When there are chunks which are not yet finished,
audio_output_all_check() returned the size of its music pipe minus
one. I can't remember exactly why I subtracted 1 from the return
value, it must have had something to do with a former meaning of this
function. Now it induces assertion failures.
After adding the container_scan() method the update_regular_file() method was quite hard to read.
Now there's update_container_file() which deals with container files.
That way normal container files (i.e. without embedded tracks) are handled by the old code like a regular file.
This will fix some of the odd behaviour observed.
If the PCM handle gets disconnected, don't close and clear it in
alsa_recover(). The MPD core will call alsa_close() anyway. This
way, we can always assume that alsa_data.pcm is always valid.
After a seek, wait until enough new chunks are decoded before starting
playback. If this takes too long, send silence chunks to the audio
outputs meanwhile.
When the audio outputs are closed, also clear the audio format. If we
don't do this, every call to audio_output_all_update() will open the
device, even if it's meant to be paused.
When playback is unpaused, pass the audio_format to
audio_output_all_open(). Don't assume that output_all.c remembers the
previous audio format. Also check if there has been an audio format
yet.
Check audio_output.command after each sub-chunk has been played. It
discards the rest of the chunk, but since all commands make the device
stop anyway, this is not a problem, but part of the improvement. This
improves the latency of audio output commands.
A larger chunk size means less overhead for managing them. 4 kB seems
to be a reasonable choice: it contains 23 ms of 44.1 kHz 16 bit stereo
data, or 3 ms of 192 kHz 24 bit stereo data. The original value of
1020 seemed to be too small, there were quite a lot of system calls
and context switches.
Instead of passing individual buffers to audio_output_all_play(), pass
music_chunk objects. Append all those chunks asynchronously to a
music_pipe instance. All output threads may then read chunks from
this pipe. This reduces MPD's internal latency by an order of
magnitude.
When a PAUSE command is received while the decoder starts, don't open
the audio device when the decoder becomes ready. It's pointless,
because MPD will close if after that.
If the header valgrind/memcheck.h is available, add
VALGRIND_MAKE_MEM_NOACCESS() and VALGRIND_MAKE_MEM_UNDEFINED()
support, which enables nice warnings in the valgrind memory checker.
This is similar to the MPD 0.14 patch "wait 10 seconds before
reopening a failed device", which only covered open() failures. This
patch adds the same feature for play().
Until now every flac file got removed unconditionally (and then re-added)
whenever the update command was issued. Now there is a check if we need
to that, so the file will only be removed if there is a embedded cuesheet
in that file
So far only seekpoints are supported, so no proper tagging yet
except for track number and track length.
Tagging should be done by parsing the cue sheet which
is often embedded as vorbis comment in flac files.
Furthermore the pathname should be configurable like "%A - %t - %T",
where %A means Artist, %t track number and %T Title or so.
In !NDEBUG, remember which audio_format is stored in every chunk and
every pipe. Check the audio_format of every new data block appended
to the music_chunk, and the format of every new chunk appended to the
music_pipe.
This patch fixes a theoretical (but practically impossible) flaw: the
variable "buffer_time" may be uninitialized when it is used.
Initialize the variable with snd_pcm_hw_params_get_buffer_time().
The default values for buffer_time and period_time were both capped by
the hardware limits on practically all chips. The result was a
period_time which was half as big as the buffer_time. On some chips,
this led to lots of underruns when using a high sample rate (192 kHz),
because MPD had very little time to send new samples to ALSA.
A period time which is one fourth of the buffer time turned out to be
much better. If no period_time is configured, see how much
buffer_time the hardware accepts, and try to configure one fourth of
it as period_time, instead of hard-coding the default period_time
value.
This is yet another attempt to provide a solution which is valid for
all sound chips. Using the SND_PCM_NONBLOCK flag also seemed to solve
the underruns, but put a lot more CPU load to MPD.
Sometimes, audio_output_update() isn't called for the second device
when the first one has succeeded. The patch
"audio_output_all_update() returns bool" broke it, because the boolean
evaluation ended after the first "true".
When the decoder chunk is empty in decoder_flush_chunk(), don't push
it into the music pipe - return it to the music buffer instead. An
empty chunk in the pipe wastes resources for no advantage.
The value of music_chunk.next is undefined for a chunk returned by
music_pipe_shift(). For more pedantic debugging, poison the reference
before returning the chunk.
This patch follows the commit 21bb10f4b.
>From Max Kellermann:
> I removed the daemonization changes in main.c. Please explain why you
> changed that. If you need it for some reason, make that a separate
> patch with a good description of your rationale.
> That's the biggest flaw of your code: it opens the mixer device in the
> init() method, while the open() method is empty. When the pulse
> daemon is not available (either during MPD startup or when it dies
> while MPD runs), the plugin will not even attempt to reconnect to
> pulse. Please move the code to the open() method, to make that work.
I changed the daemonize call as the fork losts the connection to the
pulse server. According to your remark, the init() method should be
moved to the open() ones.
With the modification, mpd is able to reconnect the pulse mixer after
restarting the pulseaudio daemon.
Signed-off-by: David Guibert <david.guibert@gmail.com>
Signed-off-by: Max Kellermann <max@duempel.org>
This patch introduces the mixer for the pulse output.
Technically speaking, the pulse index is needed to get or set
the volume. You must define callback fonctions to get this index since
the pulse output in mpd is done using the simpe api. The pulse simple api
does not provide the index of the newly defined output.
So callback fonctions are associated to the pulse context.
The list of all the sink input is then retreived.
Then we select the name of the mpd pulse output and control
its volume by its associated index number.
Signed-off-by: Patrice Linel <patnathanael@gmail.com>
Signed-off-by: David Guibert <david.guibert@gmail.com>
[mk: fixed whitespace errors and broke long lines; removed
daemonization changes from main.c]
Turn the music_pipe into a simple music_chunk queue. The music_chunk
allocation code is moved to music_buffer, and is now managed with a
linked list instead of a ring buffer. Two separate music_pipe objects
are used by the decoder for the "current" and the "next" song, which
greatly simplifies the cross-fading code.
Added music_pipe_allocate(), music_pipe_push() and
music_pipe_cancel(). Those functions allow the caller (decoder thread
in this case) to do its own chunk management. The functions
music_pipe_flush() and music_pipe_tag() can now be removed.
After the decoder command was obtained, don't wait until libflac
detects EOF (as a side effect), quit the decoder immediately. This
check was missing completely.
When the MPD core sends the decoder a command while
flac_process_single() is executed, this function fails. Abort the
decoder only if not seeking. This fixes a seeking bug.
Log the real period and buffer size. This might be useful when
debugging xruns. Note that the same information is available in
/proc/asound/card*/pcm*p/sub*/hw_params
The lastfm input plugin enables MPD to play lastfm:// URLs. This
plugin is not complete yet: it plays only the first song in the
last.fm playlist, and the playlist parser isn't even implemented
properly.
Some sound chips/drivers (e.g. Intel HDA) don't support 24 bit
samples, they want to get 32 bit instead. Now that MPD's PCM library
supports 32 bit, add a 32 bit fallback when 24 bit is not supported.
There is nothing 24 bit specific in the pcm_dither_24 struct. Since
we want to reuse the struct for 32 bit dithering, let's drop the "_24"
suffix from the struct name.
Some 24 bit code can be reused. The 32 bit variant has to use 64 bit
integers, because 32 bit integers could overflow. This may be a
performance hit on 32 bit CPUs.
This is the first patch in a series to enable 32 bit audio samples in
MPD. 32 bit samples are more tricky than 24 bit samples, because the
integer may overflow when you operate on a sample.
audio_valid_sample_format() verifies the number of channels. Let's
just say up to 8 channels is allowed (which is possible with some
consumer sound chips). I don't know if there are bigger cards, and
since I cannot test it, I'll limit it to 8 for now.
On some hardware, reading the mixer value from hardware is an
expensive operation, and MPD has to do it for every client. Throttle
access to the hardware, cache the result for one second.
time() is not a monotonic timer, and MPD might get confused by clock
skews. clock_gettime() provides a monotonic clock, but is not
portable to non-POSIX systems (i.e. Windows). This patch uses GLib's
GTimer API, which aims to be portable.
If an input_stream is not seekable, libaudiofile fails to play at all:
Audio File Library: unrecognized audio file format [error 0]
Since we know in advance whether the input_stream is seekable, just
refuse to play on a non-seekable stream.
The generic sockaddr struct is too small for some addresses. For
accept(), we have to allocate a sockaddr_storage struct on the stack,
which is large enough for all addresses.
Create the socket_util.c library, the first function is
sockaddr_to_string(): it converts a sockaddr struct to a string
containing the IP address in a human-readable form.
When checking whether database entries have been deleted, don't check
if an archive file is a directory (G_FILE_TEST_IS_DIR), use
G_FILE_TEST_IS_REGULAR for this case instead. To determine if a
"struct directory" is an archive, check for device==DEVICE_INARCHIVE.
This is always false after loading the database, so this patch is not
complete yet.
Remember the modification time of each directory. This is important
for archives (which are virtual directories right now), but may also
be useful for an automatic update mechanism.
Added the uri_remove_auth() library function which strips username
and password from a HTTP URI, and use it in song_print_url(). This
allows you to add HTTP URIs to the playlist including secret username
and password, without disclosing it to all MPD clients.
Since we introduced a GLib logging domain, the "client" string appears
twice in the log lines:
client: client 0: command returned 0
Removed the second one, now it looks like this:
client: [0] command returned 0
Still not quite good, but better than before.
MPD used to be silent when it could stat() a directory, but could not
opendir() it to read its contents. This caused a lot of support
headache with users who have wrong file permissions. Add another
warning message.
There's no point in declaring num_items as a uint8_t, it doesn't save
any space, due to padding. This allows us to lift the articial "255
items" limitation.
The warning message "problems opening audio device while playing ..."
does not help at all, and should be removed. At this point, the real
error message has already been logged by the output thread.
Use GLib's GError library for reporting output device failures.
Note that some init() methods don't clean up properly after a failure,
but that's ok for now, because the MPD core will abort anyway.
Don't call AudioOutputUnitStart() in the play() method, do it after
the device has been opened. We can eliminate the "started" property
now, because the device is always started when it's open.
ao_play() gets PCM data in the in_audio_format, and converts it to
out_audio_format. Comparing the input data with out_audio_format is
wrong.
prefixed with "STG:" will be automatically removed. STG: Trailing
empty lines will be automatically removed. STG: vi: set textwidth=75
filetype=diff nobackup:
The MPD core guarantees that the audio_output object is always
consistent, and our pa_simple!=NULL checks are superfluous. Also
don't manually close the device on error in pulse_play(), since the
MPD core does this automatically when the play() method returns 0.
The MPD core guarantees that the audio_output object is always in a
consistent state: either open or closed. When open, it will not call
the open() method again, and when closed, it will not call play().
Removed several checks and the NULL initialization.
audio_output_get_name() has been removed, which was the only function
left in output_api.h. The output plugin doesn't need the audio_output
object at all, remove the parameter from the init() method.
After much research[1][2][3] this should be the majority of currently
supported file extensions and mime-types for the currently supported
ffmpeg formats. This list maybe incomplete, but it's more complete
than anything else out there that I've been able to find. This list
needs to be updated every now and again as the ffmpeg sources support
more formats.
1. Sources
2. wiki.multimedia.cx
3. filext.com
Recursive Makefiles are inefficient and error prone (no proper way to
declare dependencies). Since there's no disadvantage in having one
single Makefile, let's do it.
The old API required an output plugin to not return until all data
passed to the play() method is consumed. Some output plugins have to
loop to fulfill that requirement, and may block during that. Simplify
these, by letting them consume only part of the buffer: make play()
return the length of the consumed data.
Now that I've found this nice function in the GLib docs, we can
finally remove our custom sleep function. Still all those callers of
g_usleep() have to be migrated one day to use events, instead of
regular polling.
Hi,
upon trying to play an MMS stream added to the play list, I got this:
mpd: /tmp/mpd/./src/input_stream.c:85: input_stream_open: Assertion `is->plugin->open == ((void *)0) || is->plugin == plugin' failed.
With the following patch applied, it works perfectly.
Thanks for having implemented MMS support :-).
Best regards,
Peter
Added an inline assembly function for the 64 bit multiplication.
Benchmark results on a Pentium II 266 MHz, 512 MB of 24 bit PCM data:
dd if=/dev/zero bs=64k count=8k |
time ./test/software_volume 48000:24:2 >/dev/null
Before this patch 22.94s, after this patch 7.24s.
Use faacDecInit2() instead of AudioSpecificConfig() to detect the AAC
track in the MP4 file. This has a great advantage: it initializes the
libfaad decoder, which the caller would normally do anyway - but now
we can go without the AudioSpecificConfig() call. When decoder==NULL
(called from mp4_tag_dup()), fall back to a mp4ff_get_track_type()==1
check, like other audio players do.
Moved the libfaad decoder initialization to mp4_faad_new(), and also
fill the audio_format struct there. This eliminates a little bit of
complexity in mp4_decode().
When a file is not seekable, MPD dropped the audio buffers before even
attempting to seek. This caused noticable sound corruption. Fix:
first attempt to seek, and only if that succeeds, call
audio_output_all_cancel().
All callers of adts_find_frame() use faad_buffer_fill() before that.
Move that faad_buffer_fill() call into adts_find_frame() instead.
adts_find_frame() will get its own logic for on-demand filling.
When I implemented the pcm_buffer library, I forgot to set the new
buffer size. This caused a new allocation in each pcm_buffer_get(),
fortunately no memory was leaked.
The decoder_plugin struct is used by both the MPD core and the decoder
plugin implementations. Move it to a shared header file, to minimize
header dependencies.
If mpd.conf specifies a user, and MPD is invoked by exactly this user,
ignore the "user" setting. Don't bother to look up its groups and
don't attempt to change uid, it won't work anyway.
Use delete_directory() for removing sub directories instead of
dirvec_clear(). This ensures that all memory occupied by
subdirectories of deleted directories is freed.
When a directory is deleted, MPD deleted only the directory from the
database; it did not bother to walk the full tree to free all memory
and to remove deleted songs from the playlist. Replace a
dirvec_delete() with delete_directory().
When you change the filesystem charset, discard the old database file
and create a new one. The old database file will most likely contain
stale or invalid information.
There are a few problems left in this plugin:
- fluidsynth decodes in real time, while MPD prefers to buffer as
quickly as possible; as a workaround, this plugin uses a timer
object to synchronize with real-time playback
- I don't know yet how fluidsynth tells me when the song has ended
- the "soundfont" configuration setting is not yet documented, and it
will likely change soon (in favor of a per-decoder configuration
block)
When MPD is not playing, it may still remember which is the "current"
song. When you switch to "random" mode, MPD will always start playing
exactly this song. This defies the goal of "random" mode a little.
Clear the "current" song when MPD is not playing during the "random"
mode switch.
The output_command library provides a command interface to the audio
outputs. It assumes the input comes from an untrusted source
(i.e. the client) and verifies all parameters.
In addition to audio_format_valid(), provide functions which validate
only one attribute of an audio_format. These functions are reused by
audio_format_parse().
Added audio_format_parse() in a separate library, with a modern
interface: return a GError instead of logging errors. This allows the
caller to deal with the error.
When MPD explicitly starts playing, ignore the "REOPEN_AFTER" timeout.
This timeout was useful when MPD attempted to reopen a failed device
over and over, but it confuses users when they explicitly tell MPD to
start playing, while MPD insists to wait for the 10 seconds to pass.
Fix a memory leak: it was not guaranteed that pcm_convert_deinit() was
called for each pcm_convert_init(). This patch always (de)initializes
the pcm_convert library when the audio_output.open flag is flipped.
Pass the music chunk as a "const void *" to the encoder, instead of a
"const char *". Actually, both encoders currently expect 16 bit
samples, passing a 8-bit character is rather pointless.
The crossfading code shouldn't depend on the audio output code. Pass
the current audio format to cross_fade_calc() and let it compare
directly, instead of using isCurrentAudioFormat().
When MPD is stopped, but the last song is still the "current song",
and you delete it, playlist->current is not updated, and becomes an
invalid value. Fix this by catching "!playlist->playing &&
playlist->current == (int)songOrder".
audio_output_config_count() returns the number of audio outputs in the
configuration file. It is only used by initAudioDriver(). The public
function audio_output_count() now returns audioOutputArraySize.
When we reset pc.next_song if there is no song queued, this might
cause a race condition: the next song to be played is cleared, while
pc.command was already set. Clear the "next_song" only if there is a
song queued for the current do_play() invocation.
If a new song is queued before calling playerSeek(), then the player
and the playlist enter an inconsistent state, because the player
discards the playlist's "queued" song in favor of the seeked song.
Call playlist_update_queued_song() after playerSeek().
After a player command (successful or not), reset pc.next_song,
because the queue is supposed to be empty then. Otherwise,
playlist.queued and pc.next_song may disagree, which triggers an
assertion failure.
Commit f78cddb4 introduced a regression: after a song was moved, the
order array was not updated (in random mode). This caused MPD to
think the "current" song has changed when you moved something to the
position of the current song.
Always assume the buffer is empty before calling the encoder. Always
flush the buffer immediately after there has been added something.
This reduces the risk of buffer overruns, because there will never be
a "rest" in the current buffer.
The non-blocking mode of libshout is sparsely documented, and MPD's
implementation had several bugs. Also removed connect throttling
code, that is done by the MPD core since 0.14.
After the state file has been loaded, the playlist version is still
"1", and "plchanges 1" returns the whole playlist. Fix this by
increasing the playlist version after the state file has been loaded.
Don't call syncPlaylistWithQueue() in nextSongInPlaylist() and
previousSongInPlaylist(). This is a relic from the time when there
was no event, and was a workaround to the timing problem.
Export the "g_playlist" variable, and pass it to all playlist
functions. This way, we can split playlist.c easier into separate
parts. The code which initializes the singleton variable is moved to
playlist_global.c.
Before every operation which modifies the playlist, remember a pointer
to the song struct. After the modification, determine the "next song"
again, and if it differs, dequeue and queue the new song.
This removes a lot of complexity from the playlist update code, and
makes it more robust.
The "current" variable is used for calculating the seek destination,
and was declared as "int". With very long song files, the 32 bit
integer can overflow. ffmpeg expects an int64_t, which is very
unlikely to overflow. Switch to int64_t.
If avcodec_decode_audio2() returns no output for an AVPacket,
libavcodec may buffer some data, and return a larger chunk of output
later. This patch disables a lot of bogus warnings.
Output the name of the codec as a debug message. During my tests,
ffmpeg never filled this struct member, but it may do so in the past,
and this debug message might become helpful.
The shout_mp3 encoder had two bugs: when no song was ever played, MPD
segfaulted during cleanup. Second bug: memory leak, each time the
shout device was opened, lame_init() was called again, and
lame_close() is only called once during shutdown.
Fix this by shutting down LAME each time the clear_encoder() method is
called.
When the update thread is started before MPD has forked (for
daemonization), it is killed, because threads do not survive a fork().
This induces an inconsistent state where MPD won't start any update
thread at all, because it thinks the thread is already running.
Move the "while" loop which checks for commands to the caller
ao_pause(). This simplifies the pause() method, and lets us remove
audio_output_is_pending().
If no ports are configured, don't overwrite the (NULL) configuration
with the port names of the first JACK server. If the server changes
after a JACK reconnect, MPD won't attempt to auto-detect again.
Currently, the JACK plugin manipulates the audio_format struct which
was passed to the open() method. This is very likely to break,
because the plugin must not permanently store this pointer. After
this patch, MPD ignores sample rate changes. It looks like other
software is doing the same, and I guess this is a non-issue.
This patch converts the audio_format pointer within jack_data into a
static audio_format struct.
Hi -
independently of libmikmod's other problems - there seems
to be a problem in mpd's wrapper: MikMod_Exit() is called
after the first file is decoded, which frees some ressources
within the mikmod library. An attempt to play a second file
leads to a crash. The appended patch fixes this for me.
(I don't know what the "dup" entry is good for - someone
who knows should review that too.)
best regards
Matthias
[mk: removed 3 more MikMod_Exit() invocations]
When there are duplicate slashes in the song paths, eliminate them;
example:
/var/lib/mpd/music//foo.mp3
becomes:
/var/lib/mpd/music/foo.mp3
The slash is only detected at the border between the music_directory
and the local part.
When the user configures a music_directory with a trailing slash, it
may break playlist loading, because MPD expects a double slash. Chop
off the trailing slash.
ffmpeg_tag_internal() does not look for a few tags that mpd
supports. Most noteably:
comment -> TAG_ITEM_COMMENT -> Description
genre -> TAG_ITEM_GENRE -> WM/Genre (not WM/GenreID)
year -> TAG_ITEM_DATE -> WM/Year
I *think* that this is the last of the tags that AVFormatContext() in
ffmpeg supports that mpd also uses.
Make those two methods optional to implement, and let input_stream.c
provide fallbacks. The buffer() method will be removed one day, and
there is now only one implementation left (input_curl.c).
The open_stream() method opens the input_stream. This allows the
archive plugin to do its own initialization, and it also allows it to
use input_stream.data. We can remove input_stream.archive now, which
was unnatural to have in the first place.
Preparation for supporting other channel numbers than stereo: use
loops instead of duplicating code for the second channel. Most
likely, gcc will unroll these loops, so the binary won't be any
different.
This patch implements the MMS protocol, by using libmms. It is quite
experimental: it does not support seeking yet, and it is currently
using synchronous I/O, which causes MPD to hang while waiting for the
server.
When the playlist is cleared, pc.errored_song is also cleared. This
causes pc_errored_song_uri() to crash, because it assumes that
pc.errored_song is set. Reset pc.error to fix that assumption.
When waiting for free space in the ring buffer, the JACK plugin
sleeped 10ms until there is enough space. This delay was too large
for low-latency setups (<10ms), and created a lot of xruns. Work
around that by reducing the sleep time to 1ms.
A proper solution for this would be to use an event based approach,
and we will do it, just not now.
When the connection failed once, you had to restart MPD, because it
never cleared the jack_data.shutdown flag. Instead, it refused to
play anything "because there is no client thread" (which is wrong at
that point).
GIOChannel is more portable than raw read()/write() calls. We're
using GIOChannel anyway, because we need it for plugging the client
into the GLib main loop.
Configure the GIOChannel to the bare minimum: no character set, no
buffering.
On some platforms, g_free() must be used for memory allocated by
GLib. This patch intends to correct a lot of occurrences, but is
probably not complete.
Both methods are always called together. There is no point in having
them separate. This simplifies the code, because the old configure()
method could be called more than once, and had to free old
allocations.
Reimplemented the legacy mixer configuration: copy the deprecated
configuration values into the audio_output section. Don't configure
the mixers twice (once for the audio_output, and a second time for the
legacy values).
This requires volume_init() to be called before initAudioDriver().
Return the default value in the conf_get_block_*() functions when
param==NULL was passed.
This simplifies a lot of code, because all initialization can be done
in one code path, regardless whether configuration is present.
Two bugs here led to a large number of interrupts being generated on the
sound card when ALSA output is being used. Because we specify no default
period_time, the sound card gives us 3000 interrupts/sec rather than a more
sane 20 or 30. This completes the revert of dd7711 already started by
4ca24f.
The larger bug was in the change to config_get_block_unsigned() and using 0
as the default value for both 'buffer_time' and 'period_time'. This means
any pre-setting of these options in newAlsaData() gets wiped out. Add a new
default for period_time, and ensure default values for buffer_time and
period_time are used if none are provided by the user.
Signed-off-by: Dan McGee <dan@archlinux.org>
[mk: set defaults in newAlsaData() to fix auto-configuration; renamed
"_MS" back to "_US" because ALSA expects microseconds, not milliseconds]
Signed-off-by: Max Kellermann <max@duempel.org>
Added all important id tags from the MusicBrainz wiki:
http://musicbrainz.org/doc/MusicBrainzTag
This should automatically enable its suport in the vorbis and flac
decoder plugins.
The input_stream API sets size to -1 when the size of the resource is
not known. The modplug decoder checked for size==0, which would be an
empty file.
You are allowed to call decoder_read() with decoder==NULL. It is a
convenience function provided by the decoder API. Don't manually fall
back to input_stream_read().
