summaryrefslogtreecommitdiffstats
path: root/DOCS/tech-overview.txt
blob: 4bb06ff764ee1822d485694fe7bc0cb2ebe6120b (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
This file intends to give a big picture overview of how mpv is structured.

player/*.c:
    Essentially makes up the player applications, including the main() function
    and the playback loop.

    Generally, it accesses all other subsystems, initializes them, and pushes
    data between them during playback.

    The structure is as follows (as of commit e13c05366557cb):
    * main():
        * basic initializations (e.g. init_libav() and more)
        * pre-parse command line (verbosity level, config file locations)
        * load config files (parse_cfgfiles())
        * parse command line, add files from the command line to playlist
          (m_config_parse_mp_command_line())
        * check help options etc. (call handle_help_options()), possibly exit
        * call mp_play_files() function that works down the playlist:
            * run idle loop (idle_loop()), until there are files in the
                playlist or an exit command was given (only if --idle it set)
            * actually load and play a file in play_current_file():
                * run all the dozens of functions to load the file and
                  initialize playback
                * run a small loop that does normal playback, until the file is
                  done or a command terminates playback
                  (on each iteration, run_playloop() is called, which is rather
                   big and complicated - it decodes some audio and video on
                   each frame, waits for input, etc.)
                * uninitialize playback
            * determine next entry on the playlist to play
            * loop, or exit if no next file or quit is requested
              (see enum stop_play_reason)
        * call mp_destroy()
    * run_playloop():
        * calls fill_audio_out_buffers()
            This checks whether new audio needs to be decoded, and pushes it
            to the AO.
        * calls write_video()
            Decode new video, and push it to the VO.
        * determines whether playback of the current file has ended
        * determines when to start playback after seeks
        * and calls a whole lot of other stuff
            (Really, this function does everything.)

    Things worth saying about the playback core:
    - most state is in MPContext (core.h), which is not available to the
      subsystems (and should not be made available)
    - the currently played tracks are in mpctx->current_tracks, and decoder
      state in track.dec/d_sub
    - the other subsystems rarely call back into the frontend, and the frontend
      polls them instead (probably a good thing)
    - one exceptions are wakeup callbacks, which notify a "higher" component
      of a changed situation in a subsystem

    I like to call the player/*.c files the "frontend".

ta.h & ta.c:
    Hierarchical memory manager inspired by talloc from Samba. It's like a
    malloc() with more features. Most importantly, each talloc allocation can
    have a parent, and if the parent is free'd, all children will be free'd as
    well. The parent is an arbitrary talloc allocation. It's either set by the
    allocation call by passing a talloc parent, usually as first argument to the
    allocation function. It can also be set or reset later by other calls (at
    least talloc_steal()). A talloc allocation that is used as parent is often
    called a talloc context.

    One very useful feature of talloc is fast tracking of memory leaks. ("Fast"
    as in it doesn't require valgrind.) You can enable it by setting the
    MPV_LEAK_REPORT environment variable to "1":
        export MPV_LEAK_REPORT=1
    Or permanently by building with --enable-ta-leak-report.
    This will list all unfree'd allocations on exit.

    Documentation can be found here:
        http://git.samba.org/?p=samba.git;a=blob;f=lib/talloc/talloc.h;hb=HEAD

    For some reason, we're still using API-compatible wrappers instead of TA
    directly. The talloc wrapper has only a subset of the functionality, and
    in particular the wrappers abort() on memory allocation failure.

    Note: unlike tcmalloc, jemalloc, etc., talloc() is not actually a malloc
          replacement. It works on top of system malloc and provides additional
          features that are supposed to make memory management easier.

player/command.c:
    This contains the implementation for client API commands and properties.
    Properties are essentially dynamic variables changed by certain commands.
    This is basically responsible for all user commands, like initiating
    seeking, switching tracks, etc. It calls into other player/*.c files,
    where most of the work is done, but also calls other parts of mpv.

player/core.h:
    Data structures and function prototypes for most of player/*.c. They are
    usually not accessed by other parts of mpv for the sake of modularization.

player/client.c:
    This implements the client API (libmpv/client.h). For the most part, this
    just calls into other parts of the player. This also manages a ringbuffer
    of events from player to clients.

options/options.h, options/options.c
    options.h contains the global option struct MPOpts. The option declarations
    (option names, types, and MPOpts offsets for the option parser) are in
    options.c. Most default values for options and MPOpts are in
    mp_default_opts at the end of options.c.

    MPOpts is unfortunately quite monolithic, but is being incrementally broken
    up into sub-structs. Many components have their own sub-option structs
    separate from MPOpts. New options should be bound to the component that uses
    them. Add a new option table/struct if needed.

    The global MPOpts still contains the sub-structs as fields, which serves to
    link them to the option parser. For example, an entry like this may be
    typical:

        {"", OPT_SUBSTRUCT(demux_opts, demux_conf)},

    This directs the option access code to include all options in demux_conf
    into the global option list, with no prefix (""), and as part of the
    MPOpts.demux_opts field. The MPOpts.demux_opts field is actually not
    accessed anywhere, and instead demux.c does this:

        struct m_config_cache *opts_cache =
            m_config_cache_alloc(demuxer, global, &demux_conf);
        struct demux_opts *opts = opts_cache->opts;

    ... to get a copy of its options.

    See m_config.h (below) how to access options.

    The actual option parser is spread over m_option.c, m_config.c, and
    parse_commandline.c, and uses the option table in options.c.

options/m_config.h & m_config.c:
    Code for querying and managing options. This (unfortunately) contains both
    declarations for the "legacy-ish" global m_config struct, and ways to access
    options in a threads-safe way anywhere, like m_config_cache_alloc().

    m_config_cache_alloc() lets anyone read, observe, and write options in any
    thread. The only state it needs is struct mpv_global, which is an opaque
    type that can be passed "down" the component hierarchy. For safety reasons,
    you should not pass down any pointers to option structs (like MPOpts), but
    instead pass down mpv_global, and use m_config_cache_alloc() (or similar)
    to get a synchronized copy of the options.

input/input.c:
    This translates keyboard input coming from VOs and other sources (such
    as remote control devices like Apple IR or client API commands) to the
    key bindings listed in the user's (or the builtin) input.conf and turns
    them into items of type struct mp_cmd. These commands are queued, and read
    by playloop.c. They get pushed with run_command() to command.c.

    Note that keyboard input and commands used by the client API are the same.
    The client API only uses the command parser though, and has its own queue
    of input commands somewhere else.

common/msg.h:
    All terminal output must go through mp_msg().

stream/*:
    File input is implemented here. stream.h/.c provides a simple stream based
    interface (like reading a number of bytes at a given offset). mpv can
    also play from http streams and such, which is implemented here.

    E.g. if mpv sees "http://something" on the command line, it will pick
    stream_lavf.c based on the prefix, and pass the rest of the filename to it.

    Some stream inputs are quite special: stream_dvd.c turns DVDs into mpeg
    streams (DVDs are actually a bunch of vob files etc. on a filesystem),
    stream_tv.c provides TV input including channel switching.

    Some stream inputs are just there to invoke special demuxers, like
    stream_mf.c. (Basically to make the prefix "mf://" do something special.)

demux/:
    Demuxers split data streams into audio/video/sub streams, which in turn
    are split in packets. Packets (see demux_packet.h) are mostly byte chunks
    tagged with a playback time (PTS). These packets are passed to the decoders.

    Most demuxers have been removed from this fork, and the only important and
    "actual" demuxers left are demux_mkv.c and demux_lavf.c (uses libavformat).
    There are some pseudo demuxers like demux_cue.c.

    The main interface is in demux.h. The stream headers are in stheader.h.
    There is a stream header for each audio/video/sub stream, and each of them
    holds codec information about the stream and other information.

    demux.c is a bit big, the main reason being that it contains the demuxer
    cache, which is implemented as a list of packets. The cache is complex
    because it support seeking, multiple ranges, prefetching, and so on.

video/:
    This contains several things related to audio/video decoding, as well as
    video filters.

    mp_image.h and img_format.h define how mpv stores decoded video frames
    internally.

video/decode/:
    vd_*.c are video decoders. (There's only vd_lavc.c left.) dec_video.c
    handles most of connecting the frontend with the actual decoder.

video/filter/:
    vf_*.c and vf.c form the video filter chain. They are fed by the video
    decoder, and output the filtered images to the VOs though vf_vo.c. By
    default, no video filters (except vf_vo) are used. vf_scale is automatically
    inserted if the video output can't handle the video format used by the
    decoder.

video/out/:
    Video output. They also create GUI windows and handle user input. In most
    cases, the windowing code is shared among VOs, like x11_common.c for X11 and
    w32_common.c for Windows. The VOs stand between frontend and windowing code.
    vo_gpu can pick a windowing system at runtime, e.g. the same binary can
    provide both X11 and Cocoa support on OSX.

    VOs can be reconfigured at runtime. A vo_reconfig() call can change the video
    resolution and format, without destroying the window.

    vo_gpu should be taken as reference.

audio/:
    format.h/format.c define the uncompressed audio formats. (As well as some
    compressed formats used for spdif.)

audio/decode/:
    ad_*.c and dec_audio.c handle audio decoding. ad_lavc.c is the
    decoder using ffmpeg. ad_spdif.c is not really a decoder, but is used for
    compressed audio passthrough.

audio/filter/:
    Audio filter chain. af_lavrresample is inserted if any form of conversion
    between audio formats is needed.

audio/out/:
    Audio outputs.

    Unlike VOs, AOs can't be reconfigured on a format change. On audio format
    changes, the AO will simply be closed and re-opened.

    There are wrappers to support for two types of audio APIs: push.c and
    pull.c. ao.c calls into one of these. They contain generic code to deal
    with the data flow these APIs impose.

    Note that mpv synchronizes the video to the audio. That's the reason
    why buggy audio drivers can have a bad influence on playback quality.

sub/:
    Contains subtitle and OSD rendering.

    osd.c/.h is actually the OSD code. It queries dec_sub.c to retrieve
    decoded/rendered subtitles. osd_libass.c is the actual implementation of
    the OSD text renderer (which uses libass, and takes care of all the tricky
    fontconfig/freetype API usage and text layouting).

    The VOs call osd.c to render OSD and subtitle (via e.g. osd_draw()). osd.c
    in turn asks dec_sub.c for subtitle overlay bitmaps, which relays the
    request to one of the sd_*.c subtitle decoders/renderers.

    Subtitle loading is in demux/. The MPlayer subreader.c is mostly gone - parts
    of it survive in demux_subreader.c. It's used as last fallback, or to handle
    some text subtitle types on Libav. It should go away eventually. Normally,
    subtitles are loaded via demux_lavf.c.

    The subtitles are passed to dec_sub.c and the subtitle decoders in sd_*.c
    as they are demuxed. All text subtitles are rendered by sd_ass.c. If text
    subtitles are not in the ASS format, the libavcodec subtitle converters are
    used (lavc_conv.c).

    Text subtitles can be preloaded, in which case they are read fully as soon
    as the subtitle is selected. In this case, they are effectively stored in
    sd_ass.c's internal state.

etc/:
    The file input.conf is actually integrated into the mpv binary by the
    build system. It contains the default keybindings.

Best practices and Concepts within mpv
======================================

General contribution etc.
-------------------------

See: DOCS/contribute.md

Error checking
--------------

If an error is relevant, it should be handled. If it's interesting, log the
error. However, mpv often keeps errors silent and reports failures somewhat
coarsely by propagating them upwards the caller chain. This is OK, as long as
the errors are not very interesting, or would require a developer to debug it
anyway (in which case using a debugger would be more convenient, and the
developer would need to add temporary debug printfs to get extremely detailed
information which would not be appropriate during normal operation).

Basically, keep a balance on error reporting. But always check them, unless you
have a good argument not to.

Memory allocation errors (OOM) are a special class of errors. Normally such
allocation failures are not handled "properly". Instead, abort() is called.
(New code should use MP_HANDLE_OOM() for this.) This is done out of laziness and
for convenience, and due to the fact that MPlayer/mplayer2 never handled it
correctly. (MPlayer varied between handling it correctly, trying to do so but
failing, and just not caring, while mplayer2 started using abort() for it.)

This is justifiable in a number of ways. Error handling paths are notoriously
untested and buggy, so merely having them won't make your program more reliable.
Having these error handling paths also complicates non-error code, due to the
need to roll back state at any point after a memory allocation.

Take any larger body of code, that is supposed to handle OOM, and test whether
the error paths actually work, for example by overriding malloc with a version
that randomly fails. You will find bugs quickly, and often they will be very
annoying to fix (if you can even reproduce them).

In addition, a clear indication that something went wrong may be missing. On
error your program may exhibit "degraded" behavior by design. Consider a video
encoder dropping frames somewhere in the middle of a video due to temporary
allocation failures, instead of just exiting with an errors. In other cases, it
may open conceptual security holes. Failing fast may be better.

mpv uses GPU APIs, which may be break on allocation errors (because driver
authors will have the same issues as described here), or don't even have a real
concept for dealing with OOM (OpenGL).

libmpv is often used by GUIs, which I predict always break if OOM happens.

Last but not least, OSes like Linux use "overcommit", which basically means that
your program may crash any time OOM happens, even if it doesn't use malloc() at
all!

But still, don't just assume malloc() always succeeds. Use MP_HANDLE_OOM(). The
ta* APIs do this for you. The reason for this is that dereferencing a NULL
pointer can have security relevant consequences if large offsets are involved.
Also, a clear error message is better than a random segfault.

Some big memory allocations are checked anyway. For example, all code must
assume that allocating video frames or packets can fail. (The above example
of dropping video frames during encoding is entirely possible in mpv.)

Undefined behavior
------------------

Undefined behavior (UB) is a concept in the C language. C is famous for being a
language that makes it almost impossible to write working code, because
undefined behavior is so easily triggered, compilers will happily abuse it to
generate "faster" code, debugging tools will shout at you, and sometimes it
even means your code doesn't work.

There is a lot of literature on this topic. Read it.

(In C's defense, UB exists in other languages too, but since they're not used
for low level infrastructure, and/or these languages are at times not rigorously
defined, simply nobody cares. However, the C standard committee is still guilty
for not addressing this. I'll admit that I can't even tell from the standard's
gibberish whether some specific behavior is UB or not. It's written like tax
law.)

In mpv, we generally try to avoid undefined behavior. For one, we want portable
and reliable operation. But more importantly, we want clean output from
debugging tools, in order to find real bugs more quickly and effectively.

Avoid the "works in practice" argument. Once debugging tools come into play, or
simply when "in practice" stops being true, this will all get back to you in a
bad way.

Global state, library safety
----------------------------

Mutable global state is when code uses global variables that are not read-only.
This must be avoided in mpv. Always use context structs that the caller of
your code needs to allocate, and whose pointers are passed to your functions.

Library safety means that your code (or library) can be used by a library
without causing conflicts with other library users in the same process. To any
piece of code, a "safe" library's API can simply be used, without having to
worry about other API users that may be around somewhere.

Libraries are often not library safe, because they they use global mutable state
or other "global" resources. Typical examples include use of signals, simple
global variables (like hsearch() in libc), or internal caches not protected by
locks.

A surprisingly high number of libraries are not library safe because they need
global initialization. Typically they provide an API function, which
"initializes" the library, and which must be called before calling any other
API functions. Often, you are to provide global configuration parameters, which
can change the behavior of the library. If two libraries A and B use library C,
but A and B initialize C with different parameters, something "bad" may happen.
In addition, these global initialization functions are often not thread-safe. So
if A and B try to initialize C at the same time (from different threads and
without knowing about each other), it may cause undefined behavior. (libcurl is
a good example of both of these issues. FFmpeg and some TLS libraries used to be
affected, but improved.)

This is so bad because library A and B from the previous example most likely
have no way to cooperate, because they're from different authors and have no
business knowing each others. They'd need a library D, which wraps library C
in a safe way. Unfortunately, typically something worse happens: libraries get
"infected" by the unsafeness of its sub-libraries, and export a global init API
just to initialize the sub-libraries. In the previous example, libraries A and B
would export global init APIs just to init library C, even though the rest of
A/B are clean and library safe. (Again, libcurl is an example of this, if you
subtract other historic anti-features.)

The main problem with library safety is that its lack propagates to all
libraries using the library.

We require libmpv to be library safe. This is not really possible, because some
libraries are not library safe (FFmpeg, Xlib, partially ALSA). However, for
ideological reasons, there is no global init API, and best effort is made to try
to avoid problems.

libmpv has some features that are not library safe, but which are disabled by
default (such as terminal usage aka stdout, or JSON IPC blocking SIGPIPE for
internal convenience).

A notable, very disgustingly library unsafe behavior of libmpv is calling
abort() on some memory allocation failure. See error checking section.

Logging
-------

All logging and terminal output in mpv goes through the functions and macros
provided in common/msg.h. This is in part for library safety, and in part to
make sure users can silence all output, or to redirect the output elsewhere,
like a log file or the internal console.lua script.

Locking
-------

See generally available literature. In mpv, we use pthread for this.

Always keep locking clean. Don't skip locking just because it will work "in
practice". (See undefined behavior section.) If your use case is simple, you may
use C11 atomics (osdep/atomic.h for partial C99 support), but most likely you
will only hurt yourself and others.

Always make clear which fields in a struct are protected by which lock. If a
field is immutable, or simply not thread-safe (e.g. state for a single worker
thread), document it as well.

Internal mpv APIs are assumed to be not thread-safe by default. If they have
special guarantees (such as being usable by more than one thread at a time),
these should be explicitly documented.

All internal mpv APIs must be free of global state. Even if a component is not
thread-safe, multiple threads can use _different_ instances of it without any
locking.

On a side note, recursive locks may seem convenient at first, but introduce
additional problems with condition variables and locking hierarchies. They
should be avoided.

Locking hierarchy
-----------------

A simple way to avoid deadlocks with classic locking is to define a locking
hierarchy or lock order. If all threads acquire locks in the same order, no
deadlocks will happen.

For example, a "leaf" lock is a lock that is below all other locks in the
hierarchy. You can acquire it any time, as long as you don't acquire other
locks while holding it.

Unfortunately, C has no way to declare or check the lock order, so you should at
least document it.

In addition, try to avoid exposing locks to the outside. Making the declaration
of a lock private to a specific .c file (and _not_ exporting accessors or
lock/unlock functions that manipulate the lock) is a good idea. Your component's
API may acquire internal locks, but should release them when returning. Keeping
the entire locking in a single file makes it easy to check it.

Avoiding callback hell
----------------------

mpv code is separated in components, like the "frontend" (i.e. MPContext mpctx),
VOs, AOs, demuxers, and more. The frontend usually calls "down" the usage
hierarchy: mpctx almost on top, then things like vo/ao, and utility code on the
very bottom.

"Callback hell" is when when components call both up and down the hierarchy,
which for example leads to accidentally recursion, reentrancy problems, or
locking nightmares. This is avoided by (mostly) calling only down the hierarchy.
Basically the call graph forms a DAG. The other direction is handled by event
queues, wakeup callbacks, and similar mechanisms.

Typically, a component provides an API, and does not know anything about its
user. The API user (component higher in the hierarchy) polls the state of the
lower component when needed.

This also enforces some level of modularization, and with some luck the locking
hierarchy. (Basically, locks of lower components automatically become leaf
locks.) Another positive effect is simpler memory management.

(Also see e.g.: http://250bpm.com/blog:24)

Wakeup callbacks
----------------

This is a common concept in mpv. Even the public API uses it. It's used when an
API has internal threads (or otherwise triggers asynchronous events), but the
component call hierarchy needs to be kept. The wakeup callback is the only
exception to the call hierarchy, and always calls up.

For example, vo spawns a thread that the API user (the mpv frontend) does not
need to know about. vo simply provides a single-threaded API (or that looks like
one). This API needs a way to notify the API user of new events. But the vo
event producer is on the vo thread - it can't simply invoke a callback back into
the API user, because then the API user has to deal with locking, despite not
using threads. In addition, this will probably cause problems like mentioned in
the "callback hell" section, especially lock order issues.

The solution is the wakeup callback. It merely unblocks the API user from
waiting, and the API user then uses the normal vo API to examine whether or
which state changed. As a concept, it documents what a wakeup callback is
allowed to do and what not, to avoid the aforementioned problems.

Generally, you are not allowed to call any API from the wakeup callback. You
just do whatever is needed to unblock your thread. For example, if it's waiting
on a mutex/condition variable, acquire the mutex, set a change flag, signal
the condition variable, unlock, return. (This mutex must not be held when
calling the API. It must be a leaf lock.)

Restricting the wakeup callback like this sidesteps any reentrancy issues and
other complexities. The API implementation can simply hold internal (and
non-recursive) locks while invoking the wakeup callback.

The API user still needs to deal with locking (probably), but there's only the
need to implement a single "receiver", that can handle the entire API of the
used component. (Or multiple APIs - MPContext for example has only 1 wakeup
callback that handles all AOs, VOs, input, demuxers, and more. It simple re-runs
the playloop.)

You could get something more advanced by turning this into a message queue. The
API would append a message to the queue, and the API user can read it. But then
you still need a way to "wakeup" the API user (unless you force the API user
to block on your API, which will make things inconvenient for the API user). You
also need to worry about what happens if the message queue overruns (you either
lose messages or have unbounded memory usage). In the mpv public API, the
distinction between message queue and wakeup callback is sort of blurry, because
it does provide a message queue, but an additional wakeup callback, so API
users are not required to call mpv_wait_event() with a high timeout.

mpv itself prefers using wakeup callbacks over a generic event queue, because
most times an event queue is not needed (or complicates things), and it is
better to do it manually.

(You could still abstract the API user side of wakeup callback handling, and
avoid reimplementing it all the time. Although mp_dispatch_queue already
provides mechanisms for this.)

Condition variables
-------------------

They're used whenever a thread needs to wait for something, without nonsense
like sleep calls or busy waiting. mpv uses the standard pthread API for this.
There's a lot of literature on it. Read it.

For initial understanding, it may be helpful to know that condition variables
are not variables that signal a condition. pthread_cond_t does not have any
state per-se. Maybe pthread_cond_t would better be named pthread_interrupt_t,
because its sole purpose is to interrupt a thread waiting via pthread_cond_wait()
(or similar). The "something" in "waiting for something" can be called
predicate (to avoid confusing it with "condition"). Consult literature for the
proper terms.

The very short version is...

Shared declarations:

    pthread_mutex_t lock;
    pthread_cond_t cond_var;
    struct something state_var; // protected by lock, changes signaled by cond_var

Waiter thread:

    pthread_mutex_lock(&lock);

    // Wait for a change in state_var. We want to wait until predicate_fulfilled()
    // returns true.
    // Must be a loop for 2 reasons:
    //  1. cond_var may be associated with other conditions too
    //  2. pthread_cond_wait() can have sporadic wakeups
    while (!predicate_fulfilled(&state_var)) {
        // This unlocks, waits for cond_var to be signaled, and then locks again.
        // The _whole_ point of cond_var is that unlocking and waiting for the
        // signal happens atomically.
        pthread_cond_wait(&cond_var, &lock);
    }

    // Here you may react to the state change. The state cannot change
    // asynchronously as long as you still hold the lock (and didn't release
    // and reacquire it).
    // ...

    pthread_mutex_unlock(&lock);

Signaler thread:

    pthread_mutex_lock(&lock);

    // Something changed. Update the shared variable with the new state.
    update_state(&state_var);

    // Notify that something changed. This will wake up the waiter thread if
    // it's blocked in pthread_cond_wait(). If not, nothing happens.
    pthread_cond_broadcast(&cond_var);

    // Fun fact: good implementations wake up the waiter only when the lock is
    // released, to reduce kernel scheduling overhead.
    pthread_mutex_unlock(&lock);

Some basic rules:
    1. Always access your state under proper locking
    2. Always check your predicate before every call to pthread_cond_wait()
       (And don't call pthread_cond_wait() if the predicate is fulfilled.)
    3. Always call pthread_cond_wait() in a loop
       (And only if your predicate failed without releasing the lock..)
    4. Always call pthread_cond_broadcast()/_signal() inside of its associated
       lock

mpv sometimes violates rule 3, and leaves "retrying" (i.e. looping) to the
caller.

Common pitfalls:
    - Thinking that pthread_cond_t is some kind of semaphore, or holds any
       application state or the user predicate (it _only_ wakes up threads
       that are at the same time blocking on pthread_cond_wait() and friends,
       nothing else)
    - Changing the predicate, but not updating all pthread_cond_broadcast()/
      _signal() calls correctly
    - Forgetting that pthread_cond_wait() unlocks the lock (other threads can
      and must acquire the lock)
    - Holding multiple nested locks while trying to wait (=> deadlock, violates
      the lock order anyway)
    - Waiting for a predicate correctly, but unlocking/relocking before acting
      on it (unlocking allows arbitrary state changes)
    - Confusing which lock/condition var. is used to manage a bit of state

Generally available literature probably has better examples and explanations.

Using condition variables the proper way is generally preferred over using more
messy variants of them. (Just saying because on win32, "SetEvent" exists, and
it's inferior to condition variables. Try to avoid the win32 primitives, even if
you're dealing with Windows-only code.)

Threads
-------

Threading should be conservatively used. Normally, mpv code pretends to be
single-threaded, and provides thread-unsafe APIs. Threads are used coarsely,
and if you can avoid messing with threads, you should. For example, VOs and AOs
do not need to deal with threads normally, even though they run on separate
threads. The glue code "isolates" them from any threading issues.