[PATCH] arm64: avoid increasing DMA masks above what hardware supports

[Robin Murphy]

On 10/01/17 14:00, Nikita Yushchenko wrote:
> There are cases when device supports wide DMA addresses wider than
> device's connection supports.
> 
> In this case driver sets DMA mask based on knowledge of device
> capabilities. That must succeed to allow drivers to initialize.
> 
> However, swiotlb or iommu still need knowledge about actual device
> capabilities. To avoid breakage, actual mask must not be set wider than
> device connection allows.
> 
> Signed-off-by: Nikita Yushchenko <nikita.yoush at cogentembedded.com>
> CC: Arnd Bergmann <arnd at arndb.de>
> CC: Robin Murphy <robin.murphy at arm.com>
> CC: Will Deacon <will.deacon at arm.com>
> ---
>  arch/arm64/Kconfig                   |  3 +++
>  arch/arm64/include/asm/device.h      |  1 +
>  arch/arm64/include/asm/dma-mapping.h |  3 +++
>  arch/arm64/mm/dma-mapping.c          | 43 ++++++++++++++++++++++++++++++++++++
>  4 files changed, 50 insertions(+)
> 
> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> index 1117421..afb2c08 100644
> --- a/arch/arm64/Kconfig
> +++ b/arch/arm64/Kconfig
> @@ -216,6 +216,9 @@ config NEED_DMA_MAP_STATE
>  config NEED_SG_DMA_LENGTH
>  	def_bool y
>  
> +config ARCH_HAS_DMA_SET_COHERENT_MASK
> +	def_bool y
> +
>  config SMP
>  	def_bool y
>  
> diff --git a/arch/arm64/include/asm/device.h b/arch/arm64/include/asm/device.h
> index 243ef25..a57e7bb 100644
> --- a/arch/arm64/include/asm/device.h
> +++ b/arch/arm64/include/asm/device.h
> @@ -22,6 +22,7 @@ struct dev_archdata {
>  	void *iommu;			/* private IOMMU data */
>  #endif
>  	bool dma_coherent;
> +	u64 parent_dma_mask;
>  };
>  
>  struct pdev_archdata {
> diff --git a/arch/arm64/include/asm/dma-mapping.h b/arch/arm64/include/asm/dma-mapping.h
> index ccea82c..eab36d2 100644
> --- a/arch/arm64/include/asm/dma-mapping.h
> +++ b/arch/arm64/include/asm/dma-mapping.h
> @@ -51,6 +51,9 @@ void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
>  			const struct iommu_ops *iommu, bool coherent);
>  #define arch_setup_dma_ops	arch_setup_dma_ops
>  
> +#define HAVE_ARCH_DMA_SET_MASK 1
> +extern int dma_set_mask(struct device *dev, u64 dma_mask);
> +
>  #ifdef CONFIG_IOMMU_DMA
>  void arch_teardown_dma_ops(struct device *dev);
>  #define arch_teardown_dma_ops	arch_teardown_dma_ops
> diff --git a/arch/arm64/mm/dma-mapping.c b/arch/arm64/mm/dma-mapping.c
> index e040827..7b1bb87 100644
> --- a/arch/arm64/mm/dma-mapping.c
> +++ b/arch/arm64/mm/dma-mapping.c
> @@ -203,6 +203,37 @@ static void __dma_free(struct device *dev, size_t size,
>  	__dma_free_coherent(dev, size, swiotlb_addr, dma_handle, attrs);
>  }
>  
> +int dma_set_mask(struct device *dev, u64 dma_mask)
> +{
> +	const struct dma_map_ops *ops = get_dma_ops(dev);
> +
> +	if (mask > dev->archdata.parent_dma_mask)
> +		mask = dev->archdata.parent_dma_mask;
> +
> +	if (ops->set_dma_mask)
> +		return ops->set_dma_mask(dev, mask);
> +
> +	if (!dev->dma_mask || !dma_supported(dev, mask))
> +		return -EIO;
> +
> +	*dev->dma_mask = mask;
> +	return 0;
> +}
> +EXPORT_SYMBOL(dma_set_mask);
> +
> +int dma_set_coherent_mask(struct device *dev, u64 mask)
> +{
> +	if (mask > dev->archdata.parent_dma_mask)
> +		mask = dev->archdata.parent_dma_mask;
> +
> +	if (!dma_supported(dev, mask))
> +		return -EIO;
> +
> +	dev->coherent_dma_mask = mask;
> +	return 0;
> +}
> +EXPORT_SYMBOL(dma_set_coherent_mask);
> +
>  static dma_addr_t __swiotlb_map_page(struct device *dev, struct page *page,
>  				     unsigned long offset, size_t size,
>  				     enum dma_data_direction dir,
> @@ -958,6 +989,18 @@ void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
>  	if (!dev->archdata.dma_ops)
>  		dev->archdata.dma_ops = &swiotlb_dma_ops;
>  
> +	/*
> +	 * we don't yet support buses that have a non-zero mapping.
> +	 *  Let's hope we won't need it
> +	 */
> +	WARN_ON(dma_base != 0);

I believe we now accomodate the bus remap bits on BCM2837 as a DMA
offset, so unfortunately I think this is no longer true.

> +	/*
> +	 * Whatever the parent bus can set. A device must not set
> +	 * a DMA mask larger than this.
> +	 */
> +	dev->archdata.parent_dma_mask = size - 1;

This will effectively constrain *all* DMA masks to be 32-bit, since for
99% of devices we're going to see a size derived from the default mask
passed in here. I worry that that's liable to lead to performance and
stability regressions (now that the block layer can apparently generate
sufficient readahead to ovflow a typical SWIOTLB bounce buffer[1]).
Whilst DT users would be able to mitigate that by putting explicit
"dma-ranges" properties on every device node, it's less clear what we'd
do for ACPI.

I reckon the easiest way forward would be to pass in some flag to
arch_setup_dma_ops to indicate whether it's an explicitly-configured
range or not - then simply initialising parent_dma_mask to ~0 for the
default case *should* keep things working as before.

Robin.

[1]:https://www.mail-archive.com/virtualization@lists.linux-foundation.org/msg26532.html

> +
>  	dev->archdata.dma_coherent = coherent;
>  	__iommu_setup_dma_ops(dev, dma_base, size, iommu);
>  }
>