On Sat, Dec 11, 2004 at 09:41:31PM -0800, Kamal R. Prasad wrote:
> I looked at the paper by Jason Thorpe sometime back
> and I BELIEVE it says it is used for kernel<-->user 
> exchange of bulk data.

No, it's for data exchange between kernel and devices, using bus-master DMA
(of course the device has to support it).

> > 
> [...]
> I have written an mmap() interface for a usb device.
> Userland does an mmap() to read in a block of data
> without involving a copyout(). I am using vtophys()
> [on freebsd for the moment] to map kernel virtual
> address to a physical interface which is returned in
> *paddr.

OK, so you have 2 different things here:
1) provide a set of physical pages to the USB device, to which
   it can do bus-master DMA.
2) provide a mmap() interface, so that this set of physical pages can be mapped
  in userland by a process.


1) requires you to use bus_dma(9). The man page, and other drivers in sys/dev/
  should give you enouth informations on how to use it. 
  For this, you need to know how you will alloc the set of physical pages.
  My guess is that bus_dmamem_alloc() is the best way to go in this case.
2) depends on how you allocated the memory in 1). If the memory comes from
 bus_dmamem_alloc(), bus_dmamem_mmap() should do what you want. You have
 examples for this in audio drivers, such as pci/auich.c.


> I have been given to understand that vtophys() will
> not be available in future and I should use bus_dma in
> future.

Yes, for the good reason that vtophys() isn't portable, and not
all architectures have it.


-- 
Manuel Bouyer <bouyer@antioche.eu.org>
     NetBSD: 26 ans d'experience feront toujours la difference
--