Re: [PATCH] mmc: sdhci-tegra: Set DMA mask

From: Alexandre Courbot
Date: Fri Feb 26 2016 - 02:24:59 EST

On Thu, Feb 25, 2016 at 11:52 PM, Arnd Bergmann <arnd@xxxxxxxx> wrote:
>> Actually even if we specify a dma-ranges on the parent DT node, the
>> DMA range will still be limited to 32 bits because of the following
>> code in of_dma_configure():
>> /*
>> * Set default coherent_dma_mask to 32 bit. Drivers are expected to
>> * setup the correct supported mask.
>> */
>> if (!dev->coherent_dma_mask)
>> dev->coherent_dma_mask = DMA_BIT_MASK(32);
>> /*
>> * Set it to coherent_dma_mask by default if the architecture
>> * code has not set it.
>> */
>> if (!dev->dma_mask)
>> dev->dma_mask = &dev->coherent_dma_mask;
>> ....
>> /* gets dma-ranges into dma_addr and size */
>> ....
>> *dev->dma_mask = min((*dev->dma_mask),
>> DMA_BIT_MASK(ilog2(dma_addr + size)));
>> So unless the DMA mask is set on the device before of_dma_configure()
>> is called, the min() statement will choose the 32 bits mask that has
>> been previously set. So IIUC in any case, the driver will need to call
>> dma_set_mask()
> Yes, the driver definitely has to call dma_set_mask(), the property of
> the parent bus is used to make that fail when the bus doesn't support
> it.

And that's where things seem to stop working: the driver's probe
function is invoked by the platform bus, *after* of_dma_configure() is
called. So unless I am missing something there is no way for the
driver to set the DMA mask in such a way that of_dma_configure() can
see it and do the right thing.

In other words, most of the DMA mask logic in of_dma_configure()
doesn't seem to have any effect (and a 32 bits mask will be set), at
least on the platform bus.

>> Can I have your thoughts on this? Am I missing something?
> One point: I think the dma_set_mask() probably should be in the
> generic part of the sdhci driver, not the tegra specific portion.

Ok, but then how does the generic part of the driver knows which DMA
mask applies to the device?

> I also forget how this really needs to interact with swiotlb. I know
> we have discussed this a couple of times, but the result currently
> is lost to me.
> Maybe the answer was that if swiotlb or iommu are enabled, then
> dma_set_mask() should always succeed, but the mask should not actually
> be updated?