Re: [PATCH] audio/dsound: fix invalid parameters error

[Zoltán Kővágó]

On 2020-01-18 07:30, Philippe Mathieu-Daudé wrote:
> On 1/17/20 7:26 PM, KJ Liew wrote:
> > QEMU Windows has broken dsound backend since the rewrite of audio API in
> > version 4.2.0. Both playback and capture buffers failed to lock with
> > invalid parameters error.
>
> Fixes: 7fa9754ac88 (dsoundaudio: port to the new audio backend api)

Hmm, I see the old code specified those parameters, but MSDN reads:

If the application passes NULL for the ppvAudioPtr2 and pdwAudioBytes2 
parameters, the lock extends no further than the end of the buffer and 
does not wrap.

Looks like this means that if the lock doesn't fit in the buffer it 
fails instead of truncating it.  I'm sure I tested the code under wine, 
and probably in a win8.1 vm too, and it worked there, maybe it's 
dependent on the windows version or sound driver?

> Cc'ing Zoltán who wrote 7fa9754ac88, and Gerd (the maintainer of this 
> file):
>
>    $ ./scripts/get_maintainer.pl -f audio/dsoundaudio.c
>    Gerd Hoffmann <address@hidden> (maintainer:Audio)
>
> > --- ../orig/qemu-4.2.0/audio/dsoundaudio.c    2019-12-12 
> > 10:20:47.000000000 -0800
> > +++ ../qemu-4.2.0/audio/dsoundaudio.c    2020-01-17 08:05:46.783966900 
> > -0800
> > @@ -53,6 +53,7 @@
> >   typedef struct {
> >       HWVoiceOut hw;
> >       LPDIRECTSOUNDBUFFER dsound_buffer;
> > +    void *last_buf;
> >       dsound *s;
> >   } DSoundVoiceOut;
> > @@ -414,10 +415,10 @@
> >       DSoundVoiceOut *ds = (DSoundVoiceOut *) hw;
> >       LPDIRECTSOUNDBUFFER dsb = ds->dsound_buffer;
> >       HRESULT hr;
> > -    DWORD ppos, act_size;
> > +    DWORD ppos, act_size, last_size;
> >       size_t req_size;
> >       int err;
> > -    void *ret;
> > +    void *ret, *last_ret;
> >       hr = IDirectSoundBuffer_GetCurrentPosition(dsb, &ppos, NULL);
> >       if (FAILED(hr)) {
> > @@ -426,17 +427,24 @@
> >           return NULL;
> >       }
> > +    if (ppos == hw->pos_emul) {
> > +        *size = 0;
> > +        return ds->last_buf;
> > +    }
> > +
> >       req_size = audio_ring_dist(ppos, hw->pos_emul, hw->size_emul);
> >       req_size = MIN(req_size, hw->size_emul - hw->pos_emul);
> > -    err = dsound_lock_out(dsb, &hw->info, hw->pos_emul, req_size, 
> > &ret, NULL,
> > -                          &act_size, NULL, false, ds->s);
> > +    err = dsound_lock_out(dsb, &hw->info, hw->pos_emul, req_size, 
> > &ret, &last_ret,
> > +                          &act_size, &last_size, false, ds->s);
> >       if (err) {
> >           dolog("Failed to lock buffer\n");
> >           *size = 0;
> >           return NULL;
> >       }
> > +    ds->last_buf = g_realloc(ds->last_buf, act_size);
> > +    memcpy(ds->last_buf, ret, act_size);
> >       *size = act_size;
> >       return ret;
> >   }

I don't really understand what's happening here, why do you need that 
memory allocation and memcpy?  This function should return a buffer 
where the caller will write data, that *size = 0; when returning 
ds->last_buf also looks incorrect to me (the calling function won't 
write anything into it).

I'm attaching a patch with a probably better (and totally untested) way 
to do this (if someone can tell me how to copy-paste a patch into 
thunderbird without it messing up long lines, please tell me).


> > @@ -445,6 +453,8 @@
> >   {
> >       DSoundVoiceOut *ds = (DSoundVoiceOut *) hw;
> >       LPDIRECTSOUNDBUFFER dsb = ds->dsound_buffer;
> > +    if (len == 0)
> > +        return 0;
> >       int err = dsound_unlock_out(dsb, buf, NULL, len, 0);
> >       if (err) {

Msdn says "The second pointer is needed even if nothing was written to 
the second pointer." so that NULL doesn't look okay.

> > @@ -508,10 +518,10 @@
> >       DSoundVoiceIn *ds = (DSoundVoiceIn *) hw;
> >       LPDIRECTSOUNDCAPTUREBUFFER dscb = ds->dsound_capture_buffer;
> >       HRESULT hr;
> > -    DWORD cpos, act_size;
> > +    DWORD cpos, act_size, last_size;
> >       size_t req_size;
> >       int err;
> > -    void *ret;
> > +    void *ret, *last_ret;
> >       hr = IDirectSoundCaptureBuffer_GetCurrentPosition(dscb, &cpos, 
> > NULL);
> >       if (FAILED(hr)) {
> > @@ -520,11 +530,16 @@
> >           return NULL;
> >       }
> > +    if (cpos == hw->pos_emul) {
> > +        *size = 0;
> > +        return NULL;
> > +    }
> > +
> >       req_size = audio_ring_dist(cpos, hw->pos_emul, hw->size_emul);
> >       req_size = MIN(req_size, hw->size_emul - hw->pos_emul);
> > -    err = dsound_lock_in(dscb, &hw->info, hw->pos_emul, req_size, 
> > &ret, NULL,
> > -                         &act_size, NULL, false, ds->s);
> > +    err = dsound_lock_in(dscb, &hw->info, hw->pos_emul, req_size, 
> > &ret, &last_ret,
> > +                         &act_size, &last_size, false, ds->s);
> >       if (err) {
> >           dolog("Failed to lock buffer\n");
> >           *size = 0;

You're completely ignoring last_ret and last_size here.  Don't you lose 
samples here?  I think it's possible to do something like I posted above 
with output here.

Regards,
Zoltan

dsound.patch
Description: Text Data