Add an ALSA output to the motown audio server.

services/media/audio/platform/linux/alsa_output.cc

+// Copyright 2015 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include <limits>
+#include <set>
+
+#include "mojo/services/media/common/cpp/local_time.h"
+#include "services/media/audio/platform/linux/alsa_output.h"
+
+namespace mojo {
+namespace media {
+namespace audio {
+
+static constexpr LocalDuration TARGET_LATENCY = local_time::from_msec(35);
+static constexpr LocalDuration LOW_BUF_THRESH = local_time::from_msec(30);
+static constexpr LocalDuration ERROR_RECOVERY_TIME = local_time::from_msec(300);
+static constexpr LocalDuration WAIT_FOR_ALSA_DELAY = local_time::from_usec(500);
+static const std::set<uint8_t> SUPPORTED_CHANNEL_COUNTS({ 1, 2 });
+static const std::set<uint32_t> SUPPORTED_SAMPLE_RATES({
+    48000, 32000, 24000, 16000, 8000, 4000,
+    44100, 22050, 11025,
+});
+
+static inline bool IsRecoverableAlsaError(int error_code) {
+  switch (error_code) {
+  case -EINTR:
+  case -EPIPE:
+  case -ESTRPIPE:
+    return true;
+  default:
+    return false;
+  }
+}
+
+AlsaOutput::AlsaOutput(AudioOutputManager* manager)
+  : StandardOutputBase(manager) {}
+
+AlsaOutput::~AlsaOutput() {
+  // We should have been cleaned up already, but in release builds, call cleanup
+  // anyway, just in case something got missed.
+  DCHECK(!alsa_device_);
+  Cleanup();
+}
+
+AudioOutputPtr AlsaOutput::New(AudioOutputManager* manager) {
+  return AudioOutputPtr(new AlsaOutput(manager));
+}
+
+MediaResult AlsaOutput::Configure(LpcmMediaTypeDetailsPtr config) {
+  if (!config) { return MediaResult::INVALID_ARGUMENT; }
+  if (output_format_) { return MediaResult::BAD_STATE; }
+
+  uint32_t bytes_per_sample;
+  switch (config->sample_format) {
+  case LpcmSampleFormat::UNSIGNED_8:
+    alsa_format_ = SND_PCM_FORMAT_U8;
+    silence_byte_ = 0x80;
+    bytes_per_sample = 1;
+    break;
+
+  case LpcmSampleFormat::SIGNED_16:
+    alsa_format_ = SND_PCM_FORMAT_S16;
+    silence_byte_ = 0x00;
+    bytes_per_sample = 2;
+    break;
+
+  case LpcmSampleFormat::SIGNED_24_IN_32:
+  default:
+    return MediaResult::UNSUPPORTED_CONFIG;
+  }
+
+  if (SUPPORTED_SAMPLE_RATES.find(config->frames_per_second) ==
+      SUPPORTED_SAMPLE_RATES.end()) {
+    return MediaResult::UNSUPPORTED_CONFIG;
+  }
+
+  if (SUPPORTED_CHANNEL_COUNTS.find(config->samples_per_frame) ==
+      SUPPORTED_CHANNEL_COUNTS.end()) {
+    return MediaResult::UNSUPPORTED_CONFIG;
+  }
+
+  // Compute the ratio between frames and local time ticks.
+  LinearTransform::Ratio sec_per_tick(LocalDuration::period::num,
+                                      LocalDuration::period::den);
+  LinearTransform::Ratio frames_per_sec(config->frames_per_second, 1);
+  bool is_precise = LinearTransform::Ratio::Compose(frames_per_sec,
+                                                    sec_per_tick,
+                                                    &frames_per_tick_);
+  DCHECK(is_precise);
+
+  // Figure out how many bytes there are per frame.
+  output_bytes_per_frame_ = bytes_per_sample * config->samples_per_frame;
+
+  // Success
+  output_format_ = config.Pass();
+  return MediaResult::OK;
+}
+
+MediaResult AlsaOutput::Init() {
+  static const char* kAlsaDevice = "default";
+
+  if (!output_format_) { return MediaResult::BAD_STATE; }
+  if (alsa_device_) { return MediaResult::BAD_STATE; }
+
+  snd_pcm_sframes_t res;
+  res = snd_pcm_open(&alsa_device_,
+                     kAlsaDevice,
+                     SND_PCM_STREAM_PLAYBACK,
+                     SND_PCM_NONBLOCK);
+  if (res != 0) {
+    LOG(ERROR) << "Failed to open ALSA device \"" << kAlsaDevice << "\".";
+    return MediaResult::INTERNAL_ERROR;
+  }
+
+  res = snd_pcm_set_params(alsa_device_,
+                           alsa_format_,
+                           SND_PCM_ACCESS_RW_INTERLEAVED,
+                           output_format_->samples_per_frame,
+                           output_format_->frames_per_second,
+                           0,   // do not allow ALSA resample
+                           local_time::to_usec<unsigned int>(TARGET_LATENCY));
+  if (res) {
+    LOG(ERROR) << "Failed to configure ALSA device \"" << kAlsaDevice << "\" "
+               << "(res = " << res << ")";
+    LOG(ERROR) << "Requested samples per frame: "
+               << output_format_->samples_per_frame;
+    LOG(ERROR) << "Requested frames per second: "
+               << output_format_->frames_per_second;
+    LOG(ERROR) << "Requested ALSA format      : " << alsa_format_;
+    Cleanup();
+    return MediaResult::INTERNAL_ERROR;
+  }
+
+  // Figure out how big our mixing buffer needs to be, then allocate it.
+  res = snd_pcm_avail_update(alsa_device_);
+  if (res <= 0) {
+    LOG(ERROR) << "[" << this << "] : "
+               << "Fatal error (" << res
+               << ") attempting to determine ALSA buffer size.";
+    Cleanup();
+    return MediaResult::INTERNAL_ERROR;
+  }
+
+  mix_buf_frames_ = res;
+  mix_buf_.reset(new uint8_t[mix_buf_frames_ * output_bytes_per_frame_]);
+
+  return MediaResult::OK;
+}
+
+void AlsaOutput::Cleanup() {
+  if (alsa_device_) {
+    snd_pcm_close(alsa_device_);
+    alsa_device_ = nullptr;
+  }
+
+  mix_buf_ = nullptr;
+  mix_buf_frames_ = 0;
+}
+
+bool AlsaOutput::StartMixJob(MixJob* job, const LocalTime& process_start) {
+  DCHECK(job);
+
+  // Are we not primed?  If so, fill a mix buffer with silence and send it to
+  // the alsa device.  Schedule a callback for a short time in the future so
+  // ALSA has a chance to start the output and we can take our best guess of the
+  // function which maps output frames to local time.
+  if (!primed_) {
+    HandleAsUnderflow();
+    return false;
+  }
+
+  // Figure out how many frames of audio we need to produce in order to top off
+  // the buffer.  If we are primed, but do not know the transformation between
+  // audio frames and local time ticks, do our best to figure it out in the
+  // process.
+  snd_pcm_sframes_t avail;
+  if (!local_to_output_known_) {
+    snd_pcm_sframes_t delay;
+
+    int res = snd_pcm_avail_delay(alsa_device_, &avail, &delay);
+    LocalTime now = LocalClock::now();
+
+    if (res < 0) {
+      HandleAlsaError(res);
+      return false;
+    }
+
+    DCHECK_GE(delay, 0);
+    int64_t now_ticks = now.time_since_epoch().count();
+    local_to_output_ = LinearTransform(now_ticks, frames_per_tick_, -delay);
+    local_to_output_known_ = true;
+    frames_sent_ = 0;
+    while (++local_to_output_gen_ == MixJob::INVALID_GENERATION) {}
+  } else {
+    avail = snd_pcm_avail_update(alsa_device_);
+    if (avail < 0) {
+      HandleAlsaError(avail);
+      return false;
+    }
+  }
+
+  // Compute the time that we think we will completely underflow, then back off
+  // from that by the low buffer threshold and use that to determine when we
+  // should mix again.
+  int64_t playout_time_ticks;
+  bool trans_ok = local_to_output_.DoReverseTransform(frames_sent_,
+                                                      &playout_time_ticks);
+  DCHECK(trans_ok);
+  LocalTime playout_time = LocalTime(LocalDuration(playout_time_ticks));
+  LocalTime low_buf_time = playout_time - LOW_BUF_THRESH;
+
+  if (process_start >= low_buf_time) {
+    // Because of the way that ALSA consumes data and updates its internal
+    // bookkeeping, it is possible that we are past our low buffer threshold,
+    // but ALSA still thinks that there is no room to write new frames.  If this
+    // is the case, just try again a short amount of time in the future.
+    DCHECK_GE(avail, 0);
+    if (!avail) {
+      SetNextSchedDelay(WAIT_FOR_ALSA_DELAY);
+      return false;
+    }
+
+    // Limit the amt that we queue to be no more than what ALSA will currently
+    // accept, or what it currently will take to fill us to our target latency.
+    //
+    // The playout target had better be ahead of the playout time, or we are
+    // almost certainly going to underflow.  If this happens, for whatever
+    // reason, just try to send a full buffer and deal with the underflow when
+    // ALSA notices it.
+    int64_t fill_amt;
+    LocalTime playout_target = LocalClock::now() + TARGET_LATENCY;
+    if (playout_target > playout_time) {
+      fill_amt = (playout_target - playout_time).count();
+    } else {
+      fill_amt = TARGET_LATENCY.count();
+    }
+
+    DCHECK_GE(fill_amt, 0);
+    DCHECK_LE(fill_amt, std::numeric_limits<int32_t>::max());
+    fill_amt *= frames_per_tick_.numerator;
+    fill_amt += frames_per_tick_.denominator - 1;
+    fill_amt /= frames_per_tick_.denominator;
+
+    job->buf_frames = (avail < fill_amt) ? avail : fill_amt;
+    if (job->buf_frames > mix_buf_frames_) {
+      job->buf_frames = mix_buf_frames_;
+    }
+
+    job->buf = mix_buf_.get();
+    job->start_pts_of = frames_sent_;
+    job->local_to_output = &local_to_output_;
+    job->local_to_output_gen = local_to_output_gen_;
+
+    // TODO(johngro): optimize this if we can.  The first buffer we mix can just
+    // put its samples directly into the output buffer, and does not need to
+    // accumulate and clip.  In theory, we only need to put silence in the
+    // places where our outputs are not going to already overwrite.
+    FillMixBufWithSilence(job->buf_frames);
+    return true;
+  }
+
+  // Wait until its time to mix some more data.
+  SetNextSchedTime(low_buf_time);
+  return false;
+}
+
+bool AlsaOutput::FinishMixJob(const MixJob& job) {
+  DCHECK(job.buf == mix_buf_.get());
+  DCHECK(job.buf_frames);
+
+  // We should always be able to write all of the data that we mixed.
+  snd_pcm_sframes_t res;
+  res = snd_pcm_writei(alsa_device_, job.buf, job.buf_frames);
+  if (res != job.buf_frames) {
+    HandleAlsaError(res);
+    return false;
+  }
+
+  frames_sent_ += res;
+  return true;
+}
+
+void AlsaOutput::FillMixBufWithSilence(uint32_t frames) {
+  DCHECK(mix_buf_);
+  DCHECK(frames <= mix_buf_frames_);
+
+  // TODO(johngro): someday, this may not be this simple.  Filling unsigned
+  // multibyte sample formats, or floating point formats, will require something
+  // more sophisticated than filling with a single byte pattern.
+  ::memset(mix_buf_.get(), silence_byte_, frames * output_bytes_per_frame_);
+}
+
+void AlsaOutput::HandleAsUnderflow() {
+  snd_pcm_sframes_t res;
+
+  // If we were already primed, then this is a legitimate underflow, not the
+  // startup case or recovery from some other error.
+  if (primed_) {
+    // TODO(johngro): come up with a way to properly throttle this.  Also, add a
+    // friendly name to the output so the log helps to identify which output
+    // underflowed.
+    LOG(WARNING) << "[" << this << "] : underflow";
+    res = snd_pcm_recover(alsa_device_, -EPIPE, true);
+    if (res < 0) {
+      HandleAsError(res);
+      return;
+    }
+  }
+
+  // TODO(johngro): We don't actually have to fill up the entire lead time with
+  // silence.  When we have better control of our thread priorities, prime this
+  // with the minimimum amt we can get away with and still be able to start
+  // mixing without underflowing.
+  FillMixBufWithSilence(mix_buf_frames_);
+  res = snd_pcm_writei(alsa_device_, mix_buf_.get(), mix_buf_frames_);
+
+  if (res < 0) {
+    HandleAsError(res);
+    return;
+  }
+
+  primed_ = true;
+  local_to_output_known_ = false;
+  SetNextSchedDelay(local_time::from_msec(1));
+}
+
+void AlsaOutput::HandleAsError(snd_pcm_sframes_t code) {
+  // TODO(johngro): Throttle this somehow.
+  LOG(WARNING) << "[" << this << "] : Attempting to recover from ALSA error "
+               << code;
+
+  if (IsRecoverableAlsaError(code)) {
+    code = snd_pcm_recover(alsa_device_, code, true);

[jeffbrown, 2015/11/10 20:13:25]

This could fail with EINTR too I think and the correct thing to do would be to
loop immediately.

[johngro, 2015/11/10 20:32:02]

Docs are fuzzy on this issue.  They basically say that if the device cannot be
recovered, that it will return the value that you passed for 'code' to the
recover call.  They do not seem to give any indication of why the system could
not be recovered (were we EINTR'ed again, or is there something else going on?)

The procedure for recovering from an error which cannot be snd_pcm_recovered
from is also unclear.  I suppose that the nuclear option is to close the device
and attempt to re-open it; I'd need to refactor some code to take that approach.
 For now, I'm just shutting down.

I would assume that the #1 reason we would see EINTR is because either someone
CTRL-C'ed us, or because someone SIGKILL'ed us, and its probably time to shut
down anyway.  For that reason, and because generally I don't want to end up with
any unbound operations on a thread like this (high priority, currently a member
of a pool) I would prefer to yield to the outer-engine.  This gives the system a
chance to re-evaluate the should-I-shutdown-now question, and also will throttle
the system so that it does not melt down.

I assume that at this point, all bets are off and that we have damaged
presentation already, so working extra hard to recover super fast (for a sound
architecture we are only casually supporting to begin with) is probably not
worth it.

I'll retry in this case, but with the same delay I use for recovered errors. 
I'll also leave a comment about needing to exercise this behavior if we plan on
officially supporting ALSA as a backend.

+    if (!code) {
+      primed_ = false;
+      local_to_output_known_ = false;
+      SetNextSchedDelay(ERROR_RECOVERY_TIME);
+    }
+  }
+
+  LOG(ERROR) << "[" << this << "] : Fatal ALSA error "
+             << code << ".  Shutting down";
+  ShutdownSelf();
+}
+
+void AlsaOutput::HandleAlsaError(snd_pcm_sframes_t code) {
+  // ALSA signals an underflow by returning -EPIPE from jobs.  If the error code
+  // is -EPIPE, treat this as an underflow and attempt to reprime the pipeline.
+  if (code == -EPIPE) {
+    HandleAsUnderflow();
+  } else {
+    HandleAsError(code);
+  }
+}
+
+}  // namespace audio
+}  // namespace media
+}  // namespace mojo
+