[Cast] Halt AudioSender transmission when too many frames are in-flight.

The approach here is to literally copy most of the logic in VideoSender
over into AudioSender, since VideoSender already solves a number of
transmission/re-transmission issues using heuristics that should work
well for audio.  (This also brings us much closer to being able to merge
AudioSender and VideoSender into one implementation.)

Testing: Confirmed correct halting and recovery behavior between
cast_sender_app and cast_receiver_app (SIGSTOPP'ed and SIGCONT'ed each
to simulate a temporary outage).  Also tested a Chromium sender with a
Chromecast receiver.

BUG=380023

Review URL: https://codereview.chromium.org/340903003

git-svn-id: svn://svn.chromium.org/chrome/trunk/src@278323 0039d316-1c4b-4281-b951-d872f2087c98

media/cast/audio_sender/audio_encoder.cc

@@ -9,14 +9,12 @@
 #include "base/bind.h"
 #include "base/bind_helpers.h"
 #include "base/location.h"
-#include "base/logging.h"
 #include "base/stl_util.h"
 #include "base/sys_byteorder.h"
 #include "base/time/time.h"
 #include "media/base/audio_bus.h"
 #include "media/cast/cast_defines.h"
 #include "media/cast/cast_environment.h"
-#include "media/cast/logging/logging_defines.h"
 #include "third_party/opus/src/include/opus.h"
 
 namespace media {
@@ -33,28 +31,6 @@ const int kFrameDurationMillis = 1000 / kFramesPerSecond;  // No remainder!
 // coming in too slow with respect to the capture timestamps.
 const int kUnderrunThresholdMillis = 3 * kFrameDurationMillis;
 
-void LogAudioFrameEncodedEvent(
-    const scoped_refptr<media::cast::CastEnvironment>& cast_environment,
-    base::TimeTicks event_time,
-    media::cast::RtpTimestamp rtp_timestamp,
-    uint32 frame_id,
-    size_t frame_size) {
-  if (!cast_environment->CurrentlyOn(CastEnvironment::MAIN)) {
-    cast_environment->PostTask(
-        CastEnvironment::MAIN,
-        FROM_HERE,
-        base::Bind(&LogAudioFrameEncodedEvent,
-                   cast_environment, event_time,
-                   rtp_timestamp, frame_id, frame_size));
-    return;
-  }
-  cast_environment->Logging()->InsertEncodedFrameEvent(
-      event_time, media::cast::FRAME_ENCODED, media::cast::AUDIO_EVENT,
-      rtp_timestamp, frame_id,
-      static_cast<int>(frame_size), /* key_frame - unused */ false,
-      /*target_bitrate - unused*/ 0);
-}
-
 }  // namespace
 
 
@@ -150,11 +126,6 @@ class AudioEncoder::ImplBase
       audio_frame->reference_time = frame_capture_time_;
 
       if (EncodeFromFilledBuffer(&audio_frame->data)) {
-        LogAudioFrameEncodedEvent(cast_environment_,
-                                  cast_environment_->Clock()->NowTicks(),
-                                  audio_frame->rtp_timestamp,
-                                  audio_frame->frame_id,
-                                  audio_frame->data.size());
         cast_environment_->PostTask(
             CastEnvironment::MAIN,
             FROM_HERE,

media/cast/audio_sender/audio_sender.h

@@ -14,6 +14,8 @@
 #include "base/time/time.h"
 #include "media/base/audio_bus.h"
 #include "media/cast/cast_config.h"
+#include "media/cast/cast_environment.h"
+#include "media/cast/logging/logging_defines.h"
 #include "media/cast/rtcp/rtcp.h"
 #include "media/cast/rtp_timestamp_helper.h"
 
@@ -22,8 +24,12 @@ namespace cast {
 
 class AudioEncoder;
 
-// This class is not thread safe.
-// It's only called from the main cast thread.
+// Not thread safe. Only called from the main cast thread.
+// This class owns all objects related to sending audio, objects that create RTP
+// packets, congestion control, audio encoder, parsing and sending of
+// RTCP packets.
+// Additionally it posts a bunch of delayed tasks to the main thread for various
+// timeouts.
 class AudioSender : public RtcpSenderFeedback,
                     public base::NonThreadSafe,
                     public base::SupportsWeakPtr<AudioSender> {
@@ -38,6 +44,10 @@ class AudioSender : public RtcpSenderFeedback,
     return cast_initialization_status_;
   }
 
+  // Note: It is not guaranteed that |audio_frame| will actually be encoded and
+  // sent, if AudioSender detects too many frames in flight.  Therefore, clients
+  // should be careful about the rate at which this method is called.
+  //
   // Note: It is invalid to call this method if InitializationResult() returns
   // anything but STATUS_AUDIO_INITIALIZED.
   void InsertAudio(scoped_ptr<AudioBus> audio_bus,
@@ -46,31 +56,98 @@ class AudioSender : public RtcpSenderFeedback,
   // Only called from the main cast thread.
   void IncomingRtcpPacket(scoped_ptr<Packet> packet);
 
- private:
-  void ResendPackets(
-      const MissingFramesAndPacketsMap& missing_frames_and_packets);
+ protected:
+  // Protected for testability.
+  virtual void OnReceivedCastFeedback(const RtcpCastMessage& cast_feedback)
+      OVERRIDE;
 
+ private:
+  // Schedule and execute periodic sending of RTCP report.
   void ScheduleNextRtcpReport();
   void SendRtcpReport(bool schedule_future_reports);
 
+  // Schedule and execute periodic checks for re-sending packets.  If no
+  // acknowledgements have been received for "too long," AudioSender will
+  // speculatively re-send certain packets of an unacked frame to kick-start
+  // re-transmission.  This is a last resort tactic to prevent the session from
+  // getting stuck after a long outage.
+  void ScheduleNextResendCheck();
+  void ResendCheck();
+  void ResendForKickstart();
+
+  // Returns true if there are too many frames in flight, as defined by the
+  // configured target playout delay plus simple logic.  When this is true,
+  // InsertAudio() will silenty drop frames instead of sending them to the audio
+  // encoder.
+  bool AreTooManyFramesInFlight() const;
+
   // Called by the |audio_encoder_| with the next EncodedFrame to send.
   void SendEncodedAudioFrame(scoped_ptr<transport::EncodedFrame> audio_frame);
 
-  virtual void OnReceivedCastFeedback(const RtcpCastMessage& cast_feedback)
-      OVERRIDE;
-
-  scoped_refptr<CastEnvironment> cast_environment_;
+  const scoped_refptr<CastEnvironment> cast_environment_;
+
+  // The total amount of time between a frame's capture/recording on the sender
+  // and its playback on the receiver (i.e., shown to a user).  This is fixed as
+  // a value large enough to give the system sufficient time to encode,
+  // transmit/retransmit, receive, decode, and render; given its run-time
+  // environment (sender/receiver hardware performance, network conditions,
+  // etc.).
+  const base::TimeDelta target_playout_delay_;
+
+  // Sends encoded frames over the configured transport (e.g., UDP).  In
+  // Chromium, this could be a proxy that first sends the frames from a renderer
+  // process to the browser process over IPC, with the browser process being
+  // responsible for "packetizing" the frames and pushing packets into the
+  // network layer.
   transport::CastTransportSender* const transport_sender_;
+
+  // Maximum number of outstanding frames before the encoding and sending of
+  // new frames shall halt.
+  const int max_unacked_frames_;
+
+  // Encodes AudioBuses into EncodedFrames.
   scoped_ptr<AudioEncoder> audio_encoder_;
-  RtpTimestampHelper rtp_timestamp_helper_;
+  const int configured_encoder_bitrate_;
+
+  // Manages sending/receiving of RTCP packets, including sender/receiver
+  // reports.
   Rtcp rtcp_;
+
+  // Records lip-sync (i.e., mapping of RTP <--> NTP timestamps), and
+  // extrapolates this mapping to any other point in time.
+  RtpTimestampHelper rtp_timestamp_helper_;
+
+  // Counts how many RTCP reports are being "aggressively" sent (i.e., one per
+  // frame) at the start of the session.  Once a threshold is reached, RTCP
+  // reports are instead sent at the configured interval + random drift.
   int num_aggressive_rtcp_reports_sent_;
 
+  // This is "null" until the first frame is sent.  Thereafter, this tracks the
+  // last time any frame was sent or re-sent.
+  base::TimeTicks last_send_time_;
+
+  // The ID of the last frame sent.  Logic throughout AudioSender assumes this
+  // can safely wrap-around.  This member is invalid until
+  // |!last_send_time_.is_null()|.
+  uint32 last_sent_frame_id_;
+
+  // The ID of the latest (not necessarily the last) frame that has been
+  // acknowledged.  Logic throughout AudioSender assumes this can safely
+  // wrap-around.  This member is invalid until |!last_send_time_.is_null()|.
+  uint32 latest_acked_frame_id_;
+
+  // Counts the number of duplicate ACK that are being received.  When this
+  // number reaches a threshold, the sender will take this as a sign that the
+  // receiver hasn't yet received the first packet of the next frame.  In this
+  // case, AudioSender will trigger a re-send of the next frame.
+  int duplicate_ack_counter_;
+
   // If this sender is ready for use, this is STATUS_AUDIO_INITIALIZED.
   CastInitializationStatus cast_initialization_status_;
 
-  // Used to map the lower 8 bits of the frame id to a RTP timestamp. This is
-  // good enough as we only use it for logging.
+  // This is a "good enough" mapping for finding the RTP timestamp associated
+  // with a video frame. The key is the lowest 8 bits of frame id (which is
+  // what is sent via RTCP). This map is used for logging purposes.
   RtpTimestamp frame_id_to_rtp_timestamp_[256];
 
   // NOTE: Weak pointers must be invalidated before all other member variables.

media/cast/video_sender/video_sender.cc

@@ -78,7 +78,7 @@ VideoSender::VideoSender(
   media::cast::transport::CastTransportVideoConfig transport_config;
   transport_config.codec = video_config.codec;
   transport_config.rtp.config = video_config.rtp_config;
-  transport_config.rtp.max_outstanding_frames = max_unacked_frames_ + 1;
+  transport_config.rtp.max_outstanding_frames = max_unacked_frames_;
   transport_sender_->InitializeVideo(transport_config);
 
   rtcp_.SetCastReceiverEventHistorySize(kReceiverRtcpEventHistorySize);

media/cast/video_sender/video_sender.h

@@ -76,15 +76,13 @@ class VideoSender : public RtcpSenderFeedback,
   void ScheduleNextRtcpReport();
   void SendRtcpReport(bool schedule_future_reports);
 
-  // Schedule and execute periodic checks for re-sending frames.  If no
+  // Schedule and execute periodic checks for re-sending packets.  If no
   // acknowledgements have been received for "too long," VideoSender will
-  // speculatively re-send the frame just after |latest_acked_frame_id_| (the
-  // whole frame).  This is a last resort tactic to prevent the session from
+  // speculatively re-send certain packets of an unacked frame to kick-start
+  // re-transmission.  This is a last resort tactic to prevent the session from
   // getting stuck after a long outage.
   void ScheduleNextResendCheck();
   void ResendCheck();
-
-  // Resend certain packets of an unacked frame to kick start re-transmission.
   void ResendForKickstart();
 
   // Returns true if there are too many frames in flight, as defined by the
@@ -169,10 +167,7 @@ class VideoSender : public RtcpSenderFeedback,
 
   // This is a "good enough" mapping for finding the RTP timestamp associated
   // with a video frame. The key is the lowest 8 bits of frame id (which is
-  // what is sent via RTCP). This map is used for logging purposes. The only
-  // time when this mapping will be incorrect is when it receives an ACK for a
-  // old enough frame such that 8-bit wrap around has already occurred, which
-  // should be pretty rare.
+  // what is sent via RTCP). This map is used for logging purposes.
   RtpTimestamp frame_id_to_rtp_timestamp_[256];
 
   // NOTE: Weak pointers must be invalidated before all other member variables.