client_write() writes a buffer to the client and buffers it if
required. client_puts() does the same for a C string. The next patch
will add more tools which will replace fdprintf() later.
clearMpdTag could be called on a tag that was still in a
tag_begin_add transaction before tag_end_add is called. This
was causing free() to attempt to operate on bulk.items; which is
un-free()-able. Now instead we unmark the bulk.busy to avoid
committing the tags to the heap only to be immediately freed.
Additionally, we need to remember to call tag_end_add() when
a song is updated before we NULL song->tag to avoid tripping
an assertion the next time tag_begin_add() is called.
Since client->fd==-1 has become our "expired" flag, it may already be
-1 when client_close() is called. Don't assert that it is still
non-negative, and call client_set_expired() instead.
During the tag library refactoring, the shout plugin was disabled, and
I forgot about adapting it to the new API. Apply the same fixes to
the oggflac decoder plugin.
While parsing the tag cache, don't allocate the directory name from
the heap, but copy it into a buffer on the stack. This reduces heap
fragmentation by 1%.
If many tag_items are added at once while the tag cache is being
loaded, manage these items in a static fixed list, instead of
reallocating the list with every newly created item. This reduces
heap fragmentation.
Massif results again:
mk before: total 12,837,632; useful 10,626,383; extra 2,211,249
mk now: total 12,736,720; useful 10,626,383; extra 2,110,337
The "useful" value is the same since this patch only changes the way
we allocate the same amount of memory, but heap fragmentation was
reduced by 5%.
Try to detect if the string needs Latin1-UTF8 conversion, or
whitespace cleanup. If not, we don't need to allocate temporary
memory, leading to decreased heap fragmentation.
At several places, we create temporary copies of non-null-terminated
strings, just to use them in functions like validUtf8String(). We can
save this temporary allocation and avoid heap fragmentation if we
add a length parameter instead of expecting a null-terminated string.
Since the inline function cannot modify its caller's variables (which
is a good thing for code readability), the new string pointer is the
return value. The resulting binary should be the same as with the
macro.
The new source tag_pool.c manages a pool of reference counted tag_item
objects. This is used to merge tag items of the same type and value,
saving lots of memory. Formerly, only the value itself was pooled,
wasting memory for all the pointers and tag_item structs.
The following results were measured with massif. Started MPD on
amd64, typed "mpc", no song being played. My music database contains
35k tagged songs. The results are what massif reports as "peak".
0.13.2: total 14,131,392; useful 11,408,972; extra 2,722,420
eric: total 18,370,696; useful 15,648,182; extra 2,722,514
mk f34f694: total 15,833,952; useful 13,111,470; extra 2,722,482
mk now: total 12,837,632; useful 10,626,383; extra 2,211,249
This patch set saves 20% memory, and does a good job in reducing heap
fragmentation.
The value is stored in the same memory allocation as the tag_item
struct; this saves memory because we do not store the value pointer
anymore. Also remove the getTagItemString()/removeTagItemString()
dummies.
This patch makes MPD consume much more memory because string pooling
is disabled, but it prepares the next bunch of patches. Replace the
code in tagTracker.c with naive algorithms without the tree code. For
now, this should do; later we should find better algorithms,
especially for getNumberOfTagItems(), which has become wasteful with
temporary memory.
Unfortunately, the C standard postulates that the argument to free()
must be non-const. This does not makes sense, and virtually prevents
every pointer which must be freed at some time to be non-const. Use
the deconst hack (sorry for that) to allow us to free constant
pointers.
Instead of passing the pointer to the "expired" flag to
processListOfCommands(), this function should use the client API to
check this flag. We can now remove the "global_expired" hack
introduced recently.
Start exporting the client struct as an opaque struct. For now, pass
it only to processCommand() and processListOfCommands(), and provide a
function to extract the socket handle. Later, we will propagate the
pointer to all command implementations, and of course to
client_print() etc.
The old code tried to write a response to the client, without even
checking if it was already closed. Now that we have added more
assertions, these may fail... perform the "expired" check earlier.
Patch bdeb8e14 ("client: moved "expired" accesses into inline
function") was created under the wrong assumption that
processListOfCommands() could modify the expired flag, which is not
the case. Although "expired" is a non-const pointer,
processListOfCommands() just reads it, using it as the break condition
in a "while" loop. I will address this issue with a better overall
solution, but for now provide a pointer to a global "expired" flag.
client_defer_output() was modified so that it can create the
deferred_send list. With this patch, the assertion on
"deferred_send!=NULL" has become invalid. Remove it.
Previously, when select() failed, we assumed that there was an invalid
file descriptor in one of the client structs. Thus we tried select()
one by one. This is bogus, because we should never have invalid file
descriptors. Remove it, and make select() errors fatal.
All of the client's resources are freed in client_close(). It is
enough to set the "expired" flag, no need to duplicate lots of
destruction code again and again.
Due to the large buffers in the client struct, the static client array
eats several megabytes of RAM with a maximum of only 10 clients. Stop
this waste and allocate each client struct from the heap.
Second patch: rename the internal struct name. We will eventually
export this type as an opaque forward-declared struct later, so we
can pass a struct pointer instead of a file descriptor, which would
save us an expensive linear lookup.
I don't believe "interface" is a good name for something like
"connection by a client to MPD", let's call it "client". This is the
first patch in the series which changes the name, beginning with the
file name.
linux/list.h is a nice doubly linked list library - it is lightweight
and powerful at the same time. It will be useful later, when we begin
to allocate client structures dynamically. Import it, and strip out
all the stuff which we are not going to use.
It should be obvious in which thread or context a function is being
executed at runtime. The code which was left in decode.c is for the
decoder thread itself; give the file a better name.
This releases several include file dependencies. As a side effect,
"CHUNK_SIZE" isn't defined by decoder_api.h anymore, so we have to
define it directly in the plugins which need it. It just isn't worth
it to add it to the decoder plugin API.
The decoder plugins need this type, so export it in the public API.
This allows is to remove "decode.h" from "decoder_api.h", uncluttering
the API namespace some more.
Unfortunately, we have to pass the DecoderControl pointer to these
inline functions, because the global variable "dc" may not be
available here. This will be fixed later.
The decoder thread is responsible for resetting dc->command after a
command was executed. As a consequence, we can assume that
dc->command is already NONE after decoder_stop().
There is no unlocked caller of clearPlayerQueue(), and the functions
lockPlaylistInteraction() and unlockPlaylistInteraction() are trivial
- merge them.
Since playerPlay() already calls playerStop(), we can remove its
invocation of playerStop() from playPlaylistOrderNumber().
We can also make playerStop a static function.
All (indirect) callers of queueSong() ensure that the queue state is
BLANK, so there is no need to check it in queueSong() again. As a
side effect, queueSong() cannot fail anymore, and can return void.
Also, playlist_queueError and all its error handling can go away.
playerKill() was marked as deprecated, but it seems like a good idea
to do proper cleanup in all threads (e.g. for usable valgrind
results). Introduce the command "EXIT" which makes the player thread
exit cleanly.
playerWait() stops the player thread (twice!) and closes the output
device. It should be well enough to just send CLOSE_AUDIO, without
STOP.
This requires a tiny change to the player thread code: make it break
when CLOSE_AUDIO is sent.
It was possible for the decoder thread to go into an endless loop
(flac and oggflac decoders): when a "STOP" command arrived, the Read()
callback would return 0, but the EOF() callback returned false. Fix:
when decoder_get_command()!=NONE, return EOF==true.
Storing local configuration in global (static) variables is obviously
a bad idea. Move all those variables into the JackData struct,
including the locks.
There is only one caller of freeJackData() left: jack_finishDriver().
This function is called by the mpd core, and is called exactly once
for every successful jack_initDriver(). We do not need to clear
audioOutput->data, since this variable is invalidated anyway.
Over the lifetime of the jack AudioOutput object, we want a single
valid JackData object, so we can persistently store data there
(configuration etc.). Allocate JackData in jack_initDriver(). After
that, we can safely remove all audioOutput->data==NULL checks (and
replace them with assertions).
No need to destroy the JackData object when an error occurs, since
jack_finishDriver() already frees it. Only deinitialize the jack
library, introduce freeJackClient() for that, and move code from
freeJackData().
Prepare the next patch: make the "!jd" check independent of the
jd->client initialization. This way we can change the "jd"
initialization semantics later.
connect_jack() invokes freeJackData() in every error handler, although
its caller also invokes this function after a failure. We can save a
lot of lines in connect_jack() by removing these redundant
freeJackData() invocations.
When we introduced decoder_read(), we added code which aborts the read
operation when a decoder command arrives. Several plugins however did
not expect that when they were converted to decoder_read(). Add
proper checks to the mp3 and flac decoder plugins.
The code said "decoder_command==STOP" because that was a conversion
from the old "dc->stop" test. As we can now check for all commands in
one test, we can simply rewrite that to decoder_command!=NONE.
This flag is used internally; it is set by decoder_seek_where(), and
indicates that the decoder plugin has begun the seek process. It is
used for the case that the decoder plugin has to read data during the
seek process. Before this patch, that was impossible, because
decoder_read() would refuse to read data unless dc->command is NONE.
This patch is kind of a dirty workaround, and needs to be redesigned
later.
The old code called can_seek() with the uninitialized pointer
"isp.is". Has this ever worked? Anyway, initialize "isp" first, then
call can_seek(&isp).
Move everything related to finding and initializing the WVC stream to
wavpack_open_wvc(). This greatly simplifies its error handling and
the function wavpack_streamdecode().
On our way to stabilize the decoder API, we will one day remove the
input stream functions. The most basic function, read() will be
provided by decoder_api.h with this patch. It already contains a loop
(still with manual polling), error/eof handling and decoder command
checks. This kind of code used to be duplicated in all decoder
plugins.
If the input stream is not seekable, the try_decode() function
consumes valuable data, which is not available to the decode()
function anymore. This means that the decode() function does not
parse the header correctly. Better skip the detection if we cannot
seek. Or implement better buffering, something like unread() or
buffered rewind().
The return value of audio_linear_dither() is always casted to
mpd_sint16. Returning long does not make sense, and consumed 8 bytes
on a 64 bit platform.
The output buffer is always flushed after being appended to, which
allows us to assume it is always empty. Always start writing at
outputBuffer, don't remember outputPtr.
Fill the whole output buffer at a time by using dither_buffer()'s
ability to decode blocks. Calculate how many samples fit into the
output buffer before each invocation.
Simplifying loops for performance: why check dropSamplesAtEnd in every
iteration, when we could modify the loop boundary? The (writable)
variable samplesLeft can be eliminated; add a write-once variable
pcm_length instead, which is used for the loop condition.
The variable samplesPerFrame is used only in one single closure. Make
it local to this closure. The compiler will probably convert it to a
register anyway.
Preparing for simplifying and thus speeding up the dithering code:
moved dithering to a separate function which contains a trivial loop.
With this patch, only one sample is dithered at a time, but the
following patches will allow us to dither a whole block at a time,
without complicated buffer length checks.
Copy some code from aac_decode() to aac_stream_decode() and apply
necessary changes to allow streaming audio data. Both functions might
be merged later.
initAacBuffer() should really only initialize the buffer; currently,
it also reads data from the input stream and parses the header. All
of the AAC buffer code should probably be moved to a separate library
anyway.
Shifting from the buffer queue is a common operation, and should be
provided as a separate function. Move code to aac_buffer_shift() and
add a bunch of assertions.
When checking for EOF, we should not check whether the read request
has been fully satisified. The InputStream API does not guarantee
that readFromInputStream() always fills the whole buffer, if EOF is
not reached. Since there is the function inputStreamAtEOF() dedicated
for this purpose, we should use it for EOF checking after
readFromInputStream()==0.
Fill the AacBuffer even when nothing has been consumed yet. The
function should not check for consumed data, but for free space at the
end of the buffer.
The flag "ready" indicates whether the input stream is ready and it
has parsed all meta data. Previously, it was impossible for
decodeStart() to see the content type of HTTP input streams, because
at that time, the HTTP response wasn't parsed yet.
With the functions decoder_plugin_register() and
decoder_plugin_unregister(), decoder plugins can register a
"secondary" plugin, like the flac input plugin does this for
"oggflac".
"decoder plugin" is a better name than "input plugin", since the
plugin does not actually do the input - InputStream does. Also don't
use typedef, so we can forward-declare it if required.
PlayerControl.command replaces the old attributes play, stop, pause,
closeAudio, lockQueue, unlockQueue, seek. The main thread waits for
each command synchronously, so there can only be one command enabled
at a time anyway.
The wavpack decoder plugin implements a hack, and it needs the song
URL for that. This API (and the hack) should be revised later, but
add that function for now.
Since we want to hide mpd internals from the decoder plugins, the
plugins should not check dc->state whether they have already called
decoder_initialized(). Use a local variable to track that.
Some decoder commands are implemented in the decoder plugins, thus
they need to have an API call to signal that their current command has
been finished. Let them use the new decoder_command_finished()
instead of the internal dc_command_finished().
Another big patch which hides internal mpd APIs from decoder plugins:
decoder plugins regularly poll dc->command; expose it with a
decoder_api.h function.
Since we moved all PCM conversions to decoder_data(), the attribute
convState isn't being used anymore by the OutputBuffer code. Move it
to struct decoder.
InputPlugin is the API which is implemented by a decoder plugin. This
belongs to the public API/ABI, so move it to decoder_api.h. It will
later be renamed to something like "decoder_plugin".
Since we have merged dc->stop, dc->seek into one variable, we don't
have to check both conditions at a time; we can replace "!stop &&
!seek" with "none".
dc->audioFormat is set once by the decoder plugins before invoking
decoder_initialized(); hide dc->audioFormat and let the decoder pass
an AudioFormat pointer to decoder_initialized().
We are now beginning to remove direct structure accesses from the
decoder plugins. decoder_clear() and decoder_flush() mask two very
common buffer functions.
Code simplification: since we are not using in-band signalling with
the chunk index anymore, we can just return a pointer to the tail
chunk instead of the index.
OutputBuffer should be a more generic low-level library, without
dependencies to the other headers. This patch adds the field
"notify", which is used to signal the player thread. It is passed in
the constructor, and removes the need to compile with the decode.h
header.
After the decoder has been initialized and the audio device has been
opened, don't sleep. The decoder plugin won't do anything special nor
will it care to wake us up for some reason.
decoder_initialized() sets the state to DECODE_STATE_DECODE and wakes
up the player thread. It is called by the decoder plugin after its
internal initialization is finished. More arguments will be added
later to prevent direct accesses to the DecoderControl struct.
The decoder struct should later be made opaque to the decoder plugin,
because maintaining a stable struct ABI is quite difficult. The ABI
should only consist of a small number of stable functions.
Don't use wrappers like player_wakeup_decoder_nb(). These have been
wrappers calling notify.c functions, for compatibility with the
existing code when we migrated to notify.c.
dc_command_finished() is invoked by the decoder thread when it has
finished a command (sent by the player thread). It resets dc.command
and wakes up the player thread. This combination was used at a lot of
places, and by introducing this function, the code will be more
readable.
Busy wait loops are a bad thing, especially when the response time can
be very long - busy waits eat a lot of CPU, and thus slow down the
other thread. Since the other thread will notify us when it's ready,
we can use notify_wait() instead.
Much of the existing code queries all three variables sequentially.
Since only one of them can be set at a time, this can be optimized and
unified by merging all of them into one enum variable. Later, the
"command" checks can be expressed in a "switch" statement.
Since pc->current_song denotes the song which the decoder should use
next, we should move it to DecoderControl. This removes one internal
PlayerControl struct access from the decoder code.
Also add pc.next_song, which is manipulated by the playlist code, and
gets copied to dc.next_song as soon as the decoder is started.
Also enable -Wunused-parameter - this forces us to add the gcc
"unused" attribute to a lot of parameters (mostly library callback
functions), but it's worth it during code refactorizations.
If nothing has been read from the input stream, we don't have to
rewind it.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7397 09075e82-0dd4-0310-85a5-a0d7c8717e4f
The variable "to_read" is never modified except in the last iteration
of the while loop. This means the while condition will never become
false, as the body will break before that may be checked.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7396 09075e82-0dd4-0310-85a5-a0d7c8717e4f
This institutes the usage of a separate thread to buffer HTTP
input. It is basically practice code for using the ringbuffer
code which I plan on reusing for the OutputBuffer as well as
further input buffering for disk (networked filesystems over
WAN, laptops on battery, etc).
Each readFromInputStream() call on an HTTP stream can take
several seconds to complete, short reads are avoided.
A single-threaded solution for systems supporting large enough
SO_RCVBUF values should also be possible and will likely be done
in the future; but this lock-free(except when full/empty)
ringbuffer is cool :)
git-svn-id: https://svn.musicpd.org/mpd/trunk@7393 09075e82-0dd4-0310-85a5-a0d7c8717e4f
We'll be using pipes when waiting for I/O, and condition
variables at other times.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7391 09075e82-0dd4-0310-85a5-a0d7c8717e4f
This will allow both the reader and writer threads to
reset the ringbuffer in a thread-safe fashion.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7390 09075e82-0dd4-0310-85a5-a0d7c8717e4f
This will eliminate unnecessary calls to ringbuf_{read,write}_space
git-svn-id: https://svn.musicpd.org/mpd/trunk@7389 09075e82-0dd4-0310-85a5-a0d7c8717e4f
The auth code also has some ugly usages of string generation
which I will eventually replace with something nicer...
git-svn-id: https://svn.musicpd.org/mpd/trunk@7387 09075e82-0dd4-0310-85a5-a0d7c8717e4f
This piece of code is from the JACK Audio Connection Kit
(trimmed down a bit for better readability).
The vector functions now reuse the common iovec struct used by
writev/readv instead of reinventing an identical but
differently-named struct.
From the comments:
> ISO/POSIX C version of Paul Davis's lock free ringbuffer C++ code.
> This is safe for the case of one read thread and one write thread.
License is LGPL 2.1 or later
git-svn-id: https://svn.musicpd.org/mpd/trunk@7386 09075e82-0dd4-0310-85a5-a0d7c8717e4f
Initialize audioOutput->data with NULL in jack_initDriver().
Previously, this was never initialized, although the other functions
relied on it being NULL prior to jack_openDevice().
This patch addresses bug 0001641[1]. In contrast to the patch provided
by the bug reporter, it moves the initialization before the "!param"
check.
[1] - http://musicpd.org/mantis/view.php?id=1641
git-svn-id: https://svn.musicpd.org/mpd/trunk@7375 09075e82-0dd4-0310-85a5-a0d7c8717e4f
Streaming was broken, beacuse the stream URL was never copied to
path_max_fs.
[ew: replaced strcpy with pathcpy_trunc for ease of auditing]
git-svn-id: https://svn.musicpd.org/mpd/trunk@7371 09075e82-0dd4-0310-85a5-a0d7c8717e4f
During the decoder thread main loop, dc.state must be
DECODE_STATE_STOP. Explicitly assigning it after the "dc.stop" check
is redundant.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7366 09075e82-0dd4-0310-85a5-a0d7c8717e4f
The function wait_main_task() is racy: if the function
wakeup_via_cond() sees the mutex is locked just before
wait_main_task() executes pthread_cond_wait(), the main thread blocks
forever.
Work around this issue by adding a "pending" flag just like in my
notify.c code. A standards-compliant solution should be implemented
later.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7365 09075e82-0dd4-0310-85a5-a0d7c8717e4f
In lazy mode (previously the default), outputBuffer.c only wakes up
the player when it was previously empty. That caused a deadlock when
the player was waiting for buffered_before_play, since the decoder
wouldn't wake up the player when buffered_before_play was reached. In
non-lazy mode, always wake up the player when a new chunk was decoded.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7364 09075e82-0dd4-0310-85a5-a0d7c8717e4f
When we are in an input plugin, dc.current_song should already be
set. Use it instead of pc.current_song.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7363 09075e82-0dd4-0310-85a5-a0d7c8717e4f
We had functions names varied between
outputBufferFoo, fooOutputBuffer, and output_buffer_foo
That was too confusing for my little brain to handle.
And the global variable was somehow named 'cb' instead of
the more obvious 'ob'...
git-svn-id: https://svn.musicpd.org/mpd/trunk@7355 09075e82-0dd4-0310-85a5-a0d7c8717e4f
All of our main singleton data structures are implicitly shared,
so there's no reason to keep passing them around and around in
the stack and making our internal API harder to deal with.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7354 09075e82-0dd4-0310-85a5-a0d7c8717e4f
This at least makes the argument list to a lot of our plugin
functions shorter and removes a good amount of line nois^W^Wcode,
hopefully making things easier to read and follow.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7353 09075e82-0dd4-0310-85a5-a0d7c8717e4f
It actually increases our image size a small bit and may even
hurt performance a very small bit, but makes the code less
verbose and easier to manage.
I don't see a reason for mpd to ever support playing multiple
files at the same time (users can run multiple instances of mpd
if they really want to play Zaireeka, but that's such an edge
case it's not worth ever supporting in our code).
git-svn-id: https://svn.musicpd.org/mpd/trunk@7352 09075e82-0dd4-0310-85a5-a0d7c8717e4f
signal is all we need since we only have one waiter and
likely faster
git-svn-id: https://svn.musicpd.org/mpd/trunk@7349 09075e82-0dd4-0310-85a5-a0d7c8717e4f
The select() in the main event loop blocks now (saving us many
unnecessary wakeups). This interacted badly with the threads
that were trying to wakeup the main task via
pthread_cond_signal() since the main task was not blocked
on a condition variable, but on select().
So now if we detect a need to wakeup the player, we write
to a pipe which select() is watching instead of blindly
calling pthread_cond_signal().
git-svn-id: https://svn.musicpd.org/mpd/trunk@7347 09075e82-0dd4-0310-85a5-a0d7c8717e4f
Since tailChunk() automatically flushes full buffers, we do not have
to check this in every iteration of sendDataToOutputBuffer().
git-svn-id: https://svn.musicpd.org/mpd/trunk@7343 09075e82-0dd4-0310-85a5-a0d7c8717e4f
Checking dc->stop in the while condition and again after the while
loop costs some CPU cycles we should save.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7340 09075e82-0dd4-0310-85a5-a0d7c8717e4f
OutputBuffer.currentChunk contains redundant data: it is either -1
when there is no chunk which is currently being written, or it equals
"end". If we always keep chunk[end] in a valid state, we can remove
OutputBuffer.currentChunk.
This patch may look a bit clumsy, especially flushOutputBuffer(), but
that will be fixed later with an major OutputBuffer API overhaul.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7339 09075e82-0dd4-0310-85a5-a0d7c8717e4f
output_buffer_expand() moves the cb->end to the new position (only its
current successor is allowed) and wakes up the player if is waiting
for the decoder. This simplifies flushOutputBuffer().
git-svn-id: https://svn.musicpd.org/mpd/trunk@7338 09075e82-0dd4-0310-85a5-a0d7c8717e4f
The current OutputBuffer object is allocated statically, i.e. it is
zeroed. To be safe for other cases in the future, also initialize the
other elements.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7337 09075e82-0dd4-0310-85a5-a0d7c8717e4f
The decoder should not wake up the player when it did not produce a
flushed chunk. Move the decoder_wakeup_player() call to
flushOutputBuffer() and invoke it only if the buffer was previously
empty.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7336 09075e82-0dd4-0310-85a5-a0d7c8717e4f
During my tests, it happened that data->position>newPosition. I have
not yet fully understood why this can happen; for now, replace this
with a run-time check.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7334 09075e82-0dd4-0310-85a5-a0d7c8717e4f
The patch "convert blocks until the buffer is full" did not update
data->chunk_length correctly: it added the number of samples, not the
number of bytes. Multiply that with bytes_per_channel
git-svn-id: https://svn.musicpd.org/mpd/trunk@7332 09075e82-0dd4-0310-85a5-a0d7c8717e4f
In the patch "special optimized case for 16bit stereo", the check for
"num_channels==2" was missing.
git-svn-id: https://svn.musicpd.org/mpd/trunk@7331 09075e82-0dd4-0310-85a5-a0d7c8717e4f
Max Kellermann