Re: [PATCH] use a nocache copy for bvecs in copy_from_iter_nocache()

["Kani, Toshimitsu" <toshi.kani@hpe.com>]

On Fri, 2016-10-07 at 18:08 +0100, Al Viro wrote:
> On Fri, Oct 07, 2016 at 10:55:11AM -0500, Brian Boylston wrote:
> > 
> > copy_from_iter_nocache() is only "nocache" for iovecs.  Enhance it
> > to also use a nocache copy for bvecs.  This improves performance by
> > 2-3X when splice()ing to a file in a DAX-mounted, pmem-backed file
> > system.
> 
> > 
> > +static void memcpy_from_page_nocache(char *to, struct page *page,
> > size_t offset, size_t len)
> > +{
> > +	char *from = kmap_atomic(page);
> > +	__copy_from_user_inatomic_nocache(to, from, len);
> > +	kunmap_atomic(from);
> > +}
> 
> At the very least, it will blow up on any architecture with split
> userland and kernel MMU contexts.  You *can't* feed a kernel pointer
> to things like that and expect it to work.  At the very least, you
> need to add memcpy_nocache() and have it default to memcpy(), with
> non-dummy version on x86.  And use _that_, rather than messing with
> __copy_from_user_inatomic_nocache()

Good point.  I think we can add memcpy_nocache() which calls
__copy_from_user_inatomic_nocache() on x86 and defauts to memcpy() on
other architectures.

Thanks,
-Toshi