Re: [RFC] dma-mapping: dma_alloc_coherent_mask return dma_addr_t

[Konrad Rzeszutek Wilk <konrad.wilk-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>]

On Tue, Dec 10, 2013 at 04:05:37PM +0000, Stefano Panella wrote:
> When running a 32bit kernel there are still some pci devices
> capable of 64bit addressing (eg. sound/pci/hda/hda_intel.c)
> 
> If these drivers are setting:
> 
> dev->coherent_dma_mask = 0xFFFFFFFFFFFFFFFF
> 
> why dma_alloc_coherent_mask is returning unsigned long instead
> of dma_addr_t resulting in truncation of the dma_mask to 32bit
> if running a 32bit kernel?

Ooops.
> 
> There are some examples where the dma_alloc_coherent running
> 32bit kernel can return a machine address bigger than 32bit.
> 
> This can be true in particular if running  the 32bit kernel as a
> pv dom0 under the Xen Hypervisor or PAE on bare metal.
> 
> In both cases it would be good if dma_alloc_coherent could return
> a machine address bigger than 32bit so more constrained 32bit address
> space could be used by devices not 64bit capable.
> 
> I am proposing this patch to eliminate this limitation but I agree
> that there might be some more reasons to do it in the current way of
> which I am not awere.
> 
> This is why I am just asking for a RFC.
> 
> The current patch apply to v3.13-rc3-74-g6c843f5

This looks to be already done for other parts of the kernel, see:
commit 364230b9952143eb2062dc071e919fb751540ae8
Author: Rob Herring <rob.herring-bsGFqQB8/DxBDgjK7y7TUQ@public.gmane.org>
Date:   Thu Aug 1 15:29:29 2013 -0500

    ARM: use phys_addr_t for DMA zone sizes
    
commit 6d7d5da7d75c6df676c8b72d32b02ff024438f0c
Author: Magnus Damm <damm-yzvPICuk2ACczHhG9Qg4qA@public.gmane.org>
Date:   Tue Oct 22 17:59:54 2013 +0100

    ARM: 7864/1: Handle 64-bit memory in case of 32-bit phys_addr_t

mmit dbd3fc3345390a989a033427aa915a0dfb62149f
Author: Bjorn Helgaas <bhelgaas-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org>
Date:   Mon Dec 10 11:24:42 2012 -0700

    PCI: Use phys_addr_t for physical ROM address

commit 7e6735c3578e76c270a2797225a4214176ba13ef
Author: Cyril Chemparathy <cyril-l0cyMroinI0@public.gmane.org>
Date:   Wed Sep 12 14:05:58 2012 -0400

    /dev/mem: use phys_addr_t for physical addresses

Where the phys_addr_t is used instead of unsigned long. This is
doing it for the dma_addr_t.

So I think this is the correct thing to do. We should CC more
people thought.
> 
> Signed-off-by: Stefano Panella <stefano.panella-Sxgqhf6Nn4DQT0dZR+AlfA@public.gmane.org>
> ---
>  Documentation/DMA-API.txt          |    5 ++---
>  arch/x86/include/asm/dma-mapping.h |    8 ++++----
>  arch/x86/kernel/pci-dma.c          |    2 +-
>  drivers/xen/swiotlb-xen.c          |    6 +++---
>  4 files changed, 10 insertions(+), 11 deletions(-)
> 
> diff --git a/Documentation/DMA-API.txt b/Documentation/DMA-API.txt
> index e865279..483458b 100644
> --- a/Documentation/DMA-API.txt
> +++ b/Documentation/DMA-API.txt
> @@ -33,9 +33,8 @@ to make sure to flush the processor's write buffers before telling
>  devices to read that memory.)
>  
>  This routine allocates a region of <size> bytes of consistent memory.
> -It also returns a <dma_handle> which may be cast to an unsigned
> -integer the same width as the bus and used as the physical address
> -base of the region.
> +It also returns a <dma_handle> which should be used as the physical 
> +address base of the region.
>  
>  Returns: a pointer to the allocated region (in the processor's virtual
>  address space) or NULL if the allocation failed.
> diff --git a/arch/x86/include/asm/dma-mapping.h b/arch/x86/include/asm/dma-mapping.h
> index 808dae6..6357810 100644
> --- a/arch/x86/include/asm/dma-mapping.h
> +++ b/arch/x86/include/asm/dma-mapping.h
> @@ -100,10 +100,10 @@ dma_cache_sync(struct device *dev, void *vaddr, size_t size,
>  	flush_write_buffers();
>  }
>  
> -static inline unsigned long dma_alloc_coherent_mask(struct device *dev,
> -						    gfp_t gfp)
> +static inline dma_addr_t dma_alloc_coherent_mask(struct device *dev,
> +						 gfp_t gfp)
>  {
> -	unsigned long dma_mask = 0;
> +	dma_addr_t dma_mask = 0;
>  
>  	dma_mask = dev->coherent_dma_mask;
>  	if (!dma_mask)
> @@ -114,7 +114,7 @@ static inline unsigned long dma_alloc_coherent_mask(struct device *dev,
>  
>  static inline gfp_t dma_alloc_coherent_gfp_flags(struct device *dev, gfp_t gfp)
>  {
> -	unsigned long dma_mask = dma_alloc_coherent_mask(dev, gfp);
> +	dma_addr_t dma_mask = dma_alloc_coherent_mask(dev, gfp);
>  
>  	if (dma_mask <= DMA_BIT_MASK(24))
>  		gfp |= GFP_DMA;
> diff --git a/arch/x86/kernel/pci-dma.c b/arch/x86/kernel/pci-dma.c
> index 872079a..a061f7b 100644
> --- a/arch/x86/kernel/pci-dma.c
> +++ b/arch/x86/kernel/pci-dma.c
> @@ -90,7 +90,7 @@ void *dma_generic_alloc_coherent(struct device *dev, size_t size,
>  				 dma_addr_t *dma_addr, gfp_t flag,
>  				 struct dma_attrs *attrs)
>  {
> -	unsigned long dma_mask;
> +	dma_addr_t dma_mask;
>  	struct page *page;
>  	unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
>  	dma_addr_t addr;
> diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c
> index 1eac073..6f75599 100644
> --- a/drivers/xen/swiotlb-xen.c
> +++ b/drivers/xen/swiotlb-xen.c
> @@ -54,10 +54,10 @@
>   */
>  
>  #ifndef CONFIG_X86
> -static unsigned long dma_alloc_coherent_mask(struct device *dev,
> -					    gfp_t gfp)
> +static dma_addr_t dma_alloc_coherent_mask(struct device *dev,
> +					  gfp_t gfp)
>  {
> -	unsigned long dma_mask = 0;
> +	dma_addr_t dma_mask = 0;
>  
>  	dma_mask = dev->coherent_dma_mask;
>  	if (!dma_mask)
> -- 
> 1.7.9.5
>