Re: [PATCH 1/1] arm64/dma-mapping: remove an unnecessary conversion

[Catalin Marinas]

On Thu, Mar 17, 2016 at 07:06:27PM +0800, Leizhen (ThunderTown) wrote:
> On 2016/3/16 9:56, Leizhen (ThunderTown) wrote:
> > On 2016/3/15 23:37, Catalin Marinas wrote:
> >> On Tue, Mar 15, 2016 at 10:12:11AM +0800, Zhen Lei wrote:
> >>> diff --git a/arch/arm64/mm/dma-mapping.c b/arch/arm64/mm/dma-mapping.c
> >>> index a6e757c..b2f2834 100644
> >>> --- a/arch/arm64/mm/dma-mapping.c
> >>> +++ b/arch/arm64/mm/dma-mapping.c
> >>> @@ -187,8 +187,6 @@ static void __dma_free(struct device *dev, size_t size,
> >>>  		       void *vaddr, dma_addr_t dma_handle,
> >>>  		       struct dma_attrs *attrs)
> >>>  {
> >>> -	void *swiotlb_addr = phys_to_virt(dma_to_phys(dev, dma_handle));
> >>> -
> >>>  	size = PAGE_ALIGN(size);
> >>>
> >>>  	if (!is_device_dma_coherent(dev)) {
> >>> @@ -196,7 +194,7 @@ static void __dma_free(struct device *dev, size_t size,
> >>>  			return;
> >>>  		vunmap(vaddr);
> >>>  	}
> >>> -	__dma_free_coherent(dev, size, swiotlb_addr, dma_handle, attrs);
> >>> +	__dma_free_coherent(dev, size, vaddr, dma_handle, attrs);
> >>>  }
> >>
> >> What happens when !is_device_dma_coherent(dev)? (hint: read two lines
> >> above __dma_free_coherent).
> 
> Do you afraid "vaddr" maybe modified by these statement?
> First, it could not be __free_from_pool. Otherwise, the function
> vunmap(which after it) can not work well. Then, it count not be vunmap
> too, the parameter is defined as "const void *".
> 
> In the call chain:
> __dma_free_coherent-->__dma_free_coherent-->swiotlb_free_coherent,
> only swiotlb_free_coherent finally use "vaddr".

Exactly. So you give swiotlb_free_coherent a vaddr which has been
unmapped. It doesn't even matter whether it's still mapped since this
address is passed further to free_pages() which performs a
virt_to_page(). The latter is *only* valid on linear map addresses (and
you would actually hit the VM_BUG_ON in free_pages; you can try running
this with CONFIG_DEBUG_VM enabled and non-coherent DMA).

For non-coherent DMA, the vaddr is not part of the linear mapping as it
has been remapped by __dma_alloc() via dma_common_contiguous_remap(),
hence for swiotlb freeing we need the actual linear map address (the
original "ptr" in __dma_alloc()). We can generate it by a
phys_to_virt(dma_to_phys(dma_handle)).

-- 
Catalin