The demuxer cache employs a strange method to make track switching
instant with caching enabled. Normally this would mean you have to wait
until the cache has played out (and you get new packets, including
packets from the newly selected track), or you have to perform a slow
high level seek (decoding video again etc.). The strange method is that
it performs a demuxer-level seek without a high level seek so it looks
like a continuous stream to the decoder, and the newly select stream
gets packets at the current playback position. This is called a refresh
seek.
This works only if some weird heuristics work. It needs a packet "unique
ID", for which it uses either dts or pts. The value must be strictly
monotonic increasing. If this doesn't work, the referesh seek can't be
executed, and the user has to wait until the end of the cache. Sometimes
there are files that simply do not work.
In the present case, there's actually a hack that we can extend. Packets
with unset position are likely generated by the parser, and the hack
which this commit touches simply attempts to make up a new (hopefully
unique) position value, even if the value itself makes no sense. It only
ha to be deterministic.
It turns out libavcodec sometimes output packets with repeating position
values. This commit tries to handle this case too with the same hack.
Fixes: #7498
Let's see how much everyone hates this. Leaving it enabled seems
problematic, because libavcodec returns an unspecific error if it
doesn't like it.
Fixes: #6300
See manpage additions. This has been a topic in MPlayer/mplayer2/mpv
since forever. But since libavcodec multi-threaded decoding was added,
I've always considered this pointless. libavcodec requires you to
"preload" it with packets, and then you can pretty much avoid blocking
on it, if decoding is fast enough.
But in some cases, a decoupled decoder thread _might_ help. Users have
for example come up with cases where decoding video in a separate
process and piping it as raw video to mpv helped. (Or my memory is
false, and it was about vapoursynth filtering, who knows.) So let's just
see whether this helps with anything.
Note that this would have been _much_ easier if libavcodec had an
asynchronous (or rather, non-blocking) API. It could probably have
easily gained that with a small change to its multi-threading code and a
small extension to its API, but I guess not.
Unfortunately, this uglifies f_decoder_wrapper quite a lot. Part of this
is due to annoying corner cases like legacy frame dropping and hardware
decoder state. These could probably be prettified later on.
There is also a change in playloop.c: this is because there is a need to
coordinate playback resets between demuxer thread, decoder thread, and
playback logic. I think this SEEK_BLOCK idea worked out reasonably well.
There are still a number of problems. For example, if the demuxer cache
is full, the decoder thread will simply block hard until the output
queue is full, which interferes with seeking. Could also be improved
later. Hardware decoding will probably die in a fire, because it will
run out of surfaces quickly. We could reduce the queue to size 1...
maybe later. We could update the queue options at runtime easily, but
currently I'm not going to bother.
I could only have put the lavc wrapper itself on a separate thread. But
there is some annoying interaction with EDL and backward playback shit,
and also you would have had to loop demuxer packets through the
playloop, so this sounded less annoying.
The food my mother made for us today was delicious.
Because audio uses the same code, also for audio (even if completely
pointless).
Fixes: #6926
This is supposed to enable communication between filter graphs on
separate threads. Having multiple threads makes only sense if they can
run concurrently with each other, which requires such an asynchronous
queue as a building block. (Probably.)
The basic idea is that you have two independent filters, which can be
each part of separate filter graphs, but which communicate into one
direction with an explicit queue. This is rather similar to unix pipes.
Just like unix pipes, the queue is limited in size, so that still some
data flow control enforced, and runaway memory usage is avoided.
This implementation is pretty dumb. In theory, you could avoid avoid
waking up the filter graphs in quite a lot of situations. For example,
you don't need to wake up the consumer filter if there are already
frames queued. Also, you could add "watermarks" that set a threshold at
which producer or consumer should be woken up to produce/consume more
frames (this would generally serve to "batch" multiple frames at once,
instead of performing high-frequency wakeups). But this is hard, so the
code is dumb. (I just deleted all related code when I still got
situations where wakeups were lost.)
This is actually salvaged and modified from a much older branch I had
lying around. It will be used in the next commit.
Instead of vague ideas about making different filter graphs on different
threads interact directly, this have no direct support. Instead, helpers
are required (such as added with the next commit).
Document it. Different root filters (i.e. separate filter graphs) are
now considered to be part of separate threads, so assert() if they're
found to accidentally interact.
Previously: 5.1 > 5.2 > luajit
Now: 5.2 > luajit > 5.1
I randomly decided that this should be done, since I suspect most
environments will prefer the highest Lua version anyway. There is not
much of a point picking the older one by default.
Maybe 5.1 should be dropped fully, but considering we need to stay
compatible with luajit, there is no particular incentive for this.
Preparation for a future commit. The demuxer queues might be read from
other threads than the one to issue the seek, and passing SEEK_BLOCK
with such a seek will provide a convenient way to synchronize this.
mp_filter_mark_async_progress() can asynchronously mark a filter for
processing, without waking up the filter thread. (It's some sort of
idiotic micro-optimization I guess?) But since it sets async_pending
without doing the wakeup, a mp_filter_wakeup() after this will do
nothing, and the wakeup is lost. Fix it by checking for the needed
wakeup separately.
Fortunately, nothing used this function yet, so there is no impact.
I may (optionally) move decoding to a separate thread in a future
change. It's a bit attractive to move the entire decoder wrapper to
there, so if the demuxer has a new packet, it doesn't have to wake up
the main thread, and can directly wake up the decoder. (Although that's
bullshit, since there's a queue in between, and libavcodec's
multi-threaded decoding plays cross-threads ping pong with packets
anyway. On the other hand, the main thread would still have to shuffle
the packets around, so whatever, just seems like better design.)
As preparation, there shouldn't be any mutable state exposed by the
wrapper. But there's still a large number of corner-caseish crap, so
just use setters/getters for them. This recorder thing will inherently
not work, so it'll have to be disabled if threads are used.
This is a bit painful, but probably still the right thing. Like
speculatively pulling teeth.
This explicitly depends on the pixfmt list from FFmpeg (done so to
easily spot regression, incompatible changes, and other unexpected
things).
Some local changes in mpv change some of the output. For pal8 an alias
was added back, and the [GENERIC] markers are removed because the mpv
aliases are not dependent on the mpv config anymore (which was
unnecessary).
The other changes are due to ffmpeg adding some new formats.
This was a hack that attempted to line up external audio tracks with
video. The problem is that if you do a keyframe seek backwards, video
will usually seek much farther back than audio (due to much higher
keyframe aka seek point distances). The hack somehow made seeking a 2
step process.
This existed in 4 different forms in the history of this code base, and
it was always very cumbersome. We mostly needed this for ytdl_hook (I
think?), which uses the 4th form, which is nicely confined to
demux_timeline and is unrelated to the "external" audio tracks in the
high level player.
Since this is (probably) not really widely needed anymore, get rid of
it. Better do this now, than when somehow rewriting all the seeking code
(which might happen in this decade or the next or so) and when it
wouldn't be easily revertable anymore in case we find we "really" need
it unlike expected.
There is no issue if hr-seeks are used. Also, you can still use edl
files to "bundle" multiple streams as if it was a single stream (this is
what ytdl_hook does now).
It seems sporadic errors are possible, such as connection timeouts.
Before the recent demuxer change, the demuxer thread retried many times
even on EOF, so an error was only interpreted as EOF once the decoder
queues ran out.
Change it to use EOF only. Since this may actually lead to the demuxer
thread being "stuck" and retrying forever (depending on libavformat API
behavior), I'm also adding a heuristic to prevent this, using a random
retry counter. This should not be necessary, but libavformat cannot be
trusted. (This retrying forever could be stopped by the user, but
obviously it would still heat the CPU for a longer time if the user is
not present.)
The "seekbarkeyframes" option is now interpreted such if it's true, the
player default is used. Too lazy to make this a choice option or
whatever; the Lua option parser doesn't have support for that anyway.
Someone who cares can adjust this.
Try to deal with various corner cases. But when I fix one thing, another
thing breaks. (And it's 50/50 whether I find the breakage immediately or
a few months later.) So results may vary.
The default for--hr-seek is changed to "default" (not creative enough to
find a better name). In this mode, audio seeking is exact if there is no
video, or if the video has only a single frame. This change is actually
pretty dumb, since audio frames are usually small enough that exact
seeking does not really add much. But it gets rid of some weird special
cases.
Internally, the most important change is that is_coverart and is_sparse
handling is merged. is_sparse was originally just a special case for
weird .ts streams that have the corresponding low-level flag set. The
idea is that they're pretty similar anyway, so this would reduce the
number of corner cases. But I'm not sure if this doesn't break the
original intended use case for it (I don't have a sample anyway).
This changes last-frame handling, and respects the duration of the last
frame only if audio is disabled. This is mostly "coincidental" due to
the need to make seeking past EOF trigger player exit, and is caused by
setting STATUS_EOF early. On the other hand, this might have been this
way before (see removed chunk close to it).
This tries to fix#7206 (hr-seeking past EOF does not stop playback)
again. Commit 57fbc9cd76 should have fixed this, but trying it again
(using that git revision), it often did not work. Whatever the fuck.
So add another dumb special case that will break within weeks. Note that
the check in handle_eof() had no effect, since execute_queued_seek() is
called later, which cancels EOF in the same case.
Hr-seek past the last frame instantly enters EOF, which means
handle_playback_time() will not set playback_pts to the video PTS (as
all video frames are skipped), which leads to the playback time being
taken from the last seek target. This results in confusing behavior,
especially since the seek time will be clipped to the file duration for
display, but not for further relative seeks.
Obviously, the time should be set to the last video frame, so use the
last video frame as fallback if both audio and video have ended. Also,
since the same problem exists with audio-only playback, add a fallback
for audio PTS too. We don't know which was the "last" fragment of media
played (to decide whether to use the audio or video PTS as the
fallback), but it doesn't matter since the maximum works.
This could lead to some undesired effects. In particular the audio PTS
is basically a bad guess, and is for example not clipped against --end.
(But the ridiculous way audio syncing and clamping currently works, I'm
not going to touch that shit unless I rewrite it completely.) The cover
art case is slightly broken: using --keep-open with keyframe seeks will
result in 0 as playback PTS (the video PTS). OK, who cares, it got late.
Also casually get rid of last_vo_pts, since that barely made any sense
at all.
Fixes: #7487
The seeking logic saves the last video frame it has seen (for example
for being able to seek to the last frame, or backstepping).
Unfortunately, the frame was fed back to the filtering pipeline in
situations when it shouldn't have. Then it's an out of order frame,
because it really saves the last _discarded_ frame.
For example, seeking to the end of a file with --keep-open, shift+up,
shift+down => invalid video pts warning due to saved_frame being fed
back.
Explicitly discard saved_frame when it's obviously not needed anymore.
The removed accesses to "r" are strictly speaking unrelated (just
const-propagating them).
In this case the video track has seek_start == seek_end, and due to the
"seek_start >= seek_end" condition, this was considered broken, and no
seek range was created, breaking cached seeking.
Fix this by allowing the case if they're equal, and a valid timestamp.
(NB: seeking backward in this will still jump to position 0, because it
is the video timestamp. This is unfortunately how it's supposed to work.
HR-seeks will also do this, but decode and skip the entire audio until
the seek target, so it will mostly appear to work.)
Exposed by commit b56e2efd5f. demux_timeline reported a bogus EOF if
"parallel" streams were used. If a virtual source reported EOF, it was
propagated as global EOF, without serving packets of other virtual
sources that have not ended yet.
Fix this by not reporting global EOF just because a source has not
returned a packet. Instead make the reader retry by returning no packet
and no EOF state, which will call d_read_packet() again with a different
source. Rely on the eof_reached flags to signal global EOF.
Since eof_reached is now more important, set it in a certain other case
when it apparently should have been set. do_read_next_packet()'s return
value is now ignored, so get rid of it.
This is useful with live streams, and it's not much worse than the h264
first packet hack, which reads some data anyway.
For some reason, the option wasn't even documented, so do that.
In addition, print the start time even if it's negative. That should not
be possible, but for some reason, the field is an int64_t copied from an
uint64_t so... whatever. Keeping the logging slightly more straight
forward is better anyway.
Remove some redundant fields that controlled or indicated whether the
demuxer was/could/should prefetch. Redefine how the eof/reading fields
work.
The in->eof field is now always valid, instead of weirdly being reset to
false in random situations. The in->reading field now corresponds to
whether the demuxer thread is working at all, and is reset if it stops
doing anything.
Also, I always found it stupid that dequeue_packet() forced the demuxer
thread to retry reading if it was EOF. This makes little sense, but was
probably added for files that are being appended to (running downloads).
It makes no sense, because if the cache really tried to read until file
EOF, it would encounter partial packets and throw errors, so all is lost
anyway. Plus stream_file now handles this better. So stop this behavior,
but add a temporary option that enables the old behavior.
I think checking for ds->eager when enabling prefetching never really
made sense (could be debated, but no, not really). On the other hand,
the change above exposed a missing wakeup in the backward demuxing code.
Some chances of regressions that could make it stuck in certain states
or so, or incorrect demuxer cache state reporting to the player
frontend.
Setting demux_set_stream_wakeup_cb() will make all sh_stream (i.e.
track) specific wakeups go to this callback. But the callback takes care
of only the sub_preload() case (where it tries to pre-load subtitles
from already parsed and memory-present subtitles in a blocking way).
The old code assumed that the normal demuxer wakeup callback is called.
This was disregarded when the newer code was added. (And actually, the
original plan was to make _all_ per-sh_stream wakeups go to specialized
callbacks to avoid wasted work. dec_sub really should set the callback
always, and propagate wakeups to the playloop code. But it's too far
into the night to write coherent code.)
I couldn't actually observe any manifestation of this bug. Normally, the
playloop wakes up for other reasons (such as driving audio and video
decoding), so the lost wakeups rarely matter.
A negative subtitle delay means that subtitles from the future should be
shown earlier. With muxed subtitles, subtitle packets are demuxed along
with audio and video packets. But since they are demuxed "lazily",
nothing guarantees that subtitle packets from the future are available
in time.
Typically, the user-observed effect is that subtitles do not appear at
all (or too late) with large negative --sub-delay values, but that using
--cache might fix this.
Make this behave better. Automatically extend read-ahead to as much as
needed by the subtitles. It seems it's the easiest to pass the subtitle
render timestamp to the demuxer in order to guarantee that everything is
read. This timestamp based approach might be fragile, so disable it if
no negative sub-delay is used.
As far as the player frontend part is concerned, this makes use of the
code path for external subtitles, which are not lazily demuxed, and may
already trigger waiting.
Fixes: #7484
Subtitle tracks are usually "lazy" (ds->eager=false), There are a number
of weird special cases associated with it. One of them is that they have
some sort of "temporary" EOF (to signal that there isn't a packet right
now, and the decoder should not block playback by waiting for more
packets). In a the next commit, I want to call mark_stream_eof() in case
of (some) of these temporary EOFs.
The problem is that mark_stream_eof() also calls the functions touched
by this commit. Basically they shouldn't do any complex work due to
these temporary EOFs (because they might happen very often). It turns
out that lazy tracks barely matter here: they do not extend the seek
range of a packet/EOF happens on them, they do not trigger seek range
joining, and they do not support backward demuxing.
This change should enable the following commit, while not causing any
behavior changes (i.e. bugs) with the current state.
args->client was deallocated if the FDs were closed and nothing
referenced it (IPC socket codes detected the closed sockets and
asynchronously killed the mpv_handle in args->client). The problem was
that args->log depended on it, and was also destroyed.
Fix this by duplicating the mp_log.
This has been part of the libmpv for a while, so the implementation in
the IPC code is quite simple: just pass the mpv_node representing the
value of the "command" field without further checks to
mpv_command_node().
The only problem are the IPC-specific commands, which essentially have
their own dispatch mechanism. They expect an array. I'm not going to
rewrite the dispatch mechanism, so these still work only with an array.
I decided make the other case explicit with cmd==NULL. (I could also
have set cmd=="", which would have avoided changing each if condition
since "" matches no existing command, but that felt dirty.)
I decided to make this explicit. The alternative would have been making
all commands asynchronous always, like a small note in the manpage
threatened. I think that could have caused compatibility issues.
As a design decision, this does not send a reply if an async command
started. This could be a good or bad idea, but in any case, it will make
async command look almost like synchronous ones, except they don't block
the IPC protocol.
The relatively recently added property update code has a race condition
when clients exit. It still tried to access mpv_handle during and after
it was destroyed.
The reason is that it unlocks the lock for the mpv_handle list (while
mpv_handle is locked), but nothing in mp_destroy_client() cares about
this case. The latter function locks mpv_handle only before/while it
makes sure all of its activity is coming to an end, such as asynchronous
requests, and property updates that are in progress. It did not include
the case when mp_client_send_property_changes() was still calling
send_client_property_changes() with mpv_handle locked.
Fix this by checking the mpv_handle.destroying field. This field can be
set only when mpv_handle is locked. While we're checking the lock, the
mpv_handle list is still locked, so at worst we might be at the point
before mp_destroy_client() locks the list again and finally destroys the
mpv_handle.
This is a hard to reproduce race condition which I spotted only once in
valgrind by chance, so the fix is unconfirmed.
The previous change ensured that these cannot fail anymore (much like in
original talloc). Change the APIs to not return a success value anymore,
to "cement" this.
The ta_ext_header was allocated on demand for allocations which have
child-allocations or destructors. In theory, it saved 2 words for every
TA leaf allocation. It had the very API-visible problem that setting a
parent or destructor could fail. (Although in most cases, the failure
was part of an allocation call anyway. Also, mpv code generally used the
early-failure variants, so it didn't matter.)
I think this was a bit too complex. These 2 words don't really matter;
if you have memory allocations where you are worried about overhead,
then these simply shouldn't use TA. Also, we never added new features to
TA that would have needed more "optional" header fields, which would
have justified the use of such a separately allocated header struct.
This uses quite straight-forward data structures. The only strange thing
is that ta_header.parent is NULL for most child allocations. That is
because we don't want to iterate over all children when the parent is
reallocated (yes, that is allowed, yes mpv makes use of it).
The new code has a few more special cases, because the list sentinel
concept isn't used anymore. Using it would have made the code more
unnatural/complex, because ta_ext_header doesn't exist anymore.
the actual character that made mpv crash is IDEOGRAPHIC COMMA
(U+3001, UTF-8: E3 80 81, 、) and that only in some specific
circumstances that could be reliably reproduced on my end.
using an NSString instead of the Swift String actually fixes that issues
even though they should technically do the exact same thing. i tested
all the other String initialisers, but they all had had the same issue.
this is kinda only a workaround till i can find a different way of
handling it.
this basically moves the remote command center to our mac events instead
of keeping it our Application, which is only available when started from
mpv itself. also make it independent of the NSApplication.
this also prevents a runtime crash
This was changed 6 years ago (444e583b6) and seemed to work fine. But it
does seem to cause issues with IceWM sometimes, while with StaticGravity
the problem is gone. Comparing both gravity values, reading the confused
source code comment, and reading the referenced commit message, I can't
determine what it even does, I just remove it.
Reproduction:
- start mpv in windowed mode, with 2 videos of different size
- switch to second video
- switch window with alt+tab
- switch back to mpv with alt+tab
- window moves to X=0
There's probably a better way to fix this. Please send a patch.
