Re: [PATCH V6 1/3] dma-mapping: Introduce device_is_coherent() as ahelper

[éåæ]

Hi, Robin,

Before 2.6.36 dma_get_cache_alignment is arch-dependent, and it is unified in commit 4565f0170dfc849b3629c27d7 ("dma-mapping: unify dma_get_cache_alignment implementations"). Should we revert to the old implementation?
 
Huacai
 
------------------ Original ------------------
From:  "Robin Murphy"<robin.murphy@xxxxxxx>;
Date:  Thu, Sep 21, 2017 06:47 PM
To:  "Huacai Chen"<chenhc@xxxxxxxxxx>; "Christoph Hellwig"<hch@xxxxxx>; 
Cc:  "Marek Szyprowski"<m.szyprowski@xxxxxxxxxxx>; "Andrew Morton"<akpm@xxxxxxxxxxxxxxxxxxxx>; "Fuxin Zhang"<zhangfx@xxxxxxxxxx>; "linux-kernel"<linux-kernel@xxxxxxxxxxxxxxx>; "James E . J . Bottomley"<jejb@xxxxxxxxxxxxxxxxxx>; "Martin K . Petersen"<martin.petersen@xxxxxxxxxx>; "linux-scsi"<linux-scsi@xxxxxxxxxxxxxxx>; "stable"<stable@xxxxxxxxxxxxxxx>; 
Subject:  Re: [PATCH V6 1/3] dma-mapping: Introduce device_is_coherent() as ahelper

 
On 19/09/17 09:52, Huacai Chen wrote:
> We will use device_is_coherent() as a helper function, which will be
> used in the next patch.
> 
> There is a MIPS-specific plat_device_is_coherent(), but we need a more
> generic solution, so add and use a new function pointer in dma_map_ops.

I think we're heading in the right direction with the series, but I
still don't like this patch. I can pretty much guarantee that driver
authors *will* abuse a generic device_is_coherent() API to mean "I can
skip other DMA API calls and just use virt_to_phys()".

I think it would be far better to allow architectures to provide their
own override of dma_get_cache_alignment(), and let the coherency detail
remain internal to the relevant arch implementations.

[...]
> @@ -697,6 +698,15 @@ static inline void *dma_zalloc_coherent(struct device *dev, size_t size,
>  }
>  
>  #ifdef CONFIG_HAS_DMA
> +static inline int device_is_coherent(struct device *dev)
> +{
> +	const struct dma_map_ops *ops = get_dma_ops(dev);
> +	if (ops && ops->device_is_coherent)
> +		return ops->device_is_coherent(dev);
> +	else
> +		return 1;    /* compatible behavior */

That is also quite scary - if someone now adds a new
dma_get_cache_alignemnt() call and dutifully passes a non-NULL device,
they will now get back an alignment of 1 on all non-coherent platforms
except MIPS: hello data corruption.

Robin.

> +}
> +
>  static inline int dma_get_cache_alignment(void)
>  {
>  #ifdef ARCH_DMA_MINALIGN
>