Re: [PATCH] [media] dib0700: get rid of on-stack dma buffers

[Florian Mickler <florian@mickler.org>]

On Sun, 6 Mar 2011 16:06:38 +0100
Oliver Neukum <oliver@neukum.org> wrote:

> Am Sonntag, 6. März 2011, 15:38:05 schrieb Florian Mickler:
> > On Sun, 6 Mar 2011 13:06:09 +0100
> > Oliver Neukum <oliver@neukum.org> wrote:
> > 
> > > Am Sonntag, 6. März 2011, 12:16:52 schrieb Florian Mickler:
> 
> > > > Please take a look at it, as I do not do that much kernel hacking
> > > > and don't wanna brake anybodys computer... :)
> > > > 
> > > > From my point of view this should _not_ go to stable even though it would
> > > > be applicable. But if someone feels strongly about that and can
> > > > take responsibility for that change...
> > > 
> > > The patch looks good and is needed in stable.
> > > It could be improved by using a buffer allocated once in the places
> > > you hold a mutex anyway.
> > > 
> > > 	Regards
> > > 		Oliver
> > 
> > Ok, I now put a buffer member in the priv dib0700_state which gets
> > allocated on the heap. 
> 
> This however is wrong. Just like DMA on the stack this breaks
> coherency rules. You may do DMA to the heap in the sense that
> you can do DMA to buffers allocated on the heap, but you cannot
> do DMA to a part of another structure allocated on the heap.
> You need a separate kmalloc for each buffer.
> You can reuse the buffer with proper locking, but you must allocate
> it seperately once.
> 
> 	Regards
> 		Oliver

Hm.. allocating the buffer
in the probe routine and deallocating it in the usb_driver disconnect
callback should work?

How come that it must be a seperate kmalloc buffer? Is it some aligning
that kmalloc garantees? 

Regards,
Flo