X-Git-Url: https://git.distorted.org.uk/~mdw/disorder/blobdiff_plain/e83d0967d4c0965eb8036248acc20d1bf12ad1d8..7f9d584755d62ebc613fe339303b453ee4210068:/server/speaker.c?ds=inline
diff --git a/server/speaker.c b/server/speaker.c
index 98ed297..e65ce91 100644
--- a/server/speaker.c
+++ b/server/speaker.c
@@ -17,14 +17,35 @@
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307
* USA
*/
-
-/* This program deliberately does not use the garbage collector even though it
- * might be convenient to do so. This is for two reasons. Firstly some libao
- * drivers are implemented using threads and we do not want to have to deal
- * with potential interactions between threading and garbage collection.
- * Secondly this process needs to be able to respond quickly and this is not
- * compatible with the collector hanging the program even relatively
- * briefly. */
+/** @file server/speaker.c
+ * @brief Speaker processs
+ *
+ * This program is responsible for transmitting a single coherent audio stream
+ * to its destination (over the network, to some sound API, to some
+ * subprocess). It receives connections from decoders via file descriptor
+ * passing from the main server and plays them in the right order.
+ *
+ * For the ALSA API, 8- and 16- bit
+ * stereo and mono are supported, with any sample rate (within the limits that
+ * ALSA can deal with.)
+ *
+ * When communicating with a subprocess, sox is invoked to convert the inbound
+ * data to a single consistent format. The same applies for network (RTP)
+ * play, though in that case currently only 44.1KHz 16-bit stereo is supported.
+ *
+ * The inbound data starts with a structure defining the data format. Note
+ * that this is NOT portable between different platforms or even necessarily
+ * between versions; the speaker is assumed to be built from the same source
+ * and run on the same host as the main server.
+ *
+ * This program deliberately does not use the garbage collector even though it
+ * might be convenient to do so. This is for two reasons. Firstly some sound
+ * APIs use thread threads and we do not want to have to deal with potential
+ * interactions between threading and garbage collection. Secondly this
+ * process needs to be able to respond quickly and this is not compatible with
+ * the collector hanging the program even relatively briefly.
+ */
#include
#include "types.h"
@@ -70,20 +91,32 @@
# define MACHINE_AO_FMT AO_FMT_LITTLE
#endif
-#define BUFFER_SECONDS 5 /* How many seconds of input to
- * buffer. */
+/** @brief How many seconds of input to buffer
+ *
+ * While any given connection has this much audio buffered, no more reads will
+ * be issued for that connection. The decoder will have to wait.
+ */
+#define BUFFER_SECONDS 5
#define FRAMES 4096 /* Frame batch size */
-#define NETWORK_BYTES 1024 /* Bytes to send per network packet */
-/* (don't make this too big or arithmetic will start to overflow) */
+/** @brief Bytes to send per network packet
+ *
+ * Don't make this too big or arithmetic will start to overflow.
+ */
+#define NETWORK_BYTES (1024+sizeof(struct rtp_header))
-#define RTP_AHEAD 2 /* Max RTP playahead (seconds) */
+/** @brief Maximum RTP playahead (ms) */
+#define RTP_AHEAD_MS 1000
-#define NFDS 256 /* Max FDs to poll for */
+/** @brief Maximum number of FDs to poll for */
+#define NFDS 256
-/* Known tracks are kept in a linked list. We don't normally to have
- * more than two - maybe three at the outside. */
+/** @brief Track structure
+ *
+ * Known tracks are kept in a linked list. Usually there will be at most two
+ * of these but rearranging the queue can cause there to be more.
+ */
static struct track {
struct track *next; /* next track */
int fd; /* input FD */
@@ -100,26 +133,110 @@ static struct track {
static time_t last_report; /* when we last reported */
static int paused; /* pause status */
-static ao_sample_format pcm_format; /* current format if aodev != 0 */
static size_t bpf; /* bytes per frame */
static struct pollfd fds[NFDS]; /* if we need more than that */
static int fdno; /* fd number */
static size_t bufsize; /* buffer size */
#if API_ALSA
-static snd_pcm_t *pcm; /* current pcm handle */
+/** @brief The current PCM handle */
+static snd_pcm_t *pcm;
static snd_pcm_uframes_t last_pcm_bufsize; /* last seen buffer size */
+static ao_sample_format pcm_format; /* current format if aodev != 0 */
#endif
-static int ready; /* ready to send audio */
+
+/** @brief Ready to send audio
+ *
+ * This is set when the destination is ready to receive audio. Generally
+ * this implies that the sound device is open. In the ALSA backend it
+ * does @b not necessarily imply that is has the right sample format.
+ */
+static int ready;
+
static int forceplay; /* frames to force play */
static int cmdfd = -1; /* child process input */
static int bfd = -1; /* broadcast FD */
-static uint32_t rtp_time; /* RTP timestamp */
-static struct timeval rtp_time_real; /* corresponding real time */
+
+/** @brief RTP timestamp
+ *
+ * This counts the number of samples played (NB not the number of frames
+ * played).
+ *
+ * The timestamp in the packet header is only 32 bits wide. With 44100Hz
+ * stereo, that only gives about half a day before wrapping, which is not
+ * particularly convenient for certain debugging purposes. Therefore the
+ * timestamp is maintained as a 64-bit integer, giving around six million years
+ * before wrapping, and truncated to 32 bits when transmitting.
+ */
+static uint64_t rtp_time;
+
+/** @brief RTP base timestamp
+ *
+ * This is the real time correspoding to an @ref rtp_time of 0. It is used
+ * to recalculate the timestamp after idle periods.
+ */
+static struct timeval rtp_time_0;
+
static uint16_t rtp_seq; /* frame sequence number */
static uint32_t rtp_id; /* RTP SSRC */
static int idled; /* set when idled */
static int audio_errors; /* audio error counter */
+/** @brief Structure of a backend */
+struct speaker_backend {
+ /** @brief Which backend this is
+ *
+ * @c -1 terminates the list.
+ */
+ int backend;
+
+ /** @brief Flags
+ *
+ * Possible values
+ * - @ref FIXED_FORMAT
+ */
+ unsigned flags;
+/** @brief Lock to configured sample format */
+#define FIXED_FORMAT 0x0001
+
+ /** @brief Initialization
+ *
+ * Called once at startup. This is responsible for one-time setup
+ * operations, for instance opening a network socket to transmit to.
+ *
+ * When writing to a native sound API this might @b not imply opening the
+ * native sound device - that might be done by @c activate below.
+ */
+ void (*init)(void);
+
+ /** @brief Activation
+ * @return 0 on success, non-0 on error
+ *
+ * Called to activate the output device.
+ *
+ * After this function succeeds, @ref ready should be non-0. As well as
+ * opening the audio device, this function is responsible for reconfiguring
+ * if it necessary to cope with different samples formats (for backends that
+ * don't demand a single fixed sample format for the lifetime of the server).
+ */
+ int (*activate)(void);
+
+ /** @brief Play sound
+ * @param frames Number of frames to play
+ * @return Number of frames actually played
+ */
+ size_t (*play)(size_t frames);
+
+ /** @brief Deactivation
+ *
+ * Called to deactivate the sound device. This is the inverse of
+ * @c activate above.
+ */
+ void (*deactivate)(void);
+};
+
+/** @brief Selected backend */
+static const struct speaker_backend *backend;
+
static const struct option options[] = {
{ "help", no_argument, 0, 'h' },
{ "version", no_argument, 0, 'V' },
@@ -152,12 +269,12 @@ static void version(void) {
exit(0);
}
-/* Return the number of bytes per frame in FORMAT. */
+/** @brief Return the number of bytes per frame in @p format */
static size_t bytes_per_frame(const ao_sample_format *format) {
return format->channels * format->bits / 8;
}
-/* Find track ID, maybe creating it if not found. */
+/** @brief Find track @p id, maybe creating it if not found */
static struct track *findtrack(const char *id, int create) {
struct track *t;
@@ -177,7 +294,7 @@ static struct track *findtrack(const char *id, int create) {
return t;
}
-/* Remove track ID (but do not destroy it). */
+/** @brief Remove track @p id (but do not destroy it) */
static struct track *removetrack(const char *id) {
struct track *t, **tt;
@@ -189,7 +306,7 @@ static struct track *removetrack(const char *id) {
return t;
}
-/* Destroy a track. */
+/** @brief Destroy a track */
static void destroy(struct track *t) {
D(("destroy %s", t->id));
if(t->fd != -1) xclose(t->fd);
@@ -197,7 +314,7 @@ static void destroy(struct track *t) {
free(t);
}
-/* Notice a new FD. */
+/** @brief Notice a new connection */
static void acquire(struct track *t, int fd) {
D(("acquire %s %d", t->id, fd));
if(t->fd != -1)
@@ -206,7 +323,102 @@ static void acquire(struct track *t, int fd) {
nonblock(fd);
}
-/* Read data into a sample buffer. Return 0 on success, -1 on EOF. */
+/** @brief Return true if A and B denote identical libao formats, else false */
+static int formats_equal(const ao_sample_format *a,
+ const ao_sample_format *b) {
+ return (a->bits == b->bits
+ && a->rate == b->rate
+ && a->channels == b->channels
+ && a->byte_format == b->byte_format);
+}
+
+/** @brief Compute arguments to sox */
+static void soxargs(const char ***pp, char **qq, ao_sample_format *ao) {
+ int n;
+
+ *(*pp)++ = "-t.raw";
+ *(*pp)++ = "-s";
+ *(*pp)++ = *qq; n = sprintf(*qq, "-r%d", ao->rate); *qq += n + 1;
+ *(*pp)++ = *qq; n = sprintf(*qq, "-c%d", ao->channels); *qq += n + 1;
+ /* sox 12.17.9 insists on -b etc; CVS sox insists on - etc; both are
+ * deployed! */
+ switch(config->sox_generation) {
+ case 0:
+ if(ao->bits != 8
+ && ao->byte_format != AO_FMT_NATIVE
+ && ao->byte_format != MACHINE_AO_FMT) {
+ *(*pp)++ = "-x";
+ }
+ switch(ao->bits) {
+ case 8: *(*pp)++ = "-b"; break;
+ case 16: *(*pp)++ = "-w"; break;
+ case 32: *(*pp)++ = "-l"; break;
+ case 64: *(*pp)++ = "-d"; break;
+ default: fatal(0, "cannot handle sample size %d", (int)ao->bits);
+ }
+ break;
+ case 1:
+ switch(ao->byte_format) {
+ case AO_FMT_NATIVE: break;
+ case AO_FMT_BIG: *(*pp)++ = "-B"; break;
+ case AO_FMT_LITTLE: *(*pp)++ = "-L"; break;
+ }
+ *(*pp)++ = *qq; n = sprintf(*qq, "-%d", ao->bits/8); *qq += n + 1;
+ break;
+ }
+}
+
+/** @brief Enable format translation
+ *
+ * If necessary, replaces a tracks inbound file descriptor with one connected
+ * to a sox invocation, which performs the required translation.
+ */
+static void enable_translation(struct track *t) {
+ if((backend->flags & FIXED_FORMAT)
+ && !formats_equal(&t->format, &config->sample_format)) {
+ char argbuf[1024], *q = argbuf;
+ const char *av[18], **pp = av;
+ int soxpipe[2];
+ pid_t soxkid;
+
+ *pp++ = "sox";
+ soxargs(&pp, &q, &t->format);
+ *pp++ = "-";
+ soxargs(&pp, &q, &config->sample_format);
+ *pp++ = "-";
+ *pp++ = 0;
+ if(debugging) {
+ for(pp = av; *pp; pp++)
+ D(("sox arg[%d] = %s", pp - av, *pp));
+ D(("end args"));
+ }
+ xpipe(soxpipe);
+ soxkid = xfork();
+ if(soxkid == 0) {
+ signal(SIGPIPE, SIG_DFL);
+ xdup2(t->fd, 0);
+ xdup2(soxpipe[1], 1);
+ fcntl(0, F_SETFL, fcntl(0, F_GETFL) & ~O_NONBLOCK);
+ close(soxpipe[0]);
+ close(soxpipe[1]);
+ close(t->fd);
+ execvp("sox", (char **)av);
+ _exit(1);
+ }
+ D(("forking sox for format conversion (kid = %d)", soxkid));
+ close(t->fd);
+ close(soxpipe[1]);
+ t->fd = soxpipe[0];
+ t->format = config->sample_format;
+ }
+}
+
+/** @brief Read data into a sample buffer
+ * @param t Pointer to track
+ * @return 0 on success, -1 on EOF
+ *
+ * This is effectively the read callback on @c t->fd.
+ */
static int fill(struct track *t) {
size_t where, left;
int n;
@@ -242,6 +454,8 @@ static int fill(struct track *t) {
/* Check that our assumptions are met. */
if(t->format.bits & 7)
fatal(0, "bits per sample not a multiple of 8");
+ /* If the input format is unsuitable, arrange to translate it */
+ enable_translation(t);
/* Make a new buffer for audio data. */
t->size = bytes_per_frame(&t->format) * t->format.rate * BUFFER_SECONDS;
t->buffer = xmalloc(t->size);
@@ -253,38 +467,16 @@ static int fill(struct track *t) {
return 0;
}
-/* Return true if A and B denote identical libao formats, else false. */
-static int formats_equal(const ao_sample_format *a,
- const ao_sample_format *b) {
- return (a->bits == b->bits
- && a->rate == b->rate
- && a->channels == b->channels
- && a->byte_format == b->byte_format);
-}
-
-/* Close the sound device. */
+/** @brief Close the sound device */
static void idle(void) {
D(("idle"));
-#if API_ALSA
- if(config->speaker_backend == BACKEND_ALSA && pcm) {
- int err;
-
- if((err = snd_pcm_nonblock(pcm, 0)) < 0)
- fatal(0, "error calling snd_pcm_nonblock: %d", err);
- D(("draining pcm"));
- snd_pcm_drain(pcm);
- D(("closing pcm"));
- snd_pcm_close(pcm);
- pcm = 0;
- forceplay = 0;
- D(("released audio device"));
- }
-#endif
+ if(backend->deactivate)
+ backend->deactivate();
idled = 1;
ready = 0;
}
-/* Abandon the current track */
+/** @brief Abandon the current track */
static void abandon(void) {
struct speaker_message sm;
@@ -300,6 +492,7 @@ static void abandon(void) {
}
#if API_ALSA
+/** @brief Log ALSA parameters */
static void log_params(snd_pcm_hw_params_t *hwparams,
snd_pcm_sw_params_t *swparams) {
snd_pcm_uframes_t f;
@@ -326,203 +519,18 @@ static void log_params(snd_pcm_hw_params_t *hwparams,
}
#endif
-static void soxargs(const char ***pp, char **qq, ao_sample_format *ao) {
- int n;
-
- *(*pp)++ = "-t.raw";
- *(*pp)++ = "-s";
- *(*pp)++ = *qq; n = sprintf(*qq, "-r%d", ao->rate); *qq += n + 1;
- *(*pp)++ = *qq; n = sprintf(*qq, "-c%d", ao->channels); *qq += n + 1;
- /* sox 12.17.9 insists on -b etc; CVS sox insists on - etc; both are
- * deployed! */
- switch(config->sox_generation) {
- case 0:
- if(ao->bits != 8
- && ao->byte_format != AO_FMT_NATIVE
- && ao->byte_format != MACHINE_AO_FMT) {
- *(*pp)++ = "-x";
- }
- switch(ao->bits) {
- case 8: *(*pp)++ = "-b"; break;
- case 16: *(*pp)++ = "-w"; break;
- case 32: *(*pp)++ = "-l"; break;
- case 64: *(*pp)++ = "-d"; break;
- default: fatal(0, "cannot handle sample size %d", (int)ao->bits);
- }
- break;
- case 1:
- switch(ao->byte_format) {
- case AO_FMT_NATIVE: break;
- case AO_FMT_BIG: *(*pp)++ = "-B"; break;
- case AO_FMT_LITTLE: *(*pp)++ = "-L"; break;
- }
- *(*pp)++ = *qq; n = sprintf(*qq, "-%d", ao->bits/8); *qq += n + 1;
- break;
- }
-}
-
-/* Make sure the sound device is open and has the right sample format. Return
- * 0 on success and -1 on error. */
+/** @brief Enable sound output
+ *
+ * Makes sure the sound device is open and has the right sample format. Return
+ * 0 on success and -1 on error.
+ */
static int activate(void) {
/* If we don't know the format yet we cannot start. */
if(!playing->got_format) {
D((" - not got format for %s", playing->id));
return -1;
}
- switch(config->speaker_backend) {
- case BACKEND_COMMAND:
- case BACKEND_NETWORK:
- /* If we pass audio on to some other agent then we enforce the configured
- * sample format on the *inbound* audio data. */
- if(!formats_equal(&playing->format, &config->sample_format)) {
- char argbuf[1024], *q = argbuf;
- const char *av[18], **pp = av;
- int soxpipe[2];
- pid_t soxkid;
- *pp++ = "sox";
- soxargs(&pp, &q, &playing->format);
- *pp++ = "-";
- soxargs(&pp, &q, &config->sample_format);
- *pp++ = "-";
- *pp++ = 0;
- if(debugging) {
- for(pp = av; *pp; pp++)
- D(("sox arg[%d] = %s", pp - av, *pp));
- D(("end args"));
- }
- xpipe(soxpipe);
- soxkid = xfork();
- if(soxkid == 0) {
- xdup2(playing->fd, 0);
- xdup2(soxpipe[1], 1);
- fcntl(0, F_SETFL, fcntl(0, F_GETFL) & ~O_NONBLOCK);
- close(soxpipe[0]);
- close(soxpipe[1]);
- close(playing->fd);
- execvp("sox", (char **)av);
- _exit(1);
- }
- D(("forking sox for format conversion (kid = %d)", soxkid));
- close(playing->fd);
- close(soxpipe[1]);
- playing->fd = soxpipe[0];
- playing->format = config->sample_format;
- ready = 0;
- }
- if(!ready) {
- pcm_format = config->sample_format;
- bufsize = 3 * FRAMES;
- bpf = bytes_per_frame(&config->sample_format);
- D(("acquired audio device"));
- ready = 1;
- }
- return 0;
- case BACKEND_ALSA:
-#if API_ALSA
- /* If we need to change format then close the current device. */
- if(pcm && !formats_equal(&playing->format, &pcm_format))
- idle();
- if(!pcm) {
- snd_pcm_hw_params_t *hwparams;
- snd_pcm_sw_params_t *swparams;
- snd_pcm_uframes_t pcm_bufsize;
- int err;
- int sample_format = 0;
- unsigned rate;
-
- D(("snd_pcm_open"));
- if((err = snd_pcm_open(&pcm,
- config->device,
- SND_PCM_STREAM_PLAYBACK,
- SND_PCM_NONBLOCK))) {
- error(0, "error from snd_pcm_open: %d", err);
- goto error;
- }
- snd_pcm_hw_params_alloca(&hwparams);
- D(("set up hw params"));
- if((err = snd_pcm_hw_params_any(pcm, hwparams)) < 0)
- fatal(0, "error from snd_pcm_hw_params_any: %d", err);
- if((err = snd_pcm_hw_params_set_access(pcm, hwparams,
- SND_PCM_ACCESS_RW_INTERLEAVED)) < 0)
- fatal(0, "error from snd_pcm_hw_params_set_access: %d", err);
- switch(playing->format.bits) {
- case 8:
- sample_format = SND_PCM_FORMAT_S8;
- break;
- case 16:
- switch(playing->format.byte_format) {
- case AO_FMT_NATIVE: sample_format = SND_PCM_FORMAT_S16; break;
- case AO_FMT_LITTLE: sample_format = SND_PCM_FORMAT_S16_LE; break;
- case AO_FMT_BIG: sample_format = SND_PCM_FORMAT_S16_BE; break;
- error(0, "unrecognized byte format %d", playing->format.byte_format);
- goto fatal;
- }
- break;
- default:
- error(0, "unsupported sample size %d", playing->format.bits);
- goto fatal;
- }
- if((err = snd_pcm_hw_params_set_format(pcm, hwparams,
- sample_format)) < 0) {
- error(0, "error from snd_pcm_hw_params_set_format (%d): %d",
- sample_format, err);
- goto fatal;
- }
- rate = playing->format.rate;
- if((err = snd_pcm_hw_params_set_rate_near(pcm, hwparams, &rate, 0)) < 0) {
- error(0, "error from snd_pcm_hw_params_set_rate (%d): %d",
- playing->format.rate, err);
- goto fatal;
- }
- if(rate != (unsigned)playing->format.rate)
- info("want rate %d, got %u", playing->format.rate, rate);
- if((err = snd_pcm_hw_params_set_channels(pcm, hwparams,
- playing->format.channels)) < 0) {
- error(0, "error from snd_pcm_hw_params_set_channels (%d): %d",
- playing->format.channels, err);
- goto fatal;
- }
- bufsize = 3 * FRAMES;
- pcm_bufsize = bufsize;
- if((err = snd_pcm_hw_params_set_buffer_size_near(pcm, hwparams,
- &pcm_bufsize)) < 0)
- fatal(0, "error from snd_pcm_hw_params_set_buffer_size (%d): %d",
- 3 * FRAMES, err);
- if(pcm_bufsize != 3 * FRAMES && pcm_bufsize != last_pcm_bufsize)
- info("asked for PCM buffer of %d frames, got %d",
- 3 * FRAMES, (int)pcm_bufsize);
- last_pcm_bufsize = pcm_bufsize;
- if((err = snd_pcm_hw_params(pcm, hwparams)) < 0)
- fatal(0, "error calling snd_pcm_hw_params: %d", err);
- D(("set up sw params"));
- snd_pcm_sw_params_alloca(&swparams);
- if((err = snd_pcm_sw_params_current(pcm, swparams)) < 0)
- fatal(0, "error calling snd_pcm_sw_params_current: %d", err);
- if((err = snd_pcm_sw_params_set_avail_min(pcm, swparams, FRAMES)) < 0)
- fatal(0, "error calling snd_pcm_sw_params_set_avail_min %d: %d",
- FRAMES, err);
- if((err = snd_pcm_sw_params(pcm, swparams)) < 0)
- fatal(0, "error calling snd_pcm_sw_params: %d", err);
- pcm_format = playing->format;
- bpf = bytes_per_frame(&pcm_format);
- D(("acquired audio device"));
- log_params(hwparams, swparams);
- ready = 1;
- }
- return 0;
- fatal:
- abandon();
- error:
- /* We assume the error is temporary and that we'll retry in a bit. */
- if(pcm) {
- snd_pcm_close(pcm);
- pcm = 0;
- }
- return -1;
-#endif
- default:
- assert(!"reached");
- }
+ return backend->activate();
}
/* Check to see whether the current track has finished playing */
@@ -541,6 +549,7 @@ static void fork_cmd(void) {
xpipe(pfd);
cmdpid = xfork();
if(!cmdpid) {
+ signal(SIGPIPE, SIG_DFL);
xdup2(pfd[0], 0);
close(pfd[0]);
close(pfd[1]);
@@ -553,11 +562,12 @@ static void fork_cmd(void) {
}
static void play(size_t frames) {
- size_t avail_bytes, written_frames;
+ size_t avail_frames, avail_bytes, write_bytes, written_frames;
ssize_t written_bytes;
- struct rtp header;
+ struct rtp_header header;
struct iovec vec[2];
+ /* Make sure the output device is activated */
if(activate()) {
if(playing)
forceplay = frames;
@@ -580,22 +590,24 @@ static void play(size_t frames) {
forceplay = 0;
/* Figure out how many frames there are available to write */
if(playing->start + playing->used > playing->size)
+ /* The ring buffer is currently wrapped, only play up to the wrap point */
avail_bytes = playing->size - playing->start;
else
+ /* The ring buffer is not wrapped, can play the lot */
avail_bytes = playing->used;
+ avail_frames = avail_bytes / bpf;
+ /* Only play up to the requested amount */
+ if(avail_frames > frames)
+ avail_frames = frames;
+ if(!avail_frames)
+ return;
switch(config->speaker_backend) {
#if API_ALSA
- case BACKEND_ALSA:
+ case BACKEND_ALSA: {
snd_pcm_sframes_t pcm_written_frames;
- size_t avail_frames;
int err;
- avail_frames = avail_bytes / bpf;
- if(avail_frames > frames)
- avail_frames = frames;
- if(!avail_frames)
- return;
pcm_written_frames = snd_pcm_writei(pcm,
playing->buffer + playing->start,
avail_frames);
@@ -618,6 +630,7 @@ static void play(size_t frames) {
written_frames = pcm_written_frames;
written_bytes = written_frames * bpf;
break;
+ }
#endif
case BACKEND_COMMAND:
if(avail_bytes > frames * bpf)
@@ -641,19 +654,67 @@ static void play(size_t frames) {
case BACKEND_NETWORK:
/* We transmit using RTP (RFC3550) and attempt to conform to the internet
* AVT profile (RFC3551). */
- if(rtp_time_real.tv_sec == 0)
- xgettimeofday(&rtp_time_real, 0);
+
if(idled) {
+ /* There may have been a gap. Fix up the RTP time accordingly. */
struct timeval now;
+ uint64_t delta;
+ uint64_t target_rtp_time;
+
+ /* Find the current time */
xgettimeofday(&now, 0);
- /* There's been a gap. Fix up the RTP time accordingly. */
- rtp_time += (((now.tv_sec + now.tv_usec /1000000.0)
- - (rtp_time_real.tv_sec + rtp_time_real.tv_usec / 1000000.0))
- * playing->format.rate * playing->format.channels);
+ /* Find the number of microseconds elapsed since rtp_time=0 */
+ delta = tvsub_us(now, rtp_time_0);
+ assert(delta <= UINT64_MAX / 88200);
+ target_rtp_time = (delta * playing->format.rate
+ * playing->format.channels) / 1000000;
+ /* Overflows at ~6 years uptime with 44100Hz stereo */
+
+ /* rtp_time is the number of samples we've played. NB that we play
+ * RTP_AHEAD_MS ahead of ourselves, so it may legitimately be ahead of
+ * the value we deduce from time comparison.
+ *
+ * Suppose we have 1s track started at t=0, and another track begins to
+ * play at t=2s. Suppose RTP_AHEAD_MS=1000 and 44100Hz stereo. In that
+ * case we'll send 1s of audio as fast as we can, giving rtp_time=88200.
+ * rtp_time stops at this point.
+ *
+ * At t=2s we'll have calculated target_rtp_time=176400. In this case we
+ * set rtp_time=176400 and the player can correctly conclude that it
+ * should leave 1s between the tracks.
+ *
+ * Suppose instead that the second track arrives at t=0.5s, and that
+ * we've managed to transmit the whole of the first track already. We'll
+ * have target_rtp_time=44100.
+ *
+ * The desired behaviour is to play the second track back to back with
+ * first. In this case therefore we do not modify rtp_time.
+ *
+ * Is it ever right to reduce rtp_time? No; for that would imply
+ * transmitting packets with overlapping timestamp ranges, which does not
+ * make sense.
+ */
+ if(target_rtp_time > rtp_time) {
+ /* More time has elapsed than we've transmitted samples. That implies
+ * we've been 'sending' silence. */
+ info("advancing rtp_time by %"PRIu64" samples",
+ target_rtp_time - rtp_time);
+ rtp_time = target_rtp_time;
+ } else if(target_rtp_time < rtp_time) {
+ const int64_t samples_ahead = ((uint64_t)RTP_AHEAD_MS
+ * config->sample_format.rate
+ * config->sample_format.channels
+ / 1000);
+
+ if(target_rtp_time + samples_ahead < rtp_time) {
+ info("reversing rtp_time by %"PRIu64" samples",
+ rtp_time - target_rtp_time);
+ }
+ }
}
header.vpxcc = 2 << 6; /* V=2, P=0, X=0, CC=0 */
header.seq = htons(rtp_seq++);
- header.timestamp = htonl(rtp_time);
+ header.timestamp = htonl((uint32_t)rtp_time);
header.ssrc = rtp_id;
header.mpt = (idled ? 0x80 : 0x00) | 10;
/* 10 = L16 = 16-bit x 2 x 44100KHz. We ought to deduce this value from
@@ -663,6 +724,7 @@ static void play(size_t frames) {
idled = 0;
if(avail_bytes > NETWORK_BYTES - sizeof header) {
avail_bytes = NETWORK_BYTES - sizeof header;
+ /* Always send a whole number of frames */
avail_bytes -= avail_bytes % bpf;
}
/* "The RTP clock rate used for generating the RTP timestamp is independent
@@ -673,45 +735,30 @@ static void play(size_t frames) {
* generated per second is then the sampling rate times the channel
* count.)"
*/
- vec[0].iov_base = (void *)&header;
- vec[0].iov_len = sizeof header;
- vec[1].iov_base = playing->buffer + playing->start;
- vec[1].iov_len = avail_bytes;
-#if 0
- {
- char buffer[3 * sizeof header + 1];
- size_t n;
- const uint8_t *ptr = (void *)&header;
-
- for(n = 0; n < sizeof header; ++n)
- sprintf(&buffer[3 * n], "%02x ", *ptr++);
- info(buffer);
- }
-#endif
- do {
- written_bytes = writev(bfd,
- vec,
- 2);
- } while(written_bytes < 0 && errno == EINTR);
- if(written_bytes < 0) {
- error(errno, "error transmitting audio data");
- ++audio_errors;
- if(audio_errors == 10)
- fatal(0, "too many audio errors");
+ write_bytes = avail_bytes;
+ if(write_bytes) {
+ vec[0].iov_base = (void *)&header;
+ vec[0].iov_len = sizeof header;
+ vec[1].iov_base = playing->buffer + playing->start;
+ vec[1].iov_len = avail_bytes;
+ do {
+ written_bytes = writev(bfd,
+ vec,
+ 2);
+ } while(written_bytes < 0 && errno == EINTR);
+ if(written_bytes < 0) {
+ error(errno, "error transmitting audio data");
+ ++audio_errors;
+ if(audio_errors == 10)
+ fatal(0, "too many audio errors");
return;
- }
+ }
+ } else
audio_errors /= 2;
written_bytes = avail_bytes;
written_frames = written_bytes / bpf;
/* Advance RTP's notion of the time */
rtp_time += written_frames * playing->format.channels;
- /* Advance the corresponding real time */
- assert(NETWORK_BYTES <= 2000); /* else risk overflowing 32 bits */
- rtp_time_real.tv_usec += written_frames * 1000000 / playing->format.rate;
- if(rtp_time_real.tv_usec >= 1000000) {
- ++rtp_time_real.tv_sec;
- rtp_time_real.tv_usec -= 1000000;
- }
break;
default:
assert(!"reached");
@@ -761,11 +808,163 @@ static int addfd(int fd, int events) {
return -1;
}
-int main(int argc, char **argv) {
- int n, fd, stdin_slot, alsa_slots, cmdfd_slot, bfd_slot, poke, timeout;
- struct timeval now, delta;
- struct track *t;
- struct speaker_message sm;
+#if API_ALSA
+/** @brief ALSA backend initialization */
+static void alsa_init(void) {
+ info("selected ALSA backend");
+}
+
+/** @brief ALSA backend activation */
+static int alsa_activate(void) {
+ /* If we need to change format then close the current device. */
+ if(pcm && !formats_equal(&playing->format, &pcm_format))
+ idle();
+ if(!pcm) {
+ snd_pcm_hw_params_t *hwparams;
+ snd_pcm_sw_params_t *swparams;
+ snd_pcm_uframes_t pcm_bufsize;
+ int err;
+ int sample_format = 0;
+ unsigned rate;
+
+ D(("snd_pcm_open"));
+ if((err = snd_pcm_open(&pcm,
+ config->device,
+ SND_PCM_STREAM_PLAYBACK,
+ SND_PCM_NONBLOCK))) {
+ error(0, "error from snd_pcm_open: %d", err);
+ goto error;
+ }
+ snd_pcm_hw_params_alloca(&hwparams);
+ D(("set up hw params"));
+ if((err = snd_pcm_hw_params_any(pcm, hwparams)) < 0)
+ fatal(0, "error from snd_pcm_hw_params_any: %d", err);
+ if((err = snd_pcm_hw_params_set_access(pcm, hwparams,
+ SND_PCM_ACCESS_RW_INTERLEAVED)) < 0)
+ fatal(0, "error from snd_pcm_hw_params_set_access: %d", err);
+ switch(playing->format.bits) {
+ case 8:
+ sample_format = SND_PCM_FORMAT_S8;
+ break;
+ case 16:
+ switch(playing->format.byte_format) {
+ case AO_FMT_NATIVE: sample_format = SND_PCM_FORMAT_S16; break;
+ case AO_FMT_LITTLE: sample_format = SND_PCM_FORMAT_S16_LE; break;
+ case AO_FMT_BIG: sample_format = SND_PCM_FORMAT_S16_BE; break;
+ error(0, "unrecognized byte format %d", playing->format.byte_format);
+ goto fatal;
+ }
+ break;
+ default:
+ error(0, "unsupported sample size %d", playing->format.bits);
+ goto fatal;
+ }
+ if((err = snd_pcm_hw_params_set_format(pcm, hwparams,
+ sample_format)) < 0) {
+ error(0, "error from snd_pcm_hw_params_set_format (%d): %d",
+ sample_format, err);
+ goto fatal;
+ }
+ rate = playing->format.rate;
+ if((err = snd_pcm_hw_params_set_rate_near(pcm, hwparams, &rate, 0)) < 0) {
+ error(0, "error from snd_pcm_hw_params_set_rate (%d): %d",
+ playing->format.rate, err);
+ goto fatal;
+ }
+ if(rate != (unsigned)playing->format.rate)
+ info("want rate %d, got %u", playing->format.rate, rate);
+ if((err = snd_pcm_hw_params_set_channels(pcm, hwparams,
+ playing->format.channels)) < 0) {
+ error(0, "error from snd_pcm_hw_params_set_channels (%d): %d",
+ playing->format.channels, err);
+ goto fatal;
+ }
+ bufsize = 3 * FRAMES;
+ pcm_bufsize = bufsize;
+ if((err = snd_pcm_hw_params_set_buffer_size_near(pcm, hwparams,
+ &pcm_bufsize)) < 0)
+ fatal(0, "error from snd_pcm_hw_params_set_buffer_size (%d): %d",
+ 3 * FRAMES, err);
+ if(pcm_bufsize != 3 * FRAMES && pcm_bufsize != last_pcm_bufsize)
+ info("asked for PCM buffer of %d frames, got %d",
+ 3 * FRAMES, (int)pcm_bufsize);
+ last_pcm_bufsize = pcm_bufsize;
+ if((err = snd_pcm_hw_params(pcm, hwparams)) < 0)
+ fatal(0, "error calling snd_pcm_hw_params: %d", err);
+ D(("set up sw params"));
+ snd_pcm_sw_params_alloca(&swparams);
+ if((err = snd_pcm_sw_params_current(pcm, swparams)) < 0)
+ fatal(0, "error calling snd_pcm_sw_params_current: %d", err);
+ if((err = snd_pcm_sw_params_set_avail_min(pcm, swparams, FRAMES)) < 0)
+ fatal(0, "error calling snd_pcm_sw_params_set_avail_min %d: %d",
+ FRAMES, err);
+ if((err = snd_pcm_sw_params(pcm, swparams)) < 0)
+ fatal(0, "error calling snd_pcm_sw_params: %d", err);
+ pcm_format = playing->format;
+ bpf = bytes_per_frame(&pcm_format);
+ D(("acquired audio device"));
+ log_params(hwparams, swparams);
+ ready = 1;
+ }
+ return 0;
+fatal:
+ abandon();
+error:
+ /* We assume the error is temporary and that we'll retry in a bit. */
+ if(pcm) {
+ snd_pcm_close(pcm);
+ pcm = 0;
+ }
+ return -1;
+}
+
+/** @brief Play via ALSA */
+static size_t alsa_play(size_t frames) {
+ return frames;
+}
+
+/** @brief ALSA deactivation */
+static void alsa_deactivate(void) {
+ if(pcm) {
+ int err;
+
+ if((err = snd_pcm_nonblock(pcm, 0)) < 0)
+ fatal(0, "error calling snd_pcm_nonblock: %d", err);
+ D(("draining pcm"));
+ snd_pcm_drain(pcm);
+ D(("closing pcm"));
+ snd_pcm_close(pcm);
+ pcm = 0;
+ forceplay = 0;
+ D(("released audio device"));
+ }
+}
+#endif
+
+/** @brief Command backend initialization */
+static void command_init(void) {
+ info("selected command backend");
+ fork_cmd();
+}
+
+/** @brief Play to a subprocess */
+static size_t command_play(size_t frames) {
+ return frames;
+}
+
+/** @brief Command/network backend activation */
+static int generic_activate(void) {
+ if(!ready) {
+ bufsize = 3 * FRAMES;
+ bpf = bytes_per_frame(&config->sample_format);
+ D(("acquired audio device"));
+ ready = 1;
+ }
+ return 0;
+}
+
+/** @brief Network backend initialization */
+static void network_init(void) {
struct addrinfo *res, *sres;
static const struct addrinfo pref = {
0,
@@ -788,7 +987,92 @@ int main(int argc, char **argv) {
0
};
static const int one = 1;
+ int sndbuf, target_sndbuf = 131072;
+ socklen_t len;
char *sockname, *ssockname;
+
+ res = get_address(&config->broadcast, &pref, &sockname);
+ if(!res) exit(-1);
+ if(config->broadcast_from.n) {
+ sres = get_address(&config->broadcast_from, &prefbind, &ssockname);
+ if(!sres) exit(-1);
+ } else
+ sres = 0;
+ if((bfd = socket(res->ai_family,
+ res->ai_socktype,
+ res->ai_protocol)) < 0)
+ fatal(errno, "error creating broadcast socket");
+ if(setsockopt(bfd, SOL_SOCKET, SO_BROADCAST, &one, sizeof one) < 0)
+ fatal(errno, "error setting SO_BROADCAST on broadcast socket");
+ len = sizeof sndbuf;
+ if(getsockopt(bfd, SOL_SOCKET, SO_SNDBUF,
+ &sndbuf, &len) < 0)
+ fatal(errno, "error getting SO_SNDBUF");
+ if(target_sndbuf > sndbuf) {
+ if(setsockopt(bfd, SOL_SOCKET, SO_SNDBUF,
+ &target_sndbuf, sizeof target_sndbuf) < 0)
+ error(errno, "error setting SO_SNDBUF to %d", target_sndbuf);
+ else
+ info("changed socket send buffer size from %d to %d",
+ sndbuf, target_sndbuf);
+ } else
+ info("default socket send buffer is %d",
+ sndbuf);
+ /* We might well want to set additional broadcast- or multicast-related
+ * options here */
+ if(sres && bind(bfd, sres->ai_addr, sres->ai_addrlen) < 0)
+ fatal(errno, "error binding broadcast socket to %s", ssockname);
+ if(connect(bfd, res->ai_addr, res->ai_addrlen) < 0)
+ fatal(errno, "error connecting broadcast socket to %s", sockname);
+ /* Select an SSRC */
+ gcry_randomize(&rtp_id, sizeof rtp_id, GCRY_STRONG_RANDOM);
+ info("selected network backend, sending to %s", sockname);
+ if(config->sample_format.byte_format != AO_FMT_BIG) {
+ info("forcing big-endian sample format");
+ config->sample_format.byte_format = AO_FMT_BIG;
+ }
+}
+
+/** @brief Play over the network */
+static size_t network_play(size_t frames) {
+ return frames;
+}
+
+/** @brief Table of speaker backends */
+static const struct speaker_backend backends[] = {
+#if API_ALSA
+ {
+ BACKEND_ALSA,
+ 0,
+ alsa_init,
+ alsa_activate,
+ alsa_play,
+ alsa_deactivate
+ },
+#endif
+ {
+ BACKEND_COMMAND,
+ FIXED_FORMAT,
+ command_init,
+ generic_activate,
+ command_play,
+ 0 /* deactivate */
+ },
+ {
+ BACKEND_NETWORK,
+ FIXED_FORMAT,
+ network_init,
+ generic_activate,
+ network_play,
+ 0 /* deactivate */
+ },
+ { -1, 0, 0, 0, 0, 0 }
+};
+
+int main(int argc, char **argv) {
+ int n, fd, stdin_slot, alsa_slots, cmdfd_slot, bfd_slot, poke, timeout;
+ struct track *t;
+ struct speaker_message sm;
#if API_ALSA
int alsa_nslots = -1, err;
#endif
@@ -822,44 +1106,15 @@ int main(int argc, char **argv) {
become_mortal();
/* make sure we're not root, whatever the config says */
if(getuid() == 0 || geteuid() == 0) fatal(0, "do not run as root");
- switch(config->speaker_backend) {
- case BACKEND_ALSA:
- info("selected ALSA backend");
- case BACKEND_COMMAND:
- info("selected command backend");
- fork_cmd();
- break;
- case BACKEND_NETWORK:
- res = get_address(&config->broadcast, &pref, &sockname);
- if(!res) return -1;
- if(config->broadcast_from.n) {
- sres = get_address(&config->broadcast_from, &prefbind, &ssockname);
- if(!sres) return -1;
- } else
- sres = 0;
- if((bfd = socket(res->ai_family,
- res->ai_socktype,
- res->ai_protocol)) < 0)
- fatal(errno, "error creating broadcast socket");
- if(setsockopt(bfd, SOL_SOCKET, SO_BROADCAST, &one, sizeof one) < 0)
- fatal(errno, "error settting SO_BROADCAST on broadcast socket");
- /* We might well want to set additional broadcast- or multicast-related
- * options here */
- if(sres && bind(bfd, sres->ai_addr, sres->ai_addrlen) < 0)
- fatal(errno, "error binding broadcast socket to %s", ssockname);
- if(connect(bfd, res->ai_addr, res->ai_addrlen) < 0)
- fatal(errno, "error connecting broadcast socket to %s", sockname);
- /* Select an SSRC */
- gcry_randomize(&rtp_id, sizeof rtp_id, GCRY_STRONG_RANDOM);
- info("selected network backend, sending to %s", sockname);
- if(config->sample_format.byte_format != AO_FMT_BIG) {
- info("forcing big-endian sample format");
- config->sample_format.byte_format = AO_FMT_BIG;
- }
- break;
- default:
- fatal(0, "unknown backend %d", config->speaker_backend);
- }
+ /* identify the backend used to play */
+ for(n = 0; backends[n].backend != -1; ++n)
+ if(backends[n].backend == config->speaker_backend)
+ break;
+ if(backends[n].backend == -1)
+ fatal(0, "unsupported backend %d", config->speaker_backend);
+ backend = &backends[n];
+ /* backend-specific initialization */
+ backend->init();
while(getppid() != 1) {
fdno = 0;
/* Always ready for commands from the main server. */
@@ -886,24 +1141,44 @@ int main(int argc, char **argv) {
if(cmdfd >= 0)
cmdfd_slot = addfd(cmdfd, POLLOUT);
break;
- case BACKEND_NETWORK:
- /* We want to keep the notional playing point somewhere in the near
- * future. If it's too near then clients that attempt even the
- * slightest amount of read-ahead will never catch up, and those that
- * don't will skip whenever there's a trivial network delay. If it's
- * too far ahead then pause latency will be too high.
- */
+ case BACKEND_NETWORK: {
+ struct timeval now;
+ uint64_t target_us;
+ uint64_t target_rtp_time;
+ const int64_t samples_ahead = ((uint64_t)RTP_AHEAD_MS
+ * config->sample_format.rate
+ * config->sample_format.channels
+ / 1000);
+#if 0
+ static unsigned logit;
+#endif
+
+ /* If we're starting then initialize the base time */
+ if(!rtp_time)
+ xgettimeofday(&rtp_time_0, 0);
+ /* We send audio data whenever we get RTP_AHEAD seconds or more
+ * behind */
xgettimeofday(&now, 0);
- delta = tvsub(rtp_time_real, now);
- if(delta.tv_sec < RTP_AHEAD) {
- D(("delta = %ld.%06ld", (long)delta.tv_sec, (long)delta.tv_usec));
+ target_us = tvsub_us(now, rtp_time_0);
+ assert(target_us <= UINT64_MAX / 88200);
+ target_rtp_time = (target_us * config->sample_format.rate
+ * config->sample_format.channels)
+
+ / 1000000;
+#if 0
+ /* TODO remove logging guff */
+ if(!(logit++ & 1023))
+ info("rtp_time %llu target %llu difference %lld [%lld]",
+ rtp_time, target_rtp_time,
+ rtp_time - target_rtp_time,
+ samples_ahead);
+#endif
+ if((int64_t)(rtp_time - target_rtp_time) < samples_ahead)
bfd_slot = addfd(bfd, POLLOUT);
- if(delta.tv_sec < 0)
- rtp_time_real = now; /* catch up */
- }
break;
+ }
#if API_ALSA
- case BACKEND_ALSA:
+ case BACKEND_ALSA: {
/* We send sample data to ALSA as fast as it can accept it, relying on
* the fact that it has a relatively small buffer to minimize pause
* latency. */
@@ -925,6 +1200,7 @@ int main(int argc, char **argv) {
if(alsa_nslots >= 0)
fdno += alsa_nslots;
break;
+ }
#endif
default:
assert(!"unknown backend");