On Thu, Nov 29, 2018 at 10:31 AM Christoph Hellwig <hch@xxxxxx> wrote:
>
> > Or, better yet, plan on removing the single-page dma mappign entirely
> > at a later date, and make the issue moot.
>
> What would be the replacement? Build a S/G list for every single page
> mapping? Not sure that would create a lot of happy campers..
It's what we ended up doing with some other cases, and it didn't
really end up hurting as much as I thought it would.
I'm thinking of the vfs functions that end up turning "buf, len" into
struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = len };
and then passing it around as a single-entry iov instead (not even
that - they end up being an iov_iter, which is not just the iov, but
the whole "what _kind_ of iov" indirection)
Maybe a very similar model could be used for just simplifying the core
dma mapping setup: sure, people will want to do single-area dma, but
how bad would it be to just turn them into single-entry SG lists on
stack, and then the dma-maping internally would just always see that?
Most of the high-performance IO is already using SG lists anyway, no?
Disk/networking/whatever.
But just an idea. And the "map_sg()" error handling isn't actually any
better, I think. It returns zero on error, no? So it's not improving
the error handling.
The whole dma-mapping layer seems full of those kinds of "inspired
error handling choices" ;)
Linus
