Reland of Enable audio streams to send padding. (patchset #4 id:60001 of https://codereview.webrtc.…

webrtc/modules/rtp_rtcp/source/rtp_sender.cc

Index: webrtc/modules/rtp_rtcp/source/rtp_sender.cc
diff --git a/webrtc/modules/rtp_rtcp/source/rtp_sender.cc b/webrtc/modules/rtp_rtcp/source/rtp_sender.cc
index add7c21e9be22e109c15227de2b7ee0ac4336c7f..02f3f72defd1b3e913fe54b422201c414870fca6 100644
--- a/webrtc/modules/rtp_rtcp/source/rtp_sender.cc
+++ b/webrtc/modules/rtp_rtcp/source/rtp_sender.cc
@@ -36,6 +36,7 @@ namespace webrtc {
 namespace {
 // Max in the RFC 3550 is 255 bytes, we limit it to be modulus 32 for SRTP.
 constexpr size_t kMaxPaddingLength = 224;
+constexpr size_t kMinAudioPaddingLength = 50;
 constexpr int kSendSideDelayWindowMs = 1000;
 constexpr size_t kRtpHeaderLength = 12;
 constexpr uint16_t kMaxInitRtpSeqNumber = 32767;  // 2^15 -1.
@@ -215,7 +216,7 @@ int32_t RTPSender::RegisterRtpHeaderExtension(RTPExtensionType type,
   return -1;
 }
 
-bool RTPSender::IsRtpHeaderExtensionRegistered(RTPExtensionType type) {
+bool RTPSender::IsRtpHeaderExtensionRegistered(RTPExtensionType type) const {
   rtc::CritScope lock(&send_critsect_);
   return rtp_header_extension_map_.IsRegistered(type);
 }
@@ -481,11 +482,20 @@ size_t RTPSender::TrySendRedundantPayloads(size_t bytes_to_send,
 }
 
 size_t RTPSender::SendPadData(size_t bytes, int probe_cluster_id) {
-  // Always send full padding packets. This is accounted for by the
-  // RtpPacketSender, which will make sure we don't send too much padding even
-  // if a single packet is larger than requested.
-  size_t padding_bytes_in_packet =
-      std::min(MaxPayloadSize(), kMaxPaddingLength);
+  size_t padding_bytes_in_packet;
+  if (audio_configured_) {
+    // Allow smaller padding packets for audio.
+    padding_bytes_in_packet =
+        std::max(std::min(bytes, MaxPayloadSize()), kMinAudioPaddingLength);

[terelius, 2017/02/03 14:26:14]

Note this order of operation allows sending padding packets that are  larger
than MaxPayloadSize(). It this the intent? Or do we assume that
kMinAudioPaddingLength < MaxPayloadSize()?

[stefan-webrtc, 2017/02/03 14:39:25]

Thanks, that's not the intent.

+    if (padding_bytes_in_packet > kMaxPaddingLength)
+      padding_bytes_in_packet = kMaxPaddingLength;
+  } else {
+    // Always send full padding packets. This is accounted for by the
+    // RtpPacketSender, which will make sure we don't send too much padding even
+    // if a single packet is larger than requested.
+    // We do this to avoid frequently sending small packets on higher bitrates.
+    padding_bytes_in_packet = std::min(MaxPayloadSize(), kMaxPaddingLength);
+  }
   size_t bytes_sent = 0;
   while (bytes_sent < bytes) {
     int64_t now_ms = clock_->TimeInMilliseconds();
@@ -502,9 +512,15 @@ size_t RTPSender::SendPadData(size_t bytes, int probe_cluster_id) {
       timestamp = last_rtp_timestamp_;
       capture_time_ms = capture_time_ms_;
       if (rtx_ == kRtxOff) {
+        if (payload_type_ == -1)
+          break;
         // Without RTX we can't send padding in the middle of frames.
-        if (!last_packet_marker_bit_)
+        // For audio marker bits doesn't mark the end of a frame and frames

[terelius, 2017/02/03 14:26:14]

What does the marker bit mean for audio? Though unlikely, what happens if we add
padding in the middle of an audio frame? Can it corrupt the decoder state?

[stefan-webrtc, 2017/02/03 14:39:25]

It has no clear meaning, but one possible interpretation is "the beginning of a
talk spurt"
http://www.cs.columbia.edu/~hgs/rtp/faq.html#marker

It should not affect the decoder state. I don't even think padding packets are
passed down to the decoder. They don't contain anything, so they won't add bits
to the bitstream.

It's a bit worse for video because we have to keep track of sequence numbers in
a more complicated way, and we also need to know when a frame starts and ends.
In theory we could remove this check also for video, but it would require us to
nack some padding packets.

[terelius, 2017/02/03 14:42:19]

Acknowledged.

+        // are usually a single packet, so for now we don't apply this rule
+        // for audio.
+        if (!audio_configured_ && !last_packet_marker_bit_) {
           break;
+        }
         ssrc = ssrc_;
         sequence_number = sequence_number_;
         ++sequence_number_;
@@ -796,7 +812,7 @@ bool RTPSender::IsFecPacket(const RtpPacketToSend& packet) const {
 }
 
 size_t RTPSender::TimeToSendPadding(size_t bytes, int probe_cluster_id) {
-  if (audio_configured_ || bytes == 0)
+  if (bytes == 0)
     return 0;
   size_t bytes_sent = TrySendRedundantPayloads(bytes, probe_cluster_id);
   if (bytes_sent < bytes)