* 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 <a href="http://www.alsa-project.org/">ALSA</a> 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, <a
+ * href="http://sox.sourceforge.net/">sox</a> 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 <config.h>
#include "types.h"
# 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 */
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 */
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;
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;
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);
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)
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 -<n> 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) {
+ switch(config->speaker_backend) {
+ case BACKEND_COMMAND:
+ case BACKEND_NETWORK:
+ /* These backends need a specific sample format */
+ break;
+ case BACKEND_ALSA:
+ /* ALSA can cope */
+ return;
+ }
+ if(!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;
+ ready = 0;
+ }
+}
+
+/** @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;
/* 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);
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
ready = 0;
}
-/* Abandon the current track */
+/** @brief Abandon the current track */
static void abandon(void) {
struct speaker_message sm;
}
#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;
}
#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 -<n> 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) {
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;
xpipe(pfd);
cmdpid = xfork();
if(!cmdpid) {
+ signal(SIGPIPE, SIG_DFL);
xdup2(pfd[0], 0);
close(pfd[0]);
close(pfd[1]);
}
static void play(size_t frames) {
- size_t avail_bytes, written_frames;
+ size_t avail_bytes, write_bytes, written_frames;
ssize_t written_bytes;
struct rtp_header header;
struct iovec vec[2];
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's 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. */
- const long offset = (((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);
- info("offset RTP timestamp by %ld", offset);
- rtp_time += offset;
+ /* 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 */
+ if(target_rtp_time > rtp_time)
+ info("advancing rtp_time by %"PRIu64" samples",
+ target_rtp_time - rtp_time);
+ else if(target_rtp_time < rtp_time)
+ info("reversing rtp_time by %"PRIu64" samples",
+ rtp_time - target_rtp_time);
+ 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
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
* 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");
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;
struct addrinfo *res, *sres;
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);
+ static unsigned logit;
+
+ /* 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 1
+ /* 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: {
/* We send sample data to ALSA as fast as it can accept it, relying on