Re: [PATCH/RFC] SPI: add DMAUNSAFE analog to David Brownell's core

[David Brownell <david-b@pacbell.net>]

On Wednesday 14 December 2005 5:50 am, Vitaly Wool wrote:
> >>It's way better to just insist that all I/O buffers (in all
> >>generic APIs) be DMA-safe.  AFAICT that's a pretty standard
> >>rule everywhere in Linux.
> >
> >I agree.  
> 
> Well, why then David doesn't insist on that in his own code?
> His synchronous transfer functions

You seem to be referring to one non-generic function that's
documented as OK to pass DMA-unsafe buffers to.  There are
several other synchronous transfer calls that don't make
such guarantees.


> are allocating transfer buffers on  
> stack which is not DMA-safe.

I think the very first version did that, but nothing since
has taken that shortcut.  (Several months now.)  It uses
a buffer that's allocated on the heap.


> Then he starts messing with allocate-or-use-preallocated stuff etc. etc.
> Why isn't he just kmalloc'ing/kfree'ing buffers each time these 
> functions are called 

So that the typical case, with little SPI contention, doesn't
hit the heap?  That's sure what I thought ... though I can't speak
for what other people may think I thought.  You were the one that
wanted to optimize the atypical case to remove a blocking path!


> (as he proposes for upper layer drivers to do)? 

No I didn't.  I actually said that the upper layer drivers should
just not use DMA-unsafe areas for I/O buffers in the first place.
Places like stacks or static data.  Doing that means there's never
a need for a new kmalloc buffer, unless maybe you're marshaling
things into a scratch buffer.


> The thing is that only controller driver is aware whether DMA is needed 
> or not, so it's controller driver that should work it out.

Given the policy that all code avoids DMA-unsafe areas for I/O buffers,
the issue has already been worked out globally.

- Dave