On 8/31/22 17:42, Al Viro wrote: > On Tue, Aug 30, 2022 at 09:18:40PM -0700, John Hubbard wrote: >> Provide two new wrapper routines that are intended for user space pages >> only: >> >> iov_iter_pin_pages() >> iov_iter_pin_pages_alloc() >> >> Internally, these routines call pin_user_pages_fast(), instead of >> get_user_pages_fast(), for user_backed_iter(i) and iov_iter_bvec(i) >> cases. >> >> As always, callers must use unpin_user_pages() or a suitable FOLL_PIN >> variant, to release the pages, if they actually were acquired via >> pin_user_pages_fast(). >> >> This is a prerequisite to converting bio/block layers over to use >> pin_user_pages_fast(). > > What of ITER_PIPE (splice from O_DIRECT fd to a to pipe, for filesystem > that uses generic_file_splice_read())? Yes. And it turns out that I sent this v2 just a little too early: it does not include Jan Kara's latest idea [1] of including ITER_PIPE and ITER_XARRAY. That should fix this up. [1] https://lore.kernel.org/r/20220831094349.boln4jjajkdtykx3@quack3 thanks, -- John Hubbard NVIDIA
