Hi Arnd,
On 12/07/22 21:28, Michael Schmitz wrote:
I think you actually have to use platform_device_register_full() to
pass
a DMA mask here: As I understand it, the dma_set_mask_and_coherent()
call in the driver fails if the device is not already marked as dma
capable by having an initial mask set.
I'll take a look at that - if true, this requires the amiga-a3000-scsi
platform device set-up be changed in the same way (the gvp11 and a2091
drivers inherit the DMA mask from the Zorro bus default, so ought to
work OK).
I think we are good with using platform_device_register_simple():
setup_pdev_dma_masks() sets the DMA mask to 32 bit when allocating a
new platform device, and platform_device_register_full() only changes
that when passed in a non-zero mask in pdevinfo.
platform_device_register_simple() leaves the pdevinfo mask zero, so
the 32 bit default set in setup_pdev_dma_masks() survives.
Verified by having pata_falcon print the default DMA mask (0xffffffff),
then setting a 24 bit DMA mask and reading it back.
That did uncover an error I made in the gvp11 driver - the devices'
default DMA mask there is given as a 32 bit integer, but the DMA API
needs 64 bit. Will send a separate fix for that.
Cheers,
Michael
Cheers,
Michael
Cheers,
Michael
The way this normally works is that the device gets created with a mask
that reflects the capabilities of the bus, while the driver sets a mask
based on what it wants to program into the device, and the dma-mapping
interfaces ensure that we only use the intersection of those.
Arnd
