On Sun, Dec 26, 2021 at 04:01:03PM +0200, Andy Shevchenko wrote:

On Sun, Dec 26, 2021 at 3:36 PM Gabriel L. Somlo [off-list ref] wrote:

quoted

On Sun, Dec 26, 2021 at 03:13:21PM +0200, Andy Shevchenko wrote:

quoted

On Sun, Dec 26, 2021 at 1:45 PM Gabriel L. Somlo [off-list ref] wrote:

quoted

On Sat, Dec 25, 2021 at 06:43:22PM +0200, Andy Shevchenko wrote:

quoted

On Wed, Dec 15, 2021 at 10:00 PM Gabriel Somlo [off-list ref] wrote:

...

quoted

quoted

quoted

quoted

quoted

+#ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT

Why under ifdeffery?

Because I only want to do it on 64-bit capable architectures.

The alternative would be to call

  dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));

on *all* architectures, but ignore the returned error (-EIO,
presumably on architetures that only support 32-bit DMA).

I don't understand why you are supposed to ignore errors and why you
expect to get such.

If I call `dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));`
on a machine where `CONFIG_ARCH_DMA_ADDR_T_64BIT` is *not* set, I
expect an error. The implicit default
(per Documentation/core-api/dma-api-howto.rst), is DMA_BIT_MASK(32).
I'm working under the impression that on machines with
CONFIG_ARCH_DMA_ADDR_T_64BIT I should increase that to DMA_BIT_MASK(64).

So if I don't #ifdef it, that call will fail on machines supporting
only 32-bits.

What am I missing?

This thread: https://lkml.org/lkml/2021/6/7/398 ?

OK, so just call `dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));`
regardless of 32- or 64-bit dma capability, *do* check the return value,
and it should *not* fail on 32-bit systems.

I'll do that in v6 (should go out in early January '22, since I'm traveling
with only occasional email access at the moment).

Thanks,
--Gabriel