On Wed, 08 Jul 2020 12:18:48 +0200, Mark Hills wrote: > > Distorted audio appears occasionally, affecting either playback or > capture and requiring the affected substream to be closed by all > applications and re-opened. > > The best way I have found to reproduce the bug is to use dmix in > combination with Chromium, which opens the audio device multiple times > in threads. Anecdotally, the problems appear to have increased with > faster CPUs. I ruled out 32-bit counter wrapping; it often happens > much earlier. > > Since applying this patch I have not had problems, where previously > they would occur several times a day. > > The patch targets the following issues: > > * Check for progress using the counter from the hardware, not after it > has been truncated to the buffer. > > This is a clean way to address a possible bug where if a whole > ringbuffer advances between interrupts, it goes unnoticed. > > * Move last_period state from chip to pipe > > This more logically belongs as part of pipe, and code is reasier to > read if it is "counter position last time a period elapsed". > > Now the code has no references to period count. A period is just > when the regular counter crosses a threshold. This increases > readability and reduces scope for bugs. > > * Treat period notification and buffer advance independently: > > This helps to clarify what is the responsibility of the interrupt > handler, and what is pcm_pointer(). > > Removing shared state between these operations means race conditions > are fixed without introducing locks. Synchronisation is only around > the read of pipe->dma_counter. There may be cache line contention > around "struct audiopipe" but I did not have cause to profile this. > > Pay attention to be robust where dma_counter wrapping is not a > multiple of period_size or buffer_size. > > This is a revised patch based on feedback from Takashi and Giuliano. > > Signed-off-by: Mark Hills <mark@xxxxxxxx> Applied now. Thanks. Takashi
