Re: [PATCH 22/23] fs: default to generic_file_splice_read for files having ->read_iter

[Christoph Hellwig]

On Thu, Jul 30, 2020 at 01:05:44AM +0100, Al Viro wrote:
> On Tue, Jul 07, 2020 at 07:48:00PM +0200, Christoph Hellwig wrote:
> > If a file implements the ->read_iter method, the iter based splice read
> > works and is always preferred over the ->read based one.  Use it by
> > default in do_splice_to and remove all the direct assignment of
> > generic_file_splice_read to file_operations.
> 
> The worst problem here is the assumption that all ->read_iter() instances
> will take pipe-backed destination; that's _not_ automatically true.
> In particular, it's almost certainly false for tap_read_iter() (as
> well as tun_chr_read_iter() in IFF_VNET_HDR case).
> 
> Other potentially interesting cases: cuse and hugetlbfs.
> 
> But in any case, that blind assertion ("iter based splice read works")
> really needs to be backed by something.

I think we need to fix that in the instances, as we really expect
->splice_read to just work instead of the caller knowing what could
work and what might not.