Re: [PATCH] CMA: Don't return a valid cma for non-cma dev

[Feng Tang <feng.tang@intel.com>]

On Wed, Aug 05, 2015 at 01:15:50PM +0200, Michal Nazarewicz wrote:
> > On Wed, Aug 05, 2015 at 12:28:03PM +0200, Michal Nazarewicz wrote:
> >> If you need several CMA areas to allocate from, create multiple struct
> >> devices.
> 
> On Wed, Aug 05 2015, Feng Tang wrote:
> > I've made a quick patch, which works ok on our multiple cma heap cases.
> 
> > ---
> > diff --git a/drivers/staging/android/ion/ion_cma_heap.c b/drivers/staging/android/ion/ion_cma_heap.c
> > index f4211f1..ee9c5d1 100644
> > --- a/drivers/staging/android/ion/ion_cma_heap.c
> > +++ b/drivers/staging/android/ion/ion_cma_heap.c
> > @@ -29,6 +29,7 @@
> >  struct ion_cma_heap {
> >  	struct ion_heap heap;
> >  	struct device *dev;
> > +	struct device default_dma_dev;
> 
> I’m unfamiliar with ION code so cannot comment in great detail, butwhy
> do you need dev and default_dma_dev fields?  Just make dev a non-pointer
> and use that.

Good point. I've thought about keeping dev to be back compatible
with current code, and only use default_dma_dev when no "dev" is
passed in platform data. But from your other comments, it may
be not necessary to keep the "dev" 

> 
> >  };
> >  
> >  #define to_cma_heap(x) container_of(x, struct ion_cma_heap, heap)
> > @@ -180,9 +181,14 @@ struct ion_heap *ion_cma_heap_create(struct ion_platform_heap *data)
> >  		return ERR_PTR(-ENOMEM);
> >  
> >  	cma_heap->heap.ops = &ion_cma_ops;
> > -	/* get device from private heaps data, later it will be
> > -	 * used to make the link with reserved CMA memory */
> > -	cma_heap->dev = data->priv;
> > +
> > +	cma_heap->dev = &cma_heap->default_dma_dev;
> > +	cma_heap->dev->coherent_dma_mask = DMA_BIT_MASK(32);
> > +	cma_heap->dev->dma_mask = &dev->coherent_dma_mask;
> > +
> > +	/* data->priv contains a pointer to struct cma */
> > +	dev_set_cma_area(cma_heap->dev, data->priv);
> 
> In the previous code, data->priv seemed to be struct device*, but in
> this code it is used as struct cma*.

As we are going to use per cma-heap's own "default_dma_dev",
the "data->priv" should be a pointer to "struct cma*", which
makes more sense, as when a cma heap is created, we'd better
provide a "struct cma *" to tell it where to request cma buffer.

> 
> > +
> >  	cma_heap->heap.type = ION_HEAP_TYPE_DMA;
> >  	return &cma_heap->heap;
> >  }
> 
> But yeah, in general, from CMA’s point of view, this looks good.

Thanks! Will try to clean the code and add more comments, then
send it out for review.

- Feng