drivers using the non-PCI dma_set_mask() on PCI devices

drivers using the non-PCI dma_set_mask() on PCI devices

[Mikael Pettersson]

Several drivers do the following:

	struct pci_dev *pdev = ...;
	...
	if (dma_set_mask(&pdev->dev, mask))
		...

But pdev->dev.dma_mask == &pdev->dma_mask, so this is essentially a
roundabout way of saying

	if (pci_set_dma_mask(pdev, mask))

except that it bypasses the PCI-specific operations pci_set_dma_mask()
may do on that platform.

Drivers doing this include drivers/scsi/aic7xxx/aic79xx_osm_pci.c,
drivers/scsi/aic7xxxx/aic7xxx_osm_pci.c, drivers/scsi/qla2xxx/qla_os.c,
drivers/net/wireless/rt2x00/rt2x00pci.c, and drivers/media/video/meye.c.

Is it considered acceptable that drivers bypass the PCI DMA API on
PCI devices like this, or are these drivers in error?

I'm doing some work on an embedded platform (ARM IXP4xx) with some
PCI DMA restrictions. To handle these the platform provides its own
versions of pci_set_dma_mask() and pci_set_consistent_dma_mask(),
but its dma_set_mask() currently does not do anything PCI-specific.
The question is: should dma_set_mask() have PCI knowledge or not?

/Mikael

Re: drivers using the non-PCI dma_set_mask() on PCI devices

[Robert Hancock]

Mikael Pettersson wrote:
> Several drivers do the following:
> 
> 	struct pci_dev *pdev = ...;
> 	...
> 	if (dma_set_mask(&pdev->dev, mask))
> 		...
> 
> But pdev->dev.dma_mask == &pdev->dma_mask, so this is essentially a
> roundabout way of saying
> 
> 	if (pci_set_dma_mask(pdev, mask))
> 
> except that it bypasses the PCI-specific operations pci_set_dma_mask()
> may do on that platform.
> 
> Drivers doing this include drivers/scsi/aic7xxx/aic79xx_osm_pci.c,
> drivers/scsi/aic7xxxx/aic7xxx_osm_pci.c, drivers/scsi/qla2xxx/qla_os.c,
> drivers/net/wireless/rt2x00/rt2x00pci.c, and drivers/media/video/meye.c.
> 
> Is it considered acceptable that drivers bypass the PCI DMA API on
> PCI devices like this, or are these drivers in error?
> 
> I'm doing some work on an embedded platform (ARM IXP4xx) with some
> PCI DMA restrictions. To handle these the platform provides its own
> versions of pci_set_dma_mask() and pci_set_consistent_dma_mask(),
> but its dma_set_mask() currently does not do anything PCI-specific.
> The question is: should dma_set_mask() have PCI knowledge or not?

AFAIK pci_set_dma_mask is somewhat deprecated, and dma_set_mask should 
be used instead. If the platform fails to do what's needed when 
dma_set_mask is called on a PCI device then it would seem the platform 
code is in error.