[PATCH v3] drm/etnaviv: Request pages from DMA32 zone on addressing_limited

[Sui Jingfeng]

Hi,


On 2024/10/1 20:17, Lucas Stach wrote:
> Hi Xiaolei,
>
> Am Dienstag, dem 03.09.2024 um 10:08 +0800 schrieb Xiaolei Wang:
>> Remove __GFP_HIGHMEM when requesting a page from DMA32 zone,
>> and since all vivante GPUs in the system will share the same
>> DMA constraints, move the check of whether to get a page from
>> DMA32 to etnaviv_bind().
>>
>> Fixes: b72af445cd38 ("drm/etnaviv: request pages from DMA32 zone when needed")
>> Suggested-by: Sui Jingfeng <sui.jingfeng at linux.dev>
>> Signed-off-by: Xiaolei Wang <xiaolei.wang at windriver.com>
>> ---
>>
>> change log
>>
>> v1:
>>    https://patchwork.kernel.org/project/dri-devel/patch/20240806104733.2018783-1-xiaolei.wang@windriver.com/
>>
>> v2:
>>    Modify the issue of not retaining GFP_USER in v1 and update the commit log.
>>
>> v3:
>>    Use "priv->shm_gfp_mask = GFP_USER | __GFP_RETRY_MAYFAIL | __GFP_NOWARN;"
>> instead of
>>    "priv->shm_gfp_mask = GFP_HIGHUSER | __GFP_RETRY_MAYFAIL | __GFP_NOWARN;"
> I don't understand this part of the changes in the new version. Why
> should we drop the HIGHMEM bit always and not only in the case where
> dma addressing is limited? This seems overly restrictive.

While reading the implementation of the dma_alloc_attrs() function,
except allocate memory from device coherent pool, the rest implementation
just mask out __GFP_DMA, __GFP_DMA32 and __GFP_HIGHMEM anyway.

So ?


```
void *dma_alloc_attrs(struct device *dev, size_t size, dma_addr_t *dma_handle,
		gfp_t flag, unsigned long attrs)
{
	const struct dma_map_ops *ops = get_dma_ops(dev);
	void *cpu_addr;

	WARN_ON_ONCE(!dev->coherent_dma_mask);

	/*
	 * DMA allocations can never be turned back into a page pointer, so
	 * requesting compound pages doesn't make sense (and can't even be
	 * supported at all by various backends).
	 */
	if (WARN_ON_ONCE(flag & __GFP_COMP))
		return NULL;

	if (dma_alloc_from_dev_coherent(dev, size, dma_handle, &cpu_addr))
		return cpu_addr;

	/* let the implementation decide on the zone to allocate from: */
	flag &= ~(__GFP_DMA | __GFP_DMA32 | __GFP_HIGHMEM);

	if (dma_alloc_direct(dev, ops))
		cpu_addr = dma_direct_alloc(dev, size, dma_handle, flag, attrs);
	else if (use_dma_iommu(dev))
		cpu_addr = iommu_dma_alloc(dev, size, dma_handle, flag, attrs);
	else if (ops->alloc)
		cpu_addr = ops->alloc(dev, size, dma_handle, flag, attrs);
	else
		return NULL;


	return cpu_addr;
}
```


> Regards,
> Lucas
>
>> and move the check of whether to get a page from DMA32 to etnaviv_bind().
>>
>>   drivers/gpu/drm/etnaviv/etnaviv_drv.c | 10 +++++++++-
>>   drivers/gpu/drm/etnaviv/etnaviv_gpu.c |  8 --------
>>   2 files changed, 9 insertions(+), 9 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/etnaviv/etnaviv_drv.c b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
>> index 6500f3999c5f..8cb2c3ec8e5d 100644
>> --- a/drivers/gpu/drm/etnaviv/etnaviv_drv.c
>> +++ b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
>> @@ -536,7 +536,15 @@ static int etnaviv_bind(struct device *dev)
>>   	mutex_init(&priv->gem_lock);
>>   	INIT_LIST_HEAD(&priv->gem_list);
>>   	priv->num_gpus = 0;
>> -	priv->shm_gfp_mask = GFP_HIGHUSER | __GFP_RETRY_MAYFAIL | __GFP_NOWARN;
>> +	priv->shm_gfp_mask = GFP_USER | __GFP_RETRY_MAYFAIL | __GFP_NOWARN;
>> +
>> +	/*
>> +	 * If the GPU is part of a system with DMA addressing limitations,
>> +	 * request pages for our SHM backend buffers from the DMA32 zone to
>> +	 * hopefully avoid performance killing SWIOTLB bounce buffering.
>> +	 */
>> +	if (dma_addressing_limited(dev))
>> +		priv->shm_gfp_mask |= GFP_DMA32;
>>   
>>   	priv->cmdbuf_suballoc = etnaviv_cmdbuf_suballoc_new(drm->dev);
>>   	if (IS_ERR(priv->cmdbuf_suballoc)) {
>> diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
>> index 7c7f97793ddd..5e753dd42f72 100644
>> --- a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
>> +++ b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
>> @@ -839,14 +839,6 @@ int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
>>   	if (ret)
>>   		goto fail;
>>   
>> -	/*
>> -	 * If the GPU is part of a system with DMA addressing limitations,
>> -	 * request pages for our SHM backend buffers from the DMA32 zone to
>> -	 * hopefully avoid performance killing SWIOTLB bounce buffering.
>> -	 */
>> -	if (dma_addressing_limited(gpu->dev))
>> -		priv->shm_gfp_mask |= GFP_DMA32;
>> -
>>   	/* Create buffer: */
>>   	ret = etnaviv_cmdbuf_init(priv->cmdbuf_suballoc, &gpu->buffer,
>>   				  PAGE_SIZE);

-- 
Best regards,
Sui