[linux-dvb] Re: More V4 Demux API Q's

[Johannes Stezenbach <js@convergence.de>]

Rob.McConnell@Zarlink.Com wrote:
> Johannes Stezenbach wrote:
> > I think usually there is one DMA buffer per section filter, whether
> > they are on the same PID, same fd or not. (The av7110 driver uses
> > software section filtering now, but usually one would use hardware
> > section filters if they are available.) The common PID filter for
> > multiple section filters on the same PID is handled within the
> > driver and not visible in the API. The V3 implementation has
> > a layer of additional buffers for the "kernel demux API" thing,
> > but I try to get rid of that in V4. But if one would allow
> > the output of multiple filters on the same fd, one would need an
> > additional output ringbuffer, plus copying.
> 
> OK, so the AV7110 functions differently to the h/w that I am familiar with.
> In fact, on this h/w there isn't one DMA buffer per section filter but
> rather the output of all section filters attached to a single PID feed get
> DMAed to a single h/w buffer.  In this case, the s/w would have to do the
> additional work of separating the tables by their ID and copying the data
> to separate circular kernel buffers each associated with a process on a
> different fd.  There is certainly more s/w overhead in the driver to
> implement this, rather than allowing the complete section tables to be
> allowed up to user space and the middleware to perform the filtering on the
> single fd.

Hm, I checked the docs for some hw and it seems that you are damn right!
Thanks for pointing this out.

It seems that if we want to match the API up with the hw caps we need
to have multiple section filters on one fd, but only if they are
on the same PID. We would also have to disallow multiple filters
for the same PID on different fds.

But I know that this would be a pita for desktop Linux users, it's
just too inconvenient if only one application can read PAT/PMT at
the same time...

(thinking...)

> I think we shouldn't be too blinded with h/w that we are familiar with on
> cards, but rather look at what is out there on real STBs as well.  Having
> both methodologies would allow one to have optimum performance wrt the
> actual h/w (less time spent in kernel mode doing copies/etc.) and would
> also provide a flexible interface to the mad user-app folk out there ; )

Actually I don't care much about the av7110, but I need to support it, too.


Johannes


-- 
Info:
To unsubscribe send a mail to ecartis@linuxtv.org with "unsubscribe linux-dvb" as subject.