Re: [RFC PATCH] m68k: set dma and coherent masks for Macintosh SONIC based ethernet

[Finn Thain <fthain@telegraphics.com.au>]

On Mon, 28 May 2018, Geert Uytterhoeven wrote:

> 
> Do we have a consensus on the way forward?

My prefered solution remains the two driver patches that I originally 
submitted, which you objected to:

https://lkml.org/lkml/2018/5/3/10
https://lkml.org/lkml/2018/5/3/9

So there is no consensus yet.

To my mind, the existing code is pretty clear: drivers should set the 
(default) mask. See also,

$ egrep -r -A1 "coherent_dma_mask.*expected" */
drivers/of/device.c:     * Set default coherent_dma_mask to 32 bit.  Drivers are expected to
drivers/of/device.c-     * setup the correct supported mask.
--
drivers/acpi/arm64/iort.c:       * Set default coherent_dma_mask to 32 bit.  Drivers are expected to
drivers/acpi/arm64/iort.c-       * setup the correct supported mask.
$ 

And drivers/usb/dwc2/platform.c:

        /*
         * Use reasonable defaults so platforms don't have to provide these.
         */
        if (!dev->dev.dma_mask)
                dev->dev.dma_mask = &dev->dev.coherent_dma_mask;
        retval = dma_set_coherent_mask(&dev->dev, DMA_BIT_MASK(32));

And FWIW,

$ egrep -wlr "dma_set_mask_and_coherent|dma_set_coherent_mask|dma_coerce_mask_and_coherent" drivers/ | wc -l
196

I suppose that a driver should avoid lengthening an existing device mask.

Since an arch gets to apply limits in the dma ops it implements, why would 
arch code also have to set a limit in the form of default platform device 
masks? Powerpc seems to be the only arch that does this.

The same line of reasoning suggests that the problematic WARN_ON should 
not appear in include/linux/ in the first place. If it is needed by 
certain architectures, it should be in arch/x.

I would send a patch to revert commit 205e1b7f51e4 ("dma-mapping: warn 
when there is no coherent_dma_mask") if I thought that arch code was not 
somehow relying on it. But I'll leave that up to Chrisoph.

--