2023-09-13 12:17:15

by Robin Murphy

Subject: Re: [PATCH v10 1/5] staging: vc04_services: vchiq_arm: Add new bus type and device type

On 2023-09-12 06:50, Umang Jain wrote:
[...]
>>> +struct vchiq_device *
>>> +vchiq_device_register(struct device *parent, const char *name)
>>> +{
>>> +    struct vchiq_device *device;
>>> +    int ret;
>>> +
>>> +    device = kzalloc(sizeof(*device), GFP_KERNEL);
>>> +    if (!device) {
>>> +        dev_err(parent, "Cannot register %s: Insufficient memory\n",
>>> +            name);
>>> +        return NULL;
>>> +    }
>>> +
>>> +    device->dev.init_name = name;
>>> +    device->dev.parent = parent;
>>> +    device->dev.bus = &vchiq_bus_type;
>>> +    device->dev.release = vchiq_device_release;
>>> +
>>> +    of_dma_configure(&device->dev, parent->of_node, true);
>>> +    ret = dma_set_mask_and_coherent(&device->dev, DMA_BIT_MASK(32));
>>> +    if (ret) {
>>> +        dev_err(&device->dev, "32-bit DMA enable failed\n");
>>> +        return NULL;
>>> +    }
>>
>> Unfortunately the call of of_dma_configure() generates warnings likes
>> this (Raspberry Pi 3A+ with multi_v7_defconfig + VCHIQ):
>>
>> [    9.206802] vchiq-bus bcm2835-audio: DMA mask not set
>> [    9.206892] vchiq-bus bcm2835-camera: DMA mask not set
>
> huh, really weird, as on my RPi-3-b I get these set correctly and I
> don't any such warning.

Can you point to the code above where device->dev.dma_mask gets
initialised between the initial kzalloc() and the call to
of_dma_configure()? ;)

BTW, bus code shouldn't be calling dma_set_mask_and_coherent() on behalf
of its children, that is for the individual drivers to do, if and when
they intend to actually use DMA. Removing that here will save you
needing to fix the memory leak as well...

Thanks,
Robin.

2023-09-13 23:38:31

by Umang Jain

Subject: Re: [PATCH v10 1/5] staging: vc04_services: vchiq_arm: Add new bus type and device type

Hi Robin

On 9/13/23 5:38 PM, Robin Murphy wrote:
> On 2023-09-12 06:50, Umang Jain wrote:
> [...]
>>>> +struct vchiq_device *
>>>> +vchiq_device_register(struct device *parent, const char *name)
>>>> +{
>>>> +    struct vchiq_device *device;
>>>> +    int ret;
>>>> +
>>>> +    device = kzalloc(sizeof(*device), GFP_KERNEL);
>>>> +    if (!device) {
>>>> +        dev_err(parent, "Cannot register %s: Insufficient memory\n",
>>>> +            name);
>>>> +        return NULL;
>>>> +    }
>>>> +
>>>> +    device->dev.init_name = name;
>>>> +    device->dev.parent = parent;
>>>> +    device->dev.bus = &vchiq_bus_type;
>>>> +    device->dev.release = vchiq_device_release;
>>>> +
>>>> +    of_dma_configure(&device->dev, parent->of_node, true);
>>>> +    ret = dma_set_mask_and_coherent(&device->dev, DMA_BIT_MASK(32));
>>>> +    if (ret) {
>>>> +        dev_err(&device->dev, "32-bit DMA enable failed\n");
>>>> +        return NULL;
>>>> +    }
>>>
>>> Unfortunately the call of of_dma_configure() generates warnings likes
>>> this (Raspberry Pi 3A+ with multi_v7_defconfig + VCHIQ):
>>>
>>> [    9.206802] vchiq-bus bcm2835-audio: DMA mask not set
>>> [    9.206892] vchiq-bus bcm2835-camera: DMA mask not set
>>
>> huh, really weird, as on my RPi-3-b I get these set correctly and I
>> don't any such warning.
>
> Can you point to the code above where device->dev.dma_mask gets
> initialised between the initial kzalloc() and the call to
> of_dma_configure()? ;)
>
> BTW, bus code shouldn't be calling dma_set_mask_and_coherent() on
> behalf of its children, that is for the individual drivers to do, if
> and when they intend to actually use DMA. Removing that here will save
> you needing to fix the memory leak as well...

Thanks for this suggestion. I have now set the dma_mask within the child
itself!
>
> Thanks,
> Robin.