On Wed, 19 Jul 2023 at 16:41, Matt Whitlock <kernel@xxxxxxxxxxxxxxxxx> wrote:
>
> Then that is my request. This entire complaint/discussion/argument would
> have been avoided if splice(2) had contained a sentence like this one from
> sendfile(2):
>
> "If out_fd refers to a socket or pipe with zero-copy support, callers must
> ensure the transferred portions of the file referred to by in_fd remain
> unmodified until the reader on the other end of out_fd has consumed the
> transferred data."
>
> That is a clear warning of the perils of the implementation under the hood,
> and it could/should be copied, more or less verbatim, to splice(2).
Ack. Internally in the kernel, the two really have always been more or
less of intermingled.
In fact, I think splice()/sendfile()/tee() could - and maybe should -
actually be a single man-page to make it clear that they are all
facets of the same thing.
The issues with TCP_CORK exist for splice too, for example, for
exactly the same reasons.
And while SPLICE_F_MORE exists, it only deals with multiple splice()
calls, it doesn't deal with the "I wrote a header before I even
started using splice()" case that is the one that is mentioned for
sendfile().
Or course, technically TCP_CORK exists for plain write() use as well,
but there the portable and historical fix is simply to use writev()
and send it all in one go.
So it's hopefully only when you use sendfile() and splice() that you
end up with "oh, but I have multiple different *kinds* of sources, and
I want to cork things until I've dealt with them all".
Linus
