Re: splice() from /dev/zero to a pipe does not work (5.9+)

[Linus Torvalds <torvalds@linux-foundation.org>]

On Fri, May 7, 2021 at 11:21 AM Kees Cook <keescook@chromium.org> wrote:
>
> So the question is likely, "do we want this for /dev/zero?"

Well, /dev/zero should at least be safe, and I guess it's actually
interesting from a performance testing standpoint (ie useful for some
kind of "what is the overhead of the splice code with no data copy").

So I'll happily take a sane patch for /dev/zero, although I think it
probably only makes sense if it's made to use the zero page explicitly
(ie exactly for that "no data copy testing" case).

So very much *not* using generic_file_splice_read(), even if that
might be the one-liner.

/dev/zero should probably also use the (already existing)
splice_write_null() function for the .splice_write case.

Anybody willing to look into this? My gu feel is that it *should* be easy to do.

That said - looking at the current 'pipe_zero()', it uses
'push_pipe()' to actually allocation regular pages, and then clear
them.

Which is basically what a generic_file_splice_read() would do, and it
feels incredibly pointless and stupid to me.

I *think* we should be able to just do something like

    len = size;
    while (len > 0) {
        struct pipe_buffer *buf;
        unsigned int tail = pipe->tail;
        unsigned int head = pipe->head;
        unsigned int mask = pipe->ring_size - 1;

        if (pipe_full(head, tail, pipe->max_usage))
            break;
        buf = &pipe->bufs[iter_head & p_mask];
        buf->ops = &zero_pipe_buf_ops;
        buf->page = ZERO_PAGE(0);
        buf->offset = 0;
        buf->len = min_t(ssize_t, len, PAGE_SIZE);
        len -= buf->len;
        pipe->head = head+1;
    }
    return size - len;

but honestly, I haven't thought a lot about it.

Al? This is another of those "right up your alley" things.

Maybe it's not worth it, and just using generic_file_splice_read() is
the way to go, but I do get the feeling that if we are splicing
/dev/null, the whole _point_ of it is about benchmarking, not "make it
work".

            Linus