[PATCH 3/7] drivers: dma-coherent: Account dma_pfn_offset when used with device tree

[Robin Murphy]

On 21/02/17 13:02, Vladimir Murzin wrote:
> On 21/02/17 12:37, Robin Murphy wrote:
>> On 15/02/17 09:59, Vladimir Murzin wrote:
>>> dma_declare_coherent_memory() and friends are designed to account
>>> difference in CPU and device addresses. However, when it is used with
>>> reserved memory regions there is assumption that CPU and device have
>>> the same view on address space. This assumption gets invalid when
>>> reserved memory for coherent DMA allocations is referenced by device
>>> with non-empty "dma-range" property.
>>>
>>> Simply feeding device address as rmem->base + dev->dma_pfn_offset
>>> would not work due to reserved memory region can be shared, so this
>>> patch turns device address to be expressed with help of CPU address
>>> and device's dma_pfn_offset.
>>>
>>> For the case where device tree is not used and device sees memory
>>> different to CPU we explicitly set device's dma_pfn_offset to
>>> accomplish such difference. The latter might look controversial, but
>>> it seems only a few drivers set device address different to CPU's:
>>> - drivers/usb/host/ohci-sm501.c
>>> - arch/sh/drivers/pci/fixups-dreamcast.c
>>> so they can be screwed only if dma_pfn_offset there is set and not in
>>> sync with device address range - we try to catch such cases with
>>> WARN_ON.
>>>
>>> Cc: Michal Nazarewicz <mina86 at mina86.com>
>>> Cc: Marek Szyprowski <m.szyprowski at samsung.com>
>>> Cc: Alan Stern <stern at rowland.harvard.edu>
>>> Cc: Yoshinori Sato <ysato at users.sourceforge.jp>
>>> Cc: Rich Felker <dalias at libc.org>
>>> Cc: Roger Quadros <rogerq at ti.com>
>>> Cc: Greg Kroah-Hartman <gregkh at linuxfoundation.org>
>>> Tested-by: Benjamin Gaignard <benjamin.gaignard at linaro.org>
>>> Tested-by: Andras Szemzo <sza at esh.hu>
>>> Tested-by: Alexandre TORGUE <alexandre.torgue at st.com>
>>> Signed-off-by: Vladimir Murzin <vladimir.murzin at arm.com>
>>> ---
>>>  drivers/base/dma-coherent.c | 17 +++++++++++++++--
>>>  1 file changed, 15 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/base/dma-coherent.c b/drivers/base/dma-coherent.c
>>> index 640a7e6..c59708c 100644
>>> --- a/drivers/base/dma-coherent.c
>>> +++ b/drivers/base/dma-coherent.c
>>> @@ -18,6 +18,12 @@ struct dma_coherent_mem {
>>>  	spinlock_t	spinlock;
>>>  };
>>>  
>>> +static inline dma_addr_t dma_get_device_base(struct device *dev,
>>> +					     struct dma_coherent_mem * mem)
>>> +{
>>> +	return (mem->pfn_base - dev->dma_pfn_offset) << PAGE_SHIFT;
>>> +}
>>> +
>>>  static bool dma_init_coherent_memory(
>>>  	phys_addr_t phys_addr, dma_addr_t device_addr, size_t size, int flags,
>>>  	struct dma_coherent_mem **mem)
>>> @@ -83,9 +89,16 @@ static void dma_release_coherent_memory(struct dma_coherent_mem *mem)
>>>  static int dma_assign_coherent_memory(struct device *dev,
>>>  				      struct dma_coherent_mem *mem)
>>>  {
>>> +	unsigned long dma_pfn_offset = mem->pfn_base - PFN_DOWN(mem->device_base);
>>> +
>>>  	if (dev->dma_mem)
>>>  		return -EBUSY;
>>>  
>>> +	if (dev->dma_pfn_offset)
>>> +		WARN_ON(dma_pfn_offset && (dev->dma_pfn_offset != dma_pfn_offset));
>>> +	else
>>> +		dev->dma_pfn_offset = dma_pfn_offset;
>>
>> This makes me rather uneasy - I can well imagine a device sharing the
>> CPU physical address map of external system memory, but having its own
>> view of its local coherent memory such that pfn_base != device_base
>> still. I know for a fact we've had internal FPGA tiles set up that way,
>> although whether it was entirely intentional is another matter... ;)
>>
>> In that situation, setting dev->dma_pfn_offset like this would break
>> streaming DMA for such devices. Could we not keep the pool-specific
>> offset and the device-specific offset independent, apply whichever is
>> non-zero, and scream if both are set?
> 
> Something like fixup bellow?
> 
> diff --git a/drivers/base/dma-coherent.c b/drivers/base/dma-coherent.c
> index 0c577ea..2060010 100644
> --- a/drivers/base/dma-coherent.c
> +++ b/drivers/base/dma-coherent.c
> @@ -30,7 +30,15 @@ static inline struct dma_coherent_mem *dev_get_coherent_memory(struct device *de
>  static inline dma_addr_t dma_get_device_base(struct device *dev,
>  					     struct dma_coherent_mem * mem)
>  {
> -	return (mem->pfn_base - dev->dma_pfn_offset) << PAGE_SHIFT;
> +	unsigned long dma_pfn_offset = mem->pfn_base - PFN_DOWN(mem->device_base);
> +
> +	if (dma_pfn_offset != dev->dma_pfn_offset )
> +		WARN_ON(device_pfn_offset && dev->dma_pfn_offset);
> +
> +	if (dma_pfn_offset)
> +		return mem->device_base;
> +	else
> +		return (mem->pfn_base - dev->dma_pfn_offset) << PAGE_SHIFT;
>  }

More or less - I can't resist the temptation to golf it down a bit
though, do you reckon this scores any higher for readability?

	unsigned long dev_pfn = PFN_DOWN(mem->device_base);

	if (dev_pfn == mem->pfn_base)
		dev_pfn -= dev->dma_pfn_offset;
	else
		WARN_ON(dev->dma_pfn_offset);

	return (dma_addr_t)dev_pfn << PAGE_SHIFT;

perhaps with a comment that it's not entirely unambiguous what it would
mean for both offsets to be nonzero (even if they were equal), so we'd
need to revisit the whole thing anyway in the vanishingly unlikely
situation that any hardware that crazy ever appears.

Another possibility is that we add something in mem->flags to indicate
if it came from of_reserved_mem, and only apply dev->dma_pfn_offset in
those cases. That feels like it might actually be a bit more robust.

Robin.

>  static bool dma_init_coherent_memory(
> @@ -98,19 +106,12 @@ static void dma_release_coherent_memory(struct dma_coherent_mem *mem)
>  static int dma_assign_coherent_memory(struct device *dev,
>  				      struct dma_coherent_mem *mem)
>  {
> -	unsigned long dma_pfn_offset = mem->pfn_base - PFN_DOWN(mem->device_base);
> -
>  	if (!dev)
>  		return -ENODEV;
>  
>  	if (dev->dma_mem)
>  		return -EBUSY;
>  
> -	if (dev->dma_pfn_offset)
> -		WARN_ON(dma_pfn_offset && (dev->dma_pfn_offset != dma_pfn_offset));
> -	else
> -		dev->dma_pfn_offset = dma_pfn_offset;
> -
>  	dev->dma_mem = mem;
>  	/* FIXME: this routine just ignores DMA_MEMORY_INCLUDES_CHILDREN */
> 
> 
> Cheers
> Vladimir
> 
>>
>> Robin.
>>
>>> +
>>>  	dev->dma_mem = mem;
>>>  	/* FIXME: this routine just ignores DMA_MEMORY_INCLUDES_CHILDREN */
>>>  
>>> @@ -133,7 +146,7 @@ void *dma_mark_declared_memory_occupied(struct device *dev,
>>>  		return ERR_PTR(-EINVAL);
>>>  
>>>  	spin_lock_irqsave(&mem->spinlock, flags);
>>> -	pos = (device_addr - mem->device_base) >> PAGE_SHIFT;
>>> +	pos = PFN_DOWN(device_addr - dma_get_device_base(dev, mem));
>>>  	err = bitmap_allocate_region(mem->bitmap, pos, get_order(size));
>>>  	spin_unlock_irqrestore(&mem->spinlock, flags);
>>>  
>>> @@ -186,7 +199,7 @@ int dma_alloc_from_coherent(struct device *dev, ssize_t size,
>>>  	/*
>>>  	 * Memory was found in the per-device area.
>>>  	 */
>>> -	*dma_handle = mem->device_base + (pageno << PAGE_SHIFT);
>>> +	*dma_handle = dma_get_device_base(dev, mem) + (pageno << PAGE_SHIFT);
>>>  	*ret = mem->virt_base + (pageno << PAGE_SHIFT);
>>>  	dma_memory_map = (mem->flags & DMA_MEMORY_MAP);
>>>  	spin_unlock_irqrestore(&mem->spinlock, flags);
>>>
>>
>>
>