ALSA: Ways to prevent underrun for speaker

I am playing a single channel audio in non-interleaved mode. I am getting underrun when I am writing audio data into speaker : ALSA lib pcm.c:7339:(snd_pcm_recover) underrun occurred

Here is how I write:

    printf("%d",snd_pcm_avail (spkhandle));  
    ret = snd_pcm_writen(spkhandle, pSpeakerBuf , framesIn18Millisec);
    if(ret < 0)
    {
        snd_pcm_recover(spkhandle, ret, 0);
    }

What are the different ways/parameter configurations to prevent ALSA under run ?

(I am using Linux 3.0, ARM )

Edit: Here is a buffer measurement using snd_pcm_avail() API

      snd_pcm_avail = 2304      << snd_pcm_writen call 1 success
      snd_pcm_avail = 2160      << snd_pcm_writen call 2 success
      snd_pcm_avail = 2016      << snd_pcm_writen call 3 success
      snd_pcm_writen error -32 Broken pipe  << snd_pcm_writen call 4 failure
      ALSA lib pcm.c:7339:(snd_pcm_recover) underrun occurred  << And displays this message

Here is the output that Marko requested for:

snd_output_t* out;
....
// Do alsa parameters init .... 
....
snd_output_stdio_attach(&out, stderr, 0);
snd_pcm_dump_sw_setup(spkhandle, out);

  tstamp_mode  : NONE
  period_step  : 1
  avail_min    : 144
  period_event : 0
  start_threshold  : 288
  stop_threshold   : 2304
  silence_threshold: 0
  silence_size : 0
  boundary     : 1207959552

Answer

I assume this code runs in a tight loop and is intended to block on snd_pcm_writen(). The sample-rate isn’t given; I assume 48kHz since the numbers all divide nicely.

What I think is going here is as follows:

• snd_pcm_write() doesn’t guarantee to write all frames provided (the return value is only ever checked for error conditions). Judging from the logging of snd_pcm_avail() it’s in fact consuming avail_min or 144 frames on each. This is 3ms of audio.
• Assuming that audio is not running at this point, after two writes, the number of frames in the buffer is equal to start_threshold – at 288 samples; audio output starts
• calls to printf() block, and I seem to remember that snd_pcm_avail() has to synchronise with the audio output hardware and might also block. Since you are now 6ms ahead of the playback, it’s entirely possible that the buffer is running dry during the time of the third call of snd_pcm_writen()

In summary, you shouldn’t be calling printf() at this point, and you probably need to compensate for fact that snd_pcm_writen() isn’t consuming all of the frames in pSpeakerBuf