Issue https://webrtc-codereview.appspot.com/4459004/ was commited as
r5259, after which flakiness was detected and a rollback was performed
at r5261.
Patch Set 1 of this issue is the code submitted in r5259. Subsequent
patch sets fixes a race condition which caused the seen problems.
The root cause was a dead lock between a thread sending rtp packets and
and a timed module processing thread:
webrtc::RTPSender::BitrateUpdated() // Get RTPSender stats lock
webrtc::Bitrate::Process() // Get Bitrate lock
webrtc::RTPSender::ProcessBitrate()
webrtc::ModuleRtpRtcpImpl::Process()
...
webrtc::Bitrate::Update() // Get Bitrate lock
webrtc::RTPSender::UpdateRtpStats() // Get RTPSender stats lock
webrtc::RTPSender::SendToNetwork()
...
This is fixed in Bitrate::Process() by releasing the lock before
calling the callback.
BUG=2235
R=mflodman@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/5619004
git-svn-id: http://webrtc.googlecode.com/svn/trunk@5281 4adac7df-926f-26a2-2b94-8c16560cd09d
This CL makes the padding packets get their own RTP timestamps,
rather than having the same timestamp as the last sent video
packet. The purpose is to solve Issue 2611, where the overuse-
detector does not react to padding packets.
A test was implemented to verify that the padding packets do
get their own timestamps.
BUG=2611
R=stefan@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/3869004
git-svn-id: http://webrtc.googlecode.com/svn/trunk@5125 4adac7df-926f-26a2-2b94-8c16560cd09d
Analogous to the recent libjingle change: http://cl/54929753-p10.
This supports scoped_ptr<T[]> and scoped_ptr<C, FreeDeleter> rather
than scoped_array and scoped_ptr_malloc respectively.
- Add Chromium's template-based COMPILE_ASSERT. We didn't have this
previously in order to support the macro in C. Instead, move the
existing macro to compile_assert_c.h.
- Additionally copy the move.h and template_util.h depedencies and add
the WARN_UNUSED_RESULT macro.
- Leave scoped_array and scoped_ptr_malloc for now, but mark as
deprecated.
- Remove scoped_ptr foo(NULL) use. The default constructor handles it.
- Remove the now redundant COMPILE_ASSERT from peerconnection_jni.cc.
- Add a CHECK_ARRAY_SIZE macro to rtp_format_vp8_unittest.cc to remove
some repeated code.
TESTED=trybots
R=pbos@webrtc.org, tommi@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/2449005
git-svn-id: http://webrtc.googlecode.com/svn/trunk@5015 4adac7df-926f-26a2-2b94-8c16560cd09d
As far as I observe, there are several reasons for this:
1. webrtc/modules/audio_coding/neteq4/neteq_impl.cc : 870
assert(new_codec_);
This is related to webrtc/modules/audio_coding/neteq4/decision_logic_normal.cc : 81
kUndefined can happen without new_codec_ being set
2. webrtc/modules/audio_coding/neteq4/neteq_impl.cc : 745
assert(sync_buffer_->FutureLength() >= expand_->overlap_length());
3. some other assert triggered.
The above happens not very often and goes away with no assertion.
3. (most common, this CL addresses this)
webrtc/modules/rtp_rtcp/source/rtp_receiver_audio.cc : 201
payload_data_length = payload_length - rtp_header->header.paddingLength;
There are situations that
payload_length < rtp_header->header.paddingLength;
OLD ACM + NetEq3 can handle this:
a) webrtc/modules/audio_coding/main/source/acm_neteq.cc : 477
int16_t payload_length = static_cast<int16_t>(length_payload);
payload_length becomes negative in this situation
b) webrtc/modules/audio_coding/neteq/recin.c
WebRtcNetEQ_RecInInternal() handles negative payload length
I do not want to touch VoE, so I tried to let ACM2 and NetEq4 handle negative payload length.
This CL does not follow the exact way of OLD ACM + NetEq3. I stopped negative payload length at ACM and did not allow it go to NetEq4.
To try this, apply my uploaded patch : https://webrtc-codereview.appspot.com/2281004/
Let me know if you see better solutions.
R=henrik.lundin@webrtc.org, stefan@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/2292005
git-svn-id: http://webrtc.googlecode.com/svn/trunk@4860 4adac7df-926f-26a2-2b94-8c16560cd09d
I mistakenly ommitted the checks when logging.h was ported from
libjingle to webrtc. This caused a significant CPU cost for logs which
were later filtered out anyway.
Verified with LS_VERBOSE logging in neteq4, running:
$ out/Release/modules_unittests \
--gtest_filter=NetEqDecodingTest.TestBitExactness \
--gtest_repeat=50 > time.txt
$ grep "case ran" time.txt | grep "[0-9]* ms" -o | sort
Results on a MacBook Retina, averaged over 5 runs:
Verbose logs disabled: 666 ms
Exisiting implementation, verbose logs enabled: 944 ms (1.42x)
New implementation, verbose logs enabled: 673 ms (1.01x)
BUG=2314
R=henrik.lundin@webrtc.org, henrike@webrtc.org, kjellander@webrtc.org, turaj@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/2160005
git-svn-id: http://webrtc.googlecode.com/svn/trunk@4682 4adac7df-926f-26a2-2b94-8c16560cd09d
- Allow padding to be sent on an ssrc which doesn't produce video, therefore
never having the last_packet_marker_bit_ set.
- Add the random timestamp offset to all padding packets.
- Store the capture time of padding packets to properly create an offset.
BUG=2258
TEST=trybots and a new test which will be committed separately.
R=mflodman@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/2060005
git-svn-id: http://webrtc.googlecode.com/svn/trunk@4566 4adac7df-926f-26a2-2b94-8c16560cd09d
