On Thu, 5 Mar 2020 at 16:23, Christoph Hellwig <hch@xxxxxx> wrote: > > On Thu, Mar 05, 2020 at 03:29:55PM +0100, Linus Walleij wrote: > > I agree. This relates to my reply to Greg K-H recently: > > https://lore.kernel.org/lkml/CACRpkdajhivkOkZ63v-hr7+6ObhTffYOx5uZP0P-MYvuVnyweA@xxxxxxxxxxxxxx/ > > > > The core of the problem is that drivers/of/platform.c is > > very simple and has no idea what kind of bus it is populating > > with devices from Device Tree. It is just guessing. > > > > For example platform.c contains this: > > > > dev->dev.coherent_dma_mask = DMA_BIT_MASK(32); > > if (!dev->dev.dma_mask) > > dev->dev.dma_mask = &dev->dev.coherent_dma_mask; > > > > And again part of the problem is that the device tree parser > > mostly just create struct platform_device's on the platform bus > > and has no real idea about any constraints on the bus where > > it will eventually end up after the platform device is probed and > > a new device on some other bus has been created. > > the plaform device code should allocate the dma_params for every > device. It should not set any limits in that, that is up to the > host drivers. That makes sense to me, but we also have other bus types to consider, like the amba bus for example. So, in principle, those buses that may get DMA capable devices attached, needs to do the allocation, that's what you are proposing? Kind regards Uffe
