X-Git-Url: https://git.distorted.org.uk/~mdw/disorder/blobdiff_plain/ea410ba1d3b06ba6f60305c7d31369da134906dd..22b9fa74de8e80471a5033ea067d3b360930b91d:/server/speaker.c diff --git a/server/speaker.c b/server/speaker.c index 2114db4..1dc90e4 100644 --- a/server/speaker.c +++ b/server/speaker.c @@ -18,7 +18,7 @@ * 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 @@ -72,10 +72,6 @@ #include #include #include -#include -#include -#include -#include #include "configuration.h" #include "syscalls.h" @@ -84,218 +80,42 @@ #include "mem.h" #include "speaker-protocol.h" #include "user.h" -#include "addr.h" -#include "timeval.h" -#include "rtp.h" +#include "speaker.h" -#if API_ALSA -#include -#endif +/** @brief Linked list of all prepared tracks */ +struct track *tracks; -#ifdef WORDS_BIGENDIAN -# define MACHINE_AO_FMT AO_FMT_BIG -#else -# define MACHINE_AO_FMT AO_FMT_LITTLE -#endif +/** @brief Playing track, or NULL */ +struct track *playing; -/** @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 +/** @brief Number of bytes pre frame */ +size_t device_bpf; -/** @brief Frame batch size - * - * This controls how many frames are written in one go. - * - * For ALSA we request a buffer of three times this size and set the low - * watermark to this amount. The goal is then to keep between 1 and 3 times - * this many frames in play. - * - * For all backends we attempt to play up to three times this many frames per - * shot. In practice we will often only send much less than this. - */ -#define FRAMES 4096 +/** @brief Array of file descriptors for poll() */ +struct pollfd fds[NFDS]; -/** @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 Next free slot in @ref fds */ +int fdno; static time_t last_report; /* when we last reported */ static int paused; /* pause status */ -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 -/** @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 -/** @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 The current device state */ +enum device_states device_state; -/** @brief Frames to force-play +/** @brief The current device sample format * - * 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. + * Only meaningful if @ref device_state = @ref device_open or perhaps @ref + * device_error. For @ref FIXED_FORMAT backends, this should always match @c + * config->sample_format. */ -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 - * - * 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; - -/** @brief RTP packet sequence number */ -static uint16_t rtp_seq; - -/** @brief RTP SSRC */ -static uint32_t rtp_id; +ao_sample_format device_format; /** @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 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 Called before poll() - * - * Called before the call to poll(). Should call addfd() to update the FD - * array and stash the slot number somewhere safe. - */ - void (*beforepoll)(void); - - /** @brief Called after poll() - * @return 0 if we could play, non-0 if not - * - * Called after the call to poll(). Should arrange to play some audio if the - * output device is ready. - * - * The return value should be 0 if the device was ready to play, or nonzero - * if it was not. - */ - int (*afterpoll)(void); -}; +int idled; /** @brief Selected backend */ static const struct speaker_backend *backend; @@ -387,8 +207,8 @@ static void acquire(struct track *t, int fd) { } /** @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) { +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 @@ -540,14 +360,15 @@ static int fill(struct track *t) { */ static void idle(void) { D(("idle")); - if(backend->deactivate) + if(backend->deactivate) backend->deactivate(); + else + device_state = device_closed; idled = 1; - ready = 0; } /** @brief Abandon the current track */ -static void abandon(void) { +void abandon(void) { struct speaker_message sm; D(("abandon")); @@ -558,49 +379,30 @@ static void abandon(void) { removetrack(playing->id); destroy(playing); playing = 0; - forceplay = 0; } -#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; - unsigned u; - - return; /* too verbose */ - if(hwparams) { - /* TODO */ - } - if(swparams) { - snd_pcm_sw_params_get_silence_size(swparams, &f); - info("sw silence_size=%lu", (unsigned long)f); - snd_pcm_sw_params_get_silence_threshold(swparams, &f); - info("sw silence_threshold=%lu", (unsigned long)f); - snd_pcm_sw_params_get_sleep_min(swparams, &u); - info("sw sleep_min=%lu", (unsigned long)u); - snd_pcm_sw_params_get_start_threshold(swparams, &f); - info("sw start_threshold=%lu", (unsigned long)f); - snd_pcm_sw_params_get_stop_threshold(swparams, &f); - info("sw stop_threshold=%lu", (unsigned long)f); - snd_pcm_sw_params_get_xfer_align(swparams, &f); - info("sw xfer_align=%lu", (unsigned long)f); - } -} -#endif - /** @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) { +static void 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; + return; + } + if(backend->flags & FIXED_FORMAT) + device_format = config->sample_format; + if(backend->activate) { + backend->activate(); + } else { + assert(backend->flags & FIXED_FORMAT); + /* ...otherwise device_format not set */ + device_state = device_open; } - return backend->activate(); + if(device_state == device_open) + device_bpf = bytes_per_frame(&device_format); } /** @brief Check whether the current track has finished @@ -618,52 +420,38 @@ static void maybe_finished(void) { abandon(); } -/** @brief Start the subprocess for @ref BACKEND_COMMAND */ -static void fork_cmd(void) { - pid_t cmdpid; - int pfd[2]; - if(cmdfd != -1) close(cmdfd); - xpipe(pfd); - cmdpid = xfork(); - if(!cmdpid) { - signal(SIGPIPE, SIG_DFL); - xdup2(pfd[0], 0); - close(pfd[0]); - close(pfd[1]); - execl("/bin/sh", "sh", "-c", config->speaker_command, (char *)0); - fatal(errno, "error execing /bin/sh"); - } - close(pfd[0]); - cmdfd = pfd[1]; - D(("forked cmd %d, fd = %d", cmdpid, cmdfd)); -} - -/** @brief Play up to @p frames frames of audio */ +/** @brief Play up to @p frames frames of audio + * + * It is always safe to call this function. + * - If @ref playing is 0 then it will just return + * - If @ref paused is non-0 then it will just return + * - If @ref device_state != @ref device_open then it will call activate() and + * return if it it fails. + * - If there is not enough audio to play then it play what is available. + * + * If there are not enough frames to play then whatever is available is played + * instead. It is up to mainloop() to ensure that play() is not called when + * unreasonably only an small amounts of data is available to play. + */ static void play(size_t frames) { size_t avail_frames, avail_bytes, written_frames; ssize_t written_bytes; - /* Make sure the output device is activated */ - if(activate()) { - if(playing) - forceplay = frames; - else - forceplay = 0; /* Must have called abandon() */ + /* Make sure there's a track to play and it is not pasued */ + if(!playing || paused) return; + /* Make sure the output device is open and has the right sample format */ + if(device_state != device_open + || !formats_equal(&device_format, &playing->format)) { + activate(); + if(device_state != device_open) + return; } - D(("play: play %zu/%zu%s %dHz %db %dc", frames, playing->used / bpf, + D(("play: play %zu/%zu%s %dHz %db %dc", frames, playing->used / device_bpf, playing->eof ? " EOF" : "", playing->format.rate, playing->format.bits, playing->format.channels)); - /* If we haven't got enough bytes yet wait until we have. Exception: when - * we are at eof. */ - if(playing->used < frames * bpf && !playing->eof) { - forceplay = frames; - return; - } - /* We have got enough data so don't force play again */ - 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 */ @@ -671,7 +459,7 @@ static void play(size_t frames) { else /* The ring buffer is not wrapped, can play the lot */ avail_bytes = playing->used; - avail_frames = avail_bytes / bpf; + avail_frames = avail_bytes / device_bpf; /* Only play up to the requested amount */ if(avail_frames > frames) avail_frames = frames; @@ -679,7 +467,7 @@ static void play(size_t frames) { return; /* Play it, Sam */ written_frames = backend->play(avail_frames); - written_bytes = written_frames * bpf; + written_bytes = written_frames * device_bpf; /* written_bytes and written_frames had better both be set and correct by * this point */ playing->start += written_bytes; @@ -690,6 +478,7 @@ static void play(size_t frames) { if(!playing->used || playing->start == playing->size) playing->start = 0; frames -= written_frames; + return; } /* Notify the server what we're up to. */ @@ -716,7 +505,7 @@ static void reap(int __attribute__((unused)) sig) { signal(SIGCHLD, reap); } -static int addfd(int fd, int events) { +int addfd(int fd, int events) { if(fdno < NFDS) { fds[fdno].fd = fd; fds[fdno].events = events; @@ -725,594 +514,62 @@ static int addfd(int fd, int events) { return -1; } -#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) { - 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 */ -static void command_init(void) { - info("selected command backend"); - fork_cmd(); -} - -/** @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) { - 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, - PF_INET, - SOCK_DGRAM, - IPPROTO_UDP, - 0, - 0, - 0, - 0 - }; - static const struct addrinfo prefbind = { - AI_PASSIVE, - PF_INET, - SOCK_DGRAM, - IPPROTO_UDP, - 0, - 0, - 0, - 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) { - 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); - } - } - } - 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 */ -static const struct speaker_backend backends[] = { +static const struct speaker_backend *backends[] = { #if API_ALSA - { - BACKEND_ALSA, - 0, - alsa_init, - alsa_activate, - alsa_play, - alsa_deactivate, - alsa_beforepoll, - alsa_afterpoll - }, + &alsa_backend, #endif - { - BACKEND_COMMAND, - FIXED_FORMAT, - command_init, - generic_activate, - command_play, - 0, /* deactivate */ - command_beforepoll, - command_afterpoll - }, - { - BACKEND_NETWORK, - FIXED_FORMAT, - network_init, - generic_activate, - network_play, - 0, /* deactivate */ - network_beforepoll, - network_afterpoll - }, - { -1, 0, 0, 0, 0, 0, 0, 0 } /* end of list */ + &command_backend, + &network_backend, + 0 }; -/** @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. +/** @brief Return nonzero if we want to play some audio * - * 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. + * We want to play audio if there is a current track; and it is not paused; and + * there are at least @ref FRAMES frames of audio to play, or we are in sight + * of the end of the current track. */ +static int playable(void) { + return playing + && !paused + && (playing->used >= FRAMES || playing->eof); +} + +/** @brief Main event loop */ static void mainloop(void) { struct track *t; struct speaker_message sm; - int n, fd, stdin_slot, poke, timeout; + int n, fd, stdin_slot, timeout; while(getppid() != 1) { fdno = 0; + /* By default we will wait up to a second before thinking about current + * state. */ + timeout = 1000; /* Always ready for commands from the main server. */ stdin_slot = addfd(0, POLLIN); /* Try to read sample data for the currently playing track if there is * buffer space. */ - if(playing && !playing->eof && playing->used < playing->size) { + if(playing && !playing->eof && playing->used < playing->size) playing->slot = addfd(playing->fd, POLLIN); - } else if(playing) + else if(playing) 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; - /* We'll break the poll as soon as the underlying sound device is ready for - * more data */ - if(ready && !forceplay) - backend->beforepoll(); + if(playable()) { + /* We want to play some audio. If the device is closed then we attempt + * to open it. */ + if(device_state == device_closed) + activate(); + /* If the device is (now) open then we will wait up until it is ready for + * more. If something went wrong then we should have device_error + * instead, but the post-poll code will cope even if it's + * device_closed. */ + if(device_state == device_open) + backend->beforepoll(); + } /* If any other tracks don't have a full buffer, try to read sample data - * from them. */ + * from them. We do this last of all, so that if we run out of slots, + * nothing important can't be monitored. */ for(t = tracks; t; t = t->next) if(t != playing) { if(!t->eof && t->used < t->size) { @@ -1327,16 +584,26 @@ static void mainloop(void) { fatal(errno, "error calling poll"); } /* Play some sound before doing anything else */ - poke = backend->afterpoll(); - if(poke) { - /* Some attempt to play must have failed */ - if(playing && !paused) - play(forceplay); - else - forceplay = 0; /* just in case */ + if(playable()) { + /* We want to play some audio */ + if(device_state == device_open) { + if(backend->ready()) + play(3 * FRAMES); + } else { + /* We must be in _closed or _error, and it should be the latter, but we + * cope with either. + * + * We most likely timed out, so now is a good time to retry. play() + * knows to re-activate the device if necessary. + */ + play(3 * FRAMES); + } } /* Perhaps we have a command to process */ if(fds[stdin_slot].revents & POLLIN) { + /* There might (in theory) be several commands queued up, but in general + * this won't be the case, so we don't bother looping around to pick them + * all up. */ n = speaker_recv(0, &sm, &fd); if(n > 0) switch(sm.type) { @@ -1352,7 +619,10 @@ static void mainloop(void) { t = findtrack(sm.id, 1); if(fd != -1) acquire(t, fd); playing = t; - play(bufsize); + /* We attempt to play straight away rather than going round the loop. + * play() is clever enough to perform any activation that is + * required. */ + play(3 * FRAMES); report(); break; case SM_PAUSE: @@ -1364,8 +634,9 @@ static void mainloop(void) { D(("SM_RESUME")); if(paused) { paused = 0; + /* As for SM_PLAY we attempt to play straight away. */ if(playing) - play(bufsize); + play(3 * FRAMES); } report(); break; @@ -1397,14 +668,11 @@ static void mainloop(void) { for(t = tracks; t; t = t->next) if(t->slot != -1 && (fds[t->slot].revents & (POLLIN | POLLHUP))) fill(t); - /* We might be able to play now */ - if(ready && forceplay && playing && !paused) - play(forceplay); /* Maybe we finished playing a track somewhere in the above */ maybe_finished(); /* If we don't need the sound device for now then close it for the benefit * of anyone else who wants it. */ - if((!playing || paused) && ready) + if((!playing || paused) && device_state == device_open) idle(); /* If we've not reported out state for a second do so now. */ if(time(0) > last_report) @@ -1445,12 +713,12 @@ int main(int argc, char **argv) { /* 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) + for(n = 0; backends[n]; ++n) + if(backends[n]->backend == config->speaker_backend) break; - if(backends[n].backend == -1) + if(!backends[n]) fatal(0, "unsupported backend %d", config->speaker_backend); - backend = &backends[n]; + backend = backends[n]; /* backend-specific initialization */ backend->init(); mainloop();