Re: [Qemu-devel] [4249] Improve audio api use in WM8750.

[Jan Kiszka <jan.kiszka@web.de>]

[-- Attachment #1: Type: text/plain, Size: 2855 bytes --]

Jan Kiszka wrote:
> andrzej zaborowski wrote:
>> On 24/04/2008, Jan Kiszka <jan.kiszka@web.de> wrote:
>>> Hi Andrzej,
>>>
>>>  Andrzej Zaborowski wrote:
>>>  > Revision: 4249
>>>  ...
>>>  >  static void wm8750_audio_out_cb(void *opaque, int free_b)
>>>  >  {
>>>  >      struct wm8750_s *s = (struct wm8750_s *) opaque;
>>>  >
>>>  > -    s->req_out = free_b;
>>>  > -    s->data_req(s->opaque, free_b >> 2, s->req_in >> 2);
>>>  > -    wm8750_out_flush(s);
>>>  > +    if (s->idx_out >= free_b) {
>>>  > +        s->idx_out = free_b;
>>>  > +        s->req_out = 0;
>>>  > +        wm8750_out_flush(s);
>>>  > +    } else
>>>  > +        s->req_out = free_b - s->idx_out;
>>>  > +
>>>  > +    s->data_req(s->opaque, s->req_out >> 2, s->req_in >> 2);
>>>
>>>  Please make sure that the callback is always issued _before_ the flush
>>>  (keep in mind: it may increase the amount of data that has to be flushed
>>>  ASAP!). And this change also leaves the MusicPal broken.
>> The idea is to output free_b bytes immediately if we have that many in
>> the buffer (it could happen assuming that free_b value can change
> 
> Wrong ordering: If there has been a _relevant_ amount of data (not a few
> 10 bytes) hanging around in our internal buffer between the guest has
> filled it during the last callbacks and the host reports demand via the
> new callback, that data was missing in the host's audio buffer! That's
> what I saw (heard) with the old version (the current one even causes
> total silence).
> 
>> between callbacks).  I'm not sure how this can break something: if
>> *inside* the data_req() callback we receive enough bytes to fill the
>> the whole buffer then dac_dat() will call out_flush().
>>
>> Without that all buffering becomes useless because we flush every
>> sample we receive and we start to crawl.
> 
> Nothing is crawling here, just use a reasonable threshold for flushing
> _after_ the callback. Need to check, but maybe we can even wait the a
> full buffer.

Of course not the full buffer, but its half is fine as it generally
leaves enough time to the guest to refill the other half:

Index: hw/wm8750.c
===================================================================
--- hw/wm8750.c	(Revision 4250)
+++ hw/wm8750.c	(Arbeitskopie)
@@ -73,14 +73,10 @@ static void wm8750_audio_out_cb(void *op
 {
     struct wm8750_s *s = (struct wm8750_s *) opaque;

-    if (s->idx_out >= free_b) {
-        s->idx_out = free_b;
-        s->req_out = 0;
+    s->req_out = free_b;
+    s->data_req(s->opaque, free_b >> 2, s->req_in >> 2);
+    if (s->idx_out >= sizeof(s->data_out)/2)
         wm8750_out_flush(s);
-    } else
-        s->req_out = free_b - s->idx_out;
-
-    s->data_req(s->opaque, s->req_out >> 2, s->req_in >> 2);
 }

 struct wm_rate_s {


Jan


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 254 bytes --]