* USA
*/
/** @file server/speaker.c
- * @brief Speaker processs
+ * @brief Speaker process
*
* 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.)
+ * @b Encodings. 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
* 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.
+ * @b Garbage @b Collection. 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.
+ *
+ * @b Units. This program thinks at various times in three different units.
+ * Bytes are obvious. A sample is a single sample on a single channel. A
+ * frame is several samples on different channels at the same point in time.
+ * So (for instance) a 16-bit stereo frame is 4 bytes and consists of a pair of
+ * 2-byte samples.
*/
#include <config.h>
#include "log.h"
#include "defs.h"
#include "mem.h"
-#include "speaker.h"
+#include "speaker-protocol.h"
#include "user.h"
#include "addr.h"
#include "timeval.h"
#include "rtp.h"
+#include "speaker.h"
#if API_ALSA
#include <alsa/asoundlib.h>
#endif
-#ifdef WORDS_BIGENDIAN
-# define MACHINE_AO_FMT AO_FMT_BIG
-#else
-# define MACHINE_AO_FMT AO_FMT_LITTLE
-#endif
+/** @brief Linked list of all prepared tracks */
+struct track *tracks;
-/** @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 */
-
-/** @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))
-
-/** @brief Maximum RTP playahead (ms) */
-#define RTP_AHEAD_MS 1000
-
-/** @brief Maximum number of FDs to poll for */
-#define NFDS 256
-
-/** @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 */
- char id[24]; /* ID */
- size_t start, used; /* start + bytes used */
- int eof; /* input is at EOF */
- int got_format; /* got format yet? */
- ao_sample_format format; /* sample format */
- unsigned long long played; /* number of frames played */
- char *buffer; /* sample buffer */
- size_t size; /* sample buffer size */
- int slot; /* poll array slot */
-} *tracks, *playing; /* all tracks + playing track */
+/** @brief Playing track, or NULL */
+struct track *playing;
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 */
-static int forceplay; /* frames to force play */
-static int cmdfd = -1; /* child process input */
-static int bfd = -1; /* broadcast FD */
+
+/** @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;
+
+/** @brief Frames to force-play
+ *
+ * If this is nonzero, and playing is enabled, then the main loop will attempt
+ * to play this many frames without checking whether the output device is
+ * ready.
+ */
+static int forceplay;
+
+/** @brief Pipe to subprocess
+ *
+ * This is the file descriptor to write to for @ref BACKEND_COMMAND.
+ */
+static int cmdfd = -1;
+
+/** @brief Network socket
+ *
+ * This is the file descriptor to write to for @ref BACKEND_NETWORK.
+ */
+static int bfd = -1;
/** @brief RTP timestamp
*
*/
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 RTP packet sequence number */
+static uint16_t rtp_seq;
-/** @brief Structure of a backend */
-struct speaker_backend {
- /** @brief Which backend this is
- *
- * @c -1 terminates the list.
- */
- int backend;
-
- /** @brief Initialization
- *
- * Called once at startup.
- */
- void (*init)(void);
+/** @brief RTP SSRC */
+static uint32_t rtp_id;
- /** @brief Activation
- * @return 0 on success, non-0 on error
- *
- * Called to activate the output device.
- */
- int (*activate)(void);
-};
+/** @brief Set when idled
+ *
+ * This is set when the sound device is deliberately closed by idle().
+ * @ref ready is set to 0 at the same time.
+ */
+static int idled; /* set when idled */
+
+/** @brief Error counter */
+static int audio_errors;
/** @brief Selected backend */
static const struct speaker_backend *backend;
* 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)) {
+ 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];
* @param t Pointer to track
* @return 0 on success, -1 on EOF
*
- * This is effectively the read callback on @c t->fd.
+ * This is effectively the read callback on @c t->fd. It is called from the
+ * main loop whenever the track's file descriptor is readable, assuming the
+ * buffer has not reached the maximum allowed occupancy.
*/
static int fill(struct track *t) {
size_t where, left;
return 0;
}
-/** @brief Close the sound device */
+/** @brief Close the sound device
+ *
+ * This is called to deactivate the output device when pausing, and also by the
+ * ALSA backend when changing encoding (in which case the sound device will be
+ * immediately reactivated).
+ */
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;
}
return backend->activate();
}
-/* Check to see whether the current track has finished playing */
+/** @brief Check whether the current track has finished
+ *
+ * The current track is determined to have finished either if the input stream
+ * eded before the format could be determined (i.e. it is malformed) or the
+ * input is at end of file and there is less than a frame left unplayed. (So
+ * it copes with decoders that crash mid-frame.)
+ */
static void maybe_finished(void) {
if(playing
&& playing->eof
abandon();
}
+/** @brief Start the subprocess for @ref BACKEND_COMMAND */
static void fork_cmd(void) {
pid_t cmdpid;
int pfd[2];
D(("forked cmd %d, fd = %d", cmdpid, cmdfd));
}
+/** @brief Play up to @p frames frames of audio */
static void play(size_t frames) {
- size_t avail_bytes, write_bytes, written_frames;
+ size_t avail_frames, avail_bytes, written_frames;
ssize_t written_bytes;
- struct rtp_header header;
- struct iovec vec[2];
+ /* Make sure the output device is activated */
if(activate()) {
if(playing)
forceplay = 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;
-
- switch(config->speaker_backend) {
-#if API_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);
- D(("actually play %zu frames, wrote %d",
- avail_frames, (int)pcm_written_frames));
- if(pcm_written_frames < 0) {
- switch(pcm_written_frames) {
- case -EPIPE: /* underrun */
- error(0, "snd_pcm_writei reports underrun");
- if((err = snd_pcm_prepare(pcm)) < 0)
- fatal(0, "error calling snd_pcm_prepare: %d", err);
- return;
- case -EAGAIN:
- return;
- default:
- fatal(0, "error calling snd_pcm_writei: %d",
- (int)pcm_written_frames);
- }
- }
- written_frames = pcm_written_frames;
- written_bytes = written_frames * bpf;
- break;
- }
-#endif
- case BACKEND_COMMAND:
- if(avail_bytes > frames * bpf)
- avail_bytes = frames * bpf;
- written_bytes = write(cmdfd, playing->buffer + playing->start,
- avail_bytes);
- D(("actually play %zu bytes, wrote %d",
- avail_bytes, (int)written_bytes));
- if(written_bytes < 0) {
- switch(errno) {
- case EPIPE:
- error(0, "hmm, command died; trying another");
- fork_cmd();
- return;
- case EAGAIN:
- return;
- }
- }
- written_frames = written_bytes / bpf; /* good enough */
- break;
- case BACKEND_NETWORK:
- /* We transmit using RTP (RFC3550) and attempt to conform to the internet
- * AVT profile (RFC3551). */
-
- 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);
- /* 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((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
- * the sample rate (in a library somewhere so that configuration.c can rule
- * out invalid rates).
- */
- 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
- * of the number of channels and the encoding; it equals the number of
- * sampling periods per second. For N-channel encodings, each sampling
- * period (say, 1/8000 of a second) generates N samples. (This terminology
- * is standard, but somewhat confusing, as the total number of samples
- * generated per second is then the sampling rate times the channel
- * count.)"
- */
- 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;
- break;
- default:
- assert(!"reached");
- }
+ avail_frames = avail_bytes / bpf;
+ /* Only play up to the requested amount */
+ if(avail_frames > frames)
+ avail_frames = frames;
+ if(!avail_frames)
+ return;
+ /* Play it, Sam */
+ written_frames = backend->play(avail_frames);
+ written_bytes = written_frames * bpf;
/* written_bytes and written_frames had better both be set and correct by
* this point */
playing->start += written_bytes;
}
return -1;
}
+
+/** @brief Play via ALSA */
+static size_t alsa_play(size_t frames) {
+ snd_pcm_sframes_t pcm_written_frames;
+ int err;
+
+ pcm_written_frames = snd_pcm_writei(pcm,
+ playing->buffer + playing->start,
+ frames);
+ D(("actually play %zu frames, wrote %d",
+ frames, (int)pcm_written_frames));
+ if(pcm_written_frames < 0) {
+ switch(pcm_written_frames) {
+ case -EPIPE: /* underrun */
+ error(0, "snd_pcm_writei reports underrun");
+ if((err = snd_pcm_prepare(pcm)) < 0)
+ fatal(0, "error calling snd_pcm_prepare: %d", err);
+ return 0;
+ case -EAGAIN:
+ return 0;
+ default:
+ fatal(0, "error calling snd_pcm_writei: %d",
+ (int)pcm_written_frames);
+ }
+ } else
+ return pcm_written_frames;
+}
+
+static int alsa_slots, alsa_nslots = -1;
+
+/** @brief Fill in poll fd array for ALSA */
+static void alsa_beforepoll(void) {
+ /* 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. */
+ int retry = 3, err;
+
+ alsa_slots = fdno;
+ do {
+ retry = 0;
+ alsa_nslots = snd_pcm_poll_descriptors(pcm, &fds[fdno], NFDS - fdno);
+ if((alsa_nslots <= 0
+ || !(fds[alsa_slots].events & POLLOUT))
+ && snd_pcm_state(pcm) == SND_PCM_STATE_XRUN) {
+ error(0, "underrun detected after call to snd_pcm_poll_descriptors()");
+ if((err = snd_pcm_prepare(pcm)))
+ fatal(0, "error calling snd_pcm_prepare: %d", err);
+ } else
+ break;
+ } while(retry-- > 0);
+ if(alsa_nslots >= 0)
+ fdno += alsa_nslots;
+}
+
+/** @brief Process poll() results for ALSA */
+static int alsa_afterpoll(void) {
+ int err;
+
+ if(alsa_slots != -1) {
+ unsigned short alsa_revents;
+
+ if((err = snd_pcm_poll_descriptors_revents(pcm,
+ &fds[alsa_slots],
+ alsa_nslots,
+ &alsa_revents)) < 0)
+ fatal(0, "error calling snd_pcm_poll_descriptors_revents: %d", err);
+ if(alsa_revents & (POLLOUT | POLLERR))
+ play(3 * FRAMES);
+ return 0;
+ } else
+ return 1;
+}
+
+/** @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 */
fork_cmd();
}
-/** @brief Command backend activation */
-static int command_activate(void) {
+/** @brief Play to a subprocess */
+static size_t command_play(size_t frames) {
+ size_t bytes = frames * bpf;
+ int written_bytes;
+
+ written_bytes = write(cmdfd, playing->buffer + playing->start, bytes);
+ D(("actually play %zu bytes, wrote %d",
+ bytes, written_bytes));
+ if(written_bytes < 0) {
+ switch(errno) {
+ case EPIPE:
+ error(0, "hmm, command died; trying another");
+ fork_cmd();
+ return 0;
+ case EAGAIN:
+ return 0;
+ default:
+ fatal(errno, "error writing to subprocess");
+ }
+ } else
+ return written_bytes / bpf;
+}
+
+static int cmdfd_slot;
+
+/** @brief Update poll array for writing to subprocess */
+static void command_beforepoll(void) {
+ /* We send sample data to the subprocess as fast as it can accept it.
+ * This isn't ideal as pause latency can be very high as a result. */
+ if(cmdfd >= 0)
+ cmdfd_slot = addfd(cmdfd, POLLOUT);
+}
+
+/** @brief Process poll() results for subprocess play */
+static int command_afterpoll(void) {
+ if(cmdfd_slot != -1) {
+ if(fds[cmdfd_slot].revents & (POLLOUT | POLLERR))
+ play(3 * FRAMES);
+ return 0;
+ } else
+ return -1;
+}
+
+/** @brief Command/network backend activation */
+static int generic_activate(void) {
if(!ready) {
- pcm_format = config->sample_format;
bufsize = 3 * FRAMES;
bpf = bytes_per_frame(&config->sample_format);
D(("acquired audio device"));
}
}
-/** @brief Network backend activation */
-static int network_activate(void) {
- if(!ready) {
- pcm_format = config->sample_format;
- bufsize = 3 * FRAMES;
- bpf = bytes_per_frame(&config->sample_format);
- D(("acquired audio device"));
- ready = 1;
+/** @brief Play over the network */
+static size_t network_play(size_t frames) {
+ struct rtp_header header;
+ struct iovec vec[2];
+ size_t bytes = frames * bpf, written_frames;
+ int written_bytes;
+ /* We transmit using RTP (RFC3550) and attempt to conform to the internet
+ * AVT profile (RFC3551). */
+
+ 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);
+ /* 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);
+ }
+ }
}
- return 0;
+ header.vpxcc = 2 << 6; /* V=2, P=0, X=0, CC=0 */
+ header.seq = htons(rtp_seq++);
+ 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
+ * the sample rate (in a library somewhere so that configuration.c can rule
+ * out invalid rates).
+ */
+ idled = 0;
+ if(bytes > NETWORK_BYTES - sizeof header) {
+ bytes = NETWORK_BYTES - sizeof header;
+ /* Always send a whole number of frames */
+ bytes -= bytes % bpf;
+ }
+ /* "The RTP clock rate used for generating the RTP timestamp is independent
+ * of the number of channels and the encoding; it equals the number of
+ * sampling periods per second. For N-channel encodings, each sampling
+ * period (say, 1/8000 of a second) generates N samples. (This terminology
+ * is standard, but somewhat confusing, as the total number of samples
+ * 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 = 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 0;
+ } else
+ audio_errors /= 2;
+ written_bytes -= sizeof (struct rtp_header);
+ written_frames = written_bytes / bpf;
+ /* Advance RTP's notion of the time */
+ rtp_time += written_frames * playing->format.channels;
+ return written_frames;
+}
+
+static int bfd_slot;
+
+/** @brief Set up poll array for network play */
+static void network_beforepoll(void) {
+ 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 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);
+ 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((int64_t)(rtp_time - target_rtp_time) < samples_ahead)
+ bfd_slot = addfd(bfd, POLLOUT);
+}
+
+/** @brief Process poll() results for network play */
+static int network_afterpoll(void) {
+ if(bfd_slot != -1) {
+ if(fds[bfd_slot].revents & (POLLOUT | POLLERR))
+ play(3 * FRAMES);
+ return 0;
+ } else
+ return 1;
}
/** @brief Table of speaker backends */
#if API_ALSA
{
BACKEND_ALSA,
+ 0,
alsa_init,
- alsa_activate
+ alsa_activate,
+ alsa_play,
+ alsa_deactivate,
+ alsa_beforepoll,
+ alsa_afterpoll
},
#endif
{
BACKEND_COMMAND,
+ FIXED_FORMAT,
command_init,
- command_activate
+ generic_activate,
+ command_play,
+ 0, /* deactivate */
+ command_beforepoll,
+ command_afterpoll
},
{
BACKEND_NETWORK,
+ FIXED_FORMAT,
network_init,
- network_activate
+ generic_activate,
+ network_play,
+ 0, /* deactivate */
+ network_beforepoll,
+ network_afterpoll
},
- { -1, 0, 0 }
+ { -1, 0, 0, 0, 0, 0, 0, 0 } /* end of list */
};
-int main(int argc, char **argv) {
- int n, fd, stdin_slot, alsa_slots, cmdfd_slot, bfd_slot, poke, timeout;
+/** @brief Main event loop
+ *
+ * This has grown in a rather bizarre and ad-hoc way is very sensitive to
+ * changes...
+ *
+ * Firstly the loop is terminated when the parent process exits. Therefore the
+ * speaker process has the same lifetime as the main server. This and the
+ * reading of data from decoders is comprehensible enough.
+ *
+ * The playing of audio is more complicated however.
+ *
+ * On the first run through when a track is ready to be played, @ref ready and
+ * @ref forceplay will both be zero. Therefore @c beforepoll is not called.
+ *
+ * @c afterpoll on the other hand @b is called and will return nonzero. The
+ * result is that we call @c play(0). This will call activate(), setting
+ * @ref ready nonzero, but otherwise has no immediate effect.
+ *
+ * We then deal with stdin and the decoders.
+ *
+ * We then reach the second place we might play some audio. @ref forceplay is
+ * 0 so nothing happens here again.
+ *
+ * On the next iteration through however @ref ready is nonzero, and @ref
+ * forceplay is 0, so we call @c beforepoll. After the @c poll() we call @c
+ * afterpoll and actually get some audio played.
+ *
+ * This is surely @b far more complicated than it needs to be!
+ *
+ * If at any call to play(), activate() fails, or if there aren't enough bytes
+ * in the buffer to satisfy the request, then @ref forceplay is set non-0. On
+ * the next pass through the event loop @c beforepoll is not called. This
+ * means that (if none of the other FDs trigger) the @c poll() call will block
+ * for up to a second. @c afterpoll will return nonzero, since @c beforepoll
+ * wasn't called, and consequently play() is called with @ref forceplay as its
+ * argument.
+ *
+ * The effect is to attempt to restart playing audio - including the activate()
+ * step, which may have failed at the previous attempt - at least once a second
+ * after an error has disabled it. The delay prevents busy-waiting on whatever
+ * condition has rendered the audio device uncooperative.
+ */
+static void mainloop(void) {
struct track *t;
struct speaker_message sm;
-#if API_ALSA
- int alsa_nslots = -1, err;
-#endif
+ int n, fd, stdin_slot, poke, timeout;
- set_progname(argv);
- if(!setlocale(LC_CTYPE, "")) fatal(errno, "error calling setlocale");
- while((n = getopt_long(argc, argv, "hVc:dD", options, 0)) >= 0) {
- switch(n) {
- case 'h': help();
- case 'V': version();
- case 'c': configfile = optarg; break;
- case 'd': debugging = 1; break;
- case 'D': debugging = 0; break;
- default: fatal(0, "invalid option");
- }
- }
- if(getenv("DISORDER_DEBUG_SPEAKER")) debugging = 1;
- /* If stderr is a TTY then log there, otherwise to syslog. */
- if(!isatty(2)) {
- openlog(progname, LOG_PID, LOG_DAEMON);
- log_default = &log_syslog;
- }
- if(config_read()) fatal(0, "cannot read configuration");
- /* ignore SIGPIPE */
- signal(SIGPIPE, SIG_IGN);
- /* reap kids */
- signal(SIGCHLD, reap);
- /* set nice value */
- xnice(config->nice_speaker);
- /* change user */
- become_mortal();
- /* make sure we're not root, whatever the config says */
- if(getuid() == 0 || geteuid() == 0) fatal(0, "do not run as root");
- /* 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. */
playing->slot = -1;
/* If forceplay is set then wait until it succeeds before waiting on the
* sound device. */
+#if API_ALSA
alsa_slots = -1;
+#endif
cmdfd_slot = -1;
bfd_slot = -1;
/* By default we will wait up to a second before thinking about current
* state. */
timeout = 1000;
- if(ready && !forceplay) {
- switch(config->speaker_backend) {
- case BACKEND_COMMAND:
- /* We send sample data to the subprocess as fast as it can accept it.
- * This isn't ideal as pause latency can be very high as a result. */
- if(cmdfd >= 0)
- cmdfd_slot = addfd(cmdfd, POLLOUT);
- break;
- 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);
- 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);
- break;
- }
-#if API_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. */
- int retry = 3;
-
- alsa_slots = fdno;
- do {
- retry = 0;
- alsa_nslots = snd_pcm_poll_descriptors(pcm, &fds[fdno], NFDS - fdno);
- if((alsa_nslots <= 0
- || !(fds[alsa_slots].events & POLLOUT))
- && snd_pcm_state(pcm) == SND_PCM_STATE_XRUN) {
- error(0, "underrun detected after call to snd_pcm_poll_descriptors()");
- if((err = snd_pcm_prepare(pcm)))
- fatal(0, "error calling snd_pcm_prepare: %d", err);
- } else
- break;
- } while(retry-- > 0);
- if(alsa_nslots >= 0)
- fdno += alsa_nslots;
- break;
- }
-#endif
- default:
- assert(!"unknown backend");
- }
- }
+ /* We'll break the poll as soon as the underlying sound device is ready for
+ * more data */
+ if(ready && !forceplay)
+ backend->beforepoll();
/* If any other tracks don't have a full buffer, try to read sample data
* from them. */
for(t = tracks; t; t = t->next)
fatal(errno, "error calling poll");
}
/* Play some sound before doing anything else */
- poke = 0;
- switch(config->speaker_backend) {
-#if API_ALSA
- case BACKEND_ALSA:
- if(alsa_slots != -1) {
- unsigned short alsa_revents;
-
- if((err = snd_pcm_poll_descriptors_revents(pcm,
- &fds[alsa_slots],
- alsa_nslots,
- &alsa_revents)) < 0)
- fatal(0, "error calling snd_pcm_poll_descriptors_revents: %d", err);
- if(alsa_revents & (POLLOUT | POLLERR))
- play(3 * FRAMES);
- } else
- poke = 1;
- break;
-#endif
- case BACKEND_COMMAND:
- if(cmdfd_slot != -1) {
- if(fds[cmdfd_slot].revents & (POLLOUT | POLLERR))
- play(3 * FRAMES);
- } else
- poke = 1;
- break;
- case BACKEND_NETWORK:
- if(bfd_slot != -1) {
- if(fds[bfd_slot].revents & (POLLOUT | POLLERR))
- play(3 * FRAMES);
- } else
- poke = 1;
- break;
- }
+ poke = backend->afterpoll();
if(poke) {
/* Some attempt to play must have failed */
if(playing && !paused)
if(time(0) > last_report)
report();
}
+}
+
+int main(int argc, char **argv) {
+ int n;
+
+ set_progname(argv);
+ if(!setlocale(LC_CTYPE, "")) fatal(errno, "error calling setlocale");
+ while((n = getopt_long(argc, argv, "hVc:dD", options, 0)) >= 0) {
+ switch(n) {
+ case 'h': help();
+ case 'V': version();
+ case 'c': configfile = optarg; break;
+ case 'd': debugging = 1; break;
+ case 'D': debugging = 0; break;
+ default: fatal(0, "invalid option");
+ }
+ }
+ if(getenv("DISORDER_DEBUG_SPEAKER")) debugging = 1;
+ /* If stderr is a TTY then log there, otherwise to syslog. */
+ if(!isatty(2)) {
+ openlog(progname, LOG_PID, LOG_DAEMON);
+ log_default = &log_syslog;
+ }
+ if(config_read()) fatal(0, "cannot read configuration");
+ /* ignore SIGPIPE */
+ signal(SIGPIPE, SIG_IGN);
+ /* reap kids */
+ signal(SIGCHLD, reap);
+ /* set nice value */
+ xnice(config->nice_speaker);
+ /* change user */
+ become_mortal();
+ /* make sure we're not root, whatever the config says */
+ if(getuid() == 0 || geteuid() == 0) fatal(0, "do not run as root");
+ /* 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();
+ mainloop();
info("stopped (parent terminated)");
exit(0);
}