Re: [RFC PATCH] net/mlx4: Get rid of page operation after dma_alloc_coherent

[Jason Gunthorpe <jgg@ziepe.ca>]

On Tue, Dec 18, 2018 at 10:45:21AM -0700, Stephen Warren wrote:

> There are two allocation paths; one using dma_alloc_coherent and one using
> alloc_pages. (The hns driver only has the dma_alloc_coherent path.) These
> both store allocations into an sg list which is stored in a table, and that
> table is searched by a single function mlx4_table_find() irrespective of
> which allocation path was used, so if one of the allocation paths is updated
> to store the CPU virtual address differently, then both paths need to be
> updated so they match, so that the single table search path can continue to
> have a single implementation.

So the problem here is that on some arches

 sg_set_buf(sg, dma_coherent_buf, size)
 p = sg_virt(sg);
 assert(p == dma_coherent_buf);

Doesn't work or crashes? Is this how sgl should work?

But if you accept this and don't do the sg_set_buf then the
scatterlist is substantially broken, many of the APIs related to it
will not work as expected at all.

So, I don't think drivers should create such a broken scatterlist or
arches should not have this problem (ie the mathematical
transformation to struct page * and back to virtual address should
work in the coherent space even if there are no backing struct pages
allocated)?

What do you think CH?

Jason