Hi Christian,

I have a question about this 40-bit restriction.

On 12/23/25 22:55, Christian König wrote:
> On 12/22/25 22:32, Arnd Bergmann wrote:
>> On Sat, Dec 20, 2025, at 17:33, Han Gao wrote:
>> [...]
>>> diff --git a/drivers/gpu/drm/radeon/radeon_irq_kms.c 
>>> b/drivers/gpu/drm/radeon/radeon_irq_kms.c
>>> index 9961251b44ba..62eb5a6968ff 100644
>>> --- a/drivers/gpu/drm/radeon/radeon_irq_kms.c
>>> +++ b/drivers/gpu/drm/radeon/radeon_irq_kms.c
>>> @@ -250,7 +250,7 @@ static bool radeon_msi_ok(struct radeon_device 
>>> *rdev)
>>>      * of address for "64-bit" MSIs which breaks on some platforms, 
>>> notably
>>>      * IBM POWER servers, so we limit them
>>>      */
>>> -   if (rdev->family < CHIP_BONAIRE) {
>>> +   if (rdev->family < CHIP_BONAIRE && !radeon_msi64) {
>>>             dev_info(rdev->dev, "radeon: MSI limited to 32-bit\n");
>>>             rdev->pdev->no_64bit_msi = 1;
>> According to the comment above it, the device can apparently
>> do 40-bit addressing but not use the entire 64-bit space.
>>
>> I assume the SG2042 chip has the irqchip somewhere above the
>> 32-bit line but below the 40-bit line, so it ends up working.
>>
>> I wonder if the msi_verify_entries() function should check
>> against dev->coherent_dma_mask instead of checking the
>> upper 32 bits for being nonzero, that probably gives you
>> the desired behavior.
> Again completely agree, that sounds like a plan to me.
>
> IIRC the modified code here is basically just a workaround because the MSI 
> control dword on older radeon HW was not setup correctly.

Does this mean that on Bonaire and onwards, MSI can reach full 64-bit
space, while DMA still only does 40-bit?
(drivers/gpu/drm/radeon/radeon_device.c sets DMA mask to at most 40 bits.)

If so, checking coherent_dma_mask would be wrong for those devices.

Do you think maybe it would be safer to introduce a msi_addr_mask for
occasions like these? We can have msi_addr_mask = DMA_BIT_MASK(40) for
pre-Bonaire, and then the ppc PCI stuff can check the mask and see if
it's usable. Probably something similar for hda.

Vivian "dramforever" Wang

Reply via email to