At Sun, 21 Nov 2010 12:09:32 +0100, Andreas Mohr wrote:

ChangeLog:

. Fix PulseAudio "ALSA driver bug" issue (if we have two alternated areas within a 64k DMA buffer, then max period size should obviously be 32k only). Back references: http://pulseaudio.org/wiki/AlsaIssues http://fedoraproject.org/wiki/Features/GlitchFreeAudio . In stop timer function, need to supply ACK in the timer control byte. . Minor log output correction

When I did my first PA testing recently, the period size bug resulted in quite precisely observeable half-period-based playback distortion.

PA-based operation is quite a bit more underrun-prone (despite its zero-copy optimizations etc.) than raw ALSA with this rather spartan sound hardware implementation on my puny Athlon.

Note that even with this patch, azt3328 still doesn't work for both cases yet, PA tsched=0 and tsched (on tsched=0 it will playback tiny fragments of periods, leading to tiny stuttering sounds with some pauses in between, whereas with timer-scheduled operation playback works fine - minus some quite increased underrun trouble on PA vs. ALSA, that is).

Patch runtime-tested on 2.6.36-rc8 git (and run through checkpatch.pl). Since 2.6 git log doesn't show patch activity it should still apply easily.

I'd suggest handling this patch whichever way you want. If you already have a couple other -rc fixes, then perhaps simply include it, and I'm quite ambivalent about -stable handling as well.

Thanks!

Signed-off-by: Andreas Mohr andi@lisas.de

Thanks, I applied this now for the next 2.6.37 pull request.

Takashi

diff --git a/sound/pci/azt3328.c b/sound/pci/azt3328.c index 4679ed8..2f3cacb 100644 --- a/sound/pci/azt3328.c +++ b/sound/pci/azt3328.c @@ -1129,10 +1129,11 @@ snd_azf3328_codec_setdmaa(struct snd_azf3328 *chip,

count_areas = size/2;
addr_area2 = addr+count_areas;

• count_areas--; /* max. index */
  

  snd_azf3328_dbgcodec("setdma: buffers %08lx[%u] / %08lx[%u]\n", addr, count_areas, addr_area2, count_areas);

• count_areas--; /* max. index */
  

• /* build combined I/O buffer length word */ lengths = (count_areas << 16) | (count_areas); spin_lock_irqsave(&chip->reg_lock, flags);

@@ -1740,11 +1741,15 @@ static const struct snd_pcm_hardware snd_azf3328_hardware = .rate_max = AZF_FREQ_66200, .channels_min = 1, .channels_max = 2,

• .buffer_bytes_max = 65536,
• .period_bytes_min = 64,
• .period_bytes_max = 65536,
• .periods_min = 1,
• .periods_max = 1024,

• .buffer_bytes_max = (64*1024),
• .period_bytes_min = 1024,
• .period_bytes_max = (32*1024),
• /* We simply have two DMA areas (instead of a list of descriptors

•   such as other cards); I believe that this is a fixed hardware
  

•   attribute and there isn't much driver magic to be done to expand it.
  

•   Thus indicate that we have at least and at most 2 periods. */
  

• .periods_min = 2,
• .periods_max = 2, /* FIXME: maybe that card actually has a FIFO?
  • Hmm, it seems newer revisions do have one, but we still don't know
  • its size... */

@@ -1980,8 +1985,13 @@ snd_azf3328_timer_stop(struct snd_timer *timer) chip = snd_timer_chip(timer); spin_lock_irqsave(&chip->reg_lock, flags); /* disable timer countdown and interrupt */

• /* FIXME: should we write TIMER_IRQ_ACK here? */
• snd_azf3328_ctrl_outb(chip, IDX_IO_TIMER_VALUE + 3, 0);

• /* Hmm, should we write TIMER_IRQ_ACK here?

•   YES indeed, otherwise a rogue timer operation - which prompts
  

•   ALSA(?) to call repeated stop() in vain, but NOT start() -
  

•   will never end (value 0x03 is kept shown in control byte).
  

•   Simply manually poking 0x04 _once_ immediately successfully stops
  

•   the hardware/ALSA interrupt activity. */
  

• snd_azf3328_ctrl_outb(chip, IDX_IO_TIMER_VALUE + 3, 0x04); spin_unlock_irqrestore(&chip->reg_lock, flags); snd_azf3328_dbgcallleave(); return 0;