On Mon, Oct 07, 2019 at 11:13:27AM -0700, Linus Torvalds wrote:
> On Mon, Oct 7, 2019 at 10:34 AM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> >
> > Tangentially related: copy_regster_to_user() and copy_regset_from_user().
>
> Not a worry. It's not performance-critical code, and if it ever is, it
> needs to be rewritten anyway.
>
> > The former variant tends to lead to few calls
> > of __copy_{to,from}_user(); the latter... On x86 it ends up doing
> > this:
>
> Just replace the __put_user() with a put_user() and be done with it.
> That code isn't acceptable, and if somebody ever complains about
> performance it's not the lack of __put_user that is the problem.
I wonder if it would be better off switching to several "copy in bulk"
like e.g. ppc does. That boilerplate with parallel "to/from kernel"
and "to/from userland" loops is asking for bugs - the calling
conventions like "pass kbuf and ubuf; exactly one must be NULL"
tend to be trouble, IME; I'm not saying we should just pass
struct iov_iter * instead of count+pos+kbuf+ubuf to ->get() and
->set(), but it might clean the things up nicely.
Let me look into that zoo a bit more...
