The "buffering" logic was active even if the stream cache was disabled.
This is contrary to what the manpage says. It also breaks playback
because of another bug: the demuxer cache is smaller than 2 seconds,
and thus the resume condition never becomes true.
Explicitly run this code only if the stream cache is enabled. Also, fix
the underlying problem of the breakage, and resume when the demuxer
thread stops reading in any case, not just on EOF.
Broken by previous commit. Unbreaks playback of local files.
Add the --cache-secs option, which literally overrides the value of
--demuxer-readahead-secs if the stream cache is active. The default
value is very high (10 seconds), which means it can act as network
cache.
Remove the old behavior of trying to pause once the byte cache runs
low. Instead, do something similar wit the demuxer cache. The nice
thing is that we can guess how many seconds of video it has cached,
and we can make better decisions. But for now, apply a relatively
naive heuristic: if the cache is below 0.5 secs, pause, and wait
until at least 2 secs are available.
Note that due to timestamp reordering, the estimated cached duration
of video might be inaccurate, depending on the file format. If the
file format has DTS, it's easy, otherwise the duration will seemingly
jump back and forth.
The purpose of the unconditional pthread_cond_signal() when reading
cached DEMUXER_CTRLs and STREAM_CTRLs was apparently to update the
stream cache state. Otherwise, the cached fields would never be updated
when the stream is e.g. paused.
The same could be said about other CTRLs, but these aren't as important,
since they are normally updated while reading packet data.
In order to reduce wakeups, make this logic explicit.
This is a simplification, because it lets us use the AVPacket
functions, instead of handling the details manually.
It also allows the libavcodec rawvideo decoder to use reference
counting, so it doesn't have to memcpy() the full image data. The change
in av_common.c enables this.
This change is somewhat risky, because we rely on the following AVPacket
implementation details and assumptions:
- av_packet_ref() doesn't access the input padding, and just copies the
data. By the API, AVPacket is always padded, and we violate this. The
lavc implementation would have to go out of its way to make this a
real problem, though.
- We hope that the way we make the AVPacket refcountable in av_common.c
is actually supported API-usage. It's hard to tell whether it is.
Of course we still use our own "old" demux_packet struct, just so that
libav* API usage is somewhat isolated.
If a packet is appended to a stream, and there were already packets
queued, nothing about the state changed, as far as the user (i.e. the
player) is concerned. Thus no wakeup is needed.
The pthread_cond_signal() call following this is not interesting - it
will simply be a NOP if there are actually no waiters.
--demuxer-readahead-secs now controls how much the demuxer should
readahead by an amount of seconds. This is based on the raw packet
timestamps. It's not always very exact. For example, h264 in Matroska
does not store any linear timestamps (only PTS values which are going
to be reordered by the decoder), so this heuristic is usually off by
several hundred milliseconds.
The decision whether to readahead is basically OR-ed with the other
--demuxer-readahead-packets options. Change the manpage descriptions
to subtly convey these semantics.
Switching tracks caused cached_demux_control() to catch the command to
switch tracks, even if no thread was running. Thus, the tracks were
never really switched, and EOF happened immediately on playback start.
Fix it by not using the cache at all if the demuxer thread is disabled.
The cache code still has to be called somewhere, though, because it
handles stream metadata update.
Regression from today.
Because why not.
This can lead to reordering of operations between seeking and track
switching (happens when the demuxer wakes up after seek and track
switching operations were queued). Do the track switching strictly
before seeks if there is a chance of reordering, which guarantees that
the seek position will always start with key frames. The reverse
(seeking, then switching) does not really have any advantages.
(Not sure if the player relies on this behavior.)
Use OPT_KEYVALUELIST() for all places where AVOptions are directly set
from mpv command line options. This allows escaping values, better
diagnostics (also no more "pal"), and somehow reduces code size.
Remove the old crappy option parser (av_opts.c).
This happens apparently randomly with rtmp:// and after seeks. This
eventually leads to audio decoding returning an EOF status, which
basically disables audio sync. This will lead to audio desync, even if
audio decoding later "recovers" when the demuxer actually returns audio
packets.
Hack-fix this by special-casing EAGAIN.
This didn't work, because the timebase was wrong. According to the
ffmpeg doxygen, if the stream index is -1 (which is what we used), the
timebase is AV_TIME_BASE. But this didn't work, and it really expected
the stream's timebase. Quite "surprising", since this feature
(avio_seek_time) is used by rtmp only.
Fixing this properly is too hard, so hack-fix our way around it.
STREAM_CTRL_SEEK_TO_TIME is also used by DVD/BD, so a new
STREAM_CTRL_AVSEEK is added. We simply pass-through the request
verbatim.
This tells the demuxer thread that it should seek, instead of waiting
until the demuxer thread is ready.
Care has to be taken about the state between seek request and actual
seeking: newly demuxed packets have to be discarded. We can't just
flush when doing the actual seek, because the user thread could read
these packets.
I'm wondering if this could lead to issues due to relaxed ordering of
operations. But it should be fine, since seeking influences packet
reading only, and seeking is always strictly done before that.
Currently, this will have no advantages; unless audio is disabled. Then
seeking as well as normal playback can be non-blocking.
Instead of starting to fill the packet queue if at least 1 stream is
selected, wait until there is at least 1 stream had new packets
requested.
In theory this is cleaner, because it allows you to e.g. do a seek and
then reselect streams without losing packets. Seeking marks all streams
as inactive, and without this new logic, the thread would read new
packets anyway right after seek.
This fixes the same symptom as the previous commit, but when the demuxer
thread is enabled. In this case, if nothing was read from the demuxer,
the STREAM_CTRLs weren't updated either. To the player, this looked like
the stream cache was never making progress, so playback was kept paused.
It can happen that read_packet() doesn't read a packet, even if it
succeeds. Typically this is because a packet was read, but then thrown
away, because it's not part of a selected stream. The result would be a
bogus EOF condition.
Fix by explicitly checking for EOF.
In corner cases, it might be possible that a demux_read_packet_async()
call fails to make the demuxer thread to read more packets.
If a packet is queued, the function will simply return a packet, without
marking the stream as active. As a consequence, read_packet() might
decide not to read any further packets, and the demuxer will never read
a packet and wake up the playback thread.
This was originally done to align it with demux_read_packet() semantics;
just drop this.
demux_read_any_packet() attempts to call read_packet(), but if no stream
is active, it can decide not to read anything. The function will return
NULL, which implies EOF. Fix this by explicitly
setting demux_stream->active if needed.
Also use dequeue_packet() instead of demux_read_packet(), because it's
cleaner. (Shouldn't change behavior.)
Possibly fixes#938.
We told the demuxer that a pipe (if stream cache is enabled) is
seekable. This is because the stream cache is technically seekable, it's
just that seeking may fail at runtime if a non-cached byte range is
requested.
This caused libavformat to issue seeks on initialization (at least when
piping mp4 youtube videos). Initialization failed completely after
spamming tons of error messages.
So, if an unseekable stream is cached, tell the demuxer that the file is
not seekable. This gets reversed later (when printing a message about
caching an unseekable stream), so the user can still try his luck by
issuing a seek command. The important part is that libavformat
initialization will not take code paths that will unnecessarily seek for
whatever reasons.
CC: @mpv-player/stable: regression from 0.3.x
It was easy to get into a wakeup feedback loop on EOF. The reason that
EOF is complicated is that we try to retry reading when EOF is reached,
in case the EOF state actually disappears (e.g. when watching a
currently downloaded file).
This feature is probably worthless, since in practice you have to do a
seek to "unstuck" it anyway, but since the old code also did this, we
want to keep this behavior for now.
Avoid the feedback loop by introducing another EOF flag (last_eof), that
contains the actual previous EOF state, and is not overwritten when
retrying reading. Wakeup is skipped if the EOF state didn't change.
Also, actually call the wakeup callback when EOF is detected.
The line that adds "ds->active = false;" actually does nothing, but in
theory it's cleaner.
The old FFmpeg API and the new Libav API disagree about mp4 display
rotation direction. Well, whatever, fix it trial-and-error-style.
CC: @mpv-player/stable: add
This adds a thread to the demuxer which reads packets asynchronously.
It will do so until a configurable minimum packet queue size is
reached. (See options.rst additions.)
For now, the thread is disabled by default. There are some corner cases
that have to be fixed, such as fixing cache behavior with webradios.
Note that most interaction with the demuxer is still blocking, so if
e.g. network dies, the player will still freeze. But this change will
make it possible to remove most causes for freezing.
Most of the new code in demux.c actually consists of weird caches to
compensate for thread-safety issues (with the previously single-threaded
design), or to avoid blocking by having to wait on the demuxer thread.
Most of the changes in the player are due to the fact that we must not
access the source stream directly. the demuxer thread already accesses
it, and the stream stuff is not thread-safe.
For timeline stuff (like ordered chapters), we enable the thread for the
current segment only. We also clear its packet queue on seek, so that
the remaining (unconsumed) readahead buffer doesn't waste memory.
Keep in mind that insane subtitles (such as ASS typesetting muxed into
mkv files) will practically disable the readahead, because the total
queue size is considered when checking whether the minimum queue size
was reached.
For OGG audio files, we usually merge the per-stream metadata back to
the file-global metadata. Don't do that for OGM, because with OGM most
metadata is actually per-stream.
libdvdnav can actually jump into the middle of the DVD (e.g. scene
selection menus do that). Then time display is incorrect: we start from
0, even though playback time is somewhere else. This really matters when
seeking. If the display time mismatches, a small relative seek will
apparently jump to the beginning of the movie.
Fix this by initializing the PTS stuff on opening. We have to do this
after some small amount of data has been read from the stream (because
libdvdnav is crap and doesn't always update the time between seeks and
the first read; also see STREAM_CTRL_GET_CURRENT_TIME remarks in
cache.c; although this was not observed when testing with scene
selection menus). On the other hand, we want to do it before opening the
demuxer, because that will read large amounts of data and likely will
change the stream position.
Also see commit 49813670.
Technically needed, but not strictly. It seems it works without in
practice, because demux_lavf.c reads exactly one packet for fill_buffer
call, so there are never packets queued.
We used a complicated and approximate method to cache the stream
timestamp, which is basically per-byte. (To reduce overhead, it was only
cached per 8KB-block, so it was approximate.)
Simplify this, and read/keep the timestamp only on discontinuities. This
is when demux_disc.c actually needs the timestamp.
Note that caching is currently disabled for dvdnav, but we still read
the timestamp only after some data is read. libdvdread behaves well, but
I don't know about libbluray, and the previous code also read the
timestamp only after reading data, so try to keep it safe.
Also drop the start_time offset. It wouldn't be correct anymore if used
with the cache, and the idea behind it wasn't very sane either (making
the player to offset the initial playback time to 0).
This should work now, at least kind of. Note that actual success depends
on the behavior of the underlying lib{dvd{nav,read},bluray}
implementation, which could go very wrong.
In the worst case, it could happen that the underlying implementation
seeks a long time before the seek target time. In this case, the player
will just decode video until the target time is reached, even if that
requires e.g. decoding 30 mintues of video before refreshing.
In the not-so-bad but still bad case, it would just miss the seek
target, and seek past it.
In my tests, it works mostly ok, though. Seeking backwards usually
fails, unless something like --hr-seek-demuxer-offset=1 is used (this
makes it seek to 1 second before the target, which may or may not be
enough to compensate for the DVD/BD imprecision).
This is a pretty big change. Instead of doing a half-hearted passthrough
of the playback timestamp, we attempt to rewrite the raw MPEG timestamps
such that they match with the playback time.
We add the offset between raw start timestamp and playback time to the
packet timestamps. This is the easy part; but the problem is with
timestamp resets. We simply detect timestamp discontinuities by checking
whether they are more than 500ms apart (large enough for all video
faster than 2 FPS and audio with reasonable framesizes/samplerates), and
adjust the timestamp offset accordingly.
This should work pretty well. There may be some problems with subtitles.
If the first packet after a timestamp reset is a subtitle instead of
video, it will fail. Also, selecting multiple audio or video streams
won't work (but mpv doesn't allow selecting several anyway). Trying to
demux subtitles with no video stream enabled will probably fail.
Untested with Bluray, because I have no Bluray sample.
Background:
libdvdnav/libdvdread/libbluray make this relatively hard. They return a
raw MPEG (PS/TS) byte stream, and additionally to that provide a
function to retrieve the current "playback" time. The playback time is
what should be displayed to the user, while the MPEG timestamps can be
completely different. Even worse, the MPEG timestamps can reset. Since
we use the libavformat demuxer (instead of parsing the MPEG packets in
the DVD/BD code), it's hard to associate between these timestamps. As a
result, the time display is special cased in the playloop, and of low
quality (updates only all 1 or 2 seconds, sometimes is incorrect). The
fact that the stream cache can be between demuxer and the stream source
makes things worse.
All the libs seem to provide an event that tells whether timestamps are
resetting. But since this signalling is byte based, it's hard to connect
it to the demuxed MPEG packets. It might be possible to create some sort
of table mapping file positions to discontinuities and new timestamps.
(For simplicity, this table could be 2 entries large, sufficient to
catch all discontinuities if the distance between them is larger than
the total buffering.)
It can happen that demux_fill_buffer() adds more than 1 packet, and then
the packets would add up. Affects demux_disc.c only (nothing else uses
this function).
Suggested by tholin on github issue #882.
This is not entirely clean, but the fields we're accessing might be
considered internal to libavformat. On the other hand, existence of the
fields is guaranteed by the ABI, and nothing in the libavformat doxygen
suggestes they're not allowed to be accessed.
CC: @mpv-player/stable
DVD and Bluray (and to some extent cdda) require awful hacks all over
the codebase to make them work. The main reason is that they act like
container, but are entirely implemented on the stream layer. The raw
mpeg data resulting from these streams must be "extended" with the
container-like metadata transported via STREAM_CTRLs. The result were
hacks all over demux.c and some higher-level parts.
Add a "disc" pseudo-demuxer, and move all these hacks and special-cases
to it.
(Again.)
This time, we simply make it event-based, as it should be. This is done
for both demuxer metadata and stream metadata.
For some ogg-over-icy streams, 2 updates are reported on stream start.
This is because libavformat reports an update right on start, while
including the same info in the "static" metadata. I don't know if that's
a bug or a feature.
It's unlikely that files with multiple audio tracks and with replaygain
actually happen, but this change might help avoid minor corner cases
with later changes.
Recently, libavformat added demuxers to open image files like normal
demuxers. This is a good thing, but for now they interfere with the
operation of demux_mf. Add them to the blacklist until there is a proper
solution.
(The list doesn't contain _all_ recognized image formats, just those
that might interfere with demux_mf.)
CC: @mpv-player/stable
This returned a stream error value directly to libavformat, which can't
make sense. For example STREAM_ERROR (0) means success in libavformat
error codes. (The meaning of the libavformat read_seek return value is
underdocumented too.)
Something like "char *s = ...; isdigit(s[0]);" triggers undefined
behavior, because char can be signed, and thus s[0] can be a negative
value. The is*() functions require unsigned char _or_ EOF. EOF is a
special value outside of unsigned char range, thus the argument to the
is*() functions can't be a char.
This undefined behavior can actually trigger crashes if the
implementation of these functions e.g. uses lookup tables, which are
then indexed with out-of-range values.
Replace all <ctype.h> uses with our own custom mp_is*() functions added
with misc/ctype.h. As a bonus, these functions are locale-independent.
(Although currently, we _require_ C locale for other reasons.)
Some of these might be security relevant.
The RealAudio code was especially bad. I'm not sure if all RealAudio
stuff still plays correctly; I didn't have that many samples for
testing. Some checks might be unnecessary or overcomplicated compared
to the (obfuscated) nature of the code.
CC: @mpv-player/stable
Also clarify the semantics.
It seems --idx didn't do anything. Possibly it used to change how the
now removed legacy demuxers like demux_avi used to behave. Or maybe
it was accidental.
--forceidx basically becomes --index=force. It's possible that new
index modes will be added in the future, so I'm keeping it
extensible, instead of e.g. creating --force-index.
Probably "needed" to get the correct alignment, although I'm not aware
of actual breakages or performance issues.
In fact we should probably always just allocate AVPackets, but for now
use the simple fix.
FFmpeg requires a bullshit padding after each input buffer, and they
just increased that padding without warning and without ABI or API bump.
We need this only in one file (although mp_image hardcodes something
similar, for which no FFmpeg API define is available), so drop our own
define.
While I'm not very fond of "const", it's important for declarations
(it decides whether a symbol is emitted in a read-only or read/write
section). Fix all these cases, so we have writeable global data only
when we really need.
Convert all these commands to properties. (Except tv_last_channel, not
sure what to do with this.) Also, internally, don't access stream
details directly, but dispatch commands with stream ctrls.
Many of the new properties are a bit strange, because they're write-
only. Also remove some OSD output these commands produced, because I
couldn't be bothered to port these.
In general, this makes everything much cleaner, and will also make it
easier to e.g. move the demuxer to its own thread.
Don't bother updating input.conf, but changes.rst documents how old
commands map to the new ones.
Mostly untested, due to lack of hardware.
In most places where af_fmt2bits is called to get the bits/sample, the
result is immediately converted to bytes/sample. Avoid this by getting
bytes/sample directly by introducing af_fmt2bps.
The i_bps members of the sh_audio and dev_video structs are mostly used
for displaying the average audio and video bitrates. Keeping them in
bits-per-second avoids truncating them to bytes-per-second and changing
them back lateron.
Stop using it in most places, and prefer STREAM_CTRL_GET_SIZE. The
advantage is that always the correct size will be used. There can be no
doubt anymore whether the end_pos value is outdated (as it happens often
with files that are being downloaded).
Some streams still use end_pos. They don't change size, and it's easier
to emulate STREAM_CTRL_GET_SIZE using end_pos, instead of adding a
STREAM_CTRL_GET_SIZE implementation to these streams.
Make sure int64_t is always used for STREAM_CTRL_GET_SIZE (it was
uint64_t before).
Remove the seek flags mess, and replace them with a seekable flag. Every
stream must set it consistently now, and an assertion in stream.c checks
this. Don't distinguish between streams that can only be forward or
backwards seeked, since we have no such stream types.
stream.start_pos was needed for optical media only, and (apparently) not
for very good reasons. Just get rid of it.
For stream_dvd, we don't need to do anything. Byte seeking was already
removed from it earlier.
For stream_cdda and stream_vcd, emulate the start_pos by offsetting the
stream pos as seen by the rest of mpv.
The bits in discnav.c and loadfile.c were for dealing with the code
seeking back to the start in demux.c. Handle this differently by
assuming the demuxer is always initialized with the stream at start
position, and instead seek back if initializing the demuxer fails.
Remove the --sb option, which worked by modifying stream.start_pos. If
someone really wants this option, it could be added back by creating a
"slice" stream (actually ffmpeg already has such a thing).
Drop: sami, vplayer, rt, pjs, mpsub, aqt, jacosub. None of these seem
to be actually in use, except sami. Sami is very complex, and the
results subreader produces are not very useful.
For all these formats, there are still parsers in FFmpeg. We remove the
subreader implementation, because it might contain security relevant
bugs and such. (This is old, unmaintained C string parsing code, written
in times where absolutely nobody cared about security. The kind of
awesome code.)
We keep the other formats, because they're (mostly) commonly used and
relatively simple, for UTF16 support (still missing in FFmpeg), and for
the sake of Libav.
Caused failure to detect .pls files, because they were misdetected as
m3u. The problem is that "forcing playlist files" and "forcing a
specific playlist format" are not really treated separate, and in both
cases p->force is set to true. This made m3u detect all files as m3u
if --playlist was used. So correctly check whether the file format is
actually being probed or not.
Because the http playlist URL I had for testing claimed to be m3u by
file extension and mime type, but didn't have the header.
Note that this actually changes behavior only in the case the format is
detected by mime type. Then p->force will be set before calling the
parser, and the header becomes optional.
mp3 has a hack lowering the probescore for format detection. This is
because detecting mp3s is hard due to their nature, and the fact that
ID3v2 tags are sometimes several megabytes big.
When playing mp3 from network, the mime-type is usually set, and that
matches the format hack entry meant for webradios, overriding the normal
mp3 entry. This can lead to network mp3s not being detected. Lower the
network case to the same probescore as on-disk mp3s. The difference is
that for network mp3s, we don't load the full probe-buffer, and we lower
the amount of audio the demuxer will read to collect data on opening
(0.5 seconds instead of typically 5 seconds).
Also remove MSGL_SMODE and friends.
Note: The indent in options.rst was added to work around a bug in
ReportLab that causes the PDF manual build to fail.
VP9 packets can contain 2 frames in some video packets (from which 1
frame is invisible). Due to a design mismatch between libvpx and the
libavcodec vp9 decoder, libvpx can take the "full" packets, but lavc vp9
can not. The consequence is that we have to split the packets if we want
to feed them to the lavc codec.
This is not entirely correct yet: timestamp handling is missing.
--demuxer=lavf and ffmpeg native utilities have the same problem. We can
fix this only once the ffmpeg VP9 parser is fixed.
wm4
Andrey Morozov
Marcoen Hirschberg
Martin Herkt