Re: [PATCH v7 05/10] mfd: sec: set DMA coherent mask

["Kaustabh Chakraborty" <kauschluss@disroot.org>]

On 2026-05-15 22:24 +00:00, sashiko-bot wrote:
> Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
> - [Medium] Shared dma_mask pointer across MFD child devices causes a data race and violates DMA API contracts.
> --
>
> commit cc615ed9ba9013f030c19361a7a4d79aad2e467d
> Author: Kaustabh Chakraborty <kauschluss@disroot.org>
>
> mfd: sec: set DMA coherent mask
>
> This commit manually sets the coherent DMA mask to 0 and points the device's
> dma_mask to it. This is intended to suppress "DMA mask not set" warnings
> for sub-devices, as the device does not use DMA for communication.
>
>> diff --git a/drivers/mfd/sec-common.c b/drivers/mfd/sec-common.c
>> index 22f6c74eb6c0e..fe92bc4a3dd26 100644
>> --- a/drivers/mfd/sec-common.c
>> +++ b/drivers/mfd/sec-common.c
>> @@ -221,6 +221,9 @@ int sec_pmic_probe(struct device *dev, int device_type, unsigned int irq,
>>  	if (IS_ERR(irq_data))
>>  		return PTR_ERR(irq_data);
>>  
>> +	dev->coherent_dma_mask = 0;
>> +	dev->dma_mask = &dev->coherent_dma_mask;
>
> When the MFD core adds child devices via mfd_add_device(), it copies the
> parent's dma_mask pointer to the child devices:
>
>     pdev->dev.dma_mask = parent->dma_mask;
>
> Because sub-devices for this PMIC probe asynchronously, they can concurrently
> execute platform_dma_configure(), which eventually calls of_dma_configure_id()
> and performs a lockless read-modify-write:
>
>     *dev->dma_mask &= mask;
>
> Since all child devices share the same dma_mask pointer referencing the
> parent's coherent_dma_mask, does this cause an unsynchronized data race on
> the parent's memory?
>
> Additionally, sharing a single dma_mask pointer across multiple devices breaks
> the driver model. If any child device were to legitimately call dma_set_mask(),
> wouldn't it silently overwrite the DMA mask for all sibling devices?
>
> Finally, manually modifying device structures to fake an invalid mask of 0
> bypasses DMA subsystem validation. Helpers like dma_coerce_mask_and_coherent()
> explicitly reject a mask of 0. Is manually assigning 0 here the recommended way
> to suppress the warning?

Note that change follows the same reasoning as in commit 0b7cbe811ca5
("mfd: cpcap: Fix cpcap dmamask not set warnings").

>> +
>>  	pm_runtime_set_active(sec_pmic->dev);
>>  
>>  	switch (sec_pmic->device_type) {