[PATCH 0/6] RISC-V: Add eMMC support for TH1520 boards

[Robin Murphy]

On 2023-10-04 19:49, Samuel Holland wrote:
> On 2023-10-04 12:16 PM, Lad, Prabhakar wrote:
>> On Wed, Oct 4, 2023 at 5:03 PM Lad, Prabhakar
>> <prabhakar.csengg at gmail.com> wrote:
>>>
>>> On Wed, Oct 4, 2023 at 3:18 PM Robin Murphy <robin.murphy at arm.com> wrote:
>>>>
>>>> On 04/10/2023 3:02 pm, Icenowy Zheng wrote:
>>>> [...]
>>>>>>>> I believe commit 484861e09f3e ("soc: renesas: Kconfig: Select the
>>>>>>>> required configs for RZ/Five SoC") can cause regression on all
>>>>>>>> non-dma-coherent riscv platforms with generic defconfig. This is
>>>>>>>> a common issue. The logic here is: generic riscv defconfig
>>>>>>>> selects
>>>>>>>> ARCH_R9A07G043 which selects DMA_GLOBAL_POOL, which assumes all
>>>>>>>> non-dma-coherent riscv platforms have a dma global pool, this
>>>>>>>> assumption
>>>>>>>> seems not correct. And I believe DMA_GLOBAL_POOL should not be
>>>>>>>> selected by ARCH_SOCFAMILIY, instead, only ARCH under some
>>>>>>>> specific
>>>>>>>> conditions can select it globaly, for example NOMMU ARM and so
>>>>>>>> on.
>>>>>>>>
>>>>>>>> Since this is a regression, what's proper fix? any suggestion is
>>>>>>>> appreciated.
>>>>>>
>>>>>> I think the answer is to not select DMA_GLOBAL_POOL, since that is
>>>>>> only
>>>>>
>>>>> Well I think for RISC-V, it's not NOMMU only but applicable for every
>>>>> core that does not support Svpbmt or vendor-specific alternatives,
>>>>> because the original RISC-V priv spec does not define memory attributes
>>>>> in page table entries.
>>>>>
>>>>> For the Renesas/Andes case I think a pool is set by OpenSBI with
>>>>> vendor-specific M-mode facility and then passed in DT, and the S-mode
>>>>> (which MMU is enabled in) just sees fixed memory attributes, in this
>>>>> case I think DMA_GLOBAL_POOL is needed.
>>>>
>>>> Oh wow, is that really a thing? In that case, either you just can't
>>>> support this platform in a multi-platform kernel, or someone needs to do
>>>> some fiddly work in dma-direct to a) introduce the notion of an optional
>>>> global pool,
>>> Looking at the code [0] we do have compile time check for
>>> CONFIG_DMA_GLOBAL_POOL irrespective of this being present in DT or
>>> not, instead if we make it compile time and runtime check ie either
>>> check for DT node or see if pool is available and only then proceed
>>> for allocation form this pool.
>>>
>>> What are your thoughts on this?
>>>
>> Something like the below:
>>
>> diff --git a/include/linux/dma-map-ops.h b/include/linux/dma-map-ops.h
>> index f2fc203fb8a1..7bf41a4634a4 100644
>> --- a/include/linux/dma-map-ops.h
>> +++ b/include/linux/dma-map-ops.h
>> @@ -198,6 +198,7 @@ int dma_release_from_global_coherent(int order,
>> void *vaddr);
>>   int dma_mmap_from_global_coherent(struct vm_area_struct *vma, void *cpu_addr,
>>                  size_t size, int *ret);
>>   int dma_init_global_coherent(phys_addr_t phys_addr, size_t size);
>> +bool dma_global_pool_available(void);
>>   #else
>>   static inline void *dma_alloc_from_global_coherent(struct device *dev,
>>                  ssize_t size, dma_addr_t *dma_handle)
>> @@ -213,6 +214,10 @@ static inline int
>> dma_mmap_from_global_coherent(struct vm_area_struct *vma,
>>   {
>>          return 0;
>>   }
>> +static inline bool dma_global_pool_available(void)
>> +{
>> +       return false;
>> +}
>>   #endif /* CONFIG_DMA_GLOBAL_POOL */
>>
>>   /*
>> diff --git a/kernel/dma/coherent.c b/kernel/dma/coherent.c
>> index c21abc77c53e..605f243b8262 100644
>> --- a/kernel/dma/coherent.c
>> +++ b/kernel/dma/coherent.c
>> @@ -277,6 +277,14 @@ int dma_mmap_from_dev_coherent(struct device
>> *dev, struct vm_area_struct *vma,
>>   #ifdef CONFIG_DMA_GLOBAL_POOL
>>   static struct dma_coherent_mem *dma_coherent_default_memory __ro_after_init;
>>
>> +bool dma_global_pool_available(void)
>> +{
>> +       if (!dma_coherent_default_memory)
>> +               return false;
>> +
>> +       return true;
>> +}
>> +
>>   void *dma_alloc_from_global_coherent(struct device *dev, ssize_t size,
>>                                       dma_addr_t *dma_handle)
>>   {
>> diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
>> index 9596ae1aa0da..a599bb731ceb 100644
>> --- a/kernel/dma/direct.c
>> +++ b/kernel/dma/direct.c
>> @@ -235,7 +235,7 @@ void *dma_direct_alloc(struct device *dev, size_t size,
>>                   * If there is a global pool, always allocate from it for
>>                   * non-coherent devices.
>>                   */
>> -               if (IS_ENABLED(CONFIG_DMA_GLOBAL_POOL))
>> +               if (IS_ENABLED(CONFIG_DMA_GLOBAL_POOL) &&
>> dma_global_pool_available())
>>                          return dma_alloc_from_global_coherent(dev, size,
>>                                          dma_handle);
> 
> dma_alloc_from_global_coherent() already checks dma_coherent_default_memory, so
> the solution could be even simpler:
> 
> --- a/kernel/dma/direct.c
> +++ b/kernel/dma/direct.c
> @@ -232,12 +232,12 @@ void *dma_direct_alloc(struct device *dev, size_t size,
>   					      attrs);
> 
>   		/*
> -		 * If there is a global pool, always allocate from it for
> +		 * If there is a global pool, always try to allocate from it for
>   		 * non-coherent devices.
>   		 */
> -		if (IS_ENABLED(CONFIG_DMA_GLOBAL_POOL))
> -			return dma_alloc_from_global_coherent(dev, size,
> -					dma_handle);
> +		ret = dma_alloc_from_global_coherent(dev, size, dma_handle);
> +		if (ret)
> +			return ret;

So if allocation fails because the pool is full, we should go ahead and 
remap something that can't work? ;)

The dma_global_pool_available() idea sort of works, but I'm still 
concerned about the case where it *should* have been available but the 
platform has been misconfigured, so again we fall through to 
DMA_DIRECT_REMAP "successfully" returning a coherent buffer that isn't, 
and the user's filesystem gets corrupted. Or at best, they get confused 
by weird errors from random devices going wrong. That's why I said it 
would be fiddly - the current state of DMA_GLOBAL_POOL as a binary 
arch-wide thing is relatively robust and easy to reason about, but 
attempting to generalise it further is... less so.

Thanks,
Robin.

> 
>   		/*
>   		 * Otherwise remap if the architecture is asking for it.  But
> 
> Regards,
> Samuel
>