_runner_->PostTask([muxer, session_id, trace_config, fd] {
    muxer->SetupTracingSession(session_id, trace_config, base::ScopedFile(fd));
  });
}

// Can be called from any thread.
void TracingMuxerImpl::TracingSessionImpl::Start() {
  auto* muxer = muxer_;
  auto session_id = session_id_;
  muxer->task_runner_->PostTask(
      [muxer, session_id] { muxer->StartTracingSession(session_id); });
}

// Can be called from any thread.
void TracingMuxerImpl::TracingSessionImpl::ChangeTraceConfig(
    const TraceConfig& cfg) {
  auto* muxer = muxer_;
  auto session_id = session_id_;
  muxer->task_runner_->PostTask([muxer, session_id, cfg] {
    muxer->ChangeTracingSessionConfig(session_id, cfg);
  });
}

// Can be called from any thread except the service thread.
void TracingMuxerImpl::TracingSessionImpl::StartBlocking() {
  PERFETTO_DCHECK(!muxer_->task_runner_->RunsTasksOnCurrentThread());
  auto* muxer = muxer_;
  auto session_id = session_id_;
  base::WaitableEvent tracing_started;
  muxer->task_runner_->PostTask([muxer, session_id, &tracing_started] {
    auto* consumer = muxer->FindConsumer(session_id);
    if (!consumer) {
      // TODO(skyostil): Signal an error to the user.
      tracing_started.Notify();
      return;
    }
    PERFETTO_DCHECK(!consumer->blocking_start_complete_callback_);
    consumer->blocking_start_complete_callback_ = [&] {
      tracing_started.Notify();
    };
    muxer->StartTracingSession(session_id);
  });
  tracing_started.Wait();
}

// Can be called from any thread.
void TracingMuxerImpl::TracingSessionImpl::Flush(
    std::function<void(bool)> user_callback,
    uint32_t timeout_ms) {
  auto* muxer = muxer_;
  auto session_id = session_id_;
  muxer->task_runner_->PostTask([muxer, session_id, timeout_ms, user_callback] {
    auto* consumer = muxer->FindConsumer(session_id);
    if (!consumer) {
      std::move(user_callback)(false);
      return;
    }
    muxer->FlushTracingSession(session_id, timeout_ms,
                               std::move(user_callback));
  });
}

// Can be called from any thread.
void TracingMuxerImpl::TracingSessionImpl::Stop() {
  auto* muxer = muxer_;
  auto session_id = session_id_;
  muxer->task_runner_->PostTask(
      [muxer, session_id] { muxer->StopTracingSession(session_id); });
}

// Can be called from any thread except the service thread.
void TracingMuxerImpl::TracingSessionImpl::StopBlocking() {
  PERFETTO_DCHECK(!muxer_->task_runner_->RunsTasksOnCurrentThread());
  auto* muxer = muxer_;
  auto session_id = session_id_;
  base::WaitableEvent tracing_stopped;
  muxer->task_runner_->PostTask([muxer, session_id, &tracing_stopped] {
    auto* consumer = muxer->FindConsumer(session_id);
    if (!consumer) {
      // TODO(skyostil): Signal an error to the user.
      tracing_stopped.Notify();
      return;
    }
    PERFETTO_DCHECK(!consumer->blocking_stop_complete_callback_);
    consumer->blocking_stop_complete_callback_ = [&] {
      tracing_stopped.Notify();
    };
    muxer->StopTracingSession(session_id);
  });
  tracing_stopped.Wait();
}

// Can be called from any thread.
void TracingMuxerImpl::TracingSessionImpl::ReadTrace(ReadTraceCallback cb) {
  auto* muxer = muxer_;
  auto session_id = session_id_;
  muxer->task_runner_->PostTask([muxer, session_id, cb] {
    muxer->ReadTracingSessionData(session_id, std::move(cb));
  });
}

// Can be called from any thread.
void TracingMuxerImpl::TracingSessionImpl::SetOnStartCallback(
    std::function<void()> cb) {
  auto* muxer = muxer_;
  auto session_id = session_id_;
  muxer->task_runner_->PostTask([muxer, session_id, cb] {
    auto* consumer = muxer->FindConsumer(session_id);
    consumer->start_complete_callback_ = cb;
  });
}

// Can be called from any thread
void TracingMuxerImpl::TracingSessionImpl::SetOnErrorCallback(
    std::function<void(TracingError)> cb) {
  auto* muxer = muxer_;
  auto session_id = session_id_;
  muxer->task_runner_->PostTask([muxer, session_id, cb] {
    auto* consumer = muxer->FindConsumer(session_id);
    if (!consumer)
      return;
    consumer->error_callback_ = cb;
  });
}

// Can be called from any thread.
void TracingMuxerImpl::TracingSessionImpl::SetOnStopCallback(
    std::function<void()> cb) {
  auto* muxer = muxer_;
  auto session_id = session_id_;
  muxer->task_runner_->PostTask([muxer, session_id, cb] {
    auto* consumer = muxer->FindConsumer(session_id);
    consumer->stop_complete_callback_ = cb;
  });
}

// Can be called from any thread.
void TracingMuxerImpl::TracingSessionImpl::GetTraceStats(
    GetTraceStatsCallback cb) {
  auto* muxer = muxer_;
  auto session_id = session_id_;
  muxer->task_runner_->PostTask([muxer, session_id, cb] {
    muxer->GetTraceStats(session_id, std::move(cb));
  });
}

// Can be called from any thread.
void TracingMuxerImpl::TracingSessionImpl::QueryServiceState(
    QueryServiceStateCallback cb) {
  auto* muxer = muxer_;
  auto session_id = session_id_;
  muxer->task_runner_->PostTask([muxer, session_id, cb] {
    muxer->QueryServiceState(session_id, std::move(cb));
  });
}

// ----- End of TracingMuxerImpl::TracingSessionImpl

// static
TracingMuxer* TracingMuxer::instance_ = nullptr;

// This is called by perfetto::Tracing::Initialize().
// Can be called on any thread. Typically, but not necessarily, that will be
// the embedder's main thread.
TracingMuxerImpl::TracingMuxerImpl(const TracingInitArgs& args)
    : TracingMuxer(args.platform ? args.platform
                                 : Platform::GetDefaultPlatform()) {
  PERFETTO_DETACH_FROM_THREAD(thread_checker_);

  // Create the thread where muxer, producers and service will live.
  task_runner_ = platform_->CreateTaskRunner({});

  // Run the initializer on that thread.
  task_runner_->PostTask([this, args] { Initialize(args); });
}

void TracingMuxerImpl::Initialize(const TracingInitArgs& args) {
  PERFETTO_DCHECK_THREAD(thread_checker_);  // Rebind the thread checker.

  auto add_backend = [this, &args](TracingBackend* backend, BackendType type) {
    if (!backend) {
      // We skip the log in release builds because the *_backend_fake.cc code
      // has already an ELOG before returning a nullptr.
      PERFETTO_DLOG("Backend creation failed, type %d", static_cast<int>(type));
      return;
    }
    TracingBackendId backend_id = backends_.size();
    backends_.emplace_back();
    RegisteredBackend& rb = backends_.back();
    rb.backend = backend;
    rb.id = backend_id;
    rb.type = type;
    rb.producer.reset(new ProducerImpl(this, backend_id,
                                       args.shmem_batch_commits_duration_ms));
    rb.producer_conn_args.producer = rb.producer.get();
    rb.producer_conn_args.producer_name = platform_->GetCurrentProcessName();
    rb.producer_conn_args.task_runner = task_runner_.get();
    rb.producer_conn_args.shmem_size_hint_bytes =
        args.shmem_size_hint_kb * 1024;
    rb.producer_conn_args.shmem_page_size_hint_bytes =
        args.shmem_page_size_hint_kb * 1024;
    rb.producer->Initialize(rb.backend->ConnectProducer(rb.producer_conn_args));
  };

  if (args.backends & kSystemBackend) {
    PERFETTO_CHECK(args.system_backend_factory_);
    add_backend(args.system_backend_factory_(), kSystemBackend);
  }

  if (args.backends & kInProcessBackend) {
    PERFETTO_CHECK(args.in_process_backend_factory_);
    add_backend(args.in_process_backend_factory_(), kInProcessBackend);
  }

  if (args.backends & kCustomBackend) {
    PERFETTO_CHECK(args.custom_backend);
    add_backend(args.custom_backend, kCustomBackend);
  }

  if (args.backends & ~(kSystemBackend | kInProcessBackend | kCustomBackend)) {
    PERFETTO_FATAL("Unsupported tracing backend type");
  }
}

// Can be called from any thread (but not concurrently).
bool TracingMuxerImpl::RegisterDataSource(
    const DataSourceDescriptor& descriptor,
    DataSourceFactory factory,
    DataSourceStaticState* static_state) {
  // Ignore repeated registrations.
  if (static_state->index != kMaxDataSources)
    return true;

  static std::atomic<uint32_t> last_id{};
  uint32_t new_index = last_id++;
  if (new_index >= kMaxDataSources) {
    PERFETTO_DLOG(
        "RegisterDataSource failed: too many data sources already registered");
    return false;
  }

  // Initialize the static state.
  static_assert(sizeof(static_state->instances[0]) >= sizeof(DataSourceState),
                "instances[] size mismatch");
  for (size_t i = 0; i < static_state->instances.size(); i++)
    new (&static_state->instances[i]) DataSourceState{};

  static_state->index = new_index;

  task_runner_->PostTask([this, descriptor, factory, static_state] {
    data_sources_.emplace_back();
    RegisteredDataSource& rds = data_sources_.back();
    rds.descriptor = descriptor;
    rds.factory = factory;
    rds.static_state = static_state;
    UpdateDataSourcesOnAllBackends();
  });
  return true;
}

// Can be called from any thread (but not concurrently).
void TracingMuxerImpl::RegisterInterceptor(
    const InterceptorDescriptor& descriptor,
    InterceptorFactory factory,
    InterceptorBase::TLSFactory tls_factory,
    InterceptorBase::TracePacketCallback packet_callback) {
  task_runner_->PostTask(
      [this, descriptor, factory, tls_factory, packet_callback] {
        // Ignore repeated registrations.
        for (const auto& interceptor : interceptors_) {
          if (interceptor.descriptor.name() == descriptor.name()) {
            PERFETTO_DCHECK(interceptor.tls_factory == tls_factory);
            PERFETTO_DCHECK(interceptor.packet_callback == packet_callback);
            return;
          }
        }
        // Only allow certain interceptors for now.
        if (descriptor.name() != "test_interceptor") {
          PERFETTO_ELOG(
              "Interceptors are experimental. If you want to use them, please "
              "get in touch with the project maintainers "
              "(https://perfetto.dev/docs/contributing/"
              "getting-started#community).");
          return;
        }
        interceptors_.emplace_back();
        RegisteredInterceptor& interceptor = interceptors_.back();
        interceptor.descriptor = descriptor;
        interceptor.factory = factory;
        interceptor.tls_factory = tls_factory;
        interceptor.packet_callback = packet_callback;
      });
}

// Called by the service of one of the backends.
void TracingMuxerImpl::SetupDataSource(TracingBackendId backend_id,
                                       uint32_t backend_connection_id,
                                       DataSourceInstanceID instance_id,
                                       const DataSourceConfig& cfg) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  PERFETTO_DLOG("Setting up data source %" PRIu64 " %s", instance_id,
                cfg.name().c_str());
  uint64_t config_hash = ComputeConfigHash(cfg);

  for (const auto& rds : data_sources_) {
    if (rds.descriptor.name() != cfg.name())
      continue;
    DataSourceStaticState& static_state = *rds.static_state;

    // If this data source is already active for this exact config, don't start
    // another instance. This happens when we have several data sources with the
    // same name, in which case the service sends one SetupDataSource event for
    // each one. Since we can't map which event maps to which data source, we
    // ensure each event only starts one data source instance.
    // TODO(skyostil): Register a unique id with each data source to the service
    // to disambiguate.
    bool active_for_config = false;
    for (uint32_t i = 0; i < kMaxDataSourceInstances; i++) {
      if (!static_state.TryGet(i))
        continue;
      auto* internal_state =
          reinterpret_cast<DataSourceState*>(&static_state.instances[i]);
      if (internal_state->backend_id == backend_id &&
          internal_state->config_hash == config_hash) {
        active_for_config = true;
        break;
      }
    }
    if (active_for_config) {
      PERFETTO_DLOG(
          "Data source %s is already active with this config, skipping",
          cfg.name().c_str());
      continue;
    }

    for (uint32_t i = 0; i < kMaxDataSourceInstances; i++) {
      // Find a free slot.
      if (static_state.TryGet(i))
        continue;

      auto* internal_state =
          reinterpret_cast<DataSourceState*>(&static_state.instances[i]);
      std::lock_guard<std::recursive_mutex> guard(internal_state->lock);
      static_assert(
          std::is_same<decltype(internal_state->data_source_instance_id),
                       DataSourceInstanceID>::value,
          "data_source_instance_id type mismatch");
      internal_state->backend_id = backend_id;
      internal_state->backend_connection_id = backend_connection_id;
      internal_state->data_source_instance_id = instance_id;
      internal_state->buffer_id =
          static_cast<internal::BufferId>(cfg.target_buffer());
      internal_state->config_hash = config_hash;
      internal_state->data_source = rds.factory();
      internal_state->interceptor = nullptr;
      internal_state->interceptor_id = 0;

      if (cfg.has_interceptor_config()) {
        for (size_t j = 0; j < interceptors_.size(); j++) {
          if (cfg.interceptor_config().name() ==
              interceptors_[j].descriptor.name()) {
            PERFETTO_DLOG("Intercepting data source %" PRIu64
                          " \"%s\" into \"%s\"",
                          instance_id, cfg.name().c_str(),
                          cfg.interceptor_config().name().c_str());
            internal_state->interceptor_id = static_cast<uint32_t>(j + 1);
            internal_state->interceptor = interceptors_[j].factory();
            internal_state->interceptor->OnSetup({cfg});
            break;
          }
        }
        if (!internal_state->interceptor_id) {
          PERFETTO_ELOG("Unknown interceptor configured for data source: %s",
                        cfg.interceptor_config().name().c_str());
        }
      }

      // This must be made at the end. See matching acquire-load in
      // DataSource::Trace().
      static_state.valid_instances.fetch_or(1 << i, std::memory_order_release);

      DataSourceBase::SetupArgs setup_args;
      setup_args.config = &cfg;
      setup_args.internal_instance_index = i;
      internal_state->data_source->OnSetup(setup_args);
      return;
    }
    PERFETTO_ELOG(
        "Maximum number of data source instances exhausted. "
        "Dropping data source %" PRIu64,
        instance_id);
    break;
  }
}

// Called by the service of one of the backends.
void TracingMuxerImpl::StartDataSource(TracingBackendId backend_id,
                                       DataSourceInstanceID instance_id) {
  PERFETTO_DLOG("Starting data source %" PRIu64, instance_id);
  PERFETTO_DCHECK_THREAD(thread_checker_);

  auto ds = FindDataSource(backend_id, instance_id);
  if (!ds) {
    PERFETTO_ELOG("Could not find data source to start");
    return;
  }

  DataSourceBase::StartArgs start_args{};
  start_args.internal_instance_index = ds.instance_idx;

  std::lock_guard<std::recursive_mutex> guard(ds.internal_state->lock);
  if (ds.internal_state->interceptor)
    ds.internal_state->interceptor->OnStart({});
  ds.internal_state->trace_lambda_enabled = true;
  ds.internal_state->data_source->OnStart(start_args);
}

// Called by the service of one of the backends.
void TracingMuxerImpl::StopDataSource_AsyncBegin(
    TracingBackendId backend_id,
    DataSourceInstanceID instance_id) {
  PERFETTO_DLOG("Stopping data source %" PRIu64, instance_id);
  PERFETTO_DCHECK_THREAD(thread_checker_);

  auto ds = FindDataSource(backend_id, instance_id);
  if (!ds) {
    PERFETTO_ELOG("Could not find data source to stop");
    return;
  }

  StopArgsImpl stop_args{};
  stop_args.internal_instance_index = ds.instance_idx;
  stop_args.async_stop_closure = [this, backend_id, instance_id] {
    // TracingMuxerImpl is long lived, capturing |this| is okay.
    // The notification closure can be moved out of the StopArgs by the
    // embedder to handle stop asynchronously. The embedder might then
    // call the closure on a different thread than the current one, hence
    // this nested PostTask().
    task_runner_->PostTask([this, backend_id, instance_id] {
      StopDataSource_AsyncEnd(backend_id, instance_id);
    });
  };

  {
    std::lock_guard<std::recursive_mutex> guard(ds.internal_state->lock);
    if (ds.internal_state->interceptor)
      ds.internal_state->interceptor->OnStop({});
    ds.internal_state->data_source->OnStop(stop_args);
  }

  // If the embedder hasn't called StopArgs.HandleStopAsynchronously() run the
  // async closure here. In theory we could avoid the PostTask and call
  // straight into CompleteDataSourceAsyncStop(). We keep that to reduce
  // divergencies between the deferred-stop vs non-deferred-stop code paths.
  if (stop_args.async_stop_closure)
    std::move(stop_args.async_stop_closure)();
}

void TracingMuxerImpl::StopDataSource_AsyncEnd(
    TracingBackendId backend_id,
    DataSourceInstanceID instance_id) {
  PERFETTO_DLOG("Ending async stop of data source %" PRIu64, instance_id);
  PERFETTO_DCHECK_THREAD(thread_checker_);

  auto ds = FindDataSource(backend_id, instance_id);
  if (!ds) {
    PERFETTO_ELOG(
        "Async stop of data source %" PRIu64
        " failed. This might be due to calling the async_stop_closure twice.",
        instance_id);
    return;
  }

  const uint32_t mask = ~(1 << ds.instance_idx);
  ds.static_state->valid_instances.fetch_and(mask, std::memory_order_acq_rel);

  // Take the mutex to prevent that the data source is in the middle of
  // a Trace() execution where it called GetDataSourceLocked() while we
  // destroy it.
  {
    std::lock_guard<std::recursive_mutex> guard(ds.internal_state->lock);
    ds.internal_state->trace_lambda_enabled = false;
    ds.internal_state->data_source.reset();
  }

  // The other fields of internal_state are deliberately *not* cleared.
  // See races-related comments of DataSource::Trace().

  TracingMuxer::generation_++;

  // |backends_| is append-only, Backend instances are always valid.
  PERFETTO_CHECK(backend_id < backends_.size());
  ProducerImpl* producer = backends_[backend_id].producer.get();
  if (!producer)
    return;
  if (producer->connected_) {
    // Flush any commits that might have been batched by SharedMemoryArbiter.
    producer->service_->MaybeSharedMemoryArbiter()
        ->FlushPendingCommitDataRequests();
    producer->service_->NotifyDataSourceStopped(instance_id);
  }
  producer->SweepDeadServices();
}

void TracingMuxerImpl::SyncProducersForTesting() {
  std::mutex mutex;
  std::condition_variable cv;

  // IPC-based producers don't report connection errors explicitly for each
  // command, but instead with an asynchronous callback
  // (ProducerImpl::OnDisconnected). This means that the sync command below
  // may have completed but failed to reach the service because of a
  // disconnection, but we can't tell until the disconnection message comes
  // through. To guard against this, we run two whole rounds of sync round-trips
  // before returning; the first one will detect any disconnected producers and
  // the second one will ensure any reconnections have completed and all data
  // sources are registered in the service again.
  for (size_t i = 0; i < 2; i++) {
    size_t countdown = std::numeric_limits<size_t>::max();
    task_runner_->PostTask([this, &mutex, &cv, &countdown] {
      {
        std::unique_lock<std::mutex> countdown_lock(mutex);
        countdown = backends_.size();
      }
      for (auto& backend : backends_) {
        auto* producer = backend.producer.get();
        producer->service_->Sync([&mutex, &cv, &countdown] {
          std::unique_lock<std::mutex> countdown_lock(mutex);
          countdown--;
          cv.notify_one();
        });
      }
    });

    {
      std::unique_lock<std::mutex> countdown_lock(mutex);
      cv.wait(countdown_lock, [&countdown] { return !countdown; });
    }
  }

  // Check that all producers are indeed connected.
  bool done = false;
  bool all_producers_connected = true;
  task_runner_->PostTask([this, &mutex, &cv, &done, &all_producers_connected] {
    for (auto& backend : backends_)
      all_producers_connected &= backend.producer->connected_;
    std::unique_lock<std::mutex> lock(mutex);
    done = true;
    cv.notify_one();
  });

  {
    std::unique_lock<std::mutex> lock(mutex);
    cv.wait(lock, [&done] { return done; });
  }
  PERFETTO_DCHECK(all_producers_connected);
}

void TracingMuxerImpl::DestroyStoppedTraceWritersForCurrentThread() {
  // Iterate across all possible data source types.
  auto cur_generation = generation_.load(std::memory_order_acquire);
  auto* root_tls = GetOrCreateTracingTLS();

  auto destroy_stopped_instances = [](DataSourceThreadLocalState& tls) {
    // |tls| has a vector of per-data-source-instance thread-local state.
    DataSourceStaticState* static_state = tls.static_state;
    if (!static_state)
      return;  // Slot not used.

    // Iterate across all possible instances for this data source.
    for (uint32_t inst = 0; inst < kMaxDataSourceInstances; inst++) {
      DataSourceInstanceThreadLocalState& ds_tls = tls.per_instance[inst];
      if (!ds_tls.trace_writer)
        continue;

      DataSourceState* ds_state = static_state->TryGet(inst);
      if (ds_state && ds_state->backend_id == ds_tls.backend_id &&
          ds_state->backend_connection_id == ds_tls.backend_connection_id &&
          ds_state->buffer_id == ds_tls.buffer_id &&
          ds_state->data_source_instance_id == ds_tls.data_source_instance_id) {
        continue;
      }

      // The DataSource instance has been destroyed or recycled.
      ds_tls.Reset();  // Will also destroy the |ds_tls.trace_writer|.
    }
  };

  for (size_t ds_idx = 0; ds_idx < kMaxDataSources; ds_idx++) {
    // |tls| has a vector of per-data-source-instance thread-local state.
    DataSourceThreadLocalState& tls = root_tls->data_sources_tls[ds_idx];
    destroy_stopped_instances(tls);
  }
  destroy_stopped_instances(root_tls->track_event_tls);
  root_tls->generation = cur_generation;
}

// Called both when a new data source is registered or when a new backend
// connects. In both cases we want to be sure we reflected the data source
// registrations on the backends.
void TracingMuxerImpl::UpdateDataSourcesOnAllBackends() {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  for (RegisteredDataSource& rds : data_sources_) {
    for (RegisteredBackend& backend : backends_) {
      // We cannot call RegisterDataSource on the backend before it connects.
      if (!backend.producer->connected_)
        continue;

      PERFETTO_DCHECK(rds.static_state->index < kMaxDataSources);
      if (backend.producer->registered_data_sources_.test(
              rds.static_state->index))
        continue;

      rds.descriptor.set_will_notify_on_start(true);
      rds.descriptor.set_will_notify_on_stop(true);
      backend.producer->service_->RegisterDataSource(rds.descriptor);
      backend.producer->registered_data_sources_.set(rds.static_state->index);
    }
  }
}

void TracingMuxerImpl::SetupTracingSession(
    TracingSessionGlobalID session_id,
    const std::shared_ptr<TraceConfig>& trace_config,
    base::ScopedFile trace_fd) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  PERFETTO_CHECK(!trace_fd || trace_config->write_into_file());

  auto* consumer = FindConsumer(session_id);
  if (!consumer)
    return;

  consumer->trace_config_ = trace_config;
  if (trace_fd)
    consumer->trace_fd_ = std::move(trace_fd);

  if (!consumer->connected_)
    return;

  // Only used in the deferred start mode.
  if (trace_config->deferred_start()) {
    consumer->service_->EnableTracing(*trace_config,
                                      std::move(consumer->trace_fd_));
  }
}

void TracingMuxerImpl::StartTracingSession(TracingSessionGlobalID session_id) {
  PERFETTO_DCHECK_THREAD(thread_checker_);

  auto* consumer = FindConsumer(session_id);

  if (!consumer)
    return;

  if (!consumer->trace_config_) {
    PERFETTO_ELOG("Must call Setup(config) first");
    return;
  }

  if (!consumer->connected_) {
    consumer->start_pending_ = true;
    return;
  }

  consumer->start_pending_ = false;
  if (consumer->trace_config_->deferred_start()) {
    consumer->service_->StartTracing();
  } else {
    consumer->service_->EnableTracing(*consumer->trace_config_,
                                      std::move(consumer->trace_fd_));
  }

  // TODO implement support for the deferred-start + fast-triggering case.
}

void TracingMuxerImpl::ChangeTracingSessionConfig(
    TracingSessionGlobalID session_id,
    const TraceConfig& trace_config) {
  PERFETTO_DCHECK_THREAD(thread_checker_);

  auto* consumer = FindConsumer(session_id);

  if (!consumer)
    return;

  if (!consumer->trace_config_) {
    // Changing the config is only supported for started sessions.
    PERFETTO_ELOG("Must call Setup(config) and Start() first");
    return;
  }

  consumer->trace_config_ = std::make_shared<TraceConfig>(trace_config);
  if (consumer->connected_)
    consumer->service_->ChangeTraceConfig(trace_config);
}

void TracingMuxerImpl::FlushTracingSession(TracingSessionGlobalID session_id,
                                           uint32_t timeout_ms,
                                           std::function<void(bool)> callback) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  auto* consumer = FindConsumer(session_id);
  if (!consumer || consumer->start_pending_ || consumer->stop_pending_ ||
      !consumer->trace_config_) {
    PERFETTO_ELOG("Flush() can be called only after Start() and before Stop()");
    std::move(callback)(false);
    return;
  }

  consumer->service_->Flush(timeout_ms, std::move(callback));
}

void TracingMuxerImpl::StopTracingSession(TracingSessionGlobalID session_id) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  auto* consumer = FindConsumer(session_id);
  if (!consumer)
    return;

  if (consumer->start_pending_) {
    // If the session hasn't started yet, wait until it does before stopping.
    consumer->stop_pending_ = true;
    return;
  }

  consumer->stop_pending_ = false;
  if (consumer->stopped_) {
    // If the session was already stopped (e.g., it failed to start), don't try
    // stopping again.
    consumer->NotifyStopComplete();
  } else if (!consumer->trace_config_) {
    PERFETTO_ELOG("Must call Setup(config) and Start() first");
    return;
  } else {
    consumer->service_->DisableTracing();
  }

  consumer->trace_config_.reset();
}

void TracingMuxerImpl::DestroyTracingSession(
    TracingSessionGlobalID session_id) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  for (RegisteredBackend& backend : backends_) {
    // We need to find the consumer (if any) and call Disconnect as we destroy
    // the tracing session. We can't call Disconnect() inside this for loop
    // because in the in-process case this will end up to a synchronous call to
    // OnConsumerDisconnect which will invalidate all the iterators to
    // |backend.consumers|.
    ConsumerImpl* consumer = nullptr;
    for (auto& con : backend.consumers) {
      if (con->session_id_ == session_id) {
        consumer = con.get();
        break;
      }
    }
    if (consumer) {
      // We broke out of the loop above on the assumption that each backend will
      // only have a single consumer per session. This DCHECK ensures that
      // this is the case.
      PERFETTO_DCHECK(
          std::count_if(backend.consumers.begin(), backend.consumers.end(),
                        [session_id](const std::unique_ptr<ConsumerImpl>& con) {
                          return con->session_id_ == session_id;
                        }) == 1u);
      consumer->Disconnect();
    }
  }
}

void TracingMuxerImpl::ReadTracingSessionData(
    TracingSessionGlobalID session_id,
    std::function<void(TracingSession::ReadTraceCallbackArgs)> callback) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  auto* consumer = FindConsumer(session_id);
  if (!consumer) {
    // TODO(skyostil): Signal an error to the user.
    TracingSession::ReadTraceCallbackArgs callback_arg{};
    callback(callback_arg);
    return;
  }
  PERFETTO_DCHECK(!consumer->read_trace_callback_);
  consumer->read_trace_callback_ = std::move(callback);
  consumer->service_->ReadBuffers();
}

void TracingMuxerImpl::GetTraceStats(
    TracingSessionGlobalID session_id,
    TracingSession::GetTraceStatsCallback callback) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  auto* consumer = FindConsumer(session_id);
  if (!consumer) {
    TracingSession::GetTraceStatsCallbackArgs callback_arg{};
    callback_arg.success = false;
    callback(std::move(callback_arg));
    return;
  }
  PERFETTO_DCHECK(!consumer->get_trace_stats_callback_);
  consumer->get_trace_stats_callback_ = std::move(callback);
  if (!consumer->connected_) {
    consumer->get_trace_stats_pending_ = true;
    return;
  }
  consumer->get_trace_stats_pending_ = false;
  consumer->service_->GetTraceStats();
}

void TracingMuxerImpl::QueryServiceState(
    TracingSessionGlobalID session_id,
    TracingSession::QueryServiceStateCallback callback) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  auto* consumer = FindConsumer(session_id);
  if (!consumer) {
    TracingSession::QueryServiceStateCallbackArgs callback_arg{};
    callback_arg.success = false;
    callback(std::move(callback_arg));
    return;
  }
  PERFETTO_DCHECK(!consumer->query_service_state_callback_);
  if (!consumer->connected_) {
    consumer->query_service_state_callback_ = std::move(callback);
    return;
  }
  auto callback_wrapper = [callback](bool success,
                                     protos::gen::TracingServiceState state) {
    TracingSession::QueryServiceStateCallbackArgs callback_arg{};
    callback_arg.success = success;
    callback_arg.service_state_data = state.SerializeAsArray();
    callback(std::move(callback_arg));
  };
  consumer->service_->QueryServiceState(std::move(callback_wrapper));
}

void TracingMuxerImpl::SetBatchCommitsDurationForTesting(
    uint32_t batch_commits_duration_ms,
    BackendType backend_type) {
  for (RegisteredBackend& backend : backends_) {
    if (backend.producer && backend.producer->connected_ &&
        backend.type == backend_type) {
      backend.producer->service_->MaybeSharedMemoryArbiter()
          ->SetBatchCommitsDuration(batch_commits_duration_ms);
    }
  }
}

bool TracingMuxerImpl::EnableDirectSMBPatchingForTesting(
    BackendType backend_type) {
  for (RegisteredBackend& backend : backends_) {
    if (backend.producer && backend.producer->connected_ &&
        backend.type == backend_type &&
        !backend.producer->service_->MaybeSharedMemoryArbiter()
             ->EnableDirectSMBPatching()) {
      return false;
    }
  }
  return true;
}

TracingMuxerImpl::ConsumerImpl* TracingMuxerImpl::FindConsumer(
    TracingSessionGlobalID session_id) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  for (RegisteredBackend& backend : backends_) {
    for (auto& consumer : backend.consumers) {
      if (consumer->session_id_ == session_id) {
        PERFETTO_DCHECK(consumer->service_);
        return consumer.get();
      }
    }
  }
  return nullptr;
}

void TracingMuxerImpl::OnConsumerDisconnected(ConsumerImpl* consumer) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  for (RegisteredBackend& backend : backends_) {
    auto pred = [consumer](const std::unique_ptr<ConsumerImpl>& con) {
      return con.get() == consumer;
    };
    backend.consumers.erase(std::remove_if(backend.consumers.begin(),
                                           backend.consumers.end(), pred),
                            backend.consumers.end());
  }
}

void TracingMuxerImpl::SetMaxProducerReconnectionsForTesting(uint32_t count) {
  max_producer_reconnections_.store(count);
}

void TracingMuxerImpl::OnProducerDisconnected(ProducerImpl* producer) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  for (RegisteredBackend& backend : backends_) {
    if (backend.producer.get() != producer)
      continue;
    // Try reconnecting the disconnected producer. If the connection succeeds,
    // all the data sources will be automatically re-registered.
    if (producer->connection_id_ > max_producer_reconnections_.load()) {
      // Avoid reconnecting a failing producer too many times. Instead we just
      // leak the producer instead of trying to avoid further complicating
      // cross-thread trace writer creation.
      PERFETTO_ELOG("Producer disconnected too many times; not reconnecting");
      continue;
    }
    backend.producer->Initialize(
        backend.backend->ConnectProducer(backend.producer_conn_args));
  }

  // Increment the generation counter to atomically ensure that:
  // 1. Old trace writers from the severed connection eventually get cleaned up
  //    by DestroyStoppedTraceWritersForCurrentThread().
  // 2. No new trace writers can be created for the SharedMemoryArbiter from the
  //    old connection.
  TracingMuxer::generation_++;
}

TracingMuxerImpl::FindDataSourceRes TracingMuxerImpl::FindDataSource(
    TracingBackendId backend_id,
    DataSourceInstanceID instance_id) {
  PERFETTO_DCHECK_THREAD(thread_checker_);
  for (const auto& rds : data_sources_) {
    DataSourceStaticState* static_state = rds.static_state;
    for (uint32_t i = 0; i < kMaxDataSourceInstances; i++) {
      auto* internal_state = static_state->TryGet(i);
      if (internal_state && internal_state->backend_id == backend_id &&
          internal_state->data_source_instance_id == instance_id) {
        return FindDataSourceRes(static_state, internal_state, i);
      }
    }
  }
  return FindDataSourceRes();
}

// Can be called from any thread.
std::unique_ptr<TraceWriterBase> TracingMuxerImpl::CreateTraceWriter(
    DataSourceStaticState* static_state,
    uint32_t data_source_instance_index,
    DataSourceState* data_source,
    BufferExhaustedPolicy buffer_exhausted_policy) {
  if (PERFETTO_UNLIKELY(data_source->interceptor_id)) {
    // If the session is being intercepted, return a heap-backed trace writer
    // instead. This is safe because all the data given to the interceptor is
    // either thread-local (|instance_index|), statically allocated
    // (|static_state|) or constant after initialization (|interceptor|). Access
    // to the interceptor instance itself through |data_source| is protected by
    // a statically allocated lock (similarly to the data source instance).
    auto& interceptor = interceptors_[data_source->interceptor_id - 1];
    return std::unique_ptr<TraceWriterBase>(new InterceptorTraceWriter(
        interceptor.tls_factory(static_state, data_source_instance_index),
        interceptor.packet_callback, static_state, data_source_instance_index));
  }
  ProducerImpl* producer = backends_[data_source->backend_id].producer.get();
  // Atomically load the current service endpoint. We keep the pointer as a
  // shared pointer on the stack to guard against it from being concurrently
  // modified on the thread by ProducerImpl::Initialize() swapping in a
  // reconnected service on the muxer task runner thread.
  //
  // The endpoint may also be concurrently modified by SweepDeadServices()
  // clearing out old disconnected services. We guard against that by
  // SharedMemoryArbiter keeping track of any outstanding trace writers. After
  // shutdown has started, the trace writer created below will be a null one
  // which will drop any written data. See SharedMemoryArbiter::TryShutdown().
  //
  // We use an atomic pointer instead of holding a lock because
  // CreateTraceWriter posts tasks under the hood.
  std::shared_ptr<ProducerEndpoint> service =
      std::atomic_load(&producer->service_);
  return service->CreateTraceWriter(data_source->buffer_id,
                                    buffer_exhausted_policy);
}

// This is called via the public API Tracing::NewTrace().
// Can be called from any thread.
std::unique_ptr<TracingSession> TracingMuxerImpl::CreateTracingSession(
    BackendType backend_type) {
  TracingSessionGlobalID session_id = ++next_tracing_session_id_;

  // |backend_type| can only specify one backend, not an OR-ed mask.
  PERFETTO_CHECK((backend_type & (backend_type - 1)) == 0);

  // Capturing |this| is fine because the TracingMuxer is a leaky singleton.
  task_runner_->PostTask([this, backend_type, session_id] {
    for (RegisteredBackend& backend : backends_) {
      if (backend_type && backend.type != backend_type)
        continue;

      backend.consumers.emplace_back(
          new ConsumerImpl(this, backend.type, backend.id, session_id));
      auto& consumer = backend.consumers.back();
      TracingBackend::ConnectConsumerArgs conn_args;
      conn_args.consumer = consumer.get();
      conn_args.task_runner = task_runner_.get();
      consumer->Initialize(backend.backend->ConnectConsumer(conn_args));
      return;
    }
    PERFETTO_ELOG(
        "Cannot create tracing session, no tracing backend ready for type=%d",
        backend_type);
  });

  return std::unique_ptr<TracingSession>(
      new TracingSessionImpl(this, session_id));
}

void TracingMuxerImpl::InitializeInstance(const TracingInitArgs& args) {
  if (instance_)
    PERFETTO_FATAL("Tracing already initialized");
  instance_ = new TracingMuxerImpl(args);
}

TracingMuxer::~TracingMuxer() = default;

static_assert(std::is_same<internal::BufferId, BufferID>::value,
              "public's BufferId and tracing/core's BufferID diverged");

}  // namespace internal
}  // namespace perfetto