[PATCH 6/6] dma-mapping: introduce DMA range map, supplanting dma_pfn_offset

[Christoph Hellwig]

On Mon, Sep 14, 2020 at 05:01:47PM -0600, Mathieu Poirier wrote:

[700 lines of the fullquote deleted..]

> > +	for (r = map; r->size; r++)
> > +		num_ranges++;
> > +
> > +	new_map = kmemdup(map, array_size(num_ranges + 1, sizeof(*map)),
> > +			  GFP_KERNEL);
> > +	if (!new_map)
> > +		return -ENOMEM;
> > +	to->dma_range_map = new_map;
> > +	return 0;
> > +}
> > +
> 
> This patch seemed Ok to me but it broke the stm32 remoteproc implementation.  When
> I tested things out function dma_coerce_mask_and_cohenrent() returns -5 and the
> rest of the initialisation fails.  I isolated things to function dma_to_pfn()
> [2].  In the original implementation __bus_to_pfn() returns 0xfffff and
> dev->dma_pfn_offset is equal to 0x38000.  As such the function returns 0x137fff
> and dma_supported() a non-zero value[3].
> 
> With this set function dma_to_pfn() received a face lift.  Function
> __bus_to_pfn() still returns 0xfffff but translate_dma_to_phys() returns 0,
> which forces dma_supported() to also return 0 and that is where the -5 (-EIO)
> comes from.
> 
> Taking a futher look at translate_dma_to_phy(), @dma_addr never falls within the
> bus_dma_region ranges and returns 0.
> 
> I'm suspecting an initialisation problem and if it occurred here, it will
> likely show up elsewhere.

Can you try this incremental patch?

diff --git a/include/linux/dma-direct.h b/include/linux/dma-direct.h
index 088c97181ab146..c6b21acba7a459 100644
--- a/include/linux/dma-direct.h
+++ b/include/linux/dma-direct.h
@@ -46,7 +46,7 @@ static inline phys_addr_t translate_dma_to_phys(struct device *dev,
 		if (dma_addr >= m->dma_start && dma_addr - m->dma_start < m->size)
 			return (phys_addr_t)dma_addr + m->offset;
 
-	return 0;
+	return (phys_addr_t)-1;
 }
 
 #ifdef CONFIG_ARCH_HAS_PHYS_TO_DMA