ha-core/homeassistant/components/stream/worker.py

602 lines
25 KiB
Python

"""Provides the worker thread needed for processing streams."""
from __future__ import annotations
from collections import defaultdict, deque
from collections.abc import Callable, Generator, Iterator, Mapping
import contextlib
import datetime
from io import SEEK_END, BytesIO
import logging
from threading import Event
from typing import Any, cast
import attr
import av
from homeassistant.core import HomeAssistant
from . import redact_credentials
from .const import (
AUDIO_CODECS,
HLS_PROVIDER,
MAX_MISSING_DTS,
MAX_TIMESTAMP_GAP,
PACKETS_TO_WAIT_FOR_AUDIO,
SEGMENT_CONTAINER_FORMAT,
SOURCE_TIMEOUT,
)
from .core import (
STREAM_SETTINGS_NON_LL_HLS,
KeyFrameConverter,
Part,
Segment,
StreamOutput,
StreamSettings,
)
from .diagnostics import Diagnostics
from .fmp4utils import read_init
from .hls import HlsStreamOutput
_LOGGER = logging.getLogger(__name__)
class StreamWorkerError(Exception):
"""An exception thrown while processing a stream."""
class StreamEndedError(StreamWorkerError):
"""Raised when the stream is complete, exposed for facilitating testing."""
class StreamState:
"""Responsible for trakcing output and playback state for a stream.
Holds state used for playback to interpret a decoded stream. A source stream
may be reset (e.g. reconnecting to an rtsp stream) and this object tracks
the state to inform the player.
"""
def __init__(
self,
hass: HomeAssistant,
outputs_callback: Callable[[], Mapping[str, StreamOutput]],
diagnostics: Diagnostics,
) -> None:
"""Initialize StreamState."""
self._stream_id: int = 0
self.hass = hass
self._outputs_callback: Callable[
[], Mapping[str, StreamOutput]
] = outputs_callback
# sequence gets incremented before the first segment so the first segment
# has a sequence number of 0.
self._sequence = -1
self._diagnostics = diagnostics
@property
def sequence(self) -> int:
"""Return the current sequence for the latest segment."""
return self._sequence
def next_sequence(self) -> int:
"""Increment the sequence number."""
self._sequence += 1
return self._sequence
@property
def stream_id(self) -> int:
"""Return the readonly stream_id attribute."""
return self._stream_id
def discontinuity(self) -> None:
"""Mark the stream as having been restarted."""
# Preserving sequence and stream_id here keep the HLS playlist logic
# simple to check for discontinuity at output time, and to determine
# the discontinuity sequence number.
self._stream_id += 1
# Call discontinuity to fix incomplete segment in HLS output
if hls_output := self._outputs_callback().get(HLS_PROVIDER):
cast(HlsStreamOutput, hls_output).discontinuity()
@property
def outputs(self) -> list[StreamOutput]:
"""Return the active stream outputs."""
return list(self._outputs_callback().values())
@property
def diagnostics(self) -> Diagnostics:
"""Return diagnostics object."""
return self._diagnostics
class StreamMuxer:
"""StreamMuxer re-packages video/audio packets for output."""
def __init__(
self,
hass: HomeAssistant,
video_stream: av.video.VideoStream,
audio_stream: av.audio.stream.AudioStream | None,
audio_bsf: av.BitStreamFilter | None,
stream_state: StreamState,
stream_settings: StreamSettings,
) -> None:
"""Initialize StreamMuxer."""
self._hass = hass
self._segment_start_dts: int = cast(int, None)
self._memory_file: BytesIO = cast(BytesIO, None)
self._av_output: av.container.OutputContainer = None
self._input_video_stream: av.video.VideoStream = video_stream
self._input_audio_stream: av.audio.stream.AudioStream | None = audio_stream
self._audio_bsf = audio_bsf
self._audio_bsf_context: av.BitStreamFilterContext = None
self._output_video_stream: av.video.VideoStream = None
self._output_audio_stream: av.audio.stream.AudioStream | None = None
self._segment: Segment | None = None
# the following 3 member variables are used for Part formation
self._memory_file_pos: int = cast(int, None)
self._part_start_dts: int = cast(int, None)
self._part_has_keyframe = False
self._stream_settings = stream_settings
self._stream_state = stream_state
self._start_time = datetime.datetime.utcnow()
def make_new_av(
self,
memory_file: BytesIO,
sequence: int,
input_vstream: av.video.VideoStream,
input_astream: av.audio.stream.AudioStream | None,
) -> tuple[
av.container.OutputContainer,
av.video.VideoStream,
av.audio.stream.AudioStream | None,
]:
"""Make a new av OutputContainer and add output streams."""
container = av.open(
memory_file,
mode="w",
format=SEGMENT_CONTAINER_FORMAT,
container_options={
**{
# Removed skip_sidx - see https://github.com/home-assistant/core/pull/39970
# "cmaf" flag replaces several of the movflags used, but too recent to use for now
"movflags": "frag_custom+empty_moov+default_base_moof+frag_discont+negative_cts_offsets+skip_trailer+delay_moov",
# Sometimes the first segment begins with negative timestamps, and this setting just
# adjusts the timestamps in the output from that segment to start from 0. Helps from
# having to make some adjustments in test_durations
"avoid_negative_ts": "make_non_negative",
"fragment_index": str(sequence + 1),
"video_track_timescale": str(int(1 / input_vstream.time_base)),
},
# Only do extra fragmenting if we are using ll_hls
# Let ffmpeg do the work using frag_duration
# Fragment durations may exceed the 15% allowed variance but it seems ok
**(
{
"movflags": "empty_moov+default_base_moof+frag_discont+negative_cts_offsets+skip_trailer+delay_moov",
# Create a fragment every TARGET_PART_DURATION. The data from each fragment is stored in
# a "Part" that can be combined with the data from all the other "Part"s, plus an init
# section, to reconstitute the data in a "Segment".
# The LL-HLS spec allows for a fragment's duration to be within the range [0.85x,1.0x]
# of the part target duration. We use the frag_duration option to tell ffmpeg to try to
# cut the fragments when they reach frag_duration. However, the resulting fragments can
# have variability in their durations and can end up being too short or too long. With a
# video track with no audio, the discrete nature of frames means that the frame at the
# end of a fragment will sometimes extend slightly beyond the desired frag_duration.
# If there are two tracks, as in the case of a video feed with audio, there is an added
# wrinkle as the fragment cut seems to be done on the first track that crosses the desired
# threshold, and cutting on the audio track may also result in a shorter video fragment
# than desired.
# Given this, our approach is to give ffmpeg a frag_duration somewhere in the middle
# of the range, hoping that the parts stay pretty well bounded, and we adjust the part
# durations a bit in the hls metadata so that everything "looks" ok.
"frag_duration": str(
self._stream_settings.part_target_duration * 9e5
),
}
if self._stream_settings.ll_hls
else {}
),
},
)
output_vstream = container.add_stream(template=input_vstream)
# Check if audio is requested
output_astream = None
if input_astream:
if self._audio_bsf:
self._audio_bsf_context = self._audio_bsf.create()
self._audio_bsf_context.set_input_stream(input_astream)
output_astream = container.add_stream(
template=self._audio_bsf_context or input_astream
)
return container, output_vstream, output_astream
def reset(self, video_dts: int) -> None:
"""Initialize a new stream segment."""
self._part_start_dts = self._segment_start_dts = video_dts
self._segment = None
self._memory_file = BytesIO()
self._memory_file_pos = 0
(
self._av_output,
self._output_video_stream,
self._output_audio_stream,
) = self.make_new_av(
memory_file=self._memory_file,
sequence=self._stream_state.next_sequence(),
input_vstream=self._input_video_stream,
input_astream=self._input_audio_stream,
)
if self._output_video_stream.name == "hevc":
self._output_video_stream.codec_tag = "hvc1"
def mux_packet(self, packet: av.Packet) -> None:
"""Mux a packet to the appropriate output stream."""
# Check for end of segment
if packet.stream == self._input_video_stream:
if (
packet.is_keyframe
and (packet.dts - self._segment_start_dts) * packet.time_base
>= self._stream_settings.min_segment_duration
):
# Flush segment (also flushes the stub part segment)
self.flush(packet, last_part=True)
# Mux the packet
packet.stream = self._output_video_stream
self._av_output.mux(packet)
self.check_flush_part(packet)
self._part_has_keyframe |= packet.is_keyframe
elif packet.stream == self._input_audio_stream:
if self._audio_bsf_context:
self._audio_bsf_context.send(packet)
while packet := self._audio_bsf_context.recv():
packet.stream = self._output_audio_stream
self._av_output.mux(packet)
return
packet.stream = self._output_audio_stream
self._av_output.mux(packet)
def create_segment(self) -> None:
"""Create a segment when the moov is ready."""
self._segment = Segment(
sequence=self._stream_state.sequence,
stream_id=self._stream_state.stream_id,
init=read_init(self._memory_file),
# Fetch the latest StreamOutputs, which may have changed since the
# worker started.
stream_outputs=self._stream_state.outputs,
start_time=self._start_time,
)
self._memory_file_pos = self._memory_file.tell()
self._memory_file.seek(0, SEEK_END)
def check_flush_part(self, packet: av.Packet) -> None:
"""Check for and mark a part segment boundary and record its duration."""
if self._memory_file_pos == self._memory_file.tell():
return
if self._segment is None:
# We have our first non-zero byte position. This means the init has just
# been written. Create a Segment and put it to the queue of each output.
self.create_segment()
# When using delay_moov, the moov is not written until a moof is also ready
# Flush the moof
self.flush(packet, last_part=False)
else: # These are the ends of the part segments
self.flush(packet, last_part=False)
def flush(self, packet: av.Packet, last_part: bool) -> None:
"""Output a part from the most recent bytes in the memory_file.
If last_part is True, also close the segment, give it a duration,
and clean up the av_output and memory_file.
There are two different ways to enter this function, and when
last_part is True, packet has not yet been muxed, while when
last_part is False, the packet has already been muxed. However,
in both cases, packet is the next packet and is not included in
the Part.
This function writes the duration metadata for the Part and
for the Segment. However, as the fragmentation done by ffmpeg
may result in fragment durations which fall outside the
[0.85x,1.0x] tolerance band allowed by LL-HLS, we need to fudge
some durations a bit by reporting them as being within that
range.
Note that repeated adjustments may cause drift between the part
durations in the metadata and those in the media and result in
playback issues in some clients.
"""
# Part durations should not exceed the part target duration
adjusted_dts = min(
packet.dts,
self._part_start_dts
+ self._stream_settings.part_target_duration / packet.time_base,
)
if last_part:
# Closing the av_output will write the remaining buffered data to the
# memory_file as a new moof/mdat.
self._av_output.close()
# With delay_moov, this may be the first time the file pointer has
# moved, so the segment may not yet have been created
if not self._segment:
self.create_segment()
elif not self._part_has_keyframe:
# Parts which are not the last part or an independent part should
# not have durations below 0.85 of the part target duration.
adjusted_dts = max(
adjusted_dts,
self._part_start_dts
+ 0.85 * self._stream_settings.part_target_duration / packet.time_base,
)
# Undo dts adjustments if we don't have ll_hls
if not self._stream_settings.ll_hls:
adjusted_dts = packet.dts
assert self._segment
self._memory_file.seek(self._memory_file_pos)
self._hass.loop.call_soon_threadsafe(
self._segment.async_add_part,
Part(
duration=float(
(adjusted_dts - self._part_start_dts) * packet.time_base
),
has_keyframe=self._part_has_keyframe,
data=self._memory_file.read(),
),
(
segment_duration := float(
(adjusted_dts - self._segment_start_dts) * packet.time_base
)
)
if last_part
else 0,
)
if last_part:
# If we've written the last part, we can close the memory_file.
self._memory_file.close() # We don't need the BytesIO object anymore
self._start_time += datetime.timedelta(seconds=segment_duration)
# Reinitialize
self.reset(packet.dts)
else:
# For the last part, these will get set again elsewhere so we can skip
# setting them here.
self._memory_file_pos = self._memory_file.tell()
self._part_start_dts = adjusted_dts
self._part_has_keyframe = False
def close(self) -> None:
"""Close stream buffer."""
self._av_output.close()
self._memory_file.close()
class PeekIterator(Iterator):
"""An Iterator that may allow multiple passes.
This may be consumed like a normal Iterator, however also supports a
peek() method that buffers consumed items from the iterator.
"""
def __init__(self, iterator: Iterator[av.Packet]) -> None:
"""Initialize PeekIterator."""
self._iterator = iterator
self._buffer: deque[av.Packet] = deque()
# A pointer to either _iterator or _buffer
self._next = self._iterator.__next__
def __iter__(self) -> Iterator:
"""Return an iterator."""
return self
def __next__(self) -> av.Packet:
"""Return and consume the next item available."""
return self._next()
def _pop_buffer(self) -> av.Packet:
"""Consume items from the buffer until exhausted."""
if self._buffer:
return self._buffer.popleft()
# The buffer is empty, so change to consume from the iterator
self._next = self._iterator.__next__
return self._next()
def peek(self) -> Generator[av.Packet, None, None]:
"""Return items without consuming from the iterator."""
# Items consumed are added to a buffer for future calls to __next__
# or peek. First iterate over the buffer from previous calls to peek.
self._next = self._pop_buffer
for packet in self._buffer:
yield packet
for packet in self._iterator:
self._buffer.append(packet)
yield packet
class TimestampValidator:
"""Validate ordering of timestamps for packets in a stream."""
def __init__(self) -> None:
"""Initialize the TimestampValidator."""
# Decompression timestamp of last packet in each stream
self._last_dts: dict[av.stream.Stream, int | float] = defaultdict(
lambda: float("-inf")
)
# Number of consecutive missing decompression timestamps
self._missing_dts = 0
def is_valid(self, packet: av.Packet) -> bool:
"""Validate the packet timestamp based on ordering within the stream."""
# Discard packets missing DTS. Terminate if too many are missing.
if packet.dts is None:
if self._missing_dts >= MAX_MISSING_DTS:
raise StreamWorkerError(
f"No dts in {MAX_MISSING_DTS+1} consecutive packets"
)
self._missing_dts += 1
return False
self._missing_dts = 0
# Discard when dts is not monotonic. Terminate if gap is too wide.
prev_dts = self._last_dts[packet.stream]
if packet.dts <= prev_dts:
gap = packet.time_base * (prev_dts - packet.dts)
if gap > MAX_TIMESTAMP_GAP:
raise StreamWorkerError(
f"Timestamp overflow detected: last dts = {prev_dts}, dts = {packet.dts}"
)
return False
self._last_dts[packet.stream] = packet.dts
return True
def is_keyframe(packet: av.Packet) -> Any:
"""Return true if the packet is a keyframe."""
return packet.is_keyframe
def get_audio_bitstream_filter(
packets: Iterator[av.Packet], audio_stream: Any
) -> av.BitStreamFilterContext | None:
"""Return the aac_adtstoasc bitstream filter if ADTS AAC is detected."""
if not audio_stream:
return None
for count, packet in enumerate(packets):
if count >= PACKETS_TO_WAIT_FOR_AUDIO:
# Some streams declare an audio stream and never send any packets
_LOGGER.warning("Audio stream not found")
break
if packet.stream == audio_stream:
# detect ADTS AAC and disable audio
if audio_stream.codec.name == "aac" and packet.size > 2:
with memoryview(packet) as packet_view:
if packet_view[0] == 0xFF and packet_view[1] & 0xF0 == 0xF0:
_LOGGER.debug(
"ADTS AAC detected. Adding aac_adtstoaac bitstream filter"
)
return av.BitStreamFilter("aac_adtstoasc")
break
return None
def stream_worker(
source: str,
pyav_options: dict[str, str],
stream_settings: StreamSettings,
stream_state: StreamState,
keyframe_converter: KeyFrameConverter,
quit_event: Event,
) -> None:
"""Handle consuming streams."""
if av.library_versions["libavformat"][0] >= 59 and "stimeout" in pyav_options:
# the stimeout option was renamed to timeout as of ffmpeg 5.0
pyav_options["timeout"] = pyav_options["stimeout"]
del pyav_options["stimeout"]
try:
container = av.open(source, options=pyav_options, timeout=SOURCE_TIMEOUT)
except av.AVError as err:
raise StreamWorkerError(
f"Error opening stream ({err.type}, {err.strerror}) {redact_credentials(str(source))}"
) from err
try:
video_stream = container.streams.video[0]
except (KeyError, IndexError) as ex:
raise StreamWorkerError("Stream has no video") from ex
keyframe_converter.create_codec_context(codec_context=video_stream.codec_context)
try:
audio_stream = container.streams.audio[0]
except (KeyError, IndexError):
audio_stream = None
if audio_stream and audio_stream.name not in AUDIO_CODECS:
audio_stream = None
# Some audio streams do not have a profile and throw errors when remuxing
if audio_stream and audio_stream.profile is None:
audio_stream = None
# Disable ll-hls for hls inputs
if container.format.name == "hls":
for field in attr.fields(StreamSettings):
setattr(
stream_settings,
field.name,
getattr(STREAM_SETTINGS_NON_LL_HLS, field.name),
)
stream_state.diagnostics.set_value("container_format", container.format.name)
stream_state.diagnostics.set_value("video_codec", video_stream.name)
if audio_stream:
stream_state.diagnostics.set_value("audio_codec", audio_stream.name)
dts_validator = TimestampValidator()
container_packets = PeekIterator(
filter(dts_validator.is_valid, container.demux((video_stream, audio_stream)))
)
def is_video(packet: av.Packet) -> Any:
"""Return true if the packet is for the video stream."""
return packet.stream.type == "video"
# Have to work around two problems with RTSP feeds in ffmpeg
# 1 - first frame has bad pts/dts https://trac.ffmpeg.org/ticket/5018
# 2 - seeking can be problematic https://trac.ffmpeg.org/ticket/7815
#
# Use a peeking iterator to peek into the start of the stream, ensuring
# everything looks good, then go back to the start when muxing below.
try:
# Get the required bitstream filter
audio_bsf = get_audio_bitstream_filter(container_packets.peek(), audio_stream)
# Advance to the first keyframe for muxing, then rewind so the muxing
# loop below can consume.
first_keyframe = next(
filter(lambda pkt: is_keyframe(pkt) and is_video(pkt), container_packets)
)
# Deal with problem #1 above (bad first packet pts/dts) by recalculating
# using pts/dts from second packet. Use the peek iterator to advance
# without consuming from container_packets. Skip over the first keyframe
# then use the duration from the second video packet to adjust dts.
next_video_packet = next(filter(is_video, container_packets.peek()))
# Since the is_valid filter has already been applied before the following
# adjustment, it does not filter out the case where the duration below is
# 0 and both the first_keyframe and next_video_packet end up with the same
# dts. Use "or 1" to deal with this.
start_dts = next_video_packet.dts - (next_video_packet.duration or 1)
first_keyframe.dts = first_keyframe.pts = start_dts
except StreamWorkerError as ex:
container.close()
raise ex
except StopIteration as ex:
container.close()
raise StreamEndedError("Stream ended; no additional packets") from ex
except av.AVError as ex:
container.close()
raise StreamWorkerError(
"Error demuxing stream while finding first packet: %s" % str(ex)
) from ex
muxer = StreamMuxer(
stream_state.hass,
video_stream,
audio_stream,
audio_bsf,
stream_state,
stream_settings,
)
muxer.reset(start_dts)
# Mux the first keyframe, then proceed through the rest of the packets
muxer.mux_packet(first_keyframe)
with contextlib.closing(container), contextlib.closing(muxer):
while not quit_event.is_set():
try:
packet = next(container_packets)
except StreamWorkerError as ex:
raise ex
except StopIteration as ex:
raise StreamEndedError("Stream ended; no additional packets") from ex
except av.AVError as ex:
raise StreamWorkerError("Error demuxing stream: %s" % str(ex)) from ex
muxer.mux_packet(packet)
if packet.is_keyframe and is_video(packet):
keyframe_converter.packet = packet