Re: [PATCH 2/9] Move dma_ops from archdata into struct device

On Wed, 2017-01-11 at 07:48 +0100, Greg Kroah-Hartman wrote:

On Tue, Jan 10, 2017 at 04:56:41PM -0800, Bart Van Assche wrote:

quoted

Several RDMA drivers, e.g. drivers/infiniband/hw/qib, use the CPU to
transfer data between memory and PCIe adapter. Because of performance
reasons it is important that the CPU cache is not flushed when such
drivers transfer data. Make this possible by allowing these drivers to
override the dma_map_ops pointer. Additionally, introduce the function
set_dma_ops() that will be used by a later patch in this series.

=20
When you say things like "additionally", that's a huge flag that this
needs to be split up into multiple patches.=A0=A0No need to add
set_dma_ops() here in this patch.

Hello Greg,

Some architectures already define a=A0set_dma_ops() function. So what this
patch does is to move both the dma_ops pointer and the set_dma_ops()
function from architecture-specific to architecture independent code. I
don't think that it is possible to separate these two changes. But I
understand that how I formulated the patch description caused confusion. I
will rewrite the patch description to make it more clear before I repost
this patch series.

And I'd argue that it should be dma_ops_set(), and dma_ops_get(), just
to keep the namespace sane, but that's probably a different set of
patches...

Every time I rebase and retest this patch series on top of a new kernel
version I have to modify some of the patches to compensate for changes in
the architecture code. So I expect that once Linus merges these patches tha=
t
he will have to resolve one or more merge conflicts. Including a rename of
the functions that query and set the dma_ops pointer in this patch series
would increase the number of merge conflicts triggered by this patch series
and would make Linus' job harder. So I hope that you will allow me to
postpone that rename until a later time ...

Bart.=