Re: [RFC][PATCHSET] iov_iter work

[Al Viro]

On Mon, Jun 07, 2021 at 04:35:46PM -0700, Linus Torvalds wrote:
> On Mon, Jun 7, 2021 at 3:01 PM Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> >
> >  (b) on all the common non-SET_FS architectures, kernel threads using
> > iov_iter_init() wouldn't work anyway, because on those architectures
> > it would always fill the thing in with an iov, not a kvec.
> 
> Thinking more about this thing, I think it means that what we *should*
> do is simply just
> 
>   void iov_iter_init(struct iov_iter *i, unsigned int direction,
>                         const struct iovec *iov, unsigned long nr_segs,
>                         size_t count)
>   {
>         WARN_ON_ONCE(direction & ~(READ | WRITE));
>         iWARN_ON_ONCE(uaccess_kernel());
>         *i = (struct iov_iter) {
>                 .iter_type = ITER_IOVEC,
>                 .data_source = direction,
>                 .iov = iov,
>                 .nr_segs = nr_segs,
>                 .iov_offset = 0,
>                 .count = count
>         };
>   }
> 
> because filling it with a kvec is simply wrong. It's wrong exactly due
> to the fact that *if* we have a kernel thread, all the modern
> non-SET_FS architectures will just ignore that entirely, and always
> use the iov meaning.

Updated and pushed out...