On Wed, Apr 03, 2019 at 07:26:40PM +0100, Russell King - ARM Linux admin wrote:

On Thu, Mar 21, 2019 at 03:52:28PM -0700, Christoph Hellwig wrote:

quoted

We still have a few drivers which pass a NULL struct device pointer
to DMA API functions, which generally is a bad idea as the API
implementations rely on the device not only for ops selection, but
also the dma mask and various other attributes, and many implementations
have been broken for NULL device support for a while.

I think I must be missing something, but...

My understanding is that ISA DMA is normally limited to 24 bits of
address

Yes.

- indeed, the x86 version only programs 24 bits of DMA address.
Looking through this series, it appears that the conversions mean that
the DMA mask for ISA becomes the full all-ones DMA mask, which would
of course lead to memory corruption if only 24 bits of the address end
up being programmed into the hardware.

In the generic dma mapping code no struct device has always meant a
32-bit DMA mask - take a look at the dma_get_mask() function.

Maybe you could say why you think this series is safe in regard to ISA
DMA?

ISA DMA has always been rather painful in a myriad of ways, and the
DMA API so far hasn't helped, given that we don't do bounce buffering
for the 24-bit limit, but just the higher limits.  So far even if you
do use the DMA API and pass a device ISA DMA so far always meant
that the higher layers had to assure things are addressable, either
by using GFP_DMA allocation in the drivers, or mid-layer hacks like
the unchecked_isa_dma flag in SCSI and/or BLK_BOUNCE_ISA in the
block layer.

This series doesn't change those facts at all.  I have some half
started series to clean some of this up but it isn't high up on
the priority list.