On Sun, 2010-06-27 at 20:01 +0200, Jaroslav Kysela wrote:
On Sun, 27 Jun 2010, David Dillow wrote:
On Sun, 2010-06-27 at 12:29 +0900, Jassi Brar wrote:
On Sun, Jun 27, 2010 at 7:13 AM, David Dillow dave@thedillows.org wrote:
This is solved by using runtime->twake as the number of samples needed for a wakeup in addition to selecting the proper wait queue to wake in snd_pcm_update_state(). This requires twake to be non-zero when used by snd_pcm_lib_{read,write}1() even if avail_min is zero.
http://mailman.alsa-project.org/pipermail/alsa-devel/2010-June/028786.html
Hmm, yes, I should have search the archives a bit. I originally tuned out the thread you listed as a request for parameter help, and missed your original postings.
I think my patch is pretty close to preserving the existing semantics as it doesn't change poll() at all, but I do see a case where the user could get a read/write back prior to avail_min samples being ready. I think that's fixable -- if the user is requesting a read/write of less than avail_min samples, then we have to wait for avail_min regardless.
Takashi, is your concern about semantics the proper honoring of avail_min in all cases, or preserving the current behavior of waiting for two periods when avail_min is set to the size of one period?
I think that the avail_min semantics is quite clear. The problem is caused by the hw transfer acknowledge interrupt jitter. It seems that the only good solution is to postpone the wake_up() call to time when the avail_min condition becomes true - using an extra timing source (system timer for example).
I think the avail_min is quite clear for poll() usage -- in that case, we get the expected behavior for wake ups. I believe the semantics are currently less well defined for read/write -- is the guarantee that at the end of the write call, there will be room for avail_min samples in the buffer, or is it guaranteed that there will be space for (avail_min - (size % avail_min)) samples, where size is the number of samples in the read/write call?
The current code gives no guarantee either way if the size of the call fits in the buffer. And there are cases where it doesn't give a guarantee that another write of avail_min will fit.
The only documentation a quick Google search turns up is the ALSA HOWTO, which only discusses poll().
Another possibility is to keep thing asis and keep to applications to handle this situation - use a different timing source than interrupts from soundcard for scheduling of I/O operations. But it's right that most of simple applications and use cases expect that I/O transfers will work even with 2 periods.
Indeed, that is how I noticed this issue -- capture running with 2 periods was causing unexpected overruns. As a user, I expected a read of avail_min (== period_size) samples to return when the period was up, not after a second period had elapsed.
Things can be easily changed so that case works as expected, the hard part is clarifying/defining the semantics for other cases, and ensuring those guarantees are met.
Perhaps documentation that recommends setting avail_min to 1 when using read/write to avoid this issue would be helpful, in lieu of changing the ALSA code or giving concrete guarantees.
Dave