[alsa-devel] [PATCH v2 09/10] ASoC: topology: Change stream formats to bitwise flag

[Takashi Iwai]

On Sat, 15 Aug 2015 15:49:42 +0200,
Lin, Mengdong wrote:
> 
> > -----Original Message-----
> > From: Takashi Iwai [mailto:tiwai at suse.de]
> > Sent: Saturday, August 15, 2015 3:38 PM
> 
> > On Fri, 14 Aug 2015 22:34:28 +0200,
> > Mark Brown wrote:
> > >
> > > On Mon, Aug 10, 2015 at 10:48:37PM +0800, mengdong.lin at intel.com
> > wrote:
> > >
> > > >  struct snd_soc_tplg_stream_caps {
> > > >  	__le32 size;		/* in bytes of this structure */
> > > >  	char name[SNDRV_CTL_ELEM_ID_NAME_MAXLEN];
> > > > -	__le64 formats[SND_SOC_TPLG_MAX_FORMATS];	/* supported
> > formats SNDRV_PCM_FMTBIT_* */
> > > > +	__le64 formats;	/* supported formats SNDRV_PCM_FORMAT_* */
> > >
> > > Argh, this is *another* ABI change which you've buried in the middle
> > > of a series you're sending right at the end of the release cycle (this
> > > was sent after -rc6, we may not even get a -rc7...).  Given how close
> > > we are to the release and the invasiveness of the changes in this
> > > series it's very lucky I even saw it before release, I'm reading this
> > > on my flight to Plumbers which means my bandwidth next week will be
> > limited.
> > 
> > Thanks for checking this.  Mengdong, if there is a change in uapi/*, this
> > means touching ABI.  So, this change isn't acceptable once when the kernel
> > is released.  At least, the backward compatibility *must* be kept no matter
> > which version is.
> 
> Thanks for your review, Mark and Takashi. 
> This series including the ABI change, is not for v4.2, but is an RFC to adding DAI/DAI links support in topology.
> I'm sorry that I should have explicitly said this in my comments.
> 
> Is there some topic branch which can accept topology patches for next kernel release?
> Can we change topology ABI in v4.3?

In general "yes, but only if it's backward-compatible".

> Now only Intel SKL driver is using topology code, so the affect should be limited.
> And I feel using bitwise flag is simpler in both user space and in kernel coding.

Yes, using FMTBIT_* is better, but this must not justify to break
ABI.  Either we disable this API/ABI for 4.2, or make the change
backward compatible.

> We can increase the topology ABI number when there is ABI change, and the topology driver always checks this version.

It's a big NO, sorry, this isn't acceptable.  Again, the kernel must
give the backward compatibility.  That is, the kernel must accept the
old dialect the user space talks if user space wants.  You cannot
force user space to communicate in a new ABI out of sudden.

So, yes, ABI version bump is needed if you change ABI.  But it means
that the kernel accepts all ABIs up to this version, not meaning
*only* this ABI version.

Hopefully this clarifies the situation.


thanks,

Takashi