On Wed, 14 Sep 2016 04:16:48 +0100
Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> [Jens and Nick Cc'd]
>
> On Fri, Sep 09, 2016 at 07:06:29PM -0700, Linus Torvalds wrote:
>
> > I think you'd be better off with just a really small on-stack case
> > (like maybe 2-3 entries), and just allocate anything bigger
> > dynamically. Or you could even see how bad it is if you just
> > force-limit it to max 4 entries or something like that and just do
> > partial writes.
>
> Umm... Right now it tries to allocate as much as the output pipe could
> possibly hold. With default being 16 buffers, you'll end up with doing
> dynamic allocation in all cases (it doesn't even look at the amount of
> data we want to transfer).
>
> The situation with splice_pipe_desc looks very odd:
>
> * all but one instance are on stack frames of some ->splice_read()
> or something called by it (exception is in vmsplice)
>
> * all but one instance (a different one - see below) go through
> splice_grow_spd / splice_to_pipe / splice_shrink_spd sequence and
> nothing else sees them. The exception is skb_splice_bits() and there we
> have MAX_SKB_FRAGS for size, don't bother with grow/shrink and the only
> thing done to that spd is splice_to_pipe() (from the callback passed to
> skb_splice_bits()).
>
> * only one ->splice_read() instance does _not_ create
> splice_pipe_descriptor. It's fuse_dev_splice_read(), and it pays for that
> by open-coding splice_to_pipe(). The only reason for open-coding is that
> we don't have a "stronger SPLICE_F_NONBLOCK" that would fail if the data
> wouldn't fit. SPLICE_F_NONBLOCK stuffs as much as possible and buggers off
> without waiting, fuse_dev_splice_read() wants all or nothing (and no waiting).
>
> * incidentally, we *can't* add new flags - splice(2)/tee(2)/vmsplice(2)
> quietly ignore all bits they do not recognize. In fact, splice(2) ends up
> passing them (unsanitized) to ->splice_read and ->splice_write instances.
>
> * for splice(2) the IO size is limited by nominal capacity of output
> pipe. Looks fairly arbitrary (the limit is the same whether the pipe is
> full or empty), but I wouldn't be surprised if userland programmers would
> get unhappy if they have to take more iterations through their loops.
>
> * the other caller of ->splice_read() is splice_direct_to_actor() and
> that can be called on a fairly deep stack. However, there we loop ourselves
> and smaller chunk size is not a problem.
>
> * in case of skb_splice_bits(), we probably want a grow/shrink pair
> as well, with well below MAX_SKB_FRAGS for a default - what's the typical
> number of fragments per skb?
>
> > So feel free to try maxing out using only a small handful of
> > pipe_buffer entries. Returning partial IO from splice() is fine.
>
> Are you sure that nobody's growing the output pipe buffer before
> doing splice() into it as a way to reduce the amount of iterations?
>
> FWIW, I would love to replace these array of page * + array of
> <offset,len,private> triples with array of pipe_buffer; for one thing,
> this ridiculous ->sbd_release() goes away (we simply call ->ops->release()
> on all unwanted buffers), which gets rid of wonders like
> static void buffer_spd_release(struct splice_pipe_desc *spd, unsigned int i)
> {
> struct buffer_ref *ref =
> (struct buffer_ref *)spd->partial[i].private;
>
> if (--ref->ref)
> return;
>
> ring_buffer_free_read_page(ref->buffer, ref->page);
> kfree(ref);
> spd->partial[i].private = 0;
> }
> static void buffer_pipe_buf_release(struct pipe_inode_info *pipe,
> struct pipe_buffer *buf)
> {
> struct buffer_ref *ref = (struct buffer_ref *)buf->private;
>
> if (--ref->ref)
> return;
>
> ring_buffer_free_read_page(ref->buffer, ref->page);
> kfree(ref);
> buf->private = 0;
> }
>
> pairs that need to be kept in sync, etc.
>
> One inconvenience created by that is stuff like
> spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages,
> spd.pages);
> in there; granted, this one will go away with __generic_file_splice_read(),
> but e.g. get_iovec_page_array() is using get_user_pages_fast(), which wants
> to put pages next to each other. That one is from vmsplice_to_pipe() guts,
> and I've no idea what the normal use patterns are. OTOH, how much overhead
> would we get from repeated calls of get_user_pages_fast() for e.g. 16 pages
> or so, compared to larger chunks? It is on a shallow stack, so it's not
> as if we couldn't afford a 16-element array of struct page * in there...
Should not be so bad, but I don't have hard numbers for you. PAGEVEC_SIZE
is 14, and that's conceptually rather similar operation (walk radix tree;
grab pages). OTOH many archs are heavier and do locking and vmas walking etc.
Documentation/features/vm/pte_special/arch-support.txt
But even for those, at 16 entries, the bulk of the cost *should* be hitting
struct page cachelines and refcounting. The rest should mostly stay in cache.
|