Re: speeding up ugen by an order of magnitude.

Julian Elischer wrote:
> 
> On Wed, 7 Jul 2004, Scott Long wrote:
> 
> 
>>On Wed, 7 Jul 2004, Poul-Henning Kamp wrote:
>>
>>>In message <200407072129.15095.mycroft_at_netbsd.org>, "Charles M. Hannum" writes:
>>>
>>>>On Wednesday 07 July 2004 20:46, Charles M. Hannum wrote:
>>>>
>>>>>1) You'll need to add an interface for assigning pipes for read and write,
>>>>>since there may be more than just bulk pipes (and may be more than one bulk
>>>>>pipe in each direction), and we only have have device node to work with.
>>>>
>>>>Seems I misspoke there.  That part looks fine.
>>>>
>>>>I think you're going to be screwed by the buffer alignment, though.  This also
>>>>causes some issues with umass if you're not using a bounce buffer.
>>>
>>>We've already had that issue with ATA for the b�rked Geode controller:
>>>physio does nothing for alignment and relies on userland doing something
>>>sensible.
>>>
>>>I think this is pretty reasonable for the kind of hardware-near
>>>work that physio is usually employed in (including if we use it for
>>>ugen).
>>>
>>>Obviously, if the alignment is not OK, EINVAL should be returned,
>>>and that means that the driver should explicitly check the alignment.
>>
>>Note that busdma can handle alignment now for loaded buffers by using
>>bounce pages.
> 
> 
> Since we have a relatively highly weighted priority on keeping this 
> code common with teh NetBSD version, how does that go for NetBSD?
> 
> 
> 
>>Scott
> 
> 
> 

I honestly don't know if NetBSD busdma honors alignment when loading 
segments.  However, I believe that our usb code has the busdma 
functionality completely broken out from the common code (in usb_mem.c,
IIRC), so it's of little consequence to portability.

I was going to mention that different semantics of bus_dmamap_load()
between FreeBSD and NetBSD might still cause problems here, but it looks
like our usb code still allocates static busdma buffers and copies the
packets in/out of these.  Wasn't this fixed a while back?

Scott