On Tue, Mar 28, 2023 at 1:38 PM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
>
> ... and all that, er, cleverness - for the sake of a single piece of shit
> driver for piece of shit hardware *and* equally beautiful ABI.
Now, I wish we didn't have any of those 'walk the iov{} array', but
sadly we do, and it's not just a single case.
It's also that pcm_native.c case, it's the qib rdma driver.
So if we didn't have people walking the iov[] array, I'd hate to add this.
But considering that we *do* have people walking the iov[] array, I'd
rather unify the two user-mode cases than have them do the whole "do
two different things for the ITER_UBUF vs ITER_IOV case".
> Is it really worth bothering with? And if anyone has doubts about the
> inturdprize kwality of the ABI in question, I suggest taking a look at
> hfi1_user_sdma_process_request() - that's where the horrors are.
Yes. I started out my email to Jens by suggesting that instead of
passing down the iov[] pointer, he should just pass down the iter
instead.
And then I looked at that code and went "yeah, no way do I want to touch it".
Which then got me to that "could we at least *unify* these two cases".
Linus
